1f5166768STheodore Ts'o // SPDX-License-Identifier: GPL-2.0 2c9de560dSAlex Tomas /* 3c9de560dSAlex Tomas * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com 4c9de560dSAlex Tomas * Written by Alex Tomas <alex@clusterfs.com> 5c9de560dSAlex Tomas */ 6c9de560dSAlex Tomas 7c9de560dSAlex Tomas 8c9de560dSAlex Tomas /* 9c9de560dSAlex Tomas * mballoc.c contains the multiblocks allocation routines 10c9de560dSAlex Tomas */ 11c9de560dSAlex Tomas 1218aadd47SBobi Jam #include "ext4_jbd2.h" 138f6e39a7SMingming Cao #include "mballoc.h" 1428623c2fSTheodore Ts'o #include <linux/log2.h> 15a0b30c12STheodore Ts'o #include <linux/module.h> 165a0e3ad6STejun Heo #include <linux/slab.h> 171a5d5e5dSJeremy Cline #include <linux/nospec.h> 1866114cadSTejun Heo #include <linux/backing-dev.h> 195229a658SJan Kara #include <linux/freezer.h> 209bffad1eSTheodore Ts'o #include <trace/events/ext4.h> 219bffad1eSTheodore Ts'o 22c9de560dSAlex Tomas /* 23c9de560dSAlex Tomas * MUSTDO: 24c9de560dSAlex Tomas * - test ext4_ext_search_left() and ext4_ext_search_right() 25c9de560dSAlex Tomas * - search for metadata in few groups 26c9de560dSAlex Tomas * 27c9de560dSAlex Tomas * TODO v4: 28c9de560dSAlex Tomas * - normalization should take into account whether file is still open 29c9de560dSAlex Tomas * - discard preallocations if no free space left (policy?) 30c9de560dSAlex Tomas * - don't normalize tails 31c9de560dSAlex Tomas * - quota 32c9de560dSAlex Tomas * - reservation for superuser 33c9de560dSAlex Tomas * 34c9de560dSAlex Tomas * TODO v3: 35c9de560dSAlex Tomas * - bitmap read-ahead (proposed by Oleg Drokin aka green) 36c9de560dSAlex Tomas * - track min/max extents in each group for better group selection 37c9de560dSAlex Tomas * - mb_mark_used() may allocate chunk right after splitting buddy 38c9de560dSAlex Tomas * - tree of groups sorted by number of free blocks 39c9de560dSAlex Tomas * - error handling 40c9de560dSAlex Tomas */ 41c9de560dSAlex Tomas 42c9de560dSAlex Tomas /* 43c9de560dSAlex Tomas * The allocation request involve request for multiple number of blocks 44c9de560dSAlex Tomas * near to the goal(block) value specified. 45c9de560dSAlex Tomas * 46b713a5ecSTheodore Ts'o * During initialization phase of the allocator we decide to use the 47b713a5ecSTheodore Ts'o * group preallocation or inode preallocation depending on the size of 48b713a5ecSTheodore Ts'o * the file. The size of the file could be the resulting file size we 49b713a5ecSTheodore Ts'o * would have after allocation, or the current file size, which ever 50b713a5ecSTheodore Ts'o * is larger. If the size is less than sbi->s_mb_stream_request we 51b713a5ecSTheodore Ts'o * select to use the group preallocation. The default value of 52b713a5ecSTheodore Ts'o * s_mb_stream_request is 16 blocks. This can also be tuned via 53b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 54b713a5ecSTheodore Ts'o * terms of number of blocks. 55c9de560dSAlex Tomas * 56c9de560dSAlex Tomas * The main motivation for having small file use group preallocation is to 57b713a5ecSTheodore Ts'o * ensure that we have small files closer together on the disk. 58c9de560dSAlex Tomas * 59b713a5ecSTheodore Ts'o * First stage the allocator looks at the inode prealloc list, 60b713a5ecSTheodore Ts'o * ext4_inode_info->i_prealloc_list, which contains list of prealloc 61b713a5ecSTheodore Ts'o * spaces for this particular inode. The inode prealloc space is 62b713a5ecSTheodore Ts'o * represented as: 63c9de560dSAlex Tomas * 64c9de560dSAlex Tomas * pa_lstart -> the logical start block for this prealloc space 65c9de560dSAlex Tomas * pa_pstart -> the physical start block for this prealloc space 6653accfa9STheodore Ts'o * pa_len -> length for this prealloc space (in clusters) 6753accfa9STheodore Ts'o * pa_free -> free space available in this prealloc space (in clusters) 68c9de560dSAlex Tomas * 69c9de560dSAlex Tomas * The inode preallocation space is used looking at the _logical_ start 70c9de560dSAlex Tomas * block. If only the logical file block falls within the range of prealloc 71caaf7a29STao Ma * space we will consume the particular prealloc space. This makes sure that 72caaf7a29STao Ma * we have contiguous physical blocks representing the file blocks 73c9de560dSAlex Tomas * 74c9de560dSAlex Tomas * The important thing to be noted in case of inode prealloc space is that 75c9de560dSAlex Tomas * we don't modify the values associated to inode prealloc space except 76c9de560dSAlex Tomas * pa_free. 77c9de560dSAlex Tomas * 78c9de560dSAlex Tomas * If we are not able to find blocks in the inode prealloc space and if we 79c9de560dSAlex Tomas * have the group allocation flag set then we look at the locality group 80caaf7a29STao Ma * prealloc space. These are per CPU prealloc list represented as 81c9de560dSAlex Tomas * 82c9de560dSAlex Tomas * ext4_sb_info.s_locality_groups[smp_processor_id()] 83c9de560dSAlex Tomas * 84c9de560dSAlex Tomas * The reason for having a per cpu locality group is to reduce the contention 85c9de560dSAlex Tomas * between CPUs. It is possible to get scheduled at this point. 86c9de560dSAlex Tomas * 87c9de560dSAlex Tomas * The locality group prealloc space is used looking at whether we have 8825985edcSLucas De Marchi * enough free space (pa_free) within the prealloc space. 89c9de560dSAlex Tomas * 90c9de560dSAlex Tomas * If we can't allocate blocks via inode prealloc or/and locality group 91c9de560dSAlex Tomas * prealloc then we look at the buddy cache. The buddy cache is represented 92c9de560dSAlex Tomas * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 93c9de560dSAlex Tomas * mapped to the buddy and bitmap information regarding different 94c9de560dSAlex Tomas * groups. The buddy information is attached to buddy cache inode so that 95c9de560dSAlex Tomas * we can access them through the page cache. The information regarding 96c9de560dSAlex Tomas * each group is loaded via ext4_mb_load_buddy. The information involve 97c9de560dSAlex Tomas * block bitmap and buddy information. The information are stored in the 98c9de560dSAlex Tomas * inode as: 99c9de560dSAlex Tomas * 100c9de560dSAlex Tomas * { page } 101c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 102c9de560dSAlex Tomas * 103c9de560dSAlex Tomas * 104c9de560dSAlex Tomas * one block each for bitmap and buddy information. So for each group we 105ea1754a0SKirill A. Shutemov * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / 106c9de560dSAlex Tomas * blocksize) blocks. So it can have information regarding groups_per_page 107c9de560dSAlex Tomas * which is blocks_per_page/2 108c9de560dSAlex Tomas * 109c9de560dSAlex Tomas * The buddy cache inode is not stored on disk. The inode is thrown 110c9de560dSAlex Tomas * away when the filesystem is unmounted. 111c9de560dSAlex Tomas * 112c9de560dSAlex Tomas * We look for count number of blocks in the buddy cache. If we were able 113c9de560dSAlex Tomas * to locate that many free blocks we return with additional information 114c9de560dSAlex Tomas * regarding rest of the contiguous physical block available 115c9de560dSAlex Tomas * 116c9de560dSAlex Tomas * Before allocating blocks via buddy cache we normalize the request 117c9de560dSAlex Tomas * blocks. This ensure we ask for more blocks that we needed. The extra 118c9de560dSAlex Tomas * blocks that we get after allocation is added to the respective prealloc 119c9de560dSAlex Tomas * list. In case of inode preallocation we follow a list of heuristics 120c9de560dSAlex Tomas * based on file size. This can be found in ext4_mb_normalize_request. If 121c9de560dSAlex Tomas * we are doing a group prealloc we try to normalize the request to 12227baebb8STheodore Ts'o * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is 12327baebb8STheodore Ts'o * dependent on the cluster size; for non-bigalloc file systems, it is 124c9de560dSAlex Tomas * 512 blocks. This can be tuned via 125d7a1fee1SDan Ehrenberg * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in 126c9de560dSAlex Tomas * terms of number of blocks. If we have mounted the file system with -O 127c9de560dSAlex Tomas * stripe=<value> option the group prealloc request is normalized to the 128b483bb77SRandy Dunlap * smallest multiple of the stripe value (sbi->s_stripe) which is 129d7a1fee1SDan Ehrenberg * greater than the default mb_group_prealloc. 130c9de560dSAlex Tomas * 131196e402aSHarshad Shirwadkar * If "mb_optimize_scan" mount option is set, we maintain in memory group info 132196e402aSHarshad Shirwadkar * structures in two data structures: 133196e402aSHarshad Shirwadkar * 134196e402aSHarshad Shirwadkar * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders) 135196e402aSHarshad Shirwadkar * 136196e402aSHarshad Shirwadkar * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks) 137196e402aSHarshad Shirwadkar * 138196e402aSHarshad Shirwadkar * This is an array of lists where the index in the array represents the 139196e402aSHarshad Shirwadkar * largest free order in the buddy bitmap of the participating group infos of 140196e402aSHarshad Shirwadkar * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total 141196e402aSHarshad Shirwadkar * number of buddy bitmap orders possible) number of lists. Group-infos are 142196e402aSHarshad Shirwadkar * placed in appropriate lists. 143196e402aSHarshad Shirwadkar * 14483e80a6eSJan Kara * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size) 145196e402aSHarshad Shirwadkar * 14683e80a6eSJan Kara * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks) 147196e402aSHarshad Shirwadkar * 14883e80a6eSJan Kara * This is an array of lists where in the i-th list there are groups with 14983e80a6eSJan Kara * average fragment size >= 2^i and < 2^(i+1). The average fragment size 15083e80a6eSJan Kara * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments. 15183e80a6eSJan Kara * Note that we don't bother with a special list for completely empty groups 15283e80a6eSJan Kara * so we only have MB_NUM_ORDERS(sb) lists. 153196e402aSHarshad Shirwadkar * 154196e402aSHarshad Shirwadkar * When "mb_optimize_scan" mount option is set, mballoc consults the above data 155196e402aSHarshad Shirwadkar * structures to decide the order in which groups are to be traversed for 156196e402aSHarshad Shirwadkar * fulfilling an allocation request. 157196e402aSHarshad Shirwadkar * 158f52f3d2bSOjaswin Mujoo * At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order 159f52f3d2bSOjaswin Mujoo * >= the order of the request. We directly look at the largest free order list 160f52f3d2bSOjaswin Mujoo * in the data structure (1) above where largest_free_order = order of the 161f52f3d2bSOjaswin Mujoo * request. If that list is empty, we look at remaining list in the increasing 162f52f3d2bSOjaswin Mujoo * order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED 163f52f3d2bSOjaswin Mujoo * lookup in O(1) time. 164196e402aSHarshad Shirwadkar * 165f52f3d2bSOjaswin Mujoo * At CR_GOAL_LEN_FAST, we only consider groups where 166f52f3d2bSOjaswin Mujoo * average fragment size > request size. So, we lookup a group which has average 167f52f3d2bSOjaswin Mujoo * fragment size just above or equal to request size using our average fragment 168f52f3d2bSOjaswin Mujoo * size group lists (data structure 2) in O(1) time. 169196e402aSHarshad Shirwadkar * 170f52f3d2bSOjaswin Mujoo * At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied 171f52f3d2bSOjaswin Mujoo * in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in 172f52f3d2bSOjaswin Mujoo * CR_GOAL_LEN_FAST suggests that there is no BG that has avg 173f52f3d2bSOjaswin Mujoo * fragment size > goal length. So before falling to the slower 174f52f3d2bSOjaswin Mujoo * CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and 175f52f3d2bSOjaswin Mujoo * then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big 176f52f3d2bSOjaswin Mujoo * enough average fragment size. This increases the chances of finding a 177f52f3d2bSOjaswin Mujoo * suitable block group in O(1) time and results in faster allocation at the 178f52f3d2bSOjaswin Mujoo * cost of reduced size of allocation. 1797e170922SOjaswin Mujoo * 180196e402aSHarshad Shirwadkar * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in 181f52f3d2bSOjaswin Mujoo * linear order which requires O(N) search time for each CR_POWER2_ALIGNED and 182f52f3d2bSOjaswin Mujoo * CR_GOAL_LEN_FAST phase. 183196e402aSHarshad Shirwadkar * 184d7a1fee1SDan Ehrenberg * The regular allocator (using the buddy cache) supports a few tunables. 185c9de560dSAlex Tomas * 186b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_min_to_scan 187b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_max_to_scan 188b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req 189196e402aSHarshad Shirwadkar * /sys/fs/ext4/<partition>/mb_linear_limit 190c9de560dSAlex Tomas * 191b713a5ecSTheodore Ts'o * The regular allocator uses buddy scan only if the request len is power of 192c9de560dSAlex Tomas * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 193c9de560dSAlex Tomas * value of s_mb_order2_reqs can be tuned via 194b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 195af901ca1SAndré Goddard Rosa * stripe size (sbi->s_stripe), we try to search for contiguous block in 196b713a5ecSTheodore Ts'o * stripe size. This should result in better allocation on RAID setups. If 197b713a5ecSTheodore Ts'o * not, we search in the specific group using bitmap for best extents. The 198b713a5ecSTheodore Ts'o * tunable min_to_scan and max_to_scan control the behaviour here. 199c9de560dSAlex Tomas * min_to_scan indicate how long the mballoc __must__ look for a best 200b713a5ecSTheodore Ts'o * extent and max_to_scan indicates how long the mballoc __can__ look for a 201c9de560dSAlex Tomas * best extent in the found extents. Searching for the blocks starts with 202c9de560dSAlex Tomas * the group specified as the goal value in allocation context via 203c9de560dSAlex Tomas * ac_g_ex. Each group is first checked based on the criteria whether it 204caaf7a29STao Ma * can be used for allocation. ext4_mb_good_group explains how the groups are 205c9de560dSAlex Tomas * checked. 206c9de560dSAlex Tomas * 207196e402aSHarshad Shirwadkar * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not 208196e402aSHarshad Shirwadkar * get traversed linearly. That may result in subsequent allocations being not 209196e402aSHarshad Shirwadkar * close to each other. And so, the underlying device may get filled up in a 210196e402aSHarshad Shirwadkar * non-linear fashion. While that may not matter on non-rotational devices, for 211196e402aSHarshad Shirwadkar * rotational devices that may result in higher seek times. "mb_linear_limit" 212196e402aSHarshad Shirwadkar * tells mballoc how many groups mballoc should search linearly before 213196e402aSHarshad Shirwadkar * performing consulting above data structures for more efficient lookups. For 214196e402aSHarshad Shirwadkar * non rotational devices, this value defaults to 0 and for rotational devices 215196e402aSHarshad Shirwadkar * this is set to MB_DEFAULT_LINEAR_LIMIT. 216196e402aSHarshad Shirwadkar * 217c9de560dSAlex Tomas * Both the prealloc space are getting populated as above. So for the first 218c9de560dSAlex Tomas * request we will hit the buddy cache which will result in this prealloc 219c9de560dSAlex Tomas * space getting filled. The prealloc space is then later used for the 220c9de560dSAlex Tomas * subsequent request. 221c9de560dSAlex Tomas */ 222c9de560dSAlex Tomas 223c9de560dSAlex Tomas /* 224c9de560dSAlex Tomas * mballoc operates on the following data: 225c9de560dSAlex Tomas * - on-disk bitmap 226c9de560dSAlex Tomas * - in-core buddy (actually includes buddy and bitmap) 227c9de560dSAlex Tomas * - preallocation descriptors (PAs) 228c9de560dSAlex Tomas * 229c9de560dSAlex Tomas * there are two types of preallocations: 230c9de560dSAlex Tomas * - inode 231c9de560dSAlex Tomas * assiged to specific inode and can be used for this inode only. 232c9de560dSAlex Tomas * it describes part of inode's space preallocated to specific 233c9de560dSAlex Tomas * physical blocks. any block from that preallocated can be used 234c9de560dSAlex Tomas * independent. the descriptor just tracks number of blocks left 235c9de560dSAlex Tomas * unused. so, before taking some block from descriptor, one must 236c9de560dSAlex Tomas * make sure corresponded logical block isn't allocated yet. this 237c9de560dSAlex Tomas * also means that freeing any block within descriptor's range 238c9de560dSAlex Tomas * must discard all preallocated blocks. 239c9de560dSAlex Tomas * - locality group 240c9de560dSAlex Tomas * assigned to specific locality group which does not translate to 241c9de560dSAlex Tomas * permanent set of inodes: inode can join and leave group. space 242c9de560dSAlex Tomas * from this type of preallocation can be used for any inode. thus 243c9de560dSAlex Tomas * it's consumed from the beginning to the end. 244c9de560dSAlex Tomas * 245c9de560dSAlex Tomas * relation between them can be expressed as: 246c9de560dSAlex Tomas * in-core buddy = on-disk bitmap + preallocation descriptors 247c9de560dSAlex Tomas * 248c9de560dSAlex Tomas * this mean blocks mballoc considers used are: 249c9de560dSAlex Tomas * - allocated blocks (persistent) 250c9de560dSAlex Tomas * - preallocated blocks (non-persistent) 251c9de560dSAlex Tomas * 252c9de560dSAlex Tomas * consistency in mballoc world means that at any time a block is either 253c9de560dSAlex Tomas * free or used in ALL structures. notice: "any time" should not be read 254c9de560dSAlex Tomas * literally -- time is discrete and delimited by locks. 255c9de560dSAlex Tomas * 256c9de560dSAlex Tomas * to keep it simple, we don't use block numbers, instead we count number of 257c9de560dSAlex Tomas * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 258c9de560dSAlex Tomas * 259c9de560dSAlex Tomas * all operations can be expressed as: 260c9de560dSAlex Tomas * - init buddy: buddy = on-disk + PAs 261c9de560dSAlex Tomas * - new PA: buddy += N; PA = N 262c9de560dSAlex Tomas * - use inode PA: on-disk += N; PA -= N 263c9de560dSAlex Tomas * - discard inode PA buddy -= on-disk - PA; PA = 0 264c9de560dSAlex Tomas * - use locality group PA on-disk += N; PA -= N 265c9de560dSAlex Tomas * - discard locality group PA buddy -= PA; PA = 0 266c9de560dSAlex Tomas * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 267c9de560dSAlex Tomas * is used in real operation because we can't know actual used 268c9de560dSAlex Tomas * bits from PA, only from on-disk bitmap 269c9de560dSAlex Tomas * 270c9de560dSAlex Tomas * if we follow this strict logic, then all operations above should be atomic. 271c9de560dSAlex Tomas * given some of them can block, we'd have to use something like semaphores 272c9de560dSAlex Tomas * killing performance on high-end SMP hardware. let's try to relax it using 273c9de560dSAlex Tomas * the following knowledge: 274c9de560dSAlex Tomas * 1) if buddy is referenced, it's already initialized 275c9de560dSAlex Tomas * 2) while block is used in buddy and the buddy is referenced, 276c9de560dSAlex Tomas * nobody can re-allocate that block 277c9de560dSAlex Tomas * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 278c9de560dSAlex Tomas * bit set and PA claims same block, it's OK. IOW, one can set bit in 279c9de560dSAlex Tomas * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 280c9de560dSAlex Tomas * block 281c9de560dSAlex Tomas * 282c9de560dSAlex Tomas * so, now we're building a concurrency table: 283c9de560dSAlex Tomas * - init buddy vs. 284c9de560dSAlex Tomas * - new PA 285c9de560dSAlex Tomas * blocks for PA are allocated in the buddy, buddy must be referenced 286c9de560dSAlex Tomas * until PA is linked to allocation group to avoid concurrent buddy init 287c9de560dSAlex Tomas * - use inode PA 288c9de560dSAlex Tomas * we need to make sure that either on-disk bitmap or PA has uptodate data 289c9de560dSAlex Tomas * given (3) we care that PA-=N operation doesn't interfere with init 290c9de560dSAlex Tomas * - discard inode PA 291c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 292c9de560dSAlex Tomas * - use locality group PA 293c9de560dSAlex Tomas * again PA-=N must be serialized with init 294c9de560dSAlex Tomas * - discard locality group PA 295c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 296c9de560dSAlex Tomas * - new PA vs. 297c9de560dSAlex Tomas * - use inode PA 298c9de560dSAlex Tomas * i_data_sem serializes them 299c9de560dSAlex Tomas * - discard inode PA 300c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 301c9de560dSAlex Tomas * - use locality group PA 302c9de560dSAlex Tomas * some mutex should serialize them 303c9de560dSAlex Tomas * - discard locality group PA 304c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 305c9de560dSAlex Tomas * - use inode PA 306c9de560dSAlex Tomas * - use inode PA 307c9de560dSAlex Tomas * i_data_sem or another mutex should serializes them 308c9de560dSAlex Tomas * - discard inode PA 309c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 310c9de560dSAlex Tomas * - use locality group PA 311c9de560dSAlex Tomas * nothing wrong here -- they're different PAs covering different blocks 312c9de560dSAlex Tomas * - discard locality group PA 313c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 314c9de560dSAlex Tomas * 315c9de560dSAlex Tomas * now we're ready to make few consequences: 316c9de560dSAlex Tomas * - PA is referenced and while it is no discard is possible 317c9de560dSAlex Tomas * - PA is referenced until block isn't marked in on-disk bitmap 318c9de560dSAlex Tomas * - PA changes only after on-disk bitmap 319c9de560dSAlex Tomas * - discard must not compete with init. either init is done before 320c9de560dSAlex Tomas * any discard or they're serialized somehow 321c9de560dSAlex Tomas * - buddy init as sum of on-disk bitmap and PAs is done atomically 322c9de560dSAlex Tomas * 323c9de560dSAlex Tomas * a special case when we've used PA to emptiness. no need to modify buddy 324c9de560dSAlex Tomas * in this case, but we should care about concurrent init 325c9de560dSAlex Tomas * 326c9de560dSAlex Tomas */ 327c9de560dSAlex Tomas 328c9de560dSAlex Tomas /* 329c9de560dSAlex Tomas * Logic in few words: 330c9de560dSAlex Tomas * 331c9de560dSAlex Tomas * - allocation: 332c9de560dSAlex Tomas * load group 333c9de560dSAlex Tomas * find blocks 334c9de560dSAlex Tomas * mark bits in on-disk bitmap 335c9de560dSAlex Tomas * release group 336c9de560dSAlex Tomas * 337c9de560dSAlex Tomas * - use preallocation: 338c9de560dSAlex Tomas * find proper PA (per-inode or group) 339c9de560dSAlex Tomas * load group 340c9de560dSAlex Tomas * mark bits in on-disk bitmap 341c9de560dSAlex Tomas * release group 342c9de560dSAlex Tomas * release PA 343c9de560dSAlex Tomas * 344c9de560dSAlex Tomas * - free: 345c9de560dSAlex Tomas * load group 346c9de560dSAlex Tomas * mark bits in on-disk bitmap 347c9de560dSAlex Tomas * release group 348c9de560dSAlex Tomas * 349c9de560dSAlex Tomas * - discard preallocations in group: 350c9de560dSAlex Tomas * mark PAs deleted 351c9de560dSAlex Tomas * move them onto local list 352c9de560dSAlex Tomas * load on-disk bitmap 353c9de560dSAlex Tomas * load group 354c9de560dSAlex Tomas * remove PA from object (inode or locality group) 355c9de560dSAlex Tomas * mark free blocks in-core 356c9de560dSAlex Tomas * 357c9de560dSAlex Tomas * - discard inode's preallocations: 358c9de560dSAlex Tomas */ 359c9de560dSAlex Tomas 360c9de560dSAlex Tomas /* 361c9de560dSAlex Tomas * Locking rules 362c9de560dSAlex Tomas * 363c9de560dSAlex Tomas * Locks: 364c9de560dSAlex Tomas * - bitlock on a group (group) 365c9de560dSAlex Tomas * - object (inode/locality) (object) 366c9de560dSAlex Tomas * - per-pa lock (pa) 367f52f3d2bSOjaswin Mujoo * - cr_power2_aligned lists lock (cr_power2_aligned) 368f52f3d2bSOjaswin Mujoo * - cr_goal_len_fast lists lock (cr_goal_len_fast) 369c9de560dSAlex Tomas * 370c9de560dSAlex Tomas * Paths: 371c9de560dSAlex Tomas * - new pa 372c9de560dSAlex Tomas * object 373c9de560dSAlex Tomas * group 374c9de560dSAlex Tomas * 375c9de560dSAlex Tomas * - find and use pa: 376c9de560dSAlex Tomas * pa 377c9de560dSAlex Tomas * 378c9de560dSAlex Tomas * - release consumed pa: 379c9de560dSAlex Tomas * pa 380c9de560dSAlex Tomas * group 381c9de560dSAlex Tomas * object 382c9de560dSAlex Tomas * 383c9de560dSAlex Tomas * - generate in-core bitmap: 384c9de560dSAlex Tomas * group 385c9de560dSAlex Tomas * pa 386c9de560dSAlex Tomas * 387c9de560dSAlex Tomas * - discard all for given object (inode, locality group): 388c9de560dSAlex Tomas * object 389c9de560dSAlex Tomas * pa 390c9de560dSAlex Tomas * group 391c9de560dSAlex Tomas * 392c9de560dSAlex Tomas * - discard all for given group: 393c9de560dSAlex Tomas * group 394c9de560dSAlex Tomas * pa 395c9de560dSAlex Tomas * group 396c9de560dSAlex Tomas * object 397c9de560dSAlex Tomas * 398196e402aSHarshad Shirwadkar * - allocation path (ext4_mb_regular_allocator) 399196e402aSHarshad Shirwadkar * group 400f52f3d2bSOjaswin Mujoo * cr_power2_aligned/cr_goal_len_fast 401c9de560dSAlex Tomas */ 402c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_pspace_cachep; 403c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_ac_cachep; 40418aadd47SBobi Jam static struct kmem_cache *ext4_free_data_cachep; 405fb1813f4SCurt Wohlgemuth 406fb1813f4SCurt Wohlgemuth /* We create slab caches for groupinfo data structures based on the 407fb1813f4SCurt Wohlgemuth * superblock block size. There will be one per mounted filesystem for 408fb1813f4SCurt Wohlgemuth * each unique s_blocksize_bits */ 4092892c15dSEric Sandeen #define NR_GRPINFO_CACHES 8 410fb1813f4SCurt Wohlgemuth static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; 411fb1813f4SCurt Wohlgemuth 412d6006186SEric Biggers static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { 4132892c15dSEric Sandeen "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", 4142892c15dSEric Sandeen "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", 4152892c15dSEric Sandeen "ext4_groupinfo_64k", "ext4_groupinfo_128k" 4162892c15dSEric Sandeen }; 4172892c15dSEric Sandeen 418c3a326a6SAneesh Kumar K.V static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 419c3a326a6SAneesh Kumar K.V ext4_group_t group); 4207a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 4217a2fcbf7SAneesh Kumar K.V ext4_group_t group); 42253f86b17SRitesh Harjani static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac); 423c3a326a6SAneesh Kumar K.V 424196e402aSHarshad Shirwadkar static bool ext4_mb_good_group(struct ext4_allocation_context *ac, 4254eb7a4a1SOjaswin Mujoo ext4_group_t group, enum criteria cr); 426196e402aSHarshad Shirwadkar 42755cdd0afSWang Jianchao static int ext4_try_to_trim_range(struct super_block *sb, 42855cdd0afSWang Jianchao struct ext4_buddy *e4b, ext4_grpblk_t start, 42955cdd0afSWang Jianchao ext4_grpblk_t max, ext4_grpblk_t minblocks); 43055cdd0afSWang Jianchao 43107b5b8e1SRitesh Harjani /* 43207b5b8e1SRitesh Harjani * The algorithm using this percpu seq counter goes below: 43307b5b8e1SRitesh Harjani * 1. We sample the percpu discard_pa_seq counter before trying for block 43407b5b8e1SRitesh Harjani * allocation in ext4_mb_new_blocks(). 43507b5b8e1SRitesh Harjani * 2. We increment this percpu discard_pa_seq counter when we either allocate 43607b5b8e1SRitesh Harjani * or free these blocks i.e. while marking those blocks as used/free in 43707b5b8e1SRitesh Harjani * mb_mark_used()/mb_free_blocks(). 43807b5b8e1SRitesh Harjani * 3. We also increment this percpu seq counter when we successfully identify 43907b5b8e1SRitesh Harjani * that the bb_prealloc_list is not empty and hence proceed for discarding 44007b5b8e1SRitesh Harjani * of those PAs inside ext4_mb_discard_group_preallocations(). 44107b5b8e1SRitesh Harjani * 44207b5b8e1SRitesh Harjani * Now to make sure that the regular fast path of block allocation is not 44307b5b8e1SRitesh Harjani * affected, as a small optimization we only sample the percpu seq counter 44407b5b8e1SRitesh Harjani * on that cpu. Only when the block allocation fails and when freed blocks 44507b5b8e1SRitesh Harjani * found were 0, that is when we sample percpu seq counter for all cpus using 44607b5b8e1SRitesh Harjani * below function ext4_get_discard_pa_seq_sum(). This happens after making 44707b5b8e1SRitesh Harjani * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty. 44807b5b8e1SRitesh Harjani */ 44907b5b8e1SRitesh Harjani static DEFINE_PER_CPU(u64, discard_pa_seq); 45007b5b8e1SRitesh Harjani static inline u64 ext4_get_discard_pa_seq_sum(void) 45107b5b8e1SRitesh Harjani { 45207b5b8e1SRitesh Harjani int __cpu; 45307b5b8e1SRitesh Harjani u64 __seq = 0; 45407b5b8e1SRitesh Harjani 45507b5b8e1SRitesh Harjani for_each_possible_cpu(__cpu) 45607b5b8e1SRitesh Harjani __seq += per_cpu(discard_pa_seq, __cpu); 45707b5b8e1SRitesh Harjani return __seq; 45807b5b8e1SRitesh Harjani } 45907b5b8e1SRitesh Harjani 460ffad0a44SAneesh Kumar K.V static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 461ffad0a44SAneesh Kumar K.V { 462c9de560dSAlex Tomas #if BITS_PER_LONG == 64 463ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 7UL) << 3; 464ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~7UL); 465c9de560dSAlex Tomas #elif BITS_PER_LONG == 32 466ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 3UL) << 3; 467ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~3UL); 468c9de560dSAlex Tomas #else 469c9de560dSAlex Tomas #error "how many bits you are?!" 470c9de560dSAlex Tomas #endif 471ffad0a44SAneesh Kumar K.V return addr; 472ffad0a44SAneesh Kumar K.V } 473c9de560dSAlex Tomas 474c9de560dSAlex Tomas static inline int mb_test_bit(int bit, void *addr) 475c9de560dSAlex Tomas { 476c9de560dSAlex Tomas /* 477c9de560dSAlex Tomas * ext4_test_bit on architecture like powerpc 478c9de560dSAlex Tomas * needs unsigned long aligned address 479c9de560dSAlex Tomas */ 480ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 481c9de560dSAlex Tomas return ext4_test_bit(bit, addr); 482c9de560dSAlex Tomas } 483c9de560dSAlex Tomas 484c9de560dSAlex Tomas static inline void mb_set_bit(int bit, void *addr) 485c9de560dSAlex Tomas { 486ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 487c9de560dSAlex Tomas ext4_set_bit(bit, addr); 488c9de560dSAlex Tomas } 489c9de560dSAlex Tomas 490c9de560dSAlex Tomas static inline void mb_clear_bit(int bit, void *addr) 491c9de560dSAlex Tomas { 492ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 493c9de560dSAlex Tomas ext4_clear_bit(bit, addr); 494c9de560dSAlex Tomas } 495c9de560dSAlex Tomas 496eabe0444SAndrey Sidorov static inline int mb_test_and_clear_bit(int bit, void *addr) 497eabe0444SAndrey Sidorov { 498eabe0444SAndrey Sidorov addr = mb_correct_addr_and_bit(&bit, addr); 499eabe0444SAndrey Sidorov return ext4_test_and_clear_bit(bit, addr); 500eabe0444SAndrey Sidorov } 501eabe0444SAndrey Sidorov 502ffad0a44SAneesh Kumar K.V static inline int mb_find_next_zero_bit(void *addr, int max, int start) 503ffad0a44SAneesh Kumar K.V { 504e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 505ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 506e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 507ffad0a44SAneesh Kumar K.V start += fix; 508ffad0a44SAneesh Kumar K.V 509e7dfb246SAneesh Kumar K.V ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 510e7dfb246SAneesh Kumar K.V if (ret > max) 511e7dfb246SAneesh Kumar K.V return max; 512e7dfb246SAneesh Kumar K.V return ret; 513ffad0a44SAneesh Kumar K.V } 514ffad0a44SAneesh Kumar K.V 515ffad0a44SAneesh Kumar K.V static inline int mb_find_next_bit(void *addr, int max, int start) 516ffad0a44SAneesh Kumar K.V { 517e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 518ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 519e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 520ffad0a44SAneesh Kumar K.V start += fix; 521ffad0a44SAneesh Kumar K.V 522e7dfb246SAneesh Kumar K.V ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 523e7dfb246SAneesh Kumar K.V if (ret > max) 524e7dfb246SAneesh Kumar K.V return max; 525e7dfb246SAneesh Kumar K.V return ret; 526ffad0a44SAneesh Kumar K.V } 527ffad0a44SAneesh Kumar K.V 528c9de560dSAlex Tomas static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 529c9de560dSAlex Tomas { 530c9de560dSAlex Tomas char *bb; 531c9de560dSAlex Tomas 532c5e8f3f3STheodore Ts'o BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 533c9de560dSAlex Tomas BUG_ON(max == NULL); 534c9de560dSAlex Tomas 535c9de560dSAlex Tomas if (order > e4b->bd_blkbits + 1) { 536c9de560dSAlex Tomas *max = 0; 537c9de560dSAlex Tomas return NULL; 538c9de560dSAlex Tomas } 539c9de560dSAlex Tomas 540c9de560dSAlex Tomas /* at order 0 we see each particular block */ 54184b775a3SColy Li if (order == 0) { 542c9de560dSAlex Tomas *max = 1 << (e4b->bd_blkbits + 3); 543c5e8f3f3STheodore Ts'o return e4b->bd_bitmap; 54484b775a3SColy Li } 545c9de560dSAlex Tomas 546c5e8f3f3STheodore Ts'o bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 547c9de560dSAlex Tomas *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 548c9de560dSAlex Tomas 549c9de560dSAlex Tomas return bb; 550c9de560dSAlex Tomas } 551c9de560dSAlex Tomas 552c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 553c9de560dSAlex Tomas static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 554c9de560dSAlex Tomas int first, int count) 555c9de560dSAlex Tomas { 556c9de560dSAlex Tomas int i; 557c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 558c9de560dSAlex Tomas 559c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 560c9de560dSAlex Tomas return; 561bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 562c9de560dSAlex Tomas for (i = 0; i < count; i++) { 563c9de560dSAlex Tomas if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 564c9de560dSAlex Tomas ext4_fsblk_t blocknr; 5655661bd68SAkinobu Mita 5665661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 56753accfa9STheodore Ts'o blocknr += EXT4_C2B(EXT4_SB(sb), first + i); 5685d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 569e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 570e29136f8STheodore Ts'o blocknr, 571e29136f8STheodore Ts'o "freeing block already freed " 572e29136f8STheodore Ts'o "(bit %u)", 573e29136f8STheodore Ts'o first + i); 574736dedbbSWang Shilong ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 575736dedbbSWang Shilong EXT4_GROUP_INFO_BBITMAP_CORRUPT); 576c9de560dSAlex Tomas } 577c9de560dSAlex Tomas mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 578c9de560dSAlex Tomas } 579c9de560dSAlex Tomas } 580c9de560dSAlex Tomas 581c9de560dSAlex Tomas static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 582c9de560dSAlex Tomas { 583c9de560dSAlex Tomas int i; 584c9de560dSAlex Tomas 585c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 586c9de560dSAlex Tomas return; 587bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 588c9de560dSAlex Tomas for (i = 0; i < count; i++) { 589c9de560dSAlex Tomas BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 590c9de560dSAlex Tomas mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 591c9de560dSAlex Tomas } 592c9de560dSAlex Tomas } 593c9de560dSAlex Tomas 594c9de560dSAlex Tomas static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 595c9de560dSAlex Tomas { 596eb2b8ebbSRitesh Harjani if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 597eb2b8ebbSRitesh Harjani return; 598c9de560dSAlex Tomas if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 599c9de560dSAlex Tomas unsigned char *b1, *b2; 600c9de560dSAlex Tomas int i; 601c9de560dSAlex Tomas b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 602c9de560dSAlex Tomas b2 = (unsigned char *) bitmap; 603c9de560dSAlex Tomas for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 604c9de560dSAlex Tomas if (b1[i] != b2[i]) { 6059d8b9ec4STheodore Ts'o ext4_msg(e4b->bd_sb, KERN_ERR, 6069d8b9ec4STheodore Ts'o "corruption in group %u " 6074776004fSTheodore Ts'o "at byte %u(%u): %x in copy != %x " 6089d8b9ec4STheodore Ts'o "on disk/prealloc", 609c9de560dSAlex Tomas e4b->bd_group, i, i * 8, b1[i], b2[i]); 610c9de560dSAlex Tomas BUG(); 611c9de560dSAlex Tomas } 612c9de560dSAlex Tomas } 613c9de560dSAlex Tomas } 614c9de560dSAlex Tomas } 615c9de560dSAlex Tomas 616a3450215SRitesh Harjani static void mb_group_bb_bitmap_alloc(struct super_block *sb, 617a3450215SRitesh Harjani struct ext4_group_info *grp, ext4_group_t group) 618a3450215SRitesh Harjani { 619a3450215SRitesh Harjani struct buffer_head *bh; 620a3450215SRitesh Harjani 621a3450215SRitesh Harjani grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); 622eb2b8ebbSRitesh Harjani if (!grp->bb_bitmap) 623eb2b8ebbSRitesh Harjani return; 624a3450215SRitesh Harjani 625a3450215SRitesh Harjani bh = ext4_read_block_bitmap(sb, group); 626eb2b8ebbSRitesh Harjani if (IS_ERR_OR_NULL(bh)) { 627eb2b8ebbSRitesh Harjani kfree(grp->bb_bitmap); 628eb2b8ebbSRitesh Harjani grp->bb_bitmap = NULL; 629eb2b8ebbSRitesh Harjani return; 630eb2b8ebbSRitesh Harjani } 631a3450215SRitesh Harjani 632a3450215SRitesh Harjani memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); 633a3450215SRitesh Harjani put_bh(bh); 634a3450215SRitesh Harjani } 635a3450215SRitesh Harjani 636a3450215SRitesh Harjani static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) 637a3450215SRitesh Harjani { 638a3450215SRitesh Harjani kfree(grp->bb_bitmap); 639a3450215SRitesh Harjani } 640a3450215SRitesh Harjani 641c9de560dSAlex Tomas #else 642c9de560dSAlex Tomas static inline void mb_free_blocks_double(struct inode *inode, 643c9de560dSAlex Tomas struct ext4_buddy *e4b, int first, int count) 644c9de560dSAlex Tomas { 645c9de560dSAlex Tomas return; 646c9de560dSAlex Tomas } 647c9de560dSAlex Tomas static inline void mb_mark_used_double(struct ext4_buddy *e4b, 648c9de560dSAlex Tomas int first, int count) 649c9de560dSAlex Tomas { 650c9de560dSAlex Tomas return; 651c9de560dSAlex Tomas } 652c9de560dSAlex Tomas static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 653c9de560dSAlex Tomas { 654c9de560dSAlex Tomas return; 655c9de560dSAlex Tomas } 656a3450215SRitesh Harjani 657a3450215SRitesh Harjani static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, 658a3450215SRitesh Harjani struct ext4_group_info *grp, ext4_group_t group) 659a3450215SRitesh Harjani { 660a3450215SRitesh Harjani return; 661a3450215SRitesh Harjani } 662a3450215SRitesh Harjani 663a3450215SRitesh Harjani static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) 664a3450215SRitesh Harjani { 665a3450215SRitesh Harjani return; 666a3450215SRitesh Harjani } 667c9de560dSAlex Tomas #endif 668c9de560dSAlex Tomas 669c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 670c9de560dSAlex Tomas 671c9de560dSAlex Tomas #define MB_CHECK_ASSERT(assert) \ 672c9de560dSAlex Tomas do { \ 673c9de560dSAlex Tomas if (!(assert)) { \ 674c9de560dSAlex Tomas printk(KERN_EMERG \ 675c9de560dSAlex Tomas "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 676c9de560dSAlex Tomas function, file, line, # assert); \ 677c9de560dSAlex Tomas BUG(); \ 678c9de560dSAlex Tomas } \ 679c9de560dSAlex Tomas } while (0) 680c9de560dSAlex Tomas 681c9de560dSAlex Tomas static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 682c9de560dSAlex Tomas const char *function, int line) 683c9de560dSAlex Tomas { 684c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 685c9de560dSAlex Tomas int order = e4b->bd_blkbits + 1; 686c9de560dSAlex Tomas int max; 687c9de560dSAlex Tomas int max2; 688c9de560dSAlex Tomas int i; 689c9de560dSAlex Tomas int j; 690c9de560dSAlex Tomas int k; 691c9de560dSAlex Tomas int count; 692c9de560dSAlex Tomas struct ext4_group_info *grp; 693c9de560dSAlex Tomas int fragments = 0; 694c9de560dSAlex Tomas int fstart; 695c9de560dSAlex Tomas struct list_head *cur; 696c9de560dSAlex Tomas void *buddy; 697c9de560dSAlex Tomas void *buddy2; 698c9de560dSAlex Tomas 699addd752cSChunguang Xu if (e4b->bd_info->bb_check_counter++ % 10) 700c9de560dSAlex Tomas return 0; 701c9de560dSAlex Tomas 702c9de560dSAlex Tomas while (order > 1) { 703c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 704c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy); 705c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order - 1, &max2); 706c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy2); 707c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy != buddy2); 708c9de560dSAlex Tomas MB_CHECK_ASSERT(max * 2 == max2); 709c9de560dSAlex Tomas 710c9de560dSAlex Tomas count = 0; 711c9de560dSAlex Tomas for (i = 0; i < max; i++) { 712c9de560dSAlex Tomas 713c9de560dSAlex Tomas if (mb_test_bit(i, buddy)) { 714af2b3275SJinke Han /* only single bit in buddy2 may be 0 */ 715c9de560dSAlex Tomas if (!mb_test_bit(i << 1, buddy2)) { 716c9de560dSAlex Tomas MB_CHECK_ASSERT( 717c9de560dSAlex Tomas mb_test_bit((i<<1)+1, buddy2)); 718c9de560dSAlex Tomas } 719c9de560dSAlex Tomas continue; 720c9de560dSAlex Tomas } 721c9de560dSAlex Tomas 7220a10da73SRobin Dong /* both bits in buddy2 must be 1 */ 723c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 724c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 725c9de560dSAlex Tomas 726c9de560dSAlex Tomas for (j = 0; j < (1 << order); j++) { 727c9de560dSAlex Tomas k = (i * (1 << order)) + j; 728c9de560dSAlex Tomas MB_CHECK_ASSERT( 729c5e8f3f3STheodore Ts'o !mb_test_bit(k, e4b->bd_bitmap)); 730c9de560dSAlex Tomas } 731c9de560dSAlex Tomas count++; 732c9de560dSAlex Tomas } 733c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 734c9de560dSAlex Tomas order--; 735c9de560dSAlex Tomas } 736c9de560dSAlex Tomas 737c9de560dSAlex Tomas fstart = -1; 738c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 739c9de560dSAlex Tomas for (i = 0; i < max; i++) { 740c9de560dSAlex Tomas if (!mb_test_bit(i, buddy)) { 741c9de560dSAlex Tomas MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 742c9de560dSAlex Tomas if (fstart == -1) { 743c9de560dSAlex Tomas fragments++; 744c9de560dSAlex Tomas fstart = i; 745c9de560dSAlex Tomas } 746c9de560dSAlex Tomas continue; 747c9de560dSAlex Tomas } 748c9de560dSAlex Tomas fstart = -1; 749c9de560dSAlex Tomas /* check used bits only */ 750c9de560dSAlex Tomas for (j = 0; j < e4b->bd_blkbits + 1; j++) { 751c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, j, &max2); 752c9de560dSAlex Tomas k = i >> j; 753c9de560dSAlex Tomas MB_CHECK_ASSERT(k < max2); 754c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 755c9de560dSAlex Tomas } 756c9de560dSAlex Tomas } 757c9de560dSAlex Tomas MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 758c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 759c9de560dSAlex Tomas 760c9de560dSAlex Tomas grp = ext4_get_group_info(sb, e4b->bd_group); 7615354b2afSTheodore Ts'o if (!grp) 7625354b2afSTheodore Ts'o return NULL; 763c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 764c9de560dSAlex Tomas ext4_group_t groupnr; 765c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 76660bd63d1SSolofo Ramangalahy pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 76760bd63d1SSolofo Ramangalahy ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 768c9de560dSAlex Tomas MB_CHECK_ASSERT(groupnr == e4b->bd_group); 76960bd63d1SSolofo Ramangalahy for (i = 0; i < pa->pa_len; i++) 770c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 771c9de560dSAlex Tomas } 772c9de560dSAlex Tomas return 0; 773c9de560dSAlex Tomas } 774c9de560dSAlex Tomas #undef MB_CHECK_ASSERT 775c9de560dSAlex Tomas #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 77646e665e9SHarvey Harrison __FILE__, __func__, __LINE__) 777c9de560dSAlex Tomas #else 778c9de560dSAlex Tomas #define mb_check_buddy(e4b) 779c9de560dSAlex Tomas #endif 780c9de560dSAlex Tomas 7817c786059SColy Li /* 7827c786059SColy Li * Divide blocks started from @first with length @len into 7837c786059SColy Li * smaller chunks with power of 2 blocks. 7847c786059SColy Li * Clear the bits in bitmap which the blocks of the chunk(s) covered, 7857c786059SColy Li * then increase bb_counters[] for corresponded chunk size. 7867c786059SColy Li */ 787c9de560dSAlex Tomas static void ext4_mb_mark_free_simple(struct super_block *sb, 788a36b4498SEric Sandeen void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 789c9de560dSAlex Tomas struct ext4_group_info *grp) 790c9de560dSAlex Tomas { 791c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 792a36b4498SEric Sandeen ext4_grpblk_t min; 793a36b4498SEric Sandeen ext4_grpblk_t max; 794a36b4498SEric Sandeen ext4_grpblk_t chunk; 79569e43e8cSChandan Rajendra unsigned int border; 796c9de560dSAlex Tomas 7977137d7a4STheodore Ts'o BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); 798c9de560dSAlex Tomas 799c9de560dSAlex Tomas border = 2 << sb->s_blocksize_bits; 800c9de560dSAlex Tomas 801c9de560dSAlex Tomas while (len > 0) { 802c9de560dSAlex Tomas /* find how many blocks can be covered since this position */ 803c9de560dSAlex Tomas max = ffs(first | border) - 1; 804c9de560dSAlex Tomas 805c9de560dSAlex Tomas /* find how many blocks of power 2 we need to mark */ 806c9de560dSAlex Tomas min = fls(len) - 1; 807c9de560dSAlex Tomas 808c9de560dSAlex Tomas if (max < min) 809c9de560dSAlex Tomas min = max; 810c9de560dSAlex Tomas chunk = 1 << min; 811c9de560dSAlex Tomas 812c9de560dSAlex Tomas /* mark multiblock chunks only */ 813c9de560dSAlex Tomas grp->bb_counters[min]++; 814c9de560dSAlex Tomas if (min > 0) 815c9de560dSAlex Tomas mb_clear_bit(first >> min, 816c9de560dSAlex Tomas buddy + sbi->s_mb_offsets[min]); 817c9de560dSAlex Tomas 818c9de560dSAlex Tomas len -= chunk; 819c9de560dSAlex Tomas first += chunk; 820c9de560dSAlex Tomas } 821c9de560dSAlex Tomas } 822c9de560dSAlex Tomas 82383e80a6eSJan Kara static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) 824196e402aSHarshad Shirwadkar { 82583e80a6eSJan Kara int order; 826196e402aSHarshad Shirwadkar 827196e402aSHarshad Shirwadkar /* 82883e80a6eSJan Kara * We don't bother with a special lists groups with only 1 block free 82983e80a6eSJan Kara * extents and for completely empty groups. 830196e402aSHarshad Shirwadkar */ 83183e80a6eSJan Kara order = fls(len) - 2; 83283e80a6eSJan Kara if (order < 0) 83383e80a6eSJan Kara return 0; 83483e80a6eSJan Kara if (order == MB_NUM_ORDERS(sb)) 83583e80a6eSJan Kara order--; 83683e80a6eSJan Kara return order; 83783e80a6eSJan Kara } 83883e80a6eSJan Kara 83983e80a6eSJan Kara /* Move group to appropriate avg_fragment_size list */ 840196e402aSHarshad Shirwadkar static void 841196e402aSHarshad Shirwadkar mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) 842196e402aSHarshad Shirwadkar { 843196e402aSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 84483e80a6eSJan Kara int new_order; 845196e402aSHarshad Shirwadkar 846196e402aSHarshad Shirwadkar if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0) 847196e402aSHarshad Shirwadkar return; 848196e402aSHarshad Shirwadkar 84983e80a6eSJan Kara new_order = mb_avg_fragment_size_order(sb, 85083e80a6eSJan Kara grp->bb_free / grp->bb_fragments); 85183e80a6eSJan Kara if (new_order == grp->bb_avg_fragment_size_order) 85283e80a6eSJan Kara return; 853196e402aSHarshad Shirwadkar 85483e80a6eSJan Kara if (grp->bb_avg_fragment_size_order != -1) { 85583e80a6eSJan Kara write_lock(&sbi->s_mb_avg_fragment_size_locks[ 85683e80a6eSJan Kara grp->bb_avg_fragment_size_order]); 85783e80a6eSJan Kara list_del(&grp->bb_avg_fragment_size_node); 85883e80a6eSJan Kara write_unlock(&sbi->s_mb_avg_fragment_size_locks[ 85983e80a6eSJan Kara grp->bb_avg_fragment_size_order]); 86083e80a6eSJan Kara } 86183e80a6eSJan Kara grp->bb_avg_fragment_size_order = new_order; 86283e80a6eSJan Kara write_lock(&sbi->s_mb_avg_fragment_size_locks[ 86383e80a6eSJan Kara grp->bb_avg_fragment_size_order]); 86483e80a6eSJan Kara list_add_tail(&grp->bb_avg_fragment_size_node, 86583e80a6eSJan Kara &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]); 86683e80a6eSJan Kara write_unlock(&sbi->s_mb_avg_fragment_size_locks[ 86783e80a6eSJan Kara grp->bb_avg_fragment_size_order]); 868196e402aSHarshad Shirwadkar } 869196e402aSHarshad Shirwadkar 870196e402aSHarshad Shirwadkar /* 871196e402aSHarshad Shirwadkar * Choose next group by traversing largest_free_order lists. Updates *new_cr if 872196e402aSHarshad Shirwadkar * cr level needs an update. 873196e402aSHarshad Shirwadkar */ 874f52f3d2bSOjaswin Mujoo static void ext4_mb_choose_next_group_p2_aligned(struct ext4_allocation_context *ac, 8754eb7a4a1SOjaswin Mujoo enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 876196e402aSHarshad Shirwadkar { 877196e402aSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 878919eb90cSKemeng Shi struct ext4_group_info *iter; 879196e402aSHarshad Shirwadkar int i; 880196e402aSHarshad Shirwadkar 881196e402aSHarshad Shirwadkar if (ac->ac_status == AC_STATUS_FOUND) 882196e402aSHarshad Shirwadkar return; 883196e402aSHarshad Shirwadkar 884f52f3d2bSOjaswin Mujoo if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED)) 885f52f3d2bSOjaswin Mujoo atomic_inc(&sbi->s_bal_p2_aligned_bad_suggestions); 886196e402aSHarshad Shirwadkar 887196e402aSHarshad Shirwadkar for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) { 888196e402aSHarshad Shirwadkar if (list_empty(&sbi->s_mb_largest_free_orders[i])) 889196e402aSHarshad Shirwadkar continue; 890196e402aSHarshad Shirwadkar read_lock(&sbi->s_mb_largest_free_orders_locks[i]); 891196e402aSHarshad Shirwadkar if (list_empty(&sbi->s_mb_largest_free_orders[i])) { 892196e402aSHarshad Shirwadkar read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); 893196e402aSHarshad Shirwadkar continue; 894196e402aSHarshad Shirwadkar } 895196e402aSHarshad Shirwadkar list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i], 896196e402aSHarshad Shirwadkar bb_largest_free_order_node) { 897196e402aSHarshad Shirwadkar if (sbi->s_mb_stats) 898f52f3d2bSOjaswin Mujoo atomic64_inc(&sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED]); 899f52f3d2bSOjaswin Mujoo if (likely(ext4_mb_good_group(ac, iter->bb_group, CR_POWER2_ALIGNED))) { 900919eb90cSKemeng Shi *group = iter->bb_group; 901919eb90cSKemeng Shi ac->ac_flags |= EXT4_MB_CR_POWER2_ALIGNED_OPTIMIZED; 902919eb90cSKemeng Shi read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); 903919eb90cSKemeng Shi return; 904196e402aSHarshad Shirwadkar } 905196e402aSHarshad Shirwadkar } 906196e402aSHarshad Shirwadkar read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); 907196e402aSHarshad Shirwadkar } 908196e402aSHarshad Shirwadkar 909919eb90cSKemeng Shi /* Increment cr and search again if no group is found */ 910f52f3d2bSOjaswin Mujoo *new_cr = CR_GOAL_LEN_FAST; 911196e402aSHarshad Shirwadkar } 912196e402aSHarshad Shirwadkar 913196e402aSHarshad Shirwadkar /* 914856d865cSOjaswin Mujoo * Find a suitable group of given order from the average fragments list. 915856d865cSOjaswin Mujoo */ 916856d865cSOjaswin Mujoo static struct ext4_group_info * 917856d865cSOjaswin Mujoo ext4_mb_find_good_group_avg_frag_lists(struct ext4_allocation_context *ac, int order) 918856d865cSOjaswin Mujoo { 919856d865cSOjaswin Mujoo struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 920856d865cSOjaswin Mujoo struct list_head *frag_list = &sbi->s_mb_avg_fragment_size[order]; 921856d865cSOjaswin Mujoo rwlock_t *frag_list_lock = &sbi->s_mb_avg_fragment_size_locks[order]; 922856d865cSOjaswin Mujoo struct ext4_group_info *grp = NULL, *iter; 923856d865cSOjaswin Mujoo enum criteria cr = ac->ac_criteria; 924856d865cSOjaswin Mujoo 925856d865cSOjaswin Mujoo if (list_empty(frag_list)) 926856d865cSOjaswin Mujoo return NULL; 927856d865cSOjaswin Mujoo read_lock(frag_list_lock); 928856d865cSOjaswin Mujoo if (list_empty(frag_list)) { 929856d865cSOjaswin Mujoo read_unlock(frag_list_lock); 930856d865cSOjaswin Mujoo return NULL; 931856d865cSOjaswin Mujoo } 932856d865cSOjaswin Mujoo list_for_each_entry(iter, frag_list, bb_avg_fragment_size_node) { 933856d865cSOjaswin Mujoo if (sbi->s_mb_stats) 934856d865cSOjaswin Mujoo atomic64_inc(&sbi->s_bal_cX_groups_considered[cr]); 935856d865cSOjaswin Mujoo if (likely(ext4_mb_good_group(ac, iter->bb_group, cr))) { 936856d865cSOjaswin Mujoo grp = iter; 937856d865cSOjaswin Mujoo break; 938856d865cSOjaswin Mujoo } 939856d865cSOjaswin Mujoo } 940856d865cSOjaswin Mujoo read_unlock(frag_list_lock); 941856d865cSOjaswin Mujoo return grp; 942856d865cSOjaswin Mujoo } 943856d865cSOjaswin Mujoo 944856d865cSOjaswin Mujoo /* 94583e80a6eSJan Kara * Choose next group by traversing average fragment size list of suitable 94683e80a6eSJan Kara * order. Updates *new_cr if cr level needs an update. 947196e402aSHarshad Shirwadkar */ 948f52f3d2bSOjaswin Mujoo static void ext4_mb_choose_next_group_goal_fast(struct ext4_allocation_context *ac, 9494eb7a4a1SOjaswin Mujoo enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 950196e402aSHarshad Shirwadkar { 951196e402aSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 952856d865cSOjaswin Mujoo struct ext4_group_info *grp = NULL; 95383e80a6eSJan Kara int i; 954196e402aSHarshad Shirwadkar 955f52f3d2bSOjaswin Mujoo if (unlikely(ac->ac_flags & EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED)) { 956196e402aSHarshad Shirwadkar if (sbi->s_mb_stats) 957f52f3d2bSOjaswin Mujoo atomic_inc(&sbi->s_bal_goal_fast_bad_suggestions); 95883e80a6eSJan Kara } 95983e80a6eSJan Kara 96083e80a6eSJan Kara for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len); 96183e80a6eSJan Kara i < MB_NUM_ORDERS(ac->ac_sb); i++) { 962856d865cSOjaswin Mujoo grp = ext4_mb_find_good_group_avg_frag_lists(ac, i); 96383e80a6eSJan Kara if (grp) { 964196e402aSHarshad Shirwadkar *group = grp->bb_group; 965f52f3d2bSOjaswin Mujoo ac->ac_flags |= EXT4_MB_CR_GOAL_LEN_FAST_OPTIMIZED; 966b50675a4SKemeng Shi return; 9677e170922SOjaswin Mujoo } 9687e170922SOjaswin Mujoo } 9697e170922SOjaswin Mujoo 970772c9f69SRitesh Harjani /* 971772c9f69SRitesh Harjani * CR_BEST_AVAIL_LEN works based on the concept that we have 972772c9f69SRitesh Harjani * a larger normalized goal len request which can be trimmed to 973772c9f69SRitesh Harjani * a smaller goal len such that it can still satisfy original 974772c9f69SRitesh Harjani * request len. However, allocation request for non-regular 975772c9f69SRitesh Harjani * files never gets normalized. 976772c9f69SRitesh Harjani * See function ext4_mb_normalize_request() (EXT4_MB_HINT_DATA). 977772c9f69SRitesh Harjani */ 978772c9f69SRitesh Harjani if (ac->ac_flags & EXT4_MB_HINT_DATA) 979b50675a4SKemeng Shi *new_cr = CR_BEST_AVAIL_LEN; 980772c9f69SRitesh Harjani else 981772c9f69SRitesh Harjani *new_cr = CR_GOAL_LEN_SLOW; 982b50675a4SKemeng Shi } 983b50675a4SKemeng Shi 9847e170922SOjaswin Mujoo /* 985f52f3d2bSOjaswin Mujoo * We couldn't find a group in CR_GOAL_LEN_FAST so try to find the highest free fragment 9867e170922SOjaswin Mujoo * order we have and proactively trim the goal request length to that order to 9877e170922SOjaswin Mujoo * find a suitable group faster. 9887e170922SOjaswin Mujoo * 9897e170922SOjaswin Mujoo * This optimizes allocation speed at the cost of slightly reduced 9907e170922SOjaswin Mujoo * preallocations. However, we make sure that we don't trim the request too 991f52f3d2bSOjaswin Mujoo * much and fall to CR_GOAL_LEN_SLOW in that case. 9927e170922SOjaswin Mujoo */ 993f52f3d2bSOjaswin Mujoo static void ext4_mb_choose_next_group_best_avail(struct ext4_allocation_context *ac, 9947e170922SOjaswin Mujoo enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 9957e170922SOjaswin Mujoo { 9967e170922SOjaswin Mujoo struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 9977e170922SOjaswin Mujoo struct ext4_group_info *grp = NULL; 9987e170922SOjaswin Mujoo int i, order, min_order; 9997e170922SOjaswin Mujoo unsigned long num_stripe_clusters = 0; 10007e170922SOjaswin Mujoo 1001f52f3d2bSOjaswin Mujoo if (unlikely(ac->ac_flags & EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED)) { 10027e170922SOjaswin Mujoo if (sbi->s_mb_stats) 1003f52f3d2bSOjaswin Mujoo atomic_inc(&sbi->s_bal_best_avail_bad_suggestions); 10047e170922SOjaswin Mujoo } 10057e170922SOjaswin Mujoo 10067e170922SOjaswin Mujoo /* 10077e170922SOjaswin Mujoo * mb_avg_fragment_size_order() returns order in a way that makes 10087e170922SOjaswin Mujoo * retrieving back the length using (1 << order) inaccurate. Hence, use 10097e170922SOjaswin Mujoo * fls() instead since we need to know the actual length while modifying 10107e170922SOjaswin Mujoo * goal length. 10117e170922SOjaswin Mujoo */ 10125d5460faSOjaswin Mujoo order = fls(ac->ac_g_ex.fe_len) - 1; 1013f52f3d2bSOjaswin Mujoo min_order = order - sbi->s_mb_best_avail_max_trim_order; 10147e170922SOjaswin Mujoo if (min_order < 0) 10157e170922SOjaswin Mujoo min_order = 0; 10167e170922SOjaswin Mujoo 10177e170922SOjaswin Mujoo if (sbi->s_stripe > 0) { 10187e170922SOjaswin Mujoo /* 10197e170922SOjaswin Mujoo * We are assuming that stripe size is always a multiple of 10207e170922SOjaswin Mujoo * cluster ratio otherwise __ext4_fill_super exists early. 10217e170922SOjaswin Mujoo */ 10227e170922SOjaswin Mujoo num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe); 10237e170922SOjaswin Mujoo if (1 << min_order < num_stripe_clusters) 10245d5460faSOjaswin Mujoo /* 10255d5460faSOjaswin Mujoo * We consider 1 order less because later we round 10265d5460faSOjaswin Mujoo * up the goal len to num_stripe_clusters 10275d5460faSOjaswin Mujoo */ 10285d5460faSOjaswin Mujoo min_order = fls(num_stripe_clusters) - 1; 10297e170922SOjaswin Mujoo } 10307e170922SOjaswin Mujoo 10315d5460faSOjaswin Mujoo if (1 << min_order < ac->ac_o_ex.fe_len) 10325d5460faSOjaswin Mujoo min_order = fls(ac->ac_o_ex.fe_len); 10335d5460faSOjaswin Mujoo 10347e170922SOjaswin Mujoo for (i = order; i >= min_order; i--) { 10357e170922SOjaswin Mujoo int frag_order; 10367e170922SOjaswin Mujoo /* 10377e170922SOjaswin Mujoo * Scale down goal len to make sure we find something 10387e170922SOjaswin Mujoo * in the free fragments list. Basically, reduce 10397e170922SOjaswin Mujoo * preallocations. 10407e170922SOjaswin Mujoo */ 10417e170922SOjaswin Mujoo ac->ac_g_ex.fe_len = 1 << i; 10427e170922SOjaswin Mujoo 10437e170922SOjaswin Mujoo if (num_stripe_clusters > 0) { 10447e170922SOjaswin Mujoo /* 10454c0cfebdSTheodore Ts'o * Try to round up the adjusted goal length to 10464c0cfebdSTheodore Ts'o * stripe size (in cluster units) multiple for 10474c0cfebdSTheodore Ts'o * efficiency. 10487e170922SOjaswin Mujoo */ 10497e170922SOjaswin Mujoo ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len, 10507e170922SOjaswin Mujoo num_stripe_clusters); 10517e170922SOjaswin Mujoo } 10527e170922SOjaswin Mujoo 10537e170922SOjaswin Mujoo frag_order = mb_avg_fragment_size_order(ac->ac_sb, 10547e170922SOjaswin Mujoo ac->ac_g_ex.fe_len); 10557e170922SOjaswin Mujoo 10567e170922SOjaswin Mujoo grp = ext4_mb_find_good_group_avg_frag_lists(ac, frag_order); 10577e170922SOjaswin Mujoo if (grp) { 10587e170922SOjaswin Mujoo *group = grp->bb_group; 1059f52f3d2bSOjaswin Mujoo ac->ac_flags |= EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED; 1060bcb123acSKemeng Shi return; 1061bcb123acSKemeng Shi } 1062bcb123acSKemeng Shi } 1063bcb123acSKemeng Shi 1064f52f3d2bSOjaswin Mujoo /* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */ 10657e170922SOjaswin Mujoo ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; 1066f52f3d2bSOjaswin Mujoo *new_cr = CR_GOAL_LEN_SLOW; 1067196e402aSHarshad Shirwadkar } 1068196e402aSHarshad Shirwadkar 1069196e402aSHarshad Shirwadkar static inline int should_optimize_scan(struct ext4_allocation_context *ac) 1070196e402aSHarshad Shirwadkar { 1071196e402aSHarshad Shirwadkar if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) 1072196e402aSHarshad Shirwadkar return 0; 1073f52f3d2bSOjaswin Mujoo if (ac->ac_criteria >= CR_GOAL_LEN_SLOW) 1074196e402aSHarshad Shirwadkar return 0; 1075077d0c2cSOjaswin Mujoo if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) 1076196e402aSHarshad Shirwadkar return 0; 1077196e402aSHarshad Shirwadkar return 1; 1078196e402aSHarshad Shirwadkar } 1079196e402aSHarshad Shirwadkar 1080196e402aSHarshad Shirwadkar /* 1081196e402aSHarshad Shirwadkar * Return next linear group for allocation. If linear traversal should not be 1082196e402aSHarshad Shirwadkar * performed, this function just returns the same group 1083196e402aSHarshad Shirwadkar */ 108460c672b7SKemeng Shi static ext4_group_t 108560c672b7SKemeng Shi next_linear_group(struct ext4_allocation_context *ac, ext4_group_t group, 108660c672b7SKemeng Shi ext4_group_t ngroups) 1087196e402aSHarshad Shirwadkar { 1088196e402aSHarshad Shirwadkar if (!should_optimize_scan(ac)) 1089196e402aSHarshad Shirwadkar goto inc_and_return; 1090196e402aSHarshad Shirwadkar 1091196e402aSHarshad Shirwadkar if (ac->ac_groups_linear_remaining) { 1092196e402aSHarshad Shirwadkar ac->ac_groups_linear_remaining--; 1093196e402aSHarshad Shirwadkar goto inc_and_return; 1094196e402aSHarshad Shirwadkar } 1095196e402aSHarshad Shirwadkar 1096196e402aSHarshad Shirwadkar return group; 1097196e402aSHarshad Shirwadkar inc_and_return: 1098196e402aSHarshad Shirwadkar /* 1099196e402aSHarshad Shirwadkar * Artificially restricted ngroups for non-extent 1100196e402aSHarshad Shirwadkar * files makes group > ngroups possible on first loop. 1101196e402aSHarshad Shirwadkar */ 1102196e402aSHarshad Shirwadkar return group + 1 >= ngroups ? 0 : group + 1; 1103196e402aSHarshad Shirwadkar } 1104196e402aSHarshad Shirwadkar 1105196e402aSHarshad Shirwadkar /* 1106196e402aSHarshad Shirwadkar * ext4_mb_choose_next_group: choose next group for allocation. 1107196e402aSHarshad Shirwadkar * 1108196e402aSHarshad Shirwadkar * @ac Allocation Context 1109196e402aSHarshad Shirwadkar * @new_cr This is an output parameter. If the there is no good group 1110196e402aSHarshad Shirwadkar * available at current CR level, this field is updated to indicate 1111196e402aSHarshad Shirwadkar * the new cr level that should be used. 1112196e402aSHarshad Shirwadkar * @group This is an input / output parameter. As an input it indicates the 1113196e402aSHarshad Shirwadkar * next group that the allocator intends to use for allocation. As 1114196e402aSHarshad Shirwadkar * output, this field indicates the next group that should be used as 1115196e402aSHarshad Shirwadkar * determined by the optimization functions. 1116196e402aSHarshad Shirwadkar * @ngroups Total number of groups 1117196e402aSHarshad Shirwadkar */ 1118196e402aSHarshad Shirwadkar static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, 11194eb7a4a1SOjaswin Mujoo enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 1120196e402aSHarshad Shirwadkar { 1121196e402aSHarshad Shirwadkar *new_cr = ac->ac_criteria; 1122196e402aSHarshad Shirwadkar 11234fca50d4SJan Kara if (!should_optimize_scan(ac) || ac->ac_groups_linear_remaining) { 11244fca50d4SJan Kara *group = next_linear_group(ac, *group, ngroups); 1125196e402aSHarshad Shirwadkar return; 11264fca50d4SJan Kara } 1127196e402aSHarshad Shirwadkar 1128f52f3d2bSOjaswin Mujoo if (*new_cr == CR_POWER2_ALIGNED) { 1129f52f3d2bSOjaswin Mujoo ext4_mb_choose_next_group_p2_aligned(ac, new_cr, group, ngroups); 1130f52f3d2bSOjaswin Mujoo } else if (*new_cr == CR_GOAL_LEN_FAST) { 1131f52f3d2bSOjaswin Mujoo ext4_mb_choose_next_group_goal_fast(ac, new_cr, group, ngroups); 1132f52f3d2bSOjaswin Mujoo } else if (*new_cr == CR_BEST_AVAIL_LEN) { 1133f52f3d2bSOjaswin Mujoo ext4_mb_choose_next_group_best_avail(ac, new_cr, group, ngroups); 1134196e402aSHarshad Shirwadkar } else { 1135196e402aSHarshad Shirwadkar /* 1136196e402aSHarshad Shirwadkar * TODO: For CR=2, we can arrange groups in an rb tree sorted by 1137196e402aSHarshad Shirwadkar * bb_free. But until that happens, we should never come here. 1138196e402aSHarshad Shirwadkar */ 1139196e402aSHarshad Shirwadkar WARN_ON(1); 1140196e402aSHarshad Shirwadkar } 1141196e402aSHarshad Shirwadkar } 1142196e402aSHarshad Shirwadkar 11438a57d9d6SCurt Wohlgemuth /* 11448a57d9d6SCurt Wohlgemuth * Cache the order of the largest free extent we have available in this block 11458a57d9d6SCurt Wohlgemuth * group. 11468a57d9d6SCurt Wohlgemuth */ 11478a57d9d6SCurt Wohlgemuth static void 11488a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) 11498a57d9d6SCurt Wohlgemuth { 1150196e402aSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 11518a57d9d6SCurt Wohlgemuth int i; 11528a57d9d6SCurt Wohlgemuth 11531940265eSJan Kara for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) 11541940265eSJan Kara if (grp->bb_counters[i] > 0) 11551940265eSJan Kara break; 11561940265eSJan Kara /* No need to move between order lists? */ 11571940265eSJan Kara if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || 11581940265eSJan Kara i == grp->bb_largest_free_order) { 11591940265eSJan Kara grp->bb_largest_free_order = i; 11601940265eSJan Kara return; 11611940265eSJan Kara } 11621940265eSJan Kara 11631940265eSJan Kara if (grp->bb_largest_free_order >= 0) { 1164196e402aSHarshad Shirwadkar write_lock(&sbi->s_mb_largest_free_orders_locks[ 1165196e402aSHarshad Shirwadkar grp->bb_largest_free_order]); 1166196e402aSHarshad Shirwadkar list_del_init(&grp->bb_largest_free_order_node); 1167196e402aSHarshad Shirwadkar write_unlock(&sbi->s_mb_largest_free_orders_locks[ 1168196e402aSHarshad Shirwadkar grp->bb_largest_free_order]); 1169196e402aSHarshad Shirwadkar } 11708a57d9d6SCurt Wohlgemuth grp->bb_largest_free_order = i; 11711940265eSJan Kara if (grp->bb_largest_free_order >= 0 && grp->bb_free) { 1172196e402aSHarshad Shirwadkar write_lock(&sbi->s_mb_largest_free_orders_locks[ 1173196e402aSHarshad Shirwadkar grp->bb_largest_free_order]); 1174196e402aSHarshad Shirwadkar list_add_tail(&grp->bb_largest_free_order_node, 1175196e402aSHarshad Shirwadkar &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); 1176196e402aSHarshad Shirwadkar write_unlock(&sbi->s_mb_largest_free_orders_locks[ 1177196e402aSHarshad Shirwadkar grp->bb_largest_free_order]); 1178196e402aSHarshad Shirwadkar } 11798a57d9d6SCurt Wohlgemuth } 11808a57d9d6SCurt Wohlgemuth 1181089ceeccSEric Sandeen static noinline_for_stack 1182089ceeccSEric Sandeen void ext4_mb_generate_buddy(struct super_block *sb, 11835354b2afSTheodore Ts'o void *buddy, void *bitmap, ext4_group_t group, 11845354b2afSTheodore Ts'o struct ext4_group_info *grp) 1185c9de560dSAlex Tomas { 1186e43bb4e6SNamjae Jeon struct ext4_sb_info *sbi = EXT4_SB(sb); 11877137d7a4STheodore Ts'o ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 1188a36b4498SEric Sandeen ext4_grpblk_t i = 0; 1189a36b4498SEric Sandeen ext4_grpblk_t first; 1190a36b4498SEric Sandeen ext4_grpblk_t len; 1191c9de560dSAlex Tomas unsigned free = 0; 1192c9de560dSAlex Tomas unsigned fragments = 0; 1193c9de560dSAlex Tomas unsigned long long period = get_cycles(); 1194c9de560dSAlex Tomas 1195c9de560dSAlex Tomas /* initialize buddy from bitmap which is aggregation 1196c9de560dSAlex Tomas * of on-disk bitmap and preallocations */ 1197ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, 0); 1198c9de560dSAlex Tomas grp->bb_first_free = i; 1199c9de560dSAlex Tomas while (i < max) { 1200c9de560dSAlex Tomas fragments++; 1201c9de560dSAlex Tomas first = i; 1202ffad0a44SAneesh Kumar K.V i = mb_find_next_bit(bitmap, max, i); 1203c9de560dSAlex Tomas len = i - first; 1204c9de560dSAlex Tomas free += len; 1205c9de560dSAlex Tomas if (len > 1) 1206c9de560dSAlex Tomas ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 1207c9de560dSAlex Tomas else 1208c9de560dSAlex Tomas grp->bb_counters[0]++; 1209c9de560dSAlex Tomas if (i < max) 1210ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, i); 1211c9de560dSAlex Tomas } 1212c9de560dSAlex Tomas grp->bb_fragments = fragments; 1213c9de560dSAlex Tomas 1214c9de560dSAlex Tomas if (free != grp->bb_free) { 1215e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, 121694d4c066STheodore Ts'o "block bitmap and bg descriptor " 121794d4c066STheodore Ts'o "inconsistent: %u vs %u free clusters", 1218e29136f8STheodore Ts'o free, grp->bb_free); 1219e56eb659SAneesh Kumar K.V /* 1220163a203dSDarrick J. Wong * If we intend to continue, we consider group descriptor 1221e56eb659SAneesh Kumar K.V * corrupt and update bb_free using bitmap value 1222e56eb659SAneesh Kumar K.V */ 1223c9de560dSAlex Tomas grp->bb_free = free; 1224db79e6d1SWang Shilong ext4_mark_group_bitmap_corrupted(sb, group, 1225db79e6d1SWang Shilong EXT4_GROUP_INFO_BBITMAP_CORRUPT); 1226c9de560dSAlex Tomas } 12278a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, grp); 122883e80a6eSJan Kara mb_update_avg_fragment_size(sb, grp); 1229c9de560dSAlex Tomas 1230c9de560dSAlex Tomas clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 1231c9de560dSAlex Tomas 1232c9de560dSAlex Tomas period = get_cycles() - period; 123367d25186SHarshad Shirwadkar atomic_inc(&sbi->s_mb_buddies_generated); 123467d25186SHarshad Shirwadkar atomic64_add(period, &sbi->s_mb_generation_time); 1235c9de560dSAlex Tomas } 1236c9de560dSAlex Tomas 1237c9de560dSAlex Tomas /* The buddy information is attached the buddy cache inode 1238c9de560dSAlex Tomas * for convenience. The information regarding each group 1239c9de560dSAlex Tomas * is loaded via ext4_mb_load_buddy. The information involve 1240c9de560dSAlex Tomas * block bitmap and buddy information. The information are 1241c9de560dSAlex Tomas * stored in the inode as 1242c9de560dSAlex Tomas * 1243c9de560dSAlex Tomas * { page } 1244c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 1245c9de560dSAlex Tomas * 1246c9de560dSAlex Tomas * 1247c9de560dSAlex Tomas * one block each for bitmap and buddy information. 1248c9de560dSAlex Tomas * So for each group we take up 2 blocks. A page can 1249ea1754a0SKirill A. Shutemov * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. 1250c9de560dSAlex Tomas * So it can have information regarding groups_per_page which 1251c9de560dSAlex Tomas * is blocks_per_page/2 12528a57d9d6SCurt Wohlgemuth * 12538a57d9d6SCurt Wohlgemuth * Locking note: This routine takes the block group lock of all groups 12548a57d9d6SCurt Wohlgemuth * for this page; do not hold this lock when calling this routine! 1255c9de560dSAlex Tomas */ 1256c9de560dSAlex Tomas 1257adb7ef60SKonstantin Khlebnikov static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) 1258c9de560dSAlex Tomas { 12598df9675fSTheodore Ts'o ext4_group_t ngroups; 126089cadf6eSLu Hongfei unsigned int blocksize; 1261c9de560dSAlex Tomas int blocks_per_page; 1262c9de560dSAlex Tomas int groups_per_page; 1263c9de560dSAlex Tomas int err = 0; 1264c9de560dSAlex Tomas int i; 1265813e5727STheodore Ts'o ext4_group_t first_group, group; 1266c9de560dSAlex Tomas int first_block; 1267c9de560dSAlex Tomas struct super_block *sb; 1268c9de560dSAlex Tomas struct buffer_head *bhs; 1269fa77dcfaSDarrick J. Wong struct buffer_head **bh = NULL; 1270c9de560dSAlex Tomas struct inode *inode; 1271c9de560dSAlex Tomas char *data; 1272c9de560dSAlex Tomas char *bitmap; 12739b8b7d35SAmir Goldstein struct ext4_group_info *grinfo; 1274c9de560dSAlex Tomas 1275c9de560dSAlex Tomas inode = page->mapping->host; 1276c9de560dSAlex Tomas sb = inode->i_sb; 12778df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 127893407472SFabian Frederick blocksize = i_blocksize(inode); 127909cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / blocksize; 1280c9de560dSAlex Tomas 1281d3df1453SRitesh Harjani mb_debug(sb, "init page %lu\n", page->index); 1282d3df1453SRitesh Harjani 1283c9de560dSAlex Tomas groups_per_page = blocks_per_page >> 1; 1284c9de560dSAlex Tomas if (groups_per_page == 0) 1285c9de560dSAlex Tomas groups_per_page = 1; 1286c9de560dSAlex Tomas 1287c9de560dSAlex Tomas /* allocate buffer_heads to read bitmaps */ 1288c9de560dSAlex Tomas if (groups_per_page > 1) { 1289c9de560dSAlex Tomas i = sizeof(struct buffer_head *) * groups_per_page; 1290adb7ef60SKonstantin Khlebnikov bh = kzalloc(i, gfp); 1291139f46d3SKemeng Shi if (bh == NULL) 1292139f46d3SKemeng Shi return -ENOMEM; 1293c9de560dSAlex Tomas } else 1294c9de560dSAlex Tomas bh = &bhs; 1295c9de560dSAlex Tomas 1296c9de560dSAlex Tomas first_group = page->index * blocks_per_page / 2; 1297c9de560dSAlex Tomas 1298c9de560dSAlex Tomas /* read all groups the page covers into the cache */ 1299813e5727STheodore Ts'o for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 1300813e5727STheodore Ts'o if (group >= ngroups) 1301c9de560dSAlex Tomas break; 1302c9de560dSAlex Tomas 1303813e5727STheodore Ts'o grinfo = ext4_get_group_info(sb, group); 13045354b2afSTheodore Ts'o if (!grinfo) 13055354b2afSTheodore Ts'o continue; 13069b8b7d35SAmir Goldstein /* 13079b8b7d35SAmir Goldstein * If page is uptodate then we came here after online resize 13089b8b7d35SAmir Goldstein * which added some new uninitialized group info structs, so 13099b8b7d35SAmir Goldstein * we must skip all initialized uptodate buddies on the page, 13109b8b7d35SAmir Goldstein * which may be currently in use by an allocating task. 13119b8b7d35SAmir Goldstein */ 13129b8b7d35SAmir Goldstein if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { 13139b8b7d35SAmir Goldstein bh[i] = NULL; 13149b8b7d35SAmir Goldstein continue; 13159b8b7d35SAmir Goldstein } 1316cfd73237SAlex Zhuravlev bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); 13179008a58eSDarrick J. Wong if (IS_ERR(bh[i])) { 13189008a58eSDarrick J. Wong err = PTR_ERR(bh[i]); 13199008a58eSDarrick J. Wong bh[i] = NULL; 1320c9de560dSAlex Tomas goto out; 13212ccb5fb9SAneesh Kumar K.V } 1322d3df1453SRitesh Harjani mb_debug(sb, "read bitmap for group %u\n", group); 1323c9de560dSAlex Tomas } 1324c9de560dSAlex Tomas 1325c9de560dSAlex Tomas /* wait for I/O completion */ 1326813e5727STheodore Ts'o for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 13279008a58eSDarrick J. Wong int err2; 13289008a58eSDarrick J. Wong 13299008a58eSDarrick J. Wong if (!bh[i]) 13309008a58eSDarrick J. Wong continue; 13319008a58eSDarrick J. Wong err2 = ext4_wait_block_bitmap(sb, group, bh[i]); 13329008a58eSDarrick J. Wong if (!err) 13339008a58eSDarrick J. Wong err = err2; 1334813e5727STheodore Ts'o } 1335c9de560dSAlex Tomas 1336c9de560dSAlex Tomas first_block = page->index * blocks_per_page; 1337c9de560dSAlex Tomas for (i = 0; i < blocks_per_page; i++) { 1338c9de560dSAlex Tomas group = (first_block + i) >> 1; 13398df9675fSTheodore Ts'o if (group >= ngroups) 1340c9de560dSAlex Tomas break; 1341c9de560dSAlex Tomas 13429b8b7d35SAmir Goldstein if (!bh[group - first_group]) 13439b8b7d35SAmir Goldstein /* skip initialized uptodate buddy */ 13449b8b7d35SAmir Goldstein continue; 13459b8b7d35SAmir Goldstein 1346bbdc322fSLukas Czerner if (!buffer_verified(bh[group - first_group])) 1347bbdc322fSLukas Czerner /* Skip faulty bitmaps */ 1348bbdc322fSLukas Czerner continue; 1349bbdc322fSLukas Czerner err = 0; 1350bbdc322fSLukas Czerner 1351c9de560dSAlex Tomas /* 1352c9de560dSAlex Tomas * data carry information regarding this 1353c9de560dSAlex Tomas * particular group in the format specified 1354c9de560dSAlex Tomas * above 1355c9de560dSAlex Tomas * 1356c9de560dSAlex Tomas */ 1357c9de560dSAlex Tomas data = page_address(page) + (i * blocksize); 1358c9de560dSAlex Tomas bitmap = bh[group - first_group]->b_data; 1359c9de560dSAlex Tomas 1360c9de560dSAlex Tomas /* 1361c9de560dSAlex Tomas * We place the buddy block and bitmap block 1362c9de560dSAlex Tomas * close together 1363c9de560dSAlex Tomas */ 1364c9de560dSAlex Tomas if ((first_block + i) & 1) { 1365c9de560dSAlex Tomas /* this is block of buddy */ 1366c9de560dSAlex Tomas BUG_ON(incore == NULL); 1367d3df1453SRitesh Harjani mb_debug(sb, "put buddy for group %u in page %lu/%x\n", 1368c9de560dSAlex Tomas group, page->index, i * blocksize); 1369f307333eSTheodore Ts'o trace_ext4_mb_buddy_bitmap_load(sb, group); 1370c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, group); 13715354b2afSTheodore Ts'o if (!grinfo) { 13725354b2afSTheodore Ts'o err = -EFSCORRUPTED; 13735354b2afSTheodore Ts'o goto out; 13745354b2afSTheodore Ts'o } 1375c9de560dSAlex Tomas grinfo->bb_fragments = 0; 1376c9de560dSAlex Tomas memset(grinfo->bb_counters, 0, 13771927805eSEric Sandeen sizeof(*grinfo->bb_counters) * 13784b68f6dfSHarshad Shirwadkar (MB_NUM_ORDERS(sb))); 1379c9de560dSAlex Tomas /* 1380c9de560dSAlex Tomas * incore got set to the group block bitmap below 1381c9de560dSAlex Tomas */ 13827a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, group); 13839b8b7d35SAmir Goldstein /* init the buddy */ 13849b8b7d35SAmir Goldstein memset(data, 0xff, blocksize); 13855354b2afSTheodore Ts'o ext4_mb_generate_buddy(sb, data, incore, group, grinfo); 13867a2fcbf7SAneesh Kumar K.V ext4_unlock_group(sb, group); 1387c9de560dSAlex Tomas incore = NULL; 1388c9de560dSAlex Tomas } else { 1389c9de560dSAlex Tomas /* this is block of bitmap */ 1390c9de560dSAlex Tomas BUG_ON(incore != NULL); 1391d3df1453SRitesh Harjani mb_debug(sb, "put bitmap for group %u in page %lu/%x\n", 1392c9de560dSAlex Tomas group, page->index, i * blocksize); 1393f307333eSTheodore Ts'o trace_ext4_mb_bitmap_load(sb, group); 1394c9de560dSAlex Tomas 1395c9de560dSAlex Tomas /* see comments in ext4_mb_put_pa() */ 1396c9de560dSAlex Tomas ext4_lock_group(sb, group); 1397c9de560dSAlex Tomas memcpy(data, bitmap, blocksize); 1398c9de560dSAlex Tomas 1399c9de560dSAlex Tomas /* mark all preallocated blks used in in-core bitmap */ 1400c9de560dSAlex Tomas ext4_mb_generate_from_pa(sb, data, group); 14017a2fcbf7SAneesh Kumar K.V ext4_mb_generate_from_freelist(sb, data, group); 1402c9de560dSAlex Tomas ext4_unlock_group(sb, group); 1403c9de560dSAlex Tomas 1404c9de560dSAlex Tomas /* set incore so that the buddy information can be 1405c9de560dSAlex Tomas * generated using this 1406c9de560dSAlex Tomas */ 1407c9de560dSAlex Tomas incore = data; 1408c9de560dSAlex Tomas } 1409c9de560dSAlex Tomas } 1410c9de560dSAlex Tomas SetPageUptodate(page); 1411c9de560dSAlex Tomas 1412c9de560dSAlex Tomas out: 1413c9de560dSAlex Tomas if (bh) { 14149b8b7d35SAmir Goldstein for (i = 0; i < groups_per_page; i++) 1415c9de560dSAlex Tomas brelse(bh[i]); 1416c9de560dSAlex Tomas if (bh != &bhs) 1417c9de560dSAlex Tomas kfree(bh); 1418c9de560dSAlex Tomas } 1419c9de560dSAlex Tomas return err; 1420c9de560dSAlex Tomas } 1421c9de560dSAlex Tomas 14228a57d9d6SCurt Wohlgemuth /* 14232de8807bSAmir Goldstein * Lock the buddy and bitmap pages. This make sure other parallel init_group 14242de8807bSAmir Goldstein * on the same buddy page doesn't happen whild holding the buddy page lock. 14252de8807bSAmir Goldstein * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap 14262de8807bSAmir Goldstein * are on the same page e4b->bd_buddy_page is NULL and return value is 0. 1427eee4adc7SEric Sandeen */ 14282de8807bSAmir Goldstein static int ext4_mb_get_buddy_page_lock(struct super_block *sb, 1429adb7ef60SKonstantin Khlebnikov ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) 1430eee4adc7SEric Sandeen { 14312de8807bSAmir Goldstein struct inode *inode = EXT4_SB(sb)->s_buddy_cache; 14322de8807bSAmir Goldstein int block, pnum, poff; 1433eee4adc7SEric Sandeen int blocks_per_page; 14342de8807bSAmir Goldstein struct page *page; 14352de8807bSAmir Goldstein 14362de8807bSAmir Goldstein e4b->bd_buddy_page = NULL; 14372de8807bSAmir Goldstein e4b->bd_bitmap_page = NULL; 1438eee4adc7SEric Sandeen 143909cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1440eee4adc7SEric Sandeen /* 1441eee4adc7SEric Sandeen * the buddy cache inode stores the block bitmap 1442eee4adc7SEric Sandeen * and buddy information in consecutive blocks. 1443eee4adc7SEric Sandeen * So for each group we need two blocks. 1444eee4adc7SEric Sandeen */ 1445eee4adc7SEric Sandeen block = group * 2; 1446eee4adc7SEric Sandeen pnum = block / blocks_per_page; 14472de8807bSAmir Goldstein poff = block % blocks_per_page; 1448adb7ef60SKonstantin Khlebnikov page = find_or_create_page(inode->i_mapping, pnum, gfp); 14492de8807bSAmir Goldstein if (!page) 1450c57ab39bSYounger Liu return -ENOMEM; 14512de8807bSAmir Goldstein BUG_ON(page->mapping != inode->i_mapping); 14522de8807bSAmir Goldstein e4b->bd_bitmap_page = page; 14532de8807bSAmir Goldstein e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1454eee4adc7SEric Sandeen 14552de8807bSAmir Goldstein if (blocks_per_page >= 2) { 14562de8807bSAmir Goldstein /* buddy and bitmap are on the same page */ 14572de8807bSAmir Goldstein return 0; 1458eee4adc7SEric Sandeen } 1459eee4adc7SEric Sandeen 14602de8807bSAmir Goldstein block++; 1461eee4adc7SEric Sandeen pnum = block / blocks_per_page; 1462adb7ef60SKonstantin Khlebnikov page = find_or_create_page(inode->i_mapping, pnum, gfp); 14632de8807bSAmir Goldstein if (!page) 1464c57ab39bSYounger Liu return -ENOMEM; 14652de8807bSAmir Goldstein BUG_ON(page->mapping != inode->i_mapping); 14662de8807bSAmir Goldstein e4b->bd_buddy_page = page; 14672de8807bSAmir Goldstein return 0; 1468eee4adc7SEric Sandeen } 1469eee4adc7SEric Sandeen 14702de8807bSAmir Goldstein static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) 14712de8807bSAmir Goldstein { 14722de8807bSAmir Goldstein if (e4b->bd_bitmap_page) { 14732de8807bSAmir Goldstein unlock_page(e4b->bd_bitmap_page); 147409cbfeafSKirill A. Shutemov put_page(e4b->bd_bitmap_page); 14752de8807bSAmir Goldstein } 14762de8807bSAmir Goldstein if (e4b->bd_buddy_page) { 14772de8807bSAmir Goldstein unlock_page(e4b->bd_buddy_page); 147809cbfeafSKirill A. Shutemov put_page(e4b->bd_buddy_page); 14792de8807bSAmir Goldstein } 1480eee4adc7SEric Sandeen } 1481eee4adc7SEric Sandeen 1482eee4adc7SEric Sandeen /* 14838a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 14848a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 14858a57d9d6SCurt Wohlgemuth * calling this routine! 14868a57d9d6SCurt Wohlgemuth */ 1487b6a758ecSAneesh Kumar K.V static noinline_for_stack 1488adb7ef60SKonstantin Khlebnikov int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) 1489b6a758ecSAneesh Kumar K.V { 1490b6a758ecSAneesh Kumar K.V 1491b6a758ecSAneesh Kumar K.V struct ext4_group_info *this_grp; 14922de8807bSAmir Goldstein struct ext4_buddy e4b; 14932de8807bSAmir Goldstein struct page *page; 14942de8807bSAmir Goldstein int ret = 0; 1495b6a758ecSAneesh Kumar K.V 1496b10a44c3STheodore Ts'o might_sleep(); 1497d3df1453SRitesh Harjani mb_debug(sb, "init group %u\n", group); 1498b6a758ecSAneesh Kumar K.V this_grp = ext4_get_group_info(sb, group); 14995354b2afSTheodore Ts'o if (!this_grp) 15005354b2afSTheodore Ts'o return -EFSCORRUPTED; 15015354b2afSTheodore Ts'o 1502b6a758ecSAneesh Kumar K.V /* 150308c3a813SAneesh Kumar K.V * This ensures that we don't reinit the buddy cache 150408c3a813SAneesh Kumar K.V * page which map to the group from which we are already 150508c3a813SAneesh Kumar K.V * allocating. If we are looking at the buddy cache we would 150608c3a813SAneesh Kumar K.V * have taken a reference using ext4_mb_load_buddy and that 15072de8807bSAmir Goldstein * would have pinned buddy page to page cache. 15082457aec6SMel Gorman * The call to ext4_mb_get_buddy_page_lock will mark the 15092457aec6SMel Gorman * page accessed. 1510b6a758ecSAneesh Kumar K.V */ 1511adb7ef60SKonstantin Khlebnikov ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); 15122de8807bSAmir Goldstein if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { 1513b6a758ecSAneesh Kumar K.V /* 1514b6a758ecSAneesh Kumar K.V * somebody initialized the group 1515b6a758ecSAneesh Kumar K.V * return without doing anything 1516b6a758ecSAneesh Kumar K.V */ 1517b6a758ecSAneesh Kumar K.V goto err; 1518b6a758ecSAneesh Kumar K.V } 15192de8807bSAmir Goldstein 15202de8807bSAmir Goldstein page = e4b.bd_bitmap_page; 1521adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_cache(page, NULL, gfp); 15222de8807bSAmir Goldstein if (ret) 1523b6a758ecSAneesh Kumar K.V goto err; 15242de8807bSAmir Goldstein if (!PageUptodate(page)) { 1525b6a758ecSAneesh Kumar K.V ret = -EIO; 1526b6a758ecSAneesh Kumar K.V goto err; 1527b6a758ecSAneesh Kumar K.V } 1528b6a758ecSAneesh Kumar K.V 15292de8807bSAmir Goldstein if (e4b.bd_buddy_page == NULL) { 1530b6a758ecSAneesh Kumar K.V /* 1531b6a758ecSAneesh Kumar K.V * If both the bitmap and buddy are in 1532b6a758ecSAneesh Kumar K.V * the same page we don't need to force 1533b6a758ecSAneesh Kumar K.V * init the buddy 1534b6a758ecSAneesh Kumar K.V */ 15352de8807bSAmir Goldstein ret = 0; 1536b6a758ecSAneesh Kumar K.V goto err; 1537b6a758ecSAneesh Kumar K.V } 15382de8807bSAmir Goldstein /* init buddy cache */ 15392de8807bSAmir Goldstein page = e4b.bd_buddy_page; 1540adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); 15412de8807bSAmir Goldstein if (ret) 15422de8807bSAmir Goldstein goto err; 15432de8807bSAmir Goldstein if (!PageUptodate(page)) { 1544b6a758ecSAneesh Kumar K.V ret = -EIO; 1545b6a758ecSAneesh Kumar K.V goto err; 1546b6a758ecSAneesh Kumar K.V } 1547b6a758ecSAneesh Kumar K.V err: 15482de8807bSAmir Goldstein ext4_mb_put_buddy_page_lock(&e4b); 1549b6a758ecSAneesh Kumar K.V return ret; 1550b6a758ecSAneesh Kumar K.V } 1551b6a758ecSAneesh Kumar K.V 15528a57d9d6SCurt Wohlgemuth /* 15538a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 15548a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 15558a57d9d6SCurt Wohlgemuth * calling this routine! 15568a57d9d6SCurt Wohlgemuth */ 15574ddfef7bSEric Sandeen static noinline_for_stack int 1558adb7ef60SKonstantin Khlebnikov ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, 1559adb7ef60SKonstantin Khlebnikov struct ext4_buddy *e4b, gfp_t gfp) 1560c9de560dSAlex Tomas { 1561c9de560dSAlex Tomas int blocks_per_page; 1562c9de560dSAlex Tomas int block; 1563c9de560dSAlex Tomas int pnum; 1564c9de560dSAlex Tomas int poff; 1565c9de560dSAlex Tomas struct page *page; 1566fdf6c7a7SShen Feng int ret; 1567920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1568920313a7SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 1569920313a7SAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 1570c9de560dSAlex Tomas 1571b10a44c3STheodore Ts'o might_sleep(); 1572d3df1453SRitesh Harjani mb_debug(sb, "load group %u\n", group); 1573c9de560dSAlex Tomas 157409cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1575920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 15765354b2afSTheodore Ts'o if (!grp) 15775354b2afSTheodore Ts'o return -EFSCORRUPTED; 1578c9de560dSAlex Tomas 1579c9de560dSAlex Tomas e4b->bd_blkbits = sb->s_blocksize_bits; 1580529da704STao Ma e4b->bd_info = grp; 1581c9de560dSAlex Tomas e4b->bd_sb = sb; 1582c9de560dSAlex Tomas e4b->bd_group = group; 1583c9de560dSAlex Tomas e4b->bd_buddy_page = NULL; 1584c9de560dSAlex Tomas e4b->bd_bitmap_page = NULL; 1585c9de560dSAlex Tomas 1586f41c0750SAneesh Kumar K.V if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1587f41c0750SAneesh Kumar K.V /* 1588f41c0750SAneesh Kumar K.V * we need full data about the group 1589f41c0750SAneesh Kumar K.V * to make a good selection 1590f41c0750SAneesh Kumar K.V */ 1591adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_group(sb, group, gfp); 1592f41c0750SAneesh Kumar K.V if (ret) 1593f41c0750SAneesh Kumar K.V return ret; 1594f41c0750SAneesh Kumar K.V } 1595f41c0750SAneesh Kumar K.V 1596c9de560dSAlex Tomas /* 1597c9de560dSAlex Tomas * the buddy cache inode stores the block bitmap 1598c9de560dSAlex Tomas * and buddy information in consecutive blocks. 1599c9de560dSAlex Tomas * So for each group we need two blocks. 1600c9de560dSAlex Tomas */ 1601c9de560dSAlex Tomas block = group * 2; 1602c9de560dSAlex Tomas pnum = block / blocks_per_page; 1603c9de560dSAlex Tomas poff = block % blocks_per_page; 1604c9de560dSAlex Tomas 1605c9de560dSAlex Tomas /* we could use find_or_create_page(), but it locks page 1606c9de560dSAlex Tomas * what we'd like to avoid in fast path ... */ 16072457aec6SMel Gorman page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1608c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1609c9de560dSAlex Tomas if (page) 1610920313a7SAneesh Kumar K.V /* 1611920313a7SAneesh Kumar K.V * drop the page reference and try 1612920313a7SAneesh Kumar K.V * to get the page with lock. If we 1613920313a7SAneesh Kumar K.V * are not uptodate that implies 1614920313a7SAneesh Kumar K.V * somebody just created the page but 1615920313a7SAneesh Kumar K.V * is yet to initialize the same. So 1616920313a7SAneesh Kumar K.V * wait for it to initialize. 1617920313a7SAneesh Kumar K.V */ 161809cbfeafSKirill A. Shutemov put_page(page); 1619adb7ef60SKonstantin Khlebnikov page = find_or_create_page(inode->i_mapping, pnum, gfp); 1620c9de560dSAlex Tomas if (page) { 162119b8b035STheodore Ts'o if (WARN_RATELIMIT(page->mapping != inode->i_mapping, 162219b8b035STheodore Ts'o "ext4: bitmap's paging->mapping != inode->i_mapping\n")) { 162319b8b035STheodore Ts'o /* should never happen */ 162419b8b035STheodore Ts'o unlock_page(page); 162519b8b035STheodore Ts'o ret = -EINVAL; 162619b8b035STheodore Ts'o goto err; 162719b8b035STheodore Ts'o } 1628c9de560dSAlex Tomas if (!PageUptodate(page)) { 1629adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_cache(page, NULL, gfp); 1630fdf6c7a7SShen Feng if (ret) { 1631fdf6c7a7SShen Feng unlock_page(page); 1632fdf6c7a7SShen Feng goto err; 1633fdf6c7a7SShen Feng } 1634c9de560dSAlex Tomas mb_cmp_bitmaps(e4b, page_address(page) + 1635c9de560dSAlex Tomas (poff * sb->s_blocksize)); 1636c9de560dSAlex Tomas } 1637c9de560dSAlex Tomas unlock_page(page); 1638c9de560dSAlex Tomas } 1639c9de560dSAlex Tomas } 1640c57ab39bSYounger Liu if (page == NULL) { 1641c57ab39bSYounger Liu ret = -ENOMEM; 1642c57ab39bSYounger Liu goto err; 1643c57ab39bSYounger Liu } 1644c57ab39bSYounger Liu if (!PageUptodate(page)) { 1645fdf6c7a7SShen Feng ret = -EIO; 1646c9de560dSAlex Tomas goto err; 1647fdf6c7a7SShen Feng } 16482457aec6SMel Gorman 16492457aec6SMel Gorman /* Pages marked accessed already */ 1650c9de560dSAlex Tomas e4b->bd_bitmap_page = page; 1651c9de560dSAlex Tomas e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1652c9de560dSAlex Tomas 1653c9de560dSAlex Tomas block++; 1654c9de560dSAlex Tomas pnum = block / blocks_per_page; 1655c9de560dSAlex Tomas poff = block % blocks_per_page; 1656c9de560dSAlex Tomas 16572457aec6SMel Gorman page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1658c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1659c9de560dSAlex Tomas if (page) 166009cbfeafSKirill A. Shutemov put_page(page); 1661adb7ef60SKonstantin Khlebnikov page = find_or_create_page(inode->i_mapping, pnum, gfp); 1662c9de560dSAlex Tomas if (page) { 166319b8b035STheodore Ts'o if (WARN_RATELIMIT(page->mapping != inode->i_mapping, 166419b8b035STheodore Ts'o "ext4: buddy bitmap's page->mapping != inode->i_mapping\n")) { 166519b8b035STheodore Ts'o /* should never happen */ 166619b8b035STheodore Ts'o unlock_page(page); 166719b8b035STheodore Ts'o ret = -EINVAL; 166819b8b035STheodore Ts'o goto err; 166919b8b035STheodore Ts'o } 1670fdf6c7a7SShen Feng if (!PageUptodate(page)) { 1671adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_cache(page, e4b->bd_bitmap, 1672adb7ef60SKonstantin Khlebnikov gfp); 1673fdf6c7a7SShen Feng if (ret) { 1674fdf6c7a7SShen Feng unlock_page(page); 1675fdf6c7a7SShen Feng goto err; 1676fdf6c7a7SShen Feng } 1677fdf6c7a7SShen Feng } 1678c9de560dSAlex Tomas unlock_page(page); 1679c9de560dSAlex Tomas } 1680c9de560dSAlex Tomas } 1681c57ab39bSYounger Liu if (page == NULL) { 1682c57ab39bSYounger Liu ret = -ENOMEM; 1683c57ab39bSYounger Liu goto err; 1684c57ab39bSYounger Liu } 1685c57ab39bSYounger Liu if (!PageUptodate(page)) { 1686fdf6c7a7SShen Feng ret = -EIO; 1687c9de560dSAlex Tomas goto err; 1688fdf6c7a7SShen Feng } 16892457aec6SMel Gorman 16902457aec6SMel Gorman /* Pages marked accessed already */ 1691c9de560dSAlex Tomas e4b->bd_buddy_page = page; 1692c9de560dSAlex Tomas e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 1693c9de560dSAlex Tomas 1694c9de560dSAlex Tomas return 0; 1695c9de560dSAlex Tomas 1696c9de560dSAlex Tomas err: 169726626f11SYang Ruirui if (page) 169809cbfeafSKirill A. Shutemov put_page(page); 1699c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 170009cbfeafSKirill A. Shutemov put_page(e4b->bd_bitmap_page); 1701285164b8SKemeng Shi 1702c9de560dSAlex Tomas e4b->bd_buddy = NULL; 1703c9de560dSAlex Tomas e4b->bd_bitmap = NULL; 1704fdf6c7a7SShen Feng return ret; 1705c9de560dSAlex Tomas } 1706c9de560dSAlex Tomas 1707adb7ef60SKonstantin Khlebnikov static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1708adb7ef60SKonstantin Khlebnikov struct ext4_buddy *e4b) 1709adb7ef60SKonstantin Khlebnikov { 1710adb7ef60SKonstantin Khlebnikov return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); 1711adb7ef60SKonstantin Khlebnikov } 1712adb7ef60SKonstantin Khlebnikov 1713e39e07fdSJing Zhang static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) 1714c9de560dSAlex Tomas { 1715c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 171609cbfeafSKirill A. Shutemov put_page(e4b->bd_bitmap_page); 1717c9de560dSAlex Tomas if (e4b->bd_buddy_page) 171809cbfeafSKirill A. Shutemov put_page(e4b->bd_buddy_page); 1719c9de560dSAlex Tomas } 1720c9de560dSAlex Tomas 1721c9de560dSAlex Tomas 1722c9de560dSAlex Tomas static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1723c9de560dSAlex Tomas { 1724ce3cca33SChunguang Xu int order = 1, max; 1725c9de560dSAlex Tomas void *bb; 1726c9de560dSAlex Tomas 1727c5e8f3f3STheodore Ts'o BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 1728c9de560dSAlex Tomas BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1729c9de560dSAlex Tomas 1730c9de560dSAlex Tomas while (order <= e4b->bd_blkbits + 1) { 1731ce3cca33SChunguang Xu bb = mb_find_buddy(e4b, order, &max); 1732ce3cca33SChunguang Xu if (!mb_test_bit(block >> order, bb)) { 1733c9de560dSAlex Tomas /* this block is part of buddy of order 'order' */ 1734c9de560dSAlex Tomas return order; 1735c9de560dSAlex Tomas } 1736c9de560dSAlex Tomas order++; 1737c9de560dSAlex Tomas } 1738c9de560dSAlex Tomas return 0; 1739c9de560dSAlex Tomas } 1740c9de560dSAlex Tomas 1741955ce5f5SAneesh Kumar K.V static void mb_clear_bits(void *bm, int cur, int len) 1742c9de560dSAlex Tomas { 1743c9de560dSAlex Tomas __u32 *addr; 1744c9de560dSAlex Tomas 1745c9de560dSAlex Tomas len = cur + len; 1746c9de560dSAlex Tomas while (cur < len) { 1747c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1748c9de560dSAlex Tomas /* fast path: clear whole word at once */ 1749c9de560dSAlex Tomas addr = bm + (cur >> 3); 1750c9de560dSAlex Tomas *addr = 0; 1751c9de560dSAlex Tomas cur += 32; 1752c9de560dSAlex Tomas continue; 1753c9de560dSAlex Tomas } 1754e8134b27SAneesh Kumar K.V mb_clear_bit(cur, bm); 1755c9de560dSAlex Tomas cur++; 1756c9de560dSAlex Tomas } 1757c9de560dSAlex Tomas } 1758c9de560dSAlex Tomas 1759eabe0444SAndrey Sidorov /* clear bits in given range 1760eabe0444SAndrey Sidorov * will return first found zero bit if any, -1 otherwise 1761eabe0444SAndrey Sidorov */ 1762eabe0444SAndrey Sidorov static int mb_test_and_clear_bits(void *bm, int cur, int len) 1763eabe0444SAndrey Sidorov { 1764eabe0444SAndrey Sidorov __u32 *addr; 1765eabe0444SAndrey Sidorov int zero_bit = -1; 1766eabe0444SAndrey Sidorov 1767eabe0444SAndrey Sidorov len = cur + len; 1768eabe0444SAndrey Sidorov while (cur < len) { 1769eabe0444SAndrey Sidorov if ((cur & 31) == 0 && (len - cur) >= 32) { 1770eabe0444SAndrey Sidorov /* fast path: clear whole word at once */ 1771eabe0444SAndrey Sidorov addr = bm + (cur >> 3); 1772eabe0444SAndrey Sidorov if (*addr != (__u32)(-1) && zero_bit == -1) 1773eabe0444SAndrey Sidorov zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); 1774eabe0444SAndrey Sidorov *addr = 0; 1775eabe0444SAndrey Sidorov cur += 32; 1776eabe0444SAndrey Sidorov continue; 1777eabe0444SAndrey Sidorov } 1778eabe0444SAndrey Sidorov if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) 1779eabe0444SAndrey Sidorov zero_bit = cur; 1780eabe0444SAndrey Sidorov cur++; 1781eabe0444SAndrey Sidorov } 1782eabe0444SAndrey Sidorov 1783eabe0444SAndrey Sidorov return zero_bit; 1784eabe0444SAndrey Sidorov } 1785eabe0444SAndrey Sidorov 1786123e3016SRitesh Harjani void mb_set_bits(void *bm, int cur, int len) 1787c9de560dSAlex Tomas { 1788c9de560dSAlex Tomas __u32 *addr; 1789c9de560dSAlex Tomas 1790c9de560dSAlex Tomas len = cur + len; 1791c9de560dSAlex Tomas while (cur < len) { 1792c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1793c9de560dSAlex Tomas /* fast path: set whole word at once */ 1794c9de560dSAlex Tomas addr = bm + (cur >> 3); 1795c9de560dSAlex Tomas *addr = 0xffffffff; 1796c9de560dSAlex Tomas cur += 32; 1797c9de560dSAlex Tomas continue; 1798c9de560dSAlex Tomas } 1799e8134b27SAneesh Kumar K.V mb_set_bit(cur, bm); 1800c9de560dSAlex Tomas cur++; 1801c9de560dSAlex Tomas } 1802c9de560dSAlex Tomas } 1803c9de560dSAlex Tomas 1804eabe0444SAndrey Sidorov static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) 1805eabe0444SAndrey Sidorov { 1806eabe0444SAndrey Sidorov if (mb_test_bit(*bit + side, bitmap)) { 1807eabe0444SAndrey Sidorov mb_clear_bit(*bit, bitmap); 1808eabe0444SAndrey Sidorov (*bit) -= side; 1809eabe0444SAndrey Sidorov return 1; 1810eabe0444SAndrey Sidorov } 1811eabe0444SAndrey Sidorov else { 1812eabe0444SAndrey Sidorov (*bit) += side; 1813eabe0444SAndrey Sidorov mb_set_bit(*bit, bitmap); 1814eabe0444SAndrey Sidorov return -1; 1815eabe0444SAndrey Sidorov } 1816eabe0444SAndrey Sidorov } 1817eabe0444SAndrey Sidorov 1818eabe0444SAndrey Sidorov static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) 1819eabe0444SAndrey Sidorov { 1820eabe0444SAndrey Sidorov int max; 1821eabe0444SAndrey Sidorov int order = 1; 1822eabe0444SAndrey Sidorov void *buddy = mb_find_buddy(e4b, order, &max); 1823eabe0444SAndrey Sidorov 1824eabe0444SAndrey Sidorov while (buddy) { 1825eabe0444SAndrey Sidorov void *buddy2; 1826eabe0444SAndrey Sidorov 1827eabe0444SAndrey Sidorov /* Bits in range [first; last] are known to be set since 1828eabe0444SAndrey Sidorov * corresponding blocks were allocated. Bits in range 1829eabe0444SAndrey Sidorov * (first; last) will stay set because they form buddies on 1830eabe0444SAndrey Sidorov * upper layer. We just deal with borders if they don't 1831eabe0444SAndrey Sidorov * align with upper layer and then go up. 1832eabe0444SAndrey Sidorov * Releasing entire group is all about clearing 1833eabe0444SAndrey Sidorov * single bit of highest order buddy. 1834eabe0444SAndrey Sidorov */ 1835eabe0444SAndrey Sidorov 1836eabe0444SAndrey Sidorov /* Example: 1837eabe0444SAndrey Sidorov * --------------------------------- 1838eabe0444SAndrey Sidorov * | 1 | 1 | 1 | 1 | 1839eabe0444SAndrey Sidorov * --------------------------------- 1840eabe0444SAndrey Sidorov * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1841eabe0444SAndrey Sidorov * --------------------------------- 1842eabe0444SAndrey Sidorov * 0 1 2 3 4 5 6 7 1843eabe0444SAndrey Sidorov * \_____________________/ 1844eabe0444SAndrey Sidorov * 1845eabe0444SAndrey Sidorov * Neither [1] nor [6] is aligned to above layer. 1846eabe0444SAndrey Sidorov * Left neighbour [0] is free, so mark it busy, 1847eabe0444SAndrey Sidorov * decrease bb_counters and extend range to 1848eabe0444SAndrey Sidorov * [0; 6] 1849eabe0444SAndrey Sidorov * Right neighbour [7] is busy. It can't be coaleasced with [6], so 1850eabe0444SAndrey Sidorov * mark [6] free, increase bb_counters and shrink range to 1851eabe0444SAndrey Sidorov * [0; 5]. 1852eabe0444SAndrey Sidorov * Then shift range to [0; 2], go up and do the same. 1853eabe0444SAndrey Sidorov */ 1854eabe0444SAndrey Sidorov 1855eabe0444SAndrey Sidorov 1856eabe0444SAndrey Sidorov if (first & 1) 1857eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); 1858eabe0444SAndrey Sidorov if (!(last & 1)) 1859eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); 1860eabe0444SAndrey Sidorov if (first > last) 1861eabe0444SAndrey Sidorov break; 1862eabe0444SAndrey Sidorov order++; 1863eabe0444SAndrey Sidorov 1864976620bdSKemeng Shi buddy2 = mb_find_buddy(e4b, order, &max); 1865976620bdSKemeng Shi if (!buddy2) { 1866eabe0444SAndrey Sidorov mb_clear_bits(buddy, first, last - first + 1); 1867eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order - 1] += last - first + 1; 1868eabe0444SAndrey Sidorov break; 1869eabe0444SAndrey Sidorov } 1870eabe0444SAndrey Sidorov first >>= 1; 1871eabe0444SAndrey Sidorov last >>= 1; 1872eabe0444SAndrey Sidorov buddy = buddy2; 1873eabe0444SAndrey Sidorov } 1874eabe0444SAndrey Sidorov } 1875eabe0444SAndrey Sidorov 18767e5a8cddSShen Feng static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1877c9de560dSAlex Tomas int first, int count) 1878c9de560dSAlex Tomas { 1879eabe0444SAndrey Sidorov int left_is_free = 0; 1880eabe0444SAndrey Sidorov int right_is_free = 0; 1881eabe0444SAndrey Sidorov int block; 1882eabe0444SAndrey Sidorov int last = first + count - 1; 1883c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 1884c9de560dSAlex Tomas 1885c99d1e6eSTheodore Ts'o if (WARN_ON(count == 0)) 1886c99d1e6eSTheodore Ts'o return; 1887eabe0444SAndrey Sidorov BUG_ON(last >= (sb->s_blocksize << 3)); 1888bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1889163a203dSDarrick J. Wong /* Don't bother if the block group is corrupt. */ 1890163a203dSDarrick J. Wong if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 1891163a203dSDarrick J. Wong return; 1892163a203dSDarrick J. Wong 1893c9de560dSAlex Tomas mb_check_buddy(e4b); 1894c9de560dSAlex Tomas mb_free_blocks_double(inode, e4b, first, count); 1895c9de560dSAlex Tomas 189607b5b8e1SRitesh Harjani this_cpu_inc(discard_pa_seq); 1897c9de560dSAlex Tomas e4b->bd_info->bb_free += count; 1898c9de560dSAlex Tomas if (first < e4b->bd_info->bb_first_free) 1899c9de560dSAlex Tomas e4b->bd_info->bb_first_free = first; 1900c9de560dSAlex Tomas 1901eabe0444SAndrey Sidorov /* access memory sequentially: check left neighbour, 1902eabe0444SAndrey Sidorov * clear range and then check right neighbour 1903eabe0444SAndrey Sidorov */ 1904c9de560dSAlex Tomas if (first != 0) 1905eabe0444SAndrey Sidorov left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); 1906eabe0444SAndrey Sidorov block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); 1907eabe0444SAndrey Sidorov if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) 1908eabe0444SAndrey Sidorov right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); 1909c9de560dSAlex Tomas 1910eabe0444SAndrey Sidorov if (unlikely(block != -1)) { 1911e43bb4e6SNamjae Jeon struct ext4_sb_info *sbi = EXT4_SB(sb); 1912c9de560dSAlex Tomas ext4_fsblk_t blocknr; 19135661bd68SAkinobu Mita 19145661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 191549598e04SJun Piao blocknr += EXT4_C2B(sbi, block); 19168016e29fSHarshad Shirwadkar if (!(sbi->s_mount_state & EXT4_FC_REPLAY)) { 19175d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 1918e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 1919e29136f8STheodore Ts'o blocknr, 19208016e29fSHarshad Shirwadkar "freeing already freed block (bit %u); block bitmap corrupt.", 1921163a203dSDarrick J. Wong block); 19228016e29fSHarshad Shirwadkar ext4_mark_group_bitmap_corrupted( 19238016e29fSHarshad Shirwadkar sb, e4b->bd_group, 1924db79e6d1SWang Shilong EXT4_GROUP_INFO_BBITMAP_CORRUPT); 19258016e29fSHarshad Shirwadkar } 1926eabe0444SAndrey Sidorov goto done; 1927c9de560dSAlex Tomas } 1928c9de560dSAlex Tomas 1929eabe0444SAndrey Sidorov /* let's maintain fragments counter */ 1930eabe0444SAndrey Sidorov if (left_is_free && right_is_free) 1931eabe0444SAndrey Sidorov e4b->bd_info->bb_fragments--; 1932eabe0444SAndrey Sidorov else if (!left_is_free && !right_is_free) 1933eabe0444SAndrey Sidorov e4b->bd_info->bb_fragments++; 1934c9de560dSAlex Tomas 1935eabe0444SAndrey Sidorov /* buddy[0] == bd_bitmap is a special case, so handle 1936eabe0444SAndrey Sidorov * it right away and let mb_buddy_mark_free stay free of 1937eabe0444SAndrey Sidorov * zero order checks. 1938eabe0444SAndrey Sidorov * Check if neighbours are to be coaleasced, 1939eabe0444SAndrey Sidorov * adjust bitmap bb_counters and borders appropriately. 1940eabe0444SAndrey Sidorov */ 1941eabe0444SAndrey Sidorov if (first & 1) { 1942eabe0444SAndrey Sidorov first += !left_is_free; 1943eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; 1944c9de560dSAlex Tomas } 1945eabe0444SAndrey Sidorov if (!(last & 1)) { 1946eabe0444SAndrey Sidorov last -= !right_is_free; 1947eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; 1948c9de560dSAlex Tomas } 1949eabe0444SAndrey Sidorov 1950eabe0444SAndrey Sidorov if (first <= last) 1951eabe0444SAndrey Sidorov mb_buddy_mark_free(e4b, first >> 1, last >> 1); 1952eabe0444SAndrey Sidorov 1953eabe0444SAndrey Sidorov done: 19548a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, e4b->bd_info); 1955196e402aSHarshad Shirwadkar mb_update_avg_fragment_size(sb, e4b->bd_info); 1956c9de560dSAlex Tomas mb_check_buddy(e4b); 1957c9de560dSAlex Tomas } 1958c9de560dSAlex Tomas 195915c006a2SRobin Dong static int mb_find_extent(struct ext4_buddy *e4b, int block, 1960c9de560dSAlex Tomas int needed, struct ext4_free_extent *ex) 1961c9de560dSAlex Tomas { 1962c9de560dSAlex Tomas int next = block; 196315c006a2SRobin Dong int max, order; 1964c9de560dSAlex Tomas void *buddy; 1965c9de560dSAlex Tomas 1966bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1967c9de560dSAlex Tomas BUG_ON(ex == NULL); 1968c9de560dSAlex Tomas 196915c006a2SRobin Dong buddy = mb_find_buddy(e4b, 0, &max); 1970c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1971c9de560dSAlex Tomas BUG_ON(block >= max); 1972c9de560dSAlex Tomas if (mb_test_bit(block, buddy)) { 1973c9de560dSAlex Tomas ex->fe_len = 0; 1974c9de560dSAlex Tomas ex->fe_start = 0; 1975c9de560dSAlex Tomas ex->fe_group = 0; 1976c9de560dSAlex Tomas return 0; 1977c9de560dSAlex Tomas } 1978c9de560dSAlex Tomas 1979c9de560dSAlex Tomas /* find actual order */ 1980c9de560dSAlex Tomas order = mb_find_order_for_block(e4b, block); 1981c9de560dSAlex Tomas block = block >> order; 1982c9de560dSAlex Tomas 1983c9de560dSAlex Tomas ex->fe_len = 1 << order; 1984c9de560dSAlex Tomas ex->fe_start = block << order; 1985c9de560dSAlex Tomas ex->fe_group = e4b->bd_group; 1986c9de560dSAlex Tomas 1987c9de560dSAlex Tomas /* calc difference from given start */ 1988c9de560dSAlex Tomas next = next - ex->fe_start; 1989c9de560dSAlex Tomas ex->fe_len -= next; 1990c9de560dSAlex Tomas ex->fe_start += next; 1991c9de560dSAlex Tomas 1992c9de560dSAlex Tomas while (needed > ex->fe_len && 1993d8ec0c39SAlan Cox mb_find_buddy(e4b, order, &max)) { 1994c9de560dSAlex Tomas 1995c9de560dSAlex Tomas if (block + 1 >= max) 1996c9de560dSAlex Tomas break; 1997c9de560dSAlex Tomas 1998c9de560dSAlex Tomas next = (block + 1) * (1 << order); 1999c5e8f3f3STheodore Ts'o if (mb_test_bit(next, e4b->bd_bitmap)) 2000c9de560dSAlex Tomas break; 2001c9de560dSAlex Tomas 2002b051d8dcSRobin Dong order = mb_find_order_for_block(e4b, next); 2003c9de560dSAlex Tomas 2004c9de560dSAlex Tomas block = next >> order; 2005c9de560dSAlex Tomas ex->fe_len += 1 << order; 2006c9de560dSAlex Tomas } 2007c9de560dSAlex Tomas 200831562b95SJan Kara if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { 200943c73221STheodore Ts'o /* Should never happen! (but apparently sometimes does?!?) */ 201043c73221STheodore Ts'o WARN_ON(1); 2011cd84bbbaSStephen Brennan ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, 2012cd84bbbaSStephen Brennan "corruption or bug in mb_find_extent " 201343c73221STheodore Ts'o "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", 201443c73221STheodore Ts'o block, order, needed, ex->fe_group, ex->fe_start, 201543c73221STheodore Ts'o ex->fe_len, ex->fe_logical); 201643c73221STheodore Ts'o ex->fe_len = 0; 201743c73221STheodore Ts'o ex->fe_start = 0; 201843c73221STheodore Ts'o ex->fe_group = 0; 201943c73221STheodore Ts'o } 2020c9de560dSAlex Tomas return ex->fe_len; 2021c9de560dSAlex Tomas } 2022c9de560dSAlex Tomas 2023c9de560dSAlex Tomas static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 2024c9de560dSAlex Tomas { 2025c9de560dSAlex Tomas int ord; 2026c9de560dSAlex Tomas int mlen = 0; 2027c9de560dSAlex Tomas int max = 0; 2028c9de560dSAlex Tomas int cur; 2029c9de560dSAlex Tomas int start = ex->fe_start; 2030c9de560dSAlex Tomas int len = ex->fe_len; 2031c9de560dSAlex Tomas unsigned ret = 0; 2032c9de560dSAlex Tomas int len0 = len; 2033c9de560dSAlex Tomas void *buddy; 2034218a6944Shanjinke bool split = false; 2035c9de560dSAlex Tomas 2036c9de560dSAlex Tomas BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 2037c9de560dSAlex Tomas BUG_ON(e4b->bd_group != ex->fe_group); 2038bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 2039c9de560dSAlex Tomas mb_check_buddy(e4b); 2040c9de560dSAlex Tomas mb_mark_used_double(e4b, start, len); 2041c9de560dSAlex Tomas 204207b5b8e1SRitesh Harjani this_cpu_inc(discard_pa_seq); 2043c9de560dSAlex Tomas e4b->bd_info->bb_free -= len; 2044c9de560dSAlex Tomas if (e4b->bd_info->bb_first_free == start) 2045c9de560dSAlex Tomas e4b->bd_info->bb_first_free += len; 2046c9de560dSAlex Tomas 2047c9de560dSAlex Tomas /* let's maintain fragments counter */ 2048c9de560dSAlex Tomas if (start != 0) 2049c5e8f3f3STheodore Ts'o mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); 2050c9de560dSAlex Tomas if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 2051c5e8f3f3STheodore Ts'o max = !mb_test_bit(start + len, e4b->bd_bitmap); 2052c9de560dSAlex Tomas if (mlen && max) 2053c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 2054c9de560dSAlex Tomas else if (!mlen && !max) 2055c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 2056c9de560dSAlex Tomas 2057c9de560dSAlex Tomas /* let's maintain buddy itself */ 2058c9de560dSAlex Tomas while (len) { 2059218a6944Shanjinke if (!split) 2060c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, start); 2061c9de560dSAlex Tomas 2062c9de560dSAlex Tomas if (((start >> ord) << ord) == start && len >= (1 << ord)) { 2063c9de560dSAlex Tomas /* the whole chunk may be allocated at once! */ 2064c9de560dSAlex Tomas mlen = 1 << ord; 2065218a6944Shanjinke if (!split) 2066c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 2067218a6944Shanjinke else 2068218a6944Shanjinke split = false; 2069c9de560dSAlex Tomas BUG_ON((start >> ord) >= max); 2070c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 2071c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 2072c9de560dSAlex Tomas start += mlen; 2073c9de560dSAlex Tomas len -= mlen; 2074c9de560dSAlex Tomas BUG_ON(len < 0); 2075c9de560dSAlex Tomas continue; 2076c9de560dSAlex Tomas } 2077c9de560dSAlex Tomas 2078c9de560dSAlex Tomas /* store for history */ 2079c9de560dSAlex Tomas if (ret == 0) 2080c9de560dSAlex Tomas ret = len | (ord << 16); 2081c9de560dSAlex Tomas 2082c9de560dSAlex Tomas /* we have to split large buddy */ 2083c9de560dSAlex Tomas BUG_ON(ord <= 0); 2084c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 2085c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 2086c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 2087c9de560dSAlex Tomas 2088c9de560dSAlex Tomas ord--; 2089c9de560dSAlex Tomas cur = (start >> ord) & ~1U; 2090c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 2091c9de560dSAlex Tomas mb_clear_bit(cur, buddy); 2092c9de560dSAlex Tomas mb_clear_bit(cur + 1, buddy); 2093c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 2094c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 2095218a6944Shanjinke split = true; 2096c9de560dSAlex Tomas } 20978a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); 2098c9de560dSAlex Tomas 2099196e402aSHarshad Shirwadkar mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); 2100123e3016SRitesh Harjani mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0); 2101c9de560dSAlex Tomas mb_check_buddy(e4b); 2102c9de560dSAlex Tomas 2103c9de560dSAlex Tomas return ret; 2104c9de560dSAlex Tomas } 2105c9de560dSAlex Tomas 2106c9de560dSAlex Tomas /* 2107c9de560dSAlex Tomas * Must be called under group lock! 2108c9de560dSAlex Tomas */ 2109c9de560dSAlex Tomas static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 2110c9de560dSAlex Tomas struct ext4_buddy *e4b) 2111c9de560dSAlex Tomas { 2112c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2113c9de560dSAlex Tomas int ret; 2114c9de560dSAlex Tomas 2115c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 2116c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2117c9de560dSAlex Tomas 2118c9de560dSAlex Tomas ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 2119c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 2120c9de560dSAlex Tomas ret = mb_mark_used(e4b, &ac->ac_b_ex); 2121c9de560dSAlex Tomas 2122c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 2123c9de560dSAlex Tomas * allocated blocks for history */ 2124c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 2125c9de560dSAlex Tomas 2126c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 2127c9de560dSAlex Tomas ac->ac_tail = ret & 0xffff; 2128c9de560dSAlex Tomas ac->ac_buddy = ret >> 16; 2129c9de560dSAlex Tomas 2130c3a326a6SAneesh Kumar K.V /* 2131c3a326a6SAneesh Kumar K.V * take the page reference. We want the page to be pinned 2132c3a326a6SAneesh Kumar K.V * so that we don't get a ext4_mb_init_cache_call for this 2133c3a326a6SAneesh Kumar K.V * group until we update the bitmap. That would mean we 2134c3a326a6SAneesh Kumar K.V * double allocate blocks. The reference is dropped 2135c3a326a6SAneesh Kumar K.V * in ext4_mb_release_context 2136c3a326a6SAneesh Kumar K.V */ 2137c9de560dSAlex Tomas ac->ac_bitmap_page = e4b->bd_bitmap_page; 2138c9de560dSAlex Tomas get_page(ac->ac_bitmap_page); 2139c9de560dSAlex Tomas ac->ac_buddy_page = e4b->bd_buddy_page; 2140c9de560dSAlex Tomas get_page(ac->ac_buddy_page); 2141c9de560dSAlex Tomas /* store last allocated for subsequent stream allocation */ 21424ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2143c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 2144c9de560dSAlex Tomas sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 2145c9de560dSAlex Tomas sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 2146c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 2147c9de560dSAlex Tomas } 214853f86b17SRitesh Harjani /* 214953f86b17SRitesh Harjani * As we've just preallocated more space than 215053f86b17SRitesh Harjani * user requested originally, we store allocated 215153f86b17SRitesh Harjani * space in a special descriptor. 215253f86b17SRitesh Harjani */ 215353f86b17SRitesh Harjani if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 215453f86b17SRitesh Harjani ext4_mb_new_preallocation(ac); 215553f86b17SRitesh Harjani 2156c9de560dSAlex Tomas } 2157c9de560dSAlex Tomas 2158c9de560dSAlex Tomas static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 2159c9de560dSAlex Tomas struct ext4_buddy *e4b, 2160c9de560dSAlex Tomas int finish_group) 2161c9de560dSAlex Tomas { 2162c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2163c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 2164c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 2165c9de560dSAlex Tomas 2166032115fcSAneesh Kumar K.V if (ac->ac_status == AC_STATUS_FOUND) 2167032115fcSAneesh Kumar K.V return; 2168c9de560dSAlex Tomas /* 2169c9de560dSAlex Tomas * We don't want to scan for a whole year 2170c9de560dSAlex Tomas */ 2171c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan && 2172c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2173c9de560dSAlex Tomas ac->ac_status = AC_STATUS_BREAK; 2174c9de560dSAlex Tomas return; 2175c9de560dSAlex Tomas } 2176c9de560dSAlex Tomas 2177c9de560dSAlex Tomas /* 2178c9de560dSAlex Tomas * Haven't found good chunk so far, let's continue 2179c9de560dSAlex Tomas */ 2180c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) 2181c9de560dSAlex Tomas return; 2182c9de560dSAlex Tomas 21833582e745SOjaswin Mujoo if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 2184c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2185c9de560dSAlex Tomas } 2186c9de560dSAlex Tomas 2187c9de560dSAlex Tomas /* 2188c9de560dSAlex Tomas * The routine checks whether found extent is good enough. If it is, 2189c9de560dSAlex Tomas * then the extent gets marked used and flag is set to the context 2190c9de560dSAlex Tomas * to stop scanning. Otherwise, the extent is compared with the 2191c9de560dSAlex Tomas * previous found extent and if new one is better, then it's stored 2192c9de560dSAlex Tomas * in the context. Later, the best found extent will be used, if 2193c9de560dSAlex Tomas * mballoc can't find good enough extent. 2194c9de560dSAlex Tomas * 21953582e745SOjaswin Mujoo * The algorithm used is roughly as follows: 21963582e745SOjaswin Mujoo * 21973582e745SOjaswin Mujoo * * If free extent found is exactly as big as goal, then 21983582e745SOjaswin Mujoo * stop the scan and use it immediately 21993582e745SOjaswin Mujoo * 22003582e745SOjaswin Mujoo * * If free extent found is smaller than goal, then keep retrying 22013582e745SOjaswin Mujoo * upto a max of sbi->s_mb_max_to_scan times (default 200). After 22023582e745SOjaswin Mujoo * that stop scanning and use whatever we have. 22033582e745SOjaswin Mujoo * 22043582e745SOjaswin Mujoo * * If free extent found is bigger than goal, then keep retrying 22053582e745SOjaswin Mujoo * upto a max of sbi->s_mb_min_to_scan times (default 10) before 22063582e745SOjaswin Mujoo * stopping the scan and using the extent. 22073582e745SOjaswin Mujoo * 22083582e745SOjaswin Mujoo * 2209c9de560dSAlex Tomas * FIXME: real allocation policy is to be designed yet! 2210c9de560dSAlex Tomas */ 2211c9de560dSAlex Tomas static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 2212c9de560dSAlex Tomas struct ext4_free_extent *ex, 2213c9de560dSAlex Tomas struct ext4_buddy *e4b) 2214c9de560dSAlex Tomas { 2215c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 2216c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 2217c9de560dSAlex Tomas 2218c9de560dSAlex Tomas BUG_ON(ex->fe_len <= 0); 22197137d7a4STheodore Ts'o BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 22207137d7a4STheodore Ts'o BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 2221c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 2222c9de560dSAlex Tomas 2223c9de560dSAlex Tomas ac->ac_found++; 2224fdd9a009SOjaswin Mujoo ac->ac_cX_found[ac->ac_criteria]++; 2225c9de560dSAlex Tomas 2226c9de560dSAlex Tomas /* 2227c9de560dSAlex Tomas * The special case - take what you catch first 2228c9de560dSAlex Tomas */ 2229c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2230c9de560dSAlex Tomas *bex = *ex; 2231c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2232c9de560dSAlex Tomas return; 2233c9de560dSAlex Tomas } 2234c9de560dSAlex Tomas 2235c9de560dSAlex Tomas /* 2236c9de560dSAlex Tomas * Let's check whether the chuck is good enough 2237c9de560dSAlex Tomas */ 2238c9de560dSAlex Tomas if (ex->fe_len == gex->fe_len) { 2239c9de560dSAlex Tomas *bex = *ex; 2240c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2241c9de560dSAlex Tomas return; 2242c9de560dSAlex Tomas } 2243c9de560dSAlex Tomas 2244c9de560dSAlex Tomas /* 2245c9de560dSAlex Tomas * If this is first found extent, just store it in the context 2246c9de560dSAlex Tomas */ 2247c9de560dSAlex Tomas if (bex->fe_len == 0) { 2248c9de560dSAlex Tomas *bex = *ex; 2249c9de560dSAlex Tomas return; 2250c9de560dSAlex Tomas } 2251c9de560dSAlex Tomas 2252c9de560dSAlex Tomas /* 2253c9de560dSAlex Tomas * If new found extent is better, store it in the context 2254c9de560dSAlex Tomas */ 2255c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) { 2256c9de560dSAlex Tomas /* if the request isn't satisfied, any found extent 2257c9de560dSAlex Tomas * larger than previous best one is better */ 2258c9de560dSAlex Tomas if (ex->fe_len > bex->fe_len) 2259c9de560dSAlex Tomas *bex = *ex; 2260c9de560dSAlex Tomas } else if (ex->fe_len > gex->fe_len) { 2261c9de560dSAlex Tomas /* if the request is satisfied, then we try to find 2262c9de560dSAlex Tomas * an extent that still satisfy the request, but is 2263c9de560dSAlex Tomas * smaller than previous one */ 2264c9de560dSAlex Tomas if (ex->fe_len < bex->fe_len) 2265c9de560dSAlex Tomas *bex = *ex; 2266c9de560dSAlex Tomas } 2267c9de560dSAlex Tomas 2268c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 0); 2269c9de560dSAlex Tomas } 2270c9de560dSAlex Tomas 2271089ceeccSEric Sandeen static noinline_for_stack 227285b67ffbSKemeng Shi void ext4_mb_try_best_found(struct ext4_allocation_context *ac, 2273c9de560dSAlex Tomas struct ext4_buddy *e4b) 2274c9de560dSAlex Tomas { 2275c9de560dSAlex Tomas struct ext4_free_extent ex = ac->ac_b_ex; 2276c9de560dSAlex Tomas ext4_group_t group = ex.fe_group; 2277c9de560dSAlex Tomas int max; 2278c9de560dSAlex Tomas int err; 2279c9de560dSAlex Tomas 2280c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 2281c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 2282c9de560dSAlex Tomas if (err) 228385b67ffbSKemeng Shi return; 2284c9de560dSAlex Tomas 2285c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 228615c006a2SRobin Dong max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); 2287c9de560dSAlex Tomas 2288c9de560dSAlex Tomas if (max > 0) { 2289c9de560dSAlex Tomas ac->ac_b_ex = ex; 2290c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2291c9de560dSAlex Tomas } 2292c9de560dSAlex Tomas 2293c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 2294e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 2295c9de560dSAlex Tomas } 2296c9de560dSAlex Tomas 2297089ceeccSEric Sandeen static noinline_for_stack 2298089ceeccSEric Sandeen int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 2299c9de560dSAlex Tomas struct ext4_buddy *e4b) 2300c9de560dSAlex Tomas { 2301c9de560dSAlex Tomas ext4_group_t group = ac->ac_g_ex.fe_group; 2302c9de560dSAlex Tomas int max; 2303c9de560dSAlex Tomas int err; 2304c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2305838cd0cfSYongqiang Yang struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2306c9de560dSAlex Tomas struct ext4_free_extent ex; 2307c9de560dSAlex Tomas 23085354b2afSTheodore Ts'o if (!grp) 23095354b2afSTheodore Ts'o return -EFSCORRUPTED; 231001e4ca29SKemeng Shi if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY))) 2311c9de560dSAlex Tomas return 0; 2312838cd0cfSYongqiang Yang if (grp->bb_free == 0) 2313838cd0cfSYongqiang Yang return 0; 2314c9de560dSAlex Tomas 2315c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 2316c9de560dSAlex Tomas if (err) 2317c9de560dSAlex Tomas return err; 2318c9de560dSAlex Tomas 2319163a203dSDarrick J. Wong if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { 2320163a203dSDarrick J. Wong ext4_mb_unload_buddy(e4b); 2321163a203dSDarrick J. Wong return 0; 2322163a203dSDarrick J. Wong } 2323163a203dSDarrick J. Wong 2324c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 232515c006a2SRobin Dong max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, 2326c9de560dSAlex Tomas ac->ac_g_ex.fe_len, &ex); 2327ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADFA11; /* debug value */ 2328c9de560dSAlex Tomas 2329c3defd99SKemeng Shi if (max >= ac->ac_g_ex.fe_len && 2330c3defd99SKemeng Shi ac->ac_g_ex.fe_len == EXT4_B2C(sbi, sbi->s_stripe)) { 2331c9de560dSAlex Tomas ext4_fsblk_t start; 2332c9de560dSAlex Tomas 233399c515e3SKemeng Shi start = ext4_grp_offs_to_block(ac->ac_sb, &ex); 2334c9de560dSAlex Tomas /* use do_div to get remainder (would be 64-bit modulo) */ 2335c9de560dSAlex Tomas if (do_div(start, sbi->s_stripe) == 0) { 2336c9de560dSAlex Tomas ac->ac_found++; 2337c9de560dSAlex Tomas ac->ac_b_ex = ex; 2338c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2339c9de560dSAlex Tomas } 2340c9de560dSAlex Tomas } else if (max >= ac->ac_g_ex.fe_len) { 2341c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 2342c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 2343c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 2344c9de560dSAlex Tomas ac->ac_found++; 2345c9de560dSAlex Tomas ac->ac_b_ex = ex; 2346c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2347c9de560dSAlex Tomas } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 2348c9de560dSAlex Tomas /* Sometimes, caller may want to merge even small 2349c9de560dSAlex Tomas * number of blocks to an existing extent */ 2350c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 2351c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 2352c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 2353c9de560dSAlex Tomas ac->ac_found++; 2354c9de560dSAlex Tomas ac->ac_b_ex = ex; 2355c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2356c9de560dSAlex Tomas } 2357c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 2358e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 2359c9de560dSAlex Tomas 2360c9de560dSAlex Tomas return 0; 2361c9de560dSAlex Tomas } 2362c9de560dSAlex Tomas 2363c9de560dSAlex Tomas /* 2364c9de560dSAlex Tomas * The routine scans buddy structures (not bitmap!) from given order 2365c9de560dSAlex Tomas * to max order and tries to find big enough chunk to satisfy the req 2366c9de560dSAlex Tomas */ 2367089ceeccSEric Sandeen static noinline_for_stack 2368089ceeccSEric Sandeen void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 2369c9de560dSAlex Tomas struct ext4_buddy *e4b) 2370c9de560dSAlex Tomas { 2371c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2372c9de560dSAlex Tomas struct ext4_group_info *grp = e4b->bd_info; 2373c9de560dSAlex Tomas void *buddy; 2374c9de560dSAlex Tomas int i; 2375c9de560dSAlex Tomas int k; 2376c9de560dSAlex Tomas int max; 2377c9de560dSAlex Tomas 2378c9de560dSAlex Tomas BUG_ON(ac->ac_2order <= 0); 23794b68f6dfSHarshad Shirwadkar for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { 2380c9de560dSAlex Tomas if (grp->bb_counters[i] == 0) 2381c9de560dSAlex Tomas continue; 2382c9de560dSAlex Tomas 2383c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, i, &max); 238419b8b035STheodore Ts'o if (WARN_RATELIMIT(buddy == NULL, 238519b8b035STheodore Ts'o "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i)) 238619b8b035STheodore Ts'o continue; 2387c9de560dSAlex Tomas 2388ffad0a44SAneesh Kumar K.V k = mb_find_next_zero_bit(buddy, max, 0); 2389eb576086SDmitry Monakhov if (k >= max) { 2390eb576086SDmitry Monakhov ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, 2391eb576086SDmitry Monakhov "%d free clusters of order %d. But found 0", 2392eb576086SDmitry Monakhov grp->bb_counters[i], i); 2393eb576086SDmitry Monakhov ext4_mark_group_bitmap_corrupted(ac->ac_sb, 2394eb576086SDmitry Monakhov e4b->bd_group, 2395eb576086SDmitry Monakhov EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2396eb576086SDmitry Monakhov break; 2397eb576086SDmitry Monakhov } 2398c9de560dSAlex Tomas ac->ac_found++; 2399fdd9a009SOjaswin Mujoo ac->ac_cX_found[ac->ac_criteria]++; 2400c9de560dSAlex Tomas 2401c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 1 << i; 2402c9de560dSAlex Tomas ac->ac_b_ex.fe_start = k << i; 2403c9de560dSAlex Tomas ac->ac_b_ex.fe_group = e4b->bd_group; 2404c9de560dSAlex Tomas 2405c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2406c9de560dSAlex Tomas 240753f86b17SRitesh Harjani BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); 2408c9de560dSAlex Tomas 2409c9de560dSAlex Tomas if (EXT4_SB(sb)->s_mb_stats) 2410c9de560dSAlex Tomas atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 2411c9de560dSAlex Tomas 2412c9de560dSAlex Tomas break; 2413c9de560dSAlex Tomas } 2414c9de560dSAlex Tomas } 2415c9de560dSAlex Tomas 2416c9de560dSAlex Tomas /* 2417c9de560dSAlex Tomas * The routine scans the group and measures all found extents. 2418c9de560dSAlex Tomas * In order to optimize scanning, caller must pass number of 2419c9de560dSAlex Tomas * free blocks in the group, so the routine can know upper limit. 2420c9de560dSAlex Tomas */ 2421089ceeccSEric Sandeen static noinline_for_stack 2422089ceeccSEric Sandeen void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 2423c9de560dSAlex Tomas struct ext4_buddy *e4b) 2424c9de560dSAlex Tomas { 2425c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2426c5e8f3f3STheodore Ts'o void *bitmap = e4b->bd_bitmap; 2427c9de560dSAlex Tomas struct ext4_free_extent ex; 24281b420011SOjaswin Mujoo int i, j, freelen; 2429c9de560dSAlex Tomas int free; 2430c9de560dSAlex Tomas 2431c9de560dSAlex Tomas free = e4b->bd_info->bb_free; 2432907ea529STheodore Ts'o if (WARN_ON(free <= 0)) 2433907ea529STheodore Ts'o return; 2434c9de560dSAlex Tomas 2435c9de560dSAlex Tomas i = e4b->bd_info->bb_first_free; 2436c9de560dSAlex Tomas 2437c9de560dSAlex Tomas while (free && ac->ac_status == AC_STATUS_CONTINUE) { 2438ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, 24397137d7a4STheodore Ts'o EXT4_CLUSTERS_PER_GROUP(sb), i); 24407137d7a4STheodore Ts'o if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { 244126346ff6SAneesh Kumar K.V /* 2442e56eb659SAneesh Kumar K.V * IF we have corrupt bitmap, we won't find any 244326346ff6SAneesh Kumar K.V * free blocks even though group info says we 2444b483bb77SRandy Dunlap * have free blocks 244526346ff6SAneesh Kumar K.V */ 2446e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 244753accfa9STheodore Ts'o "%d free clusters as per " 2448fde4d95aSTheodore Ts'o "group info. But bitmap says 0", 244926346ff6SAneesh Kumar K.V free); 2450736dedbbSWang Shilong ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 2451736dedbbSWang Shilong EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2452c9de560dSAlex Tomas break; 2453c9de560dSAlex Tomas } 2454c9de560dSAlex Tomas 2455304749c0SOjaswin Mujoo if (!ext4_mb_cr_expensive(ac->ac_criteria)) { 24561b420011SOjaswin Mujoo /* 2457f52f3d2bSOjaswin Mujoo * In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are 2458f52f3d2bSOjaswin Mujoo * sure that this group will have a large enough 2459f52f3d2bSOjaswin Mujoo * continuous free extent, so skip over the smaller free 2460f52f3d2bSOjaswin Mujoo * extents 24611b420011SOjaswin Mujoo */ 24621b420011SOjaswin Mujoo j = mb_find_next_bit(bitmap, 24631b420011SOjaswin Mujoo EXT4_CLUSTERS_PER_GROUP(sb), i); 24641b420011SOjaswin Mujoo freelen = j - i; 24651b420011SOjaswin Mujoo 24661b420011SOjaswin Mujoo if (freelen < ac->ac_g_ex.fe_len) { 24671b420011SOjaswin Mujoo i = j; 24681b420011SOjaswin Mujoo free -= freelen; 24691b420011SOjaswin Mujoo continue; 24701b420011SOjaswin Mujoo } 24711b420011SOjaswin Mujoo } 24721b420011SOjaswin Mujoo 247315c006a2SRobin Dong mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); 2474907ea529STheodore Ts'o if (WARN_ON(ex.fe_len <= 0)) 2475907ea529STheodore Ts'o break; 247626346ff6SAneesh Kumar K.V if (free < ex.fe_len) { 2477e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 247853accfa9STheodore Ts'o "%d free clusters as per " 2479fde4d95aSTheodore Ts'o "group info. But got %d blocks", 248026346ff6SAneesh Kumar K.V free, ex.fe_len); 2481736dedbbSWang Shilong ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 2482736dedbbSWang Shilong EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2483e56eb659SAneesh Kumar K.V /* 2484e56eb659SAneesh Kumar K.V * The number of free blocks differs. This mostly 2485e56eb659SAneesh Kumar K.V * indicate that the bitmap is corrupt. So exit 2486e56eb659SAneesh Kumar K.V * without claiming the space. 2487e56eb659SAneesh Kumar K.V */ 2488e56eb659SAneesh Kumar K.V break; 248926346ff6SAneesh Kumar K.V } 2490ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADC0DE; /* debug value */ 2491c9de560dSAlex Tomas ext4_mb_measure_extent(ac, &ex, e4b); 2492c9de560dSAlex Tomas 2493c9de560dSAlex Tomas i += ex.fe_len; 2494c9de560dSAlex Tomas free -= ex.fe_len; 2495c9de560dSAlex Tomas } 2496c9de560dSAlex Tomas 2497c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 1); 2498c9de560dSAlex Tomas } 2499c9de560dSAlex Tomas 2500c9de560dSAlex Tomas /* 2501c9de560dSAlex Tomas * This is a special case for storages like raid5 2502506bf2d8SEric Sandeen * we try to find stripe-aligned chunks for stripe-size-multiple requests 2503c9de560dSAlex Tomas */ 2504089ceeccSEric Sandeen static noinline_for_stack 2505089ceeccSEric Sandeen void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 2506c9de560dSAlex Tomas struct ext4_buddy *e4b) 2507c9de560dSAlex Tomas { 2508c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2509c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2510c5e8f3f3STheodore Ts'o void *bitmap = e4b->bd_bitmap; 2511c9de560dSAlex Tomas struct ext4_free_extent ex; 2512c9de560dSAlex Tomas ext4_fsblk_t first_group_block; 2513c9de560dSAlex Tomas ext4_fsblk_t a; 2514c3defd99SKemeng Shi ext4_grpblk_t i, stripe; 2515c9de560dSAlex Tomas int max; 2516c9de560dSAlex Tomas 2517c9de560dSAlex Tomas BUG_ON(sbi->s_stripe == 0); 2518c9de560dSAlex Tomas 2519c9de560dSAlex Tomas /* find first stripe-aligned block in group */ 25205661bd68SAkinobu Mita first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); 25215661bd68SAkinobu Mita 2522c9de560dSAlex Tomas a = first_group_block + sbi->s_stripe - 1; 2523c9de560dSAlex Tomas do_div(a, sbi->s_stripe); 2524c9de560dSAlex Tomas i = (a * sbi->s_stripe) - first_group_block; 2525c9de560dSAlex Tomas 2526c3defd99SKemeng Shi stripe = EXT4_B2C(sbi, sbi->s_stripe); 2527c3defd99SKemeng Shi i = EXT4_B2C(sbi, i); 25287137d7a4STheodore Ts'o while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { 2529c9de560dSAlex Tomas if (!mb_test_bit(i, bitmap)) { 2530c3defd99SKemeng Shi max = mb_find_extent(e4b, i, stripe, &ex); 2531c3defd99SKemeng Shi if (max >= stripe) { 2532c9de560dSAlex Tomas ac->ac_found++; 2533fdd9a009SOjaswin Mujoo ac->ac_cX_found[ac->ac_criteria]++; 2534ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADF00D; /* debug value */ 2535c9de560dSAlex Tomas ac->ac_b_ex = ex; 2536c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2537c9de560dSAlex Tomas break; 2538c9de560dSAlex Tomas } 2539c9de560dSAlex Tomas } 2540c3defd99SKemeng Shi i += stripe; 2541c9de560dSAlex Tomas } 2542c9de560dSAlex Tomas } 2543c9de560dSAlex Tomas 254442ac1848SLukas Czerner /* 25458ef123feSRitesh Harjani * This is also called BEFORE we load the buddy bitmap. 254642ac1848SLukas Czerner * Returns either 1 or 0 indicating that the group is either suitable 25478ef123feSRitesh Harjani * for the allocation or not. 254842ac1848SLukas Czerner */ 25498ef123feSRitesh Harjani static bool ext4_mb_good_group(struct ext4_allocation_context *ac, 25504eb7a4a1SOjaswin Mujoo ext4_group_t group, enum criteria cr) 2551c9de560dSAlex Tomas { 25528ef123feSRitesh Harjani ext4_grpblk_t free, fragments; 2553a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 2554c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2555c9de560dSAlex Tomas 2556f52f3d2bSOjaswin Mujoo BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS); 25578a57d9d6SCurt Wohlgemuth 2558a9ce5993SKemeng Shi if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 25598ef123feSRitesh Harjani return false; 256001fc48e8STheodore Ts'o 2561dddcd2f9Sbrookxu free = grp->bb_free; 2562dddcd2f9Sbrookxu if (free == 0) 25638ef123feSRitesh Harjani return false; 2564c9de560dSAlex Tomas 2565c9de560dSAlex Tomas fragments = grp->bb_fragments; 2566c9de560dSAlex Tomas if (fragments == 0) 25678ef123feSRitesh Harjani return false; 2568c9de560dSAlex Tomas 2569c9de560dSAlex Tomas switch (cr) { 2570f52f3d2bSOjaswin Mujoo case CR_POWER2_ALIGNED: 2571c9de560dSAlex Tomas BUG_ON(ac->ac_2order == 0); 2572c9de560dSAlex Tomas 2573a4912123STheodore Ts'o /* Avoid using the first bg of a flexgroup for data files */ 2574a4912123STheodore Ts'o if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 2575a4912123STheodore Ts'o (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 2576a4912123STheodore Ts'o ((group % flex_size) == 0)) 25778ef123feSRitesh Harjani return false; 2578a4912123STheodore Ts'o 2579dddcd2f9Sbrookxu if (free < ac->ac_g_ex.fe_len) 2580dddcd2f9Sbrookxu return false; 2581dddcd2f9Sbrookxu 25824b68f6dfSHarshad Shirwadkar if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) 25838ef123feSRitesh Harjani return true; 258440ae3487STheodore Ts'o 258540ae3487STheodore Ts'o if (grp->bb_largest_free_order < ac->ac_2order) 25868ef123feSRitesh Harjani return false; 258740ae3487STheodore Ts'o 25888ef123feSRitesh Harjani return true; 2589f52f3d2bSOjaswin Mujoo case CR_GOAL_LEN_FAST: 2590f52f3d2bSOjaswin Mujoo case CR_BEST_AVAIL_LEN: 2591c9de560dSAlex Tomas if ((free / fragments) >= ac->ac_g_ex.fe_len) 25928ef123feSRitesh Harjani return true; 2593c9de560dSAlex Tomas break; 2594f52f3d2bSOjaswin Mujoo case CR_GOAL_LEN_SLOW: 2595c9de560dSAlex Tomas if (free >= ac->ac_g_ex.fe_len) 25968ef123feSRitesh Harjani return true; 2597c9de560dSAlex Tomas break; 2598f52f3d2bSOjaswin Mujoo case CR_ANY_FREE: 25998ef123feSRitesh Harjani return true; 2600c9de560dSAlex Tomas default: 2601c9de560dSAlex Tomas BUG(); 2602c9de560dSAlex Tomas } 2603c9de560dSAlex Tomas 26048ef123feSRitesh Harjani return false; 26058ef123feSRitesh Harjani } 26068ef123feSRitesh Harjani 26078ef123feSRitesh Harjani /* 26088ef123feSRitesh Harjani * This could return negative error code if something goes wrong 26098ef123feSRitesh Harjani * during ext4_mb_init_group(). This should not be called with 26108ef123feSRitesh Harjani * ext4_lock_group() held. 2611a5fda113STheodore Ts'o * 2612a5fda113STheodore Ts'o * Note: because we are conditionally operating with the group lock in 2613a5fda113STheodore Ts'o * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this 2614a5fda113STheodore Ts'o * function using __acquire and __release. This means we need to be 2615a5fda113STheodore Ts'o * super careful before messing with the error path handling via "goto 2616a5fda113STheodore Ts'o * out"! 26178ef123feSRitesh Harjani */ 26188ef123feSRitesh Harjani static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, 26194eb7a4a1SOjaswin Mujoo ext4_group_t group, enum criteria cr) 26208ef123feSRitesh Harjani { 26218ef123feSRitesh Harjani struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 262299377830SRitesh Harjani struct super_block *sb = ac->ac_sb; 2623c1d2c7d4SAlex Zhuravlev struct ext4_sb_info *sbi = EXT4_SB(sb); 262499377830SRitesh Harjani bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; 26258ef123feSRitesh Harjani ext4_grpblk_t free; 26268ef123feSRitesh Harjani int ret = 0; 26278ef123feSRitesh Harjani 26285354b2afSTheodore Ts'o if (!grp) 26295354b2afSTheodore Ts'o return -EFSCORRUPTED; 2630a6c75eafSHarshad Shirwadkar if (sbi->s_mb_stats) 2631a6c75eafSHarshad Shirwadkar atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); 2632a5fda113STheodore Ts'o if (should_lock) { 263399377830SRitesh Harjani ext4_lock_group(sb, group); 2634a5fda113STheodore Ts'o __release(ext4_group_lock_ptr(sb, group)); 2635a5fda113STheodore Ts'o } 26368ef123feSRitesh Harjani free = grp->bb_free; 26378ef123feSRitesh Harjani if (free == 0) 26388ef123feSRitesh Harjani goto out; 2639304749c0SOjaswin Mujoo /* 2640304749c0SOjaswin Mujoo * In all criterias except CR_ANY_FREE we try to avoid groups that 2641304749c0SOjaswin Mujoo * can't possibly satisfy the full goal request due to insufficient 2642304749c0SOjaswin Mujoo * free blocks. 2643304749c0SOjaswin Mujoo */ 2644304749c0SOjaswin Mujoo if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len) 26458ef123feSRitesh Harjani goto out; 26468ef123feSRitesh Harjani if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 26478ef123feSRitesh Harjani goto out; 2648a5fda113STheodore Ts'o if (should_lock) { 2649a5fda113STheodore Ts'o __acquire(ext4_group_lock_ptr(sb, group)); 265099377830SRitesh Harjani ext4_unlock_group(sb, group); 2651a5fda113STheodore Ts'o } 26528ef123feSRitesh Harjani 26538ef123feSRitesh Harjani /* We only do this if the grp has never been initialized */ 26548ef123feSRitesh Harjani if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 2655c1d2c7d4SAlex Zhuravlev struct ext4_group_desc *gdp = 2656c1d2c7d4SAlex Zhuravlev ext4_get_group_desc(sb, group, NULL); 2657c1d2c7d4SAlex Zhuravlev int ret; 2658c1d2c7d4SAlex Zhuravlev 2659f52f3d2bSOjaswin Mujoo /* 2660f52f3d2bSOjaswin Mujoo * cr=CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic 2661f52f3d2bSOjaswin Mujoo * search to find large good chunks almost for free. If buddy 2662f52f3d2bSOjaswin Mujoo * data is not ready, then this optimization makes no sense. But 2663f52f3d2bSOjaswin Mujoo * we never skip the first block group in a flex_bg, since this 2664f52f3d2bSOjaswin Mujoo * gets used for metadata block allocation, and we want to make 2665f52f3d2bSOjaswin Mujoo * sure we locate metadata blocks in the first block group in 2666f52f3d2bSOjaswin Mujoo * the flex_bg if possible. 2667c1d2c7d4SAlex Zhuravlev */ 2668304749c0SOjaswin Mujoo if (!ext4_mb_cr_expensive(cr) && 2669c1d2c7d4SAlex Zhuravlev (!sbi->s_log_groups_per_flex || 2670c1d2c7d4SAlex Zhuravlev ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && 2671c1d2c7d4SAlex Zhuravlev !(ext4_has_group_desc_csum(sb) && 2672c1d2c7d4SAlex Zhuravlev (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) 2673c1d2c7d4SAlex Zhuravlev return 0; 2674c1d2c7d4SAlex Zhuravlev ret = ext4_mb_init_group(sb, group, GFP_NOFS); 26758ef123feSRitesh Harjani if (ret) 26768ef123feSRitesh Harjani return ret; 26778ef123feSRitesh Harjani } 26788ef123feSRitesh Harjani 2679a5fda113STheodore Ts'o if (should_lock) { 268099377830SRitesh Harjani ext4_lock_group(sb, group); 2681a5fda113STheodore Ts'o __release(ext4_group_lock_ptr(sb, group)); 2682a5fda113STheodore Ts'o } 26838ef123feSRitesh Harjani ret = ext4_mb_good_group(ac, group, cr); 26848ef123feSRitesh Harjani out: 2685a5fda113STheodore Ts'o if (should_lock) { 2686a5fda113STheodore Ts'o __acquire(ext4_group_lock_ptr(sb, group)); 268799377830SRitesh Harjani ext4_unlock_group(sb, group); 2688a5fda113STheodore Ts'o } 26898ef123feSRitesh Harjani return ret; 2690c9de560dSAlex Tomas } 2691c9de560dSAlex Tomas 2692cfd73237SAlex Zhuravlev /* 2693cfd73237SAlex Zhuravlev * Start prefetching @nr block bitmaps starting at @group. 2694cfd73237SAlex Zhuravlev * Return the next group which needs to be prefetched. 2695cfd73237SAlex Zhuravlev */ 26963d392b26STheodore Ts'o ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, 2697cfd73237SAlex Zhuravlev unsigned int nr, int *cnt) 2698cfd73237SAlex Zhuravlev { 2699cfd73237SAlex Zhuravlev ext4_group_t ngroups = ext4_get_groups_count(sb); 2700cfd73237SAlex Zhuravlev struct buffer_head *bh; 2701cfd73237SAlex Zhuravlev struct blk_plug plug; 2702cfd73237SAlex Zhuravlev 2703cfd73237SAlex Zhuravlev blk_start_plug(&plug); 2704cfd73237SAlex Zhuravlev while (nr-- > 0) { 2705cfd73237SAlex Zhuravlev struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, 2706cfd73237SAlex Zhuravlev NULL); 2707cfd73237SAlex Zhuravlev struct ext4_group_info *grp = ext4_get_group_info(sb, group); 2708cfd73237SAlex Zhuravlev 2709cfd73237SAlex Zhuravlev /* 2710cfd73237SAlex Zhuravlev * Prefetch block groups with free blocks; but don't 2711cfd73237SAlex Zhuravlev * bother if it is marked uninitialized on disk, since 2712cfd73237SAlex Zhuravlev * it won't require I/O to read. Also only try to 2713cfd73237SAlex Zhuravlev * prefetch once, so we avoid getblk() call, which can 2714cfd73237SAlex Zhuravlev * be expensive. 2715cfd73237SAlex Zhuravlev */ 27165354b2afSTheodore Ts'o if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) && 2717cfd73237SAlex Zhuravlev EXT4_MB_GRP_NEED_INIT(grp) && 27183c629604SOjaswin Mujoo ext4_free_group_clusters(sb, gdp) > 0 ) { 2719cfd73237SAlex Zhuravlev bh = ext4_read_block_bitmap_nowait(sb, group, true); 2720cfd73237SAlex Zhuravlev if (bh && !IS_ERR(bh)) { 2721cfd73237SAlex Zhuravlev if (!buffer_uptodate(bh) && cnt) 2722cfd73237SAlex Zhuravlev (*cnt)++; 2723cfd73237SAlex Zhuravlev brelse(bh); 2724cfd73237SAlex Zhuravlev } 2725cfd73237SAlex Zhuravlev } 2726cfd73237SAlex Zhuravlev if (++group >= ngroups) 2727cfd73237SAlex Zhuravlev group = 0; 2728cfd73237SAlex Zhuravlev } 2729cfd73237SAlex Zhuravlev blk_finish_plug(&plug); 2730cfd73237SAlex Zhuravlev return group; 2731cfd73237SAlex Zhuravlev } 2732cfd73237SAlex Zhuravlev 2733cfd73237SAlex Zhuravlev /* 2734cfd73237SAlex Zhuravlev * Prefetching reads the block bitmap into the buffer cache; but we 2735cfd73237SAlex Zhuravlev * need to make sure that the buddy bitmap in the page cache has been 2736cfd73237SAlex Zhuravlev * initialized. Note that ext4_mb_init_group() will block if the I/O 2737cfd73237SAlex Zhuravlev * is not yet completed, or indeed if it was not initiated by 2738cfd73237SAlex Zhuravlev * ext4_mb_prefetch did not start the I/O. 2739cfd73237SAlex Zhuravlev * 2740cfd73237SAlex Zhuravlev * TODO: We should actually kick off the buddy bitmap setup in a work 2741cfd73237SAlex Zhuravlev * queue when the buffer I/O is completed, so that we don't block 2742cfd73237SAlex Zhuravlev * waiting for the block allocation bitmap read to finish when 2743cfd73237SAlex Zhuravlev * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). 2744cfd73237SAlex Zhuravlev */ 27453d392b26STheodore Ts'o void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, 2746cfd73237SAlex Zhuravlev unsigned int nr) 2747cfd73237SAlex Zhuravlev { 274822fab984SKemeng Shi struct ext4_group_desc *gdp; 274922fab984SKemeng Shi struct ext4_group_info *grp; 2750cfd73237SAlex Zhuravlev 275122fab984SKemeng Shi while (nr-- > 0) { 2752cfd73237SAlex Zhuravlev if (!group) 2753cfd73237SAlex Zhuravlev group = ext4_get_groups_count(sb); 2754cfd73237SAlex Zhuravlev group--; 275522fab984SKemeng Shi gdp = ext4_get_group_desc(sb, group, NULL); 2756cfd73237SAlex Zhuravlev grp = ext4_get_group_info(sb, group); 2757cfd73237SAlex Zhuravlev 27585354b2afSTheodore Ts'o if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) && 27593c629604SOjaswin Mujoo ext4_free_group_clusters(sb, gdp) > 0) { 2760cfd73237SAlex Zhuravlev if (ext4_mb_init_group(sb, group, GFP_NOFS)) 2761cfd73237SAlex Zhuravlev break; 2762cfd73237SAlex Zhuravlev } 2763cfd73237SAlex Zhuravlev } 2764cfd73237SAlex Zhuravlev } 2765cfd73237SAlex Zhuravlev 27664ddfef7bSEric Sandeen static noinline_for_stack int 27674ddfef7bSEric Sandeen ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 2768c9de560dSAlex Tomas { 2769cfd73237SAlex Zhuravlev ext4_group_t prefetch_grp = 0, ngroups, group, i; 27704c0cfebdSTheodore Ts'o enum criteria new_cr, cr = CR_GOAL_LEN_FAST; 277142ac1848SLukas Czerner int err = 0, first_err = 0; 2772cfd73237SAlex Zhuravlev unsigned int nr = 0, prefetch_ios = 0; 2773c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2774c9de560dSAlex Tomas struct super_block *sb; 2775c9de560dSAlex Tomas struct ext4_buddy e4b; 277666d5e027Sbrookxu int lost; 2777c9de560dSAlex Tomas 2778c9de560dSAlex Tomas sb = ac->ac_sb; 2779c9de560dSAlex Tomas sbi = EXT4_SB(sb); 27808df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 2781fb0a387dSEric Sandeen /* non-extent files are limited to low blocks/groups */ 278212e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) 2783fb0a387dSEric Sandeen ngroups = sbi->s_blockfile_groups; 2784fb0a387dSEric Sandeen 2785c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2786c9de560dSAlex Tomas 2787c9de560dSAlex Tomas /* first, try the goal */ 2788c9de560dSAlex Tomas err = ext4_mb_find_by_goal(ac, &e4b); 2789c9de560dSAlex Tomas if (err || ac->ac_status == AC_STATUS_FOUND) 2790c9de560dSAlex Tomas goto out; 2791c9de560dSAlex Tomas 2792c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2793c9de560dSAlex Tomas goto out; 2794c9de560dSAlex Tomas 2795c9de560dSAlex Tomas /* 2796e9a3cd48Sbrookxu * ac->ac_2order is set only if the fe_len is a power of 2 27974eea9fbeSKemeng Shi * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED 27984eea9fbeSKemeng Shi * so that we try exact allocation using buddy. 2799c9de560dSAlex Tomas */ 2800c9de560dSAlex Tomas i = fls(ac->ac_g_ex.fe_len); 2801c9de560dSAlex Tomas ac->ac_2order = 0; 2802c9de560dSAlex Tomas /* 2803c9de560dSAlex Tomas * We search using buddy data only if the order of the request 2804c9de560dSAlex Tomas * is greater than equal to the sbi_s_mb_order2_reqs 2805b713a5ecSTheodore Ts'o * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 2806d9b22cf9SJan Kara * We also support searching for power-of-two requests only for 2807d9b22cf9SJan Kara * requests upto maximum buddy size we have constructed. 2808c9de560dSAlex Tomas */ 28094b68f6dfSHarshad Shirwadkar if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { 2810bb60caa2SKemeng Shi if (is_power_of_2(ac->ac_g_ex.fe_len)) 28111a5d5e5dSJeremy Cline ac->ac_2order = array_index_nospec(i - 1, 28124b68f6dfSHarshad Shirwadkar MB_NUM_ORDERS(sb)); 2813c9de560dSAlex Tomas } 2814c9de560dSAlex Tomas 28154ba74d00STheodore Ts'o /* if stream allocation is enabled, use global goal */ 28164ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2817c9de560dSAlex Tomas /* TBD: may be hot point */ 2818c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 2819c9de560dSAlex Tomas ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 2820c9de560dSAlex Tomas ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 2821c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 2822c9de560dSAlex Tomas } 28234ba74d00STheodore Ts'o 2824c9de560dSAlex Tomas /* 28254c0cfebdSTheodore Ts'o * Let's just scan groups to find more-less suitable blocks We 28264c0cfebdSTheodore Ts'o * start with CR_GOAL_LEN_FAST, unless it is power of 2 28274c0cfebdSTheodore Ts'o * aligned, in which case let's do that faster approach first. 2828c9de560dSAlex Tomas */ 28294c0cfebdSTheodore Ts'o if (ac->ac_2order) 28304c0cfebdSTheodore Ts'o cr = CR_POWER2_ALIGNED; 2831c9de560dSAlex Tomas repeat: 28324eb7a4a1SOjaswin Mujoo for (; cr < EXT4_MB_NUM_CRS && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 2833c9de560dSAlex Tomas ac->ac_criteria = cr; 2834ed8f9c75SAneesh Kumar K.V /* 2835ed8f9c75SAneesh Kumar K.V * searching for the right group start 2836ed8f9c75SAneesh Kumar K.V * from the goal value specified 2837ed8f9c75SAneesh Kumar K.V */ 2838ed8f9c75SAneesh Kumar K.V group = ac->ac_g_ex.fe_group; 2839196e402aSHarshad Shirwadkar ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups; 2840cfd73237SAlex Zhuravlev prefetch_grp = group; 2841ed8f9c75SAneesh Kumar K.V 28424fca50d4SJan Kara for (i = 0, new_cr = cr; i < ngroups; i++, 28434fca50d4SJan Kara ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) { 28444fca50d4SJan Kara int ret = 0; 2845196e402aSHarshad Shirwadkar 28462ed5724dSTheodore Ts'o cond_resched(); 2847196e402aSHarshad Shirwadkar if (new_cr != cr) { 2848196e402aSHarshad Shirwadkar cr = new_cr; 2849196e402aSHarshad Shirwadkar goto repeat; 2850196e402aSHarshad Shirwadkar } 2851c9de560dSAlex Tomas 2852cfd73237SAlex Zhuravlev /* 2853cfd73237SAlex Zhuravlev * Batch reads of the block allocation bitmaps 2854cfd73237SAlex Zhuravlev * to get multiple READs in flight; limit 28554eea9fbeSKemeng Shi * prefetching at inexpensive CR, otherwise mballoc 28564eea9fbeSKemeng Shi * can spend a lot of time loading imperfect groups 2857cfd73237SAlex Zhuravlev */ 2858cfd73237SAlex Zhuravlev if ((prefetch_grp == group) && 2859304749c0SOjaswin Mujoo (ext4_mb_cr_expensive(cr) || 2860cfd73237SAlex Zhuravlev prefetch_ios < sbi->s_mb_prefetch_limit)) { 2861cfd73237SAlex Zhuravlev nr = sbi->s_mb_prefetch; 2862cfd73237SAlex Zhuravlev if (ext4_has_feature_flex_bg(sb)) { 286382ef1370SChunguang Xu nr = 1 << sbi->s_log_groups_per_flex; 286482ef1370SChunguang Xu nr -= group & (nr - 1); 286582ef1370SChunguang Xu nr = min(nr, sbi->s_mb_prefetch); 2866cfd73237SAlex Zhuravlev } 2867cfd73237SAlex Zhuravlev prefetch_grp = ext4_mb_prefetch(sb, group, 2868cfd73237SAlex Zhuravlev nr, &prefetch_ios); 2869cfd73237SAlex Zhuravlev } 2870cfd73237SAlex Zhuravlev 28718a57d9d6SCurt Wohlgemuth /* This now checks without needing the buddy page */ 28728ef123feSRitesh Harjani ret = ext4_mb_good_group_nolock(ac, group, cr); 287342ac1848SLukas Czerner if (ret <= 0) { 287442ac1848SLukas Czerner if (!first_err) 287542ac1848SLukas Czerner first_err = ret; 2876c9de560dSAlex Tomas continue; 287742ac1848SLukas Czerner } 2878c9de560dSAlex Tomas 2879c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2880c9de560dSAlex Tomas if (err) 2881c9de560dSAlex Tomas goto out; 2882c9de560dSAlex Tomas 2883c9de560dSAlex Tomas ext4_lock_group(sb, group); 28848a57d9d6SCurt Wohlgemuth 28858a57d9d6SCurt Wohlgemuth /* 28868a57d9d6SCurt Wohlgemuth * We need to check again after locking the 28878a57d9d6SCurt Wohlgemuth * block group 28888a57d9d6SCurt Wohlgemuth */ 288942ac1848SLukas Czerner ret = ext4_mb_good_group(ac, group, cr); 28908ef123feSRitesh Harjani if (ret == 0) { 2891c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2892e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2893c9de560dSAlex Tomas continue; 2894c9de560dSAlex Tomas } 2895c9de560dSAlex Tomas 2896c9de560dSAlex Tomas ac->ac_groups_scanned++; 2897f52f3d2bSOjaswin Mujoo if (cr == CR_POWER2_ALIGNED) 2898c9de560dSAlex Tomas ext4_mb_simple_scan_group(ac, &e4b); 2899f52f3d2bSOjaswin Mujoo else if ((cr == CR_GOAL_LEN_FAST || 2900f52f3d2bSOjaswin Mujoo cr == CR_BEST_AVAIL_LEN) && 2901f52f3d2bSOjaswin Mujoo sbi->s_stripe && 2902c3defd99SKemeng Shi !(ac->ac_g_ex.fe_len % 2903c3defd99SKemeng Shi EXT4_B2C(sbi, sbi->s_stripe))) 2904c9de560dSAlex Tomas ext4_mb_scan_aligned(ac, &e4b); 2905c9de560dSAlex Tomas else 2906c9de560dSAlex Tomas ext4_mb_complex_scan_group(ac, &e4b); 2907c9de560dSAlex Tomas 2908c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2909e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2910c9de560dSAlex Tomas 2911c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_CONTINUE) 2912c9de560dSAlex Tomas break; 2913c9de560dSAlex Tomas } 2914a6c75eafSHarshad Shirwadkar /* Processed all groups and haven't found blocks */ 2915a6c75eafSHarshad Shirwadkar if (sbi->s_mb_stats && i == ngroups) 2916a6c75eafSHarshad Shirwadkar atomic64_inc(&sbi->s_bal_cX_failed[cr]); 29177e170922SOjaswin Mujoo 2918f52f3d2bSOjaswin Mujoo if (i == ngroups && ac->ac_criteria == CR_BEST_AVAIL_LEN) 29197e170922SOjaswin Mujoo /* Reset goal length to original goal length before 2920f52f3d2bSOjaswin Mujoo * falling into CR_GOAL_LEN_SLOW */ 29217e170922SOjaswin Mujoo ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; 2922c9de560dSAlex Tomas } 2923c9de560dSAlex Tomas 2924c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2925c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2926c9de560dSAlex Tomas /* 2927c9de560dSAlex Tomas * We've been searching too long. Let's try to allocate 2928c9de560dSAlex Tomas * the best chunk we've found so far 2929c9de560dSAlex Tomas */ 2930c9de560dSAlex Tomas ext4_mb_try_best_found(ac, &e4b); 2931c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_FOUND) { 2932c9de560dSAlex Tomas /* 2933c9de560dSAlex Tomas * Someone more lucky has already allocated it. 2934c9de560dSAlex Tomas * The only thing we can do is just take first 2935c9de560dSAlex Tomas * found block(s) 2936c9de560dSAlex Tomas */ 293766d5e027Sbrookxu lost = atomic_inc_return(&sbi->s_mb_lost_chunks); 293866d5e027Sbrookxu mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", 2939c55ee7d2Sbrookxu ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, 2940c55ee7d2Sbrookxu ac->ac_b_ex.fe_len, lost); 2941c55ee7d2Sbrookxu 2942c9de560dSAlex Tomas ac->ac_b_ex.fe_group = 0; 2943c9de560dSAlex Tomas ac->ac_b_ex.fe_start = 0; 2944c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 0; 2945c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 2946c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_FIRST; 2947f52f3d2bSOjaswin Mujoo cr = CR_ANY_FREE; 2948c9de560dSAlex Tomas goto repeat; 2949c9de560dSAlex Tomas } 2950c9de560dSAlex Tomas } 2951a6c75eafSHarshad Shirwadkar 2952a6c75eafSHarshad Shirwadkar if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) 2953a6c75eafSHarshad Shirwadkar atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); 2954c9de560dSAlex Tomas out: 295542ac1848SLukas Czerner if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) 295642ac1848SLukas Czerner err = first_err; 2957bbc4ec77SRitesh Harjani 2958d3df1453SRitesh Harjani mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", 2959bbc4ec77SRitesh Harjani ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, 2960bbc4ec77SRitesh Harjani ac->ac_flags, cr, err); 2961cfd73237SAlex Zhuravlev 2962cfd73237SAlex Zhuravlev if (nr) 2963cfd73237SAlex Zhuravlev ext4_mb_prefetch_fini(sb, prefetch_grp, nr); 2964cfd73237SAlex Zhuravlev 2965c9de560dSAlex Tomas return err; 2966c9de560dSAlex Tomas } 2967c9de560dSAlex Tomas 2968c9de560dSAlex Tomas static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2969c9de560dSAlex Tomas { 2970359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 2971c9de560dSAlex Tomas ext4_group_t group; 2972c9de560dSAlex Tomas 29738df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2974c9de560dSAlex Tomas return NULL; 2975c9de560dSAlex Tomas group = *pos + 1; 2976a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2977c9de560dSAlex Tomas } 2978c9de560dSAlex Tomas 2979c9de560dSAlex Tomas static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2980c9de560dSAlex Tomas { 2981359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 2982c9de560dSAlex Tomas ext4_group_t group; 2983c9de560dSAlex Tomas 2984c9de560dSAlex Tomas ++*pos; 29858df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2986c9de560dSAlex Tomas return NULL; 2987c9de560dSAlex Tomas group = *pos + 1; 2988a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2989c9de560dSAlex Tomas } 2990c9de560dSAlex Tomas 2991c9de560dSAlex Tomas static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2992c9de560dSAlex Tomas { 2993359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 2994a9df9a49STheodore Ts'o ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2995c9de560dSAlex Tomas int i; 29961c8457caSAditya Kali int err, buddy_loaded = 0; 2997c9de560dSAlex Tomas struct ext4_buddy e4b; 29981c8457caSAditya Kali struct ext4_group_info *grinfo; 29992df2c340SArnd Bergmann unsigned char blocksize_bits = min_t(unsigned char, 30002df2c340SArnd Bergmann sb->s_blocksize_bits, 30012df2c340SArnd Bergmann EXT4_MAX_BLOCK_LOG_SIZE); 3002c9de560dSAlex Tomas struct sg { 3003c9de560dSAlex Tomas struct ext4_group_info info; 3004b80b32b6STheodore Ts'o ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; 3005c9de560dSAlex Tomas } sg; 3006c9de560dSAlex Tomas 3007c9de560dSAlex Tomas group--; 3008c9de560dSAlex Tomas if (group == 0) 300997b4af2fSRasmus Villemoes seq_puts(seq, "#group: free frags first [" 301097b4af2fSRasmus Villemoes " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " 3011802cf1f9SHuaitong Han " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); 3012c9de560dSAlex Tomas 3013b80b32b6STheodore Ts'o i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 3014b80b32b6STheodore Ts'o sizeof(struct ext4_group_info); 3015b80b32b6STheodore Ts'o 30161c8457caSAditya Kali grinfo = ext4_get_group_info(sb, group); 30175354b2afSTheodore Ts'o if (!grinfo) 30185354b2afSTheodore Ts'o return 0; 30191c8457caSAditya Kali /* Load the group info in memory only if not already loaded. */ 30201c8457caSAditya Kali if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { 3021c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3022c9de560dSAlex Tomas if (err) { 3023a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: I/O error\n", group); 3024c9de560dSAlex Tomas return 0; 3025c9de560dSAlex Tomas } 30261c8457caSAditya Kali buddy_loaded = 1; 30271c8457caSAditya Kali } 30281c8457caSAditya Kali 30295354b2afSTheodore Ts'o memcpy(&sg, grinfo, i); 30301c8457caSAditya Kali 30311c8457caSAditya Kali if (buddy_loaded) 3032e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3033c9de560dSAlex Tomas 3034a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 3035c9de560dSAlex Tomas sg.info.bb_fragments, sg.info.bb_first_free); 3036c9de560dSAlex Tomas for (i = 0; i <= 13; i++) 30372df2c340SArnd Bergmann seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? 3038c9de560dSAlex Tomas sg.info.bb_counters[i] : 0); 3039e0d438c7SXu Wang seq_puts(seq, " ]\n"); 3040c9de560dSAlex Tomas 3041c9de560dSAlex Tomas return 0; 3042c9de560dSAlex Tomas } 3043c9de560dSAlex Tomas 3044c9de560dSAlex Tomas static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 3045c9de560dSAlex Tomas { 3046c9de560dSAlex Tomas } 3047c9de560dSAlex Tomas 3048247dbed8SChristoph Hellwig const struct seq_operations ext4_mb_seq_groups_ops = { 3049c9de560dSAlex Tomas .start = ext4_mb_seq_groups_start, 3050c9de560dSAlex Tomas .next = ext4_mb_seq_groups_next, 3051c9de560dSAlex Tomas .stop = ext4_mb_seq_groups_stop, 3052c9de560dSAlex Tomas .show = ext4_mb_seq_groups_show, 3053c9de560dSAlex Tomas }; 3054c9de560dSAlex Tomas 3055a6c75eafSHarshad Shirwadkar int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) 3056a6c75eafSHarshad Shirwadkar { 3057c30365b9SYu Zhe struct super_block *sb = seq->private; 3058a6c75eafSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 3059a6c75eafSHarshad Shirwadkar 3060a6c75eafSHarshad Shirwadkar seq_puts(seq, "mballoc:\n"); 3061a6c75eafSHarshad Shirwadkar if (!sbi->s_mb_stats) { 3062a6c75eafSHarshad Shirwadkar seq_puts(seq, "\tmb stats collection turned off.\n"); 3063f52f3d2bSOjaswin Mujoo seq_puts( 3064f52f3d2bSOjaswin Mujoo seq, 3065f52f3d2bSOjaswin Mujoo "\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); 3066a6c75eafSHarshad Shirwadkar return 0; 3067a6c75eafSHarshad Shirwadkar } 3068a6c75eafSHarshad Shirwadkar seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); 3069a6c75eafSHarshad Shirwadkar seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); 3070a6c75eafSHarshad Shirwadkar 3071f52f3d2bSOjaswin Mujoo seq_printf(seq, "\tgroups_scanned: %u\n", 3072f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_groups_scanned)); 3073a6c75eafSHarshad Shirwadkar 3074f52f3d2bSOjaswin Mujoo /* CR_POWER2_ALIGNED stats */ 3075f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_p2_aligned_stats:\n"); 3076f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3077f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED])); 3078f52f3d2bSOjaswin Mujoo seq_printf( 3079f52f3d2bSOjaswin Mujoo seq, "\t\tgroups_considered: %llu\n", 3080f52f3d2bSOjaswin Mujoo atomic64_read( 3081f52f3d2bSOjaswin Mujoo &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED])); 3082f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3083f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED])); 3084a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tuseless_loops: %llu\n", 3085f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED])); 3086196e402aSHarshad Shirwadkar seq_printf(seq, "\t\tbad_suggestions: %u\n", 3087f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_p2_aligned_bad_suggestions)); 3088a6c75eafSHarshad Shirwadkar 3089f52f3d2bSOjaswin Mujoo /* CR_GOAL_LEN_FAST stats */ 3090f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_goal_fast_stats:\n"); 3091f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3092f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST])); 3093a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tgroups_considered: %llu\n", 3094f52f3d2bSOjaswin Mujoo atomic64_read( 3095f52f3d2bSOjaswin Mujoo &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST])); 3096f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3097f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST])); 3098a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tuseless_loops: %llu\n", 3099f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST])); 3100196e402aSHarshad Shirwadkar seq_printf(seq, "\t\tbad_suggestions: %u\n", 3101f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_goal_fast_bad_suggestions)); 3102a6c75eafSHarshad Shirwadkar 3103f52f3d2bSOjaswin Mujoo /* CR_BEST_AVAIL_LEN stats */ 3104f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_best_avail_stats:\n"); 3105f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3106f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN])); 3107f52f3d2bSOjaswin Mujoo seq_printf( 3108f52f3d2bSOjaswin Mujoo seq, "\t\tgroups_considered: %llu\n", 3109f52f3d2bSOjaswin Mujoo atomic64_read( 3110f52f3d2bSOjaswin Mujoo &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN])); 3111f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3112f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN])); 31137e170922SOjaswin Mujoo seq_printf(seq, "\t\tuseless_loops: %llu\n", 3114f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN])); 31157e170922SOjaswin Mujoo seq_printf(seq, "\t\tbad_suggestions: %u\n", 3116f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_best_avail_bad_suggestions)); 31177e170922SOjaswin Mujoo 3118f52f3d2bSOjaswin Mujoo /* CR_GOAL_LEN_SLOW stats */ 3119f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_goal_slow_stats:\n"); 3120f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3121f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW])); 3122a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tgroups_considered: %llu\n", 3123f52f3d2bSOjaswin Mujoo atomic64_read( 3124f52f3d2bSOjaswin Mujoo &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW])); 3125f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3126f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW])); 3127a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tuseless_loops: %llu\n", 3128f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW])); 3129a6c75eafSHarshad Shirwadkar 3130f52f3d2bSOjaswin Mujoo /* CR_ANY_FREE stats */ 3131f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_any_free_stats:\n"); 3132f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3133f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE])); 3134f52f3d2bSOjaswin Mujoo seq_printf( 3135f52f3d2bSOjaswin Mujoo seq, "\t\tgroups_considered: %llu\n", 3136f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE])); 3137f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3138f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE])); 3139a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tuseless_loops: %llu\n", 3140f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE])); 3141f52f3d2bSOjaswin Mujoo 3142f52f3d2bSOjaswin Mujoo /* Aggregates */ 3143f52f3d2bSOjaswin Mujoo seq_printf(seq, "\textents_scanned: %u\n", 3144f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_ex_scanned)); 3145a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); 3146f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\tlen_goal_hits: %u\n", 3147f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_len_goals)); 3148a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); 3149a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); 3150a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); 3151a6c75eafSHarshad Shirwadkar seq_printf(seq, "\tbuddies_generated: %u/%u\n", 3152a6c75eafSHarshad Shirwadkar atomic_read(&sbi->s_mb_buddies_generated), 3153a6c75eafSHarshad Shirwadkar ext4_get_groups_count(sb)); 3154a6c75eafSHarshad Shirwadkar seq_printf(seq, "\tbuddies_time_used: %llu\n", 3155a6c75eafSHarshad Shirwadkar atomic64_read(&sbi->s_mb_generation_time)); 3156a6c75eafSHarshad Shirwadkar seq_printf(seq, "\tpreallocated: %u\n", 3157a6c75eafSHarshad Shirwadkar atomic_read(&sbi->s_mb_preallocated)); 3158f52f3d2bSOjaswin Mujoo seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded)); 3159a6c75eafSHarshad Shirwadkar return 0; 3160a6c75eafSHarshad Shirwadkar } 3161a6c75eafSHarshad Shirwadkar 3162f68f4063SHarshad Shirwadkar static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) 3163a5fda113STheodore Ts'o __acquires(&EXT4_SB(sb)->s_mb_rb_lock) 3164f68f4063SHarshad Shirwadkar { 3165359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3166f68f4063SHarshad Shirwadkar unsigned long position; 3167f68f4063SHarshad Shirwadkar 316883e80a6eSJan Kara if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) 3169f68f4063SHarshad Shirwadkar return NULL; 3170f68f4063SHarshad Shirwadkar position = *pos + 1; 3171f68f4063SHarshad Shirwadkar return (void *) ((unsigned long) position); 3172f68f4063SHarshad Shirwadkar } 3173f68f4063SHarshad Shirwadkar 3174f68f4063SHarshad Shirwadkar static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) 3175f68f4063SHarshad Shirwadkar { 3176359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3177f68f4063SHarshad Shirwadkar unsigned long position; 3178f68f4063SHarshad Shirwadkar 3179f68f4063SHarshad Shirwadkar ++*pos; 318083e80a6eSJan Kara if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) 3181f68f4063SHarshad Shirwadkar return NULL; 3182f68f4063SHarshad Shirwadkar position = *pos + 1; 3183f68f4063SHarshad Shirwadkar return (void *) ((unsigned long) position); 3184f68f4063SHarshad Shirwadkar } 3185f68f4063SHarshad Shirwadkar 3186f68f4063SHarshad Shirwadkar static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) 3187f68f4063SHarshad Shirwadkar { 3188359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3189f68f4063SHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 3190f68f4063SHarshad Shirwadkar unsigned long position = ((unsigned long) v); 3191f68f4063SHarshad Shirwadkar struct ext4_group_info *grp; 319283e80a6eSJan Kara unsigned int count; 3193f68f4063SHarshad Shirwadkar 3194f68f4063SHarshad Shirwadkar position--; 3195f68f4063SHarshad Shirwadkar if (position >= MB_NUM_ORDERS(sb)) { 319683e80a6eSJan Kara position -= MB_NUM_ORDERS(sb); 319783e80a6eSJan Kara if (position == 0) 319883e80a6eSJan Kara seq_puts(seq, "avg_fragment_size_lists:\n"); 3199f68f4063SHarshad Shirwadkar 320083e80a6eSJan Kara count = 0; 320183e80a6eSJan Kara read_lock(&sbi->s_mb_avg_fragment_size_locks[position]); 320283e80a6eSJan Kara list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position], 320383e80a6eSJan Kara bb_avg_fragment_size_node) 320483e80a6eSJan Kara count++; 320583e80a6eSJan Kara read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]); 320683e80a6eSJan Kara seq_printf(seq, "\tlist_order_%u_groups: %u\n", 320783e80a6eSJan Kara (unsigned int)position, count); 3208f68f4063SHarshad Shirwadkar return 0; 3209f68f4063SHarshad Shirwadkar } 3210f68f4063SHarshad Shirwadkar 3211f68f4063SHarshad Shirwadkar if (position == 0) { 3212f68f4063SHarshad Shirwadkar seq_printf(seq, "optimize_scan: %d\n", 3213f68f4063SHarshad Shirwadkar test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); 3214f68f4063SHarshad Shirwadkar seq_puts(seq, "max_free_order_lists:\n"); 3215f68f4063SHarshad Shirwadkar } 3216f68f4063SHarshad Shirwadkar count = 0; 321783e80a6eSJan Kara read_lock(&sbi->s_mb_largest_free_orders_locks[position]); 3218f68f4063SHarshad Shirwadkar list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position], 3219f68f4063SHarshad Shirwadkar bb_largest_free_order_node) 3220f68f4063SHarshad Shirwadkar count++; 322183e80a6eSJan Kara read_unlock(&sbi->s_mb_largest_free_orders_locks[position]); 3222f68f4063SHarshad Shirwadkar seq_printf(seq, "\tlist_order_%u_groups: %u\n", 3223f68f4063SHarshad Shirwadkar (unsigned int)position, count); 3224f68f4063SHarshad Shirwadkar 3225f68f4063SHarshad Shirwadkar return 0; 3226f68f4063SHarshad Shirwadkar } 3227f68f4063SHarshad Shirwadkar 3228f68f4063SHarshad Shirwadkar static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) 3229f68f4063SHarshad Shirwadkar { 3230f68f4063SHarshad Shirwadkar } 3231f68f4063SHarshad Shirwadkar 3232f68f4063SHarshad Shirwadkar const struct seq_operations ext4_mb_seq_structs_summary_ops = { 3233f68f4063SHarshad Shirwadkar .start = ext4_mb_seq_structs_summary_start, 3234f68f4063SHarshad Shirwadkar .next = ext4_mb_seq_structs_summary_next, 3235f68f4063SHarshad Shirwadkar .stop = ext4_mb_seq_structs_summary_stop, 3236f68f4063SHarshad Shirwadkar .show = ext4_mb_seq_structs_summary_show, 3237f68f4063SHarshad Shirwadkar }; 3238f68f4063SHarshad Shirwadkar 3239fb1813f4SCurt Wohlgemuth static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) 3240fb1813f4SCurt Wohlgemuth { 3241fb1813f4SCurt Wohlgemuth int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 3242fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; 3243fb1813f4SCurt Wohlgemuth 3244fb1813f4SCurt Wohlgemuth BUG_ON(!cachep); 3245fb1813f4SCurt Wohlgemuth return cachep; 3246fb1813f4SCurt Wohlgemuth } 32475f21b0e6SFrederic Bohe 324828623c2fSTheodore Ts'o /* 324928623c2fSTheodore Ts'o * Allocate the top-level s_group_info array for the specified number 325028623c2fSTheodore Ts'o * of groups 325128623c2fSTheodore Ts'o */ 325228623c2fSTheodore Ts'o int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) 325328623c2fSTheodore Ts'o { 325428623c2fSTheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(sb); 325528623c2fSTheodore Ts'o unsigned size; 3256df3da4eaSSuraj Jitindar Singh struct ext4_group_info ***old_groupinfo, ***new_groupinfo; 325728623c2fSTheodore Ts'o 325828623c2fSTheodore Ts'o size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> 325928623c2fSTheodore Ts'o EXT4_DESC_PER_BLOCK_BITS(sb); 326028623c2fSTheodore Ts'o if (size <= sbi->s_group_info_size) 326128623c2fSTheodore Ts'o return 0; 326228623c2fSTheodore Ts'o 326328623c2fSTheodore Ts'o size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); 3264a7c3e901SMichal Hocko new_groupinfo = kvzalloc(size, GFP_KERNEL); 326528623c2fSTheodore Ts'o if (!new_groupinfo) { 326628623c2fSTheodore Ts'o ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); 326728623c2fSTheodore Ts'o return -ENOMEM; 326828623c2fSTheodore Ts'o } 3269df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3270df3da4eaSSuraj Jitindar Singh old_groupinfo = rcu_dereference(sbi->s_group_info); 3271df3da4eaSSuraj Jitindar Singh if (old_groupinfo) 3272df3da4eaSSuraj Jitindar Singh memcpy(new_groupinfo, old_groupinfo, 327328623c2fSTheodore Ts'o sbi->s_group_info_size * sizeof(*sbi->s_group_info)); 3274df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3275df3da4eaSSuraj Jitindar Singh rcu_assign_pointer(sbi->s_group_info, new_groupinfo); 327628623c2fSTheodore Ts'o sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); 3277df3da4eaSSuraj Jitindar Singh if (old_groupinfo) 3278df3da4eaSSuraj Jitindar Singh ext4_kvfree_array_rcu(old_groupinfo); 327928623c2fSTheodore Ts'o ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 328028623c2fSTheodore Ts'o sbi->s_group_info_size); 328128623c2fSTheodore Ts'o return 0; 328228623c2fSTheodore Ts'o } 328328623c2fSTheodore Ts'o 32845f21b0e6SFrederic Bohe /* Create and initialize ext4_group_info data for the given group. */ 3285920313a7SAneesh Kumar K.V int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 32865f21b0e6SFrederic Bohe struct ext4_group_desc *desc) 32875f21b0e6SFrederic Bohe { 3288fb1813f4SCurt Wohlgemuth int i; 32895f21b0e6SFrederic Bohe int metalen = 0; 3290df3da4eaSSuraj Jitindar Singh int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); 32915f21b0e6SFrederic Bohe struct ext4_sb_info *sbi = EXT4_SB(sb); 32925f21b0e6SFrederic Bohe struct ext4_group_info **meta_group_info; 3293fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 32945f21b0e6SFrederic Bohe 32955f21b0e6SFrederic Bohe /* 32965f21b0e6SFrederic Bohe * First check if this group is the first of a reserved block. 32975f21b0e6SFrederic Bohe * If it's true, we have to allocate a new table of pointers 32985f21b0e6SFrederic Bohe * to ext4_group_info structures 32995f21b0e6SFrederic Bohe */ 33005f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 33015f21b0e6SFrederic Bohe metalen = sizeof(*meta_group_info) << 33025f21b0e6SFrederic Bohe EXT4_DESC_PER_BLOCK_BITS(sb); 33034fdb5543SDmitry Monakhov meta_group_info = kmalloc(metalen, GFP_NOFS); 33045f21b0e6SFrederic Bohe if (meta_group_info == NULL) { 33057f6a11e7SJoe Perches ext4_msg(sb, KERN_ERR, "can't allocate mem " 33069d8b9ec4STheodore Ts'o "for a buddy group"); 3307df119095SKemeng Shi return -ENOMEM; 33085f21b0e6SFrederic Bohe } 3309df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3310df3da4eaSSuraj Jitindar Singh rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; 3311df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 33125f21b0e6SFrederic Bohe } 33135f21b0e6SFrederic Bohe 3314df3da4eaSSuraj Jitindar Singh meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); 33155f21b0e6SFrederic Bohe i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 33165f21b0e6SFrederic Bohe 33174fdb5543SDmitry Monakhov meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); 33185f21b0e6SFrederic Bohe if (meta_group_info[i] == NULL) { 33197f6a11e7SJoe Perches ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); 33205f21b0e6SFrederic Bohe goto exit_group_info; 33215f21b0e6SFrederic Bohe } 33225f21b0e6SFrederic Bohe set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 33235f21b0e6SFrederic Bohe &(meta_group_info[i]->bb_state)); 33245f21b0e6SFrederic Bohe 33255f21b0e6SFrederic Bohe /* 33265f21b0e6SFrederic Bohe * initialize bb_free to be able to skip 33275f21b0e6SFrederic Bohe * empty groups without initialization 33285f21b0e6SFrederic Bohe */ 33298844618dSTheodore Ts'o if (ext4_has_group_desc_csum(sb) && 33308844618dSTheodore Ts'o (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 33315f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 3332cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, group, desc); 33335f21b0e6SFrederic Bohe } else { 33345f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 3335021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc); 33365f21b0e6SFrederic Bohe } 33375f21b0e6SFrederic Bohe 33385f21b0e6SFrederic Bohe INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 3339920313a7SAneesh Kumar K.V init_rwsem(&meta_group_info[i]->alloc_sem); 334064e290ecSVenkatesh Pallipadi meta_group_info[i]->bb_free_root = RB_ROOT; 3341196e402aSHarshad Shirwadkar INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); 334283e80a6eSJan Kara INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node); 33438a57d9d6SCurt Wohlgemuth meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ 334483e80a6eSJan Kara meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */ 3345196e402aSHarshad Shirwadkar meta_group_info[i]->bb_group = group; 33465f21b0e6SFrederic Bohe 3347a3450215SRitesh Harjani mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); 33485f21b0e6SFrederic Bohe return 0; 33495f21b0e6SFrederic Bohe 33505f21b0e6SFrederic Bohe exit_group_info: 33515f21b0e6SFrederic Bohe /* If a meta_group_info table has been allocated, release it now */ 3352caaf7a29STao Ma if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 3353df3da4eaSSuraj Jitindar Singh struct ext4_group_info ***group_info; 3354df3da4eaSSuraj Jitindar Singh 3355df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3356df3da4eaSSuraj Jitindar Singh group_info = rcu_dereference(sbi->s_group_info); 3357df3da4eaSSuraj Jitindar Singh kfree(group_info[idx]); 3358df3da4eaSSuraj Jitindar Singh group_info[idx] = NULL; 3359df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3360caaf7a29STao Ma } 33615f21b0e6SFrederic Bohe return -ENOMEM; 33625f21b0e6SFrederic Bohe } /* ext4_mb_add_groupinfo */ 33635f21b0e6SFrederic Bohe 3364c9de560dSAlex Tomas static int ext4_mb_init_backend(struct super_block *sb) 3365c9de560dSAlex Tomas { 33668df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 3367c9de560dSAlex Tomas ext4_group_t i; 3368c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 336928623c2fSTheodore Ts'o int err; 33705f21b0e6SFrederic Bohe struct ext4_group_desc *desc; 3371df3da4eaSSuraj Jitindar Singh struct ext4_group_info ***group_info; 3372fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep; 3373c9de560dSAlex Tomas 337428623c2fSTheodore Ts'o err = ext4_mb_alloc_groupinfo(sb, ngroups); 337528623c2fSTheodore Ts'o if (err) 337628623c2fSTheodore Ts'o return err; 33775f21b0e6SFrederic Bohe 3378c9de560dSAlex Tomas sbi->s_buddy_cache = new_inode(sb); 3379c9de560dSAlex Tomas if (sbi->s_buddy_cache == NULL) { 33809d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't get new inode"); 3381c9de560dSAlex Tomas goto err_freesgi; 3382c9de560dSAlex Tomas } 338348e6061bSYu Jian /* To avoid potentially colliding with an valid on-disk inode number, 338448e6061bSYu Jian * use EXT4_BAD_INO for the buddy cache inode number. This inode is 338548e6061bSYu Jian * not in the inode hash, so it should never be found by iget(), but 338648e6061bSYu Jian * this will avoid confusion if it ever shows up during debugging. */ 338748e6061bSYu Jian sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; 3388c9de560dSAlex Tomas EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 33898df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 33904b99faa2SKhazhismel Kumykov cond_resched(); 3391c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, i, NULL); 3392c9de560dSAlex Tomas if (desc == NULL) { 33939d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); 3394c9de560dSAlex Tomas goto err_freebuddy; 3395c9de560dSAlex Tomas } 33965f21b0e6SFrederic Bohe if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 33975f21b0e6SFrederic Bohe goto err_freebuddy; 3398c9de560dSAlex Tomas } 3399c9de560dSAlex Tomas 3400cfd73237SAlex Zhuravlev if (ext4_has_feature_flex_bg(sb)) { 3401f91436d5SSabyrzhan Tasbolatov /* a single flex group is supposed to be read by a single IO. 3402f91436d5SSabyrzhan Tasbolatov * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is 3403f91436d5SSabyrzhan Tasbolatov * unsigned integer, so the maximum shift is 32. 3404f91436d5SSabyrzhan Tasbolatov */ 3405f91436d5SSabyrzhan Tasbolatov if (sbi->s_es->s_log_groups_per_flex >= 32) { 3406f91436d5SSabyrzhan Tasbolatov ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); 3407a8867f4eSPhillip Potter goto err_freebuddy; 3408f91436d5SSabyrzhan Tasbolatov } 3409f91436d5SSabyrzhan Tasbolatov sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, 341082ef1370SChunguang Xu BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); 3411cfd73237SAlex Zhuravlev sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ 3412cfd73237SAlex Zhuravlev } else { 3413cfd73237SAlex Zhuravlev sbi->s_mb_prefetch = 32; 3414cfd73237SAlex Zhuravlev } 3415cfd73237SAlex Zhuravlev if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) 3416cfd73237SAlex Zhuravlev sbi->s_mb_prefetch = ext4_get_groups_count(sb); 3417cfd73237SAlex Zhuravlev /* now many real IOs to prefetch within a single allocation at cr=0 3418cfd73237SAlex Zhuravlev * given cr=0 is an CPU-related optimization we shouldn't try to 3419cfd73237SAlex Zhuravlev * load too many groups, at some point we should start to use what 3420cfd73237SAlex Zhuravlev * we've got in memory. 3421cfd73237SAlex Zhuravlev * with an average random access time 5ms, it'd take a second to get 3422cfd73237SAlex Zhuravlev * 200 groups (* N with flex_bg), so let's make this limit 4 3423cfd73237SAlex Zhuravlev */ 3424cfd73237SAlex Zhuravlev sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; 3425cfd73237SAlex Zhuravlev if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) 3426cfd73237SAlex Zhuravlev sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); 3427cfd73237SAlex Zhuravlev 3428c9de560dSAlex Tomas return 0; 3429c9de560dSAlex Tomas 3430c9de560dSAlex Tomas err_freebuddy: 3431fb1813f4SCurt Wohlgemuth cachep = get_groupinfo_cache(sb->s_blocksize_bits); 34325354b2afSTheodore Ts'o while (i-- > 0) { 34335354b2afSTheodore Ts'o struct ext4_group_info *grp = ext4_get_group_info(sb, i); 34345354b2afSTheodore Ts'o 34355354b2afSTheodore Ts'o if (grp) 34365354b2afSTheodore Ts'o kmem_cache_free(cachep, grp); 34375354b2afSTheodore Ts'o } 343828623c2fSTheodore Ts'o i = sbi->s_group_info_size; 3439df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3440df3da4eaSSuraj Jitindar Singh group_info = rcu_dereference(sbi->s_group_info); 3441f1fa3342SRoel Kluin while (i-- > 0) 3442df3da4eaSSuraj Jitindar Singh kfree(group_info[i]); 3443df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3444c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 3445c9de560dSAlex Tomas err_freesgi: 3446df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3447df3da4eaSSuraj Jitindar Singh kvfree(rcu_dereference(sbi->s_group_info)); 3448df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3449c9de560dSAlex Tomas return -ENOMEM; 3450c9de560dSAlex Tomas } 3451c9de560dSAlex Tomas 34522892c15dSEric Sandeen static void ext4_groupinfo_destroy_slabs(void) 34532892c15dSEric Sandeen { 34542892c15dSEric Sandeen int i; 34552892c15dSEric Sandeen 34562892c15dSEric Sandeen for (i = 0; i < NR_GRPINFO_CACHES; i++) { 34572892c15dSEric Sandeen kmem_cache_destroy(ext4_groupinfo_caches[i]); 34582892c15dSEric Sandeen ext4_groupinfo_caches[i] = NULL; 34592892c15dSEric Sandeen } 34602892c15dSEric Sandeen } 34612892c15dSEric Sandeen 34622892c15dSEric Sandeen static int ext4_groupinfo_create_slab(size_t size) 34632892c15dSEric Sandeen { 34642892c15dSEric Sandeen static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); 34652892c15dSEric Sandeen int slab_size; 34662892c15dSEric Sandeen int blocksize_bits = order_base_2(size); 34672892c15dSEric Sandeen int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 34682892c15dSEric Sandeen struct kmem_cache *cachep; 34692892c15dSEric Sandeen 34702892c15dSEric Sandeen if (cache_index >= NR_GRPINFO_CACHES) 34712892c15dSEric Sandeen return -EINVAL; 34722892c15dSEric Sandeen 34732892c15dSEric Sandeen if (unlikely(cache_index < 0)) 34742892c15dSEric Sandeen cache_index = 0; 34752892c15dSEric Sandeen 34762892c15dSEric Sandeen mutex_lock(&ext4_grpinfo_slab_create_mutex); 34772892c15dSEric Sandeen if (ext4_groupinfo_caches[cache_index]) { 34782892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 34792892c15dSEric Sandeen return 0; /* Already created */ 34802892c15dSEric Sandeen } 34812892c15dSEric Sandeen 34822892c15dSEric Sandeen slab_size = offsetof(struct ext4_group_info, 34832892c15dSEric Sandeen bb_counters[blocksize_bits + 2]); 34842892c15dSEric Sandeen 34852892c15dSEric Sandeen cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], 34862892c15dSEric Sandeen slab_size, 0, SLAB_RECLAIM_ACCOUNT, 34872892c15dSEric Sandeen NULL); 34882892c15dSEric Sandeen 3489823ba01fSTao Ma ext4_groupinfo_caches[cache_index] = cachep; 3490823ba01fSTao Ma 34912892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 34922892c15dSEric Sandeen if (!cachep) { 34939d8b9ec4STheodore Ts'o printk(KERN_EMERG 34949d8b9ec4STheodore Ts'o "EXT4-fs: no memory for groupinfo slab cache\n"); 34952892c15dSEric Sandeen return -ENOMEM; 34962892c15dSEric Sandeen } 34972892c15dSEric Sandeen 34982892c15dSEric Sandeen return 0; 34992892c15dSEric Sandeen } 35002892c15dSEric Sandeen 350155cdd0afSWang Jianchao static void ext4_discard_work(struct work_struct *work) 350255cdd0afSWang Jianchao { 350355cdd0afSWang Jianchao struct ext4_sb_info *sbi = container_of(work, 350455cdd0afSWang Jianchao struct ext4_sb_info, s_discard_work); 350555cdd0afSWang Jianchao struct super_block *sb = sbi->s_sb; 350655cdd0afSWang Jianchao struct ext4_free_data *fd, *nfd; 350755cdd0afSWang Jianchao struct ext4_buddy e4b; 35080f6bc579SRuan Jinjie LIST_HEAD(discard_list); 350955cdd0afSWang Jianchao ext4_group_t grp, load_grp; 351055cdd0afSWang Jianchao int err = 0; 351155cdd0afSWang Jianchao 351255cdd0afSWang Jianchao spin_lock(&sbi->s_md_lock); 351355cdd0afSWang Jianchao list_splice_init(&sbi->s_discard_list, &discard_list); 351455cdd0afSWang Jianchao spin_unlock(&sbi->s_md_lock); 351555cdd0afSWang Jianchao 351655cdd0afSWang Jianchao load_grp = UINT_MAX; 351755cdd0afSWang Jianchao list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { 351855cdd0afSWang Jianchao /* 35195036ab8dSWang Jianchao * If filesystem is umounting or no memory or suffering 35205036ab8dSWang Jianchao * from no space, give up the discard 352155cdd0afSWang Jianchao */ 35225036ab8dSWang Jianchao if ((sb->s_flags & SB_ACTIVE) && !err && 35235036ab8dSWang Jianchao !atomic_read(&sbi->s_retry_alloc_pending)) { 352455cdd0afSWang Jianchao grp = fd->efd_group; 352555cdd0afSWang Jianchao if (grp != load_grp) { 352655cdd0afSWang Jianchao if (load_grp != UINT_MAX) 352755cdd0afSWang Jianchao ext4_mb_unload_buddy(&e4b); 352855cdd0afSWang Jianchao 352955cdd0afSWang Jianchao err = ext4_mb_load_buddy(sb, grp, &e4b); 353055cdd0afSWang Jianchao if (err) { 353155cdd0afSWang Jianchao kmem_cache_free(ext4_free_data_cachep, fd); 353255cdd0afSWang Jianchao load_grp = UINT_MAX; 353355cdd0afSWang Jianchao continue; 353455cdd0afSWang Jianchao } else { 353555cdd0afSWang Jianchao load_grp = grp; 353655cdd0afSWang Jianchao } 353755cdd0afSWang Jianchao } 353855cdd0afSWang Jianchao 353955cdd0afSWang Jianchao ext4_lock_group(sb, grp); 354055cdd0afSWang Jianchao ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, 354155cdd0afSWang Jianchao fd->efd_start_cluster + fd->efd_count - 1, 1); 354255cdd0afSWang Jianchao ext4_unlock_group(sb, grp); 354355cdd0afSWang Jianchao } 354455cdd0afSWang Jianchao kmem_cache_free(ext4_free_data_cachep, fd); 354555cdd0afSWang Jianchao } 354655cdd0afSWang Jianchao 354755cdd0afSWang Jianchao if (load_grp != UINT_MAX) 354855cdd0afSWang Jianchao ext4_mb_unload_buddy(&e4b); 354955cdd0afSWang Jianchao } 355055cdd0afSWang Jianchao 35519d99012fSAkira Fujita int ext4_mb_init(struct super_block *sb) 3552c9de560dSAlex Tomas { 3553c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 35546be2ded1SAneesh Kumar K.V unsigned i, j; 3555935244cdSNicolai Stange unsigned offset, offset_incr; 3556c9de560dSAlex Tomas unsigned max; 355774767c5aSShen Feng int ret; 3558c9de560dSAlex Tomas 35594b68f6dfSHarshad Shirwadkar i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); 3560c9de560dSAlex Tomas 3561c9de560dSAlex Tomas sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 3562c9de560dSAlex Tomas if (sbi->s_mb_offsets == NULL) { 3563fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 3564fb1813f4SCurt Wohlgemuth goto out; 3565c9de560dSAlex Tomas } 3566ff7ef329SYasunori Goto 35674b68f6dfSHarshad Shirwadkar i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); 3568c9de560dSAlex Tomas sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 3569c9de560dSAlex Tomas if (sbi->s_mb_maxs == NULL) { 3570fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 3571fb1813f4SCurt Wohlgemuth goto out; 3572fb1813f4SCurt Wohlgemuth } 3573fb1813f4SCurt Wohlgemuth 35742892c15dSEric Sandeen ret = ext4_groupinfo_create_slab(sb->s_blocksize); 35752892c15dSEric Sandeen if (ret < 0) 3576fb1813f4SCurt Wohlgemuth goto out; 3577c9de560dSAlex Tomas 3578c9de560dSAlex Tomas /* order 0 is regular bitmap */ 3579c9de560dSAlex Tomas sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 3580c9de560dSAlex Tomas sbi->s_mb_offsets[0] = 0; 3581c9de560dSAlex Tomas 3582c9de560dSAlex Tomas i = 1; 3583c9de560dSAlex Tomas offset = 0; 3584935244cdSNicolai Stange offset_incr = 1 << (sb->s_blocksize_bits - 1); 3585c9de560dSAlex Tomas max = sb->s_blocksize << 2; 3586c9de560dSAlex Tomas do { 3587c9de560dSAlex Tomas sbi->s_mb_offsets[i] = offset; 3588c9de560dSAlex Tomas sbi->s_mb_maxs[i] = max; 3589935244cdSNicolai Stange offset += offset_incr; 3590935244cdSNicolai Stange offset_incr = offset_incr >> 1; 3591c9de560dSAlex Tomas max = max >> 1; 3592c9de560dSAlex Tomas i++; 35934b68f6dfSHarshad Shirwadkar } while (i < MB_NUM_ORDERS(sb)); 35944b68f6dfSHarshad Shirwadkar 359583e80a6eSJan Kara sbi->s_mb_avg_fragment_size = 359683e80a6eSJan Kara kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), 359783e80a6eSJan Kara GFP_KERNEL); 359883e80a6eSJan Kara if (!sbi->s_mb_avg_fragment_size) { 359983e80a6eSJan Kara ret = -ENOMEM; 360083e80a6eSJan Kara goto out; 360183e80a6eSJan Kara } 360283e80a6eSJan Kara sbi->s_mb_avg_fragment_size_locks = 360383e80a6eSJan Kara kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), 360483e80a6eSJan Kara GFP_KERNEL); 360583e80a6eSJan Kara if (!sbi->s_mb_avg_fragment_size_locks) { 360683e80a6eSJan Kara ret = -ENOMEM; 360783e80a6eSJan Kara goto out; 360883e80a6eSJan Kara } 360983e80a6eSJan Kara for (i = 0; i < MB_NUM_ORDERS(sb); i++) { 361083e80a6eSJan Kara INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]); 361183e80a6eSJan Kara rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]); 361283e80a6eSJan Kara } 3613196e402aSHarshad Shirwadkar sbi->s_mb_largest_free_orders = 3614196e402aSHarshad Shirwadkar kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), 3615196e402aSHarshad Shirwadkar GFP_KERNEL); 3616196e402aSHarshad Shirwadkar if (!sbi->s_mb_largest_free_orders) { 3617196e402aSHarshad Shirwadkar ret = -ENOMEM; 3618196e402aSHarshad Shirwadkar goto out; 3619196e402aSHarshad Shirwadkar } 3620196e402aSHarshad Shirwadkar sbi->s_mb_largest_free_orders_locks = 3621196e402aSHarshad Shirwadkar kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), 3622196e402aSHarshad Shirwadkar GFP_KERNEL); 3623196e402aSHarshad Shirwadkar if (!sbi->s_mb_largest_free_orders_locks) { 3624196e402aSHarshad Shirwadkar ret = -ENOMEM; 3625196e402aSHarshad Shirwadkar goto out; 3626196e402aSHarshad Shirwadkar } 3627196e402aSHarshad Shirwadkar for (i = 0; i < MB_NUM_ORDERS(sb); i++) { 3628196e402aSHarshad Shirwadkar INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); 3629196e402aSHarshad Shirwadkar rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); 3630196e402aSHarshad Shirwadkar } 3631c9de560dSAlex Tomas 3632c9de560dSAlex Tomas spin_lock_init(&sbi->s_md_lock); 3633d08854f5STheodore Ts'o sbi->s_mb_free_pending = 0; 3634*ce774e53SJinke Han INIT_LIST_HEAD(&sbi->s_freed_data_list[0]); 3635*ce774e53SJinke Han INIT_LIST_HEAD(&sbi->s_freed_data_list[1]); 363655cdd0afSWang Jianchao INIT_LIST_HEAD(&sbi->s_discard_list); 363755cdd0afSWang Jianchao INIT_WORK(&sbi->s_discard_work, ext4_discard_work); 36385036ab8dSWang Jianchao atomic_set(&sbi->s_retry_alloc_pending, 0); 3639c9de560dSAlex Tomas 3640c9de560dSAlex Tomas sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 3641c9de560dSAlex Tomas sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 3642c9de560dSAlex Tomas sbi->s_mb_stats = MB_DEFAULT_STATS; 3643c9de560dSAlex Tomas sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 3644c9de560dSAlex Tomas sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 3645f52f3d2bSOjaswin Mujoo sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER; 36467e170922SOjaswin Mujoo 364727baebb8STheodore Ts'o /* 364827baebb8STheodore Ts'o * The default group preallocation is 512, which for 4k block 364927baebb8STheodore Ts'o * sizes translates to 2 megabytes. However for bigalloc file 365027baebb8STheodore Ts'o * systems, this is probably too big (i.e, if the cluster size 365127baebb8STheodore Ts'o * is 1 megabyte, then group preallocation size becomes half a 365227baebb8STheodore Ts'o * gigabyte!). As a default, we will keep a two megabyte 365327baebb8STheodore Ts'o * group pralloc size for cluster sizes up to 64k, and after 365427baebb8STheodore Ts'o * that, we will force a minimum group preallocation size of 365527baebb8STheodore Ts'o * 32 clusters. This translates to 8 megs when the cluster 365627baebb8STheodore Ts'o * size is 256k, and 32 megs when the cluster size is 1 meg, 365727baebb8STheodore Ts'o * which seems reasonable as a default. 365827baebb8STheodore Ts'o */ 365927baebb8STheodore Ts'o sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> 366027baebb8STheodore Ts'o sbi->s_cluster_bits, 32); 3661d7a1fee1SDan Ehrenberg /* 3662d7a1fee1SDan Ehrenberg * If there is a s_stripe > 1, then we set the s_mb_group_prealloc 3663d7a1fee1SDan Ehrenberg * to the lowest multiple of s_stripe which is bigger than 3664d7a1fee1SDan Ehrenberg * the s_mb_group_prealloc as determined above. We want 3665d7a1fee1SDan Ehrenberg * the preallocation size to be an exact multiple of the 3666d7a1fee1SDan Ehrenberg * RAID stripe size so that preallocations don't fragment 3667d7a1fee1SDan Ehrenberg * the stripes. 3668d7a1fee1SDan Ehrenberg */ 3669d7a1fee1SDan Ehrenberg if (sbi->s_stripe > 1) { 3670d7a1fee1SDan Ehrenberg sbi->s_mb_group_prealloc = roundup( 3671c3defd99SKemeng Shi sbi->s_mb_group_prealloc, EXT4_B2C(sbi, sbi->s_stripe)); 3672d7a1fee1SDan Ehrenberg } 3673c9de560dSAlex Tomas 3674730c213cSEric Sandeen sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 3675c9de560dSAlex Tomas if (sbi->s_locality_groups == NULL) { 3676fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 3677029b10c5SAndrey Tsyvarev goto out; 3678c9de560dSAlex Tomas } 3679730c213cSEric Sandeen for_each_possible_cpu(i) { 3680c9de560dSAlex Tomas struct ext4_locality_group *lg; 3681730c213cSEric Sandeen lg = per_cpu_ptr(sbi->s_locality_groups, i); 3682c9de560dSAlex Tomas mutex_init(&lg->lg_mutex); 36836be2ded1SAneesh Kumar K.V for (j = 0; j < PREALLOC_TB_SIZE; j++) 36846be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 3685c9de560dSAlex Tomas spin_lock_init(&lg->lg_prealloc_lock); 3686c9de560dSAlex Tomas } 3687c9de560dSAlex Tomas 368810f0d2a5SChristoph Hellwig if (bdev_nonrot(sb->s_bdev)) 3689196e402aSHarshad Shirwadkar sbi->s_mb_max_linear_groups = 0; 3690196e402aSHarshad Shirwadkar else 3691196e402aSHarshad Shirwadkar sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; 369279a77c5aSYu Jian /* init file for buddy data */ 369379a77c5aSYu Jian ret = ext4_mb_init_backend(sb); 36947aa0baeaSTao Ma if (ret != 0) 36957aa0baeaSTao Ma goto out_free_locality_groups; 369679a77c5aSYu Jian 36977aa0baeaSTao Ma return 0; 36987aa0baeaSTao Ma 36997aa0baeaSTao Ma out_free_locality_groups: 37007aa0baeaSTao Ma free_percpu(sbi->s_locality_groups); 37017aa0baeaSTao Ma sbi->s_locality_groups = NULL; 3702fb1813f4SCurt Wohlgemuth out: 370383e80a6eSJan Kara kfree(sbi->s_mb_avg_fragment_size); 370483e80a6eSJan Kara kfree(sbi->s_mb_avg_fragment_size_locks); 3705196e402aSHarshad Shirwadkar kfree(sbi->s_mb_largest_free_orders); 3706196e402aSHarshad Shirwadkar kfree(sbi->s_mb_largest_free_orders_locks); 3707fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_offsets); 37087aa0baeaSTao Ma sbi->s_mb_offsets = NULL; 3709fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_maxs); 37107aa0baeaSTao Ma sbi->s_mb_maxs = NULL; 3711fb1813f4SCurt Wohlgemuth return ret; 3712c9de560dSAlex Tomas } 3713c9de560dSAlex Tomas 3714955ce5f5SAneesh Kumar K.V /* need to called with the ext4 group lock held */ 3715d3df1453SRitesh Harjani static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) 3716c9de560dSAlex Tomas { 3717c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3718c9de560dSAlex Tomas struct list_head *cur, *tmp; 3719c9de560dSAlex Tomas int count = 0; 3720c9de560dSAlex Tomas 3721c9de560dSAlex Tomas list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 3722c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3723c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3724c9de560dSAlex Tomas count++; 3725688f05a0SAneesh Kumar K.V kmem_cache_free(ext4_pspace_cachep, pa); 3726c9de560dSAlex Tomas } 3727d3df1453SRitesh Harjani return count; 3728c9de560dSAlex Tomas } 3729c9de560dSAlex Tomas 3730c9de560dSAlex Tomas int ext4_mb_release(struct super_block *sb) 3731c9de560dSAlex Tomas { 37328df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 3733c9de560dSAlex Tomas ext4_group_t i; 3734c9de560dSAlex Tomas int num_meta_group_infos; 3735df3da4eaSSuraj Jitindar Singh struct ext4_group_info *grinfo, ***group_info; 3736c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3737fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 3738d3df1453SRitesh Harjani int count; 3739c9de560dSAlex Tomas 374055cdd0afSWang Jianchao if (test_opt(sb, DISCARD)) { 374155cdd0afSWang Jianchao /* 374255cdd0afSWang Jianchao * wait the discard work to drain all of ext4_free_data 374355cdd0afSWang Jianchao */ 374455cdd0afSWang Jianchao flush_work(&sbi->s_discard_work); 374555cdd0afSWang Jianchao WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); 374655cdd0afSWang Jianchao } 374755cdd0afSWang Jianchao 3748c9de560dSAlex Tomas if (sbi->s_group_info) { 37498df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 37504b99faa2SKhazhismel Kumykov cond_resched(); 3751c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, i); 37525354b2afSTheodore Ts'o if (!grinfo) 37535354b2afSTheodore Ts'o continue; 3754a3450215SRitesh Harjani mb_group_bb_bitmap_free(grinfo); 3755c9de560dSAlex Tomas ext4_lock_group(sb, i); 3756d3df1453SRitesh Harjani count = ext4_mb_cleanup_pa(grinfo); 3757d3df1453SRitesh Harjani if (count) 3758d3df1453SRitesh Harjani mb_debug(sb, "mballoc: %d PAs left\n", 3759d3df1453SRitesh Harjani count); 3760c9de560dSAlex Tomas ext4_unlock_group(sb, i); 3761fb1813f4SCurt Wohlgemuth kmem_cache_free(cachep, grinfo); 3762c9de560dSAlex Tomas } 37638df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + 3764c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK(sb) - 1) >> 3765c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK_BITS(sb); 3766df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3767df3da4eaSSuraj Jitindar Singh group_info = rcu_dereference(sbi->s_group_info); 3768c9de560dSAlex Tomas for (i = 0; i < num_meta_group_infos; i++) 3769df3da4eaSSuraj Jitindar Singh kfree(group_info[i]); 3770df3da4eaSSuraj Jitindar Singh kvfree(group_info); 3771df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3772c9de560dSAlex Tomas } 377383e80a6eSJan Kara kfree(sbi->s_mb_avg_fragment_size); 377483e80a6eSJan Kara kfree(sbi->s_mb_avg_fragment_size_locks); 3775196e402aSHarshad Shirwadkar kfree(sbi->s_mb_largest_free_orders); 3776196e402aSHarshad Shirwadkar kfree(sbi->s_mb_largest_free_orders_locks); 3777c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 3778c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 3779c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 3780c9de560dSAlex Tomas if (sbi->s_mb_stats) { 37819d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 37829d8b9ec4STheodore Ts'o "mballoc: %u blocks %u reqs (%u success)", 3783c9de560dSAlex Tomas atomic_read(&sbi->s_bal_allocated), 3784c9de560dSAlex Tomas atomic_read(&sbi->s_bal_reqs), 3785c9de560dSAlex Tomas atomic_read(&sbi->s_bal_success)); 37869d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 3787a6c75eafSHarshad Shirwadkar "mballoc: %u extents scanned, %u groups scanned, %u goal hits, " 37889d8b9ec4STheodore Ts'o "%u 2^N hits, %u breaks, %u lost", 3789c9de560dSAlex Tomas atomic_read(&sbi->s_bal_ex_scanned), 3790a6c75eafSHarshad Shirwadkar atomic_read(&sbi->s_bal_groups_scanned), 3791c9de560dSAlex Tomas atomic_read(&sbi->s_bal_goals), 3792c9de560dSAlex Tomas atomic_read(&sbi->s_bal_2orders), 3793c9de560dSAlex Tomas atomic_read(&sbi->s_bal_breaks), 3794c9de560dSAlex Tomas atomic_read(&sbi->s_mb_lost_chunks)); 37959d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 379667d25186SHarshad Shirwadkar "mballoc: %u generated and it took %llu", 379767d25186SHarshad Shirwadkar atomic_read(&sbi->s_mb_buddies_generated), 379867d25186SHarshad Shirwadkar atomic64_read(&sbi->s_mb_generation_time)); 37999d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 38009d8b9ec4STheodore Ts'o "mballoc: %u preallocated, %u discarded", 3801c9de560dSAlex Tomas atomic_read(&sbi->s_mb_preallocated), 3802c9de560dSAlex Tomas atomic_read(&sbi->s_mb_discarded)); 3803c9de560dSAlex Tomas } 3804c9de560dSAlex Tomas 3805730c213cSEric Sandeen free_percpu(sbi->s_locality_groups); 3806c9de560dSAlex Tomas 3807c9de560dSAlex Tomas return 0; 3808c9de560dSAlex Tomas } 3809c9de560dSAlex Tomas 381077ca6cdfSLukas Czerner static inline int ext4_issue_discard(struct super_block *sb, 3811a0154344SDaeho Jeong ext4_group_t block_group, ext4_grpblk_t cluster, int count, 3812a0154344SDaeho Jeong struct bio **biop) 38135c521830SJiaying Zhang { 38145c521830SJiaying Zhang ext4_fsblk_t discard_block; 38155c521830SJiaying Zhang 381684130193STheodore Ts'o discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + 381784130193STheodore Ts'o ext4_group_first_block_no(sb, block_group)); 381884130193STheodore Ts'o count = EXT4_C2B(EXT4_SB(sb), count); 38195c521830SJiaying Zhang trace_ext4_discard_blocks(sb, 38205c521830SJiaying Zhang (unsigned long long) discard_block, count); 3821a0154344SDaeho Jeong if (biop) { 3822a0154344SDaeho Jeong return __blkdev_issue_discard(sb->s_bdev, 3823a0154344SDaeho Jeong (sector_t)discard_block << (sb->s_blocksize_bits - 9), 3824a0154344SDaeho Jeong (sector_t)count << (sb->s_blocksize_bits - 9), 382544abff2cSChristoph Hellwig GFP_NOFS, biop); 3826a0154344SDaeho Jeong } else 382793259636SLukas Czerner return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); 38285c521830SJiaying Zhang } 38295c521830SJiaying Zhang 3830a0154344SDaeho Jeong static void ext4_free_data_in_buddy(struct super_block *sb, 3831a0154344SDaeho Jeong struct ext4_free_data *entry) 3832c9de560dSAlex Tomas { 3833c9de560dSAlex Tomas struct ext4_buddy e4b; 3834c894058dSAneesh Kumar K.V struct ext4_group_info *db; 3835c7f2bafaSKemeng Shi int err, count = 0; 3836c9de560dSAlex Tomas 3837d3df1453SRitesh Harjani mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", 383818aadd47SBobi Jam entry->efd_count, entry->efd_group, entry); 3839c9de560dSAlex Tomas 384018aadd47SBobi Jam err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); 3841c9de560dSAlex Tomas /* we expect to find existing buddy because it's pinned */ 3842c9de560dSAlex Tomas BUG_ON(err != 0); 3843c9de560dSAlex Tomas 3844d08854f5STheodore Ts'o spin_lock(&EXT4_SB(sb)->s_md_lock); 3845d08854f5STheodore Ts'o EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; 3846d08854f5STheodore Ts'o spin_unlock(&EXT4_SB(sb)->s_md_lock); 384718aadd47SBobi Jam 3848c894058dSAneesh Kumar K.V db = e4b.bd_info; 3849c9de560dSAlex Tomas /* there are blocks to put in buddy to make them really free */ 385018aadd47SBobi Jam count += entry->efd_count; 385118aadd47SBobi Jam ext4_lock_group(sb, entry->efd_group); 3852c894058dSAneesh Kumar K.V /* Take it out of per group rb tree */ 385318aadd47SBobi Jam rb_erase(&entry->efd_node, &(db->bb_free_root)); 385418aadd47SBobi Jam mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); 3855c9de560dSAlex Tomas 38563d56b8d2STao Ma /* 38573d56b8d2STao Ma * Clear the trimmed flag for the group so that the next 38583d56b8d2STao Ma * ext4_trim_fs can trim it. 38593d56b8d2STao Ma * If the volume is mounted with -o discard, online discard 38603d56b8d2STao Ma * is supported and the free blocks will be trimmed online. 38613d56b8d2STao Ma */ 38623d56b8d2STao Ma if (!test_opt(sb, DISCARD)) 38633d56b8d2STao Ma EXT4_MB_GRP_CLEAR_TRIMMED(db); 38643d56b8d2STao Ma 3865c894058dSAneesh Kumar K.V if (!db->bb_free_root.rb_node) { 3866c894058dSAneesh Kumar K.V /* No more items in the per group rb tree 3867c894058dSAneesh Kumar K.V * balance refcounts from ext4_mb_free_metadata() 3868c894058dSAneesh Kumar K.V */ 386909cbfeafSKirill A. Shutemov put_page(e4b.bd_buddy_page); 387009cbfeafSKirill A. Shutemov put_page(e4b.bd_bitmap_page); 3871c894058dSAneesh Kumar K.V } 387218aadd47SBobi Jam ext4_unlock_group(sb, entry->efd_group); 3873e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3874c9de560dSAlex Tomas 3875c7f2bafaSKemeng Shi mb_debug(sb, "freed %d blocks in 1 structures\n", count); 3876c9de560dSAlex Tomas } 3877c9de560dSAlex Tomas 3878a0154344SDaeho Jeong /* 3879a0154344SDaeho Jeong * This function is called by the jbd2 layer once the commit has finished, 3880a0154344SDaeho Jeong * so we know we can free the blocks that were released with that commit. 3881a0154344SDaeho Jeong */ 3882a0154344SDaeho Jeong void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) 3883a0154344SDaeho Jeong { 3884a0154344SDaeho Jeong struct ext4_sb_info *sbi = EXT4_SB(sb); 3885a0154344SDaeho Jeong struct ext4_free_data *entry, *tmp; 38860f6bc579SRuan Jinjie LIST_HEAD(freed_data_list); 3887*ce774e53SJinke Han struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1]; 388855cdd0afSWang Jianchao bool wake; 3889a0154344SDaeho Jeong 3890*ce774e53SJinke Han list_replace_init(s_freed_head, &freed_data_list); 3891a0154344SDaeho Jeong 389255cdd0afSWang Jianchao list_for_each_entry(entry, &freed_data_list, efd_list) 3893a0154344SDaeho Jeong ext4_free_data_in_buddy(sb, entry); 389455cdd0afSWang Jianchao 389555cdd0afSWang Jianchao if (test_opt(sb, DISCARD)) { 389655cdd0afSWang Jianchao spin_lock(&sbi->s_md_lock); 389755cdd0afSWang Jianchao wake = list_empty(&sbi->s_discard_list); 389855cdd0afSWang Jianchao list_splice_tail(&freed_data_list, &sbi->s_discard_list); 389955cdd0afSWang Jianchao spin_unlock(&sbi->s_md_lock); 390055cdd0afSWang Jianchao if (wake) 390155cdd0afSWang Jianchao queue_work(system_unbound_wq, &sbi->s_discard_work); 390255cdd0afSWang Jianchao } else { 390355cdd0afSWang Jianchao list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) 390455cdd0afSWang Jianchao kmem_cache_free(ext4_free_data_cachep, entry); 390555cdd0afSWang Jianchao } 3906a0154344SDaeho Jeong } 3907a0154344SDaeho Jeong 39085dabfc78STheodore Ts'o int __init ext4_init_mballoc(void) 3909c9de560dSAlex Tomas { 391016828088STheodore Ts'o ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, 391116828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 3912c9de560dSAlex Tomas if (ext4_pspace_cachep == NULL) 3913f283529aSRitesh Harjani goto out; 3914c9de560dSAlex Tomas 391516828088STheodore Ts'o ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, 391616828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 3917f283529aSRitesh Harjani if (ext4_ac_cachep == NULL) 3918f283529aSRitesh Harjani goto out_pa_free; 3919c894058dSAneesh Kumar K.V 392018aadd47SBobi Jam ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, 392116828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 3922f283529aSRitesh Harjani if (ext4_free_data_cachep == NULL) 3923f283529aSRitesh Harjani goto out_ac_free; 3924f283529aSRitesh Harjani 3925c9de560dSAlex Tomas return 0; 3926f283529aSRitesh Harjani 3927f283529aSRitesh Harjani out_ac_free: 3928f283529aSRitesh Harjani kmem_cache_destroy(ext4_ac_cachep); 3929f283529aSRitesh Harjani out_pa_free: 3930f283529aSRitesh Harjani kmem_cache_destroy(ext4_pspace_cachep); 3931f283529aSRitesh Harjani out: 3932f283529aSRitesh Harjani return -ENOMEM; 3933c9de560dSAlex Tomas } 3934c9de560dSAlex Tomas 39355dabfc78STheodore Ts'o void ext4_exit_mballoc(void) 3936c9de560dSAlex Tomas { 39373e03f9caSJesper Dangaard Brouer /* 39383e03f9caSJesper Dangaard Brouer * Wait for completion of call_rcu()'s on ext4_pspace_cachep 39393e03f9caSJesper Dangaard Brouer * before destroying the slab cache. 39403e03f9caSJesper Dangaard Brouer */ 39413e03f9caSJesper Dangaard Brouer rcu_barrier(); 3942c9de560dSAlex Tomas kmem_cache_destroy(ext4_pspace_cachep); 3943256bdb49SEric Sandeen kmem_cache_destroy(ext4_ac_cachep); 394418aadd47SBobi Jam kmem_cache_destroy(ext4_free_data_cachep); 39452892c15dSEric Sandeen ext4_groupinfo_destroy_slabs(); 3946c9de560dSAlex Tomas } 3947c9de560dSAlex Tomas 3948c9de560dSAlex Tomas 3949c9de560dSAlex Tomas /* 395073b2c716SUwe Kleine-König * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps 3951c9de560dSAlex Tomas * Returns 0 if success or error code 3952c9de560dSAlex Tomas */ 39534ddfef7bSEric Sandeen static noinline_for_stack int 39544ddfef7bSEric Sandeen ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 395553accfa9STheodore Ts'o handle_t *handle, unsigned int reserv_clstrs) 3956c9de560dSAlex Tomas { 3957c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3958c9de560dSAlex Tomas struct ext4_group_desc *gdp; 3959c9de560dSAlex Tomas struct buffer_head *gdp_bh; 3960c9de560dSAlex Tomas struct ext4_sb_info *sbi; 3961c9de560dSAlex Tomas struct super_block *sb; 3962c9de560dSAlex Tomas ext4_fsblk_t block; 3963519deca0SAneesh Kumar K.V int err, len; 3964c9de560dSAlex Tomas 3965c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3966c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len <= 0); 3967c9de560dSAlex Tomas 3968c9de560dSAlex Tomas sb = ac->ac_sb; 3969c9de560dSAlex Tomas sbi = EXT4_SB(sb); 3970c9de560dSAlex Tomas 3971574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 39729008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 3973fb28f9ceSKemeng Shi return PTR_ERR(bitmap_bh); 39749008a58eSDarrick J. Wong } 3975c9de560dSAlex Tomas 39765d601255Sliang xie BUFFER_TRACE(bitmap_bh, "getting write access"); 3977188c299eSJan Kara err = ext4_journal_get_write_access(handle, sb, bitmap_bh, 3978188c299eSJan Kara EXT4_JTR_NONE); 3979c9de560dSAlex Tomas if (err) 3980c9de560dSAlex Tomas goto out_err; 3981c9de560dSAlex Tomas 3982c9de560dSAlex Tomas err = -EIO; 3983c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 3984c9de560dSAlex Tomas if (!gdp) 3985c9de560dSAlex Tomas goto out_err; 3986c9de560dSAlex Tomas 3987a9df9a49STheodore Ts'o ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 3988021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, gdp)); 398903cddb80SAneesh Kumar K.V 39905d601255Sliang xie BUFFER_TRACE(gdp_bh, "get_write_access"); 3991188c299eSJan Kara err = ext4_journal_get_write_access(handle, sb, gdp_bh, EXT4_JTR_NONE); 3992c9de560dSAlex Tomas if (err) 3993c9de560dSAlex Tomas goto out_err; 3994c9de560dSAlex Tomas 3995bda00de7SAkinobu Mita block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3996c9de560dSAlex Tomas 399753accfa9STheodore Ts'o len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 3998ce9f24ccSJan Kara if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { 399912062dddSEric Sandeen ext4_error(sb, "Allocating blocks %llu-%llu which overlap " 40001084f252STheodore Ts'o "fs metadata", block, block+len); 4001519deca0SAneesh Kumar K.V /* File system mounted not to panic on error 4002554a5cccSVegard Nossum * Fix the bitmap and return EFSCORRUPTED 4003519deca0SAneesh Kumar K.V * We leak some of the blocks here. 4004519deca0SAneesh Kumar K.V */ 4005955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 4006123e3016SRitesh Harjani mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 4007519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len); 4008955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 40090390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 4010519deca0SAneesh Kumar K.V if (!err) 4011554a5cccSVegard Nossum err = -EFSCORRUPTED; 4012519deca0SAneesh Kumar K.V goto out_err; 4013c9de560dSAlex Tomas } 4014955ce5f5SAneesh Kumar K.V 4015955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 4016c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 4017c9de560dSAlex Tomas { 4018c9de560dSAlex Tomas int i; 4019c9de560dSAlex Tomas for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 4020c9de560dSAlex Tomas BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 4021c9de560dSAlex Tomas bitmap_bh->b_data)); 4022c9de560dSAlex Tomas } 4023c9de560dSAlex Tomas } 4024c9de560dSAlex Tomas #endif 4025123e3016SRitesh Harjani mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 4026c3e94d1dSYongqiang Yang ac->ac_b_ex.fe_len); 40278844618dSTheodore Ts'o if (ext4_has_group_desc_csum(sb) && 40288844618dSTheodore Ts'o (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 4029c9de560dSAlex Tomas gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 4030021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, 4031cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, 4032560671a0SAneesh Kumar K.V ac->ac_b_ex.fe_group, gdp)); 4033c9de560dSAlex Tomas } 4034021b65bbSTheodore Ts'o len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; 4035021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, len); 40361df9bde4SKemeng Shi ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 4037feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); 4038955ce5f5SAneesh Kumar K.V 4039955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 404057042651STheodore Ts'o percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); 4041d2a17637SMingming Cao /* 40426bc6e63fSAneesh Kumar K.V * Now reduce the dirty block count also. Should not go negative 4043d2a17637SMingming Cao */ 40446bc6e63fSAneesh Kumar K.V if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 40456bc6e63fSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 404657042651STheodore Ts'o percpu_counter_sub(&sbi->s_dirtyclusters_counter, 404757042651STheodore Ts'o reserv_clstrs); 4048c9de560dSAlex Tomas 4049772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 4050772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, 4051772cb7c8SJose R. Santos ac->ac_b_ex.fe_group); 405290ba983fSTheodore Ts'o atomic64_sub(ac->ac_b_ex.fe_len, 40537c990728SSuraj Jitindar Singh &sbi_array_rcu_deref(sbi, s_flex_groups, 40547c990728SSuraj Jitindar Singh flex_group)->free_clusters); 4055772cb7c8SJose R. Santos } 4056772cb7c8SJose R. Santos 40570390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 4058c9de560dSAlex Tomas if (err) 4059c9de560dSAlex Tomas goto out_err; 40600390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 4061c9de560dSAlex Tomas 4062c9de560dSAlex Tomas out_err: 406342a10addSAneesh Kumar K.V brelse(bitmap_bh); 4064c9de560dSAlex Tomas return err; 4065c9de560dSAlex Tomas } 4066c9de560dSAlex Tomas 4067c9de560dSAlex Tomas /* 40688016e29fSHarshad Shirwadkar * Idempotent helper for Ext4 fast commit replay path to set the state of 40698016e29fSHarshad Shirwadkar * blocks in bitmaps and update counters. 40708016e29fSHarshad Shirwadkar */ 40718016e29fSHarshad Shirwadkar void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, 40728016e29fSHarshad Shirwadkar int len, int state) 40738016e29fSHarshad Shirwadkar { 40748016e29fSHarshad Shirwadkar struct buffer_head *bitmap_bh = NULL; 40758016e29fSHarshad Shirwadkar struct ext4_group_desc *gdp; 40768016e29fSHarshad Shirwadkar struct buffer_head *gdp_bh; 40778016e29fSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 40788016e29fSHarshad Shirwadkar ext4_group_t group; 40798016e29fSHarshad Shirwadkar ext4_grpblk_t blkoff; 4080a50bda14SSu Hui int i, err = 0; 40818016e29fSHarshad Shirwadkar int already; 4082bfdc502aSRitesh Harjani unsigned int clen, clen_changed, thisgrp_len; 40838016e29fSHarshad Shirwadkar 4084bfdc502aSRitesh Harjani while (len > 0) { 40858016e29fSHarshad Shirwadkar ext4_get_group_no_and_offset(sb, block, &group, &blkoff); 4086bfdc502aSRitesh Harjani 4087bfdc502aSRitesh Harjani /* 4088bfdc502aSRitesh Harjani * Check to see if we are freeing blocks across a group 4089bfdc502aSRitesh Harjani * boundary. 4090bfdc502aSRitesh Harjani * In case of flex_bg, this can happen that (block, len) may 4091bfdc502aSRitesh Harjani * span across more than one group. In that case we need to 4092bfdc502aSRitesh Harjani * get the corresponding group metadata to work with. 4093bfdc502aSRitesh Harjani * For this we have goto again loop. 4094bfdc502aSRitesh Harjani */ 4095bfdc502aSRitesh Harjani thisgrp_len = min_t(unsigned int, (unsigned int)len, 4096bfdc502aSRitesh Harjani EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff)); 4097bfdc502aSRitesh Harjani clen = EXT4_NUM_B2C(sbi, thisgrp_len); 4098bfdc502aSRitesh Harjani 40998c91c579SRitesh Harjani if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) { 41008c91c579SRitesh Harjani ext4_error(sb, "Marking blocks in system zone - " 41018c91c579SRitesh Harjani "Block = %llu, len = %u", 41028c91c579SRitesh Harjani block, thisgrp_len); 41038c91c579SRitesh Harjani bitmap_bh = NULL; 41048c91c579SRitesh Harjani break; 41058c91c579SRitesh Harjani } 41068c91c579SRitesh Harjani 41078016e29fSHarshad Shirwadkar bitmap_bh = ext4_read_block_bitmap(sb, group); 41088016e29fSHarshad Shirwadkar if (IS_ERR(bitmap_bh)) { 41098016e29fSHarshad Shirwadkar err = PTR_ERR(bitmap_bh); 41108016e29fSHarshad Shirwadkar bitmap_bh = NULL; 4111bfdc502aSRitesh Harjani break; 41128016e29fSHarshad Shirwadkar } 41138016e29fSHarshad Shirwadkar 41148016e29fSHarshad Shirwadkar err = -EIO; 41158016e29fSHarshad Shirwadkar gdp = ext4_get_group_desc(sb, group, &gdp_bh); 41168016e29fSHarshad Shirwadkar if (!gdp) 4117bfdc502aSRitesh Harjani break; 41188016e29fSHarshad Shirwadkar 41198016e29fSHarshad Shirwadkar ext4_lock_group(sb, group); 41208016e29fSHarshad Shirwadkar already = 0; 41218016e29fSHarshad Shirwadkar for (i = 0; i < clen; i++) 4122bfdc502aSRitesh Harjani if (!mb_test_bit(blkoff + i, bitmap_bh->b_data) == 4123bfdc502aSRitesh Harjani !state) 41248016e29fSHarshad Shirwadkar already++; 41258016e29fSHarshad Shirwadkar 4126a5c0e2fdSRitesh Harjani clen_changed = clen - already; 41278016e29fSHarshad Shirwadkar if (state) 4128123e3016SRitesh Harjani mb_set_bits(bitmap_bh->b_data, blkoff, clen); 41298016e29fSHarshad Shirwadkar else 4130bd8247eeSRitesh Harjani mb_clear_bits(bitmap_bh->b_data, blkoff, clen); 41318016e29fSHarshad Shirwadkar if (ext4_has_group_desc_csum(sb) && 41328016e29fSHarshad Shirwadkar (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 41338016e29fSHarshad Shirwadkar gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 41348016e29fSHarshad Shirwadkar ext4_free_group_clusters_set(sb, gdp, 4135bfdc502aSRitesh Harjani ext4_free_clusters_after_init(sb, group, gdp)); 41368016e29fSHarshad Shirwadkar } 41378016e29fSHarshad Shirwadkar if (state) 4138a5c0e2fdSRitesh Harjani clen = ext4_free_group_clusters(sb, gdp) - clen_changed; 41398016e29fSHarshad Shirwadkar else 4140a5c0e2fdSRitesh Harjani clen = ext4_free_group_clusters(sb, gdp) + clen_changed; 41418016e29fSHarshad Shirwadkar 41428016e29fSHarshad Shirwadkar ext4_free_group_clusters_set(sb, gdp, clen); 41431df9bde4SKemeng Shi ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 41448016e29fSHarshad Shirwadkar ext4_group_desc_csum_set(sb, group, gdp); 41458016e29fSHarshad Shirwadkar 41468016e29fSHarshad Shirwadkar ext4_unlock_group(sb, group); 41478016e29fSHarshad Shirwadkar 41488016e29fSHarshad Shirwadkar if (sbi->s_log_groups_per_flex) { 41498016e29fSHarshad Shirwadkar ext4_group_t flex_group = ext4_flex_group(sbi, group); 4150a5c0e2fdSRitesh Harjani struct flex_groups *fg = sbi_array_rcu_deref(sbi, 4151a5c0e2fdSRitesh Harjani s_flex_groups, flex_group); 41528016e29fSHarshad Shirwadkar 4153a5c0e2fdSRitesh Harjani if (state) 4154a5c0e2fdSRitesh Harjani atomic64_sub(clen_changed, &fg->free_clusters); 4155a5c0e2fdSRitesh Harjani else 4156a5c0e2fdSRitesh Harjani atomic64_add(clen_changed, &fg->free_clusters); 4157bfdc502aSRitesh Harjani 41588016e29fSHarshad Shirwadkar } 41598016e29fSHarshad Shirwadkar 41608016e29fSHarshad Shirwadkar err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); 41618016e29fSHarshad Shirwadkar if (err) 4162bfdc502aSRitesh Harjani break; 41638016e29fSHarshad Shirwadkar sync_dirty_buffer(bitmap_bh); 41648016e29fSHarshad Shirwadkar err = ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); 41658016e29fSHarshad Shirwadkar sync_dirty_buffer(gdp_bh); 4166bfdc502aSRitesh Harjani if (err) 4167bfdc502aSRitesh Harjani break; 41688016e29fSHarshad Shirwadkar 4169bfdc502aSRitesh Harjani block += thisgrp_len; 4170bfdc502aSRitesh Harjani len -= thisgrp_len; 4171bfdc502aSRitesh Harjani brelse(bitmap_bh); 4172bfdc502aSRitesh Harjani BUG_ON(len < 0); 4173bfdc502aSRitesh Harjani } 4174bfdc502aSRitesh Harjani 4175bfdc502aSRitesh Harjani if (err) 41768016e29fSHarshad Shirwadkar brelse(bitmap_bh); 41778016e29fSHarshad Shirwadkar } 41788016e29fSHarshad Shirwadkar 41798016e29fSHarshad Shirwadkar /* 4180c9de560dSAlex Tomas * here we normalize request for locality group 4181d7a1fee1SDan Ehrenberg * Group request are normalized to s_mb_group_prealloc, which goes to 4182d7a1fee1SDan Ehrenberg * s_strip if we set the same via mount option. 4183d7a1fee1SDan Ehrenberg * s_mb_group_prealloc can be configured via 4184b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_group_prealloc 4185c9de560dSAlex Tomas * 4186c9de560dSAlex Tomas * XXX: should we try to preallocate more than the group has now? 4187c9de560dSAlex Tomas */ 4188c9de560dSAlex Tomas static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 4189c9de560dSAlex Tomas { 4190c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 4191c9de560dSAlex Tomas struct ext4_locality_group *lg = ac->ac_lg; 4192c9de560dSAlex Tomas 4193c9de560dSAlex Tomas BUG_ON(lg == NULL); 4194c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 4195d3df1453SRitesh Harjani mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); 4196c9de560dSAlex Tomas } 4197c9de560dSAlex Tomas 419838727786SOjaswin Mujoo /* 419938727786SOjaswin Mujoo * This function returns the next element to look at during inode 420038727786SOjaswin Mujoo * PA rbtree walk. We assume that we have held the inode PA rbtree lock 420138727786SOjaswin Mujoo * (ei->i_prealloc_lock) 420238727786SOjaswin Mujoo * 420338727786SOjaswin Mujoo * new_start The start of the range we want to compare 420438727786SOjaswin Mujoo * cur_start The existing start that we are comparing against 420538727786SOjaswin Mujoo * node The node of the rb_tree 420638727786SOjaswin Mujoo */ 420738727786SOjaswin Mujoo static inline struct rb_node* 420838727786SOjaswin Mujoo ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node) 420938727786SOjaswin Mujoo { 421038727786SOjaswin Mujoo if (new_start < cur_start) 421138727786SOjaswin Mujoo return node->rb_left; 421238727786SOjaswin Mujoo else 421338727786SOjaswin Mujoo return node->rb_right; 421438727786SOjaswin Mujoo } 421538727786SOjaswin Mujoo 42167692094aSOjaswin Mujoo static inline void 42177692094aSOjaswin Mujoo ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac, 4218bedc5d34SBaokun Li ext4_lblk_t start, loff_t end) 42197692094aSOjaswin Mujoo { 42207692094aSOjaswin Mujoo struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 42217692094aSOjaswin Mujoo struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 42227692094aSOjaswin Mujoo struct ext4_prealloc_space *tmp_pa; 4223bedc5d34SBaokun Li ext4_lblk_t tmp_pa_start; 4224bedc5d34SBaokun Li loff_t tmp_pa_end; 422538727786SOjaswin Mujoo struct rb_node *iter; 42267692094aSOjaswin Mujoo 422738727786SOjaswin Mujoo read_lock(&ei->i_prealloc_lock); 422838727786SOjaswin Mujoo for (iter = ei->i_prealloc_node.rb_node; iter; 422938727786SOjaswin Mujoo iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) { 423038727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 423138727786SOjaswin Mujoo pa_node.inode_node); 42327692094aSOjaswin Mujoo tmp_pa_start = tmp_pa->pa_lstart; 4233bedc5d34SBaokun Li tmp_pa_end = pa_logical_end(sbi, tmp_pa); 42347692094aSOjaswin Mujoo 423538727786SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 423638727786SOjaswin Mujoo if (tmp_pa->pa_deleted == 0) 42377692094aSOjaswin Mujoo BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start)); 42387692094aSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 42397692094aSOjaswin Mujoo } 424038727786SOjaswin Mujoo read_unlock(&ei->i_prealloc_lock); 42417692094aSOjaswin Mujoo } 42427692094aSOjaswin Mujoo 4243c9de560dSAlex Tomas /* 42440830344cSOjaswin Mujoo * Given an allocation context "ac" and a range "start", "end", check 42450830344cSOjaswin Mujoo * and adjust boundaries if the range overlaps with any of the existing 42460830344cSOjaswin Mujoo * preallocatoins stored in the corresponding inode of the allocation context. 42470830344cSOjaswin Mujoo * 42480830344cSOjaswin Mujoo * Parameters: 42490830344cSOjaswin Mujoo * ac allocation context 42500830344cSOjaswin Mujoo * start start of the new range 42510830344cSOjaswin Mujoo * end end of the new range 42520830344cSOjaswin Mujoo */ 42530830344cSOjaswin Mujoo static inline void 42540830344cSOjaswin Mujoo ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac, 4255bedc5d34SBaokun Li ext4_lblk_t *start, loff_t *end) 42560830344cSOjaswin Mujoo { 42570830344cSOjaswin Mujoo struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 42580830344cSOjaswin Mujoo struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 425938727786SOjaswin Mujoo struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL; 426038727786SOjaswin Mujoo struct rb_node *iter; 4261bedc5d34SBaokun Li ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1; 4262bedc5d34SBaokun Li loff_t new_end, tmp_pa_end, left_pa_end = -1; 42630830344cSOjaswin Mujoo 42640830344cSOjaswin Mujoo new_start = *start; 42650830344cSOjaswin Mujoo new_end = *end; 42660830344cSOjaswin Mujoo 426738727786SOjaswin Mujoo /* 426838727786SOjaswin Mujoo * Adjust the normalized range so that it doesn't overlap with any 426938727786SOjaswin Mujoo * existing preallocated blocks(PAs). Make sure to hold the rbtree lock 427038727786SOjaswin Mujoo * so it doesn't change underneath us. 427138727786SOjaswin Mujoo */ 427238727786SOjaswin Mujoo read_lock(&ei->i_prealloc_lock); 42730830344cSOjaswin Mujoo 427438727786SOjaswin Mujoo /* Step 1: find any one immediate neighboring PA of the normalized range */ 427538727786SOjaswin Mujoo for (iter = ei->i_prealloc_node.rb_node; iter; 427638727786SOjaswin Mujoo iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, 427738727786SOjaswin Mujoo tmp_pa_start, iter)) { 427838727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 427938727786SOjaswin Mujoo pa_node.inode_node); 42800830344cSOjaswin Mujoo tmp_pa_start = tmp_pa->pa_lstart; 4281bedc5d34SBaokun Li tmp_pa_end = pa_logical_end(sbi, tmp_pa); 42820830344cSOjaswin Mujoo 42830830344cSOjaswin Mujoo /* PA must not overlap original request */ 428438727786SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 428538727786SOjaswin Mujoo if (tmp_pa->pa_deleted == 0) 42860830344cSOjaswin Mujoo BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end || 42870830344cSOjaswin Mujoo ac->ac_o_ex.fe_logical < tmp_pa_start)); 42880830344cSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 42890830344cSOjaswin Mujoo } 42900830344cSOjaswin Mujoo 429138727786SOjaswin Mujoo /* 429238727786SOjaswin Mujoo * Step 2: check if the found PA is left or right neighbor and 429338727786SOjaswin Mujoo * get the other neighbor 429438727786SOjaswin Mujoo */ 429538727786SOjaswin Mujoo if (tmp_pa) { 429638727786SOjaswin Mujoo if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) { 429738727786SOjaswin Mujoo struct rb_node *tmp; 429838727786SOjaswin Mujoo 429938727786SOjaswin Mujoo left_pa = tmp_pa; 430038727786SOjaswin Mujoo tmp = rb_next(&left_pa->pa_node.inode_node); 430138727786SOjaswin Mujoo if (tmp) { 430238727786SOjaswin Mujoo right_pa = rb_entry(tmp, 430338727786SOjaswin Mujoo struct ext4_prealloc_space, 430438727786SOjaswin Mujoo pa_node.inode_node); 430538727786SOjaswin Mujoo } 430638727786SOjaswin Mujoo } else { 430738727786SOjaswin Mujoo struct rb_node *tmp; 430838727786SOjaswin Mujoo 430938727786SOjaswin Mujoo right_pa = tmp_pa; 431038727786SOjaswin Mujoo tmp = rb_prev(&right_pa->pa_node.inode_node); 431138727786SOjaswin Mujoo if (tmp) { 431238727786SOjaswin Mujoo left_pa = rb_entry(tmp, 431338727786SOjaswin Mujoo struct ext4_prealloc_space, 431438727786SOjaswin Mujoo pa_node.inode_node); 431538727786SOjaswin Mujoo } 431638727786SOjaswin Mujoo } 431738727786SOjaswin Mujoo } 431838727786SOjaswin Mujoo 431938727786SOjaswin Mujoo /* Step 3: get the non deleted neighbors */ 432038727786SOjaswin Mujoo if (left_pa) { 432138727786SOjaswin Mujoo for (iter = &left_pa->pa_node.inode_node;; 432238727786SOjaswin Mujoo iter = rb_prev(iter)) { 432338727786SOjaswin Mujoo if (!iter) { 432438727786SOjaswin Mujoo left_pa = NULL; 432538727786SOjaswin Mujoo break; 432638727786SOjaswin Mujoo } 432738727786SOjaswin Mujoo 432838727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 432938727786SOjaswin Mujoo pa_node.inode_node); 433038727786SOjaswin Mujoo left_pa = tmp_pa; 433138727786SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 433238727786SOjaswin Mujoo if (tmp_pa->pa_deleted == 0) { 433338727786SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 433438727786SOjaswin Mujoo break; 43350830344cSOjaswin Mujoo } 43360830344cSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 43370830344cSOjaswin Mujoo } 433838727786SOjaswin Mujoo } 433938727786SOjaswin Mujoo 434038727786SOjaswin Mujoo if (right_pa) { 434138727786SOjaswin Mujoo for (iter = &right_pa->pa_node.inode_node;; 434238727786SOjaswin Mujoo iter = rb_next(iter)) { 434338727786SOjaswin Mujoo if (!iter) { 434438727786SOjaswin Mujoo right_pa = NULL; 434538727786SOjaswin Mujoo break; 434638727786SOjaswin Mujoo } 434738727786SOjaswin Mujoo 434838727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 434938727786SOjaswin Mujoo pa_node.inode_node); 435038727786SOjaswin Mujoo right_pa = tmp_pa; 435138727786SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 435238727786SOjaswin Mujoo if (tmp_pa->pa_deleted == 0) { 435338727786SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 435438727786SOjaswin Mujoo break; 435538727786SOjaswin Mujoo } 435638727786SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 435738727786SOjaswin Mujoo } 435838727786SOjaswin Mujoo } 435938727786SOjaswin Mujoo 436038727786SOjaswin Mujoo if (left_pa) { 4361bedc5d34SBaokun Li left_pa_end = pa_logical_end(sbi, left_pa); 436238727786SOjaswin Mujoo BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical); 436338727786SOjaswin Mujoo } 436438727786SOjaswin Mujoo 436538727786SOjaswin Mujoo if (right_pa) { 436638727786SOjaswin Mujoo right_pa_start = right_pa->pa_lstart; 436738727786SOjaswin Mujoo BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical); 436838727786SOjaswin Mujoo } 436938727786SOjaswin Mujoo 437038727786SOjaswin Mujoo /* Step 4: trim our normalized range to not overlap with the neighbors */ 437138727786SOjaswin Mujoo if (left_pa) { 437238727786SOjaswin Mujoo if (left_pa_end > new_start) 437338727786SOjaswin Mujoo new_start = left_pa_end; 437438727786SOjaswin Mujoo } 437538727786SOjaswin Mujoo 437638727786SOjaswin Mujoo if (right_pa) { 437738727786SOjaswin Mujoo if (right_pa_start < new_end) 437838727786SOjaswin Mujoo new_end = right_pa_start; 437938727786SOjaswin Mujoo } 438038727786SOjaswin Mujoo read_unlock(&ei->i_prealloc_lock); 43810830344cSOjaswin Mujoo 43820830344cSOjaswin Mujoo /* XXX: extra loop to check we really don't overlap preallocations */ 43830830344cSOjaswin Mujoo ext4_mb_pa_assert_overlap(ac, new_start, new_end); 43840830344cSOjaswin Mujoo 43850830344cSOjaswin Mujoo *start = new_start; 43860830344cSOjaswin Mujoo *end = new_end; 43870830344cSOjaswin Mujoo } 43880830344cSOjaswin Mujoo 43890830344cSOjaswin Mujoo /* 4390c9de560dSAlex Tomas * Normalization means making request better in terms of 4391c9de560dSAlex Tomas * size and alignment 4392c9de560dSAlex Tomas */ 43934ddfef7bSEric Sandeen static noinline_for_stack void 43944ddfef7bSEric Sandeen ext4_mb_normalize_request(struct ext4_allocation_context *ac, 4395c9de560dSAlex Tomas struct ext4_allocation_request *ar) 4396c9de560dSAlex Tomas { 439753accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4398b07ffe69SKemeng Shi struct ext4_super_block *es = sbi->s_es; 4399c9de560dSAlex Tomas int bsbits, max; 4400bedc5d34SBaokun Li loff_t size, start_off, end; 44011592d2c5SCurt Wohlgemuth loff_t orig_size __maybe_unused; 44025a0790c2SAndi Kleen ext4_lblk_t start; 4403c9de560dSAlex Tomas 4404c9de560dSAlex Tomas /* do normalize only data requests, metadata requests 4405c9de560dSAlex Tomas do not need preallocation */ 4406c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 4407c9de560dSAlex Tomas return; 4408c9de560dSAlex Tomas 4409c9de560dSAlex Tomas /* sometime caller may want exact blocks */ 4410c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 4411c9de560dSAlex Tomas return; 4412c9de560dSAlex Tomas 4413c9de560dSAlex Tomas /* caller may indicate that preallocation isn't 4414c9de560dSAlex Tomas * required (it's a tail, for example) */ 4415c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 4416c9de560dSAlex Tomas return; 4417c9de560dSAlex Tomas 4418c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 4419c9de560dSAlex Tomas ext4_mb_normalize_group_request(ac); 4420c9de560dSAlex Tomas return ; 4421c9de560dSAlex Tomas } 4422c9de560dSAlex Tomas 4423c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 4424c9de560dSAlex Tomas 4425c9de560dSAlex Tomas /* first, let's learn actual file size 4426c9de560dSAlex Tomas * given current request is allocated */ 442743bbddc0SBaokun Li size = extent_logical_end(sbi, &ac->ac_o_ex); 4428c9de560dSAlex Tomas size = size << bsbits; 4429c9de560dSAlex Tomas if (size < i_size_read(ac->ac_inode)) 4430c9de560dSAlex Tomas size = i_size_read(ac->ac_inode); 44315a0790c2SAndi Kleen orig_size = size; 4432c9de560dSAlex Tomas 44331930479cSValerie Clement /* max size of free chunks */ 44341930479cSValerie Clement max = 2 << bsbits; 4435c9de560dSAlex Tomas 44361930479cSValerie Clement #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 44371930479cSValerie Clement (req <= (size) || max <= (chunk_size)) 4438c9de560dSAlex Tomas 4439c9de560dSAlex Tomas /* first, try to predict filesize */ 4440c9de560dSAlex Tomas /* XXX: should this table be tunable? */ 4441c9de560dSAlex Tomas start_off = 0; 4442c9de560dSAlex Tomas if (size <= 16 * 1024) { 4443c9de560dSAlex Tomas size = 16 * 1024; 4444c9de560dSAlex Tomas } else if (size <= 32 * 1024) { 4445c9de560dSAlex Tomas size = 32 * 1024; 4446c9de560dSAlex Tomas } else if (size <= 64 * 1024) { 4447c9de560dSAlex Tomas size = 64 * 1024; 4448c9de560dSAlex Tomas } else if (size <= 128 * 1024) { 4449c9de560dSAlex Tomas size = 128 * 1024; 4450c9de560dSAlex Tomas } else if (size <= 256 * 1024) { 4451c9de560dSAlex Tomas size = 256 * 1024; 4452c9de560dSAlex Tomas } else if (size <= 512 * 1024) { 4453c9de560dSAlex Tomas size = 512 * 1024; 4454c9de560dSAlex Tomas } else if (size <= 1024 * 1024) { 4455c9de560dSAlex Tomas size = 1024 * 1024; 44561930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 4457c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 44581930479cSValerie Clement (21 - bsbits)) << 21; 44591930479cSValerie Clement size = 2 * 1024 * 1024; 44601930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 4461c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 4462c9de560dSAlex Tomas (22 - bsbits)) << 22; 4463c9de560dSAlex Tomas size = 4 * 1024 * 1024; 4464b3916da0SKemeng Shi } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len), 44651930479cSValerie Clement (8<<20)>>bsbits, max, 8 * 1024)) { 4466c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 4467c9de560dSAlex Tomas (23 - bsbits)) << 23; 4468c9de560dSAlex Tomas size = 8 * 1024 * 1024; 4469c9de560dSAlex Tomas } else { 4470c9de560dSAlex Tomas start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; 447191a48aafSKemeng Shi size = (loff_t) EXT4_C2B(sbi, 4472b27b1535SXiaoguang Wang ac->ac_o_ex.fe_len) << bsbits; 4473c9de560dSAlex Tomas } 44745a0790c2SAndi Kleen size = size >> bsbits; 44755a0790c2SAndi Kleen start = start_off >> bsbits; 4476c9de560dSAlex Tomas 4477a08f789dSBaokun Li /* 4478a08f789dSBaokun Li * For tiny groups (smaller than 8MB) the chosen allocation 4479a08f789dSBaokun Li * alignment may be larger than group size. Make sure the 4480a08f789dSBaokun Li * alignment does not move allocation to a different group which 4481a08f789dSBaokun Li * makes mballoc fail assertions later. 4482a08f789dSBaokun Li */ 4483a08f789dSBaokun Li start = max(start, rounddown(ac->ac_o_ex.fe_logical, 4484a08f789dSBaokun Li (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb))); 4485a08f789dSBaokun Li 4486c9de560dSAlex Tomas /* don't cover already allocated blocks in selected range */ 4487c9de560dSAlex Tomas if (ar->pleft && start <= ar->lleft) { 4488c9de560dSAlex Tomas size -= ar->lleft + 1 - start; 4489c9de560dSAlex Tomas start = ar->lleft + 1; 4490c9de560dSAlex Tomas } 4491c9de560dSAlex Tomas if (ar->pright && start + size - 1 >= ar->lright) 4492c9de560dSAlex Tomas size -= start + size - ar->lright; 4493c9de560dSAlex Tomas 4494cd648b8aSJan Kara /* 4495cd648b8aSJan Kara * Trim allocation request for filesystems with artificially small 4496cd648b8aSJan Kara * groups. 4497cd648b8aSJan Kara */ 4498cd648b8aSJan Kara if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) 4499cd648b8aSJan Kara size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); 4500cd648b8aSJan Kara 4501c9de560dSAlex Tomas end = start + size; 4502c9de560dSAlex Tomas 45030830344cSOjaswin Mujoo ext4_mb_pa_adjust_overlap(ac, &start, &end); 4504c9de560dSAlex Tomas 4505c9de560dSAlex Tomas size = end - start; 4506c9de560dSAlex Tomas 4507cf4ff938SBaokun Li /* 4508cf4ff938SBaokun Li * In this function "start" and "size" are normalized for better 4509cf4ff938SBaokun Li * alignment and length such that we could preallocate more blocks. 4510cf4ff938SBaokun Li * This normalization is done such that original request of 4511cf4ff938SBaokun Li * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and 4512cf4ff938SBaokun Li * "size" boundaries. 4513cf4ff938SBaokun Li * (Note fe_len can be relaxed since FS block allocation API does not 4514cf4ff938SBaokun Li * provide gurantee on number of contiguous blocks allocation since that 4515cf4ff938SBaokun Li * depends upon free space left, etc). 4516cf4ff938SBaokun Li * In case of inode pa, later we use the allocated blocks 45171221b235SKemeng Shi * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated 4518cf4ff938SBaokun Li * range of goal/best blocks [start, size] to put it at the 4519cf4ff938SBaokun Li * ac_o_ex.fe_logical extent of this inode. 4520cf4ff938SBaokun Li * (See ext4_mb_use_inode_pa() for more details) 4521cf4ff938SBaokun Li */ 4522cf4ff938SBaokun Li if (start + size <= ac->ac_o_ex.fe_logical || 4523c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical) { 45249d8b9ec4STheodore Ts'o ext4_msg(ac->ac_sb, KERN_ERR, 45259d8b9ec4STheodore Ts'o "start %lu, size %lu, fe_logical %lu", 4526c9de560dSAlex Tomas (unsigned long) start, (unsigned long) size, 4527c9de560dSAlex Tomas (unsigned long) ac->ac_o_ex.fe_logical); 4528dfe076c1SDmitry Monakhov BUG(); 4529c9de560dSAlex Tomas } 4530b5b60778SMaurizio Lombardi BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 4531c9de560dSAlex Tomas 4532c9de560dSAlex Tomas /* now prepare goal request */ 4533c9de560dSAlex Tomas 4534c9de560dSAlex Tomas /* XXX: is it better to align blocks WRT to logical 4535c9de560dSAlex Tomas * placement or satisfy big request as is */ 4536c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = start; 453753accfa9STheodore Ts'o ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); 45387e170922SOjaswin Mujoo ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; 4539c9de560dSAlex Tomas 4540c9de560dSAlex Tomas /* define goal start in order to merge */ 4541b07ffe69SKemeng Shi if (ar->pright && (ar->lright == (start + size)) && 4542b07ffe69SKemeng Shi ar->pright >= size && 4543b07ffe69SKemeng Shi ar->pright - size >= le32_to_cpu(es->s_first_data_block)) { 4544c9de560dSAlex Tomas /* merge to the right */ 4545c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 4546b07ffe69SKemeng Shi &ac->ac_g_ex.fe_group, 4547b07ffe69SKemeng Shi &ac->ac_g_ex.fe_start); 4548c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 4549c9de560dSAlex Tomas } 4550b07ffe69SKemeng Shi if (ar->pleft && (ar->lleft + 1 == start) && 4551b07ffe69SKemeng Shi ar->pleft + 1 < ext4_blocks_count(es)) { 4552c9de560dSAlex Tomas /* merge to the left */ 4553c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 4554b07ffe69SKemeng Shi &ac->ac_g_ex.fe_group, 4555b07ffe69SKemeng Shi &ac->ac_g_ex.fe_start); 4556c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 4557c9de560dSAlex Tomas } 4558c9de560dSAlex Tomas 4559d3df1453SRitesh Harjani mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, 4560d3df1453SRitesh Harjani orig_size, start); 4561c9de560dSAlex Tomas } 4562c9de560dSAlex Tomas 4563c9de560dSAlex Tomas static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 4564c9de560dSAlex Tomas { 4565c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4566c9de560dSAlex Tomas 4567a6c75eafSHarshad Shirwadkar if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { 4568c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_reqs); 4569c9de560dSAlex Tomas atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 4570291dae47SCurt Wohlgemuth if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) 4571c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_success); 4572fdd9a009SOjaswin Mujoo 4573c9de560dSAlex Tomas atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 4574fdd9a009SOjaswin Mujoo for (int i=0; i<EXT4_MB_NUM_CRS; i++) { 4575fdd9a009SOjaswin Mujoo atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]); 4576fdd9a009SOjaswin Mujoo } 4577fdd9a009SOjaswin Mujoo 4578a6c75eafSHarshad Shirwadkar atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); 4579c9de560dSAlex Tomas if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 4580c9de560dSAlex Tomas ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 4581c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_goals); 45827e170922SOjaswin Mujoo /* did we allocate as much as normalizer originally wanted? */ 45837e170922SOjaswin Mujoo if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len) 45843ef5d263SOjaswin Mujoo atomic_inc(&sbi->s_bal_len_goals); 45857e170922SOjaswin Mujoo 4586c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan) 4587c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_breaks); 4588c9de560dSAlex Tomas } 4589c9de560dSAlex Tomas 4590296c355cSTheodore Ts'o if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 4591296c355cSTheodore Ts'o trace_ext4_mballoc_alloc(ac); 4592296c355cSTheodore Ts'o else 4593296c355cSTheodore Ts'o trace_ext4_mballoc_prealloc(ac); 4594c9de560dSAlex Tomas } 4595c9de560dSAlex Tomas 4596c9de560dSAlex Tomas /* 4597b844167eSCurt Wohlgemuth * Called on failure; free up any blocks from the inode PA for this 4598b844167eSCurt Wohlgemuth * context. We don't need this for MB_GROUP_PA because we only change 4599b844167eSCurt Wohlgemuth * pa_free in ext4_mb_release_context(), but on failure, we've already 4600b844167eSCurt Wohlgemuth * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. 4601b844167eSCurt Wohlgemuth */ 4602b844167eSCurt Wohlgemuth static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) 4603b844167eSCurt Wohlgemuth { 4604b844167eSCurt Wohlgemuth struct ext4_prealloc_space *pa = ac->ac_pa; 460586f0afd4STheodore Ts'o struct ext4_buddy e4b; 460686f0afd4STheodore Ts'o int err; 4607b844167eSCurt Wohlgemuth 460886f0afd4STheodore Ts'o if (pa == NULL) { 4609c99d1e6eSTheodore Ts'o if (ac->ac_f_ex.fe_len == 0) 4610c99d1e6eSTheodore Ts'o return; 461186f0afd4STheodore Ts'o err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); 461219b8b035STheodore Ts'o if (WARN_RATELIMIT(err, 461319b8b035STheodore Ts'o "ext4: mb_load_buddy failed (%d)", err)) 461486f0afd4STheodore Ts'o /* 461586f0afd4STheodore Ts'o * This should never happen since we pin the 461686f0afd4STheodore Ts'o * pages in the ext4_allocation_context so 461786f0afd4STheodore Ts'o * ext4_mb_load_buddy() should never fail. 461886f0afd4STheodore Ts'o */ 461986f0afd4STheodore Ts'o return; 462086f0afd4STheodore Ts'o ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 462186f0afd4STheodore Ts'o mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, 462286f0afd4STheodore Ts'o ac->ac_f_ex.fe_len); 462386f0afd4STheodore Ts'o ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 4624c99d1e6eSTheodore Ts'o ext4_mb_unload_buddy(&e4b); 462586f0afd4STheodore Ts'o return; 462686f0afd4STheodore Ts'o } 462736cb0f52SKemeng Shi if (pa->pa_type == MB_INODE_PA) { 462836cb0f52SKemeng Shi spin_lock(&pa->pa_lock); 4629400db9d3SZheng Liu pa->pa_free += ac->ac_b_ex.fe_len; 463036cb0f52SKemeng Shi spin_unlock(&pa->pa_lock); 463136cb0f52SKemeng Shi } 4632b844167eSCurt Wohlgemuth } 4633b844167eSCurt Wohlgemuth 4634b844167eSCurt Wohlgemuth /* 4635c9de560dSAlex Tomas * use blocks preallocated to inode 4636c9de560dSAlex Tomas */ 4637c9de560dSAlex Tomas static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 4638c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 4639c9de560dSAlex Tomas { 464053accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4641c9de560dSAlex Tomas ext4_fsblk_t start; 4642c9de560dSAlex Tomas ext4_fsblk_t end; 4643c9de560dSAlex Tomas int len; 4644c9de560dSAlex Tomas 4645c9de560dSAlex Tomas /* found preallocated blocks, use them */ 4646c9de560dSAlex Tomas start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 464753accfa9STheodore Ts'o end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), 464853accfa9STheodore Ts'o start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); 464953accfa9STheodore Ts'o len = EXT4_NUM_B2C(sbi, end - start); 4650c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 4651c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 4652c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 4653c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 4654c9de560dSAlex Tomas ac->ac_pa = pa; 4655c9de560dSAlex Tomas 4656c9de560dSAlex Tomas BUG_ON(start < pa->pa_pstart); 465753accfa9STheodore Ts'o BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); 4658c9de560dSAlex Tomas BUG_ON(pa->pa_free < len); 465993cdf49fSOjaswin Mujoo BUG_ON(ac->ac_b_ex.fe_len <= 0); 4660c9de560dSAlex Tomas pa->pa_free -= len; 4661c9de560dSAlex Tomas 4662d3df1453SRitesh Harjani mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); 4663c9de560dSAlex Tomas } 4664c9de560dSAlex Tomas 4665c9de560dSAlex Tomas /* 4666c9de560dSAlex Tomas * use blocks preallocated to locality group 4667c9de560dSAlex Tomas */ 4668c9de560dSAlex Tomas static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 4669c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 4670c9de560dSAlex Tomas { 467103cddb80SAneesh Kumar K.V unsigned int len = ac->ac_o_ex.fe_len; 46726be2ded1SAneesh Kumar K.V 4673c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 4674c9de560dSAlex Tomas &ac->ac_b_ex.fe_group, 4675c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 4676c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 4677c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 4678c9de560dSAlex Tomas ac->ac_pa = pa; 4679c9de560dSAlex Tomas 46801221b235SKemeng Shi /* we don't correct pa_pstart or pa_len here to avoid 468126346ff6SAneesh Kumar K.V * possible race when the group is being loaded concurrently 4682c9de560dSAlex Tomas * instead we correct pa later, after blocks are marked 468326346ff6SAneesh Kumar K.V * in on-disk bitmap -- see ext4_mb_release_context() 468426346ff6SAneesh Kumar K.V * Other CPUs are prevented from allocating from this pa by lg_mutex 4685c9de560dSAlex Tomas */ 4686d3df1453SRitesh Harjani mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", 46871afdc588SKemeng Shi pa->pa_lstart, len, pa); 4688c9de560dSAlex Tomas } 4689c9de560dSAlex Tomas 4690c9de560dSAlex Tomas /* 46915e745b04SAneesh Kumar K.V * Return the prealloc space that have minimal distance 46925e745b04SAneesh Kumar K.V * from the goal block. @cpa is the prealloc 46935e745b04SAneesh Kumar K.V * space that is having currently known minimal distance 46945e745b04SAneesh Kumar K.V * from the goal block. 46955e745b04SAneesh Kumar K.V */ 46965e745b04SAneesh Kumar K.V static struct ext4_prealloc_space * 46975e745b04SAneesh Kumar K.V ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 46985e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, 46995e745b04SAneesh Kumar K.V struct ext4_prealloc_space *cpa) 47005e745b04SAneesh Kumar K.V { 47015e745b04SAneesh Kumar K.V ext4_fsblk_t cur_distance, new_distance; 47025e745b04SAneesh Kumar K.V 47035e745b04SAneesh Kumar K.V if (cpa == NULL) { 47045e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 47055e745b04SAneesh Kumar K.V return pa; 47065e745b04SAneesh Kumar K.V } 470779211c8eSAndrew Morton cur_distance = abs(goal_block - cpa->pa_pstart); 470879211c8eSAndrew Morton new_distance = abs(goal_block - pa->pa_pstart); 47095e745b04SAneesh Kumar K.V 47105a54b2f1SColy Li if (cur_distance <= new_distance) 47115e745b04SAneesh Kumar K.V return cpa; 47125e745b04SAneesh Kumar K.V 47135e745b04SAneesh Kumar K.V /* drop the previous reference */ 47145e745b04SAneesh Kumar K.V atomic_dec(&cpa->pa_count); 47155e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 47165e745b04SAneesh Kumar K.V return pa; 47175e745b04SAneesh Kumar K.V } 47185e745b04SAneesh Kumar K.V 47195e745b04SAneesh Kumar K.V /* 47201eff5904SKemeng Shi * check if found pa meets EXT4_MB_HINT_GOAL_ONLY 47211eff5904SKemeng Shi */ 47221eff5904SKemeng Shi static bool 47231eff5904SKemeng Shi ext4_mb_pa_goal_check(struct ext4_allocation_context *ac, 47241eff5904SKemeng Shi struct ext4_prealloc_space *pa) 47251eff5904SKemeng Shi { 47261eff5904SKemeng Shi struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 47271eff5904SKemeng Shi ext4_fsblk_t start; 47281eff5904SKemeng Shi 47291eff5904SKemeng Shi if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))) 47301eff5904SKemeng Shi return true; 47311eff5904SKemeng Shi 47321eff5904SKemeng Shi /* 47331eff5904SKemeng Shi * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted 47341eff5904SKemeng Shi * in ext4_mb_normalize_request and will keep same with ac_o_ex 47351eff5904SKemeng Shi * from ext4_mb_initialize_context. Choose ac_g_ex here to keep 47361eff5904SKemeng Shi * consistent with ext4_mb_find_by_goal. 47371eff5904SKemeng Shi */ 47381eff5904SKemeng Shi start = pa->pa_pstart + 47391eff5904SKemeng Shi (ac->ac_g_ex.fe_logical - pa->pa_lstart); 47401eff5904SKemeng Shi if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start) 47411eff5904SKemeng Shi return false; 47421eff5904SKemeng Shi 47431eff5904SKemeng Shi if (ac->ac_g_ex.fe_len > pa->pa_len - 47441eff5904SKemeng Shi EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart)) 47451eff5904SKemeng Shi return false; 47461eff5904SKemeng Shi 47471eff5904SKemeng Shi return true; 47481eff5904SKemeng Shi } 47491eff5904SKemeng Shi 47501eff5904SKemeng Shi /* 4751c9de560dSAlex Tomas * search goal blocks in preallocated space 4752c9de560dSAlex Tomas */ 47534fca8f07SRitesh Harjani static noinline_for_stack bool 47544ddfef7bSEric Sandeen ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 4755c9de560dSAlex Tomas { 475653accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 47576be2ded1SAneesh Kumar K.V int order, i; 4758c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 4759c9de560dSAlex Tomas struct ext4_locality_group *lg; 47609d3de7eeSOjaswin Mujoo struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL; 476138727786SOjaswin Mujoo struct rb_node *iter; 47625e745b04SAneesh Kumar K.V ext4_fsblk_t goal_block; 4763c9de560dSAlex Tomas 4764c9de560dSAlex Tomas /* only data can be preallocated */ 4765c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 47664fca8f07SRitesh Harjani return false; 4767c9de560dSAlex Tomas 47689d3de7eeSOjaswin Mujoo /* 47699d3de7eeSOjaswin Mujoo * first, try per-file preallocation by searching the inode pa rbtree. 47709d3de7eeSOjaswin Mujoo * 47719d3de7eeSOjaswin Mujoo * Here, we can't do a direct traversal of the tree because 47729d3de7eeSOjaswin Mujoo * ext4_mb_discard_group_preallocation() can paralelly mark the pa 47739d3de7eeSOjaswin Mujoo * deleted and that can cause direct traversal to skip some entries. 47749d3de7eeSOjaswin Mujoo */ 477538727786SOjaswin Mujoo read_lock(&ei->i_prealloc_lock); 47769d3de7eeSOjaswin Mujoo 47779d3de7eeSOjaswin Mujoo if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) { 47789d3de7eeSOjaswin Mujoo goto try_group_pa; 47799d3de7eeSOjaswin Mujoo } 47809d3de7eeSOjaswin Mujoo 47819d3de7eeSOjaswin Mujoo /* 47829d3de7eeSOjaswin Mujoo * Step 1: Find a pa with logical start immediately adjacent to the 47839d3de7eeSOjaswin Mujoo * original logical start. This could be on the left or right. 47849d3de7eeSOjaswin Mujoo * 47859d3de7eeSOjaswin Mujoo * (tmp_pa->pa_lstart never changes so we can skip locking for it). 47869d3de7eeSOjaswin Mujoo */ 478738727786SOjaswin Mujoo for (iter = ei->i_prealloc_node.rb_node; iter; 478838727786SOjaswin Mujoo iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, 47899d3de7eeSOjaswin Mujoo tmp_pa->pa_lstart, iter)) { 479038727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 479138727786SOjaswin Mujoo pa_node.inode_node); 47929d3de7eeSOjaswin Mujoo } 4793c9de560dSAlex Tomas 47949d3de7eeSOjaswin Mujoo /* 47959d3de7eeSOjaswin Mujoo * Step 2: The adjacent pa might be to the right of logical start, find 47969d3de7eeSOjaswin Mujoo * the left adjacent pa. After this step we'd have a valid tmp_pa whose 47979d3de7eeSOjaswin Mujoo * logical start is towards the left of original request's logical start 47989d3de7eeSOjaswin Mujoo */ 47999d3de7eeSOjaswin Mujoo if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) { 48009d3de7eeSOjaswin Mujoo struct rb_node *tmp; 48019d3de7eeSOjaswin Mujoo tmp = rb_prev(&tmp_pa->pa_node.inode_node); 4802bcf43499SOjaswin Mujoo 48039d3de7eeSOjaswin Mujoo if (tmp) { 48049d3de7eeSOjaswin Mujoo tmp_pa = rb_entry(tmp, struct ext4_prealloc_space, 48059d3de7eeSOjaswin Mujoo pa_node.inode_node); 48069d3de7eeSOjaswin Mujoo } else { 48079d3de7eeSOjaswin Mujoo /* 48089d3de7eeSOjaswin Mujoo * If there is no adjacent pa to the left then finding 48099d3de7eeSOjaswin Mujoo * an overlapping pa is not possible hence stop searching 48109d3de7eeSOjaswin Mujoo * inode pa tree 48119d3de7eeSOjaswin Mujoo */ 48129d3de7eeSOjaswin Mujoo goto try_group_pa; 48139d3de7eeSOjaswin Mujoo } 48149d3de7eeSOjaswin Mujoo } 48159d3de7eeSOjaswin Mujoo 48169d3de7eeSOjaswin Mujoo BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); 48179d3de7eeSOjaswin Mujoo 48189d3de7eeSOjaswin Mujoo /* 48199d3de7eeSOjaswin Mujoo * Step 3: If the left adjacent pa is deleted, keep moving left to find 48209d3de7eeSOjaswin Mujoo * the first non deleted adjacent pa. After this step we should have a 48219d3de7eeSOjaswin Mujoo * valid tmp_pa which is guaranteed to be non deleted. 48229d3de7eeSOjaswin Mujoo */ 48239d3de7eeSOjaswin Mujoo for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) { 48249d3de7eeSOjaswin Mujoo if (!iter) { 48259d3de7eeSOjaswin Mujoo /* 48269d3de7eeSOjaswin Mujoo * no non deleted left adjacent pa, so stop searching 48279d3de7eeSOjaswin Mujoo * inode pa tree 48289d3de7eeSOjaswin Mujoo */ 48299d3de7eeSOjaswin Mujoo goto try_group_pa; 48309d3de7eeSOjaswin Mujoo } 48319d3de7eeSOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 48329d3de7eeSOjaswin Mujoo pa_node.inode_node); 48339d3de7eeSOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 48349d3de7eeSOjaswin Mujoo if (tmp_pa->pa_deleted == 0) { 48359d3de7eeSOjaswin Mujoo /* 48369d3de7eeSOjaswin Mujoo * We will keep holding the pa_lock from 48379d3de7eeSOjaswin Mujoo * this point on because we don't want group discard 48389d3de7eeSOjaswin Mujoo * to delete this pa underneath us. Since group 48399d3de7eeSOjaswin Mujoo * discard is anyways an ENOSPC operation it 48409d3de7eeSOjaswin Mujoo * should be okay for it to wait a few more cycles. 48419d3de7eeSOjaswin Mujoo */ 48429d3de7eeSOjaswin Mujoo break; 48439d3de7eeSOjaswin Mujoo } else { 48449d3de7eeSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 48459d3de7eeSOjaswin Mujoo } 48469d3de7eeSOjaswin Mujoo } 48479d3de7eeSOjaswin Mujoo 48489d3de7eeSOjaswin Mujoo BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); 48499d3de7eeSOjaswin Mujoo BUG_ON(tmp_pa->pa_deleted == 1); 48509d3de7eeSOjaswin Mujoo 48519d3de7eeSOjaswin Mujoo /* 48529d3de7eeSOjaswin Mujoo * Step 4: We now have the non deleted left adjacent pa. Only this 48539d3de7eeSOjaswin Mujoo * pa can possibly satisfy the request hence check if it overlaps 48549d3de7eeSOjaswin Mujoo * original logical start and stop searching if it doesn't. 48559d3de7eeSOjaswin Mujoo */ 485643bbddc0SBaokun Li if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) { 48579d3de7eeSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 48589d3de7eeSOjaswin Mujoo goto try_group_pa; 48599d3de7eeSOjaswin Mujoo } 4860c9de560dSAlex Tomas 4861fb0a387dSEric Sandeen /* non-extent files can't have physical blocks past 2^32 */ 486212e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && 4863bcf43499SOjaswin Mujoo (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) > 4864e86a7182SOjaswin Mujoo EXT4_MAX_BLOCK_FILE_PHYS)) { 4865e86a7182SOjaswin Mujoo /* 48669d3de7eeSOjaswin Mujoo * Since PAs don't overlap, we won't find any other PA to 48679d3de7eeSOjaswin Mujoo * satisfy this. 4868e86a7182SOjaswin Mujoo */ 48699d3de7eeSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 48709d3de7eeSOjaswin Mujoo goto try_group_pa; 4871e86a7182SOjaswin Mujoo } 4872fb0a387dSEric Sandeen 48739d3de7eeSOjaswin Mujoo if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) { 4874bcf43499SOjaswin Mujoo atomic_inc(&tmp_pa->pa_count); 4875bcf43499SOjaswin Mujoo ext4_mb_use_inode_pa(ac, tmp_pa); 4876bcf43499SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 487738727786SOjaswin Mujoo read_unlock(&ei->i_prealloc_lock); 48784fca8f07SRitesh Harjani return true; 48799d3de7eeSOjaswin Mujoo } else { 48809d3de7eeSOjaswin Mujoo /* 48819d3de7eeSOjaswin Mujoo * We found a valid overlapping pa but couldn't use it because 48829d3de7eeSOjaswin Mujoo * it had no free blocks. This should ideally never happen 48839d3de7eeSOjaswin Mujoo * because: 48849d3de7eeSOjaswin Mujoo * 48859d3de7eeSOjaswin Mujoo * 1. When a new inode pa is added to rbtree it must have 48869d3de7eeSOjaswin Mujoo * pa_free > 0 since otherwise we won't actually need 48879d3de7eeSOjaswin Mujoo * preallocation. 48889d3de7eeSOjaswin Mujoo * 48899d3de7eeSOjaswin Mujoo * 2. An inode pa that is in the rbtree can only have it's 48909d3de7eeSOjaswin Mujoo * pa_free become zero when another thread calls: 48919d3de7eeSOjaswin Mujoo * ext4_mb_new_blocks 48929d3de7eeSOjaswin Mujoo * ext4_mb_use_preallocated 48939d3de7eeSOjaswin Mujoo * ext4_mb_use_inode_pa 48949d3de7eeSOjaswin Mujoo * 48959d3de7eeSOjaswin Mujoo * 3. Further, after the above calls make pa_free == 0, we will 48969d3de7eeSOjaswin Mujoo * immediately remove it from the rbtree in: 48979d3de7eeSOjaswin Mujoo * ext4_mb_new_blocks 48989d3de7eeSOjaswin Mujoo * ext4_mb_release_context 48999d3de7eeSOjaswin Mujoo * ext4_mb_put_pa 49009d3de7eeSOjaswin Mujoo * 49019d3de7eeSOjaswin Mujoo * 4. Since the pa_free becoming 0 and pa_free getting removed 49029d3de7eeSOjaswin Mujoo * from tree both happen in ext4_mb_new_blocks, which is always 49039d3de7eeSOjaswin Mujoo * called with i_data_sem held for data allocations, we can be 49049d3de7eeSOjaswin Mujoo * sure that another process will never see a pa in rbtree with 49059d3de7eeSOjaswin Mujoo * pa_free == 0. 49069d3de7eeSOjaswin Mujoo */ 49079d3de7eeSOjaswin Mujoo WARN_ON_ONCE(tmp_pa->pa_free == 0); 4908c9de560dSAlex Tomas } 4909bcf43499SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 49109d3de7eeSOjaswin Mujoo try_group_pa: 491138727786SOjaswin Mujoo read_unlock(&ei->i_prealloc_lock); 4912c9de560dSAlex Tomas 4913c9de560dSAlex Tomas /* can we use group allocation? */ 4914c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 49154fca8f07SRitesh Harjani return false; 4916c9de560dSAlex Tomas 4917c9de560dSAlex Tomas /* inode may have no locality group for some reason */ 4918c9de560dSAlex Tomas lg = ac->ac_lg; 4919c9de560dSAlex Tomas if (lg == NULL) 49204fca8f07SRitesh Harjani return false; 49216be2ded1SAneesh Kumar K.V order = fls(ac->ac_o_ex.fe_len) - 1; 49226be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 49236be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 49246be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 4925c9de560dSAlex Tomas 4926bda00de7SAkinobu Mita goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); 49275e745b04SAneesh Kumar K.V /* 49285e745b04SAneesh Kumar K.V * search for the prealloc space that is having 49295e745b04SAneesh Kumar K.V * minimal distance from the goal block. 49305e745b04SAneesh Kumar K.V */ 49316be2ded1SAneesh Kumar K.V for (i = order; i < PREALLOC_TB_SIZE; i++) { 4932c9de560dSAlex Tomas rcu_read_lock(); 4933bcf43499SOjaswin Mujoo list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i], 4934a8e38fd3SOjaswin Mujoo pa_node.lg_list) { 4935bcf43499SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 4936bcf43499SOjaswin Mujoo if (tmp_pa->pa_deleted == 0 && 4937bcf43499SOjaswin Mujoo tmp_pa->pa_free >= ac->ac_o_ex.fe_len) { 49385e745b04SAneesh Kumar K.V 49395e745b04SAneesh Kumar K.V cpa = ext4_mb_check_group_pa(goal_block, 4940bcf43499SOjaswin Mujoo tmp_pa, cpa); 49415e745b04SAneesh Kumar K.V } 4942bcf43499SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 49435e745b04SAneesh Kumar K.V } 49445e745b04SAneesh Kumar K.V rcu_read_unlock(); 49455e745b04SAneesh Kumar K.V } 49465e745b04SAneesh Kumar K.V if (cpa) { 49475e745b04SAneesh Kumar K.V ext4_mb_use_group_pa(ac, cpa); 49484fca8f07SRitesh Harjani return true; 4949c9de560dSAlex Tomas } 49504fca8f07SRitesh Harjani return false; 4951c9de560dSAlex Tomas } 4952c9de560dSAlex Tomas 4953c9de560dSAlex Tomas /* 49547a2fcbf7SAneesh Kumar K.V * the function goes through all block freed in the group 49557a2fcbf7SAneesh Kumar K.V * but not yet committed and marks them used in in-core bitmap. 49567a2fcbf7SAneesh Kumar K.V * buddy must be generated from this bitmap 4957955ce5f5SAneesh Kumar K.V * Need to be called with the ext4 group lock held 49587a2fcbf7SAneesh Kumar K.V */ 49597a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 49607a2fcbf7SAneesh Kumar K.V ext4_group_t group) 49617a2fcbf7SAneesh Kumar K.V { 49627a2fcbf7SAneesh Kumar K.V struct rb_node *n; 49637a2fcbf7SAneesh Kumar K.V struct ext4_group_info *grp; 49647a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 49657a2fcbf7SAneesh Kumar K.V 49667a2fcbf7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 49675354b2afSTheodore Ts'o if (!grp) 49685354b2afSTheodore Ts'o return; 49697a2fcbf7SAneesh Kumar K.V n = rb_first(&(grp->bb_free_root)); 49707a2fcbf7SAneesh Kumar K.V 49717a2fcbf7SAneesh Kumar K.V while (n) { 497218aadd47SBobi Jam entry = rb_entry(n, struct ext4_free_data, efd_node); 4973123e3016SRitesh Harjani mb_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); 49747a2fcbf7SAneesh Kumar K.V n = rb_next(n); 49757a2fcbf7SAneesh Kumar K.V } 49767a2fcbf7SAneesh Kumar K.V } 49777a2fcbf7SAneesh Kumar K.V 49787a2fcbf7SAneesh Kumar K.V /* 4979c9de560dSAlex Tomas * the function goes through all preallocation in this group and marks them 4980c9de560dSAlex Tomas * used in in-core bitmap. buddy must be generated from this bitmap 4981955ce5f5SAneesh Kumar K.V * Need to be called with ext4 group lock held 4982c9de560dSAlex Tomas */ 4983089ceeccSEric Sandeen static noinline_for_stack 4984089ceeccSEric Sandeen void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 4985c9de560dSAlex Tomas ext4_group_t group) 4986c9de560dSAlex Tomas { 4987c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 4988c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 4989c9de560dSAlex Tomas struct list_head *cur; 4990c9de560dSAlex Tomas ext4_group_t groupnr; 4991c9de560dSAlex Tomas ext4_grpblk_t start; 4992c9de560dSAlex Tomas int preallocated = 0; 4993c9de560dSAlex Tomas int len; 4994c9de560dSAlex Tomas 49955354b2afSTheodore Ts'o if (!grp) 49965354b2afSTheodore Ts'o return; 49975354b2afSTheodore Ts'o 4998c9de560dSAlex Tomas /* all form of preallocation discards first load group, 4999c9de560dSAlex Tomas * so the only competing code is preallocation use. 5000c9de560dSAlex Tomas * we don't need any locking here 5001c9de560dSAlex Tomas * notice we do NOT ignore preallocations with pa_deleted 5002c9de560dSAlex Tomas * otherwise we could leave used blocks available for 5003c9de560dSAlex Tomas * allocation in buddy when concurrent ext4_mb_put_pa() 5004c9de560dSAlex Tomas * is dropping preallocation 5005c9de560dSAlex Tomas */ 5006c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 5007c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 5008c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 5009c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 5010c9de560dSAlex Tomas &groupnr, &start); 5011c9de560dSAlex Tomas len = pa->pa_len; 5012c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5013c9de560dSAlex Tomas if (unlikely(len == 0)) 5014c9de560dSAlex Tomas continue; 5015c9de560dSAlex Tomas BUG_ON(groupnr != group); 5016123e3016SRitesh Harjani mb_set_bits(bitmap, start, len); 5017c9de560dSAlex Tomas preallocated += len; 5018c9de560dSAlex Tomas } 5019d3df1453SRitesh Harjani mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); 5020c9de560dSAlex Tomas } 5021c9de560dSAlex Tomas 502227bc446eSbrookxu static void ext4_mb_mark_pa_deleted(struct super_block *sb, 502327bc446eSbrookxu struct ext4_prealloc_space *pa) 502427bc446eSbrookxu { 502527bc446eSbrookxu struct ext4_inode_info *ei; 502627bc446eSbrookxu 502727bc446eSbrookxu if (pa->pa_deleted) { 502827bc446eSbrookxu ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", 502927bc446eSbrookxu pa->pa_type, pa->pa_pstart, pa->pa_lstart, 503027bc446eSbrookxu pa->pa_len); 503127bc446eSbrookxu return; 503227bc446eSbrookxu } 503327bc446eSbrookxu 503427bc446eSbrookxu pa->pa_deleted = 1; 503527bc446eSbrookxu 503627bc446eSbrookxu if (pa->pa_type == MB_INODE_PA) { 503727bc446eSbrookxu ei = EXT4_I(pa->pa_inode); 503827bc446eSbrookxu atomic_dec(&ei->i_prealloc_active); 503927bc446eSbrookxu } 504027bc446eSbrookxu } 504127bc446eSbrookxu 504282089725SOjaswin Mujoo static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa) 5043c9de560dSAlex Tomas { 504482089725SOjaswin Mujoo BUG_ON(!pa); 50454e8d2139SJunho Ryu BUG_ON(atomic_read(&pa->pa_count)); 50464e8d2139SJunho Ryu BUG_ON(pa->pa_deleted == 0); 5047c9de560dSAlex Tomas kmem_cache_free(ext4_pspace_cachep, pa); 5048c9de560dSAlex Tomas } 5049c9de560dSAlex Tomas 505082089725SOjaswin Mujoo static void ext4_mb_pa_callback(struct rcu_head *head) 505182089725SOjaswin Mujoo { 505282089725SOjaswin Mujoo struct ext4_prealloc_space *pa; 505382089725SOjaswin Mujoo 505482089725SOjaswin Mujoo pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 505582089725SOjaswin Mujoo ext4_mb_pa_free(pa); 505682089725SOjaswin Mujoo } 505782089725SOjaswin Mujoo 5058c9de560dSAlex Tomas /* 5059c9de560dSAlex Tomas * drops a reference to preallocated space descriptor 5060c9de560dSAlex Tomas * if this was the last reference and the space is consumed 5061c9de560dSAlex Tomas */ 5062c9de560dSAlex Tomas static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 5063c9de560dSAlex Tomas struct super_block *sb, struct ext4_prealloc_space *pa) 5064c9de560dSAlex Tomas { 5065a9df9a49STheodore Ts'o ext4_group_t grp; 5066d33a1976SEric Sandeen ext4_fsblk_t grp_blk; 506738727786SOjaswin Mujoo struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 5068c9de560dSAlex Tomas 5069c9de560dSAlex Tomas /* in this short window concurrent discard can set pa_deleted */ 5070c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 50714e8d2139SJunho Ryu if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { 50724e8d2139SJunho Ryu spin_unlock(&pa->pa_lock); 50734e8d2139SJunho Ryu return; 50744e8d2139SJunho Ryu } 50754e8d2139SJunho Ryu 5076c9de560dSAlex Tomas if (pa->pa_deleted == 1) { 5077c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5078c9de560dSAlex Tomas return; 5079c9de560dSAlex Tomas } 5080c9de560dSAlex Tomas 508127bc446eSbrookxu ext4_mb_mark_pa_deleted(sb, pa); 5082c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5083c9de560dSAlex Tomas 5084d33a1976SEric Sandeen grp_blk = pa->pa_pstart; 5085cc0fb9adSAneesh Kumar K.V /* 5086cc0fb9adSAneesh Kumar K.V * If doing group-based preallocation, pa_pstart may be in the 5087cc0fb9adSAneesh Kumar K.V * next group when pa is used up 5088cc0fb9adSAneesh Kumar K.V */ 5089cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 5090d33a1976SEric Sandeen grp_blk--; 5091d33a1976SEric Sandeen 5092bd86298eSLukas Czerner grp = ext4_get_group_number(sb, grp_blk); 5093c9de560dSAlex Tomas 5094c9de560dSAlex Tomas /* 5095c9de560dSAlex Tomas * possible race: 5096c9de560dSAlex Tomas * 5097c9de560dSAlex Tomas * P1 (buddy init) P2 (regular allocation) 5098c9de560dSAlex Tomas * find block B in PA 5099c9de560dSAlex Tomas * copy on-disk bitmap to buddy 5100c9de560dSAlex Tomas * mark B in on-disk bitmap 5101c9de560dSAlex Tomas * drop PA from group 5102c9de560dSAlex Tomas * mark all PAs in buddy 5103c9de560dSAlex Tomas * 5104c9de560dSAlex Tomas * thus, P1 initializes buddy with B available. to prevent this 5105c9de560dSAlex Tomas * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 5106c9de560dSAlex Tomas * against that pair 5107c9de560dSAlex Tomas */ 5108c9de560dSAlex Tomas ext4_lock_group(sb, grp); 5109c9de560dSAlex Tomas list_del(&pa->pa_group_list); 5110c9de560dSAlex Tomas ext4_unlock_group(sb, grp); 5111c9de560dSAlex Tomas 5112a8e38fd3SOjaswin Mujoo if (pa->pa_type == MB_INODE_PA) { 511338727786SOjaswin Mujoo write_lock(pa->pa_node_lock.inode_lock); 511438727786SOjaswin Mujoo rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 511538727786SOjaswin Mujoo write_unlock(pa->pa_node_lock.inode_lock); 511638727786SOjaswin Mujoo ext4_mb_pa_free(pa); 5117a8e38fd3SOjaswin Mujoo } else { 5118a8e38fd3SOjaswin Mujoo spin_lock(pa->pa_node_lock.lg_lock); 5119a8e38fd3SOjaswin Mujoo list_del_rcu(&pa->pa_node.lg_list); 5120a8e38fd3SOjaswin Mujoo spin_unlock(pa->pa_node_lock.lg_lock); 512138727786SOjaswin Mujoo call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 512238727786SOjaswin Mujoo } 5123a8e38fd3SOjaswin Mujoo } 5124c9de560dSAlex Tomas 512538727786SOjaswin Mujoo static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new) 512638727786SOjaswin Mujoo { 512738727786SOjaswin Mujoo struct rb_node **iter = &root->rb_node, *parent = NULL; 512838727786SOjaswin Mujoo struct ext4_prealloc_space *iter_pa, *new_pa; 512938727786SOjaswin Mujoo ext4_lblk_t iter_start, new_start; 513038727786SOjaswin Mujoo 513138727786SOjaswin Mujoo while (*iter) { 513238727786SOjaswin Mujoo iter_pa = rb_entry(*iter, struct ext4_prealloc_space, 513338727786SOjaswin Mujoo pa_node.inode_node); 513438727786SOjaswin Mujoo new_pa = rb_entry(new, struct ext4_prealloc_space, 513538727786SOjaswin Mujoo pa_node.inode_node); 513638727786SOjaswin Mujoo iter_start = iter_pa->pa_lstart; 513738727786SOjaswin Mujoo new_start = new_pa->pa_lstart; 513838727786SOjaswin Mujoo 513938727786SOjaswin Mujoo parent = *iter; 514038727786SOjaswin Mujoo if (new_start < iter_start) 514138727786SOjaswin Mujoo iter = &((*iter)->rb_left); 514238727786SOjaswin Mujoo else 514338727786SOjaswin Mujoo iter = &((*iter)->rb_right); 514438727786SOjaswin Mujoo } 514538727786SOjaswin Mujoo 514638727786SOjaswin Mujoo rb_link_node(new, parent, iter); 514738727786SOjaswin Mujoo rb_insert_color(new, root); 5148c9de560dSAlex Tomas } 5149c9de560dSAlex Tomas 5150c9de560dSAlex Tomas /* 5151c9de560dSAlex Tomas * creates new preallocated space for given inode 5152c9de560dSAlex Tomas */ 515353f86b17SRitesh Harjani static noinline_for_stack void 51544ddfef7bSEric Sandeen ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 5155c9de560dSAlex Tomas { 5156c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 515753accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(sb); 5158c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 5159c9de560dSAlex Tomas struct ext4_group_info *grp; 5160c9de560dSAlex Tomas struct ext4_inode_info *ei; 5161c9de560dSAlex Tomas 5162c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 5163c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 5164c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 5165c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 516653f86b17SRitesh Harjani BUG_ON(ac->ac_pa == NULL); 5167c9de560dSAlex Tomas 516853f86b17SRitesh Harjani pa = ac->ac_pa; 5169c9de560dSAlex Tomas 51707e170922SOjaswin Mujoo if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) { 5171bc056e71SBaokun Li struct ext4_free_extent ex = { 5172bc056e71SBaokun Li .fe_logical = ac->ac_g_ex.fe_logical, 5173bc056e71SBaokun Li .fe_len = ac->ac_orig_goal_len, 5174bc056e71SBaokun Li }; 5175bc056e71SBaokun Li loff_t orig_goal_end = extent_logical_end(sbi, &ex); 5176c9de560dSAlex Tomas 5177c9de560dSAlex Tomas /* we can't allocate as much as normalizer wants. 5178c9de560dSAlex Tomas * so, found space must get proper lstart 5179c9de560dSAlex Tomas * to cover original request */ 5180c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 5181c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 5182c9de560dSAlex Tomas 518393cdf49fSOjaswin Mujoo /* 518493cdf49fSOjaswin Mujoo * Use the below logic for adjusting best extent as it keeps 518593cdf49fSOjaswin Mujoo * fragmentation in check while ensuring logical range of best 518693cdf49fSOjaswin Mujoo * extent doesn't overflow out of goal extent: 518793cdf49fSOjaswin Mujoo * 51887e170922SOjaswin Mujoo * 1. Check if best ex can be kept at end of goal (before 51897e170922SOjaswin Mujoo * cr_best_avail trimmed it) and still cover original start 519093cdf49fSOjaswin Mujoo * 2. Else, check if best ex can be kept at start of goal and 519193cdf49fSOjaswin Mujoo * still cover original start 519293cdf49fSOjaswin Mujoo * 3. Else, keep the best ex at start of original request. 519393cdf49fSOjaswin Mujoo */ 5194bc056e71SBaokun Li ex.fe_len = ac->ac_b_ex.fe_len; 5195bc056e71SBaokun Li 5196bc056e71SBaokun Li ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len); 5197bc056e71SBaokun Li if (ac->ac_o_ex.fe_logical >= ex.fe_logical) 519893cdf49fSOjaswin Mujoo goto adjust_bex; 5199c9de560dSAlex Tomas 5200bc056e71SBaokun Li ex.fe_logical = ac->ac_g_ex.fe_logical; 5201bc056e71SBaokun Li if (ac->ac_o_ex.fe_logical < extent_logical_end(sbi, &ex)) 520293cdf49fSOjaswin Mujoo goto adjust_bex; 5203c9de560dSAlex Tomas 5204bc056e71SBaokun Li ex.fe_logical = ac->ac_o_ex.fe_logical; 520593cdf49fSOjaswin Mujoo adjust_bex: 5206bc056e71SBaokun Li ac->ac_b_ex.fe_logical = ex.fe_logical; 5207c9de560dSAlex Tomas 5208c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 5209c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 5210bc056e71SBaokun Li BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end); 5211c9de560dSAlex Tomas } 5212c9de560dSAlex Tomas 5213c9de560dSAlex Tomas pa->pa_lstart = ac->ac_b_ex.fe_logical; 5214c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 5215c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 5216c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 5217c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 5218d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 5219c9de560dSAlex Tomas pa->pa_deleted = 0; 5220cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_INODE_PA; 5221c9de560dSAlex Tomas 5222d3df1453SRitesh Harjani mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, 5223d3df1453SRitesh Harjani pa->pa_len, pa->pa_lstart); 52249bffad1eSTheodore Ts'o trace_ext4_mb_new_inode_pa(ac, pa); 5225c9de560dSAlex Tomas 522653accfa9STheodore Ts'o atomic_add(pa->pa_free, &sbi->s_mb_preallocated); 5227abc075d4SKemeng Shi ext4_mb_use_inode_pa(ac, pa); 5228c9de560dSAlex Tomas 5229c9de560dSAlex Tomas ei = EXT4_I(ac->ac_inode); 5230c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 52315354b2afSTheodore Ts'o if (!grp) 52325354b2afSTheodore Ts'o return; 5233c9de560dSAlex Tomas 5234a8e38fd3SOjaswin Mujoo pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock; 5235c9de560dSAlex Tomas pa->pa_inode = ac->ac_inode; 5236c9de560dSAlex Tomas 5237c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 5238c9de560dSAlex Tomas 523938727786SOjaswin Mujoo write_lock(pa->pa_node_lock.inode_lock); 524038727786SOjaswin Mujoo ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node); 524138727786SOjaswin Mujoo write_unlock(pa->pa_node_lock.inode_lock); 524227bc446eSbrookxu atomic_inc(&ei->i_prealloc_active); 5243c9de560dSAlex Tomas } 5244c9de560dSAlex Tomas 5245c9de560dSAlex Tomas /* 5246c9de560dSAlex Tomas * creates new preallocated space for locality group inodes belongs to 5247c9de560dSAlex Tomas */ 524853f86b17SRitesh Harjani static noinline_for_stack void 52494ddfef7bSEric Sandeen ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 5250c9de560dSAlex Tomas { 5251c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 5252c9de560dSAlex Tomas struct ext4_locality_group *lg; 5253c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 5254c9de560dSAlex Tomas struct ext4_group_info *grp; 5255c9de560dSAlex Tomas 5256c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 5257c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 5258c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 5259c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 526053f86b17SRitesh Harjani BUG_ON(ac->ac_pa == NULL); 5261c9de560dSAlex Tomas 526253f86b17SRitesh Harjani pa = ac->ac_pa; 5263c9de560dSAlex Tomas 5264c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 5265c9de560dSAlex Tomas pa->pa_lstart = pa->pa_pstart; 5266c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 5267c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 5268c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 5269a8e38fd3SOjaswin Mujoo INIT_LIST_HEAD(&pa->pa_node.lg_list); 5270d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 5271c9de560dSAlex Tomas pa->pa_deleted = 0; 5272cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_GROUP_PA; 5273c9de560dSAlex Tomas 5274d3df1453SRitesh Harjani mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, 5275d3df1453SRitesh Harjani pa->pa_len, pa->pa_lstart); 52769bffad1eSTheodore Ts'o trace_ext4_mb_new_group_pa(ac, pa); 5277c9de560dSAlex Tomas 5278c9de560dSAlex Tomas ext4_mb_use_group_pa(ac, pa); 5279c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 5280c9de560dSAlex Tomas 5281c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 52825354b2afSTheodore Ts'o if (!grp) 52835354b2afSTheodore Ts'o return; 5284c9de560dSAlex Tomas lg = ac->ac_lg; 5285c9de560dSAlex Tomas BUG_ON(lg == NULL); 5286c9de560dSAlex Tomas 5287a8e38fd3SOjaswin Mujoo pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock; 5288c9de560dSAlex Tomas pa->pa_inode = NULL; 5289c9de560dSAlex Tomas 5290c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 5291c9de560dSAlex Tomas 52926be2ded1SAneesh Kumar K.V /* 52936be2ded1SAneesh Kumar K.V * We will later add the new pa to the right bucket 52946be2ded1SAneesh Kumar K.V * after updating the pa_free in ext4_mb_release_context 52956be2ded1SAneesh Kumar K.V */ 5296c9de560dSAlex Tomas } 5297c9de560dSAlex Tomas 529853f86b17SRitesh Harjani static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 5299c9de560dSAlex Tomas { 5300c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 530153f86b17SRitesh Harjani ext4_mb_new_group_pa(ac); 5302c9de560dSAlex Tomas else 530353f86b17SRitesh Harjani ext4_mb_new_inode_pa(ac); 5304c9de560dSAlex Tomas } 5305c9de560dSAlex Tomas 5306c9de560dSAlex Tomas /* 5307c9de560dSAlex Tomas * finds all unused blocks in on-disk bitmap, frees them in 5308c9de560dSAlex Tomas * in-core bitmap and buddy. 5309c9de560dSAlex Tomas * @pa must be unlinked from inode and group lists, so that 5310c9de560dSAlex Tomas * nobody else can find/use it. 5311c9de560dSAlex Tomas * the caller MUST hold group/inode locks. 5312c9de560dSAlex Tomas * TODO: optimize the case when there are no in-core structures yet 5313c9de560dSAlex Tomas */ 53144ddfef7bSEric Sandeen static noinline_for_stack int 53154ddfef7bSEric Sandeen ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 53163e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 5317c9de560dSAlex Tomas { 5318c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 5319c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 5320498e5f24STheodore Ts'o unsigned int end; 5321498e5f24STheodore Ts'o unsigned int next; 5322c9de560dSAlex Tomas ext4_group_t group; 5323c9de560dSAlex Tomas ext4_grpblk_t bit; 5324ba80b101STheodore Ts'o unsigned long long grp_blk_start; 5325c9de560dSAlex Tomas int free = 0; 5326c9de560dSAlex Tomas 5327c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 5328c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 532953accfa9STheodore Ts'o grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); 5330c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 5331c9de560dSAlex Tomas end = bit + pa->pa_len; 5332c9de560dSAlex Tomas 5333c9de560dSAlex Tomas while (bit < end) { 5334ffad0a44SAneesh Kumar K.V bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 5335c9de560dSAlex Tomas if (bit >= end) 5336c9de560dSAlex Tomas break; 5337ffad0a44SAneesh Kumar K.V next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 5338d3df1453SRitesh Harjani mb_debug(sb, "free preallocated %u/%u in group %u\n", 53395a0790c2SAndi Kleen (unsigned) ext4_group_first_block_no(sb, group) + bit, 53405a0790c2SAndi Kleen (unsigned) next - bit, (unsigned) group); 5341c9de560dSAlex Tomas free += next - bit; 5342c9de560dSAlex Tomas 53433e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); 534453accfa9STheodore Ts'o trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + 534553accfa9STheodore Ts'o EXT4_C2B(sbi, bit)), 5346a9c667f8SLukas Czerner next - bit); 5347c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 5348c9de560dSAlex Tomas bit = next + 1; 5349c9de560dSAlex Tomas } 5350c9de560dSAlex Tomas if (free != pa->pa_free) { 53519d8b9ec4STheodore Ts'o ext4_msg(e4b->bd_sb, KERN_CRIT, 535236bad423SRitesh Harjani "pa %p: logic %lu, phys. %lu, len %d", 5353c9de560dSAlex Tomas pa, (unsigned long) pa->pa_lstart, 5354c9de560dSAlex Tomas (unsigned long) pa->pa_pstart, 535536bad423SRitesh Harjani pa->pa_len); 5356e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", 535726346ff6SAneesh Kumar K.V free, pa->pa_free); 5358e56eb659SAneesh Kumar K.V /* 5359e56eb659SAneesh Kumar K.V * pa is already deleted so we use the value obtained 5360e56eb659SAneesh Kumar K.V * from the bitmap and continue. 5361e56eb659SAneesh Kumar K.V */ 5362c9de560dSAlex Tomas } 5363c9de560dSAlex Tomas atomic_add(free, &sbi->s_mb_discarded); 5364c9de560dSAlex Tomas 5365863c37fcSzhong jiang return 0; 5366c9de560dSAlex Tomas } 5367c9de560dSAlex Tomas 53684ddfef7bSEric Sandeen static noinline_for_stack int 53694ddfef7bSEric Sandeen ext4_mb_release_group_pa(struct ext4_buddy *e4b, 53703e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 5371c9de560dSAlex Tomas { 5372c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 5373c9de560dSAlex Tomas ext4_group_t group; 5374c9de560dSAlex Tomas ext4_grpblk_t bit; 5375c9de560dSAlex Tomas 537660e07cf5SYongqiang Yang trace_ext4_mb_release_group_pa(sb, pa); 5377c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 5378c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 5379463808f2STheodore Ts'o if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) { 5380463808f2STheodore Ts'o ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu", 5381463808f2STheodore Ts'o e4b->bd_group, group, pa->pa_pstart); 5382463808f2STheodore Ts'o return 0; 5383463808f2STheodore Ts'o } 5384c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 5385c9de560dSAlex Tomas atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 53863e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); 5387c9de560dSAlex Tomas 5388c9de560dSAlex Tomas return 0; 5389c9de560dSAlex Tomas } 5390c9de560dSAlex Tomas 5391c9de560dSAlex Tomas /* 5392c9de560dSAlex Tomas * releases all preallocations in given group 5393c9de560dSAlex Tomas * 5394c9de560dSAlex Tomas * first, we need to decide discard policy: 5395c9de560dSAlex Tomas * - when do we discard 5396c9de560dSAlex Tomas * 1) ENOSPC 5397c9de560dSAlex Tomas * - how many do we discard 5398c9de560dSAlex Tomas * 1) how many requested 5399c9de560dSAlex Tomas */ 54004ddfef7bSEric Sandeen static noinline_for_stack int 54014ddfef7bSEric Sandeen ext4_mb_discard_group_preallocations(struct super_block *sb, 54028c80fb31SChunguang Xu ext4_group_t group, int *busy) 5403c9de560dSAlex Tomas { 5404c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 5405c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 5406c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 54070f6bc579SRuan Jinjie LIST_HEAD(list); 5408c9de560dSAlex Tomas struct ext4_buddy e4b; 540938727786SOjaswin Mujoo struct ext4_inode_info *ei; 5410c9de560dSAlex Tomas int err; 54118c80fb31SChunguang Xu int free = 0; 5412c9de560dSAlex Tomas 54135354b2afSTheodore Ts'o if (!grp) 54145354b2afSTheodore Ts'o return 0; 5415d3df1453SRitesh Harjani mb_debug(sb, "discard preallocation for group %u\n", group); 5416c9de560dSAlex Tomas if (list_empty(&grp->bb_prealloc_list)) 5417bbc4ec77SRitesh Harjani goto out_dbg; 5418c9de560dSAlex Tomas 5419574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 54209008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 54219008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 542254d3adbcSTheodore Ts'o ext4_error_err(sb, -err, 542354d3adbcSTheodore Ts'o "Error %d reading block bitmap for %u", 54249008a58eSDarrick J. Wong err, group); 5425bbc4ec77SRitesh Harjani goto out_dbg; 5426c9de560dSAlex Tomas } 5427c9de560dSAlex Tomas 5428c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 5429ce89f46cSAneesh Kumar K.V if (err) { 54309651e6b2SKonstantin Khlebnikov ext4_warning(sb, "Error %d loading buddy information for %u", 54319651e6b2SKonstantin Khlebnikov err, group); 5432ce89f46cSAneesh Kumar K.V put_bh(bitmap_bh); 5433bbc4ec77SRitesh Harjani goto out_dbg; 5434ce89f46cSAneesh Kumar K.V } 5435c9de560dSAlex Tomas 5436c9de560dSAlex Tomas ext4_lock_group(sb, group); 5437c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, 5438c9de560dSAlex Tomas &grp->bb_prealloc_list, pa_group_list) { 5439c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 5440c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 5441c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 54428c80fb31SChunguang Xu *busy = 1; 5443c9de560dSAlex Tomas continue; 5444c9de560dSAlex Tomas } 5445c9de560dSAlex Tomas if (pa->pa_deleted) { 5446c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5447c9de560dSAlex Tomas continue; 5448c9de560dSAlex Tomas } 5449c9de560dSAlex Tomas 5450c9de560dSAlex Tomas /* seems this one can be freed ... */ 545127bc446eSbrookxu ext4_mb_mark_pa_deleted(sb, pa); 5452c9de560dSAlex Tomas 545370022da8SYe Bin if (!free) 545470022da8SYe Bin this_cpu_inc(discard_pa_seq); 545570022da8SYe Bin 5456c9de560dSAlex Tomas /* we can trust pa_free ... */ 5457c9de560dSAlex Tomas free += pa->pa_free; 5458c9de560dSAlex Tomas 5459c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5460c9de560dSAlex Tomas 5461c9de560dSAlex Tomas list_del(&pa->pa_group_list); 5462c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 5463c9de560dSAlex Tomas } 5464c9de560dSAlex Tomas 5465c9de560dSAlex Tomas /* now free all selected PAs */ 5466c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 5467c9de560dSAlex Tomas 5468c9de560dSAlex Tomas /* remove from object (inode or locality group) */ 5469a8e38fd3SOjaswin Mujoo if (pa->pa_type == MB_GROUP_PA) { 5470a8e38fd3SOjaswin Mujoo spin_lock(pa->pa_node_lock.lg_lock); 5471a8e38fd3SOjaswin Mujoo list_del_rcu(&pa->pa_node.lg_list); 5472a8e38fd3SOjaswin Mujoo spin_unlock(pa->pa_node_lock.lg_lock); 5473a8e38fd3SOjaswin Mujoo } else { 547438727786SOjaswin Mujoo write_lock(pa->pa_node_lock.inode_lock); 547538727786SOjaswin Mujoo ei = EXT4_I(pa->pa_inode); 547638727786SOjaswin Mujoo rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 547738727786SOjaswin Mujoo write_unlock(pa->pa_node_lock.inode_lock); 5478a8e38fd3SOjaswin Mujoo } 5479c9de560dSAlex Tomas 5480c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 548138727786SOjaswin Mujoo 548238727786SOjaswin Mujoo if (pa->pa_type == MB_GROUP_PA) { 548338727786SOjaswin Mujoo ext4_mb_release_group_pa(&e4b, pa); 5484c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 548538727786SOjaswin Mujoo } else { 548638727786SOjaswin Mujoo ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 548738727786SOjaswin Mujoo ext4_mb_pa_free(pa); 548838727786SOjaswin Mujoo } 5489c9de560dSAlex Tomas } 5490c9de560dSAlex Tomas 5491c9de560dSAlex Tomas ext4_unlock_group(sb, group); 5492e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 5493c9de560dSAlex Tomas put_bh(bitmap_bh); 5494bbc4ec77SRitesh Harjani out_dbg: 5495d3df1453SRitesh Harjani mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", 54968c80fb31SChunguang Xu free, group, grp->bb_free); 54978c80fb31SChunguang Xu return free; 5498c9de560dSAlex Tomas } 5499c9de560dSAlex Tomas 5500c9de560dSAlex Tomas /* 5501c9de560dSAlex Tomas * releases all non-used preallocated blocks for given inode 5502c9de560dSAlex Tomas * 5503c9de560dSAlex Tomas * It's important to discard preallocations under i_data_sem 5504c9de560dSAlex Tomas * We don't want another block to be served from the prealloc 5505c9de560dSAlex Tomas * space when we are discarding the inode prealloc space. 5506c9de560dSAlex Tomas * 5507c9de560dSAlex Tomas * FIXME!! Make sure it is valid at all the call sites 5508c9de560dSAlex Tomas */ 550927bc446eSbrookxu void ext4_discard_preallocations(struct inode *inode, unsigned int needed) 5510c9de560dSAlex Tomas { 5511c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(inode); 5512c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 5513c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 5514c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 5515c9de560dSAlex Tomas ext4_group_t group = 0; 55160f6bc579SRuan Jinjie LIST_HEAD(list); 5517c9de560dSAlex Tomas struct ext4_buddy e4b; 551838727786SOjaswin Mujoo struct rb_node *iter; 5519c9de560dSAlex Tomas int err; 5520c9de560dSAlex Tomas 5521c2ea3fdeSTheodore Ts'o if (!S_ISREG(inode->i_mode)) { 5522c9de560dSAlex Tomas return; 5523c9de560dSAlex Tomas } 5524c9de560dSAlex Tomas 55258016e29fSHarshad Shirwadkar if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) 55268016e29fSHarshad Shirwadkar return; 55278016e29fSHarshad Shirwadkar 5528d3df1453SRitesh Harjani mb_debug(sb, "discard preallocation for inode %lu\n", 5529d3df1453SRitesh Harjani inode->i_ino); 553027bc446eSbrookxu trace_ext4_discard_preallocations(inode, 553127bc446eSbrookxu atomic_read(&ei->i_prealloc_active), needed); 5532c9de560dSAlex Tomas 553327bc446eSbrookxu if (needed == 0) 553427bc446eSbrookxu needed = UINT_MAX; 553527bc446eSbrookxu 5536c9de560dSAlex Tomas repeat: 5537c9de560dSAlex Tomas /* first, collect all pa's in the inode */ 553838727786SOjaswin Mujoo write_lock(&ei->i_prealloc_lock); 553938727786SOjaswin Mujoo for (iter = rb_first(&ei->i_prealloc_node); iter && needed; 554038727786SOjaswin Mujoo iter = rb_next(iter)) { 554138727786SOjaswin Mujoo pa = rb_entry(iter, struct ext4_prealloc_space, 554238727786SOjaswin Mujoo pa_node.inode_node); 5543a8e38fd3SOjaswin Mujoo BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock); 554438727786SOjaswin Mujoo 5545c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 5546c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 5547c9de560dSAlex Tomas /* this shouldn't happen often - nobody should 5548c9de560dSAlex Tomas * use preallocation while we're discarding it */ 5549c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 555038727786SOjaswin Mujoo write_unlock(&ei->i_prealloc_lock); 55519d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, 55529d8b9ec4STheodore Ts'o "uh-oh! used pa while discarding"); 5553c9de560dSAlex Tomas WARN_ON(1); 5554c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 5555c9de560dSAlex Tomas goto repeat; 5556c9de560dSAlex Tomas 5557c9de560dSAlex Tomas } 5558c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 555927bc446eSbrookxu ext4_mb_mark_pa_deleted(sb, pa); 5560c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 556138727786SOjaswin Mujoo rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 5562c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 556327bc446eSbrookxu needed--; 5564c9de560dSAlex Tomas continue; 5565c9de560dSAlex Tomas } 5566c9de560dSAlex Tomas 5567c9de560dSAlex Tomas /* someone is deleting pa right now */ 5568c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 556938727786SOjaswin Mujoo write_unlock(&ei->i_prealloc_lock); 5570c9de560dSAlex Tomas 5571c9de560dSAlex Tomas /* we have to wait here because pa_deleted 5572c9de560dSAlex Tomas * doesn't mean pa is already unlinked from 5573c9de560dSAlex Tomas * the list. as we might be called from 5574c9de560dSAlex Tomas * ->clear_inode() the inode will get freed 5575c9de560dSAlex Tomas * and concurrent thread which is unlinking 5576c9de560dSAlex Tomas * pa from inode's list may access already 5577c9de560dSAlex Tomas * freed memory, bad-bad-bad */ 5578c9de560dSAlex Tomas 5579c9de560dSAlex Tomas /* XXX: if this happens too often, we can 5580c9de560dSAlex Tomas * add a flag to force wait only in case 5581c9de560dSAlex Tomas * of ->clear_inode(), but not in case of 5582c9de560dSAlex Tomas * regular truncate */ 5583c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 5584c9de560dSAlex Tomas goto repeat; 5585c9de560dSAlex Tomas } 558638727786SOjaswin Mujoo write_unlock(&ei->i_prealloc_lock); 5587c9de560dSAlex Tomas 5588c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 5589cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_INODE_PA); 5590bd86298eSLukas Czerner group = ext4_get_group_number(sb, pa->pa_pstart); 5591c9de560dSAlex Tomas 55929651e6b2SKonstantin Khlebnikov err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 55939651e6b2SKonstantin Khlebnikov GFP_NOFS|__GFP_NOFAIL); 5594ce89f46cSAneesh Kumar K.V if (err) { 559554d3adbcSTheodore Ts'o ext4_error_err(sb, -err, "Error %d loading buddy information for %u", 55969651e6b2SKonstantin Khlebnikov err, group); 5597ce89f46cSAneesh Kumar K.V continue; 5598ce89f46cSAneesh Kumar K.V } 5599c9de560dSAlex Tomas 5600574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 56019008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 56029008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 560354d3adbcSTheodore Ts'o ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", 56049008a58eSDarrick J. Wong err, group); 5605e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 5606ce89f46cSAneesh Kumar K.V continue; 5607c9de560dSAlex Tomas } 5608c9de560dSAlex Tomas 5609c9de560dSAlex Tomas ext4_lock_group(sb, group); 5610c9de560dSAlex Tomas list_del(&pa->pa_group_list); 56113e1e5f50SEric Sandeen ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 5612c9de560dSAlex Tomas ext4_unlock_group(sb, group); 5613c9de560dSAlex Tomas 5614e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 5615c9de560dSAlex Tomas put_bh(bitmap_bh); 5616c9de560dSAlex Tomas 5617c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 561838727786SOjaswin Mujoo ext4_mb_pa_free(pa); 5619c9de560dSAlex Tomas } 5620c9de560dSAlex Tomas } 5621c9de560dSAlex Tomas 562253f86b17SRitesh Harjani static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) 562353f86b17SRitesh Harjani { 562453f86b17SRitesh Harjani struct ext4_prealloc_space *pa; 562553f86b17SRitesh Harjani 562653f86b17SRitesh Harjani BUG_ON(ext4_pspace_cachep == NULL); 562753f86b17SRitesh Harjani pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); 562853f86b17SRitesh Harjani if (!pa) 562953f86b17SRitesh Harjani return -ENOMEM; 563053f86b17SRitesh Harjani atomic_set(&pa->pa_count, 1); 563153f86b17SRitesh Harjani ac->ac_pa = pa; 563253f86b17SRitesh Harjani return 0; 563353f86b17SRitesh Harjani } 563453f86b17SRitesh Harjani 563582089725SOjaswin Mujoo static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac) 563653f86b17SRitesh Harjani { 563753f86b17SRitesh Harjani struct ext4_prealloc_space *pa = ac->ac_pa; 563853f86b17SRitesh Harjani 563953f86b17SRitesh Harjani BUG_ON(!pa); 564053f86b17SRitesh Harjani ac->ac_pa = NULL; 564153f86b17SRitesh Harjani WARN_ON(!atomic_dec_and_test(&pa->pa_count)); 564282089725SOjaswin Mujoo /* 564382089725SOjaswin Mujoo * current function is only called due to an error or due to 564482089725SOjaswin Mujoo * len of found blocks < len of requested blocks hence the PA has not 564582089725SOjaswin Mujoo * been added to grp->bb_prealloc_list. So we don't need to lock it 564682089725SOjaswin Mujoo */ 564782089725SOjaswin Mujoo pa->pa_deleted = 1; 564882089725SOjaswin Mujoo ext4_mb_pa_free(pa); 564953f86b17SRitesh Harjani } 565053f86b17SRitesh Harjani 56516ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 5652e68cf40cSRitesh Harjani static inline void ext4_mb_show_pa(struct super_block *sb) 5653c9de560dSAlex Tomas { 5654e68cf40cSRitesh Harjani ext4_group_t i, ngroups; 5655c9de560dSAlex Tomas 565695257987SJan Kara if (ext4_forced_shutdown(sb)) 5657e3570639SEric Sandeen return; 5658e3570639SEric Sandeen 56598df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 5660d3df1453SRitesh Harjani mb_debug(sb, "groups: "); 56618df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 5662c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, i); 5663c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 5664c9de560dSAlex Tomas ext4_grpblk_t start; 5665c9de560dSAlex Tomas struct list_head *cur; 56665354b2afSTheodore Ts'o 56675354b2afSTheodore Ts'o if (!grp) 56685354b2afSTheodore Ts'o continue; 5669c9de560dSAlex Tomas ext4_lock_group(sb, i); 5670c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 5671c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, 5672c9de560dSAlex Tomas pa_group_list); 5673c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 5674c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 5675c9de560dSAlex Tomas NULL, &start); 5676c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5677d3df1453SRitesh Harjani mb_debug(sb, "PA:%u:%d:%d\n", i, start, 5678d3df1453SRitesh Harjani pa->pa_len); 5679c9de560dSAlex Tomas } 568060bd63d1SSolofo Ramangalahy ext4_unlock_group(sb, i); 5681d3df1453SRitesh Harjani mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, 5682d3df1453SRitesh Harjani grp->bb_fragments); 5683c9de560dSAlex Tomas } 5684c9de560dSAlex Tomas } 5685e68cf40cSRitesh Harjani 5686e68cf40cSRitesh Harjani static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 5687e68cf40cSRitesh Harjani { 5688e68cf40cSRitesh Harjani struct super_block *sb = ac->ac_sb; 5689e68cf40cSRitesh Harjani 569095257987SJan Kara if (ext4_forced_shutdown(sb)) 5691e68cf40cSRitesh Harjani return; 5692e68cf40cSRitesh Harjani 5693d3df1453SRitesh Harjani mb_debug(sb, "Can't allocate:" 5694e68cf40cSRitesh Harjani " Allocation context details:"); 5695d3df1453SRitesh Harjani mb_debug(sb, "status %u flags 0x%x", 5696e68cf40cSRitesh Harjani ac->ac_status, ac->ac_flags); 5697d3df1453SRitesh Harjani mb_debug(sb, "orig %lu/%lu/%lu@%lu, " 5698e68cf40cSRitesh Harjani "goal %lu/%lu/%lu@%lu, " 5699e68cf40cSRitesh Harjani "best %lu/%lu/%lu@%lu cr %d", 5700e68cf40cSRitesh Harjani (unsigned long)ac->ac_o_ex.fe_group, 5701e68cf40cSRitesh Harjani (unsigned long)ac->ac_o_ex.fe_start, 5702e68cf40cSRitesh Harjani (unsigned long)ac->ac_o_ex.fe_len, 5703e68cf40cSRitesh Harjani (unsigned long)ac->ac_o_ex.fe_logical, 5704e68cf40cSRitesh Harjani (unsigned long)ac->ac_g_ex.fe_group, 5705e68cf40cSRitesh Harjani (unsigned long)ac->ac_g_ex.fe_start, 5706e68cf40cSRitesh Harjani (unsigned long)ac->ac_g_ex.fe_len, 5707e68cf40cSRitesh Harjani (unsigned long)ac->ac_g_ex.fe_logical, 5708e68cf40cSRitesh Harjani (unsigned long)ac->ac_b_ex.fe_group, 5709e68cf40cSRitesh Harjani (unsigned long)ac->ac_b_ex.fe_start, 5710e68cf40cSRitesh Harjani (unsigned long)ac->ac_b_ex.fe_len, 5711e68cf40cSRitesh Harjani (unsigned long)ac->ac_b_ex.fe_logical, 5712e68cf40cSRitesh Harjani (int)ac->ac_criteria); 5713d3df1453SRitesh Harjani mb_debug(sb, "%u found", ac->ac_found); 5714569f196fSRitesh Harjani mb_debug(sb, "used pa: %s, ", ac->ac_pa ? "yes" : "no"); 5715569f196fSRitesh Harjani if (ac->ac_pa) 5716569f196fSRitesh Harjani mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ? 5717569f196fSRitesh Harjani "group pa" : "inode pa"); 5718e68cf40cSRitesh Harjani ext4_mb_show_pa(sb); 5719e68cf40cSRitesh Harjani } 5720c9de560dSAlex Tomas #else 5721e68cf40cSRitesh Harjani static inline void ext4_mb_show_pa(struct super_block *sb) 5722e68cf40cSRitesh Harjani { 5723e68cf40cSRitesh Harjani } 5724c9de560dSAlex Tomas static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 5725c9de560dSAlex Tomas { 5726e68cf40cSRitesh Harjani ext4_mb_show_pa(ac->ac_sb); 5727c9de560dSAlex Tomas } 5728c9de560dSAlex Tomas #endif 5729c9de560dSAlex Tomas 5730c9de560dSAlex Tomas /* 5731c9de560dSAlex Tomas * We use locality group preallocation for small size file. The size of the 5732c9de560dSAlex Tomas * file is determined by the current size or the resulting size after 5733c9de560dSAlex Tomas * allocation which ever is larger 5734c9de560dSAlex Tomas * 5735b713a5ecSTheodore Ts'o * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 5736c9de560dSAlex Tomas */ 5737c9de560dSAlex Tomas static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 5738c9de560dSAlex Tomas { 5739c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 5740c9de560dSAlex Tomas int bsbits = ac->ac_sb->s_blocksize_bits; 5741c9de560dSAlex Tomas loff_t size, isize; 5742a9f2a293SJan Kara bool inode_pa_eligible, group_pa_eligible; 5743c9de560dSAlex Tomas 5744c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 5745c9de560dSAlex Tomas return; 5746c9de560dSAlex Tomas 57474ba74d00STheodore Ts'o if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 57484ba74d00STheodore Ts'o return; 57494ba74d00STheodore Ts'o 5750a9f2a293SJan Kara group_pa_eligible = sbi->s_mb_group_prealloc > 0; 5751a9f2a293SJan Kara inode_pa_eligible = true; 575243bbddc0SBaokun Li size = extent_logical_end(sbi, &ac->ac_o_ex); 575350797481STheodore Ts'o isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 575450797481STheodore Ts'o >> bsbits; 5755c9de560dSAlex Tomas 5756a9f2a293SJan Kara /* No point in using inode preallocation for closed files */ 575782dd124cSNikolay Borisov if ((size == isize) && !ext4_fs_is_busy(sbi) && 5758a9f2a293SJan Kara !inode_is_open_for_write(ac->ac_inode)) 5759a9f2a293SJan Kara inode_pa_eligible = false; 576050797481STheodore Ts'o 576171780577STheodore Ts'o size = max(size, isize); 5762a9f2a293SJan Kara /* Don't use group allocation for large files */ 5763a9f2a293SJan Kara if (size > sbi->s_mb_stream_request) 5764a9f2a293SJan Kara group_pa_eligible = false; 5765a9f2a293SJan Kara 5766a9f2a293SJan Kara if (!group_pa_eligible) { 5767a9f2a293SJan Kara if (inode_pa_eligible) 57684ba74d00STheodore Ts'o ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 5769a9f2a293SJan Kara else 5770a9f2a293SJan Kara ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 5771c9de560dSAlex Tomas return; 57724ba74d00STheodore Ts'o } 5773c9de560dSAlex Tomas 5774c9de560dSAlex Tomas BUG_ON(ac->ac_lg != NULL); 5775c9de560dSAlex Tomas /* 5776c9de560dSAlex Tomas * locality group prealloc space are per cpu. The reason for having 5777c9de560dSAlex Tomas * per cpu locality group is to reduce the contention between block 5778c9de560dSAlex Tomas * request from multiple CPUs. 5779c9de560dSAlex Tomas */ 5780a0b6bc63SChristoph Lameter ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); 5781c9de560dSAlex Tomas 5782c9de560dSAlex Tomas /* we're going to use group allocation */ 5783c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 5784c9de560dSAlex Tomas 5785c9de560dSAlex Tomas /* serialize all allocations in the group */ 5786c9de560dSAlex Tomas mutex_lock(&ac->ac_lg->lg_mutex); 5787c9de560dSAlex Tomas } 5788c9de560dSAlex Tomas 5789d73eff68SGuoqing Jiang static noinline_for_stack void 57904ddfef7bSEric Sandeen ext4_mb_initialize_context(struct ext4_allocation_context *ac, 5791c9de560dSAlex Tomas struct ext4_allocation_request *ar) 5792c9de560dSAlex Tomas { 5793c9de560dSAlex Tomas struct super_block *sb = ar->inode->i_sb; 5794c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 5795c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 5796c9de560dSAlex Tomas ext4_group_t group; 5797498e5f24STheodore Ts'o unsigned int len; 5798498e5f24STheodore Ts'o ext4_fsblk_t goal; 5799c9de560dSAlex Tomas ext4_grpblk_t block; 5800c9de560dSAlex Tomas 5801c9de560dSAlex Tomas /* we can't allocate > group size */ 5802c9de560dSAlex Tomas len = ar->len; 5803c9de560dSAlex Tomas 5804c9de560dSAlex Tomas /* just a dirty hack to filter too big requests */ 580540ae3487STheodore Ts'o if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) 580640ae3487STheodore Ts'o len = EXT4_CLUSTERS_PER_GROUP(sb); 5807c9de560dSAlex Tomas 5808c9de560dSAlex Tomas /* start searching from the goal */ 5809c9de560dSAlex Tomas goal = ar->goal; 5810c9de560dSAlex Tomas if (goal < le32_to_cpu(es->s_first_data_block) || 5811c9de560dSAlex Tomas goal >= ext4_blocks_count(es)) 5812c9de560dSAlex Tomas goal = le32_to_cpu(es->s_first_data_block); 5813c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, goal, &group, &block); 5814c9de560dSAlex Tomas 5815c9de560dSAlex Tomas /* set up allocation goals */ 5816f5a44db5STheodore Ts'o ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); 5817c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 5818c9de560dSAlex Tomas ac->ac_sb = sb; 5819c9de560dSAlex Tomas ac->ac_inode = ar->inode; 582053accfa9STheodore Ts'o ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; 5821c9de560dSAlex Tomas ac->ac_o_ex.fe_group = group; 5822c9de560dSAlex Tomas ac->ac_o_ex.fe_start = block; 5823c9de560dSAlex Tomas ac->ac_o_ex.fe_len = len; 582453accfa9STheodore Ts'o ac->ac_g_ex = ac->ac_o_ex; 58257e170922SOjaswin Mujoo ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; 5826c9de560dSAlex Tomas ac->ac_flags = ar->flags; 5827c9de560dSAlex Tomas 58283cb77bd2Sbrookxu /* we have to define context: we'll work with a file or 5829c9de560dSAlex Tomas * locality group. this is a policy, actually */ 5830c9de560dSAlex Tomas ext4_mb_group_or_file(ac); 5831c9de560dSAlex Tomas 5832d3df1453SRitesh Harjani mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " 5833c9de560dSAlex Tomas "left: %u/%u, right %u/%u to %swritable\n", 5834c9de560dSAlex Tomas (unsigned) ar->len, (unsigned) ar->logical, 5835c9de560dSAlex Tomas (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 5836c9de560dSAlex Tomas (unsigned) ar->lleft, (unsigned) ar->pleft, 5837c9de560dSAlex Tomas (unsigned) ar->lright, (unsigned) ar->pright, 583882dd124cSNikolay Borisov inode_is_open_for_write(ar->inode) ? "" : "non-"); 5839c9de560dSAlex Tomas } 5840c9de560dSAlex Tomas 58416be2ded1SAneesh Kumar K.V static noinline_for_stack void 58426be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(struct super_block *sb, 58436be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg, 58446be2ded1SAneesh Kumar K.V int order, int total_entries) 58456be2ded1SAneesh Kumar K.V { 58466be2ded1SAneesh Kumar K.V ext4_group_t group = 0; 58476be2ded1SAneesh Kumar K.V struct ext4_buddy e4b; 58480f6bc579SRuan Jinjie LIST_HEAD(discard_list); 58496be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa, *tmp; 58506be2ded1SAneesh Kumar K.V 5851d3df1453SRitesh Harjani mb_debug(sb, "discard locality group preallocation\n"); 58526be2ded1SAneesh Kumar K.V 58536be2ded1SAneesh Kumar K.V spin_lock(&lg->lg_prealloc_lock); 58546be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 5855a8e38fd3SOjaswin Mujoo pa_node.lg_list, 585692e9c58cSMadhuparna Bhowmik lockdep_is_held(&lg->lg_prealloc_lock)) { 58576be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 58586be2ded1SAneesh Kumar K.V if (atomic_read(&pa->pa_count)) { 58596be2ded1SAneesh Kumar K.V /* 58606be2ded1SAneesh Kumar K.V * This is the pa that we just used 58616be2ded1SAneesh Kumar K.V * for block allocation. So don't 58626be2ded1SAneesh Kumar K.V * free that 58636be2ded1SAneesh Kumar K.V */ 58646be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 58656be2ded1SAneesh Kumar K.V continue; 58666be2ded1SAneesh Kumar K.V } 58676be2ded1SAneesh Kumar K.V if (pa->pa_deleted) { 58686be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 58696be2ded1SAneesh Kumar K.V continue; 58706be2ded1SAneesh Kumar K.V } 58716be2ded1SAneesh Kumar K.V /* only lg prealloc space */ 5872cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_GROUP_PA); 58736be2ded1SAneesh Kumar K.V 58746be2ded1SAneesh Kumar K.V /* seems this one can be freed ... */ 587527bc446eSbrookxu ext4_mb_mark_pa_deleted(sb, pa); 58766be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 58776be2ded1SAneesh Kumar K.V 5878a8e38fd3SOjaswin Mujoo list_del_rcu(&pa->pa_node.lg_list); 58796be2ded1SAneesh Kumar K.V list_add(&pa->u.pa_tmp_list, &discard_list); 58806be2ded1SAneesh Kumar K.V 58816be2ded1SAneesh Kumar K.V total_entries--; 58826be2ded1SAneesh Kumar K.V if (total_entries <= 5) { 58836be2ded1SAneesh Kumar K.V /* 58846be2ded1SAneesh Kumar K.V * we want to keep only 5 entries 58856be2ded1SAneesh Kumar K.V * allowing it to grow to 8. This 58866be2ded1SAneesh Kumar K.V * mak sure we don't call discard 58876be2ded1SAneesh Kumar K.V * soon for this list. 58886be2ded1SAneesh Kumar K.V */ 58896be2ded1SAneesh Kumar K.V break; 58906be2ded1SAneesh Kumar K.V } 58916be2ded1SAneesh Kumar K.V } 58926be2ded1SAneesh Kumar K.V spin_unlock(&lg->lg_prealloc_lock); 58936be2ded1SAneesh Kumar K.V 58946be2ded1SAneesh Kumar K.V list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 58959651e6b2SKonstantin Khlebnikov int err; 58966be2ded1SAneesh Kumar K.V 5897bd86298eSLukas Czerner group = ext4_get_group_number(sb, pa->pa_pstart); 58989651e6b2SKonstantin Khlebnikov err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 58999651e6b2SKonstantin Khlebnikov GFP_NOFS|__GFP_NOFAIL); 59009651e6b2SKonstantin Khlebnikov if (err) { 590154d3adbcSTheodore Ts'o ext4_error_err(sb, -err, "Error %d loading buddy information for %u", 59029651e6b2SKonstantin Khlebnikov err, group); 59036be2ded1SAneesh Kumar K.V continue; 59046be2ded1SAneesh Kumar K.V } 59056be2ded1SAneesh Kumar K.V ext4_lock_group(sb, group); 59066be2ded1SAneesh Kumar K.V list_del(&pa->pa_group_list); 59073e1e5f50SEric Sandeen ext4_mb_release_group_pa(&e4b, pa); 59086be2ded1SAneesh Kumar K.V ext4_unlock_group(sb, group); 59096be2ded1SAneesh Kumar K.V 5910e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 59116be2ded1SAneesh Kumar K.V list_del(&pa->u.pa_tmp_list); 59126be2ded1SAneesh Kumar K.V call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 59136be2ded1SAneesh Kumar K.V } 59146be2ded1SAneesh Kumar K.V } 59156be2ded1SAneesh Kumar K.V 59166be2ded1SAneesh Kumar K.V /* 59176be2ded1SAneesh Kumar K.V * We have incremented pa_count. So it cannot be freed at this 59186be2ded1SAneesh Kumar K.V * point. Also we hold lg_mutex. So no parallel allocation is 59196be2ded1SAneesh Kumar K.V * possible from this lg. That means pa_free cannot be updated. 59206be2ded1SAneesh Kumar K.V * 59216be2ded1SAneesh Kumar K.V * A parallel ext4_mb_discard_group_preallocations is possible. 59226be2ded1SAneesh Kumar K.V * which can cause the lg_prealloc_list to be updated. 59236be2ded1SAneesh Kumar K.V */ 59246be2ded1SAneesh Kumar K.V 59256be2ded1SAneesh Kumar K.V static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 59266be2ded1SAneesh Kumar K.V { 59276be2ded1SAneesh Kumar K.V int order, added = 0, lg_prealloc_count = 1; 59286be2ded1SAneesh Kumar K.V struct super_block *sb = ac->ac_sb; 59296be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg = ac->ac_lg; 59306be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 59316be2ded1SAneesh Kumar K.V 59326be2ded1SAneesh Kumar K.V order = fls(pa->pa_free) - 1; 59336be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 59346be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 59356be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 59366be2ded1SAneesh Kumar K.V /* Add the prealloc space to lg */ 5937f1167009SNiu Yawei spin_lock(&lg->lg_prealloc_lock); 59386be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 5939a8e38fd3SOjaswin Mujoo pa_node.lg_list, 594092e9c58cSMadhuparna Bhowmik lockdep_is_held(&lg->lg_prealloc_lock)) { 59416be2ded1SAneesh Kumar K.V spin_lock(&tmp_pa->pa_lock); 59426be2ded1SAneesh Kumar K.V if (tmp_pa->pa_deleted) { 5943e7c9e3e9STheodore Ts'o spin_unlock(&tmp_pa->pa_lock); 59446be2ded1SAneesh Kumar K.V continue; 59456be2ded1SAneesh Kumar K.V } 59466be2ded1SAneesh Kumar K.V if (!added && pa->pa_free < tmp_pa->pa_free) { 59476be2ded1SAneesh Kumar K.V /* Add to the tail of the previous entry */ 5948a8e38fd3SOjaswin Mujoo list_add_tail_rcu(&pa->pa_node.lg_list, 5949a8e38fd3SOjaswin Mujoo &tmp_pa->pa_node.lg_list); 59506be2ded1SAneesh Kumar K.V added = 1; 59516be2ded1SAneesh Kumar K.V /* 59526be2ded1SAneesh Kumar K.V * we want to count the total 59536be2ded1SAneesh Kumar K.V * number of entries in the list 59546be2ded1SAneesh Kumar K.V */ 59556be2ded1SAneesh Kumar K.V } 59566be2ded1SAneesh Kumar K.V spin_unlock(&tmp_pa->pa_lock); 59576be2ded1SAneesh Kumar K.V lg_prealloc_count++; 59586be2ded1SAneesh Kumar K.V } 59596be2ded1SAneesh Kumar K.V if (!added) 5960a8e38fd3SOjaswin Mujoo list_add_tail_rcu(&pa->pa_node.lg_list, 59616be2ded1SAneesh Kumar K.V &lg->lg_prealloc_list[order]); 5962f1167009SNiu Yawei spin_unlock(&lg->lg_prealloc_lock); 59636be2ded1SAneesh Kumar K.V 59646be2ded1SAneesh Kumar K.V /* Now trim the list to be not more than 8 elements */ 5965ad635507SKemeng Shi if (lg_prealloc_count > 8) 59666be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(sb, lg, 59676be2ded1SAneesh Kumar K.V order, lg_prealloc_count); 59686be2ded1SAneesh Kumar K.V } 59696be2ded1SAneesh Kumar K.V 5970c9de560dSAlex Tomas /* 5971c9de560dSAlex Tomas * release all resource we used in allocation 5972c9de560dSAlex Tomas */ 5973c9de560dSAlex Tomas static int ext4_mb_release_context(struct ext4_allocation_context *ac) 5974c9de560dSAlex Tomas { 597553accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 59766be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa = ac->ac_pa; 59776be2ded1SAneesh Kumar K.V if (pa) { 5978cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) { 5979c9de560dSAlex Tomas /* see comment in ext4_mb_use_group_pa() */ 59806be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 598153accfa9STheodore Ts'o pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 598253accfa9STheodore Ts'o pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 59836be2ded1SAneesh Kumar K.V pa->pa_free -= ac->ac_b_ex.fe_len; 59846be2ded1SAneesh Kumar K.V pa->pa_len -= ac->ac_b_ex.fe_len; 59856be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 598666d5e027Sbrookxu 59876be2ded1SAneesh Kumar K.V /* 59886be2ded1SAneesh Kumar K.V * We want to add the pa to the right bucket. 59896be2ded1SAneesh Kumar K.V * Remove it from the list and while adding 59906be2ded1SAneesh Kumar K.V * make sure the list to which we are adding 599144183d42SAmir Goldstein * doesn't grow big. 59926be2ded1SAneesh Kumar K.V */ 599366d5e027Sbrookxu if (likely(pa->pa_free)) { 5994a8e38fd3SOjaswin Mujoo spin_lock(pa->pa_node_lock.lg_lock); 5995a8e38fd3SOjaswin Mujoo list_del_rcu(&pa->pa_node.lg_list); 5996a8e38fd3SOjaswin Mujoo spin_unlock(pa->pa_node_lock.lg_lock); 59976be2ded1SAneesh Kumar K.V ext4_mb_add_n_trim(ac); 5998c9de560dSAlex Tomas } 599966d5e027Sbrookxu } 600027bc446eSbrookxu 60016be2ded1SAneesh Kumar K.V ext4_mb_put_pa(ac, ac->ac_sb, pa); 6002c9de560dSAlex Tomas } 6003c9de560dSAlex Tomas if (ac->ac_bitmap_page) 600409cbfeafSKirill A. Shutemov put_page(ac->ac_bitmap_page); 6005c9de560dSAlex Tomas if (ac->ac_buddy_page) 600609cbfeafSKirill A. Shutemov put_page(ac->ac_buddy_page); 6007c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 6008c9de560dSAlex Tomas mutex_unlock(&ac->ac_lg->lg_mutex); 6009c9de560dSAlex Tomas ext4_mb_collect_stats(ac); 6010c9de560dSAlex Tomas return 0; 6011c9de560dSAlex Tomas } 6012c9de560dSAlex Tomas 6013c9de560dSAlex Tomas static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 6014c9de560dSAlex Tomas { 60158df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 6016c9de560dSAlex Tomas int ret; 60178c80fb31SChunguang Xu int freed = 0, busy = 0; 60188c80fb31SChunguang Xu int retry = 0; 6019c9de560dSAlex Tomas 60209bffad1eSTheodore Ts'o trace_ext4_mb_discard_preallocations(sb, needed); 60218c80fb31SChunguang Xu 60228c80fb31SChunguang Xu if (needed == 0) 60238c80fb31SChunguang Xu needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; 60248c80fb31SChunguang Xu repeat: 60258df9675fSTheodore Ts'o for (i = 0; i < ngroups && needed > 0; i++) { 60268c80fb31SChunguang Xu ret = ext4_mb_discard_group_preallocations(sb, i, &busy); 6027c9de560dSAlex Tomas freed += ret; 6028c9de560dSAlex Tomas needed -= ret; 60298c80fb31SChunguang Xu cond_resched(); 60308c80fb31SChunguang Xu } 60318c80fb31SChunguang Xu 60328c80fb31SChunguang Xu if (needed > 0 && busy && ++retry < 3) { 60338c80fb31SChunguang Xu busy = 0; 60348c80fb31SChunguang Xu goto repeat; 6035c9de560dSAlex Tomas } 6036c9de560dSAlex Tomas 6037c9de560dSAlex Tomas return freed; 6038c9de560dSAlex Tomas } 6039c9de560dSAlex Tomas 6040cf5e2ca6SRitesh Harjani static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, 604107b5b8e1SRitesh Harjani struct ext4_allocation_context *ac, u64 *seq) 6042cf5e2ca6SRitesh Harjani { 6043cf5e2ca6SRitesh Harjani int freed; 604407b5b8e1SRitesh Harjani u64 seq_retry = 0; 604507b5b8e1SRitesh Harjani bool ret = false; 6046cf5e2ca6SRitesh Harjani 6047cf5e2ca6SRitesh Harjani freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 604807b5b8e1SRitesh Harjani if (freed) { 604907b5b8e1SRitesh Harjani ret = true; 605007b5b8e1SRitesh Harjani goto out_dbg; 605107b5b8e1SRitesh Harjani } 605207b5b8e1SRitesh Harjani seq_retry = ext4_get_discard_pa_seq_sum(); 605399377830SRitesh Harjani if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { 605499377830SRitesh Harjani ac->ac_flags |= EXT4_MB_STRICT_CHECK; 605507b5b8e1SRitesh Harjani *seq = seq_retry; 605607b5b8e1SRitesh Harjani ret = true; 605707b5b8e1SRitesh Harjani } 605807b5b8e1SRitesh Harjani 605907b5b8e1SRitesh Harjani out_dbg: 606007b5b8e1SRitesh Harjani mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); 606107b5b8e1SRitesh Harjani return ret; 6062cf5e2ca6SRitesh Harjani } 6063cf5e2ca6SRitesh Harjani 6064ad78b5efSKemeng Shi /* 6065ad78b5efSKemeng Shi * Simple allocator for Ext4 fast commit replay path. It searches for blocks 6066ad78b5efSKemeng Shi * linearly starting at the goal block and also excludes the blocks which 6067ad78b5efSKemeng Shi * are going to be in use after fast commit replay. 6068ad78b5efSKemeng Shi */ 6069ad78b5efSKemeng Shi static ext4_fsblk_t 6070ad78b5efSKemeng Shi ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp) 6071ad78b5efSKemeng Shi { 6072ad78b5efSKemeng Shi struct buffer_head *bitmap_bh; 6073ad78b5efSKemeng Shi struct super_block *sb = ar->inode->i_sb; 6074ad78b5efSKemeng Shi struct ext4_sb_info *sbi = EXT4_SB(sb); 6075ad78b5efSKemeng Shi ext4_group_t group, nr; 6076ad78b5efSKemeng Shi ext4_grpblk_t blkoff; 6077ad78b5efSKemeng Shi ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 6078ad78b5efSKemeng Shi ext4_grpblk_t i = 0; 6079ad78b5efSKemeng Shi ext4_fsblk_t goal, block; 608079ebf48cSLu Hongfei struct ext4_super_block *es = sbi->s_es; 6081ad78b5efSKemeng Shi 6082ad78b5efSKemeng Shi goal = ar->goal; 6083ad78b5efSKemeng Shi if (goal < le32_to_cpu(es->s_first_data_block) || 6084ad78b5efSKemeng Shi goal >= ext4_blocks_count(es)) 6085ad78b5efSKemeng Shi goal = le32_to_cpu(es->s_first_data_block); 6086ad78b5efSKemeng Shi 6087ad78b5efSKemeng Shi ar->len = 0; 6088ad78b5efSKemeng Shi ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); 6089ad78b5efSKemeng Shi for (nr = ext4_get_groups_count(sb); nr > 0; nr--) { 6090ad78b5efSKemeng Shi bitmap_bh = ext4_read_block_bitmap(sb, group); 6091ad78b5efSKemeng Shi if (IS_ERR(bitmap_bh)) { 6092ad78b5efSKemeng Shi *errp = PTR_ERR(bitmap_bh); 6093ad78b5efSKemeng Shi pr_warn("Failed to read block bitmap\n"); 6094ad78b5efSKemeng Shi return 0; 6095ad78b5efSKemeng Shi } 6096ad78b5efSKemeng Shi 6097ad78b5efSKemeng Shi while (1) { 6098ad78b5efSKemeng Shi i = mb_find_next_zero_bit(bitmap_bh->b_data, max, 6099ad78b5efSKemeng Shi blkoff); 6100ad78b5efSKemeng Shi if (i >= max) 6101ad78b5efSKemeng Shi break; 6102ad78b5efSKemeng Shi if (ext4_fc_replay_check_excluded(sb, 6103ad78b5efSKemeng Shi ext4_group_first_block_no(sb, group) + 6104ad78b5efSKemeng Shi EXT4_C2B(sbi, i))) { 6105ad78b5efSKemeng Shi blkoff = i + 1; 6106ad78b5efSKemeng Shi } else 6107ad78b5efSKemeng Shi break; 6108ad78b5efSKemeng Shi } 6109ad78b5efSKemeng Shi brelse(bitmap_bh); 6110ad78b5efSKemeng Shi if (i < max) 6111ad78b5efSKemeng Shi break; 6112ad78b5efSKemeng Shi 6113ad78b5efSKemeng Shi if (++group >= ext4_get_groups_count(sb)) 6114ad78b5efSKemeng Shi group = 0; 6115ad78b5efSKemeng Shi 6116ad78b5efSKemeng Shi blkoff = 0; 6117ad78b5efSKemeng Shi } 6118ad78b5efSKemeng Shi 6119ad78b5efSKemeng Shi if (i >= max) { 6120ad78b5efSKemeng Shi *errp = -ENOSPC; 6121ad78b5efSKemeng Shi return 0; 6122ad78b5efSKemeng Shi } 6123ad78b5efSKemeng Shi 6124ad78b5efSKemeng Shi block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i); 6125ad78b5efSKemeng Shi ext4_mb_mark_bb(sb, block, 1, 1); 6126ad78b5efSKemeng Shi ar->len = 1; 6127ad78b5efSKemeng Shi 6128ad78b5efSKemeng Shi return block; 6129ad78b5efSKemeng Shi } 61308016e29fSHarshad Shirwadkar 6131c9de560dSAlex Tomas /* 6132c9de560dSAlex Tomas * Main entry point into mballoc to allocate blocks 6133c9de560dSAlex Tomas * it tries to use preallocation first, then falls back 6134c9de560dSAlex Tomas * to usual allocation 6135c9de560dSAlex Tomas */ 6136c9de560dSAlex Tomas ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 6137c9de560dSAlex Tomas struct ext4_allocation_request *ar, int *errp) 6138c9de560dSAlex Tomas { 6139256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 6140c9de560dSAlex Tomas struct ext4_sb_info *sbi; 6141c9de560dSAlex Tomas struct super_block *sb; 6142c9de560dSAlex Tomas ext4_fsblk_t block = 0; 614360e58e0fSMingming Cao unsigned int inquota = 0; 614453accfa9STheodore Ts'o unsigned int reserv_clstrs = 0; 614580fa46d6STheodore Ts'o int retries = 0; 614607b5b8e1SRitesh Harjani u64 seq; 6147c9de560dSAlex Tomas 6148b10a44c3STheodore Ts'o might_sleep(); 6149c9de560dSAlex Tomas sb = ar->inode->i_sb; 6150c9de560dSAlex Tomas sbi = EXT4_SB(sb); 6151c9de560dSAlex Tomas 61529bffad1eSTheodore Ts'o trace_ext4_request_blocks(ar); 61538016e29fSHarshad Shirwadkar if (sbi->s_mount_state & EXT4_FC_REPLAY) 6154ad78b5efSKemeng Shi return ext4_mb_new_blocks_simple(ar, errp); 6155ba80b101STheodore Ts'o 615645dc63e7SDmitry Monakhov /* Allow to use superuser reservation for quota file */ 615702749a4cSTahsin Erdogan if (ext4_is_quota_file(ar->inode)) 615845dc63e7SDmitry Monakhov ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; 615945dc63e7SDmitry Monakhov 6160e3cf5d5dSTheodore Ts'o if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { 616160e58e0fSMingming Cao /* Without delayed allocation we need to verify 616260e58e0fSMingming Cao * there is enough free blocks to do block allocation 616360e58e0fSMingming Cao * and verify allocation doesn't exceed the quota limits. 6164d2a17637SMingming Cao */ 616555f020dbSAllison Henderson while (ar->len && 6166e7d5f315STheodore Ts'o ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { 616755f020dbSAllison Henderson 6168030ba6bcSAneesh Kumar K.V /* let others to free the space */ 6169bb8b20edSLukas Czerner cond_resched(); 6170030ba6bcSAneesh Kumar K.V ar->len = ar->len >> 1; 6171030ba6bcSAneesh Kumar K.V } 6172030ba6bcSAneesh Kumar K.V if (!ar->len) { 6173bbc4ec77SRitesh Harjani ext4_mb_show_pa(sb); 617407031431SMingming Cao *errp = -ENOSPC; 617507031431SMingming Cao return 0; 617607031431SMingming Cao } 617753accfa9STheodore Ts'o reserv_clstrs = ar->len; 617855f020dbSAllison Henderson if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { 617953accfa9STheodore Ts'o dquot_alloc_block_nofail(ar->inode, 618053accfa9STheodore Ts'o EXT4_C2B(sbi, ar->len)); 618155f020dbSAllison Henderson } else { 618255f020dbSAllison Henderson while (ar->len && 618353accfa9STheodore Ts'o dquot_alloc_block(ar->inode, 618453accfa9STheodore Ts'o EXT4_C2B(sbi, ar->len))) { 618555f020dbSAllison Henderson 6186c9de560dSAlex Tomas ar->flags |= EXT4_MB_HINT_NOPREALLOC; 6187c9de560dSAlex Tomas ar->len--; 6188c9de560dSAlex Tomas } 618955f020dbSAllison Henderson } 619060e58e0fSMingming Cao inquota = ar->len; 6191c9de560dSAlex Tomas if (ar->len == 0) { 6192c9de560dSAlex Tomas *errp = -EDQUOT; 61936c7a120aSAditya Kali goto out; 6194c9de560dSAlex Tomas } 619560e58e0fSMingming Cao } 6196d2a17637SMingming Cao 619785556c9aSWei Yongjun ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); 6198833576b3STheodore Ts'o if (!ac) { 6199363d4251SShen Feng ar->len = 0; 6200256bdb49SEric Sandeen *errp = -ENOMEM; 62016c7a120aSAditya Kali goto out; 6202256bdb49SEric Sandeen } 6203256bdb49SEric Sandeen 6204d73eff68SGuoqing Jiang ext4_mb_initialize_context(ac, ar); 6205c9de560dSAlex Tomas 6206256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 620781198536SRitesh Harjani seq = this_cpu_read(discard_pa_seq); 6208256bdb49SEric Sandeen if (!ext4_mb_use_preallocated(ac)) { 6209256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_ALLOC; 6210256bdb49SEric Sandeen ext4_mb_normalize_request(ac, ar); 621153f86b17SRitesh Harjani 621253f86b17SRitesh Harjani *errp = ext4_mb_pa_alloc(ac); 621353f86b17SRitesh Harjani if (*errp) 621453f86b17SRitesh Harjani goto errout; 6215c9de560dSAlex Tomas repeat: 6216c9de560dSAlex Tomas /* allocate space in core */ 62176c7a120aSAditya Kali *errp = ext4_mb_regular_allocator(ac); 621853f86b17SRitesh Harjani /* 621953f86b17SRitesh Harjani * pa allocated above is added to grp->bb_prealloc_list only 622053f86b17SRitesh Harjani * when we were able to allocate some block i.e. when 622153f86b17SRitesh Harjani * ac->ac_status == AC_STATUS_FOUND. 622253f86b17SRitesh Harjani * And error from above mean ac->ac_status != AC_STATUS_FOUND 622353f86b17SRitesh Harjani * So we have to free this pa here itself. 622453f86b17SRitesh Harjani */ 62252c00ef3eSAlexey Khoroshilov if (*errp) { 622682089725SOjaswin Mujoo ext4_mb_pa_put_free(ac); 62272c00ef3eSAlexey Khoroshilov ext4_discard_allocated_blocks(ac); 62282c00ef3eSAlexey Khoroshilov goto errout; 62292c00ef3eSAlexey Khoroshilov } 623053f86b17SRitesh Harjani if (ac->ac_status == AC_STATUS_FOUND && 623153f86b17SRitesh Harjani ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) 623282089725SOjaswin Mujoo ext4_mb_pa_put_free(ac); 6233c9de560dSAlex Tomas } 6234256bdb49SEric Sandeen if (likely(ac->ac_status == AC_STATUS_FOUND)) { 623553accfa9STheodore Ts'o *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); 6236554a5cccSVegard Nossum if (*errp) { 6237b844167eSCurt Wohlgemuth ext4_discard_allocated_blocks(ac); 62386d138cedSEric Sandeen goto errout; 62396d138cedSEric Sandeen } else { 6240256bdb49SEric Sandeen block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 6241256bdb49SEric Sandeen ar->len = ac->ac_b_ex.fe_len; 6242519deca0SAneesh Kumar K.V } 6243c9de560dSAlex Tomas } else { 624480fa46d6STheodore Ts'o if (++retries < 3 && 624580fa46d6STheodore Ts'o ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) 6246c9de560dSAlex Tomas goto repeat; 624753f86b17SRitesh Harjani /* 624853f86b17SRitesh Harjani * If block allocation fails then the pa allocated above 624953f86b17SRitesh Harjani * needs to be freed here itself. 625053f86b17SRitesh Harjani */ 625182089725SOjaswin Mujoo ext4_mb_pa_put_free(ac); 6252c9de560dSAlex Tomas *errp = -ENOSPC; 62536c7a120aSAditya Kali } 62546c7a120aSAditya Kali 62556c7a120aSAditya Kali if (*errp) { 6256aaae558dSKemeng Shi errout: 6257256bdb49SEric Sandeen ac->ac_b_ex.fe_len = 0; 6258c9de560dSAlex Tomas ar->len = 0; 6259256bdb49SEric Sandeen ext4_mb_show_ac(ac); 6260c9de560dSAlex Tomas } 6261256bdb49SEric Sandeen ext4_mb_release_context(ac); 6262363d4251SShen Feng kmem_cache_free(ext4_ac_cachep, ac); 6263aaae558dSKemeng Shi out: 626460e58e0fSMingming Cao if (inquota && ar->len < inquota) 626553accfa9STheodore Ts'o dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); 62660087d9fbSAneesh Kumar K.V if (!ar->len) { 6267e3cf5d5dSTheodore Ts'o if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) 62680087d9fbSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 626957042651STheodore Ts'o percpu_counter_sub(&sbi->s_dirtyclusters_counter, 627053accfa9STheodore Ts'o reserv_clstrs); 62710087d9fbSAneesh Kumar K.V } 6272c9de560dSAlex Tomas 62739bffad1eSTheodore Ts'o trace_ext4_allocate_blocks(ar, (unsigned long long)block); 6274ba80b101STheodore Ts'o 6275c9de560dSAlex Tomas return block; 6276c9de560dSAlex Tomas } 6277c9de560dSAlex Tomas 6278c894058dSAneesh Kumar K.V /* 6279c894058dSAneesh Kumar K.V * We can merge two free data extents only if the physical blocks 6280c894058dSAneesh Kumar K.V * are contiguous, AND the extents were freed by the same transaction, 6281c894058dSAneesh Kumar K.V * AND the blocks are associated with the same group. 6282c894058dSAneesh Kumar K.V */ 6283a0154344SDaeho Jeong static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, 6284a0154344SDaeho Jeong struct ext4_free_data *entry, 6285a0154344SDaeho Jeong struct ext4_free_data *new_entry, 6286a0154344SDaeho Jeong struct rb_root *entry_rb_root) 6287c894058dSAneesh Kumar K.V { 6288a0154344SDaeho Jeong if ((entry->efd_tid != new_entry->efd_tid) || 6289a0154344SDaeho Jeong (entry->efd_group != new_entry->efd_group)) 6290a0154344SDaeho Jeong return; 6291a0154344SDaeho Jeong if (entry->efd_start_cluster + entry->efd_count == 6292a0154344SDaeho Jeong new_entry->efd_start_cluster) { 6293a0154344SDaeho Jeong new_entry->efd_start_cluster = entry->efd_start_cluster; 6294a0154344SDaeho Jeong new_entry->efd_count += entry->efd_count; 6295a0154344SDaeho Jeong } else if (new_entry->efd_start_cluster + new_entry->efd_count == 6296a0154344SDaeho Jeong entry->efd_start_cluster) { 6297a0154344SDaeho Jeong new_entry->efd_count += entry->efd_count; 6298a0154344SDaeho Jeong } else 6299a0154344SDaeho Jeong return; 6300a0154344SDaeho Jeong spin_lock(&sbi->s_md_lock); 6301a0154344SDaeho Jeong list_del(&entry->efd_list); 6302a0154344SDaeho Jeong spin_unlock(&sbi->s_md_lock); 6303a0154344SDaeho Jeong rb_erase(&entry->efd_node, entry_rb_root); 6304a0154344SDaeho Jeong kmem_cache_free(ext4_free_data_cachep, entry); 6305c894058dSAneesh Kumar K.V } 6306c894058dSAneesh Kumar K.V 630785b67ffbSKemeng Shi static noinline_for_stack void 63084ddfef7bSEric Sandeen ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 63097a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry) 6310c9de560dSAlex Tomas { 6311e29136f8STheodore Ts'o ext4_group_t group = e4b->bd_group; 631284130193STheodore Ts'o ext4_grpblk_t cluster; 6313d08854f5STheodore Ts'o ext4_grpblk_t clusters = new_entry->efd_count; 63147a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 6315c9de560dSAlex Tomas struct ext4_group_info *db = e4b->bd_info; 6316c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 6317c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 6318c894058dSAneesh Kumar K.V struct rb_node **n = &db->bb_free_root.rb_node, *node; 6319c894058dSAneesh Kumar K.V struct rb_node *parent = NULL, *new_node; 6320c894058dSAneesh Kumar K.V 63210390131bSFrank Mayhar BUG_ON(!ext4_handle_valid(handle)); 6322c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 6323c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 6324c9de560dSAlex Tomas 632518aadd47SBobi Jam new_node = &new_entry->efd_node; 632618aadd47SBobi Jam cluster = new_entry->efd_start_cluster; 6327c9de560dSAlex Tomas 6328c894058dSAneesh Kumar K.V if (!*n) { 6329c894058dSAneesh Kumar K.V /* first free block exent. We need to 6330c894058dSAneesh Kumar K.V protect buddy cache from being freed, 6331c9de560dSAlex Tomas * otherwise we'll refresh it from 6332c9de560dSAlex Tomas * on-disk bitmap and lose not-yet-available 6333c9de560dSAlex Tomas * blocks */ 633409cbfeafSKirill A. Shutemov get_page(e4b->bd_buddy_page); 633509cbfeafSKirill A. Shutemov get_page(e4b->bd_bitmap_page); 6336c894058dSAneesh Kumar K.V } 6337c894058dSAneesh Kumar K.V while (*n) { 6338c894058dSAneesh Kumar K.V parent = *n; 633918aadd47SBobi Jam entry = rb_entry(parent, struct ext4_free_data, efd_node); 634018aadd47SBobi Jam if (cluster < entry->efd_start_cluster) 6341c894058dSAneesh Kumar K.V n = &(*n)->rb_left; 634218aadd47SBobi Jam else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) 6343c894058dSAneesh Kumar K.V n = &(*n)->rb_right; 6344c894058dSAneesh Kumar K.V else { 6345e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 634684130193STheodore Ts'o ext4_group_first_block_no(sb, group) + 634784130193STheodore Ts'o EXT4_C2B(sbi, cluster), 6348e29136f8STheodore Ts'o "Block already on to-be-freed list"); 6349cca41553SChunguang Xu kmem_cache_free(ext4_free_data_cachep, new_entry); 635085b67ffbSKemeng Shi return; 6351c9de560dSAlex Tomas } 6352c9de560dSAlex Tomas } 6353c9de560dSAlex Tomas 6354c894058dSAneesh Kumar K.V rb_link_node(new_node, parent, n); 6355c894058dSAneesh Kumar K.V rb_insert_color(new_node, &db->bb_free_root); 6356c894058dSAneesh Kumar K.V 6357c894058dSAneesh Kumar K.V /* Now try to see the extent can be merged to left and right */ 6358c894058dSAneesh Kumar K.V node = rb_prev(new_node); 6359c894058dSAneesh Kumar K.V if (node) { 636018aadd47SBobi Jam entry = rb_entry(node, struct ext4_free_data, efd_node); 6361a0154344SDaeho Jeong ext4_try_merge_freed_extent(sbi, entry, new_entry, 6362a0154344SDaeho Jeong &(db->bb_free_root)); 6363c9de560dSAlex Tomas } 6364c894058dSAneesh Kumar K.V 6365c894058dSAneesh Kumar K.V node = rb_next(new_node); 6366c894058dSAneesh Kumar K.V if (node) { 636718aadd47SBobi Jam entry = rb_entry(node, struct ext4_free_data, efd_node); 6368a0154344SDaeho Jeong ext4_try_merge_freed_extent(sbi, entry, new_entry, 6369a0154344SDaeho Jeong &(db->bb_free_root)); 6370c894058dSAneesh Kumar K.V } 6371a0154344SDaeho Jeong 6372d08854f5STheodore Ts'o spin_lock(&sbi->s_md_lock); 6373*ce774e53SJinke Han list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]); 6374d08854f5STheodore Ts'o sbi->s_mb_free_pending += clusters; 6375d08854f5STheodore Ts'o spin_unlock(&sbi->s_md_lock); 6376c9de560dSAlex Tomas } 6377c9de560dSAlex Tomas 63788016e29fSHarshad Shirwadkar static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, 63798016e29fSHarshad Shirwadkar unsigned long count) 63808016e29fSHarshad Shirwadkar { 63818016e29fSHarshad Shirwadkar struct buffer_head *bitmap_bh; 63828016e29fSHarshad Shirwadkar struct super_block *sb = inode->i_sb; 63838016e29fSHarshad Shirwadkar struct ext4_group_desc *gdp; 63848016e29fSHarshad Shirwadkar struct buffer_head *gdp_bh; 63858016e29fSHarshad Shirwadkar ext4_group_t group; 63868016e29fSHarshad Shirwadkar ext4_grpblk_t blkoff; 63878016e29fSHarshad Shirwadkar int already_freed = 0, err, i; 63888016e29fSHarshad Shirwadkar 63898016e29fSHarshad Shirwadkar ext4_get_group_no_and_offset(sb, block, &group, &blkoff); 63908016e29fSHarshad Shirwadkar bitmap_bh = ext4_read_block_bitmap(sb, group); 63918016e29fSHarshad Shirwadkar if (IS_ERR(bitmap_bh)) { 63928016e29fSHarshad Shirwadkar pr_warn("Failed to read block bitmap\n"); 63938016e29fSHarshad Shirwadkar return; 63948016e29fSHarshad Shirwadkar } 63958016e29fSHarshad Shirwadkar gdp = ext4_get_group_desc(sb, group, &gdp_bh); 63968016e29fSHarshad Shirwadkar if (!gdp) 63971b5c9d34SKemeng Shi goto err_out; 63988016e29fSHarshad Shirwadkar 63998016e29fSHarshad Shirwadkar for (i = 0; i < count; i++) { 64008016e29fSHarshad Shirwadkar if (!mb_test_bit(blkoff + i, bitmap_bh->b_data)) 64018016e29fSHarshad Shirwadkar already_freed++; 64028016e29fSHarshad Shirwadkar } 64038016e29fSHarshad Shirwadkar mb_clear_bits(bitmap_bh->b_data, blkoff, count); 64048016e29fSHarshad Shirwadkar err = ext4_handle_dirty_metadata(NULL, NULL, bitmap_bh); 64058016e29fSHarshad Shirwadkar if (err) 64061b5c9d34SKemeng Shi goto err_out; 64078016e29fSHarshad Shirwadkar ext4_free_group_clusters_set( 64088016e29fSHarshad Shirwadkar sb, gdp, ext4_free_group_clusters(sb, gdp) + 64098016e29fSHarshad Shirwadkar count - already_freed); 64101df9bde4SKemeng Shi ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 64118016e29fSHarshad Shirwadkar ext4_group_desc_csum_set(sb, group, gdp); 64128016e29fSHarshad Shirwadkar ext4_handle_dirty_metadata(NULL, NULL, gdp_bh); 64138016e29fSHarshad Shirwadkar sync_dirty_buffer(bitmap_bh); 64148016e29fSHarshad Shirwadkar sync_dirty_buffer(gdp_bh); 64151b5c9d34SKemeng Shi 64161b5c9d34SKemeng Shi err_out: 64178016e29fSHarshad Shirwadkar brelse(bitmap_bh); 64188016e29fSHarshad Shirwadkar } 64198016e29fSHarshad Shirwadkar 642044338711STheodore Ts'o /** 64218ac3939dSRitesh Harjani * ext4_mb_clear_bb() -- helper function for freeing blocks. 64228ac3939dSRitesh Harjani * Used by ext4_free_blocks() 642344338711STheodore Ts'o * @handle: handle for this transaction 642444338711STheodore Ts'o * @inode: inode 6425c60990b3STheodore Ts'o * @block: starting physical block to be freed 6426c60990b3STheodore Ts'o * @count: number of blocks to be freed 64275def1360SYongqiang Yang * @flags: flags used by ext4_free_blocks 6428c9de560dSAlex Tomas */ 64298ac3939dSRitesh Harjani static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode, 64308ac3939dSRitesh Harjani ext4_fsblk_t block, unsigned long count, 64318ac3939dSRitesh Harjani int flags) 6432c9de560dSAlex Tomas { 643326346ff6SAneesh Kumar K.V struct buffer_head *bitmap_bh = NULL; 6434c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 6435c9de560dSAlex Tomas struct ext4_group_desc *gdp; 64365354b2afSTheodore Ts'o struct ext4_group_info *grp; 6437498e5f24STheodore Ts'o unsigned int overflow; 6438c9de560dSAlex Tomas ext4_grpblk_t bit; 6439c9de560dSAlex Tomas struct buffer_head *gd_bh; 6440c9de560dSAlex Tomas ext4_group_t block_group; 6441c9de560dSAlex Tomas struct ext4_sb_info *sbi; 6442c9de560dSAlex Tomas struct ext4_buddy e4b; 644384130193STheodore Ts'o unsigned int count_clusters; 6444c9de560dSAlex Tomas int err = 0; 6445c9de560dSAlex Tomas int ret; 6446c9de560dSAlex Tomas 64478016e29fSHarshad Shirwadkar sbi = EXT4_SB(sb); 64488016e29fSHarshad Shirwadkar 64491e1c2b86SLukas Czerner if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 64501e1c2b86SLukas Czerner !ext4_inode_block_valid(inode, block, count)) { 64511e1c2b86SLukas Czerner ext4_error(sb, "Freeing blocks in system zone - " 64521e1c2b86SLukas Czerner "Block = %llu, count = %lu", block, count); 64531e1c2b86SLukas Czerner /* err = 0. ext4_std_error should be a no op */ 64541e1c2b86SLukas Czerner goto error_return; 64551e1c2b86SLukas Czerner } 64561e1c2b86SLukas Czerner flags |= EXT4_FREE_BLOCKS_VALIDATED; 64571e1c2b86SLukas Czerner 6458c9de560dSAlex Tomas do_more: 6459c9de560dSAlex Tomas overflow = 0; 6460c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 6461c9de560dSAlex Tomas 64625354b2afSTheodore Ts'o grp = ext4_get_group_info(sb, block_group); 64635354b2afSTheodore Ts'o if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 6464163a203dSDarrick J. Wong return; 6465163a203dSDarrick J. Wong 6466c9de560dSAlex Tomas /* 6467c9de560dSAlex Tomas * Check to see if we are freeing blocks across a group 6468c9de560dSAlex Tomas * boundary. 6469c9de560dSAlex Tomas */ 647084130193STheodore Ts'o if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { 647184130193STheodore Ts'o overflow = EXT4_C2B(sbi, bit) + count - 647284130193STheodore Ts'o EXT4_BLOCKS_PER_GROUP(sb); 6473c9de560dSAlex Tomas count -= overflow; 64741e1c2b86SLukas Czerner /* The range changed so it's no longer validated */ 64751e1c2b86SLukas Czerner flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 6476c9de560dSAlex Tomas } 6477810da240SLukas Czerner count_clusters = EXT4_NUM_B2C(sbi, count); 6478574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, block_group); 64799008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 64809008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 64819008a58eSDarrick J. Wong bitmap_bh = NULL; 6482c9de560dSAlex Tomas goto error_return; 6483ce89f46cSAneesh Kumar K.V } 6484c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 6485ce89f46cSAneesh Kumar K.V if (!gdp) { 6486ce89f46cSAneesh Kumar K.V err = -EIO; 6487c9de560dSAlex Tomas goto error_return; 6488ce89f46cSAneesh Kumar K.V } 6489c9de560dSAlex Tomas 64901e1c2b86SLukas Czerner if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 64911e1c2b86SLukas Czerner !ext4_inode_block_valid(inode, block, count)) { 649212062dddSEric Sandeen ext4_error(sb, "Freeing blocks in system zone - " 64930610b6e9STheodore Ts'o "Block = %llu, count = %lu", block, count); 6494519deca0SAneesh Kumar K.V /* err = 0. ext4_std_error should be a no op */ 6495519deca0SAneesh Kumar K.V goto error_return; 6496c9de560dSAlex Tomas } 6497c9de560dSAlex Tomas 6498c9de560dSAlex Tomas BUFFER_TRACE(bitmap_bh, "getting write access"); 6499188c299eSJan Kara err = ext4_journal_get_write_access(handle, sb, bitmap_bh, 6500188c299eSJan Kara EXT4_JTR_NONE); 6501c9de560dSAlex Tomas if (err) 6502c9de560dSAlex Tomas goto error_return; 6503c9de560dSAlex Tomas 6504c9de560dSAlex Tomas /* 6505c9de560dSAlex Tomas * We are about to modify some metadata. Call the journal APIs 6506c9de560dSAlex Tomas * to unshare ->b_data if a currently-committing transaction is 6507c9de560dSAlex Tomas * using it 6508c9de560dSAlex Tomas */ 6509c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "get_write_access"); 6510188c299eSJan Kara err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); 6511c9de560dSAlex Tomas if (err) 6512c9de560dSAlex Tomas goto error_return; 6513c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 6514c9de560dSAlex Tomas { 6515c9de560dSAlex Tomas int i; 651684130193STheodore Ts'o for (i = 0; i < count_clusters; i++) 6517c9de560dSAlex Tomas BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 6518c9de560dSAlex Tomas } 6519c9de560dSAlex Tomas #endif 652084130193STheodore Ts'o trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); 6521c9de560dSAlex Tomas 6522adb7ef60SKonstantin Khlebnikov /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ 6523adb7ef60SKonstantin Khlebnikov err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, 6524adb7ef60SKonstantin Khlebnikov GFP_NOFS|__GFP_NOFAIL); 6525920313a7SAneesh Kumar K.V if (err) 6526920313a7SAneesh Kumar K.V goto error_return; 6527e6362609STheodore Ts'o 6528f96c450dSDaeho Jeong /* 6529f96c450dSDaeho Jeong * We need to make sure we don't reuse the freed block until after the 6530f96c450dSDaeho Jeong * transaction is committed. We make an exception if the inode is to be 6531f96c450dSDaeho Jeong * written in writeback mode since writeback mode has weak data 6532f96c450dSDaeho Jeong * consistency guarantees. 6533f96c450dSDaeho Jeong */ 6534f96c450dSDaeho Jeong if (ext4_handle_valid(handle) && 6535f96c450dSDaeho Jeong ((flags & EXT4_FREE_BLOCKS_METADATA) || 6536f96c450dSDaeho Jeong !ext4_should_writeback_data(inode))) { 65377a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry; 65387a2fcbf7SAneesh Kumar K.V /* 65397444a072SMichal Hocko * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed 65407444a072SMichal Hocko * to fail. 65417a2fcbf7SAneesh Kumar K.V */ 65427444a072SMichal Hocko new_entry = kmem_cache_alloc(ext4_free_data_cachep, 65437444a072SMichal Hocko GFP_NOFS|__GFP_NOFAIL); 654418aadd47SBobi Jam new_entry->efd_start_cluster = bit; 654518aadd47SBobi Jam new_entry->efd_group = block_group; 654618aadd47SBobi Jam new_entry->efd_count = count_clusters; 654718aadd47SBobi Jam new_entry->efd_tid = handle->h_transaction->t_tid; 6548955ce5f5SAneesh Kumar K.V 65497a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, block_group); 655084130193STheodore Ts'o mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 65517a2fcbf7SAneesh Kumar K.V ext4_mb_free_metadata(handle, &e4b, new_entry); 6552c9de560dSAlex Tomas } else { 65537a2fcbf7SAneesh Kumar K.V /* need to update group_info->bb_free and bitmap 65547a2fcbf7SAneesh Kumar K.V * with group lock held. generate_buddy look at 65557a2fcbf7SAneesh Kumar K.V * them with group lock_held 65567a2fcbf7SAneesh Kumar K.V */ 6557d71c1ae2SLukas Czerner if (test_opt(sb, DISCARD)) { 6558247c3d21SKemeng Shi err = ext4_issue_discard(sb, block_group, bit, 6559247c3d21SKemeng Shi count_clusters, NULL); 6560d71c1ae2SLukas Czerner if (err && err != -EOPNOTSUPP) 6561d71c1ae2SLukas Czerner ext4_msg(sb, KERN_WARNING, "discard request in" 6562a00b482bSRitesh Harjani " group:%u block:%d count:%lu failed" 6563d71c1ae2SLukas Czerner " with %d", block_group, bit, count, 6564d71c1ae2SLukas Czerner err); 65658f9ff189SLukas Czerner } else 65668f9ff189SLukas Czerner EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); 6567d71c1ae2SLukas Czerner 6568955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 656984130193STheodore Ts'o mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 657084130193STheodore Ts'o mb_free_blocks(inode, &e4b, bit, count_clusters); 6571c9de560dSAlex Tomas } 6572c9de560dSAlex Tomas 6573021b65bbSTheodore Ts'o ret = ext4_free_group_clusters(sb, gdp) + count_clusters; 6574021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, ret); 65751df9bde4SKemeng Shi ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 6576feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, block_group, gdp); 6577955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 6578c9de560dSAlex Tomas 6579772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 6580772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 658190ba983fSTheodore Ts'o atomic64_add(count_clusters, 65827c990728SSuraj Jitindar Singh &sbi_array_rcu_deref(sbi, s_flex_groups, 65837c990728SSuraj Jitindar Singh flex_group)->free_clusters); 6584772cb7c8SJose R. Santos } 6585772cb7c8SJose R. Santos 65869fe67149SEric Whitney /* 65879fe67149SEric Whitney * on a bigalloc file system, defer the s_freeclusters_counter 65889fe67149SEric Whitney * update to the caller (ext4_remove_space and friends) so they 65899fe67149SEric Whitney * can determine if a cluster freed here should be rereserved 65909fe67149SEric Whitney */ 65919fe67149SEric Whitney if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { 65927b415bf6SAditya Kali if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) 65937b415bf6SAditya Kali dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); 65949fe67149SEric Whitney percpu_counter_add(&sbi->s_freeclusters_counter, 65959fe67149SEric Whitney count_clusters); 65969fe67149SEric Whitney } 65977d734532SJan Kara 65987d734532SJan Kara ext4_mb_unload_buddy(&e4b); 65997b415bf6SAditya Kali 66007a2fcbf7SAneesh Kumar K.V /* We dirtied the bitmap block */ 66017a2fcbf7SAneesh Kumar K.V BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 66027a2fcbf7SAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 66037a2fcbf7SAneesh Kumar K.V 6604c9de560dSAlex Tomas /* And the group descriptor block */ 6605c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 66060390131bSFrank Mayhar ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 6607c9de560dSAlex Tomas if (!err) 6608c9de560dSAlex Tomas err = ret; 6609c9de560dSAlex Tomas 6610c9de560dSAlex Tomas if (overflow && !err) { 6611c9de560dSAlex Tomas block += count; 6612c9de560dSAlex Tomas count = overflow; 6613c9de560dSAlex Tomas put_bh(bitmap_bh); 66141e1c2b86SLukas Czerner /* The range changed so it's no longer validated */ 66151e1c2b86SLukas Czerner flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 6616c9de560dSAlex Tomas goto do_more; 6617c9de560dSAlex Tomas } 6618c9de560dSAlex Tomas error_return: 6619c9de560dSAlex Tomas brelse(bitmap_bh); 6620c9de560dSAlex Tomas ext4_std_error(sb, err); 6621c9de560dSAlex Tomas } 66227360d173SLukas Czerner 66237360d173SLukas Czerner /** 66248ac3939dSRitesh Harjani * ext4_free_blocks() -- Free given blocks and update quota 66258ac3939dSRitesh Harjani * @handle: handle for this transaction 66268ac3939dSRitesh Harjani * @inode: inode 66278ac3939dSRitesh Harjani * @bh: optional buffer of the block to be freed 66288ac3939dSRitesh Harjani * @block: starting physical block to be freed 66298ac3939dSRitesh Harjani * @count: number of blocks to be freed 66308ac3939dSRitesh Harjani * @flags: flags used by ext4_free_blocks 66318ac3939dSRitesh Harjani */ 66328ac3939dSRitesh Harjani void ext4_free_blocks(handle_t *handle, struct inode *inode, 66338ac3939dSRitesh Harjani struct buffer_head *bh, ext4_fsblk_t block, 66348ac3939dSRitesh Harjani unsigned long count, int flags) 66358ac3939dSRitesh Harjani { 66368ac3939dSRitesh Harjani struct super_block *sb = inode->i_sb; 66378ac3939dSRitesh Harjani unsigned int overflow; 66388ac3939dSRitesh Harjani struct ext4_sb_info *sbi; 66398ac3939dSRitesh Harjani 66408ac3939dSRitesh Harjani sbi = EXT4_SB(sb); 66418ac3939dSRitesh Harjani 66428ac3939dSRitesh Harjani if (bh) { 66438ac3939dSRitesh Harjani if (block) 66448ac3939dSRitesh Harjani BUG_ON(block != bh->b_blocknr); 66458ac3939dSRitesh Harjani else 66468ac3939dSRitesh Harjani block = bh->b_blocknr; 66478ac3939dSRitesh Harjani } 66488ac3939dSRitesh Harjani 664911b6890bSKemeng Shi if (sbi->s_mount_state & EXT4_FC_REPLAY) { 66502ec6d0a5SKemeng Shi ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count)); 665111b6890bSKemeng Shi return; 665211b6890bSKemeng Shi } 665311b6890bSKemeng Shi 665411b6890bSKemeng Shi might_sleep(); 665511b6890bSKemeng Shi 66568ac3939dSRitesh Harjani if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 66578ac3939dSRitesh Harjani !ext4_inode_block_valid(inode, block, count)) { 66588ac3939dSRitesh Harjani ext4_error(sb, "Freeing blocks not in datazone - " 66598ac3939dSRitesh Harjani "block = %llu, count = %lu", block, count); 66608ac3939dSRitesh Harjani return; 66618ac3939dSRitesh Harjani } 66621e1c2b86SLukas Czerner flags |= EXT4_FREE_BLOCKS_VALIDATED; 66638ac3939dSRitesh Harjani 66648ac3939dSRitesh Harjani ext4_debug("freeing block %llu\n", block); 66658ac3939dSRitesh Harjani trace_ext4_free_blocks(inode, block, count, flags); 66668ac3939dSRitesh Harjani 66678ac3939dSRitesh Harjani if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 66688ac3939dSRitesh Harjani BUG_ON(count > 1); 66698ac3939dSRitesh Harjani 66708ac3939dSRitesh Harjani ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, 66718ac3939dSRitesh Harjani inode, bh, block); 66728ac3939dSRitesh Harjani } 66738ac3939dSRitesh Harjani 66748ac3939dSRitesh Harjani /* 66758ac3939dSRitesh Harjani * If the extent to be freed does not begin on a cluster 66768ac3939dSRitesh Harjani * boundary, we need to deal with partial clusters at the 66778ac3939dSRitesh Harjani * beginning and end of the extent. Normally we will free 66788ac3939dSRitesh Harjani * blocks at the beginning or the end unless we are explicitly 66798ac3939dSRitesh Harjani * requested to avoid doing so. 66808ac3939dSRitesh Harjani */ 66818ac3939dSRitesh Harjani overflow = EXT4_PBLK_COFF(sbi, block); 66828ac3939dSRitesh Harjani if (overflow) { 66838ac3939dSRitesh Harjani if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { 66848ac3939dSRitesh Harjani overflow = sbi->s_cluster_ratio - overflow; 66858ac3939dSRitesh Harjani block += overflow; 66868ac3939dSRitesh Harjani if (count > overflow) 66878ac3939dSRitesh Harjani count -= overflow; 66888ac3939dSRitesh Harjani else 66898ac3939dSRitesh Harjani return; 66908ac3939dSRitesh Harjani } else { 66918ac3939dSRitesh Harjani block -= overflow; 66928ac3939dSRitesh Harjani count += overflow; 66938ac3939dSRitesh Harjani } 66941e1c2b86SLukas Czerner /* The range changed so it's no longer validated */ 66951e1c2b86SLukas Czerner flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 66968ac3939dSRitesh Harjani } 66978ac3939dSRitesh Harjani overflow = EXT4_LBLK_COFF(sbi, count); 66988ac3939dSRitesh Harjani if (overflow) { 66998ac3939dSRitesh Harjani if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { 67008ac3939dSRitesh Harjani if (count > overflow) 67018ac3939dSRitesh Harjani count -= overflow; 67028ac3939dSRitesh Harjani else 67038ac3939dSRitesh Harjani return; 67048ac3939dSRitesh Harjani } else 67058ac3939dSRitesh Harjani count += sbi->s_cluster_ratio - overflow; 67061e1c2b86SLukas Czerner /* The range changed so it's no longer validated */ 67071e1c2b86SLukas Czerner flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 67088ac3939dSRitesh Harjani } 67098ac3939dSRitesh Harjani 67108ac3939dSRitesh Harjani if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 67118ac3939dSRitesh Harjani int i; 67128ac3939dSRitesh Harjani int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; 67138ac3939dSRitesh Harjani 67148ac3939dSRitesh Harjani for (i = 0; i < count; i++) { 67158ac3939dSRitesh Harjani cond_resched(); 67168ac3939dSRitesh Harjani if (is_metadata) 67178ac3939dSRitesh Harjani bh = sb_find_get_block(inode->i_sb, block + i); 67188ac3939dSRitesh Harjani ext4_forget(handle, is_metadata, inode, bh, block + i); 67198ac3939dSRitesh Harjani } 67208ac3939dSRitesh Harjani } 67218ac3939dSRitesh Harjani 67228ac3939dSRitesh Harjani ext4_mb_clear_bb(handle, inode, block, count, flags); 67238ac3939dSRitesh Harjani } 67248ac3939dSRitesh Harjani 67258ac3939dSRitesh Harjani /** 67260529155eSYongqiang Yang * ext4_group_add_blocks() -- Add given blocks to an existing group 67272846e820SAmir Goldstein * @handle: handle to this transaction 67282846e820SAmir Goldstein * @sb: super block 67294907cb7bSAnatol Pomozov * @block: start physical block to add to the block group 67302846e820SAmir Goldstein * @count: number of blocks to free 67312846e820SAmir Goldstein * 6732e73a347bSAmir Goldstein * This marks the blocks as free in the bitmap and buddy. 67332846e820SAmir Goldstein */ 6734cc7365dfSYongqiang Yang int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, 67352846e820SAmir Goldstein ext4_fsblk_t block, unsigned long count) 67362846e820SAmir Goldstein { 67372846e820SAmir Goldstein struct buffer_head *bitmap_bh = NULL; 67382846e820SAmir Goldstein struct buffer_head *gd_bh; 67392846e820SAmir Goldstein ext4_group_t block_group; 67402846e820SAmir Goldstein ext4_grpblk_t bit; 67412846e820SAmir Goldstein unsigned int i; 67422846e820SAmir Goldstein struct ext4_group_desc *desc; 67432846e820SAmir Goldstein struct ext4_sb_info *sbi = EXT4_SB(sb); 6744e73a347bSAmir Goldstein struct ext4_buddy e4b; 6745d77147ffSharshads int err = 0, ret, free_clusters_count; 6746d77147ffSharshads ext4_grpblk_t clusters_freed; 6747d77147ffSharshads ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); 6748d77147ffSharshads ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); 6749d77147ffSharshads unsigned long cluster_count = last_cluster - first_cluster + 1; 67502846e820SAmir Goldstein 67512846e820SAmir Goldstein ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); 67522846e820SAmir Goldstein 67534740b830SYongqiang Yang if (count == 0) 67544740b830SYongqiang Yang return 0; 67554740b830SYongqiang Yang 67562846e820SAmir Goldstein ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 67572846e820SAmir Goldstein /* 67582846e820SAmir Goldstein * Check to see if we are freeing blocks across a group 67592846e820SAmir Goldstein * boundary. 67602846e820SAmir Goldstein */ 6761d77147ffSharshads if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { 6762d77147ffSharshads ext4_warning(sb, "too many blocks added to group %u", 6763cc7365dfSYongqiang Yang block_group); 6764cc7365dfSYongqiang Yang err = -EINVAL; 67652846e820SAmir Goldstein goto error_return; 6766cc7365dfSYongqiang Yang } 67672cd05cc3STheodore Ts'o 67682846e820SAmir Goldstein bitmap_bh = ext4_read_block_bitmap(sb, block_group); 67699008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 67709008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 67719008a58eSDarrick J. Wong bitmap_bh = NULL; 67722846e820SAmir Goldstein goto error_return; 6773cc7365dfSYongqiang Yang } 6774cc7365dfSYongqiang Yang 67752846e820SAmir Goldstein desc = ext4_get_group_desc(sb, block_group, &gd_bh); 6776cc7365dfSYongqiang Yang if (!desc) { 6777cc7365dfSYongqiang Yang err = -EIO; 67782846e820SAmir Goldstein goto error_return; 6779cc7365dfSYongqiang Yang } 67802846e820SAmir Goldstein 6781a00b482bSRitesh Harjani if (!ext4_sb_block_valid(sb, NULL, block, count)) { 67822846e820SAmir Goldstein ext4_error(sb, "Adding blocks in system zones - " 67832846e820SAmir Goldstein "Block = %llu, count = %lu", 67842846e820SAmir Goldstein block, count); 6785cc7365dfSYongqiang Yang err = -EINVAL; 67862846e820SAmir Goldstein goto error_return; 67872846e820SAmir Goldstein } 67882846e820SAmir Goldstein 67892cd05cc3STheodore Ts'o BUFFER_TRACE(bitmap_bh, "getting write access"); 6790188c299eSJan Kara err = ext4_journal_get_write_access(handle, sb, bitmap_bh, 6791188c299eSJan Kara EXT4_JTR_NONE); 67922846e820SAmir Goldstein if (err) 67932846e820SAmir Goldstein goto error_return; 67942846e820SAmir Goldstein 67952846e820SAmir Goldstein /* 67962846e820SAmir Goldstein * We are about to modify some metadata. Call the journal APIs 67972846e820SAmir Goldstein * to unshare ->b_data if a currently-committing transaction is 67982846e820SAmir Goldstein * using it 67992846e820SAmir Goldstein */ 68002846e820SAmir Goldstein BUFFER_TRACE(gd_bh, "get_write_access"); 6801188c299eSJan Kara err = ext4_journal_get_write_access(handle, sb, gd_bh, EXT4_JTR_NONE); 68022846e820SAmir Goldstein if (err) 68032846e820SAmir Goldstein goto error_return; 6804e73a347bSAmir Goldstein 6805d77147ffSharshads for (i = 0, clusters_freed = 0; i < cluster_count; i++) { 68062846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "clear bit"); 6807e73a347bSAmir Goldstein if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { 68082846e820SAmir Goldstein ext4_error(sb, "bit already cleared for block %llu", 68092846e820SAmir Goldstein (ext4_fsblk_t)(block + i)); 68102846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "bit already cleared"); 68112846e820SAmir Goldstein } else { 6812d77147ffSharshads clusters_freed++; 68132846e820SAmir Goldstein } 68142846e820SAmir Goldstein } 6815e73a347bSAmir Goldstein 6816e73a347bSAmir Goldstein err = ext4_mb_load_buddy(sb, block_group, &e4b); 6817e73a347bSAmir Goldstein if (err) 6818e73a347bSAmir Goldstein goto error_return; 6819e73a347bSAmir Goldstein 6820e73a347bSAmir Goldstein /* 6821e73a347bSAmir Goldstein * need to update group_info->bb_free and bitmap 6822e73a347bSAmir Goldstein * with group lock held. generate_buddy look at 6823e73a347bSAmir Goldstein * them with group lock_held 6824e73a347bSAmir Goldstein */ 68252846e820SAmir Goldstein ext4_lock_group(sb, block_group); 6826d77147ffSharshads mb_clear_bits(bitmap_bh->b_data, bit, cluster_count); 6827d77147ffSharshads mb_free_blocks(NULL, &e4b, bit, cluster_count); 6828d77147ffSharshads free_clusters_count = clusters_freed + 6829d77147ffSharshads ext4_free_group_clusters(sb, desc); 6830d77147ffSharshads ext4_free_group_clusters_set(sb, desc, free_clusters_count); 68311df9bde4SKemeng Shi ext4_block_bitmap_csum_set(sb, desc, bitmap_bh); 6832feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, block_group, desc); 68332846e820SAmir Goldstein ext4_unlock_group(sb, block_group); 683457042651STheodore Ts'o percpu_counter_add(&sbi->s_freeclusters_counter, 6835d77147ffSharshads clusters_freed); 68362846e820SAmir Goldstein 68372846e820SAmir Goldstein if (sbi->s_log_groups_per_flex) { 68382846e820SAmir Goldstein ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 6839d77147ffSharshads atomic64_add(clusters_freed, 68407c990728SSuraj Jitindar Singh &sbi_array_rcu_deref(sbi, s_flex_groups, 68417c990728SSuraj Jitindar Singh flex_group)->free_clusters); 68422846e820SAmir Goldstein } 6843e73a347bSAmir Goldstein 6844e73a347bSAmir Goldstein ext4_mb_unload_buddy(&e4b); 68452846e820SAmir Goldstein 68462846e820SAmir Goldstein /* We dirtied the bitmap block */ 68472846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 68482846e820SAmir Goldstein err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 68492846e820SAmir Goldstein 68502846e820SAmir Goldstein /* And the group descriptor block */ 68512846e820SAmir Goldstein BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 68522846e820SAmir Goldstein ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 68532846e820SAmir Goldstein if (!err) 68542846e820SAmir Goldstein err = ret; 68552846e820SAmir Goldstein 68562846e820SAmir Goldstein error_return: 68572846e820SAmir Goldstein brelse(bitmap_bh); 68582846e820SAmir Goldstein ext4_std_error(sb, err); 6859cc7365dfSYongqiang Yang return err; 68602846e820SAmir Goldstein } 68612846e820SAmir Goldstein 68622846e820SAmir Goldstein /** 68637360d173SLukas Czerner * ext4_trim_extent -- function to TRIM one single free extent in the group 68647360d173SLukas Czerner * @sb: super block for the file system 68657360d173SLukas Czerner * @start: starting block of the free extent in the alloc. group 68667360d173SLukas Czerner * @count: number of blocks to TRIM 68677360d173SLukas Czerner * @e4b: ext4 buddy for the group 68687360d173SLukas Czerner * 68697360d173SLukas Czerner * Trim "count" blocks starting at "start" in the "group". To assure that no 68707360d173SLukas Czerner * one will allocate those blocks, mark it as used in buddy bitmap. This must 68717360d173SLukas Czerner * be called with under the group lock. 68727360d173SLukas Czerner */ 6873bd2eea8dSWang Jianchao static int ext4_trim_extent(struct super_block *sb, 6874bd2eea8dSWang Jianchao int start, int count, struct ext4_buddy *e4b) 6875e2cbd587Sjon ernst __releases(bitlock) 6876e2cbd587Sjon ernst __acquires(bitlock) 68777360d173SLukas Czerner { 68787360d173SLukas Czerner struct ext4_free_extent ex; 6879bd2eea8dSWang Jianchao ext4_group_t group = e4b->bd_group; 6880d71c1ae2SLukas Czerner int ret = 0; 68817360d173SLukas Czerner 6882b3d4c2b1STao Ma trace_ext4_trim_extent(sb, group, start, count); 6883b3d4c2b1STao Ma 68847360d173SLukas Czerner assert_spin_locked(ext4_group_lock_ptr(sb, group)); 68857360d173SLukas Czerner 68867360d173SLukas Czerner ex.fe_start = start; 68877360d173SLukas Czerner ex.fe_group = group; 68887360d173SLukas Czerner ex.fe_len = count; 68897360d173SLukas Czerner 68907360d173SLukas Czerner /* 68917360d173SLukas Czerner * Mark blocks used, so no one can reuse them while 68927360d173SLukas Czerner * being trimmed. 68937360d173SLukas Czerner */ 68947360d173SLukas Czerner mb_mark_used(e4b, &ex); 68957360d173SLukas Czerner ext4_unlock_group(sb, group); 6896a0154344SDaeho Jeong ret = ext4_issue_discard(sb, group, start, count, NULL); 68977360d173SLukas Czerner ext4_lock_group(sb, group); 68987360d173SLukas Czerner mb_free_blocks(NULL, e4b, start, ex.fe_len); 6899d71c1ae2SLukas Czerner return ret; 69007360d173SLukas Czerner } 69017360d173SLukas Czerner 690245e4ab32SJan Kara static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb, 690345e4ab32SJan Kara ext4_group_t grp) 690445e4ab32SJan Kara { 690545e4ab32SJan Kara if (grp < ext4_get_groups_count(sb)) 690645e4ab32SJan Kara return EXT4_CLUSTERS_PER_GROUP(sb) - 1; 690745e4ab32SJan Kara return (ext4_blocks_count(EXT4_SB(sb)->s_es) - 690845e4ab32SJan Kara ext4_group_first_block_no(sb, grp) - 1) >> 690945e4ab32SJan Kara EXT4_CLUSTER_BITS(sb); 691045e4ab32SJan Kara } 691145e4ab32SJan Kara 69125229a658SJan Kara static bool ext4_trim_interrupted(void) 69135229a658SJan Kara { 69145229a658SJan Kara return fatal_signal_pending(current) || freezing(current); 69155229a658SJan Kara } 69165229a658SJan Kara 69176920b391SWang Jianchao static int ext4_try_to_trim_range(struct super_block *sb, 69186920b391SWang Jianchao struct ext4_buddy *e4b, ext4_grpblk_t start, 69196920b391SWang Jianchao ext4_grpblk_t max, ext4_grpblk_t minblocks) 6920a5fda113STheodore Ts'o __acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) 6921a5fda113STheodore Ts'o __releases(ext4_group_lock_ptr(sb, e4b->bd_group)) 69226920b391SWang Jianchao { 69236920b391SWang Jianchao ext4_grpblk_t next, count, free_count; 692445e4ab32SJan Kara bool set_trimmed = false; 69256920b391SWang Jianchao void *bitmap; 69266920b391SWang Jianchao 69276920b391SWang Jianchao bitmap = e4b->bd_bitmap; 692845e4ab32SJan Kara if (start == 0 && max >= ext4_last_grp_cluster(sb, e4b->bd_group)) 692945e4ab32SJan Kara set_trimmed = true; 6930de8bf0e5SKemeng Shi start = max(e4b->bd_info->bb_first_free, start); 69316920b391SWang Jianchao count = 0; 69326920b391SWang Jianchao free_count = 0; 69336920b391SWang Jianchao 69346920b391SWang Jianchao while (start <= max) { 69356920b391SWang Jianchao start = mb_find_next_zero_bit(bitmap, max + 1, start); 69366920b391SWang Jianchao if (start > max) 69376920b391SWang Jianchao break; 69386920b391SWang Jianchao next = mb_find_next_bit(bitmap, max + 1, start); 69396920b391SWang Jianchao 69406920b391SWang Jianchao if ((next - start) >= minblocks) { 6941afcc4e32SLukas Bulwahn int ret = ext4_trim_extent(sb, start, next - start, e4b); 6942afcc4e32SLukas Bulwahn 69436920b391SWang Jianchao if (ret && ret != -EOPNOTSUPP) 694445e4ab32SJan Kara return count; 69456920b391SWang Jianchao count += next - start; 69466920b391SWang Jianchao } 69476920b391SWang Jianchao free_count += next - start; 69486920b391SWang Jianchao start = next + 1; 69496920b391SWang Jianchao 69505229a658SJan Kara if (ext4_trim_interrupted()) 69515229a658SJan Kara return count; 69526920b391SWang Jianchao 69536920b391SWang Jianchao if (need_resched()) { 69546920b391SWang Jianchao ext4_unlock_group(sb, e4b->bd_group); 69556920b391SWang Jianchao cond_resched(); 69566920b391SWang Jianchao ext4_lock_group(sb, e4b->bd_group); 69576920b391SWang Jianchao } 69586920b391SWang Jianchao 69596920b391SWang Jianchao if ((e4b->bd_info->bb_free - free_count) < minblocks) 69606920b391SWang Jianchao break; 69616920b391SWang Jianchao } 69626920b391SWang Jianchao 696345e4ab32SJan Kara if (set_trimmed) 696445e4ab32SJan Kara EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info); 696545e4ab32SJan Kara 69666920b391SWang Jianchao return count; 69676920b391SWang Jianchao } 69686920b391SWang Jianchao 69697360d173SLukas Czerner /** 69707360d173SLukas Czerner * ext4_trim_all_free -- function to trim all free space in alloc. group 69717360d173SLukas Czerner * @sb: super block for file system 697222612283STao Ma * @group: group to be trimmed 69737360d173SLukas Czerner * @start: first group block to examine 69747360d173SLukas Czerner * @max: last group block to examine 69757360d173SLukas Czerner * @minblocks: minimum extent block count 69767360d173SLukas Czerner * 69777360d173SLukas Czerner * ext4_trim_all_free walks through group's block bitmap searching for free 69787360d173SLukas Czerner * extents. When the free extent is found, mark it as used in group buddy 69797360d173SLukas Czerner * bitmap. Then issue a TRIM command on this extent and free the extent in 6980b6f5558cSWang Jianchao * the group buddy bitmap. 69817360d173SLukas Czerner */ 69820b75a840SLukas Czerner static ext4_grpblk_t 698378944086SLukas Czerner ext4_trim_all_free(struct super_block *sb, ext4_group_t group, 698478944086SLukas Czerner ext4_grpblk_t start, ext4_grpblk_t max, 698545e4ab32SJan Kara ext4_grpblk_t minblocks) 69867360d173SLukas Czerner { 698778944086SLukas Czerner struct ext4_buddy e4b; 69886920b391SWang Jianchao int ret; 69897360d173SLukas Czerner 6990b3d4c2b1STao Ma trace_ext4_trim_all_free(sb, group, start, max); 6991b3d4c2b1STao Ma 699278944086SLukas Czerner ret = ext4_mb_load_buddy(sb, group, &e4b); 699378944086SLukas Czerner if (ret) { 69949651e6b2SKonstantin Khlebnikov ext4_warning(sb, "Error %d loading buddy information for %u", 69959651e6b2SKonstantin Khlebnikov ret, group); 699678944086SLukas Czerner return ret; 699778944086SLukas Czerner } 699828739eeaSLukas Czerner 699928739eeaSLukas Czerner ext4_lock_group(sb, group); 70003d56b8d2STao Ma 70016920b391SWang Jianchao if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || 700245e4ab32SJan Kara minblocks < EXT4_SB(sb)->s_last_trim_minblks) 70036920b391SWang Jianchao ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); 700445e4ab32SJan Kara else 70056920b391SWang Jianchao ret = 0; 70066920b391SWang Jianchao 70077360d173SLukas Czerner ext4_unlock_group(sb, group); 700878944086SLukas Czerner ext4_mb_unload_buddy(&e4b); 70097360d173SLukas Czerner 70107360d173SLukas Czerner ext4_debug("trimmed %d blocks in the group %d\n", 70116920b391SWang Jianchao ret, group); 70127360d173SLukas Czerner 7013d71c1ae2SLukas Czerner return ret; 70147360d173SLukas Czerner } 70157360d173SLukas Czerner 70167360d173SLukas Czerner /** 70177360d173SLukas Czerner * ext4_trim_fs() -- trim ioctl handle function 70187360d173SLukas Czerner * @sb: superblock for filesystem 70197360d173SLukas Czerner * @range: fstrim_range structure 70207360d173SLukas Czerner * 70217360d173SLukas Czerner * start: First Byte to trim 70227360d173SLukas Czerner * len: number of Bytes to trim from start 70237360d173SLukas Czerner * minlen: minimum extent length in Bytes 70247360d173SLukas Czerner * ext4_trim_fs goes through all allocation groups containing Bytes from 70257360d173SLukas Czerner * start to start+len. For each such a group ext4_trim_all_free function 70267360d173SLukas Czerner * is invoked to trim all free space. 70277360d173SLukas Czerner */ 70287360d173SLukas Czerner int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) 70297360d173SLukas Czerner { 70307b47ef52SChristoph Hellwig unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev); 703178944086SLukas Czerner struct ext4_group_info *grp; 7032913eed83SLukas Czerner ext4_group_t group, first_group, last_group; 70337137d7a4STheodore Ts'o ext4_grpblk_t cnt = 0, first_cluster, last_cluster; 7034913eed83SLukas Czerner uint64_t start, end, minlen, trimmed = 0; 70350f0a25bfSJan Kara ext4_fsblk_t first_data_blk = 70360f0a25bfSJan Kara le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 7037913eed83SLukas Czerner ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); 70387360d173SLukas Czerner int ret = 0; 70397360d173SLukas Czerner 70407360d173SLukas Czerner start = range->start >> sb->s_blocksize_bits; 7041913eed83SLukas Czerner end = start + (range->len >> sb->s_blocksize_bits) - 1; 7042aaf7d73eSLukas Czerner minlen = EXT4_NUM_B2C(EXT4_SB(sb), 7043aaf7d73eSLukas Czerner range->minlen >> sb->s_blocksize_bits); 70447360d173SLukas Czerner 70455de35e8dSLukas Czerner if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || 70465de35e8dSLukas Czerner start >= max_blks || 70475de35e8dSLukas Czerner range->len < sb->s_blocksize) 70487360d173SLukas Czerner return -EINVAL; 7049173b6e38SJan Kara /* No point to try to trim less than discard granularity */ 70507b47ef52SChristoph Hellwig if (range->minlen < discard_granularity) { 7051173b6e38SJan Kara minlen = EXT4_NUM_B2C(EXT4_SB(sb), 70527b47ef52SChristoph Hellwig discard_granularity >> sb->s_blocksize_bits); 7053173b6e38SJan Kara if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) 7054173b6e38SJan Kara goto out; 7055173b6e38SJan Kara } 705645e4ab32SJan Kara if (end >= max_blks - 1) 7057913eed83SLukas Czerner end = max_blks - 1; 7058913eed83SLukas Czerner if (end <= first_data_blk) 705922f10457STao Ma goto out; 7060913eed83SLukas Czerner if (start < first_data_blk) 70610f0a25bfSJan Kara start = first_data_blk; 70627360d173SLukas Czerner 7063913eed83SLukas Czerner /* Determine first and last group to examine based on start and end */ 70647360d173SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, 70657137d7a4STheodore Ts'o &first_group, &first_cluster); 7066913eed83SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, 70677137d7a4STheodore Ts'o &last_group, &last_cluster); 70687360d173SLukas Czerner 7069913eed83SLukas Czerner /* end now represents the last cluster to discard in this group */ 7070913eed83SLukas Czerner end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 70717360d173SLukas Czerner 70727360d173SLukas Czerner for (group = first_group; group <= last_group; group++) { 70735229a658SJan Kara if (ext4_trim_interrupted()) 70745229a658SJan Kara break; 707578944086SLukas Czerner grp = ext4_get_group_info(sb, group); 70765354b2afSTheodore Ts'o if (!grp) 70775354b2afSTheodore Ts'o continue; 707878944086SLukas Czerner /* We only do this if the grp has never been initialized */ 707978944086SLukas Czerner if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 7080adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_group(sb, group, GFP_NOFS); 708178944086SLukas Czerner if (ret) 70827360d173SLukas Czerner break; 70837360d173SLukas Czerner } 70847360d173SLukas Czerner 70850ba08517STao Ma /* 7086913eed83SLukas Czerner * For all the groups except the last one, last cluster will 7087913eed83SLukas Czerner * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to 7088913eed83SLukas Czerner * change it for the last group, note that last_cluster is 7089913eed83SLukas Czerner * already computed earlier by ext4_get_group_no_and_offset() 70900ba08517STao Ma */ 709145e4ab32SJan Kara if (group == last_group) 7092913eed83SLukas Czerner end = last_cluster; 709378944086SLukas Czerner if (grp->bb_free >= minlen) { 70947137d7a4STheodore Ts'o cnt = ext4_trim_all_free(sb, group, first_cluster, 709545e4ab32SJan Kara end, minlen); 70967360d173SLukas Czerner if (cnt < 0) { 70977360d173SLukas Czerner ret = cnt; 70987360d173SLukas Czerner break; 70997360d173SLukas Czerner } 71007360d173SLukas Czerner trimmed += cnt; 710121e7fd22SLukas Czerner } 7102913eed83SLukas Czerner 7103913eed83SLukas Czerner /* 7104913eed83SLukas Czerner * For every group except the first one, we are sure 7105913eed83SLukas Czerner * that the first cluster to discard will be cluster #0. 7106913eed83SLukas Czerner */ 71077137d7a4STheodore Ts'o first_cluster = 0; 71087360d173SLukas Czerner } 71097360d173SLukas Czerner 71103d56b8d2STao Ma if (!ret) 71112327fb2eSLukas Czerner EXT4_SB(sb)->s_last_trim_minblks = minlen; 71123d56b8d2STao Ma 711322f10457STao Ma out: 7114aaf7d73eSLukas Czerner range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; 71157360d173SLukas Czerner return ret; 71167360d173SLukas Czerner } 71170c9ec4beSDarrick J. Wong 71180c9ec4beSDarrick J. Wong /* Iterate all the free extents in the group. */ 71190c9ec4beSDarrick J. Wong int 71200c9ec4beSDarrick J. Wong ext4_mballoc_query_range( 71210c9ec4beSDarrick J. Wong struct super_block *sb, 71220c9ec4beSDarrick J. Wong ext4_group_t group, 71230c9ec4beSDarrick J. Wong ext4_grpblk_t start, 71240c9ec4beSDarrick J. Wong ext4_grpblk_t end, 71250c9ec4beSDarrick J. Wong ext4_mballoc_query_range_fn formatter, 71260c9ec4beSDarrick J. Wong void *priv) 71270c9ec4beSDarrick J. Wong { 71280c9ec4beSDarrick J. Wong void *bitmap; 71290c9ec4beSDarrick J. Wong ext4_grpblk_t next; 71300c9ec4beSDarrick J. Wong struct ext4_buddy e4b; 71310c9ec4beSDarrick J. Wong int error; 71320c9ec4beSDarrick J. Wong 71330c9ec4beSDarrick J. Wong error = ext4_mb_load_buddy(sb, group, &e4b); 71340c9ec4beSDarrick J. Wong if (error) 71350c9ec4beSDarrick J. Wong return error; 71360c9ec4beSDarrick J. Wong bitmap = e4b.bd_bitmap; 71370c9ec4beSDarrick J. Wong 71380c9ec4beSDarrick J. Wong ext4_lock_group(sb, group); 71390c9ec4beSDarrick J. Wong 7140de8bf0e5SKemeng Shi start = max(e4b.bd_info->bb_first_free, start); 71410c9ec4beSDarrick J. Wong if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) 71420c9ec4beSDarrick J. Wong end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 71430c9ec4beSDarrick J. Wong 71440c9ec4beSDarrick J. Wong while (start <= end) { 71450c9ec4beSDarrick J. Wong start = mb_find_next_zero_bit(bitmap, end + 1, start); 71460c9ec4beSDarrick J. Wong if (start > end) 71470c9ec4beSDarrick J. Wong break; 71480c9ec4beSDarrick J. Wong next = mb_find_next_bit(bitmap, end + 1, start); 71490c9ec4beSDarrick J. Wong 71500c9ec4beSDarrick J. Wong ext4_unlock_group(sb, group); 71510c9ec4beSDarrick J. Wong error = formatter(sb, group, start, next - start, priv); 71520c9ec4beSDarrick J. Wong if (error) 71530c9ec4beSDarrick J. Wong goto out_unload; 71540c9ec4beSDarrick J. Wong ext4_lock_group(sb, group); 71550c9ec4beSDarrick J. Wong 71560c9ec4beSDarrick J. Wong start = next + 1; 71570c9ec4beSDarrick J. Wong } 71580c9ec4beSDarrick J. Wong 71590c9ec4beSDarrick J. Wong ext4_unlock_group(sb, group); 71600c9ec4beSDarrick J. Wong out_unload: 71610c9ec4beSDarrick J. Wong ext4_mb_unload_buddy(&e4b); 71620c9ec4beSDarrick J. Wong 71630c9ec4beSDarrick J. Wong return error; 71640c9ec4beSDarrick J. Wong } 7165