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> 21bdefd689SKemeng Shi #include <kunit/static_stub.h> 229bffad1eSTheodore Ts'o 23c9de560dSAlex Tomas /* 24c9de560dSAlex Tomas * MUSTDO: 25c9de560dSAlex Tomas * - test ext4_ext_search_left() and ext4_ext_search_right() 26c9de560dSAlex Tomas * - search for metadata in few groups 27c9de560dSAlex Tomas * 28c9de560dSAlex Tomas * TODO v4: 29c9de560dSAlex Tomas * - normalization should take into account whether file is still open 30c9de560dSAlex Tomas * - discard preallocations if no free space left (policy?) 31c9de560dSAlex Tomas * - don't normalize tails 32c9de560dSAlex Tomas * - quota 33c9de560dSAlex Tomas * - reservation for superuser 34c9de560dSAlex Tomas * 35c9de560dSAlex Tomas * TODO v3: 36c9de560dSAlex Tomas * - bitmap read-ahead (proposed by Oleg Drokin aka green) 37c9de560dSAlex Tomas * - track min/max extents in each group for better group selection 38c9de560dSAlex Tomas * - mb_mark_used() may allocate chunk right after splitting buddy 39c9de560dSAlex Tomas * - tree of groups sorted by number of free blocks 40c9de560dSAlex Tomas * - error handling 41c9de560dSAlex Tomas */ 42c9de560dSAlex Tomas 43c9de560dSAlex Tomas /* 44c9de560dSAlex Tomas * The allocation request involve request for multiple number of blocks 45c9de560dSAlex Tomas * near to the goal(block) value specified. 46c9de560dSAlex Tomas * 47b713a5ecSTheodore Ts'o * During initialization phase of the allocator we decide to use the 48b713a5ecSTheodore Ts'o * group preallocation or inode preallocation depending on the size of 49b713a5ecSTheodore Ts'o * the file. The size of the file could be the resulting file size we 50b713a5ecSTheodore Ts'o * would have after allocation, or the current file size, which ever 51b713a5ecSTheodore Ts'o * is larger. If the size is less than sbi->s_mb_stream_request we 52b713a5ecSTheodore Ts'o * select to use the group preallocation. The default value of 53b713a5ecSTheodore Ts'o * s_mb_stream_request is 16 blocks. This can also be tuned via 54b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 55b713a5ecSTheodore Ts'o * terms of number of blocks. 56c9de560dSAlex Tomas * 57c9de560dSAlex Tomas * The main motivation for having small file use group preallocation is to 58b713a5ecSTheodore Ts'o * ensure that we have small files closer together on the disk. 59c9de560dSAlex Tomas * 60b713a5ecSTheodore Ts'o * First stage the allocator looks at the inode prealloc list, 61b713a5ecSTheodore Ts'o * ext4_inode_info->i_prealloc_list, which contains list of prealloc 62b713a5ecSTheodore Ts'o * spaces for this particular inode. The inode prealloc space is 63b713a5ecSTheodore Ts'o * represented as: 64c9de560dSAlex Tomas * 65c9de560dSAlex Tomas * pa_lstart -> the logical start block for this prealloc space 66c9de560dSAlex Tomas * pa_pstart -> the physical start block for this prealloc space 6753accfa9STheodore Ts'o * pa_len -> length for this prealloc space (in clusters) 6853accfa9STheodore Ts'o * pa_free -> free space available in this prealloc space (in clusters) 69c9de560dSAlex Tomas * 70c9de560dSAlex Tomas * The inode preallocation space is used looking at the _logical_ start 71c9de560dSAlex Tomas * block. If only the logical file block falls within the range of prealloc 72caaf7a29STao Ma * space we will consume the particular prealloc space. This makes sure that 73caaf7a29STao Ma * we have contiguous physical blocks representing the file blocks 74c9de560dSAlex Tomas * 75c9de560dSAlex Tomas * The important thing to be noted in case of inode prealloc space is that 76c9de560dSAlex Tomas * we don't modify the values associated to inode prealloc space except 77c9de560dSAlex Tomas * pa_free. 78c9de560dSAlex Tomas * 79c9de560dSAlex Tomas * If we are not able to find blocks in the inode prealloc space and if we 80c9de560dSAlex Tomas * have the group allocation flag set then we look at the locality group 81caaf7a29STao Ma * prealloc space. These are per CPU prealloc list represented as 82c9de560dSAlex Tomas * 83c9de560dSAlex Tomas * ext4_sb_info.s_locality_groups[smp_processor_id()] 84c9de560dSAlex Tomas * 85c9de560dSAlex Tomas * The reason for having a per cpu locality group is to reduce the contention 86c9de560dSAlex Tomas * between CPUs. It is possible to get scheduled at this point. 87c9de560dSAlex Tomas * 88c9de560dSAlex Tomas * The locality group prealloc space is used looking at whether we have 8925985edcSLucas De Marchi * enough free space (pa_free) within the prealloc space. 90c9de560dSAlex Tomas * 91c9de560dSAlex Tomas * If we can't allocate blocks via inode prealloc or/and locality group 92c9de560dSAlex Tomas * prealloc then we look at the buddy cache. The buddy cache is represented 93c9de560dSAlex Tomas * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 94c9de560dSAlex Tomas * mapped to the buddy and bitmap information regarding different 95c9de560dSAlex Tomas * groups. The buddy information is attached to buddy cache inode so that 96c9de560dSAlex Tomas * we can access them through the page cache. The information regarding 97c9de560dSAlex Tomas * each group is loaded via ext4_mb_load_buddy. The information involve 98c9de560dSAlex Tomas * block bitmap and buddy information. The information are stored in the 99c9de560dSAlex Tomas * inode as: 100c9de560dSAlex Tomas * 101c9de560dSAlex Tomas * { page } 102c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 103c9de560dSAlex Tomas * 104c9de560dSAlex Tomas * 105c9de560dSAlex Tomas * one block each for bitmap and buddy information. So for each group we 106ea1754a0SKirill A. Shutemov * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / 107c9de560dSAlex Tomas * blocksize) blocks. So it can have information regarding groups_per_page 108c9de560dSAlex Tomas * which is blocks_per_page/2 109c9de560dSAlex Tomas * 110c9de560dSAlex Tomas * The buddy cache inode is not stored on disk. The inode is thrown 111c9de560dSAlex Tomas * away when the filesystem is unmounted. 112c9de560dSAlex Tomas * 113c9de560dSAlex Tomas * We look for count number of blocks in the buddy cache. If we were able 114c9de560dSAlex Tomas * to locate that many free blocks we return with additional information 115c9de560dSAlex Tomas * regarding rest of the contiguous physical block available 116c9de560dSAlex Tomas * 117c9de560dSAlex Tomas * Before allocating blocks via buddy cache we normalize the request 118c9de560dSAlex Tomas * blocks. This ensure we ask for more blocks that we needed. The extra 119c9de560dSAlex Tomas * blocks that we get after allocation is added to the respective prealloc 120c9de560dSAlex Tomas * list. In case of inode preallocation we follow a list of heuristics 121c9de560dSAlex Tomas * based on file size. This can be found in ext4_mb_normalize_request. If 122c9de560dSAlex Tomas * we are doing a group prealloc we try to normalize the request to 12327baebb8STheodore Ts'o * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is 12427baebb8STheodore Ts'o * dependent on the cluster size; for non-bigalloc file systems, it is 125c9de560dSAlex Tomas * 512 blocks. This can be tuned via 126d7a1fee1SDan Ehrenberg * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in 127c9de560dSAlex Tomas * terms of number of blocks. If we have mounted the file system with -O 128c9de560dSAlex Tomas * stripe=<value> option the group prealloc request is normalized to the 129b483bb77SRandy Dunlap * smallest multiple of the stripe value (sbi->s_stripe) which is 130d7a1fee1SDan Ehrenberg * greater than the default mb_group_prealloc. 131c9de560dSAlex Tomas * 132196e402aSHarshad Shirwadkar * If "mb_optimize_scan" mount option is set, we maintain in memory group info 133196e402aSHarshad Shirwadkar * structures in two data structures: 134196e402aSHarshad Shirwadkar * 135196e402aSHarshad Shirwadkar * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders) 136196e402aSHarshad Shirwadkar * 137196e402aSHarshad Shirwadkar * Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks) 138196e402aSHarshad Shirwadkar * 139196e402aSHarshad Shirwadkar * This is an array of lists where the index in the array represents the 140196e402aSHarshad Shirwadkar * largest free order in the buddy bitmap of the participating group infos of 141196e402aSHarshad Shirwadkar * that list. So, there are exactly MB_NUM_ORDERS(sb) (which means total 142196e402aSHarshad Shirwadkar * number of buddy bitmap orders possible) number of lists. Group-infos are 143196e402aSHarshad Shirwadkar * placed in appropriate lists. 144196e402aSHarshad Shirwadkar * 14583e80a6eSJan Kara * 2) Average fragment size lists (sbi->s_mb_avg_fragment_size) 146196e402aSHarshad Shirwadkar * 14783e80a6eSJan Kara * Locking: sbi->s_mb_avg_fragment_size_locks(array of rw locks) 148196e402aSHarshad Shirwadkar * 14983e80a6eSJan Kara * This is an array of lists where in the i-th list there are groups with 15083e80a6eSJan Kara * average fragment size >= 2^i and < 2^(i+1). The average fragment size 15183e80a6eSJan Kara * is computed as ext4_group_info->bb_free / ext4_group_info->bb_fragments. 15283e80a6eSJan Kara * Note that we don't bother with a special list for completely empty groups 15383e80a6eSJan Kara * so we only have MB_NUM_ORDERS(sb) lists. 154196e402aSHarshad Shirwadkar * 155196e402aSHarshad Shirwadkar * When "mb_optimize_scan" mount option is set, mballoc consults the above data 156196e402aSHarshad Shirwadkar * structures to decide the order in which groups are to be traversed for 157196e402aSHarshad Shirwadkar * fulfilling an allocation request. 158196e402aSHarshad Shirwadkar * 159f52f3d2bSOjaswin Mujoo * At CR_POWER2_ALIGNED , we look for groups which have the largest_free_order 160f52f3d2bSOjaswin Mujoo * >= the order of the request. We directly look at the largest free order list 161f52f3d2bSOjaswin Mujoo * in the data structure (1) above where largest_free_order = order of the 162f52f3d2bSOjaswin Mujoo * request. If that list is empty, we look at remaining list in the increasing 163f52f3d2bSOjaswin Mujoo * order of largest_free_order. This allows us to perform CR_POWER2_ALIGNED 164f52f3d2bSOjaswin Mujoo * lookup in O(1) time. 165196e402aSHarshad Shirwadkar * 166f52f3d2bSOjaswin Mujoo * At CR_GOAL_LEN_FAST, we only consider groups where 167f52f3d2bSOjaswin Mujoo * average fragment size > request size. So, we lookup a group which has average 168f52f3d2bSOjaswin Mujoo * fragment size just above or equal to request size using our average fragment 169f52f3d2bSOjaswin Mujoo * size group lists (data structure 2) in O(1) time. 170196e402aSHarshad Shirwadkar * 171f52f3d2bSOjaswin Mujoo * At CR_BEST_AVAIL_LEN, we aim to optimize allocations which can't be satisfied 172f52f3d2bSOjaswin Mujoo * in CR_GOAL_LEN_FAST. The fact that we couldn't find a group in 173f52f3d2bSOjaswin Mujoo * CR_GOAL_LEN_FAST suggests that there is no BG that has avg 174f52f3d2bSOjaswin Mujoo * fragment size > goal length. So before falling to the slower 175f52f3d2bSOjaswin Mujoo * CR_GOAL_LEN_SLOW, in CR_BEST_AVAIL_LEN we proactively trim goal length and 176f52f3d2bSOjaswin Mujoo * then use the same fragment lists as CR_GOAL_LEN_FAST to find a BG with a big 177f52f3d2bSOjaswin Mujoo * enough average fragment size. This increases the chances of finding a 178f52f3d2bSOjaswin Mujoo * suitable block group in O(1) time and results in faster allocation at the 179f52f3d2bSOjaswin Mujoo * cost of reduced size of allocation. 1807e170922SOjaswin Mujoo * 181196e402aSHarshad Shirwadkar * If "mb_optimize_scan" mount option is not set, mballoc traverses groups in 182f52f3d2bSOjaswin Mujoo * linear order which requires O(N) search time for each CR_POWER2_ALIGNED and 183f52f3d2bSOjaswin Mujoo * CR_GOAL_LEN_FAST phase. 184196e402aSHarshad Shirwadkar * 185d7a1fee1SDan Ehrenberg * The regular allocator (using the buddy cache) supports a few tunables. 186c9de560dSAlex Tomas * 187b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_min_to_scan 188b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_max_to_scan 189b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req 190196e402aSHarshad Shirwadkar * /sys/fs/ext4/<partition>/mb_linear_limit 191c9de560dSAlex Tomas * 192b713a5ecSTheodore Ts'o * The regular allocator uses buddy scan only if the request len is power of 193c9de560dSAlex Tomas * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 194c9de560dSAlex Tomas * value of s_mb_order2_reqs can be tuned via 195b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 196af901ca1SAndré Goddard Rosa * stripe size (sbi->s_stripe), we try to search for contiguous block in 197b713a5ecSTheodore Ts'o * stripe size. This should result in better allocation on RAID setups. If 198b713a5ecSTheodore Ts'o * not, we search in the specific group using bitmap for best extents. The 199b713a5ecSTheodore Ts'o * tunable min_to_scan and max_to_scan control the behaviour here. 200c9de560dSAlex Tomas * min_to_scan indicate how long the mballoc __must__ look for a best 201b713a5ecSTheodore Ts'o * extent and max_to_scan indicates how long the mballoc __can__ look for a 202c9de560dSAlex Tomas * best extent in the found extents. Searching for the blocks starts with 203c9de560dSAlex Tomas * the group specified as the goal value in allocation context via 204c9de560dSAlex Tomas * ac_g_ex. Each group is first checked based on the criteria whether it 205caaf7a29STao Ma * can be used for allocation. ext4_mb_good_group explains how the groups are 206c9de560dSAlex Tomas * checked. 207c9de560dSAlex Tomas * 208196e402aSHarshad Shirwadkar * When "mb_optimize_scan" is turned on, as mentioned above, the groups may not 209196e402aSHarshad Shirwadkar * get traversed linearly. That may result in subsequent allocations being not 210196e402aSHarshad Shirwadkar * close to each other. And so, the underlying device may get filled up in a 211196e402aSHarshad Shirwadkar * non-linear fashion. While that may not matter on non-rotational devices, for 212196e402aSHarshad Shirwadkar * rotational devices that may result in higher seek times. "mb_linear_limit" 213196e402aSHarshad Shirwadkar * tells mballoc how many groups mballoc should search linearly before 214196e402aSHarshad Shirwadkar * performing consulting above data structures for more efficient lookups. For 215196e402aSHarshad Shirwadkar * non rotational devices, this value defaults to 0 and for rotational devices 216196e402aSHarshad Shirwadkar * this is set to MB_DEFAULT_LINEAR_LIMIT. 217196e402aSHarshad Shirwadkar * 218c9de560dSAlex Tomas * Both the prealloc space are getting populated as above. So for the first 219c9de560dSAlex Tomas * request we will hit the buddy cache which will result in this prealloc 220c9de560dSAlex Tomas * space getting filled. The prealloc space is then later used for the 221c9de560dSAlex Tomas * subsequent request. 222c9de560dSAlex Tomas */ 223c9de560dSAlex Tomas 224c9de560dSAlex Tomas /* 225c9de560dSAlex Tomas * mballoc operates on the following data: 226c9de560dSAlex Tomas * - on-disk bitmap 227c9de560dSAlex Tomas * - in-core buddy (actually includes buddy and bitmap) 228c9de560dSAlex Tomas * - preallocation descriptors (PAs) 229c9de560dSAlex Tomas * 230c9de560dSAlex Tomas * there are two types of preallocations: 231c9de560dSAlex Tomas * - inode 232c9de560dSAlex Tomas * assiged to specific inode and can be used for this inode only. 233c9de560dSAlex Tomas * it describes part of inode's space preallocated to specific 234c9de560dSAlex Tomas * physical blocks. any block from that preallocated can be used 235c9de560dSAlex Tomas * independent. the descriptor just tracks number of blocks left 236c9de560dSAlex Tomas * unused. so, before taking some block from descriptor, one must 237c9de560dSAlex Tomas * make sure corresponded logical block isn't allocated yet. this 238c9de560dSAlex Tomas * also means that freeing any block within descriptor's range 239c9de560dSAlex Tomas * must discard all preallocated blocks. 240c9de560dSAlex Tomas * - locality group 241c9de560dSAlex Tomas * assigned to specific locality group which does not translate to 242c9de560dSAlex Tomas * permanent set of inodes: inode can join and leave group. space 243c9de560dSAlex Tomas * from this type of preallocation can be used for any inode. thus 244c9de560dSAlex Tomas * it's consumed from the beginning to the end. 245c9de560dSAlex Tomas * 246c9de560dSAlex Tomas * relation between them can be expressed as: 247c9de560dSAlex Tomas * in-core buddy = on-disk bitmap + preallocation descriptors 248c9de560dSAlex Tomas * 249c9de560dSAlex Tomas * this mean blocks mballoc considers used are: 250c9de560dSAlex Tomas * - allocated blocks (persistent) 251c9de560dSAlex Tomas * - preallocated blocks (non-persistent) 252c9de560dSAlex Tomas * 253c9de560dSAlex Tomas * consistency in mballoc world means that at any time a block is either 254c9de560dSAlex Tomas * free or used in ALL structures. notice: "any time" should not be read 255c9de560dSAlex Tomas * literally -- time is discrete and delimited by locks. 256c9de560dSAlex Tomas * 257c9de560dSAlex Tomas * to keep it simple, we don't use block numbers, instead we count number of 258c9de560dSAlex Tomas * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 259c9de560dSAlex Tomas * 260c9de560dSAlex Tomas * all operations can be expressed as: 261c9de560dSAlex Tomas * - init buddy: buddy = on-disk + PAs 262c9de560dSAlex Tomas * - new PA: buddy += N; PA = N 263c9de560dSAlex Tomas * - use inode PA: on-disk += N; PA -= N 264c9de560dSAlex Tomas * - discard inode PA buddy -= on-disk - PA; PA = 0 265c9de560dSAlex Tomas * - use locality group PA on-disk += N; PA -= N 266c9de560dSAlex Tomas * - discard locality group PA buddy -= PA; PA = 0 267c9de560dSAlex Tomas * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 268c9de560dSAlex Tomas * is used in real operation because we can't know actual used 269c9de560dSAlex Tomas * bits from PA, only from on-disk bitmap 270c9de560dSAlex Tomas * 271c9de560dSAlex Tomas * if we follow this strict logic, then all operations above should be atomic. 272c9de560dSAlex Tomas * given some of them can block, we'd have to use something like semaphores 273c9de560dSAlex Tomas * killing performance on high-end SMP hardware. let's try to relax it using 274c9de560dSAlex Tomas * the following knowledge: 275c9de560dSAlex Tomas * 1) if buddy is referenced, it's already initialized 276c9de560dSAlex Tomas * 2) while block is used in buddy and the buddy is referenced, 277c9de560dSAlex Tomas * nobody can re-allocate that block 278c9de560dSAlex Tomas * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 279c9de560dSAlex Tomas * bit set and PA claims same block, it's OK. IOW, one can set bit in 280c9de560dSAlex Tomas * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 281c9de560dSAlex Tomas * block 282c9de560dSAlex Tomas * 283c9de560dSAlex Tomas * so, now we're building a concurrency table: 284c9de560dSAlex Tomas * - init buddy vs. 285c9de560dSAlex Tomas * - new PA 286c9de560dSAlex Tomas * blocks for PA are allocated in the buddy, buddy must be referenced 287c9de560dSAlex Tomas * until PA is linked to allocation group to avoid concurrent buddy init 288c9de560dSAlex Tomas * - use inode PA 289c9de560dSAlex Tomas * we need to make sure that either on-disk bitmap or PA has uptodate data 290c9de560dSAlex Tomas * given (3) we care that PA-=N operation doesn't interfere with init 291c9de560dSAlex Tomas * - discard inode PA 292c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 293c9de560dSAlex Tomas * - use locality group PA 294c9de560dSAlex Tomas * again PA-=N must be serialized with init 295c9de560dSAlex Tomas * - discard locality group PA 296c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 297c9de560dSAlex Tomas * - new PA vs. 298c9de560dSAlex Tomas * - use inode PA 299c9de560dSAlex Tomas * i_data_sem serializes them 300c9de560dSAlex Tomas * - discard inode PA 301c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 302c9de560dSAlex Tomas * - use locality group PA 303c9de560dSAlex Tomas * some mutex should serialize them 304c9de560dSAlex Tomas * - discard locality group PA 305c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 306c9de560dSAlex Tomas * - use inode PA 307c9de560dSAlex Tomas * - use inode PA 308c9de560dSAlex Tomas * i_data_sem or another mutex should serializes them 309c9de560dSAlex Tomas * - discard inode PA 310c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 311c9de560dSAlex Tomas * - use locality group PA 312c9de560dSAlex Tomas * nothing wrong here -- they're different PAs covering different blocks 313c9de560dSAlex Tomas * - discard locality group PA 314c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 315c9de560dSAlex Tomas * 316c9de560dSAlex Tomas * now we're ready to make few consequences: 317c9de560dSAlex Tomas * - PA is referenced and while it is no discard is possible 318c9de560dSAlex Tomas * - PA is referenced until block isn't marked in on-disk bitmap 319c9de560dSAlex Tomas * - PA changes only after on-disk bitmap 320c9de560dSAlex Tomas * - discard must not compete with init. either init is done before 321c9de560dSAlex Tomas * any discard or they're serialized somehow 322c9de560dSAlex Tomas * - buddy init as sum of on-disk bitmap and PAs is done atomically 323c9de560dSAlex Tomas * 324c9de560dSAlex Tomas * a special case when we've used PA to emptiness. no need to modify buddy 325c9de560dSAlex Tomas * in this case, but we should care about concurrent init 326c9de560dSAlex Tomas * 327c9de560dSAlex Tomas */ 328c9de560dSAlex Tomas 329c9de560dSAlex Tomas /* 330c9de560dSAlex Tomas * Logic in few words: 331c9de560dSAlex Tomas * 332c9de560dSAlex Tomas * - allocation: 333c9de560dSAlex Tomas * load group 334c9de560dSAlex Tomas * find blocks 335c9de560dSAlex Tomas * mark bits in on-disk bitmap 336c9de560dSAlex Tomas * release group 337c9de560dSAlex Tomas * 338c9de560dSAlex Tomas * - use preallocation: 339c9de560dSAlex Tomas * find proper PA (per-inode or group) 340c9de560dSAlex Tomas * load group 341c9de560dSAlex Tomas * mark bits in on-disk bitmap 342c9de560dSAlex Tomas * release group 343c9de560dSAlex Tomas * release PA 344c9de560dSAlex Tomas * 345c9de560dSAlex Tomas * - free: 346c9de560dSAlex Tomas * load group 347c9de560dSAlex Tomas * mark bits in on-disk bitmap 348c9de560dSAlex Tomas * release group 349c9de560dSAlex Tomas * 350c9de560dSAlex Tomas * - discard preallocations in group: 351c9de560dSAlex Tomas * mark PAs deleted 352c9de560dSAlex Tomas * move them onto local list 353c9de560dSAlex Tomas * load on-disk bitmap 354c9de560dSAlex Tomas * load group 355c9de560dSAlex Tomas * remove PA from object (inode or locality group) 356c9de560dSAlex Tomas * mark free blocks in-core 357c9de560dSAlex Tomas * 358c9de560dSAlex Tomas * - discard inode's preallocations: 359c9de560dSAlex Tomas */ 360c9de560dSAlex Tomas 361c9de560dSAlex Tomas /* 362c9de560dSAlex Tomas * Locking rules 363c9de560dSAlex Tomas * 364c9de560dSAlex Tomas * Locks: 365c9de560dSAlex Tomas * - bitlock on a group (group) 366c9de560dSAlex Tomas * - object (inode/locality) (object) 367c9de560dSAlex Tomas * - per-pa lock (pa) 368f52f3d2bSOjaswin Mujoo * - cr_power2_aligned lists lock (cr_power2_aligned) 369f52f3d2bSOjaswin Mujoo * - cr_goal_len_fast lists lock (cr_goal_len_fast) 370c9de560dSAlex Tomas * 371c9de560dSAlex Tomas * Paths: 372c9de560dSAlex Tomas * - new pa 373c9de560dSAlex Tomas * object 374c9de560dSAlex Tomas * group 375c9de560dSAlex Tomas * 376c9de560dSAlex Tomas * - find and use pa: 377c9de560dSAlex Tomas * pa 378c9de560dSAlex Tomas * 379c9de560dSAlex Tomas * - release consumed pa: 380c9de560dSAlex Tomas * pa 381c9de560dSAlex Tomas * group 382c9de560dSAlex Tomas * object 383c9de560dSAlex Tomas * 384c9de560dSAlex Tomas * - generate in-core bitmap: 385c9de560dSAlex Tomas * group 386c9de560dSAlex Tomas * pa 387c9de560dSAlex Tomas * 388c9de560dSAlex Tomas * - discard all for given object (inode, locality group): 389c9de560dSAlex Tomas * object 390c9de560dSAlex Tomas * pa 391c9de560dSAlex Tomas * group 392c9de560dSAlex Tomas * 393c9de560dSAlex Tomas * - discard all for given group: 394c9de560dSAlex Tomas * group 395c9de560dSAlex Tomas * pa 396c9de560dSAlex Tomas * group 397c9de560dSAlex Tomas * object 398c9de560dSAlex Tomas * 399196e402aSHarshad Shirwadkar * - allocation path (ext4_mb_regular_allocator) 400196e402aSHarshad Shirwadkar * group 401f52f3d2bSOjaswin Mujoo * cr_power2_aligned/cr_goal_len_fast 402c9de560dSAlex Tomas */ 403c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_pspace_cachep; 404c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_ac_cachep; 40518aadd47SBobi Jam static struct kmem_cache *ext4_free_data_cachep; 406fb1813f4SCurt Wohlgemuth 407fb1813f4SCurt Wohlgemuth /* We create slab caches for groupinfo data structures based on the 408fb1813f4SCurt Wohlgemuth * superblock block size. There will be one per mounted filesystem for 409fb1813f4SCurt Wohlgemuth * each unique s_blocksize_bits */ 4102892c15dSEric Sandeen #define NR_GRPINFO_CACHES 8 411fb1813f4SCurt Wohlgemuth static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; 412fb1813f4SCurt Wohlgemuth 413d6006186SEric Biggers static const char * const ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { 4142892c15dSEric Sandeen "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", 4152892c15dSEric Sandeen "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", 4162892c15dSEric Sandeen "ext4_groupinfo_64k", "ext4_groupinfo_128k" 4172892c15dSEric Sandeen }; 4182892c15dSEric Sandeen 419c3a326a6SAneesh Kumar K.V static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 420c3a326a6SAneesh Kumar K.V ext4_group_t group); 42153f86b17SRitesh Harjani static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac); 422c3a326a6SAneesh Kumar K.V 423196e402aSHarshad Shirwadkar static bool ext4_mb_good_group(struct ext4_allocation_context *ac, 4244eb7a4a1SOjaswin Mujoo ext4_group_t group, enum criteria cr); 425196e402aSHarshad Shirwadkar 42655cdd0afSWang Jianchao static int ext4_try_to_trim_range(struct super_block *sb, 42755cdd0afSWang Jianchao struct ext4_buddy *e4b, ext4_grpblk_t start, 42855cdd0afSWang Jianchao ext4_grpblk_t max, ext4_grpblk_t minblocks); 42955cdd0afSWang Jianchao 43007b5b8e1SRitesh Harjani /* 43107b5b8e1SRitesh Harjani * The algorithm using this percpu seq counter goes below: 43207b5b8e1SRitesh Harjani * 1. We sample the percpu discard_pa_seq counter before trying for block 43307b5b8e1SRitesh Harjani * allocation in ext4_mb_new_blocks(). 43407b5b8e1SRitesh Harjani * 2. We increment this percpu discard_pa_seq counter when we either allocate 43507b5b8e1SRitesh Harjani * or free these blocks i.e. while marking those blocks as used/free in 43607b5b8e1SRitesh Harjani * mb_mark_used()/mb_free_blocks(). 43707b5b8e1SRitesh Harjani * 3. We also increment this percpu seq counter when we successfully identify 43807b5b8e1SRitesh Harjani * that the bb_prealloc_list is not empty and hence proceed for discarding 43907b5b8e1SRitesh Harjani * of those PAs inside ext4_mb_discard_group_preallocations(). 44007b5b8e1SRitesh Harjani * 44107b5b8e1SRitesh Harjani * Now to make sure that the regular fast path of block allocation is not 44207b5b8e1SRitesh Harjani * affected, as a small optimization we only sample the percpu seq counter 44307b5b8e1SRitesh Harjani * on that cpu. Only when the block allocation fails and when freed blocks 44407b5b8e1SRitesh Harjani * found were 0, that is when we sample percpu seq counter for all cpus using 44507b5b8e1SRitesh Harjani * below function ext4_get_discard_pa_seq_sum(). This happens after making 44607b5b8e1SRitesh Harjani * sure that all the PAs on grp->bb_prealloc_list got freed or if it's empty. 44707b5b8e1SRitesh Harjani */ 44807b5b8e1SRitesh Harjani static DEFINE_PER_CPU(u64, discard_pa_seq); 44907b5b8e1SRitesh Harjani static inline u64 ext4_get_discard_pa_seq_sum(void) 45007b5b8e1SRitesh Harjani { 45107b5b8e1SRitesh Harjani int __cpu; 45207b5b8e1SRitesh Harjani u64 __seq = 0; 45307b5b8e1SRitesh Harjani 45407b5b8e1SRitesh Harjani for_each_possible_cpu(__cpu) 45507b5b8e1SRitesh Harjani __seq += per_cpu(discard_pa_seq, __cpu); 45607b5b8e1SRitesh Harjani return __seq; 45707b5b8e1SRitesh Harjani } 45807b5b8e1SRitesh Harjani 459ffad0a44SAneesh Kumar K.V static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 460ffad0a44SAneesh Kumar K.V { 461c9de560dSAlex Tomas #if BITS_PER_LONG == 64 462ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 7UL) << 3; 463ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~7UL); 464c9de560dSAlex Tomas #elif BITS_PER_LONG == 32 465ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 3UL) << 3; 466ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~3UL); 467c9de560dSAlex Tomas #else 468c9de560dSAlex Tomas #error "how many bits you are?!" 469c9de560dSAlex Tomas #endif 470ffad0a44SAneesh Kumar K.V return addr; 471ffad0a44SAneesh Kumar K.V } 472c9de560dSAlex Tomas 473c9de560dSAlex Tomas static inline int mb_test_bit(int bit, void *addr) 474c9de560dSAlex Tomas { 475c9de560dSAlex Tomas /* 476c9de560dSAlex Tomas * ext4_test_bit on architecture like powerpc 477c9de560dSAlex Tomas * needs unsigned long aligned address 478c9de560dSAlex Tomas */ 479ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 480c9de560dSAlex Tomas return ext4_test_bit(bit, addr); 481c9de560dSAlex Tomas } 482c9de560dSAlex Tomas 483c9de560dSAlex Tomas static inline void mb_set_bit(int bit, void *addr) 484c9de560dSAlex Tomas { 485ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 486c9de560dSAlex Tomas ext4_set_bit(bit, addr); 487c9de560dSAlex Tomas } 488c9de560dSAlex Tomas 489c9de560dSAlex Tomas static inline void mb_clear_bit(int bit, void *addr) 490c9de560dSAlex Tomas { 491ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 492c9de560dSAlex Tomas ext4_clear_bit(bit, addr); 493c9de560dSAlex Tomas } 494c9de560dSAlex Tomas 495eabe0444SAndrey Sidorov static inline int mb_test_and_clear_bit(int bit, void *addr) 496eabe0444SAndrey Sidorov { 497eabe0444SAndrey Sidorov addr = mb_correct_addr_and_bit(&bit, addr); 498eabe0444SAndrey Sidorov return ext4_test_and_clear_bit(bit, addr); 499eabe0444SAndrey Sidorov } 500eabe0444SAndrey Sidorov 501ffad0a44SAneesh Kumar K.V static inline int mb_find_next_zero_bit(void *addr, int max, int start) 502ffad0a44SAneesh Kumar K.V { 503e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 504ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 505e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 506ffad0a44SAneesh Kumar K.V start += fix; 507ffad0a44SAneesh Kumar K.V 508e7dfb246SAneesh Kumar K.V ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 509e7dfb246SAneesh Kumar K.V if (ret > max) 510e7dfb246SAneesh Kumar K.V return max; 511e7dfb246SAneesh Kumar K.V return ret; 512ffad0a44SAneesh Kumar K.V } 513ffad0a44SAneesh Kumar K.V 514ffad0a44SAneesh Kumar K.V static inline int mb_find_next_bit(void *addr, int max, int start) 515ffad0a44SAneesh Kumar K.V { 516e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 517ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 518e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 519ffad0a44SAneesh Kumar K.V start += fix; 520ffad0a44SAneesh Kumar K.V 521e7dfb246SAneesh Kumar K.V ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 522e7dfb246SAneesh Kumar K.V if (ret > max) 523e7dfb246SAneesh Kumar K.V return max; 524e7dfb246SAneesh Kumar K.V return ret; 525ffad0a44SAneesh Kumar K.V } 526ffad0a44SAneesh Kumar K.V 527c9de560dSAlex Tomas static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 528c9de560dSAlex Tomas { 529c9de560dSAlex Tomas char *bb; 530c9de560dSAlex Tomas 531c5e8f3f3STheodore Ts'o BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 532c9de560dSAlex Tomas BUG_ON(max == NULL); 533c9de560dSAlex Tomas 534c9de560dSAlex Tomas if (order > e4b->bd_blkbits + 1) { 535c9de560dSAlex Tomas *max = 0; 536c9de560dSAlex Tomas return NULL; 537c9de560dSAlex Tomas } 538c9de560dSAlex Tomas 539c9de560dSAlex Tomas /* at order 0 we see each particular block */ 54084b775a3SColy Li if (order == 0) { 541c9de560dSAlex Tomas *max = 1 << (e4b->bd_blkbits + 3); 542c5e8f3f3STheodore Ts'o return e4b->bd_bitmap; 54384b775a3SColy Li } 544c9de560dSAlex Tomas 545c5e8f3f3STheodore Ts'o bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 546c9de560dSAlex Tomas *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 547c9de560dSAlex Tomas 548c9de560dSAlex Tomas return bb; 549c9de560dSAlex Tomas } 550c9de560dSAlex Tomas 551c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 552c9de560dSAlex Tomas static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 553c9de560dSAlex Tomas int first, int count) 554c9de560dSAlex Tomas { 555c9de560dSAlex Tomas int i; 556c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 557c9de560dSAlex Tomas 558c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 559c9de560dSAlex Tomas return; 560bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 561c9de560dSAlex Tomas for (i = 0; i < count; i++) { 562c9de560dSAlex Tomas if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 563c9de560dSAlex Tomas ext4_fsblk_t blocknr; 5645661bd68SAkinobu Mita 5655661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 56653accfa9STheodore Ts'o blocknr += EXT4_C2B(EXT4_SB(sb), first + i); 567c5f3a382SBaokun Li ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 568c5f3a382SBaokun Li EXT4_GROUP_INFO_BBITMAP_CORRUPT); 5695d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 570e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 571e29136f8STheodore Ts'o blocknr, 572e29136f8STheodore Ts'o "freeing block already freed " 573e29136f8STheodore Ts'o "(bit %u)", 574e29136f8STheodore Ts'o first + i); 575c9de560dSAlex Tomas } 576c9de560dSAlex Tomas mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 577c9de560dSAlex Tomas } 578c9de560dSAlex Tomas } 579c9de560dSAlex Tomas 580c9de560dSAlex Tomas static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 581c9de560dSAlex Tomas { 582c9de560dSAlex Tomas int i; 583c9de560dSAlex Tomas 584c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 585c9de560dSAlex Tomas return; 586bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 587c9de560dSAlex Tomas for (i = 0; i < count; i++) { 588c9de560dSAlex Tomas BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 589c9de560dSAlex Tomas mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 590c9de560dSAlex Tomas } 591c9de560dSAlex Tomas } 592c9de560dSAlex Tomas 593c9de560dSAlex Tomas static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 594c9de560dSAlex Tomas { 595eb2b8ebbSRitesh Harjani if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 596eb2b8ebbSRitesh Harjani return; 597c9de560dSAlex Tomas if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 598c9de560dSAlex Tomas unsigned char *b1, *b2; 599c9de560dSAlex Tomas int i; 600c9de560dSAlex Tomas b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 601c9de560dSAlex Tomas b2 = (unsigned char *) bitmap; 602c9de560dSAlex Tomas for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 603c9de560dSAlex Tomas if (b1[i] != b2[i]) { 6049d8b9ec4STheodore Ts'o ext4_msg(e4b->bd_sb, KERN_ERR, 6059d8b9ec4STheodore Ts'o "corruption in group %u " 6064776004fSTheodore Ts'o "at byte %u(%u): %x in copy != %x " 6079d8b9ec4STheodore Ts'o "on disk/prealloc", 608c9de560dSAlex Tomas e4b->bd_group, i, i * 8, b1[i], b2[i]); 609c9de560dSAlex Tomas BUG(); 610c9de560dSAlex Tomas } 611c9de560dSAlex Tomas } 612c9de560dSAlex Tomas } 613c9de560dSAlex Tomas } 614c9de560dSAlex Tomas 615a3450215SRitesh Harjani static void mb_group_bb_bitmap_alloc(struct super_block *sb, 616a3450215SRitesh Harjani struct ext4_group_info *grp, ext4_group_t group) 617a3450215SRitesh Harjani { 618a3450215SRitesh Harjani struct buffer_head *bh; 619a3450215SRitesh Harjani 620a3450215SRitesh Harjani grp->bb_bitmap = kmalloc(sb->s_blocksize, GFP_NOFS); 621eb2b8ebbSRitesh Harjani if (!grp->bb_bitmap) 622eb2b8ebbSRitesh Harjani return; 623a3450215SRitesh Harjani 624a3450215SRitesh Harjani bh = ext4_read_block_bitmap(sb, group); 625eb2b8ebbSRitesh Harjani if (IS_ERR_OR_NULL(bh)) { 626eb2b8ebbSRitesh Harjani kfree(grp->bb_bitmap); 627eb2b8ebbSRitesh Harjani grp->bb_bitmap = NULL; 628eb2b8ebbSRitesh Harjani return; 629eb2b8ebbSRitesh Harjani } 630a3450215SRitesh Harjani 631a3450215SRitesh Harjani memcpy(grp->bb_bitmap, bh->b_data, sb->s_blocksize); 632a3450215SRitesh Harjani put_bh(bh); 633a3450215SRitesh Harjani } 634a3450215SRitesh Harjani 635a3450215SRitesh Harjani static void mb_group_bb_bitmap_free(struct ext4_group_info *grp) 636a3450215SRitesh Harjani { 637a3450215SRitesh Harjani kfree(grp->bb_bitmap); 638a3450215SRitesh Harjani } 639a3450215SRitesh Harjani 640c9de560dSAlex Tomas #else 641c9de560dSAlex Tomas static inline void mb_free_blocks_double(struct inode *inode, 642c9de560dSAlex Tomas struct ext4_buddy *e4b, int first, int count) 643c9de560dSAlex Tomas { 644c9de560dSAlex Tomas return; 645c9de560dSAlex Tomas } 646c9de560dSAlex Tomas static inline void mb_mark_used_double(struct ext4_buddy *e4b, 647c9de560dSAlex Tomas int first, int count) 648c9de560dSAlex Tomas { 649c9de560dSAlex Tomas return; 650c9de560dSAlex Tomas } 651c9de560dSAlex Tomas static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 652c9de560dSAlex Tomas { 653c9de560dSAlex Tomas return; 654c9de560dSAlex Tomas } 655a3450215SRitesh Harjani 656a3450215SRitesh Harjani static inline void mb_group_bb_bitmap_alloc(struct super_block *sb, 657a3450215SRitesh Harjani struct ext4_group_info *grp, ext4_group_t group) 658a3450215SRitesh Harjani { 659a3450215SRitesh Harjani return; 660a3450215SRitesh Harjani } 661a3450215SRitesh Harjani 662a3450215SRitesh Harjani static inline void mb_group_bb_bitmap_free(struct ext4_group_info *grp) 663a3450215SRitesh Harjani { 664a3450215SRitesh Harjani return; 665a3450215SRitesh Harjani } 666c9de560dSAlex Tomas #endif 667c9de560dSAlex Tomas 668c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 669c9de560dSAlex Tomas 670c9de560dSAlex Tomas #define MB_CHECK_ASSERT(assert) \ 671c9de560dSAlex Tomas do { \ 672c9de560dSAlex Tomas if (!(assert)) { \ 673c9de560dSAlex Tomas printk(KERN_EMERG \ 674c9de560dSAlex Tomas "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 675c9de560dSAlex Tomas function, file, line, # assert); \ 676c9de560dSAlex Tomas BUG(); \ 677c9de560dSAlex Tomas } \ 678c9de560dSAlex Tomas } while (0) 679c9de560dSAlex Tomas 680133de5a0SKemeng Shi static void __mb_check_buddy(struct ext4_buddy *e4b, char *file, 681c9de560dSAlex Tomas const char *function, int line) 682c9de560dSAlex Tomas { 683c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 684c9de560dSAlex Tomas int order = e4b->bd_blkbits + 1; 685c9de560dSAlex Tomas int max; 686c9de560dSAlex Tomas int max2; 687c9de560dSAlex Tomas int i; 688c9de560dSAlex Tomas int j; 689c9de560dSAlex Tomas int k; 690c9de560dSAlex Tomas int count; 691c9de560dSAlex Tomas struct ext4_group_info *grp; 692c9de560dSAlex Tomas int fragments = 0; 693c9de560dSAlex Tomas int fstart; 694c9de560dSAlex Tomas struct list_head *cur; 695c9de560dSAlex Tomas void *buddy; 696c9de560dSAlex Tomas void *buddy2; 697c9de560dSAlex Tomas 698addd752cSChunguang Xu if (e4b->bd_info->bb_check_counter++ % 10) 699133de5a0SKemeng Shi return; 700c9de560dSAlex Tomas 701c9de560dSAlex Tomas while (order > 1) { 702c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 703c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy); 704c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order - 1, &max2); 705c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy2); 706c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy != buddy2); 707c9de560dSAlex Tomas MB_CHECK_ASSERT(max * 2 == max2); 708c9de560dSAlex Tomas 709c9de560dSAlex Tomas count = 0; 710c9de560dSAlex Tomas for (i = 0; i < max; i++) { 711c9de560dSAlex Tomas 712c9de560dSAlex Tomas if (mb_test_bit(i, buddy)) { 713af2b3275SJinke Han /* only single bit in buddy2 may be 0 */ 714c9de560dSAlex Tomas if (!mb_test_bit(i << 1, buddy2)) { 715c9de560dSAlex Tomas MB_CHECK_ASSERT( 716c9de560dSAlex Tomas mb_test_bit((i<<1)+1, buddy2)); 717c9de560dSAlex Tomas } 718c9de560dSAlex Tomas continue; 719c9de560dSAlex Tomas } 720c9de560dSAlex Tomas 7210a10da73SRobin Dong /* both bits in buddy2 must be 1 */ 722c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 723c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 724c9de560dSAlex Tomas 725c9de560dSAlex Tomas for (j = 0; j < (1 << order); j++) { 726c9de560dSAlex Tomas k = (i * (1 << order)) + j; 727c9de560dSAlex Tomas MB_CHECK_ASSERT( 728c5e8f3f3STheodore Ts'o !mb_test_bit(k, e4b->bd_bitmap)); 729c9de560dSAlex Tomas } 730c9de560dSAlex Tomas count++; 731c9de560dSAlex Tomas } 732c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 733c9de560dSAlex Tomas order--; 734c9de560dSAlex Tomas } 735c9de560dSAlex Tomas 736c9de560dSAlex Tomas fstart = -1; 737c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 738c9de560dSAlex Tomas for (i = 0; i < max; i++) { 739c9de560dSAlex Tomas if (!mb_test_bit(i, buddy)) { 740c9de560dSAlex Tomas MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 741c9de560dSAlex Tomas if (fstart == -1) { 742c9de560dSAlex Tomas fragments++; 743c9de560dSAlex Tomas fstart = i; 744c9de560dSAlex Tomas } 745c9de560dSAlex Tomas continue; 746c9de560dSAlex Tomas } 747c9de560dSAlex Tomas fstart = -1; 748c9de560dSAlex Tomas /* check used bits only */ 749c9de560dSAlex Tomas for (j = 0; j < e4b->bd_blkbits + 1; j++) { 750c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, j, &max2); 751c9de560dSAlex Tomas k = i >> j; 752c9de560dSAlex Tomas MB_CHECK_ASSERT(k < max2); 753c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 754c9de560dSAlex Tomas } 755c9de560dSAlex Tomas } 756c9de560dSAlex Tomas MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 757c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 758c9de560dSAlex Tomas 759c9de560dSAlex Tomas grp = ext4_get_group_info(sb, e4b->bd_group); 7605354b2afSTheodore Ts'o if (!grp) 761133de5a0SKemeng Shi return; 762c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 763c9de560dSAlex Tomas ext4_group_t groupnr; 764c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 76560bd63d1SSolofo Ramangalahy pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 76660bd63d1SSolofo Ramangalahy ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 767c9de560dSAlex Tomas MB_CHECK_ASSERT(groupnr == e4b->bd_group); 76860bd63d1SSolofo Ramangalahy for (i = 0; i < pa->pa_len; i++) 769c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 770c9de560dSAlex Tomas } 771c9de560dSAlex Tomas } 772c9de560dSAlex Tomas #undef MB_CHECK_ASSERT 773c9de560dSAlex Tomas #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 77446e665e9SHarvey Harrison __FILE__, __func__, __LINE__) 775c9de560dSAlex Tomas #else 776c9de560dSAlex Tomas #define mb_check_buddy(e4b) 777c9de560dSAlex Tomas #endif 778c9de560dSAlex Tomas 7797c786059SColy Li /* 7807c786059SColy Li * Divide blocks started from @first with length @len into 7817c786059SColy Li * smaller chunks with power of 2 blocks. 7827c786059SColy Li * Clear the bits in bitmap which the blocks of the chunk(s) covered, 7837c786059SColy Li * then increase bb_counters[] for corresponded chunk size. 7847c786059SColy Li */ 785c9de560dSAlex Tomas static void ext4_mb_mark_free_simple(struct super_block *sb, 786a36b4498SEric Sandeen void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 787c9de560dSAlex Tomas struct ext4_group_info *grp) 788c9de560dSAlex Tomas { 789c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 790a36b4498SEric Sandeen ext4_grpblk_t min; 791a36b4498SEric Sandeen ext4_grpblk_t max; 792a36b4498SEric Sandeen ext4_grpblk_t chunk; 79369e43e8cSChandan Rajendra unsigned int border; 794c9de560dSAlex Tomas 7957137d7a4STheodore Ts'o BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); 796c9de560dSAlex Tomas 797c9de560dSAlex Tomas border = 2 << sb->s_blocksize_bits; 798c9de560dSAlex Tomas 799c9de560dSAlex Tomas while (len > 0) { 800c9de560dSAlex Tomas /* find how many blocks can be covered since this position */ 801c9de560dSAlex Tomas max = ffs(first | border) - 1; 802c9de560dSAlex Tomas 803c9de560dSAlex Tomas /* find how many blocks of power 2 we need to mark */ 804c9de560dSAlex Tomas min = fls(len) - 1; 805c9de560dSAlex Tomas 806c9de560dSAlex Tomas if (max < min) 807c9de560dSAlex Tomas min = max; 808c9de560dSAlex Tomas chunk = 1 << min; 809c9de560dSAlex Tomas 810c9de560dSAlex Tomas /* mark multiblock chunks only */ 811c9de560dSAlex Tomas grp->bb_counters[min]++; 812c9de560dSAlex Tomas if (min > 0) 813c9de560dSAlex Tomas mb_clear_bit(first >> min, 814c9de560dSAlex Tomas buddy + sbi->s_mb_offsets[min]); 815c9de560dSAlex Tomas 816c9de560dSAlex Tomas len -= chunk; 817c9de560dSAlex Tomas first += chunk; 818c9de560dSAlex Tomas } 819c9de560dSAlex Tomas } 820c9de560dSAlex Tomas 82183e80a6eSJan Kara static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) 822196e402aSHarshad Shirwadkar { 82383e80a6eSJan Kara int order; 824196e402aSHarshad Shirwadkar 825196e402aSHarshad Shirwadkar /* 82683e80a6eSJan Kara * We don't bother with a special lists groups with only 1 block free 82783e80a6eSJan Kara * extents and for completely empty groups. 828196e402aSHarshad Shirwadkar */ 82983e80a6eSJan Kara order = fls(len) - 2; 83083e80a6eSJan Kara if (order < 0) 83183e80a6eSJan Kara return 0; 83283e80a6eSJan Kara if (order == MB_NUM_ORDERS(sb)) 83383e80a6eSJan Kara order--; 83413df4d44SBaokun Li if (WARN_ON_ONCE(order > MB_NUM_ORDERS(sb))) 83513df4d44SBaokun Li order = MB_NUM_ORDERS(sb) - 1; 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 846993bf0f4SBaokun Li if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_fragments == 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, 875438a35e7SKemeng Shi enum criteria *new_cr, ext4_group_t *group) 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, 949438a35e7SKemeng Shi enum criteria *new_cr, ext4_group_t *group) 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, 994438a35e7SKemeng Shi enum criteria *new_cr, ext4_group_t *group) 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; 101313df4d44SBaokun Li if (WARN_ON_ONCE(order - 1 > MB_NUM_ORDERS(ac->ac_sb))) 101413df4d44SBaokun Li order = MB_NUM_ORDERS(ac->ac_sb); 1015f52f3d2bSOjaswin Mujoo min_order = order - sbi->s_mb_best_avail_max_trim_order; 10167e170922SOjaswin Mujoo if (min_order < 0) 10177e170922SOjaswin Mujoo min_order = 0; 10187e170922SOjaswin Mujoo 10197e170922SOjaswin Mujoo if (sbi->s_stripe > 0) { 10207e170922SOjaswin Mujoo /* 10217e170922SOjaswin Mujoo * We are assuming that stripe size is always a multiple of 10227e170922SOjaswin Mujoo * cluster ratio otherwise __ext4_fill_super exists early. 10237e170922SOjaswin Mujoo */ 10247e170922SOjaswin Mujoo num_stripe_clusters = EXT4_NUM_B2C(sbi, sbi->s_stripe); 10257e170922SOjaswin Mujoo if (1 << min_order < num_stripe_clusters) 10265d5460faSOjaswin Mujoo /* 10275d5460faSOjaswin Mujoo * We consider 1 order less because later we round 10285d5460faSOjaswin Mujoo * up the goal len to num_stripe_clusters 10295d5460faSOjaswin Mujoo */ 10305d5460faSOjaswin Mujoo min_order = fls(num_stripe_clusters) - 1; 10317e170922SOjaswin Mujoo } 10327e170922SOjaswin Mujoo 10335d5460faSOjaswin Mujoo if (1 << min_order < ac->ac_o_ex.fe_len) 10345d5460faSOjaswin Mujoo min_order = fls(ac->ac_o_ex.fe_len); 10355d5460faSOjaswin Mujoo 10367e170922SOjaswin Mujoo for (i = order; i >= min_order; i--) { 10377e170922SOjaswin Mujoo int frag_order; 10387e170922SOjaswin Mujoo /* 10397e170922SOjaswin Mujoo * Scale down goal len to make sure we find something 10407e170922SOjaswin Mujoo * in the free fragments list. Basically, reduce 10417e170922SOjaswin Mujoo * preallocations. 10427e170922SOjaswin Mujoo */ 10437e170922SOjaswin Mujoo ac->ac_g_ex.fe_len = 1 << i; 10447e170922SOjaswin Mujoo 10457e170922SOjaswin Mujoo if (num_stripe_clusters > 0) { 10467e170922SOjaswin Mujoo /* 10474c0cfebdSTheodore Ts'o * Try to round up the adjusted goal length to 10484c0cfebdSTheodore Ts'o * stripe size (in cluster units) multiple for 10494c0cfebdSTheodore Ts'o * efficiency. 10507e170922SOjaswin Mujoo */ 10517e170922SOjaswin Mujoo ac->ac_g_ex.fe_len = roundup(ac->ac_g_ex.fe_len, 10527e170922SOjaswin Mujoo num_stripe_clusters); 10537e170922SOjaswin Mujoo } 10547e170922SOjaswin Mujoo 10557e170922SOjaswin Mujoo frag_order = mb_avg_fragment_size_order(ac->ac_sb, 10567e170922SOjaswin Mujoo ac->ac_g_ex.fe_len); 10577e170922SOjaswin Mujoo 10587e170922SOjaswin Mujoo grp = ext4_mb_find_good_group_avg_frag_lists(ac, frag_order); 10597e170922SOjaswin Mujoo if (grp) { 10607e170922SOjaswin Mujoo *group = grp->bb_group; 1061f52f3d2bSOjaswin Mujoo ac->ac_flags |= EXT4_MB_CR_BEST_AVAIL_LEN_OPTIMIZED; 1062bcb123acSKemeng Shi return; 1063bcb123acSKemeng Shi } 1064bcb123acSKemeng Shi } 1065bcb123acSKemeng Shi 1066f52f3d2bSOjaswin Mujoo /* Reset goal length to original goal length before falling into CR_GOAL_LEN_SLOW */ 10677e170922SOjaswin Mujoo ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; 1068f52f3d2bSOjaswin Mujoo *new_cr = CR_GOAL_LEN_SLOW; 1069196e402aSHarshad Shirwadkar } 1070196e402aSHarshad Shirwadkar 1071196e402aSHarshad Shirwadkar static inline int should_optimize_scan(struct ext4_allocation_context *ac) 1072196e402aSHarshad Shirwadkar { 1073196e402aSHarshad Shirwadkar if (unlikely(!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN))) 1074196e402aSHarshad Shirwadkar return 0; 1075f52f3d2bSOjaswin Mujoo if (ac->ac_criteria >= CR_GOAL_LEN_SLOW) 1076196e402aSHarshad Shirwadkar return 0; 1077077d0c2cSOjaswin Mujoo if (!ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) 1078196e402aSHarshad Shirwadkar return 0; 1079196e402aSHarshad Shirwadkar return 1; 1080196e402aSHarshad Shirwadkar } 1081196e402aSHarshad Shirwadkar 1082196e402aSHarshad Shirwadkar /* 1083da5704eeSKemeng Shi * Return next linear group for allocation. 1084196e402aSHarshad Shirwadkar */ 108560c672b7SKemeng Shi static ext4_group_t 1086da5704eeSKemeng Shi next_linear_group(ext4_group_t group, ext4_group_t ngroups) 1087196e402aSHarshad Shirwadkar { 1088196e402aSHarshad Shirwadkar /* 1089196e402aSHarshad Shirwadkar * Artificially restricted ngroups for non-extent 1090196e402aSHarshad Shirwadkar * files makes group > ngroups possible on first loop. 1091196e402aSHarshad Shirwadkar */ 1092196e402aSHarshad Shirwadkar return group + 1 >= ngroups ? 0 : group + 1; 1093196e402aSHarshad Shirwadkar } 1094196e402aSHarshad Shirwadkar 1095196e402aSHarshad Shirwadkar /* 1096196e402aSHarshad Shirwadkar * ext4_mb_choose_next_group: choose next group for allocation. 1097196e402aSHarshad Shirwadkar * 1098196e402aSHarshad Shirwadkar * @ac Allocation Context 1099196e402aSHarshad Shirwadkar * @new_cr This is an output parameter. If the there is no good group 1100196e402aSHarshad Shirwadkar * available at current CR level, this field is updated to indicate 1101196e402aSHarshad Shirwadkar * the new cr level that should be used. 1102196e402aSHarshad Shirwadkar * @group This is an input / output parameter. As an input it indicates the 1103196e402aSHarshad Shirwadkar * next group that the allocator intends to use for allocation. As 1104196e402aSHarshad Shirwadkar * output, this field indicates the next group that should be used as 1105196e402aSHarshad Shirwadkar * determined by the optimization functions. 1106196e402aSHarshad Shirwadkar * @ngroups Total number of groups 1107196e402aSHarshad Shirwadkar */ 1108196e402aSHarshad Shirwadkar static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, 11094eb7a4a1SOjaswin Mujoo enum criteria *new_cr, ext4_group_t *group, ext4_group_t ngroups) 1110196e402aSHarshad Shirwadkar { 1111196e402aSHarshad Shirwadkar *new_cr = ac->ac_criteria; 1112196e402aSHarshad Shirwadkar 1113da5704eeSKemeng Shi if (!should_optimize_scan(ac)) { 1114da5704eeSKemeng Shi *group = next_linear_group(*group, ngroups); 1115da5704eeSKemeng Shi return; 1116da5704eeSKemeng Shi } 1117da5704eeSKemeng Shi 1118da5704eeSKemeng Shi /* 1119da5704eeSKemeng Shi * Optimized scanning can return non adjacent groups which can cause 1120da5704eeSKemeng Shi * seek overhead for rotational disks. So try few linear groups before 1121da5704eeSKemeng Shi * trying optimized scan. 1122da5704eeSKemeng Shi */ 1123da5704eeSKemeng Shi if (ac->ac_groups_linear_remaining) { 1124da5704eeSKemeng Shi *group = next_linear_group(*group, ngroups); 1125da5704eeSKemeng Shi ac->ac_groups_linear_remaining--; 1126196e402aSHarshad Shirwadkar return; 11274fca50d4SJan Kara } 1128196e402aSHarshad Shirwadkar 1129f52f3d2bSOjaswin Mujoo if (*new_cr == CR_POWER2_ALIGNED) { 1130438a35e7SKemeng Shi ext4_mb_choose_next_group_p2_aligned(ac, new_cr, group); 1131f52f3d2bSOjaswin Mujoo } else if (*new_cr == CR_GOAL_LEN_FAST) { 1132438a35e7SKemeng Shi ext4_mb_choose_next_group_goal_fast(ac, new_cr, group); 1133f52f3d2bSOjaswin Mujoo } else if (*new_cr == CR_BEST_AVAIL_LEN) { 1134438a35e7SKemeng Shi ext4_mb_choose_next_group_best_avail(ac, new_cr, group); 1135196e402aSHarshad Shirwadkar } else { 1136196e402aSHarshad Shirwadkar /* 11372caffb6aSKemeng Shi * TODO: For CR_GOAL_LEN_SLOW, we can arrange groups in an 11382caffb6aSKemeng Shi * rb tree sorted by bb_free. But until that happens, we should 11392caffb6aSKemeng Shi * never come here. 1140196e402aSHarshad Shirwadkar */ 1141196e402aSHarshad Shirwadkar WARN_ON(1); 1142196e402aSHarshad Shirwadkar } 1143196e402aSHarshad Shirwadkar } 1144196e402aSHarshad Shirwadkar 11458a57d9d6SCurt Wohlgemuth /* 11468a57d9d6SCurt Wohlgemuth * Cache the order of the largest free extent we have available in this block 11478a57d9d6SCurt Wohlgemuth * group. 11488a57d9d6SCurt Wohlgemuth */ 11498a57d9d6SCurt Wohlgemuth static void 11508a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) 11518a57d9d6SCurt Wohlgemuth { 1152196e402aSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 11538a57d9d6SCurt Wohlgemuth int i; 11548a57d9d6SCurt Wohlgemuth 11551940265eSJan Kara for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) 11561940265eSJan Kara if (grp->bb_counters[i] > 0) 11571940265eSJan Kara break; 11581940265eSJan Kara /* No need to move between order lists? */ 11591940265eSJan Kara if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || 11601940265eSJan Kara i == grp->bb_largest_free_order) { 11611940265eSJan Kara grp->bb_largest_free_order = i; 11621940265eSJan Kara return; 11631940265eSJan Kara } 11641940265eSJan Kara 11651940265eSJan Kara if (grp->bb_largest_free_order >= 0) { 1166196e402aSHarshad Shirwadkar write_lock(&sbi->s_mb_largest_free_orders_locks[ 1167196e402aSHarshad Shirwadkar grp->bb_largest_free_order]); 1168196e402aSHarshad Shirwadkar list_del_init(&grp->bb_largest_free_order_node); 1169196e402aSHarshad Shirwadkar write_unlock(&sbi->s_mb_largest_free_orders_locks[ 1170196e402aSHarshad Shirwadkar grp->bb_largest_free_order]); 1171196e402aSHarshad Shirwadkar } 11728a57d9d6SCurt Wohlgemuth grp->bb_largest_free_order = i; 11731940265eSJan Kara if (grp->bb_largest_free_order >= 0 && grp->bb_free) { 1174196e402aSHarshad Shirwadkar write_lock(&sbi->s_mb_largest_free_orders_locks[ 1175196e402aSHarshad Shirwadkar grp->bb_largest_free_order]); 1176196e402aSHarshad Shirwadkar list_add_tail(&grp->bb_largest_free_order_node, 1177196e402aSHarshad Shirwadkar &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); 1178196e402aSHarshad Shirwadkar write_unlock(&sbi->s_mb_largest_free_orders_locks[ 1179196e402aSHarshad Shirwadkar grp->bb_largest_free_order]); 1180196e402aSHarshad Shirwadkar } 11818a57d9d6SCurt Wohlgemuth } 11828a57d9d6SCurt Wohlgemuth 1183089ceeccSEric Sandeen static noinline_for_stack 1184089ceeccSEric Sandeen void ext4_mb_generate_buddy(struct super_block *sb, 11855354b2afSTheodore Ts'o void *buddy, void *bitmap, ext4_group_t group, 11865354b2afSTheodore Ts'o struct ext4_group_info *grp) 1187c9de560dSAlex Tomas { 1188e43bb4e6SNamjae Jeon struct ext4_sb_info *sbi = EXT4_SB(sb); 11897137d7a4STheodore Ts'o ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 1190a36b4498SEric Sandeen ext4_grpblk_t i = 0; 1191a36b4498SEric Sandeen ext4_grpblk_t first; 1192a36b4498SEric Sandeen ext4_grpblk_t len; 1193c9de560dSAlex Tomas unsigned free = 0; 1194c9de560dSAlex Tomas unsigned fragments = 0; 1195c9de560dSAlex Tomas unsigned long long period = get_cycles(); 1196c9de560dSAlex Tomas 1197c9de560dSAlex Tomas /* initialize buddy from bitmap which is aggregation 1198c9de560dSAlex Tomas * of on-disk bitmap and preallocations */ 1199ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, 0); 1200c9de560dSAlex Tomas grp->bb_first_free = i; 1201c9de560dSAlex Tomas while (i < max) { 1202c9de560dSAlex Tomas fragments++; 1203c9de560dSAlex Tomas first = i; 1204ffad0a44SAneesh Kumar K.V i = mb_find_next_bit(bitmap, max, i); 1205c9de560dSAlex Tomas len = i - first; 1206c9de560dSAlex Tomas free += len; 1207c9de560dSAlex Tomas if (len > 1) 1208c9de560dSAlex Tomas ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 1209c9de560dSAlex Tomas else 1210c9de560dSAlex Tomas grp->bb_counters[0]++; 1211c9de560dSAlex Tomas if (i < max) 1212ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, i); 1213c9de560dSAlex Tomas } 1214c9de560dSAlex Tomas grp->bb_fragments = fragments; 1215c9de560dSAlex Tomas 1216c9de560dSAlex Tomas if (free != grp->bb_free) { 1217e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, 121894d4c066STheodore Ts'o "block bitmap and bg descriptor " 121994d4c066STheodore Ts'o "inconsistent: %u vs %u free clusters", 1220e29136f8STheodore Ts'o free, grp->bb_free); 1221e56eb659SAneesh Kumar K.V /* 1222163a203dSDarrick J. Wong * If we intend to continue, we consider group descriptor 1223e56eb659SAneesh Kumar K.V * corrupt and update bb_free using bitmap value 1224e56eb659SAneesh Kumar K.V */ 1225c9de560dSAlex Tomas grp->bb_free = free; 1226db79e6d1SWang Shilong ext4_mark_group_bitmap_corrupted(sb, group, 1227db79e6d1SWang Shilong EXT4_GROUP_INFO_BBITMAP_CORRUPT); 1228c9de560dSAlex Tomas } 12298a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, grp); 123083e80a6eSJan Kara mb_update_avg_fragment_size(sb, grp); 1231c9de560dSAlex Tomas 1232c9de560dSAlex Tomas clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 1233c9de560dSAlex Tomas 1234c9de560dSAlex Tomas period = get_cycles() - period; 123567d25186SHarshad Shirwadkar atomic_inc(&sbi->s_mb_buddies_generated); 123667d25186SHarshad Shirwadkar atomic64_add(period, &sbi->s_mb_generation_time); 1237c9de560dSAlex Tomas } 1238c9de560dSAlex Tomas 1239c9b528c3SBaokun Li static void mb_regenerate_buddy(struct ext4_buddy *e4b) 1240c9b528c3SBaokun Li { 1241c9b528c3SBaokun Li int count; 1242c9b528c3SBaokun Li int order = 1; 1243c9b528c3SBaokun Li void *buddy; 1244c9b528c3SBaokun Li 1245c9b528c3SBaokun Li while ((buddy = mb_find_buddy(e4b, order++, &count))) 1246c9b528c3SBaokun Li mb_set_bits(buddy, 0, count); 1247c9b528c3SBaokun Li 1248c9b528c3SBaokun Li e4b->bd_info->bb_fragments = 0; 1249c9b528c3SBaokun Li memset(e4b->bd_info->bb_counters, 0, 1250c9b528c3SBaokun Li sizeof(*e4b->bd_info->bb_counters) * 1251c9b528c3SBaokun Li (e4b->bd_sb->s_blocksize_bits + 2)); 1252c9b528c3SBaokun Li 1253c9b528c3SBaokun Li ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, 1254c9b528c3SBaokun Li e4b->bd_bitmap, e4b->bd_group, e4b->bd_info); 1255c9b528c3SBaokun Li } 1256c9b528c3SBaokun Li 1257c9de560dSAlex Tomas /* The buddy information is attached the buddy cache inode 1258c9de560dSAlex Tomas * for convenience. The information regarding each group 1259c9de560dSAlex Tomas * is loaded via ext4_mb_load_buddy. The information involve 1260c9de560dSAlex Tomas * block bitmap and buddy information. The information are 1261c9de560dSAlex Tomas * stored in the inode as 1262c9de560dSAlex Tomas * 1263c9de560dSAlex Tomas * { page } 1264c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 1265c9de560dSAlex Tomas * 1266c9de560dSAlex Tomas * 1267c9de560dSAlex Tomas * one block each for bitmap and buddy information. 1268c9de560dSAlex Tomas * So for each group we take up 2 blocks. A page can 1269ea1754a0SKirill A. Shutemov * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. 1270c9de560dSAlex Tomas * So it can have information regarding groups_per_page which 1271c9de560dSAlex Tomas * is blocks_per_page/2 12728a57d9d6SCurt Wohlgemuth * 12738a57d9d6SCurt Wohlgemuth * Locking note: This routine takes the block group lock of all groups 12748a57d9d6SCurt Wohlgemuth * for this page; do not hold this lock when calling this routine! 1275c9de560dSAlex Tomas */ 1276c9de560dSAlex Tomas 1277e1622a0dSMatthew Wilcox (Oracle) static int ext4_mb_init_cache(struct folio *folio, char *incore, gfp_t gfp) 1278c9de560dSAlex Tomas { 12798df9675fSTheodore Ts'o ext4_group_t ngroups; 128089cadf6eSLu Hongfei unsigned int blocksize; 1281c9de560dSAlex Tomas int blocks_per_page; 1282c9de560dSAlex Tomas int groups_per_page; 1283c9de560dSAlex Tomas int err = 0; 1284c9de560dSAlex Tomas int i; 1285813e5727STheodore Ts'o ext4_group_t first_group, group; 1286c9de560dSAlex Tomas int first_block; 1287c9de560dSAlex Tomas struct super_block *sb; 1288c9de560dSAlex Tomas struct buffer_head *bhs; 1289fa77dcfaSDarrick J. Wong struct buffer_head **bh = NULL; 1290c9de560dSAlex Tomas struct inode *inode; 1291c9de560dSAlex Tomas char *data; 1292c9de560dSAlex Tomas char *bitmap; 12939b8b7d35SAmir Goldstein struct ext4_group_info *grinfo; 1294c9de560dSAlex Tomas 1295e1622a0dSMatthew Wilcox (Oracle) inode = folio->mapping->host; 1296c9de560dSAlex Tomas sb = inode->i_sb; 12978df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 129893407472SFabian Frederick blocksize = i_blocksize(inode); 129909cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / blocksize; 1300c9de560dSAlex Tomas 1301e1622a0dSMatthew Wilcox (Oracle) mb_debug(sb, "init folio %lu\n", folio->index); 1302d3df1453SRitesh Harjani 1303c9de560dSAlex Tomas groups_per_page = blocks_per_page >> 1; 1304c9de560dSAlex Tomas if (groups_per_page == 0) 1305c9de560dSAlex Tomas groups_per_page = 1; 1306c9de560dSAlex Tomas 1307c9de560dSAlex Tomas /* allocate buffer_heads to read bitmaps */ 1308c9de560dSAlex Tomas if (groups_per_page > 1) { 1309c9de560dSAlex Tomas i = sizeof(struct buffer_head *) * groups_per_page; 1310adb7ef60SKonstantin Khlebnikov bh = kzalloc(i, gfp); 1311139f46d3SKemeng Shi if (bh == NULL) 1312139f46d3SKemeng Shi return -ENOMEM; 1313c9de560dSAlex Tomas } else 1314c9de560dSAlex Tomas bh = &bhs; 1315c9de560dSAlex Tomas 1316e1622a0dSMatthew Wilcox (Oracle) first_group = folio->index * blocks_per_page / 2; 1317c9de560dSAlex Tomas 1318e1622a0dSMatthew Wilcox (Oracle) /* read all groups the folio covers into the cache */ 1319813e5727STheodore Ts'o for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 1320813e5727STheodore Ts'o if (group >= ngroups) 1321c9de560dSAlex Tomas break; 1322c9de560dSAlex Tomas 1323813e5727STheodore Ts'o grinfo = ext4_get_group_info(sb, group); 13245354b2afSTheodore Ts'o if (!grinfo) 13255354b2afSTheodore Ts'o continue; 13269b8b7d35SAmir Goldstein /* 13279b8b7d35SAmir Goldstein * If page is uptodate then we came here after online resize 13289b8b7d35SAmir Goldstein * which added some new uninitialized group info structs, so 1329e1622a0dSMatthew Wilcox (Oracle) * we must skip all initialized uptodate buddies on the folio, 13309b8b7d35SAmir Goldstein * which may be currently in use by an allocating task. 13319b8b7d35SAmir Goldstein */ 1332e1622a0dSMatthew Wilcox (Oracle) if (folio_test_uptodate(folio) && 1333e1622a0dSMatthew Wilcox (Oracle) !EXT4_MB_GRP_NEED_INIT(grinfo)) { 13349b8b7d35SAmir Goldstein bh[i] = NULL; 13359b8b7d35SAmir Goldstein continue; 13369b8b7d35SAmir Goldstein } 1337cfd73237SAlex Zhuravlev bh[i] = ext4_read_block_bitmap_nowait(sb, group, false); 13389008a58eSDarrick J. Wong if (IS_ERR(bh[i])) { 13399008a58eSDarrick J. Wong err = PTR_ERR(bh[i]); 13409008a58eSDarrick J. Wong bh[i] = NULL; 1341c9de560dSAlex Tomas goto out; 13422ccb5fb9SAneesh Kumar K.V } 1343d3df1453SRitesh Harjani mb_debug(sb, "read bitmap for group %u\n", group); 1344c9de560dSAlex Tomas } 1345c9de560dSAlex Tomas 1346c9de560dSAlex Tomas /* wait for I/O completion */ 1347813e5727STheodore Ts'o for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 13489008a58eSDarrick J. Wong int err2; 13499008a58eSDarrick J. Wong 13509008a58eSDarrick J. Wong if (!bh[i]) 13519008a58eSDarrick J. Wong continue; 13529008a58eSDarrick J. Wong err2 = ext4_wait_block_bitmap(sb, group, bh[i]); 13539008a58eSDarrick J. Wong if (!err) 13549008a58eSDarrick J. Wong err = err2; 1355813e5727STheodore Ts'o } 1356c9de560dSAlex Tomas 1357e1622a0dSMatthew Wilcox (Oracle) first_block = folio->index * blocks_per_page; 1358c9de560dSAlex Tomas for (i = 0; i < blocks_per_page; i++) { 1359c9de560dSAlex Tomas group = (first_block + i) >> 1; 13608df9675fSTheodore Ts'o if (group >= ngroups) 1361c9de560dSAlex Tomas break; 1362c9de560dSAlex Tomas 13639b8b7d35SAmir Goldstein if (!bh[group - first_group]) 13649b8b7d35SAmir Goldstein /* skip initialized uptodate buddy */ 13659b8b7d35SAmir Goldstein continue; 13669b8b7d35SAmir Goldstein 1367bbdc322fSLukas Czerner if (!buffer_verified(bh[group - first_group])) 1368bbdc322fSLukas Czerner /* Skip faulty bitmaps */ 1369bbdc322fSLukas Czerner continue; 1370bbdc322fSLukas Czerner err = 0; 1371bbdc322fSLukas Czerner 1372c9de560dSAlex Tomas /* 1373c9de560dSAlex Tomas * data carry information regarding this 1374c9de560dSAlex Tomas * particular group in the format specified 1375c9de560dSAlex Tomas * above 1376c9de560dSAlex Tomas * 1377c9de560dSAlex Tomas */ 1378e1622a0dSMatthew Wilcox (Oracle) data = folio_address(folio) + (i * blocksize); 1379c9de560dSAlex Tomas bitmap = bh[group - first_group]->b_data; 1380c9de560dSAlex Tomas 1381c9de560dSAlex Tomas /* 1382c9de560dSAlex Tomas * We place the buddy block and bitmap block 1383c9de560dSAlex Tomas * close together 1384c9de560dSAlex Tomas */ 1385ebf6cb7cSWang Jianjian grinfo = ext4_get_group_info(sb, group); 1386ebf6cb7cSWang Jianjian if (!grinfo) { 1387ebf6cb7cSWang Jianjian err = -EFSCORRUPTED; 1388ebf6cb7cSWang Jianjian goto out; 1389ebf6cb7cSWang Jianjian } 1390c9de560dSAlex Tomas if ((first_block + i) & 1) { 1391c9de560dSAlex Tomas /* this is block of buddy */ 1392c9de560dSAlex Tomas BUG_ON(incore == NULL); 1393e1622a0dSMatthew Wilcox (Oracle) mb_debug(sb, "put buddy for group %u in folio %lu/%x\n", 1394e1622a0dSMatthew Wilcox (Oracle) group, folio->index, i * blocksize); 1395f307333eSTheodore Ts'o trace_ext4_mb_buddy_bitmap_load(sb, group); 1396c9de560dSAlex Tomas grinfo->bb_fragments = 0; 1397c9de560dSAlex Tomas memset(grinfo->bb_counters, 0, 13981927805eSEric Sandeen sizeof(*grinfo->bb_counters) * 13994b68f6dfSHarshad Shirwadkar (MB_NUM_ORDERS(sb))); 1400c9de560dSAlex Tomas /* 1401c9de560dSAlex Tomas * incore got set to the group block bitmap below 1402c9de560dSAlex Tomas */ 14037a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, group); 14049b8b7d35SAmir Goldstein /* init the buddy */ 14059b8b7d35SAmir Goldstein memset(data, 0xff, blocksize); 14065354b2afSTheodore Ts'o ext4_mb_generate_buddy(sb, data, incore, group, grinfo); 14077a2fcbf7SAneesh Kumar K.V ext4_unlock_group(sb, group); 1408c9de560dSAlex Tomas incore = NULL; 1409c9de560dSAlex Tomas } else { 1410c9de560dSAlex Tomas /* this is block of bitmap */ 1411c9de560dSAlex Tomas BUG_ON(incore != NULL); 1412e1622a0dSMatthew Wilcox (Oracle) mb_debug(sb, "put bitmap for group %u in folio %lu/%x\n", 1413e1622a0dSMatthew Wilcox (Oracle) group, folio->index, i * blocksize); 1414f307333eSTheodore Ts'o trace_ext4_mb_bitmap_load(sb, group); 1415c9de560dSAlex Tomas 1416c9de560dSAlex Tomas /* see comments in ext4_mb_put_pa() */ 1417c9de560dSAlex Tomas ext4_lock_group(sb, group); 1418c9de560dSAlex Tomas memcpy(data, bitmap, blocksize); 1419c9de560dSAlex Tomas 1420c9de560dSAlex Tomas /* mark all preallocated blks used in in-core bitmap */ 1421c9de560dSAlex Tomas ext4_mb_generate_from_pa(sb, data, group); 1422ebf6cb7cSWang Jianjian WARN_ON_ONCE(!RB_EMPTY_ROOT(&grinfo->bb_free_root)); 1423c9de560dSAlex Tomas ext4_unlock_group(sb, group); 1424c9de560dSAlex Tomas 1425c9de560dSAlex Tomas /* set incore so that the buddy information can be 1426c9de560dSAlex Tomas * generated using this 1427c9de560dSAlex Tomas */ 1428c9de560dSAlex Tomas incore = data; 1429c9de560dSAlex Tomas } 1430c9de560dSAlex Tomas } 1431e1622a0dSMatthew Wilcox (Oracle) folio_mark_uptodate(folio); 1432c9de560dSAlex Tomas 1433c9de560dSAlex Tomas out: 1434c9de560dSAlex Tomas if (bh) { 14359b8b7d35SAmir Goldstein for (i = 0; i < groups_per_page; i++) 1436c9de560dSAlex Tomas brelse(bh[i]); 1437c9de560dSAlex Tomas if (bh != &bhs) 1438c9de560dSAlex Tomas kfree(bh); 1439c9de560dSAlex Tomas } 1440c9de560dSAlex Tomas return err; 1441c9de560dSAlex Tomas } 1442c9de560dSAlex Tomas 14438a57d9d6SCurt Wohlgemuth /* 14442de8807bSAmir Goldstein * Lock the buddy and bitmap pages. This make sure other parallel init_group 14452de8807bSAmir Goldstein * on the same buddy page doesn't happen whild holding the buddy page lock. 14462de8807bSAmir Goldstein * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap 14475eea586bSMatthew Wilcox (Oracle) * are on the same page e4b->bd_buddy_folio is NULL and return value is 0. 1448eee4adc7SEric Sandeen */ 14492de8807bSAmir Goldstein static int ext4_mb_get_buddy_page_lock(struct super_block *sb, 1450adb7ef60SKonstantin Khlebnikov ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) 1451eee4adc7SEric Sandeen { 14522de8807bSAmir Goldstein struct inode *inode = EXT4_SB(sb)->s_buddy_cache; 14532de8807bSAmir Goldstein int block, pnum, poff; 1454eee4adc7SEric Sandeen int blocks_per_page; 145599b150d8SMatthew Wilcox (Oracle) struct folio *folio; 14562de8807bSAmir Goldstein 14575eea586bSMatthew Wilcox (Oracle) e4b->bd_buddy_folio = NULL; 145899b150d8SMatthew Wilcox (Oracle) e4b->bd_bitmap_folio = NULL; 1459eee4adc7SEric Sandeen 146009cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1461eee4adc7SEric Sandeen /* 1462eee4adc7SEric Sandeen * the buddy cache inode stores the block bitmap 1463eee4adc7SEric Sandeen * and buddy information in consecutive blocks. 1464eee4adc7SEric Sandeen * So for each group we need two blocks. 1465eee4adc7SEric Sandeen */ 1466eee4adc7SEric Sandeen block = group * 2; 1467eee4adc7SEric Sandeen pnum = block / blocks_per_page; 14682de8807bSAmir Goldstein poff = block % blocks_per_page; 146999b150d8SMatthew Wilcox (Oracle) folio = __filemap_get_folio(inode->i_mapping, pnum, 147099b150d8SMatthew Wilcox (Oracle) FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); 147199b150d8SMatthew Wilcox (Oracle) if (IS_ERR(folio)) 147299b150d8SMatthew Wilcox (Oracle) return PTR_ERR(folio); 147399b150d8SMatthew Wilcox (Oracle) BUG_ON(folio->mapping != inode->i_mapping); 147499b150d8SMatthew Wilcox (Oracle) e4b->bd_bitmap_folio = folio; 147599b150d8SMatthew Wilcox (Oracle) e4b->bd_bitmap = folio_address(folio) + (poff * sb->s_blocksize); 1476eee4adc7SEric Sandeen 14772de8807bSAmir Goldstein if (blocks_per_page >= 2) { 14782de8807bSAmir Goldstein /* buddy and bitmap are on the same page */ 14792de8807bSAmir Goldstein return 0; 1480eee4adc7SEric Sandeen } 1481eee4adc7SEric Sandeen 1482f2fec3e9SGou Hao /* blocks_per_page == 1, hence we need another page for the buddy */ 14835eea586bSMatthew Wilcox (Oracle) folio = __filemap_get_folio(inode->i_mapping, block + 1, 14845eea586bSMatthew Wilcox (Oracle) FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); 14855eea586bSMatthew Wilcox (Oracle) if (IS_ERR(folio)) 14865eea586bSMatthew Wilcox (Oracle) return PTR_ERR(folio); 14875eea586bSMatthew Wilcox (Oracle) BUG_ON(folio->mapping != inode->i_mapping); 14885eea586bSMatthew Wilcox (Oracle) e4b->bd_buddy_folio = folio; 14892de8807bSAmir Goldstein return 0; 1490eee4adc7SEric Sandeen } 1491eee4adc7SEric Sandeen 14922de8807bSAmir Goldstein static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) 14932de8807bSAmir Goldstein { 149499b150d8SMatthew Wilcox (Oracle) if (e4b->bd_bitmap_folio) { 149599b150d8SMatthew Wilcox (Oracle) folio_unlock(e4b->bd_bitmap_folio); 149699b150d8SMatthew Wilcox (Oracle) folio_put(e4b->bd_bitmap_folio); 14972de8807bSAmir Goldstein } 14985eea586bSMatthew Wilcox (Oracle) if (e4b->bd_buddy_folio) { 14995eea586bSMatthew Wilcox (Oracle) folio_unlock(e4b->bd_buddy_folio); 15005eea586bSMatthew Wilcox (Oracle) folio_put(e4b->bd_buddy_folio); 15012de8807bSAmir Goldstein } 1502eee4adc7SEric Sandeen } 1503eee4adc7SEric Sandeen 1504eee4adc7SEric Sandeen /* 15058a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 15068a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 15078a57d9d6SCurt Wohlgemuth * calling this routine! 15088a57d9d6SCurt Wohlgemuth */ 1509b6a758ecSAneesh Kumar K.V static noinline_for_stack 1510adb7ef60SKonstantin Khlebnikov int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) 1511b6a758ecSAneesh Kumar K.V { 1512b6a758ecSAneesh Kumar K.V 1513b6a758ecSAneesh Kumar K.V struct ext4_group_info *this_grp; 15142de8807bSAmir Goldstein struct ext4_buddy e4b; 151599b150d8SMatthew Wilcox (Oracle) struct folio *folio; 15162de8807bSAmir Goldstein int ret = 0; 1517b6a758ecSAneesh Kumar K.V 1518b10a44c3STheodore Ts'o might_sleep(); 1519d3df1453SRitesh Harjani mb_debug(sb, "init group %u\n", group); 1520b6a758ecSAneesh Kumar K.V this_grp = ext4_get_group_info(sb, group); 15215354b2afSTheodore Ts'o if (!this_grp) 15225354b2afSTheodore Ts'o return -EFSCORRUPTED; 15235354b2afSTheodore Ts'o 1524b6a758ecSAneesh Kumar K.V /* 152508c3a813SAneesh Kumar K.V * This ensures that we don't reinit the buddy cache 152608c3a813SAneesh Kumar K.V * page which map to the group from which we are already 152708c3a813SAneesh Kumar K.V * allocating. If we are looking at the buddy cache we would 152808c3a813SAneesh Kumar K.V * have taken a reference using ext4_mb_load_buddy and that 15292de8807bSAmir Goldstein * would have pinned buddy page to page cache. 15302457aec6SMel Gorman * The call to ext4_mb_get_buddy_page_lock will mark the 15312457aec6SMel Gorman * page accessed. 1532b6a758ecSAneesh Kumar K.V */ 1533adb7ef60SKonstantin Khlebnikov ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); 15342de8807bSAmir Goldstein if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { 1535b6a758ecSAneesh Kumar K.V /* 1536b6a758ecSAneesh Kumar K.V * somebody initialized the group 1537b6a758ecSAneesh Kumar K.V * return without doing anything 1538b6a758ecSAneesh Kumar K.V */ 1539b6a758ecSAneesh Kumar K.V goto err; 1540b6a758ecSAneesh Kumar K.V } 15412de8807bSAmir Goldstein 154299b150d8SMatthew Wilcox (Oracle) folio = e4b.bd_bitmap_folio; 1543e1622a0dSMatthew Wilcox (Oracle) ret = ext4_mb_init_cache(folio, NULL, gfp); 15442de8807bSAmir Goldstein if (ret) 1545b6a758ecSAneesh Kumar K.V goto err; 154699b150d8SMatthew Wilcox (Oracle) if (!folio_test_uptodate(folio)) { 1547b6a758ecSAneesh Kumar K.V ret = -EIO; 1548b6a758ecSAneesh Kumar K.V goto err; 1549b6a758ecSAneesh Kumar K.V } 1550b6a758ecSAneesh Kumar K.V 15515eea586bSMatthew Wilcox (Oracle) if (e4b.bd_buddy_folio == NULL) { 1552b6a758ecSAneesh Kumar K.V /* 1553b6a758ecSAneesh Kumar K.V * If both the bitmap and buddy are in 1554b6a758ecSAneesh Kumar K.V * the same page we don't need to force 1555b6a758ecSAneesh Kumar K.V * init the buddy 1556b6a758ecSAneesh Kumar K.V */ 15572de8807bSAmir Goldstein ret = 0; 1558b6a758ecSAneesh Kumar K.V goto err; 1559b6a758ecSAneesh Kumar K.V } 15602de8807bSAmir Goldstein /* init buddy cache */ 15615eea586bSMatthew Wilcox (Oracle) folio = e4b.bd_buddy_folio; 1562e1622a0dSMatthew Wilcox (Oracle) ret = ext4_mb_init_cache(folio, e4b.bd_bitmap, gfp); 15632de8807bSAmir Goldstein if (ret) 15642de8807bSAmir Goldstein goto err; 15655eea586bSMatthew Wilcox (Oracle) if (!folio_test_uptodate(folio)) { 1566b6a758ecSAneesh Kumar K.V ret = -EIO; 1567b6a758ecSAneesh Kumar K.V goto err; 1568b6a758ecSAneesh Kumar K.V } 1569b6a758ecSAneesh Kumar K.V err: 15702de8807bSAmir Goldstein ext4_mb_put_buddy_page_lock(&e4b); 1571b6a758ecSAneesh Kumar K.V return ret; 1572b6a758ecSAneesh Kumar K.V } 1573b6a758ecSAneesh Kumar K.V 15748a57d9d6SCurt Wohlgemuth /* 15758a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 15768a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 15778a57d9d6SCurt Wohlgemuth * calling this routine! 15788a57d9d6SCurt Wohlgemuth */ 15794ddfef7bSEric Sandeen static noinline_for_stack int 1580adb7ef60SKonstantin Khlebnikov ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, 1581adb7ef60SKonstantin Khlebnikov struct ext4_buddy *e4b, gfp_t gfp) 1582c9de560dSAlex Tomas { 1583c9de560dSAlex Tomas int blocks_per_page; 1584c9de560dSAlex Tomas int block; 1585c9de560dSAlex Tomas int pnum; 1586c9de560dSAlex Tomas int poff; 158799b150d8SMatthew Wilcox (Oracle) struct folio *folio; 1588fdf6c7a7SShen Feng int ret; 1589920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1590920313a7SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 1591920313a7SAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 1592c9de560dSAlex Tomas 1593b10a44c3STheodore Ts'o might_sleep(); 1594d3df1453SRitesh Harjani mb_debug(sb, "load group %u\n", group); 1595c9de560dSAlex Tomas 159609cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1597920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 15985354b2afSTheodore Ts'o if (!grp) 15995354b2afSTheodore Ts'o return -EFSCORRUPTED; 1600c9de560dSAlex Tomas 1601c9de560dSAlex Tomas e4b->bd_blkbits = sb->s_blocksize_bits; 1602529da704STao Ma e4b->bd_info = grp; 1603c9de560dSAlex Tomas e4b->bd_sb = sb; 1604c9de560dSAlex Tomas e4b->bd_group = group; 16055eea586bSMatthew Wilcox (Oracle) e4b->bd_buddy_folio = NULL; 160699b150d8SMatthew Wilcox (Oracle) e4b->bd_bitmap_folio = NULL; 1607c9de560dSAlex Tomas 1608f41c0750SAneesh Kumar K.V if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1609f41c0750SAneesh Kumar K.V /* 1610f41c0750SAneesh Kumar K.V * we need full data about the group 1611f41c0750SAneesh Kumar K.V * to make a good selection 1612f41c0750SAneesh Kumar K.V */ 1613adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_group(sb, group, gfp); 1614f41c0750SAneesh Kumar K.V if (ret) 1615f41c0750SAneesh Kumar K.V return ret; 1616f41c0750SAneesh Kumar K.V } 1617f41c0750SAneesh Kumar K.V 1618c9de560dSAlex Tomas /* 1619c9de560dSAlex Tomas * the buddy cache inode stores the block bitmap 1620c9de560dSAlex Tomas * and buddy information in consecutive blocks. 1621c9de560dSAlex Tomas * So for each group we need two blocks. 1622c9de560dSAlex Tomas */ 1623c9de560dSAlex Tomas block = group * 2; 1624c9de560dSAlex Tomas pnum = block / blocks_per_page; 1625c9de560dSAlex Tomas poff = block % blocks_per_page; 1626c9de560dSAlex Tomas 162799b150d8SMatthew Wilcox (Oracle) /* Avoid locking the folio in the fast path ... */ 162899b150d8SMatthew Wilcox (Oracle) folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0); 162999b150d8SMatthew Wilcox (Oracle) if (IS_ERR(folio) || !folio_test_uptodate(folio)) { 163099b150d8SMatthew Wilcox (Oracle) if (!IS_ERR(folio)) 1631920313a7SAneesh Kumar K.V /* 163299b150d8SMatthew Wilcox (Oracle) * drop the folio reference and try 163399b150d8SMatthew Wilcox (Oracle) * to get the folio with lock. If we 1634920313a7SAneesh Kumar K.V * are not uptodate that implies 163599b150d8SMatthew Wilcox (Oracle) * somebody just created the folio but 163699b150d8SMatthew Wilcox (Oracle) * is yet to initialize it. So 1637920313a7SAneesh Kumar K.V * wait for it to initialize. 1638920313a7SAneesh Kumar K.V */ 163999b150d8SMatthew Wilcox (Oracle) folio_put(folio); 164099b150d8SMatthew Wilcox (Oracle) folio = __filemap_get_folio(inode->i_mapping, pnum, 164199b150d8SMatthew Wilcox (Oracle) FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); 164299b150d8SMatthew Wilcox (Oracle) if (!IS_ERR(folio)) { 164399b150d8SMatthew Wilcox (Oracle) if (WARN_RATELIMIT(folio->mapping != inode->i_mapping, 164499b150d8SMatthew Wilcox (Oracle) "ext4: bitmap's mapping != inode->i_mapping\n")) { 164519b8b035STheodore Ts'o /* should never happen */ 164699b150d8SMatthew Wilcox (Oracle) folio_unlock(folio); 164719b8b035STheodore Ts'o ret = -EINVAL; 164819b8b035STheodore Ts'o goto err; 164919b8b035STheodore Ts'o } 165099b150d8SMatthew Wilcox (Oracle) if (!folio_test_uptodate(folio)) { 1651e1622a0dSMatthew Wilcox (Oracle) ret = ext4_mb_init_cache(folio, NULL, gfp); 1652fdf6c7a7SShen Feng if (ret) { 165399b150d8SMatthew Wilcox (Oracle) folio_unlock(folio); 1654fdf6c7a7SShen Feng goto err; 1655fdf6c7a7SShen Feng } 165699b150d8SMatthew Wilcox (Oracle) mb_cmp_bitmaps(e4b, folio_address(folio) + 1657c9de560dSAlex Tomas (poff * sb->s_blocksize)); 1658c9de560dSAlex Tomas } 165999b150d8SMatthew Wilcox (Oracle) folio_unlock(folio); 1660c9de560dSAlex Tomas } 1661c9de560dSAlex Tomas } 166299b150d8SMatthew Wilcox (Oracle) if (IS_ERR(folio)) { 166399b150d8SMatthew Wilcox (Oracle) ret = PTR_ERR(folio); 1664c57ab39bSYounger Liu goto err; 1665c57ab39bSYounger Liu } 166699b150d8SMatthew Wilcox (Oracle) if (!folio_test_uptodate(folio)) { 1667fdf6c7a7SShen Feng ret = -EIO; 1668c9de560dSAlex Tomas goto err; 1669fdf6c7a7SShen Feng } 16702457aec6SMel Gorman 16715eea586bSMatthew Wilcox (Oracle) /* Folios marked accessed already */ 167299b150d8SMatthew Wilcox (Oracle) e4b->bd_bitmap_folio = folio; 167399b150d8SMatthew Wilcox (Oracle) e4b->bd_bitmap = folio_address(folio) + (poff * sb->s_blocksize); 1674c9de560dSAlex Tomas 1675c9de560dSAlex Tomas block++; 1676c9de560dSAlex Tomas pnum = block / blocks_per_page; 1677c9de560dSAlex Tomas poff = block % blocks_per_page; 1678c9de560dSAlex Tomas 16795eea586bSMatthew Wilcox (Oracle) folio = __filemap_get_folio(inode->i_mapping, pnum, FGP_ACCESSED, 0); 16805eea586bSMatthew Wilcox (Oracle) if (IS_ERR(folio) || !folio_test_uptodate(folio)) { 16815eea586bSMatthew Wilcox (Oracle) if (!IS_ERR(folio)) 16825eea586bSMatthew Wilcox (Oracle) folio_put(folio); 16835eea586bSMatthew Wilcox (Oracle) folio = __filemap_get_folio(inode->i_mapping, pnum, 16845eea586bSMatthew Wilcox (Oracle) FGP_LOCK | FGP_ACCESSED | FGP_CREAT, gfp); 16855eea586bSMatthew Wilcox (Oracle) if (!IS_ERR(folio)) { 16865eea586bSMatthew Wilcox (Oracle) if (WARN_RATELIMIT(folio->mapping != inode->i_mapping, 16875eea586bSMatthew Wilcox (Oracle) "ext4: buddy bitmap's mapping != inode->i_mapping\n")) { 168819b8b035STheodore Ts'o /* should never happen */ 16895eea586bSMatthew Wilcox (Oracle) folio_unlock(folio); 169019b8b035STheodore Ts'o ret = -EINVAL; 169119b8b035STheodore Ts'o goto err; 169219b8b035STheodore Ts'o } 16935eea586bSMatthew Wilcox (Oracle) if (!folio_test_uptodate(folio)) { 1694e1622a0dSMatthew Wilcox (Oracle) ret = ext4_mb_init_cache(folio, e4b->bd_bitmap, 1695adb7ef60SKonstantin Khlebnikov gfp); 1696fdf6c7a7SShen Feng if (ret) { 16975eea586bSMatthew Wilcox (Oracle) folio_unlock(folio); 1698fdf6c7a7SShen Feng goto err; 1699fdf6c7a7SShen Feng } 1700fdf6c7a7SShen Feng } 17015eea586bSMatthew Wilcox (Oracle) folio_unlock(folio); 1702c9de560dSAlex Tomas } 1703c9de560dSAlex Tomas } 17045eea586bSMatthew Wilcox (Oracle) if (IS_ERR(folio)) { 17055eea586bSMatthew Wilcox (Oracle) ret = PTR_ERR(folio); 1706c57ab39bSYounger Liu goto err; 1707c57ab39bSYounger Liu } 17085eea586bSMatthew Wilcox (Oracle) if (!folio_test_uptodate(folio)) { 1709fdf6c7a7SShen Feng ret = -EIO; 1710c9de560dSAlex Tomas goto err; 1711fdf6c7a7SShen Feng } 17122457aec6SMel Gorman 17135eea586bSMatthew Wilcox (Oracle) /* Folios marked accessed already */ 17145eea586bSMatthew Wilcox (Oracle) e4b->bd_buddy_folio = folio; 17155eea586bSMatthew Wilcox (Oracle) e4b->bd_buddy = folio_address(folio) + (poff * sb->s_blocksize); 1716c9de560dSAlex Tomas 1717c9de560dSAlex Tomas return 0; 1718c9de560dSAlex Tomas 1719c9de560dSAlex Tomas err: 1720*c6a6c969SDan Carpenter if (!IS_ERR_OR_NULL(folio)) 17215eea586bSMatthew Wilcox (Oracle) folio_put(folio); 172299b150d8SMatthew Wilcox (Oracle) if (e4b->bd_bitmap_folio) 172399b150d8SMatthew Wilcox (Oracle) folio_put(e4b->bd_bitmap_folio); 1724285164b8SKemeng Shi 1725c9de560dSAlex Tomas e4b->bd_buddy = NULL; 1726c9de560dSAlex Tomas e4b->bd_bitmap = NULL; 1727fdf6c7a7SShen Feng return ret; 1728c9de560dSAlex Tomas } 1729c9de560dSAlex Tomas 1730adb7ef60SKonstantin Khlebnikov static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1731adb7ef60SKonstantin Khlebnikov struct ext4_buddy *e4b) 1732adb7ef60SKonstantin Khlebnikov { 1733adb7ef60SKonstantin Khlebnikov return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); 1734adb7ef60SKonstantin Khlebnikov } 1735adb7ef60SKonstantin Khlebnikov 1736e39e07fdSJing Zhang static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) 1737c9de560dSAlex Tomas { 173899b150d8SMatthew Wilcox (Oracle) if (e4b->bd_bitmap_folio) 173999b150d8SMatthew Wilcox (Oracle) folio_put(e4b->bd_bitmap_folio); 17405eea586bSMatthew Wilcox (Oracle) if (e4b->bd_buddy_folio) 17415eea586bSMatthew Wilcox (Oracle) folio_put(e4b->bd_buddy_folio); 1742c9de560dSAlex Tomas } 1743c9de560dSAlex Tomas 1744c9de560dSAlex Tomas 1745c9de560dSAlex Tomas static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1746c9de560dSAlex Tomas { 1747ce3cca33SChunguang Xu int order = 1, max; 1748c9de560dSAlex Tomas void *bb; 1749c9de560dSAlex Tomas 1750c5e8f3f3STheodore Ts'o BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 1751c9de560dSAlex Tomas BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1752c9de560dSAlex Tomas 1753c9de560dSAlex Tomas while (order <= e4b->bd_blkbits + 1) { 1754ce3cca33SChunguang Xu bb = mb_find_buddy(e4b, order, &max); 1755ce3cca33SChunguang Xu if (!mb_test_bit(block >> order, bb)) { 1756c9de560dSAlex Tomas /* this block is part of buddy of order 'order' */ 1757c9de560dSAlex Tomas return order; 1758c9de560dSAlex Tomas } 1759c9de560dSAlex Tomas order++; 1760c9de560dSAlex Tomas } 1761c9de560dSAlex Tomas return 0; 1762c9de560dSAlex Tomas } 1763c9de560dSAlex Tomas 1764955ce5f5SAneesh Kumar K.V static void mb_clear_bits(void *bm, int cur, int len) 1765c9de560dSAlex Tomas { 1766c9de560dSAlex Tomas __u32 *addr; 1767c9de560dSAlex Tomas 1768c9de560dSAlex Tomas len = cur + len; 1769c9de560dSAlex Tomas while (cur < len) { 1770c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1771c9de560dSAlex Tomas /* fast path: clear whole word at once */ 1772c9de560dSAlex Tomas addr = bm + (cur >> 3); 1773c9de560dSAlex Tomas *addr = 0; 1774c9de560dSAlex Tomas cur += 32; 1775c9de560dSAlex Tomas continue; 1776c9de560dSAlex Tomas } 1777e8134b27SAneesh Kumar K.V mb_clear_bit(cur, bm); 1778c9de560dSAlex Tomas cur++; 1779c9de560dSAlex Tomas } 1780c9de560dSAlex Tomas } 1781c9de560dSAlex Tomas 1782eabe0444SAndrey Sidorov /* clear bits in given range 1783eabe0444SAndrey Sidorov * will return first found zero bit if any, -1 otherwise 1784eabe0444SAndrey Sidorov */ 1785eabe0444SAndrey Sidorov static int mb_test_and_clear_bits(void *bm, int cur, int len) 1786eabe0444SAndrey Sidorov { 1787eabe0444SAndrey Sidorov __u32 *addr; 1788eabe0444SAndrey Sidorov int zero_bit = -1; 1789eabe0444SAndrey Sidorov 1790eabe0444SAndrey Sidorov len = cur + len; 1791eabe0444SAndrey Sidorov while (cur < len) { 1792eabe0444SAndrey Sidorov if ((cur & 31) == 0 && (len - cur) >= 32) { 1793eabe0444SAndrey Sidorov /* fast path: clear whole word at once */ 1794eabe0444SAndrey Sidorov addr = bm + (cur >> 3); 1795eabe0444SAndrey Sidorov if (*addr != (__u32)(-1) && zero_bit == -1) 1796eabe0444SAndrey Sidorov zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); 1797eabe0444SAndrey Sidorov *addr = 0; 1798eabe0444SAndrey Sidorov cur += 32; 1799eabe0444SAndrey Sidorov continue; 1800eabe0444SAndrey Sidorov } 1801eabe0444SAndrey Sidorov if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) 1802eabe0444SAndrey Sidorov zero_bit = cur; 1803eabe0444SAndrey Sidorov cur++; 1804eabe0444SAndrey Sidorov } 1805eabe0444SAndrey Sidorov 1806eabe0444SAndrey Sidorov return zero_bit; 1807eabe0444SAndrey Sidorov } 1808eabe0444SAndrey Sidorov 1809123e3016SRitesh Harjani void mb_set_bits(void *bm, int cur, int len) 1810c9de560dSAlex Tomas { 1811c9de560dSAlex Tomas __u32 *addr; 1812c9de560dSAlex Tomas 1813c9de560dSAlex Tomas len = cur + len; 1814c9de560dSAlex Tomas while (cur < len) { 1815c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1816c9de560dSAlex Tomas /* fast path: set whole word at once */ 1817c9de560dSAlex Tomas addr = bm + (cur >> 3); 1818c9de560dSAlex Tomas *addr = 0xffffffff; 1819c9de560dSAlex Tomas cur += 32; 1820c9de560dSAlex Tomas continue; 1821c9de560dSAlex Tomas } 1822e8134b27SAneesh Kumar K.V mb_set_bit(cur, bm); 1823c9de560dSAlex Tomas cur++; 1824c9de560dSAlex Tomas } 1825c9de560dSAlex Tomas } 1826c9de560dSAlex Tomas 1827eabe0444SAndrey Sidorov static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) 1828eabe0444SAndrey Sidorov { 1829eabe0444SAndrey Sidorov if (mb_test_bit(*bit + side, bitmap)) { 1830eabe0444SAndrey Sidorov mb_clear_bit(*bit, bitmap); 1831eabe0444SAndrey Sidorov (*bit) -= side; 1832eabe0444SAndrey Sidorov return 1; 1833eabe0444SAndrey Sidorov } 1834eabe0444SAndrey Sidorov else { 1835eabe0444SAndrey Sidorov (*bit) += side; 1836eabe0444SAndrey Sidorov mb_set_bit(*bit, bitmap); 1837eabe0444SAndrey Sidorov return -1; 1838eabe0444SAndrey Sidorov } 1839eabe0444SAndrey Sidorov } 1840eabe0444SAndrey Sidorov 1841eabe0444SAndrey Sidorov static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) 1842eabe0444SAndrey Sidorov { 1843eabe0444SAndrey Sidorov int max; 1844eabe0444SAndrey Sidorov int order = 1; 1845eabe0444SAndrey Sidorov void *buddy = mb_find_buddy(e4b, order, &max); 1846eabe0444SAndrey Sidorov 1847eabe0444SAndrey Sidorov while (buddy) { 1848eabe0444SAndrey Sidorov void *buddy2; 1849eabe0444SAndrey Sidorov 1850eabe0444SAndrey Sidorov /* Bits in range [first; last] are known to be set since 1851eabe0444SAndrey Sidorov * corresponding blocks were allocated. Bits in range 1852eabe0444SAndrey Sidorov * (first; last) will stay set because they form buddies on 1853eabe0444SAndrey Sidorov * upper layer. We just deal with borders if they don't 1854eabe0444SAndrey Sidorov * align with upper layer and then go up. 1855eabe0444SAndrey Sidorov * Releasing entire group is all about clearing 1856eabe0444SAndrey Sidorov * single bit of highest order buddy. 1857eabe0444SAndrey Sidorov */ 1858eabe0444SAndrey Sidorov 1859eabe0444SAndrey Sidorov /* Example: 1860eabe0444SAndrey Sidorov * --------------------------------- 1861eabe0444SAndrey Sidorov * | 1 | 1 | 1 | 1 | 1862eabe0444SAndrey Sidorov * --------------------------------- 1863eabe0444SAndrey Sidorov * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1864eabe0444SAndrey Sidorov * --------------------------------- 1865eabe0444SAndrey Sidorov * 0 1 2 3 4 5 6 7 1866eabe0444SAndrey Sidorov * \_____________________/ 1867eabe0444SAndrey Sidorov * 1868eabe0444SAndrey Sidorov * Neither [1] nor [6] is aligned to above layer. 1869eabe0444SAndrey Sidorov * Left neighbour [0] is free, so mark it busy, 1870eabe0444SAndrey Sidorov * decrease bb_counters and extend range to 1871eabe0444SAndrey Sidorov * [0; 6] 1872eabe0444SAndrey Sidorov * Right neighbour [7] is busy. It can't be coaleasced with [6], so 1873eabe0444SAndrey Sidorov * mark [6] free, increase bb_counters and shrink range to 1874eabe0444SAndrey Sidorov * [0; 5]. 1875eabe0444SAndrey Sidorov * Then shift range to [0; 2], go up and do the same. 1876eabe0444SAndrey Sidorov */ 1877eabe0444SAndrey Sidorov 1878eabe0444SAndrey Sidorov 1879eabe0444SAndrey Sidorov if (first & 1) 1880eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); 1881eabe0444SAndrey Sidorov if (!(last & 1)) 1882eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); 1883eabe0444SAndrey Sidorov if (first > last) 1884eabe0444SAndrey Sidorov break; 1885eabe0444SAndrey Sidorov order++; 1886eabe0444SAndrey Sidorov 1887976620bdSKemeng Shi buddy2 = mb_find_buddy(e4b, order, &max); 1888976620bdSKemeng Shi if (!buddy2) { 1889eabe0444SAndrey Sidorov mb_clear_bits(buddy, first, last - first + 1); 1890eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order - 1] += last - first + 1; 1891eabe0444SAndrey Sidorov break; 1892eabe0444SAndrey Sidorov } 1893eabe0444SAndrey Sidorov first >>= 1; 1894eabe0444SAndrey Sidorov last >>= 1; 1895eabe0444SAndrey Sidorov buddy = buddy2; 1896eabe0444SAndrey Sidorov } 1897eabe0444SAndrey Sidorov } 1898eabe0444SAndrey Sidorov 18997e5a8cddSShen Feng static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1900c9de560dSAlex Tomas int first, int count) 1901c9de560dSAlex Tomas { 1902eabe0444SAndrey Sidorov int left_is_free = 0; 1903eabe0444SAndrey Sidorov int right_is_free = 0; 1904eabe0444SAndrey Sidorov int block; 1905eabe0444SAndrey Sidorov int last = first + count - 1; 1906c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 1907c9de560dSAlex Tomas 1908c99d1e6eSTheodore Ts'o if (WARN_ON(count == 0)) 1909c99d1e6eSTheodore Ts'o return; 1910eabe0444SAndrey Sidorov BUG_ON(last >= (sb->s_blocksize << 3)); 1911bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1912163a203dSDarrick J. Wong /* Don't bother if the block group is corrupt. */ 1913163a203dSDarrick J. Wong if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 1914163a203dSDarrick J. Wong return; 1915163a203dSDarrick J. Wong 1916c9de560dSAlex Tomas mb_check_buddy(e4b); 1917c9de560dSAlex Tomas mb_free_blocks_double(inode, e4b, first, count); 1918c9de560dSAlex Tomas 1919eabe0444SAndrey Sidorov /* access memory sequentially: check left neighbour, 1920eabe0444SAndrey Sidorov * clear range and then check right neighbour 1921eabe0444SAndrey Sidorov */ 1922c9de560dSAlex Tomas if (first != 0) 1923eabe0444SAndrey Sidorov left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); 1924eabe0444SAndrey Sidorov block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); 1925eabe0444SAndrey Sidorov if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) 1926eabe0444SAndrey Sidorov right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); 1927c9de560dSAlex Tomas 1928eabe0444SAndrey Sidorov if (unlikely(block != -1)) { 1929e43bb4e6SNamjae Jeon struct ext4_sb_info *sbi = EXT4_SB(sb); 1930c9de560dSAlex Tomas ext4_fsblk_t blocknr; 19315661bd68SAkinobu Mita 19322331fd4aSBaokun Li /* 19332331fd4aSBaokun Li * Fastcommit replay can free already freed blocks which 19342331fd4aSBaokun Li * corrupts allocation info. Regenerate it. 19352331fd4aSBaokun Li */ 19362331fd4aSBaokun Li if (sbi->s_mount_state & EXT4_FC_REPLAY) { 19372331fd4aSBaokun Li mb_regenerate_buddy(e4b); 19382331fd4aSBaokun Li goto check; 19392331fd4aSBaokun Li } 19402331fd4aSBaokun Li 19415661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 194249598e04SJun Piao blocknr += EXT4_C2B(sbi, block); 1943c5f3a382SBaokun Li ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 1944c5f3a382SBaokun Li EXT4_GROUP_INFO_BBITMAP_CORRUPT); 19455d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 19462331fd4aSBaokun Li inode ? inode->i_ino : 0, blocknr, 19478016e29fSHarshad Shirwadkar "freeing already freed block (bit %u); block bitmap corrupt.", 1948163a203dSDarrick J. Wong block); 19492331fd4aSBaokun Li return; 19508016e29fSHarshad Shirwadkar } 19512331fd4aSBaokun Li 19522331fd4aSBaokun Li this_cpu_inc(discard_pa_seq); 19532331fd4aSBaokun Li e4b->bd_info->bb_free += count; 19542331fd4aSBaokun Li if (first < e4b->bd_info->bb_first_free) 19552331fd4aSBaokun Li e4b->bd_info->bb_first_free = first; 1956c9de560dSAlex Tomas 1957eabe0444SAndrey Sidorov /* let's maintain fragments counter */ 1958eabe0444SAndrey Sidorov if (left_is_free && right_is_free) 1959eabe0444SAndrey Sidorov e4b->bd_info->bb_fragments--; 1960eabe0444SAndrey Sidorov else if (!left_is_free && !right_is_free) 1961eabe0444SAndrey Sidorov e4b->bd_info->bb_fragments++; 1962c9de560dSAlex Tomas 1963eabe0444SAndrey Sidorov /* buddy[0] == bd_bitmap is a special case, so handle 1964eabe0444SAndrey Sidorov * it right away and let mb_buddy_mark_free stay free of 1965eabe0444SAndrey Sidorov * zero order checks. 1966eabe0444SAndrey Sidorov * Check if neighbours are to be coaleasced, 1967eabe0444SAndrey Sidorov * adjust bitmap bb_counters and borders appropriately. 1968eabe0444SAndrey Sidorov */ 1969eabe0444SAndrey Sidorov if (first & 1) { 1970eabe0444SAndrey Sidorov first += !left_is_free; 1971eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; 1972c9de560dSAlex Tomas } 1973eabe0444SAndrey Sidorov if (!(last & 1)) { 1974eabe0444SAndrey Sidorov last -= !right_is_free; 1975eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; 1976c9de560dSAlex Tomas } 1977eabe0444SAndrey Sidorov 1978eabe0444SAndrey Sidorov if (first <= last) 1979eabe0444SAndrey Sidorov mb_buddy_mark_free(e4b, first >> 1, last >> 1); 1980eabe0444SAndrey Sidorov 19818a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, e4b->bd_info); 1982196e402aSHarshad Shirwadkar mb_update_avg_fragment_size(sb, e4b->bd_info); 19832331fd4aSBaokun Li check: 1984c9de560dSAlex Tomas mb_check_buddy(e4b); 1985c9de560dSAlex Tomas } 1986c9de560dSAlex Tomas 198715c006a2SRobin Dong static int mb_find_extent(struct ext4_buddy *e4b, int block, 1988c9de560dSAlex Tomas int needed, struct ext4_free_extent *ex) 1989c9de560dSAlex Tomas { 19902bf5eb2aSGou Hao int max, order, next; 1991c9de560dSAlex Tomas void *buddy; 1992c9de560dSAlex Tomas 1993bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1994c9de560dSAlex Tomas BUG_ON(ex == NULL); 1995c9de560dSAlex Tomas 199615c006a2SRobin Dong buddy = mb_find_buddy(e4b, 0, &max); 1997c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1998c9de560dSAlex Tomas BUG_ON(block >= max); 1999c9de560dSAlex Tomas if (mb_test_bit(block, buddy)) { 2000c9de560dSAlex Tomas ex->fe_len = 0; 2001c9de560dSAlex Tomas ex->fe_start = 0; 2002c9de560dSAlex Tomas ex->fe_group = 0; 2003c9de560dSAlex Tomas return 0; 2004c9de560dSAlex Tomas } 2005c9de560dSAlex Tomas 2006c9de560dSAlex Tomas /* find actual order */ 2007c9de560dSAlex Tomas order = mb_find_order_for_block(e4b, block); 2008c9de560dSAlex Tomas 20092bf5eb2aSGou Hao ex->fe_len = (1 << order) - (block & ((1 << order) - 1)); 20102bf5eb2aSGou Hao ex->fe_start = block; 2011c9de560dSAlex Tomas ex->fe_group = e4b->bd_group; 2012c9de560dSAlex Tomas 20132bf5eb2aSGou Hao block = block >> order; 2014c9de560dSAlex Tomas 2015c9de560dSAlex Tomas while (needed > ex->fe_len && 2016d8ec0c39SAlan Cox mb_find_buddy(e4b, order, &max)) { 2017c9de560dSAlex Tomas 2018c9de560dSAlex Tomas if (block + 1 >= max) 2019c9de560dSAlex Tomas break; 2020c9de560dSAlex Tomas 2021c9de560dSAlex Tomas next = (block + 1) * (1 << order); 2022c5e8f3f3STheodore Ts'o if (mb_test_bit(next, e4b->bd_bitmap)) 2023c9de560dSAlex Tomas break; 2024c9de560dSAlex Tomas 2025b051d8dcSRobin Dong order = mb_find_order_for_block(e4b, next); 2026c9de560dSAlex Tomas 2027c9de560dSAlex Tomas block = next >> order; 2028c9de560dSAlex Tomas ex->fe_len += 1 << order; 2029c9de560dSAlex Tomas } 2030c9de560dSAlex Tomas 203131562b95SJan Kara if (ex->fe_start + ex->fe_len > EXT4_CLUSTERS_PER_GROUP(e4b->bd_sb)) { 203243c73221STheodore Ts'o /* Should never happen! (but apparently sometimes does?!?) */ 203343c73221STheodore Ts'o WARN_ON(1); 2034cd84bbbaSStephen Brennan ext4_grp_locked_error(e4b->bd_sb, e4b->bd_group, 0, 0, 2035cd84bbbaSStephen Brennan "corruption or bug in mb_find_extent " 203643c73221STheodore Ts'o "block=%d, order=%d needed=%d ex=%u/%d/%d@%u", 203743c73221STheodore Ts'o block, order, needed, ex->fe_group, ex->fe_start, 203843c73221STheodore Ts'o ex->fe_len, ex->fe_logical); 203943c73221STheodore Ts'o ex->fe_len = 0; 204043c73221STheodore Ts'o ex->fe_start = 0; 204143c73221STheodore Ts'o ex->fe_group = 0; 204243c73221STheodore Ts'o } 2043c9de560dSAlex Tomas return ex->fe_len; 2044c9de560dSAlex Tomas } 2045c9de560dSAlex Tomas 2046c9de560dSAlex Tomas static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 2047c9de560dSAlex Tomas { 2048c9de560dSAlex Tomas int ord; 2049c9de560dSAlex Tomas int mlen = 0; 2050c9de560dSAlex Tomas int max = 0; 2051c9de560dSAlex Tomas int start = ex->fe_start; 2052c9de560dSAlex Tomas int len = ex->fe_len; 2053c9de560dSAlex Tomas unsigned ret = 0; 2054c9de560dSAlex Tomas int len0 = len; 2055c9de560dSAlex Tomas void *buddy; 2056d1a3924eSKemeng Shi int ord_start, ord_end; 2057c9de560dSAlex Tomas 2058c9de560dSAlex Tomas BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 2059c9de560dSAlex Tomas BUG_ON(e4b->bd_group != ex->fe_group); 2060bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 2061c9de560dSAlex Tomas mb_check_buddy(e4b); 2062c9de560dSAlex Tomas mb_mark_used_double(e4b, start, len); 2063c9de560dSAlex Tomas 206407b5b8e1SRitesh Harjani this_cpu_inc(discard_pa_seq); 2065c9de560dSAlex Tomas e4b->bd_info->bb_free -= len; 2066c9de560dSAlex Tomas if (e4b->bd_info->bb_first_free == start) 2067c9de560dSAlex Tomas e4b->bd_info->bb_first_free += len; 2068c9de560dSAlex Tomas 2069c9de560dSAlex Tomas /* let's maintain fragments counter */ 2070c9de560dSAlex Tomas if (start != 0) 2071c5e8f3f3STheodore Ts'o mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); 2072c9de560dSAlex Tomas if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 2073c5e8f3f3STheodore Ts'o max = !mb_test_bit(start + len, e4b->bd_bitmap); 2074c9de560dSAlex Tomas if (mlen && max) 2075c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 2076c9de560dSAlex Tomas else if (!mlen && !max) 2077c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 2078c9de560dSAlex Tomas 2079c9de560dSAlex Tomas /* let's maintain buddy itself */ 2080c9de560dSAlex Tomas while (len) { 2081c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, start); 2082c9de560dSAlex Tomas 2083c9de560dSAlex Tomas if (((start >> ord) << ord) == start && len >= (1 << ord)) { 2084c9de560dSAlex Tomas /* the whole chunk may be allocated at once! */ 2085c9de560dSAlex Tomas mlen = 1 << ord; 2086c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 2087c9de560dSAlex Tomas BUG_ON((start >> ord) >= max); 2088c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 2089c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 2090c9de560dSAlex Tomas start += mlen; 2091c9de560dSAlex Tomas len -= mlen; 2092c9de560dSAlex Tomas BUG_ON(len < 0); 2093c9de560dSAlex Tomas continue; 2094c9de560dSAlex Tomas } 2095c9de560dSAlex Tomas 2096c9de560dSAlex Tomas /* store for history */ 2097c9de560dSAlex Tomas if (ret == 0) 2098c9de560dSAlex Tomas ret = len | (ord << 16); 2099c9de560dSAlex Tomas 2100c9de560dSAlex Tomas BUG_ON(ord <= 0); 2101c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 2102c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 2103c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 2104c9de560dSAlex Tomas 2105d1a3924eSKemeng Shi ord_start = (start >> ord) << ord; 2106d1a3924eSKemeng Shi ord_end = ord_start + (1 << ord); 2107d1a3924eSKemeng Shi /* first chunk */ 2108d1a3924eSKemeng Shi if (start > ord_start) 2109d1a3924eSKemeng Shi ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy, 2110d1a3924eSKemeng Shi ord_start, start - ord_start, 2111d1a3924eSKemeng Shi e4b->bd_info); 2112d1a3924eSKemeng Shi 2113d1a3924eSKemeng Shi /* last chunk */ 2114d1a3924eSKemeng Shi if (start + len < ord_end) { 2115d1a3924eSKemeng Shi ext4_mb_mark_free_simple(e4b->bd_sb, e4b->bd_buddy, 2116d1a3924eSKemeng Shi start + len, 2117d1a3924eSKemeng Shi ord_end - (start + len), 2118d1a3924eSKemeng Shi e4b->bd_info); 2119d1a3924eSKemeng Shi break; 2120d1a3924eSKemeng Shi } 2121d1a3924eSKemeng Shi len = start + len - ord_end; 2122d1a3924eSKemeng Shi start = ord_end; 2123c9de560dSAlex Tomas } 21248a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); 2125c9de560dSAlex Tomas 2126196e402aSHarshad Shirwadkar mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info); 2127123e3016SRitesh Harjani mb_set_bits(e4b->bd_bitmap, ex->fe_start, len0); 2128c9de560dSAlex Tomas mb_check_buddy(e4b); 2129c9de560dSAlex Tomas 2130c9de560dSAlex Tomas return ret; 2131c9de560dSAlex Tomas } 2132c9de560dSAlex Tomas 2133c9de560dSAlex Tomas /* 2134c9de560dSAlex Tomas * Must be called under group lock! 2135c9de560dSAlex Tomas */ 2136c9de560dSAlex Tomas static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 2137c9de560dSAlex Tomas struct ext4_buddy *e4b) 2138c9de560dSAlex Tomas { 2139c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2140c9de560dSAlex Tomas int ret; 2141c9de560dSAlex Tomas 2142c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 2143c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2144c9de560dSAlex Tomas 2145c9de560dSAlex Tomas ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 2146c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 2147c9de560dSAlex Tomas ret = mb_mark_used(e4b, &ac->ac_b_ex); 2148c9de560dSAlex Tomas 2149c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 2150c9de560dSAlex Tomas * allocated blocks for history */ 2151c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 2152c9de560dSAlex Tomas 2153c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 2154c9de560dSAlex Tomas ac->ac_tail = ret & 0xffff; 2155c9de560dSAlex Tomas ac->ac_buddy = ret >> 16; 2156c9de560dSAlex Tomas 2157c3a326a6SAneesh Kumar K.V /* 2158c3a326a6SAneesh Kumar K.V * take the page reference. We want the page to be pinned 2159c3a326a6SAneesh Kumar K.V * so that we don't get a ext4_mb_init_cache_call for this 2160c3a326a6SAneesh Kumar K.V * group until we update the bitmap. That would mean we 2161c3a326a6SAneesh Kumar K.V * double allocate blocks. The reference is dropped 2162c3a326a6SAneesh Kumar K.V * in ext4_mb_release_context 2163c3a326a6SAneesh Kumar K.V */ 2164ccedf35bSMatthew Wilcox (Oracle) ac->ac_bitmap_folio = e4b->bd_bitmap_folio; 2165ccedf35bSMatthew Wilcox (Oracle) folio_get(ac->ac_bitmap_folio); 2166c84f1510SMatthew Wilcox (Oracle) ac->ac_buddy_folio = e4b->bd_buddy_folio; 2167c84f1510SMatthew Wilcox (Oracle) folio_get(ac->ac_buddy_folio); 2168c9de560dSAlex Tomas /* store last allocated for subsequent stream allocation */ 21694ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2170c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 2171c9de560dSAlex Tomas sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 2172c9de560dSAlex Tomas sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 2173c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 2174c9de560dSAlex Tomas } 217553f86b17SRitesh Harjani /* 217653f86b17SRitesh Harjani * As we've just preallocated more space than 217753f86b17SRitesh Harjani * user requested originally, we store allocated 217853f86b17SRitesh Harjani * space in a special descriptor. 217953f86b17SRitesh Harjani */ 218053f86b17SRitesh Harjani if (ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 218153f86b17SRitesh Harjani ext4_mb_new_preallocation(ac); 218253f86b17SRitesh Harjani 2183c9de560dSAlex Tomas } 2184c9de560dSAlex Tomas 2185c9de560dSAlex Tomas static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 2186c9de560dSAlex Tomas struct ext4_buddy *e4b, 2187c9de560dSAlex Tomas int finish_group) 2188c9de560dSAlex Tomas { 2189c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2190c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 2191c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 2192c9de560dSAlex Tomas 2193032115fcSAneesh Kumar K.V if (ac->ac_status == AC_STATUS_FOUND) 2194032115fcSAneesh Kumar K.V return; 2195c9de560dSAlex Tomas /* 2196c9de560dSAlex Tomas * We don't want to scan for a whole year 2197c9de560dSAlex Tomas */ 2198c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan && 2199c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2200c9de560dSAlex Tomas ac->ac_status = AC_STATUS_BREAK; 2201c9de560dSAlex Tomas return; 2202c9de560dSAlex Tomas } 2203c9de560dSAlex Tomas 2204c9de560dSAlex Tomas /* 2205c9de560dSAlex Tomas * Haven't found good chunk so far, let's continue 2206c9de560dSAlex Tomas */ 2207c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) 2208c9de560dSAlex Tomas return; 2209c9de560dSAlex Tomas 22103582e745SOjaswin Mujoo if (finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 2211c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2212c9de560dSAlex Tomas } 2213c9de560dSAlex Tomas 2214c9de560dSAlex Tomas /* 2215c9de560dSAlex Tomas * The routine checks whether found extent is good enough. If it is, 2216c9de560dSAlex Tomas * then the extent gets marked used and flag is set to the context 2217c9de560dSAlex Tomas * to stop scanning. Otherwise, the extent is compared with the 2218c9de560dSAlex Tomas * previous found extent and if new one is better, then it's stored 2219c9de560dSAlex Tomas * in the context. Later, the best found extent will be used, if 2220c9de560dSAlex Tomas * mballoc can't find good enough extent. 2221c9de560dSAlex Tomas * 22223582e745SOjaswin Mujoo * The algorithm used is roughly as follows: 22233582e745SOjaswin Mujoo * 22243582e745SOjaswin Mujoo * * If free extent found is exactly as big as goal, then 22253582e745SOjaswin Mujoo * stop the scan and use it immediately 22263582e745SOjaswin Mujoo * 22273582e745SOjaswin Mujoo * * If free extent found is smaller than goal, then keep retrying 22283582e745SOjaswin Mujoo * upto a max of sbi->s_mb_max_to_scan times (default 200). After 22293582e745SOjaswin Mujoo * that stop scanning and use whatever we have. 22303582e745SOjaswin Mujoo * 22313582e745SOjaswin Mujoo * * If free extent found is bigger than goal, then keep retrying 22323582e745SOjaswin Mujoo * upto a max of sbi->s_mb_min_to_scan times (default 10) before 22333582e745SOjaswin Mujoo * stopping the scan and using the extent. 22343582e745SOjaswin Mujoo * 22353582e745SOjaswin Mujoo * 2236c9de560dSAlex Tomas * FIXME: real allocation policy is to be designed yet! 2237c9de560dSAlex Tomas */ 2238c9de560dSAlex Tomas static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 2239c9de560dSAlex Tomas struct ext4_free_extent *ex, 2240c9de560dSAlex Tomas struct ext4_buddy *e4b) 2241c9de560dSAlex Tomas { 2242c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 2243c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 2244c9de560dSAlex Tomas 2245c9de560dSAlex Tomas BUG_ON(ex->fe_len <= 0); 22467137d7a4STheodore Ts'o BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 22477137d7a4STheodore Ts'o BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 2248c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 2249c9de560dSAlex Tomas 2250c9de560dSAlex Tomas ac->ac_found++; 2251fdd9a009SOjaswin Mujoo ac->ac_cX_found[ac->ac_criteria]++; 2252c9de560dSAlex Tomas 2253c9de560dSAlex Tomas /* 2254c9de560dSAlex Tomas * The special case - take what you catch first 2255c9de560dSAlex Tomas */ 2256c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2257c9de560dSAlex Tomas *bex = *ex; 2258c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2259c9de560dSAlex Tomas return; 2260c9de560dSAlex Tomas } 2261c9de560dSAlex Tomas 2262c9de560dSAlex Tomas /* 2263c9de560dSAlex Tomas * Let's check whether the chuck is good enough 2264c9de560dSAlex Tomas */ 2265c9de560dSAlex Tomas if (ex->fe_len == gex->fe_len) { 2266c9de560dSAlex Tomas *bex = *ex; 2267c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2268c9de560dSAlex Tomas return; 2269c9de560dSAlex Tomas } 2270c9de560dSAlex Tomas 2271c9de560dSAlex Tomas /* 2272c9de560dSAlex Tomas * If this is first found extent, just store it in the context 2273c9de560dSAlex Tomas */ 2274c9de560dSAlex Tomas if (bex->fe_len == 0) { 2275c9de560dSAlex Tomas *bex = *ex; 2276c9de560dSAlex Tomas return; 2277c9de560dSAlex Tomas } 2278c9de560dSAlex Tomas 2279c9de560dSAlex Tomas /* 2280c9de560dSAlex Tomas * If new found extent is better, store it in the context 2281c9de560dSAlex Tomas */ 2282c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) { 2283c9de560dSAlex Tomas /* if the request isn't satisfied, any found extent 2284c9de560dSAlex Tomas * larger than previous best one is better */ 2285c9de560dSAlex Tomas if (ex->fe_len > bex->fe_len) 2286c9de560dSAlex Tomas *bex = *ex; 2287c9de560dSAlex Tomas } else if (ex->fe_len > gex->fe_len) { 2288c9de560dSAlex Tomas /* if the request is satisfied, then we try to find 2289c9de560dSAlex Tomas * an extent that still satisfy the request, but is 2290c9de560dSAlex Tomas * smaller than previous one */ 2291c9de560dSAlex Tomas if (ex->fe_len < bex->fe_len) 2292c9de560dSAlex Tomas *bex = *ex; 2293c9de560dSAlex Tomas } 2294c9de560dSAlex Tomas 2295c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 0); 2296c9de560dSAlex Tomas } 2297c9de560dSAlex Tomas 2298089ceeccSEric Sandeen static noinline_for_stack 229985b67ffbSKemeng Shi void ext4_mb_try_best_found(struct ext4_allocation_context *ac, 2300c9de560dSAlex Tomas struct ext4_buddy *e4b) 2301c9de560dSAlex Tomas { 2302c9de560dSAlex Tomas struct ext4_free_extent ex = ac->ac_b_ex; 2303c9de560dSAlex Tomas ext4_group_t group = ex.fe_group; 2304c9de560dSAlex Tomas int max; 2305c9de560dSAlex Tomas int err; 2306c9de560dSAlex Tomas 2307c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 2308c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 2309c9de560dSAlex Tomas if (err) 231085b67ffbSKemeng Shi return; 2311c9de560dSAlex Tomas 2312c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 23134530b366SBaokun Li if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 23144530b366SBaokun Li goto out; 23154530b366SBaokun Li 231615c006a2SRobin Dong max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); 2317c9de560dSAlex Tomas 2318c9de560dSAlex Tomas if (max > 0) { 2319c9de560dSAlex Tomas ac->ac_b_ex = ex; 2320c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2321c9de560dSAlex Tomas } 2322c9de560dSAlex Tomas 23234530b366SBaokun Li out: 2324c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 2325e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 2326c9de560dSAlex Tomas } 2327c9de560dSAlex Tomas 2328089ceeccSEric Sandeen static noinline_for_stack 2329089ceeccSEric Sandeen int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 2330c9de560dSAlex Tomas struct ext4_buddy *e4b) 2331c9de560dSAlex Tomas { 2332c9de560dSAlex Tomas ext4_group_t group = ac->ac_g_ex.fe_group; 2333c9de560dSAlex Tomas int max; 2334c9de560dSAlex Tomas int err; 2335c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2336838cd0cfSYongqiang Yang struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2337c9de560dSAlex Tomas struct ext4_free_extent ex; 2338c9de560dSAlex Tomas 23395354b2afSTheodore Ts'o if (!grp) 23405354b2afSTheodore Ts'o return -EFSCORRUPTED; 234101e4ca29SKemeng Shi if (!(ac->ac_flags & (EXT4_MB_HINT_TRY_GOAL | EXT4_MB_HINT_GOAL_ONLY))) 2342c9de560dSAlex Tomas return 0; 2343838cd0cfSYongqiang Yang if (grp->bb_free == 0) 2344838cd0cfSYongqiang Yang return 0; 2345c9de560dSAlex Tomas 2346c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 2347c9de560dSAlex Tomas if (err) 2348c9de560dSAlex Tomas return err; 2349c9de560dSAlex Tomas 2350c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 235183269837SBaokun Li if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 235283269837SBaokun Li goto out; 235383269837SBaokun Li 235415c006a2SRobin Dong max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, 2355c9de560dSAlex Tomas ac->ac_g_ex.fe_len, &ex); 2356ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADFA11; /* debug value */ 2357c9de560dSAlex Tomas 2358c3defd99SKemeng Shi if (max >= ac->ac_g_ex.fe_len && 2359c3defd99SKemeng Shi ac->ac_g_ex.fe_len == EXT4_B2C(sbi, sbi->s_stripe)) { 2360c9de560dSAlex Tomas ext4_fsblk_t start; 2361c9de560dSAlex Tomas 236299c515e3SKemeng Shi start = ext4_grp_offs_to_block(ac->ac_sb, &ex); 2363c9de560dSAlex Tomas /* use do_div to get remainder (would be 64-bit modulo) */ 2364c9de560dSAlex Tomas if (do_div(start, sbi->s_stripe) == 0) { 2365c9de560dSAlex Tomas ac->ac_found++; 2366c9de560dSAlex Tomas ac->ac_b_ex = ex; 2367c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2368c9de560dSAlex Tomas } 2369c9de560dSAlex Tomas } else if (max >= ac->ac_g_ex.fe_len) { 2370c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 2371c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 2372c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 2373c9de560dSAlex Tomas ac->ac_found++; 2374c9de560dSAlex Tomas ac->ac_b_ex = ex; 2375c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2376c9de560dSAlex Tomas } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 2377c9de560dSAlex Tomas /* Sometimes, caller may want to merge even small 2378c9de560dSAlex Tomas * number of blocks to an existing extent */ 2379c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 2380c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 2381c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 2382c9de560dSAlex Tomas ac->ac_found++; 2383c9de560dSAlex Tomas ac->ac_b_ex = ex; 2384c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2385c9de560dSAlex Tomas } 238683269837SBaokun Li out: 2387c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 2388e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 2389c9de560dSAlex Tomas 2390c9de560dSAlex Tomas return 0; 2391c9de560dSAlex Tomas } 2392c9de560dSAlex Tomas 2393c9de560dSAlex Tomas /* 2394c9de560dSAlex Tomas * The routine scans buddy structures (not bitmap!) from given order 2395c9de560dSAlex Tomas * to max order and tries to find big enough chunk to satisfy the req 2396c9de560dSAlex Tomas */ 2397089ceeccSEric Sandeen static noinline_for_stack 2398089ceeccSEric Sandeen void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 2399c9de560dSAlex Tomas struct ext4_buddy *e4b) 2400c9de560dSAlex Tomas { 2401c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2402c9de560dSAlex Tomas struct ext4_group_info *grp = e4b->bd_info; 2403c9de560dSAlex Tomas void *buddy; 2404c9de560dSAlex Tomas int i; 2405c9de560dSAlex Tomas int k; 2406c9de560dSAlex Tomas int max; 2407c9de560dSAlex Tomas 2408c9de560dSAlex Tomas BUG_ON(ac->ac_2order <= 0); 24094b68f6dfSHarshad Shirwadkar for (i = ac->ac_2order; i < MB_NUM_ORDERS(sb); i++) { 2410c9de560dSAlex Tomas if (grp->bb_counters[i] == 0) 2411c9de560dSAlex Tomas continue; 2412c9de560dSAlex Tomas 2413c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, i, &max); 241419b8b035STheodore Ts'o if (WARN_RATELIMIT(buddy == NULL, 241519b8b035STheodore Ts'o "ext4: mb_simple_scan_group: mb_find_buddy failed, (%d)\n", i)) 241619b8b035STheodore Ts'o continue; 2417c9de560dSAlex Tomas 2418ffad0a44SAneesh Kumar K.V k = mb_find_next_zero_bit(buddy, max, 0); 2419eb576086SDmitry Monakhov if (k >= max) { 2420eb576086SDmitry Monakhov ext4_mark_group_bitmap_corrupted(ac->ac_sb, 2421eb576086SDmitry Monakhov e4b->bd_group, 2422eb576086SDmitry Monakhov EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2423c5f3a382SBaokun Li ext4_grp_locked_error(ac->ac_sb, e4b->bd_group, 0, 0, 2424c5f3a382SBaokun Li "%d free clusters of order %d. But found 0", 2425c5f3a382SBaokun Li grp->bb_counters[i], i); 2426eb576086SDmitry Monakhov break; 2427eb576086SDmitry Monakhov } 2428c9de560dSAlex Tomas ac->ac_found++; 2429fdd9a009SOjaswin Mujoo ac->ac_cX_found[ac->ac_criteria]++; 2430c9de560dSAlex Tomas 2431c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 1 << i; 2432c9de560dSAlex Tomas ac->ac_b_ex.fe_start = k << i; 2433c9de560dSAlex Tomas ac->ac_b_ex.fe_group = e4b->bd_group; 2434c9de560dSAlex Tomas 2435c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2436c9de560dSAlex Tomas 243753f86b17SRitesh Harjani BUG_ON(ac->ac_f_ex.fe_len != ac->ac_g_ex.fe_len); 2438c9de560dSAlex Tomas 2439c9de560dSAlex Tomas if (EXT4_SB(sb)->s_mb_stats) 2440c9de560dSAlex Tomas atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 2441c9de560dSAlex Tomas 2442c9de560dSAlex Tomas break; 2443c9de560dSAlex Tomas } 2444c9de560dSAlex Tomas } 2445c9de560dSAlex Tomas 2446c9de560dSAlex Tomas /* 2447c9de560dSAlex Tomas * The routine scans the group and measures all found extents. 2448c9de560dSAlex Tomas * In order to optimize scanning, caller must pass number of 2449c9de560dSAlex Tomas * free blocks in the group, so the routine can know upper limit. 2450c9de560dSAlex Tomas */ 2451089ceeccSEric Sandeen static noinline_for_stack 2452089ceeccSEric Sandeen void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 2453c9de560dSAlex Tomas struct ext4_buddy *e4b) 2454c9de560dSAlex Tomas { 2455c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2456c5e8f3f3STheodore Ts'o void *bitmap = e4b->bd_bitmap; 2457c9de560dSAlex Tomas struct ext4_free_extent ex; 24581b420011SOjaswin Mujoo int i, j, freelen; 2459c9de560dSAlex Tomas int free; 2460c9de560dSAlex Tomas 2461c9de560dSAlex Tomas free = e4b->bd_info->bb_free; 2462907ea529STheodore Ts'o if (WARN_ON(free <= 0)) 2463907ea529STheodore Ts'o return; 2464c9de560dSAlex Tomas 2465c9de560dSAlex Tomas i = e4b->bd_info->bb_first_free; 2466c9de560dSAlex Tomas 2467c9de560dSAlex Tomas while (free && ac->ac_status == AC_STATUS_CONTINUE) { 2468ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, 24697137d7a4STheodore Ts'o EXT4_CLUSTERS_PER_GROUP(sb), i); 24707137d7a4STheodore Ts'o if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { 247126346ff6SAneesh Kumar K.V /* 2472e56eb659SAneesh Kumar K.V * IF we have corrupt bitmap, we won't find any 247326346ff6SAneesh Kumar K.V * free blocks even though group info says we 2474b483bb77SRandy Dunlap * have free blocks 247526346ff6SAneesh Kumar K.V */ 2476c5f3a382SBaokun Li ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 2477c5f3a382SBaokun Li EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2478e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 247953accfa9STheodore Ts'o "%d free clusters as per " 2480fde4d95aSTheodore Ts'o "group info. But bitmap says 0", 248126346ff6SAneesh Kumar K.V free); 2482c9de560dSAlex Tomas break; 2483c9de560dSAlex Tomas } 2484c9de560dSAlex Tomas 2485304749c0SOjaswin Mujoo if (!ext4_mb_cr_expensive(ac->ac_criteria)) { 24861b420011SOjaswin Mujoo /* 2487f52f3d2bSOjaswin Mujoo * In CR_GOAL_LEN_FAST and CR_BEST_AVAIL_LEN, we are 2488f52f3d2bSOjaswin Mujoo * sure that this group will have a large enough 2489f52f3d2bSOjaswin Mujoo * continuous free extent, so skip over the smaller free 2490f52f3d2bSOjaswin Mujoo * extents 24911b420011SOjaswin Mujoo */ 24921b420011SOjaswin Mujoo j = mb_find_next_bit(bitmap, 24931b420011SOjaswin Mujoo EXT4_CLUSTERS_PER_GROUP(sb), i); 24941b420011SOjaswin Mujoo freelen = j - i; 24951b420011SOjaswin Mujoo 24961b420011SOjaswin Mujoo if (freelen < ac->ac_g_ex.fe_len) { 24971b420011SOjaswin Mujoo i = j; 24981b420011SOjaswin Mujoo free -= freelen; 24991b420011SOjaswin Mujoo continue; 25001b420011SOjaswin Mujoo } 25011b420011SOjaswin Mujoo } 25021b420011SOjaswin Mujoo 250315c006a2SRobin Dong mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); 2504907ea529STheodore Ts'o if (WARN_ON(ex.fe_len <= 0)) 2505907ea529STheodore Ts'o break; 250626346ff6SAneesh Kumar K.V if (free < ex.fe_len) { 2507c5f3a382SBaokun Li ext4_mark_group_bitmap_corrupted(sb, e4b->bd_group, 2508c5f3a382SBaokun Li EXT4_GROUP_INFO_BBITMAP_CORRUPT); 2509e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 251053accfa9STheodore Ts'o "%d free clusters as per " 2511fde4d95aSTheodore Ts'o "group info. But got %d blocks", 251226346ff6SAneesh Kumar K.V free, ex.fe_len); 2513e56eb659SAneesh Kumar K.V /* 2514e56eb659SAneesh Kumar K.V * The number of free blocks differs. This mostly 2515e56eb659SAneesh Kumar K.V * indicate that the bitmap is corrupt. So exit 2516e56eb659SAneesh Kumar K.V * without claiming the space. 2517e56eb659SAneesh Kumar K.V */ 2518e56eb659SAneesh Kumar K.V break; 251926346ff6SAneesh Kumar K.V } 2520ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADC0DE; /* debug value */ 2521c9de560dSAlex Tomas ext4_mb_measure_extent(ac, &ex, e4b); 2522c9de560dSAlex Tomas 2523c9de560dSAlex Tomas i += ex.fe_len; 2524c9de560dSAlex Tomas free -= ex.fe_len; 2525c9de560dSAlex Tomas } 2526c9de560dSAlex Tomas 2527c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 1); 2528c9de560dSAlex Tomas } 2529c9de560dSAlex Tomas 2530c9de560dSAlex Tomas /* 2531c9de560dSAlex Tomas * This is a special case for storages like raid5 2532506bf2d8SEric Sandeen * we try to find stripe-aligned chunks for stripe-size-multiple requests 2533c9de560dSAlex Tomas */ 2534089ceeccSEric Sandeen static noinline_for_stack 2535089ceeccSEric Sandeen void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 2536c9de560dSAlex Tomas struct ext4_buddy *e4b) 2537c9de560dSAlex Tomas { 2538c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2539c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2540c5e8f3f3STheodore Ts'o void *bitmap = e4b->bd_bitmap; 2541c9de560dSAlex Tomas struct ext4_free_extent ex; 2542c9de560dSAlex Tomas ext4_fsblk_t first_group_block; 2543c9de560dSAlex Tomas ext4_fsblk_t a; 2544c3defd99SKemeng Shi ext4_grpblk_t i, stripe; 2545c9de560dSAlex Tomas int max; 2546c9de560dSAlex Tomas 2547c9de560dSAlex Tomas BUG_ON(sbi->s_stripe == 0); 2548c9de560dSAlex Tomas 2549c9de560dSAlex Tomas /* find first stripe-aligned block in group */ 25505661bd68SAkinobu Mita first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); 25515661bd68SAkinobu Mita 2552c9de560dSAlex Tomas a = first_group_block + sbi->s_stripe - 1; 2553c9de560dSAlex Tomas do_div(a, sbi->s_stripe); 2554c9de560dSAlex Tomas i = (a * sbi->s_stripe) - first_group_block; 2555c9de560dSAlex Tomas 2556c3defd99SKemeng Shi stripe = EXT4_B2C(sbi, sbi->s_stripe); 2557c3defd99SKemeng Shi i = EXT4_B2C(sbi, i); 25587137d7a4STheodore Ts'o while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { 2559c9de560dSAlex Tomas if (!mb_test_bit(i, bitmap)) { 2560c3defd99SKemeng Shi max = mb_find_extent(e4b, i, stripe, &ex); 2561c3defd99SKemeng Shi if (max >= stripe) { 2562c9de560dSAlex Tomas ac->ac_found++; 2563fdd9a009SOjaswin Mujoo ac->ac_cX_found[ac->ac_criteria]++; 2564ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADF00D; /* debug value */ 2565c9de560dSAlex Tomas ac->ac_b_ex = ex; 2566c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2567c9de560dSAlex Tomas break; 2568c9de560dSAlex Tomas } 2569c9de560dSAlex Tomas } 2570c3defd99SKemeng Shi i += stripe; 2571c9de560dSAlex Tomas } 2572c9de560dSAlex Tomas } 2573c9de560dSAlex Tomas 257442ac1848SLukas Czerner /* 25758ef123feSRitesh Harjani * This is also called BEFORE we load the buddy bitmap. 257642ac1848SLukas Czerner * Returns either 1 or 0 indicating that the group is either suitable 25778ef123feSRitesh Harjani * for the allocation or not. 257842ac1848SLukas Czerner */ 25798ef123feSRitesh Harjani static bool ext4_mb_good_group(struct ext4_allocation_context *ac, 25804eb7a4a1SOjaswin Mujoo ext4_group_t group, enum criteria cr) 2581c9de560dSAlex Tomas { 25828ef123feSRitesh Harjani ext4_grpblk_t free, fragments; 2583a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 2584c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2585c9de560dSAlex Tomas 2586f52f3d2bSOjaswin Mujoo BUG_ON(cr < CR_POWER2_ALIGNED || cr >= EXT4_MB_NUM_CRS); 25878a57d9d6SCurt Wohlgemuth 2588a9ce5993SKemeng Shi if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 25898ef123feSRitesh Harjani return false; 259001fc48e8STheodore Ts'o 2591dddcd2f9Sbrookxu free = grp->bb_free; 2592dddcd2f9Sbrookxu if (free == 0) 25938ef123feSRitesh Harjani return false; 2594c9de560dSAlex Tomas 2595c9de560dSAlex Tomas fragments = grp->bb_fragments; 2596c9de560dSAlex Tomas if (fragments == 0) 25978ef123feSRitesh Harjani return false; 2598c9de560dSAlex Tomas 2599c9de560dSAlex Tomas switch (cr) { 2600f52f3d2bSOjaswin Mujoo case CR_POWER2_ALIGNED: 2601c9de560dSAlex Tomas BUG_ON(ac->ac_2order == 0); 2602c9de560dSAlex Tomas 2603a4912123STheodore Ts'o /* Avoid using the first bg of a flexgroup for data files */ 2604a4912123STheodore Ts'o if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 2605a4912123STheodore Ts'o (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 2606a4912123STheodore Ts'o ((group % flex_size) == 0)) 26078ef123feSRitesh Harjani return false; 2608a4912123STheodore Ts'o 2609dddcd2f9Sbrookxu if (free < ac->ac_g_ex.fe_len) 2610dddcd2f9Sbrookxu return false; 2611dddcd2f9Sbrookxu 26124b68f6dfSHarshad Shirwadkar if (ac->ac_2order >= MB_NUM_ORDERS(ac->ac_sb)) 26138ef123feSRitesh Harjani return true; 261440ae3487STheodore Ts'o 261540ae3487STheodore Ts'o if (grp->bb_largest_free_order < ac->ac_2order) 26168ef123feSRitesh Harjani return false; 261740ae3487STheodore Ts'o 26188ef123feSRitesh Harjani return true; 2619f52f3d2bSOjaswin Mujoo case CR_GOAL_LEN_FAST: 2620f52f3d2bSOjaswin Mujoo case CR_BEST_AVAIL_LEN: 2621c9de560dSAlex Tomas if ((free / fragments) >= ac->ac_g_ex.fe_len) 26228ef123feSRitesh Harjani return true; 2623c9de560dSAlex Tomas break; 2624f52f3d2bSOjaswin Mujoo case CR_GOAL_LEN_SLOW: 2625c9de560dSAlex Tomas if (free >= ac->ac_g_ex.fe_len) 26268ef123feSRitesh Harjani return true; 2627c9de560dSAlex Tomas break; 2628f52f3d2bSOjaswin Mujoo case CR_ANY_FREE: 26298ef123feSRitesh Harjani return true; 2630c9de560dSAlex Tomas default: 2631c9de560dSAlex Tomas BUG(); 2632c9de560dSAlex Tomas } 2633c9de560dSAlex Tomas 26348ef123feSRitesh Harjani return false; 26358ef123feSRitesh Harjani } 26368ef123feSRitesh Harjani 26378ef123feSRitesh Harjani /* 26388ef123feSRitesh Harjani * This could return negative error code if something goes wrong 26398ef123feSRitesh Harjani * during ext4_mb_init_group(). This should not be called with 26408ef123feSRitesh Harjani * ext4_lock_group() held. 2641a5fda113STheodore Ts'o * 2642a5fda113STheodore Ts'o * Note: because we are conditionally operating with the group lock in 2643a5fda113STheodore Ts'o * the EXT4_MB_STRICT_CHECK case, we need to fake out sparse in this 2644a5fda113STheodore Ts'o * function using __acquire and __release. This means we need to be 2645a5fda113STheodore Ts'o * super careful before messing with the error path handling via "goto 2646a5fda113STheodore Ts'o * out"! 26478ef123feSRitesh Harjani */ 26488ef123feSRitesh Harjani static int ext4_mb_good_group_nolock(struct ext4_allocation_context *ac, 26494eb7a4a1SOjaswin Mujoo ext4_group_t group, enum criteria cr) 26508ef123feSRitesh Harjani { 26518ef123feSRitesh Harjani struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 265299377830SRitesh Harjani struct super_block *sb = ac->ac_sb; 2653c1d2c7d4SAlex Zhuravlev struct ext4_sb_info *sbi = EXT4_SB(sb); 265499377830SRitesh Harjani bool should_lock = ac->ac_flags & EXT4_MB_STRICT_CHECK; 26558ef123feSRitesh Harjani ext4_grpblk_t free; 26568ef123feSRitesh Harjani int ret = 0; 26578ef123feSRitesh Harjani 26585354b2afSTheodore Ts'o if (!grp) 26595354b2afSTheodore Ts'o return -EFSCORRUPTED; 2660a6c75eafSHarshad Shirwadkar if (sbi->s_mb_stats) 2661a6c75eafSHarshad Shirwadkar atomic64_inc(&sbi->s_bal_cX_groups_considered[ac->ac_criteria]); 2662a5fda113STheodore Ts'o if (should_lock) { 266399377830SRitesh Harjani ext4_lock_group(sb, group); 2664a5fda113STheodore Ts'o __release(ext4_group_lock_ptr(sb, group)); 2665a5fda113STheodore Ts'o } 26668ef123feSRitesh Harjani free = grp->bb_free; 26678ef123feSRitesh Harjani if (free == 0) 26688ef123feSRitesh Harjani goto out; 2669304749c0SOjaswin Mujoo /* 2670304749c0SOjaswin Mujoo * In all criterias except CR_ANY_FREE we try to avoid groups that 2671304749c0SOjaswin Mujoo * can't possibly satisfy the full goal request due to insufficient 2672304749c0SOjaswin Mujoo * free blocks. 2673304749c0SOjaswin Mujoo */ 2674304749c0SOjaswin Mujoo if (cr < CR_ANY_FREE && free < ac->ac_g_ex.fe_len) 26758ef123feSRitesh Harjani goto out; 26768ef123feSRitesh Harjani if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 26778ef123feSRitesh Harjani goto out; 2678a5fda113STheodore Ts'o if (should_lock) { 2679a5fda113STheodore Ts'o __acquire(ext4_group_lock_ptr(sb, group)); 268099377830SRitesh Harjani ext4_unlock_group(sb, group); 2681a5fda113STheodore Ts'o } 26828ef123feSRitesh Harjani 26838ef123feSRitesh Harjani /* We only do this if the grp has never been initialized */ 26848ef123feSRitesh Harjani if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 2685c1d2c7d4SAlex Zhuravlev struct ext4_group_desc *gdp = 2686c1d2c7d4SAlex Zhuravlev ext4_get_group_desc(sb, group, NULL); 2687c1d2c7d4SAlex Zhuravlev int ret; 2688c1d2c7d4SAlex Zhuravlev 2689f52f3d2bSOjaswin Mujoo /* 26902caffb6aSKemeng Shi * CR_POWER2_ALIGNED/CR_GOAL_LEN_FAST is a very optimistic 2691f52f3d2bSOjaswin Mujoo * search to find large good chunks almost for free. If buddy 2692f52f3d2bSOjaswin Mujoo * data is not ready, then this optimization makes no sense. But 2693f52f3d2bSOjaswin Mujoo * we never skip the first block group in a flex_bg, since this 2694f52f3d2bSOjaswin Mujoo * gets used for metadata block allocation, and we want to make 2695f52f3d2bSOjaswin Mujoo * sure we locate metadata blocks in the first block group in 2696f52f3d2bSOjaswin Mujoo * the flex_bg if possible. 2697c1d2c7d4SAlex Zhuravlev */ 2698304749c0SOjaswin Mujoo if (!ext4_mb_cr_expensive(cr) && 2699c1d2c7d4SAlex Zhuravlev (!sbi->s_log_groups_per_flex || 2700c1d2c7d4SAlex Zhuravlev ((group & ((1 << sbi->s_log_groups_per_flex) - 1)) != 0)) && 2701c1d2c7d4SAlex Zhuravlev !(ext4_has_group_desc_csum(sb) && 2702c1d2c7d4SAlex Zhuravlev (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) 2703c1d2c7d4SAlex Zhuravlev return 0; 2704c1d2c7d4SAlex Zhuravlev ret = ext4_mb_init_group(sb, group, GFP_NOFS); 27058ef123feSRitesh Harjani if (ret) 27068ef123feSRitesh Harjani return ret; 27078ef123feSRitesh Harjani } 27088ef123feSRitesh Harjani 2709a5fda113STheodore Ts'o if (should_lock) { 271099377830SRitesh Harjani ext4_lock_group(sb, group); 2711a5fda113STheodore Ts'o __release(ext4_group_lock_ptr(sb, group)); 2712a5fda113STheodore Ts'o } 27138ef123feSRitesh Harjani ret = ext4_mb_good_group(ac, group, cr); 27148ef123feSRitesh Harjani out: 2715a5fda113STheodore Ts'o if (should_lock) { 2716a5fda113STheodore Ts'o __acquire(ext4_group_lock_ptr(sb, group)); 271799377830SRitesh Harjani ext4_unlock_group(sb, group); 2718a5fda113STheodore Ts'o } 27198ef123feSRitesh Harjani return ret; 2720c9de560dSAlex Tomas } 2721c9de560dSAlex Tomas 2722cfd73237SAlex Zhuravlev /* 2723cfd73237SAlex Zhuravlev * Start prefetching @nr block bitmaps starting at @group. 2724cfd73237SAlex Zhuravlev * Return the next group which needs to be prefetched. 2725cfd73237SAlex Zhuravlev */ 27263d392b26STheodore Ts'o ext4_group_t ext4_mb_prefetch(struct super_block *sb, ext4_group_t group, 2727cfd73237SAlex Zhuravlev unsigned int nr, int *cnt) 2728cfd73237SAlex Zhuravlev { 2729cfd73237SAlex Zhuravlev ext4_group_t ngroups = ext4_get_groups_count(sb); 2730cfd73237SAlex Zhuravlev struct buffer_head *bh; 2731cfd73237SAlex Zhuravlev struct blk_plug plug; 2732cfd73237SAlex Zhuravlev 2733cfd73237SAlex Zhuravlev blk_start_plug(&plug); 2734cfd73237SAlex Zhuravlev while (nr-- > 0) { 2735cfd73237SAlex Zhuravlev struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, 2736cfd73237SAlex Zhuravlev NULL); 2737cfd73237SAlex Zhuravlev struct ext4_group_info *grp = ext4_get_group_info(sb, group); 2738cfd73237SAlex Zhuravlev 2739cfd73237SAlex Zhuravlev /* 2740cfd73237SAlex Zhuravlev * Prefetch block groups with free blocks; but don't 2741cfd73237SAlex Zhuravlev * bother if it is marked uninitialized on disk, since 2742cfd73237SAlex Zhuravlev * it won't require I/O to read. Also only try to 2743cfd73237SAlex Zhuravlev * prefetch once, so we avoid getblk() call, which can 2744cfd73237SAlex Zhuravlev * be expensive. 2745cfd73237SAlex Zhuravlev */ 27465354b2afSTheodore Ts'o if (gdp && grp && !EXT4_MB_GRP_TEST_AND_SET_READ(grp) && 2747cfd73237SAlex Zhuravlev EXT4_MB_GRP_NEED_INIT(grp) && 27483c629604SOjaswin Mujoo ext4_free_group_clusters(sb, gdp) > 0 ) { 2749cfd73237SAlex Zhuravlev bh = ext4_read_block_bitmap_nowait(sb, group, true); 2750cfd73237SAlex Zhuravlev if (bh && !IS_ERR(bh)) { 2751cfd73237SAlex Zhuravlev if (!buffer_uptodate(bh) && cnt) 2752cfd73237SAlex Zhuravlev (*cnt)++; 2753cfd73237SAlex Zhuravlev brelse(bh); 2754cfd73237SAlex Zhuravlev } 2755cfd73237SAlex Zhuravlev } 2756cfd73237SAlex Zhuravlev if (++group >= ngroups) 2757cfd73237SAlex Zhuravlev group = 0; 2758cfd73237SAlex Zhuravlev } 2759cfd73237SAlex Zhuravlev blk_finish_plug(&plug); 2760cfd73237SAlex Zhuravlev return group; 2761cfd73237SAlex Zhuravlev } 2762cfd73237SAlex Zhuravlev 2763cfd73237SAlex Zhuravlev /* 2764cfd73237SAlex Zhuravlev * Prefetching reads the block bitmap into the buffer cache; but we 2765cfd73237SAlex Zhuravlev * need to make sure that the buddy bitmap in the page cache has been 2766cfd73237SAlex Zhuravlev * initialized. Note that ext4_mb_init_group() will block if the I/O 2767cfd73237SAlex Zhuravlev * is not yet completed, or indeed if it was not initiated by 2768cfd73237SAlex Zhuravlev * ext4_mb_prefetch did not start the I/O. 2769cfd73237SAlex Zhuravlev * 2770cfd73237SAlex Zhuravlev * TODO: We should actually kick off the buddy bitmap setup in a work 2771cfd73237SAlex Zhuravlev * queue when the buffer I/O is completed, so that we don't block 2772cfd73237SAlex Zhuravlev * waiting for the block allocation bitmap read to finish when 2773cfd73237SAlex Zhuravlev * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator(). 2774cfd73237SAlex Zhuravlev */ 27753d392b26STheodore Ts'o void ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group, 2776cfd73237SAlex Zhuravlev unsigned int nr) 2777cfd73237SAlex Zhuravlev { 277822fab984SKemeng Shi struct ext4_group_desc *gdp; 277922fab984SKemeng Shi struct ext4_group_info *grp; 2780cfd73237SAlex Zhuravlev 278122fab984SKemeng Shi while (nr-- > 0) { 2782cfd73237SAlex Zhuravlev if (!group) 2783cfd73237SAlex Zhuravlev group = ext4_get_groups_count(sb); 2784cfd73237SAlex Zhuravlev group--; 278522fab984SKemeng Shi gdp = ext4_get_group_desc(sb, group, NULL); 2786cfd73237SAlex Zhuravlev grp = ext4_get_group_info(sb, group); 2787cfd73237SAlex Zhuravlev 27885354b2afSTheodore Ts'o if (grp && gdp && EXT4_MB_GRP_NEED_INIT(grp) && 27893c629604SOjaswin Mujoo ext4_free_group_clusters(sb, gdp) > 0) { 2790cfd73237SAlex Zhuravlev if (ext4_mb_init_group(sb, group, GFP_NOFS)) 2791cfd73237SAlex Zhuravlev break; 2792cfd73237SAlex Zhuravlev } 2793cfd73237SAlex Zhuravlev } 2794cfd73237SAlex Zhuravlev } 2795cfd73237SAlex Zhuravlev 27964ddfef7bSEric Sandeen static noinline_for_stack int 27974ddfef7bSEric Sandeen ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 2798c9de560dSAlex Tomas { 2799cfd73237SAlex Zhuravlev ext4_group_t prefetch_grp = 0, ngroups, group, i; 28004c0cfebdSTheodore Ts'o enum criteria new_cr, cr = CR_GOAL_LEN_FAST; 280142ac1848SLukas Czerner int err = 0, first_err = 0; 2802cfd73237SAlex Zhuravlev unsigned int nr = 0, prefetch_ios = 0; 2803c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2804c9de560dSAlex Tomas struct super_block *sb; 2805c9de560dSAlex Tomas struct ext4_buddy e4b; 280666d5e027Sbrookxu int lost; 2807c9de560dSAlex Tomas 2808c9de560dSAlex Tomas sb = ac->ac_sb; 2809c9de560dSAlex Tomas sbi = EXT4_SB(sb); 28108df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 2811fb0a387dSEric Sandeen /* non-extent files are limited to low blocks/groups */ 281212e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) 2813fb0a387dSEric Sandeen ngroups = sbi->s_blockfile_groups; 2814fb0a387dSEric Sandeen 2815c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2816c9de560dSAlex Tomas 2817c9de560dSAlex Tomas /* first, try the goal */ 2818c9de560dSAlex Tomas err = ext4_mb_find_by_goal(ac, &e4b); 2819c9de560dSAlex Tomas if (err || ac->ac_status == AC_STATUS_FOUND) 2820c9de560dSAlex Tomas goto out; 2821c9de560dSAlex Tomas 2822c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2823c9de560dSAlex Tomas goto out; 2824c9de560dSAlex Tomas 2825c9de560dSAlex Tomas /* 2826e9a3cd48Sbrookxu * ac->ac_2order is set only if the fe_len is a power of 2 28274eea9fbeSKemeng Shi * if ac->ac_2order is set we also set criteria to CR_POWER2_ALIGNED 28284eea9fbeSKemeng Shi * so that we try exact allocation using buddy. 2829c9de560dSAlex Tomas */ 2830c9de560dSAlex Tomas i = fls(ac->ac_g_ex.fe_len); 2831c9de560dSAlex Tomas ac->ac_2order = 0; 2832c9de560dSAlex Tomas /* 2833c9de560dSAlex Tomas * We search using buddy data only if the order of the request 2834c9de560dSAlex Tomas * is greater than equal to the sbi_s_mb_order2_reqs 2835b713a5ecSTheodore Ts'o * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 2836d9b22cf9SJan Kara * We also support searching for power-of-two requests only for 2837d9b22cf9SJan Kara * requests upto maximum buddy size we have constructed. 2838c9de560dSAlex Tomas */ 28394b68f6dfSHarshad Shirwadkar if (i >= sbi->s_mb_order2_reqs && i <= MB_NUM_ORDERS(sb)) { 2840bb60caa2SKemeng Shi if (is_power_of_2(ac->ac_g_ex.fe_len)) 28411a5d5e5dSJeremy Cline ac->ac_2order = array_index_nospec(i - 1, 28424b68f6dfSHarshad Shirwadkar MB_NUM_ORDERS(sb)); 2843c9de560dSAlex Tomas } 2844c9de560dSAlex Tomas 28454ba74d00STheodore Ts'o /* if stream allocation is enabled, use global goal */ 28464ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2847c9de560dSAlex Tomas /* TBD: may be hot point */ 2848c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 2849c9de560dSAlex Tomas ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 2850c9de560dSAlex Tomas ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 2851c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 2852c9de560dSAlex Tomas } 28534ba74d00STheodore Ts'o 2854c9de560dSAlex Tomas /* 28554c0cfebdSTheodore Ts'o * Let's just scan groups to find more-less suitable blocks We 28564c0cfebdSTheodore Ts'o * start with CR_GOAL_LEN_FAST, unless it is power of 2 28574c0cfebdSTheodore Ts'o * aligned, in which case let's do that faster approach first. 2858c9de560dSAlex Tomas */ 28594c0cfebdSTheodore Ts'o if (ac->ac_2order) 28604c0cfebdSTheodore Ts'o cr = CR_POWER2_ALIGNED; 2861c9de560dSAlex Tomas repeat: 28624eb7a4a1SOjaswin Mujoo for (; cr < EXT4_MB_NUM_CRS && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 2863c9de560dSAlex Tomas ac->ac_criteria = cr; 2864ed8f9c75SAneesh Kumar K.V /* 2865ed8f9c75SAneesh Kumar K.V * searching for the right group start 2866ed8f9c75SAneesh Kumar K.V * from the goal value specified 2867ed8f9c75SAneesh Kumar K.V */ 2868ed8f9c75SAneesh Kumar K.V group = ac->ac_g_ex.fe_group; 2869196e402aSHarshad Shirwadkar ac->ac_groups_linear_remaining = sbi->s_mb_max_linear_groups; 2870cfd73237SAlex Zhuravlev prefetch_grp = group; 28719c97c34aSKemeng Shi nr = 0; 2872ed8f9c75SAneesh Kumar K.V 28734fca50d4SJan Kara for (i = 0, new_cr = cr; i < ngroups; i++, 28744fca50d4SJan Kara ext4_mb_choose_next_group(ac, &new_cr, &group, ngroups)) { 28754fca50d4SJan Kara int ret = 0; 2876196e402aSHarshad Shirwadkar 28772ed5724dSTheodore Ts'o cond_resched(); 2878196e402aSHarshad Shirwadkar if (new_cr != cr) { 2879196e402aSHarshad Shirwadkar cr = new_cr; 2880196e402aSHarshad Shirwadkar goto repeat; 2881196e402aSHarshad Shirwadkar } 2882c9de560dSAlex Tomas 2883cfd73237SAlex Zhuravlev /* 2884cfd73237SAlex Zhuravlev * Batch reads of the block allocation bitmaps 2885cfd73237SAlex Zhuravlev * to get multiple READs in flight; limit 28864eea9fbeSKemeng Shi * prefetching at inexpensive CR, otherwise mballoc 28874eea9fbeSKemeng Shi * can spend a lot of time loading imperfect groups 2888cfd73237SAlex Zhuravlev */ 2889cfd73237SAlex Zhuravlev if ((prefetch_grp == group) && 2890304749c0SOjaswin Mujoo (ext4_mb_cr_expensive(cr) || 2891cfd73237SAlex Zhuravlev prefetch_ios < sbi->s_mb_prefetch_limit)) { 2892cfd73237SAlex Zhuravlev nr = sbi->s_mb_prefetch; 2893cfd73237SAlex Zhuravlev if (ext4_has_feature_flex_bg(sb)) { 289482ef1370SChunguang Xu nr = 1 << sbi->s_log_groups_per_flex; 289582ef1370SChunguang Xu nr -= group & (nr - 1); 289682ef1370SChunguang Xu nr = min(nr, sbi->s_mb_prefetch); 2897cfd73237SAlex Zhuravlev } 2898cfd73237SAlex Zhuravlev prefetch_grp = ext4_mb_prefetch(sb, group, 2899cfd73237SAlex Zhuravlev nr, &prefetch_ios); 2900cfd73237SAlex Zhuravlev } 2901cfd73237SAlex Zhuravlev 29028a57d9d6SCurt Wohlgemuth /* This now checks without needing the buddy page */ 29038ef123feSRitesh Harjani ret = ext4_mb_good_group_nolock(ac, group, cr); 290442ac1848SLukas Czerner if (ret <= 0) { 290542ac1848SLukas Czerner if (!first_err) 290642ac1848SLukas Czerner first_err = ret; 2907c9de560dSAlex Tomas continue; 290842ac1848SLukas Czerner } 2909c9de560dSAlex Tomas 2910c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2911c9de560dSAlex Tomas if (err) 2912c9de560dSAlex Tomas goto out; 2913c9de560dSAlex Tomas 2914c9de560dSAlex Tomas ext4_lock_group(sb, group); 29158a57d9d6SCurt Wohlgemuth 29168a57d9d6SCurt Wohlgemuth /* 29178a57d9d6SCurt Wohlgemuth * We need to check again after locking the 29188a57d9d6SCurt Wohlgemuth * block group 29198a57d9d6SCurt Wohlgemuth */ 292042ac1848SLukas Czerner ret = ext4_mb_good_group(ac, group, cr); 29218ef123feSRitesh Harjani if (ret == 0) { 2922c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2923e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2924c9de560dSAlex Tomas continue; 2925c9de560dSAlex Tomas } 2926c9de560dSAlex Tomas 2927c9de560dSAlex Tomas ac->ac_groups_scanned++; 2928f52f3d2bSOjaswin Mujoo if (cr == CR_POWER2_ALIGNED) 2929c9de560dSAlex Tomas ext4_mb_simple_scan_group(ac, &e4b); 29301f6bc02fSOjaswin Mujoo else { 29311f6bc02fSOjaswin Mujoo bool is_stripe_aligned = sbi->s_stripe && 2932c3defd99SKemeng Shi !(ac->ac_g_ex.fe_len % 29331f6bc02fSOjaswin Mujoo EXT4_B2C(sbi, sbi->s_stripe)); 29341f6bc02fSOjaswin Mujoo 29351f6bc02fSOjaswin Mujoo if ((cr == CR_GOAL_LEN_FAST || 29361f6bc02fSOjaswin Mujoo cr == CR_BEST_AVAIL_LEN) && 29371f6bc02fSOjaswin Mujoo is_stripe_aligned) 2938c9de560dSAlex Tomas ext4_mb_scan_aligned(ac, &e4b); 29391f6bc02fSOjaswin Mujoo 29401f6bc02fSOjaswin Mujoo if (ac->ac_status == AC_STATUS_CONTINUE) 2941c9de560dSAlex Tomas ext4_mb_complex_scan_group(ac, &e4b); 29421f6bc02fSOjaswin Mujoo } 2943c9de560dSAlex Tomas 2944c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2945e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2946c9de560dSAlex Tomas 2947c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_CONTINUE) 2948c9de560dSAlex Tomas break; 2949c9de560dSAlex Tomas } 2950a6c75eafSHarshad Shirwadkar /* Processed all groups and haven't found blocks */ 2951a6c75eafSHarshad Shirwadkar if (sbi->s_mb_stats && i == ngroups) 2952a6c75eafSHarshad Shirwadkar atomic64_inc(&sbi->s_bal_cX_failed[cr]); 29537e170922SOjaswin Mujoo 2954f52f3d2bSOjaswin Mujoo if (i == ngroups && ac->ac_criteria == CR_BEST_AVAIL_LEN) 29557e170922SOjaswin Mujoo /* Reset goal length to original goal length before 2956f52f3d2bSOjaswin Mujoo * falling into CR_GOAL_LEN_SLOW */ 29577e170922SOjaswin Mujoo ac->ac_g_ex.fe_len = ac->ac_orig_goal_len; 2958c9de560dSAlex Tomas } 2959c9de560dSAlex Tomas 2960c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2961c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2962c9de560dSAlex Tomas /* 2963c9de560dSAlex Tomas * We've been searching too long. Let's try to allocate 2964c9de560dSAlex Tomas * the best chunk we've found so far 2965c9de560dSAlex Tomas */ 2966c9de560dSAlex Tomas ext4_mb_try_best_found(ac, &e4b); 2967c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_FOUND) { 2968c9de560dSAlex Tomas /* 2969c9de560dSAlex Tomas * Someone more lucky has already allocated it. 2970c9de560dSAlex Tomas * The only thing we can do is just take first 2971c9de560dSAlex Tomas * found block(s) 2972c9de560dSAlex Tomas */ 297366d5e027Sbrookxu lost = atomic_inc_return(&sbi->s_mb_lost_chunks); 297466d5e027Sbrookxu mb_debug(sb, "lost chunk, group: %u, start: %d, len: %d, lost: %d\n", 2975c55ee7d2Sbrookxu ac->ac_b_ex.fe_group, ac->ac_b_ex.fe_start, 2976c55ee7d2Sbrookxu ac->ac_b_ex.fe_len, lost); 2977c55ee7d2Sbrookxu 2978c9de560dSAlex Tomas ac->ac_b_ex.fe_group = 0; 2979c9de560dSAlex Tomas ac->ac_b_ex.fe_start = 0; 2980c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 0; 2981c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 2982c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_FIRST; 2983f52f3d2bSOjaswin Mujoo cr = CR_ANY_FREE; 2984c9de560dSAlex Tomas goto repeat; 2985c9de560dSAlex Tomas } 2986c9de560dSAlex Tomas } 2987a6c75eafSHarshad Shirwadkar 2988a6c75eafSHarshad Shirwadkar if (sbi->s_mb_stats && ac->ac_status == AC_STATUS_FOUND) 2989a6c75eafSHarshad Shirwadkar atomic64_inc(&sbi->s_bal_cX_hits[ac->ac_criteria]); 2990c9de560dSAlex Tomas out: 299142ac1848SLukas Czerner if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) 299242ac1848SLukas Czerner err = first_err; 2993bbc4ec77SRitesh Harjani 2994d3df1453SRitesh Harjani mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n", 2995bbc4ec77SRitesh Harjani ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status, 2996bbc4ec77SRitesh Harjani ac->ac_flags, cr, err); 2997cfd73237SAlex Zhuravlev 2998cfd73237SAlex Zhuravlev if (nr) 2999cfd73237SAlex Zhuravlev ext4_mb_prefetch_fini(sb, prefetch_grp, nr); 3000cfd73237SAlex Zhuravlev 3001c9de560dSAlex Tomas return err; 3002c9de560dSAlex Tomas } 3003c9de560dSAlex Tomas 3004c9de560dSAlex Tomas static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 3005c9de560dSAlex Tomas { 3006359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3007c9de560dSAlex Tomas ext4_group_t group; 3008c9de560dSAlex Tomas 30098df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 3010c9de560dSAlex Tomas return NULL; 3011c9de560dSAlex Tomas group = *pos + 1; 3012a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 3013c9de560dSAlex Tomas } 3014c9de560dSAlex Tomas 3015c9de560dSAlex Tomas static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 3016c9de560dSAlex Tomas { 3017359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3018c9de560dSAlex Tomas ext4_group_t group; 3019c9de560dSAlex Tomas 3020c9de560dSAlex Tomas ++*pos; 30218df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 3022c9de560dSAlex Tomas return NULL; 3023c9de560dSAlex Tomas group = *pos + 1; 3024a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 3025c9de560dSAlex Tomas } 3026c9de560dSAlex Tomas 3027c9de560dSAlex Tomas static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 3028c9de560dSAlex Tomas { 3029359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3030a9df9a49STheodore Ts'o ext4_group_t group = (ext4_group_t) ((unsigned long) v); 303125044880Syangerkun int i, err; 30324b55d343Syangerkun char nbuf[16]; 3033c9de560dSAlex Tomas struct ext4_buddy e4b; 30341c8457caSAditya Kali struct ext4_group_info *grinfo; 30352df2c340SArnd Bergmann unsigned char blocksize_bits = min_t(unsigned char, 30362df2c340SArnd Bergmann sb->s_blocksize_bits, 30372df2c340SArnd Bergmann EXT4_MAX_BLOCK_LOG_SIZE); 3038c9de560dSAlex Tomas struct sg { 3039c9de560dSAlex Tomas struct ext4_group_info info; 3040b80b32b6STheodore Ts'o ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2]; 3041c9de560dSAlex Tomas } sg; 3042c9de560dSAlex Tomas 3043c9de560dSAlex Tomas group--; 3044c9de560dSAlex Tomas if (group == 0) 304597b4af2fSRasmus Villemoes seq_puts(seq, "#group: free frags first [" 304697b4af2fSRasmus Villemoes " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " 3047802cf1f9SHuaitong Han " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); 3048c9de560dSAlex Tomas 3049b80b32b6STheodore Ts'o i = (blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 3050b80b32b6STheodore Ts'o sizeof(struct ext4_group_info); 3051b80b32b6STheodore Ts'o 30521c8457caSAditya Kali grinfo = ext4_get_group_info(sb, group); 30535354b2afSTheodore Ts'o if (!grinfo) 30545354b2afSTheodore Ts'o return 0; 30551c8457caSAditya Kali /* Load the group info in memory only if not already loaded. */ 30561c8457caSAditya Kali if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { 3057c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3058c9de560dSAlex Tomas if (err) { 30594b55d343Syangerkun seq_printf(seq, "#%-5u: %s\n", group, ext4_decode_error(NULL, err, nbuf)); 3060c9de560dSAlex Tomas return 0; 3061c9de560dSAlex Tomas } 306225044880Syangerkun ext4_mb_unload_buddy(&e4b); 30631c8457caSAditya Kali } 30641c8457caSAditya Kali 306525044880Syangerkun /* 306625044880Syangerkun * We care only about free space counters in the group info and 306725044880Syangerkun * these are safe to access even after the buddy has been unloaded 306825044880Syangerkun */ 30695354b2afSTheodore Ts'o memcpy(&sg, grinfo, i); 3070a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 3071c9de560dSAlex Tomas sg.info.bb_fragments, sg.info.bb_first_free); 3072c9de560dSAlex Tomas for (i = 0; i <= 13; i++) 30732df2c340SArnd Bergmann seq_printf(seq, " %-5u", i <= blocksize_bits + 1 ? 3074c9de560dSAlex Tomas sg.info.bb_counters[i] : 0); 307568ee261fSZhang Yi seq_puts(seq, " ]"); 307668ee261fSZhang Yi if (EXT4_MB_GRP_BBITMAP_CORRUPT(&sg.info)) 307768ee261fSZhang Yi seq_puts(seq, " Block bitmap corrupted!"); 307868ee261fSZhang Yi seq_puts(seq, "\n"); 3079c9de560dSAlex Tomas 3080c9de560dSAlex Tomas return 0; 3081c9de560dSAlex Tomas } 3082c9de560dSAlex Tomas 3083c9de560dSAlex Tomas static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 3084c9de560dSAlex Tomas { 3085c9de560dSAlex Tomas } 3086c9de560dSAlex Tomas 3087247dbed8SChristoph Hellwig const struct seq_operations ext4_mb_seq_groups_ops = { 3088c9de560dSAlex Tomas .start = ext4_mb_seq_groups_start, 3089c9de560dSAlex Tomas .next = ext4_mb_seq_groups_next, 3090c9de560dSAlex Tomas .stop = ext4_mb_seq_groups_stop, 3091c9de560dSAlex Tomas .show = ext4_mb_seq_groups_show, 3092c9de560dSAlex Tomas }; 3093c9de560dSAlex Tomas 3094a6c75eafSHarshad Shirwadkar int ext4_seq_mb_stats_show(struct seq_file *seq, void *offset) 3095a6c75eafSHarshad Shirwadkar { 3096c30365b9SYu Zhe struct super_block *sb = seq->private; 3097a6c75eafSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 3098a6c75eafSHarshad Shirwadkar 3099a6c75eafSHarshad Shirwadkar seq_puts(seq, "mballoc:\n"); 3100a6c75eafSHarshad Shirwadkar if (!sbi->s_mb_stats) { 3101a6c75eafSHarshad Shirwadkar seq_puts(seq, "\tmb stats collection turned off.\n"); 3102f52f3d2bSOjaswin Mujoo seq_puts( 3103f52f3d2bSOjaswin Mujoo seq, 3104f52f3d2bSOjaswin Mujoo "\tTo enable, please write \"1\" to sysfs file mb_stats.\n"); 3105a6c75eafSHarshad Shirwadkar return 0; 3106a6c75eafSHarshad Shirwadkar } 3107a6c75eafSHarshad Shirwadkar seq_printf(seq, "\treqs: %u\n", atomic_read(&sbi->s_bal_reqs)); 3108a6c75eafSHarshad Shirwadkar seq_printf(seq, "\tsuccess: %u\n", atomic_read(&sbi->s_bal_success)); 3109a6c75eafSHarshad Shirwadkar 3110f52f3d2bSOjaswin Mujoo seq_printf(seq, "\tgroups_scanned: %u\n", 3111f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_groups_scanned)); 3112a6c75eafSHarshad Shirwadkar 3113f52f3d2bSOjaswin Mujoo /* CR_POWER2_ALIGNED stats */ 3114f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_p2_aligned_stats:\n"); 3115f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3116f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_POWER2_ALIGNED])); 3117f52f3d2bSOjaswin Mujoo seq_printf( 3118f52f3d2bSOjaswin Mujoo seq, "\t\tgroups_considered: %llu\n", 3119f52f3d2bSOjaswin Mujoo atomic64_read( 3120f52f3d2bSOjaswin Mujoo &sbi->s_bal_cX_groups_considered[CR_POWER2_ALIGNED])); 3121f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3122f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_POWER2_ALIGNED])); 3123a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tuseless_loops: %llu\n", 3124f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_POWER2_ALIGNED])); 3125196e402aSHarshad Shirwadkar seq_printf(seq, "\t\tbad_suggestions: %u\n", 3126f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_p2_aligned_bad_suggestions)); 3127a6c75eafSHarshad Shirwadkar 3128f52f3d2bSOjaswin Mujoo /* CR_GOAL_LEN_FAST stats */ 3129f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_goal_fast_stats:\n"); 3130f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3131f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_FAST])); 3132a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tgroups_considered: %llu\n", 3133f52f3d2bSOjaswin Mujoo atomic64_read( 3134f52f3d2bSOjaswin Mujoo &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_FAST])); 3135f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3136f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_FAST])); 3137a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tuseless_loops: %llu\n", 3138f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_FAST])); 3139196e402aSHarshad Shirwadkar seq_printf(seq, "\t\tbad_suggestions: %u\n", 3140f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_goal_fast_bad_suggestions)); 3141a6c75eafSHarshad Shirwadkar 3142f52f3d2bSOjaswin Mujoo /* CR_BEST_AVAIL_LEN stats */ 3143f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_best_avail_stats:\n"); 3144f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3145f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_BEST_AVAIL_LEN])); 3146f52f3d2bSOjaswin Mujoo seq_printf( 3147f52f3d2bSOjaswin Mujoo seq, "\t\tgroups_considered: %llu\n", 3148f52f3d2bSOjaswin Mujoo atomic64_read( 3149f52f3d2bSOjaswin Mujoo &sbi->s_bal_cX_groups_considered[CR_BEST_AVAIL_LEN])); 3150f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3151f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_BEST_AVAIL_LEN])); 31527e170922SOjaswin Mujoo seq_printf(seq, "\t\tuseless_loops: %llu\n", 3153f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_BEST_AVAIL_LEN])); 31547e170922SOjaswin Mujoo seq_printf(seq, "\t\tbad_suggestions: %u\n", 3155f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_best_avail_bad_suggestions)); 31567e170922SOjaswin Mujoo 3157f52f3d2bSOjaswin Mujoo /* CR_GOAL_LEN_SLOW stats */ 3158f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_goal_slow_stats:\n"); 3159f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3160f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_GOAL_LEN_SLOW])); 3161a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tgroups_considered: %llu\n", 3162f52f3d2bSOjaswin Mujoo atomic64_read( 3163f52f3d2bSOjaswin Mujoo &sbi->s_bal_cX_groups_considered[CR_GOAL_LEN_SLOW])); 3164f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3165f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_GOAL_LEN_SLOW])); 3166a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tuseless_loops: %llu\n", 3167f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_GOAL_LEN_SLOW])); 3168a6c75eafSHarshad Shirwadkar 3169f52f3d2bSOjaswin Mujoo /* CR_ANY_FREE stats */ 3170f52f3d2bSOjaswin Mujoo seq_puts(seq, "\tcr_any_free_stats:\n"); 3171f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\thits: %llu\n", 3172f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_hits[CR_ANY_FREE])); 3173f52f3d2bSOjaswin Mujoo seq_printf( 3174f52f3d2bSOjaswin Mujoo seq, "\t\tgroups_considered: %llu\n", 3175f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_groups_considered[CR_ANY_FREE])); 3176f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\textents_scanned: %u\n", 3177f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_cX_ex_scanned[CR_ANY_FREE])); 3178a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tuseless_loops: %llu\n", 3179f52f3d2bSOjaswin Mujoo atomic64_read(&sbi->s_bal_cX_failed[CR_ANY_FREE])); 3180f52f3d2bSOjaswin Mujoo 3181f52f3d2bSOjaswin Mujoo /* Aggregates */ 3182f52f3d2bSOjaswin Mujoo seq_printf(seq, "\textents_scanned: %u\n", 3183f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_ex_scanned)); 3184a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tgoal_hits: %u\n", atomic_read(&sbi->s_bal_goals)); 3185f52f3d2bSOjaswin Mujoo seq_printf(seq, "\t\tlen_goal_hits: %u\n", 3186f52f3d2bSOjaswin Mujoo atomic_read(&sbi->s_bal_len_goals)); 3187a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\t2^n_hits: %u\n", atomic_read(&sbi->s_bal_2orders)); 3188a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tbreaks: %u\n", atomic_read(&sbi->s_bal_breaks)); 3189a6c75eafSHarshad Shirwadkar seq_printf(seq, "\t\tlost: %u\n", atomic_read(&sbi->s_mb_lost_chunks)); 3190a6c75eafSHarshad Shirwadkar seq_printf(seq, "\tbuddies_generated: %u/%u\n", 3191a6c75eafSHarshad Shirwadkar atomic_read(&sbi->s_mb_buddies_generated), 3192a6c75eafSHarshad Shirwadkar ext4_get_groups_count(sb)); 3193a6c75eafSHarshad Shirwadkar seq_printf(seq, "\tbuddies_time_used: %llu\n", 3194a6c75eafSHarshad Shirwadkar atomic64_read(&sbi->s_mb_generation_time)); 3195a6c75eafSHarshad Shirwadkar seq_printf(seq, "\tpreallocated: %u\n", 3196a6c75eafSHarshad Shirwadkar atomic_read(&sbi->s_mb_preallocated)); 3197f52f3d2bSOjaswin Mujoo seq_printf(seq, "\tdiscarded: %u\n", atomic_read(&sbi->s_mb_discarded)); 3198a6c75eafSHarshad Shirwadkar return 0; 3199a6c75eafSHarshad Shirwadkar } 3200a6c75eafSHarshad Shirwadkar 3201f68f4063SHarshad Shirwadkar static void *ext4_mb_seq_structs_summary_start(struct seq_file *seq, loff_t *pos) 3202f68f4063SHarshad Shirwadkar { 3203359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3204f68f4063SHarshad Shirwadkar unsigned long position; 3205f68f4063SHarshad Shirwadkar 320683e80a6eSJan Kara if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) 3207f68f4063SHarshad Shirwadkar return NULL; 3208f68f4063SHarshad Shirwadkar position = *pos + 1; 3209f68f4063SHarshad Shirwadkar return (void *) ((unsigned long) position); 3210f68f4063SHarshad Shirwadkar } 3211f68f4063SHarshad Shirwadkar 3212f68f4063SHarshad Shirwadkar static void *ext4_mb_seq_structs_summary_next(struct seq_file *seq, void *v, loff_t *pos) 3213f68f4063SHarshad Shirwadkar { 3214359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3215f68f4063SHarshad Shirwadkar unsigned long position; 3216f68f4063SHarshad Shirwadkar 3217f68f4063SHarshad Shirwadkar ++*pos; 321883e80a6eSJan Kara if (*pos < 0 || *pos >= 2*MB_NUM_ORDERS(sb)) 3219f68f4063SHarshad Shirwadkar return NULL; 3220f68f4063SHarshad Shirwadkar position = *pos + 1; 3221f68f4063SHarshad Shirwadkar return (void *) ((unsigned long) position); 3222f68f4063SHarshad Shirwadkar } 3223f68f4063SHarshad Shirwadkar 3224f68f4063SHarshad Shirwadkar static int ext4_mb_seq_structs_summary_show(struct seq_file *seq, void *v) 3225f68f4063SHarshad Shirwadkar { 3226359745d7SMuchun Song struct super_block *sb = pde_data(file_inode(seq->file)); 3227f68f4063SHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 3228f68f4063SHarshad Shirwadkar unsigned long position = ((unsigned long) v); 3229f68f4063SHarshad Shirwadkar struct ext4_group_info *grp; 323083e80a6eSJan Kara unsigned int count; 3231f68f4063SHarshad Shirwadkar 3232f68f4063SHarshad Shirwadkar position--; 3233f68f4063SHarshad Shirwadkar if (position >= MB_NUM_ORDERS(sb)) { 323483e80a6eSJan Kara position -= MB_NUM_ORDERS(sb); 323583e80a6eSJan Kara if (position == 0) 323683e80a6eSJan Kara seq_puts(seq, "avg_fragment_size_lists:\n"); 3237f68f4063SHarshad Shirwadkar 323883e80a6eSJan Kara count = 0; 323983e80a6eSJan Kara read_lock(&sbi->s_mb_avg_fragment_size_locks[position]); 324083e80a6eSJan Kara list_for_each_entry(grp, &sbi->s_mb_avg_fragment_size[position], 324183e80a6eSJan Kara bb_avg_fragment_size_node) 324283e80a6eSJan Kara count++; 324383e80a6eSJan Kara read_unlock(&sbi->s_mb_avg_fragment_size_locks[position]); 324483e80a6eSJan Kara seq_printf(seq, "\tlist_order_%u_groups: %u\n", 324583e80a6eSJan Kara (unsigned int)position, count); 3246f68f4063SHarshad Shirwadkar return 0; 3247f68f4063SHarshad Shirwadkar } 3248f68f4063SHarshad Shirwadkar 3249f68f4063SHarshad Shirwadkar if (position == 0) { 3250f68f4063SHarshad Shirwadkar seq_printf(seq, "optimize_scan: %d\n", 3251f68f4063SHarshad Shirwadkar test_opt2(sb, MB_OPTIMIZE_SCAN) ? 1 : 0); 3252f68f4063SHarshad Shirwadkar seq_puts(seq, "max_free_order_lists:\n"); 3253f68f4063SHarshad Shirwadkar } 3254f68f4063SHarshad Shirwadkar count = 0; 325583e80a6eSJan Kara read_lock(&sbi->s_mb_largest_free_orders_locks[position]); 3256f68f4063SHarshad Shirwadkar list_for_each_entry(grp, &sbi->s_mb_largest_free_orders[position], 3257f68f4063SHarshad Shirwadkar bb_largest_free_order_node) 3258f68f4063SHarshad Shirwadkar count++; 325983e80a6eSJan Kara read_unlock(&sbi->s_mb_largest_free_orders_locks[position]); 3260f68f4063SHarshad Shirwadkar seq_printf(seq, "\tlist_order_%u_groups: %u\n", 3261f68f4063SHarshad Shirwadkar (unsigned int)position, count); 3262f68f4063SHarshad Shirwadkar 3263f68f4063SHarshad Shirwadkar return 0; 3264f68f4063SHarshad Shirwadkar } 3265f68f4063SHarshad Shirwadkar 3266f68f4063SHarshad Shirwadkar static void ext4_mb_seq_structs_summary_stop(struct seq_file *seq, void *v) 3267f68f4063SHarshad Shirwadkar { 3268f68f4063SHarshad Shirwadkar } 3269f68f4063SHarshad Shirwadkar 3270f68f4063SHarshad Shirwadkar const struct seq_operations ext4_mb_seq_structs_summary_ops = { 3271f68f4063SHarshad Shirwadkar .start = ext4_mb_seq_structs_summary_start, 3272f68f4063SHarshad Shirwadkar .next = ext4_mb_seq_structs_summary_next, 3273f68f4063SHarshad Shirwadkar .stop = ext4_mb_seq_structs_summary_stop, 3274f68f4063SHarshad Shirwadkar .show = ext4_mb_seq_structs_summary_show, 3275f68f4063SHarshad Shirwadkar }; 3276f68f4063SHarshad Shirwadkar 3277fb1813f4SCurt Wohlgemuth static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) 3278fb1813f4SCurt Wohlgemuth { 3279fb1813f4SCurt Wohlgemuth int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 3280fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; 3281fb1813f4SCurt Wohlgemuth 3282fb1813f4SCurt Wohlgemuth BUG_ON(!cachep); 3283fb1813f4SCurt Wohlgemuth return cachep; 3284fb1813f4SCurt Wohlgemuth } 32855f21b0e6SFrederic Bohe 328628623c2fSTheodore Ts'o /* 328728623c2fSTheodore Ts'o * Allocate the top-level s_group_info array for the specified number 328828623c2fSTheodore Ts'o * of groups 328928623c2fSTheodore Ts'o */ 329028623c2fSTheodore Ts'o int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) 329128623c2fSTheodore Ts'o { 329228623c2fSTheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(sb); 329328623c2fSTheodore Ts'o unsigned size; 3294df3da4eaSSuraj Jitindar Singh struct ext4_group_info ***old_groupinfo, ***new_groupinfo; 329528623c2fSTheodore Ts'o 329628623c2fSTheodore Ts'o size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> 329728623c2fSTheodore Ts'o EXT4_DESC_PER_BLOCK_BITS(sb); 329828623c2fSTheodore Ts'o if (size <= sbi->s_group_info_size) 329928623c2fSTheodore Ts'o return 0; 330028623c2fSTheodore Ts'o 330128623c2fSTheodore Ts'o size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); 3302a7c3e901SMichal Hocko new_groupinfo = kvzalloc(size, GFP_KERNEL); 330328623c2fSTheodore Ts'o if (!new_groupinfo) { 330428623c2fSTheodore Ts'o ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); 330528623c2fSTheodore Ts'o return -ENOMEM; 330628623c2fSTheodore Ts'o } 3307df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3308df3da4eaSSuraj Jitindar Singh old_groupinfo = rcu_dereference(sbi->s_group_info); 3309df3da4eaSSuraj Jitindar Singh if (old_groupinfo) 3310df3da4eaSSuraj Jitindar Singh memcpy(new_groupinfo, old_groupinfo, 331128623c2fSTheodore Ts'o sbi->s_group_info_size * sizeof(*sbi->s_group_info)); 3312df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3313df3da4eaSSuraj Jitindar Singh rcu_assign_pointer(sbi->s_group_info, new_groupinfo); 331428623c2fSTheodore Ts'o sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); 3315df3da4eaSSuraj Jitindar Singh if (old_groupinfo) 3316df3da4eaSSuraj Jitindar Singh ext4_kvfree_array_rcu(old_groupinfo); 331728623c2fSTheodore Ts'o ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 331828623c2fSTheodore Ts'o sbi->s_group_info_size); 331928623c2fSTheodore Ts'o return 0; 332028623c2fSTheodore Ts'o } 332128623c2fSTheodore Ts'o 33225f21b0e6SFrederic Bohe /* Create and initialize ext4_group_info data for the given group. */ 3323920313a7SAneesh Kumar K.V int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 33245f21b0e6SFrederic Bohe struct ext4_group_desc *desc) 33255f21b0e6SFrederic Bohe { 3326fb1813f4SCurt Wohlgemuth int i; 33275f21b0e6SFrederic Bohe int metalen = 0; 3328df3da4eaSSuraj Jitindar Singh int idx = group >> EXT4_DESC_PER_BLOCK_BITS(sb); 33295f21b0e6SFrederic Bohe struct ext4_sb_info *sbi = EXT4_SB(sb); 33305f21b0e6SFrederic Bohe struct ext4_group_info **meta_group_info; 3331fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 33325f21b0e6SFrederic Bohe 33335f21b0e6SFrederic Bohe /* 33345f21b0e6SFrederic Bohe * First check if this group is the first of a reserved block. 33355f21b0e6SFrederic Bohe * If it's true, we have to allocate a new table of pointers 33365f21b0e6SFrederic Bohe * to ext4_group_info structures 33375f21b0e6SFrederic Bohe */ 33385f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 33395f21b0e6SFrederic Bohe metalen = sizeof(*meta_group_info) << 33405f21b0e6SFrederic Bohe EXT4_DESC_PER_BLOCK_BITS(sb); 33414fdb5543SDmitry Monakhov meta_group_info = kmalloc(metalen, GFP_NOFS); 33425f21b0e6SFrederic Bohe if (meta_group_info == NULL) { 33437f6a11e7SJoe Perches ext4_msg(sb, KERN_ERR, "can't allocate mem " 33449d8b9ec4STheodore Ts'o "for a buddy group"); 3345df119095SKemeng Shi return -ENOMEM; 33465f21b0e6SFrederic Bohe } 3347df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3348df3da4eaSSuraj Jitindar Singh rcu_dereference(sbi->s_group_info)[idx] = meta_group_info; 3349df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 33505f21b0e6SFrederic Bohe } 33515f21b0e6SFrederic Bohe 3352df3da4eaSSuraj Jitindar Singh meta_group_info = sbi_array_rcu_deref(sbi, s_group_info, idx); 33535f21b0e6SFrederic Bohe i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 33545f21b0e6SFrederic Bohe 33554fdb5543SDmitry Monakhov meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); 33565f21b0e6SFrederic Bohe if (meta_group_info[i] == NULL) { 33577f6a11e7SJoe Perches ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); 33585f21b0e6SFrederic Bohe goto exit_group_info; 33595f21b0e6SFrederic Bohe } 33605f21b0e6SFrederic Bohe set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 33615f21b0e6SFrederic Bohe &(meta_group_info[i]->bb_state)); 33625f21b0e6SFrederic Bohe 33635f21b0e6SFrederic Bohe /* 33645f21b0e6SFrederic Bohe * initialize bb_free to be able to skip 33655f21b0e6SFrederic Bohe * empty groups without initialization 33665f21b0e6SFrederic Bohe */ 33678844618dSTheodore Ts'o if (ext4_has_group_desc_csum(sb) && 33688844618dSTheodore Ts'o (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 33695f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 3370cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, group, desc); 33715f21b0e6SFrederic Bohe } else { 33725f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 3373021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc); 33745f21b0e6SFrederic Bohe } 33755f21b0e6SFrederic Bohe 33765f21b0e6SFrederic Bohe INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 3377920313a7SAneesh Kumar K.V init_rwsem(&meta_group_info[i]->alloc_sem); 337864e290ecSVenkatesh Pallipadi meta_group_info[i]->bb_free_root = RB_ROOT; 3379196e402aSHarshad Shirwadkar INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); 338083e80a6eSJan Kara INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node); 33818a57d9d6SCurt Wohlgemuth meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ 338283e80a6eSJan Kara meta_group_info[i]->bb_avg_fragment_size_order = -1; /* uninit */ 3383196e402aSHarshad Shirwadkar meta_group_info[i]->bb_group = group; 33845f21b0e6SFrederic Bohe 3385a3450215SRitesh Harjani mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group); 33865f21b0e6SFrederic Bohe return 0; 33875f21b0e6SFrederic Bohe 33885f21b0e6SFrederic Bohe exit_group_info: 33895f21b0e6SFrederic Bohe /* If a meta_group_info table has been allocated, release it now */ 3390caaf7a29STao Ma if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 3391df3da4eaSSuraj Jitindar Singh struct ext4_group_info ***group_info; 3392df3da4eaSSuraj Jitindar Singh 3393df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3394df3da4eaSSuraj Jitindar Singh group_info = rcu_dereference(sbi->s_group_info); 3395df3da4eaSSuraj Jitindar Singh kfree(group_info[idx]); 3396df3da4eaSSuraj Jitindar Singh group_info[idx] = NULL; 3397df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3398caaf7a29STao Ma } 33995f21b0e6SFrederic Bohe return -ENOMEM; 34005f21b0e6SFrederic Bohe } /* ext4_mb_add_groupinfo */ 34015f21b0e6SFrederic Bohe 3402c9de560dSAlex Tomas static int ext4_mb_init_backend(struct super_block *sb) 3403c9de560dSAlex Tomas { 34048df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 3405c9de560dSAlex Tomas ext4_group_t i; 3406c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 340728623c2fSTheodore Ts'o int err; 34085f21b0e6SFrederic Bohe struct ext4_group_desc *desc; 3409df3da4eaSSuraj Jitindar Singh struct ext4_group_info ***group_info; 3410fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep; 3411c9de560dSAlex Tomas 341228623c2fSTheodore Ts'o err = ext4_mb_alloc_groupinfo(sb, ngroups); 341328623c2fSTheodore Ts'o if (err) 341428623c2fSTheodore Ts'o return err; 34155f21b0e6SFrederic Bohe 3416c9de560dSAlex Tomas sbi->s_buddy_cache = new_inode(sb); 3417c9de560dSAlex Tomas if (sbi->s_buddy_cache == NULL) { 34189d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't get new inode"); 3419c9de560dSAlex Tomas goto err_freesgi; 3420c9de560dSAlex Tomas } 342148e6061bSYu Jian /* To avoid potentially colliding with an valid on-disk inode number, 342248e6061bSYu Jian * use EXT4_BAD_INO for the buddy cache inode number. This inode is 342348e6061bSYu Jian * not in the inode hash, so it should never be found by iget(), but 342448e6061bSYu Jian * this will avoid confusion if it ever shows up during debugging. */ 342548e6061bSYu Jian sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; 3426c9de560dSAlex Tomas EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 34278df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 34284b99faa2SKhazhismel Kumykov cond_resched(); 3429c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, i, NULL); 3430c9de560dSAlex Tomas if (desc == NULL) { 34319d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); 3432c9de560dSAlex Tomas goto err_freebuddy; 3433c9de560dSAlex Tomas } 34345f21b0e6SFrederic Bohe if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 34355f21b0e6SFrederic Bohe goto err_freebuddy; 3436c9de560dSAlex Tomas } 3437c9de560dSAlex Tomas 3438cfd73237SAlex Zhuravlev if (ext4_has_feature_flex_bg(sb)) { 3439f91436d5SSabyrzhan Tasbolatov /* a single flex group is supposed to be read by a single IO. 3440f91436d5SSabyrzhan Tasbolatov * 2 ^ s_log_groups_per_flex != UINT_MAX as s_mb_prefetch is 3441f91436d5SSabyrzhan Tasbolatov * unsigned integer, so the maximum shift is 32. 3442f91436d5SSabyrzhan Tasbolatov */ 3443f91436d5SSabyrzhan Tasbolatov if (sbi->s_es->s_log_groups_per_flex >= 32) { 3444f91436d5SSabyrzhan Tasbolatov ext4_msg(sb, KERN_ERR, "too many log groups per flexible block group"); 3445a8867f4eSPhillip Potter goto err_freebuddy; 3446f91436d5SSabyrzhan Tasbolatov } 3447f91436d5SSabyrzhan Tasbolatov sbi->s_mb_prefetch = min_t(uint, 1 << sbi->s_es->s_log_groups_per_flex, 344882ef1370SChunguang Xu BLK_MAX_SEGMENT_SIZE >> (sb->s_blocksize_bits - 9)); 3449cfd73237SAlex Zhuravlev sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */ 3450cfd73237SAlex Zhuravlev } else { 3451cfd73237SAlex Zhuravlev sbi->s_mb_prefetch = 32; 3452cfd73237SAlex Zhuravlev } 3453cfd73237SAlex Zhuravlev if (sbi->s_mb_prefetch > ext4_get_groups_count(sb)) 3454cfd73237SAlex Zhuravlev sbi->s_mb_prefetch = ext4_get_groups_count(sb); 34552caffb6aSKemeng Shi /* 34562caffb6aSKemeng Shi * now many real IOs to prefetch within a single allocation at 34572caffb6aSKemeng Shi * CR_POWER2_ALIGNED. Given CR_POWER2_ALIGNED is an CPU-related 34582caffb6aSKemeng Shi * optimization we shouldn't try to load too many groups, at some point 34592caffb6aSKemeng Shi * we should start to use what we've got in memory. 3460cfd73237SAlex Zhuravlev * with an average random access time 5ms, it'd take a second to get 3461cfd73237SAlex Zhuravlev * 200 groups (* N with flex_bg), so let's make this limit 4 3462cfd73237SAlex Zhuravlev */ 3463cfd73237SAlex Zhuravlev sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4; 3464cfd73237SAlex Zhuravlev if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb)) 3465cfd73237SAlex Zhuravlev sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb); 3466cfd73237SAlex Zhuravlev 3467c9de560dSAlex Tomas return 0; 3468c9de560dSAlex Tomas 3469c9de560dSAlex Tomas err_freebuddy: 3470fb1813f4SCurt Wohlgemuth cachep = get_groupinfo_cache(sb->s_blocksize_bits); 34715354b2afSTheodore Ts'o while (i-- > 0) { 34725354b2afSTheodore Ts'o struct ext4_group_info *grp = ext4_get_group_info(sb, i); 34735354b2afSTheodore Ts'o 34745354b2afSTheodore Ts'o if (grp) 34755354b2afSTheodore Ts'o kmem_cache_free(cachep, grp); 34765354b2afSTheodore Ts'o } 347728623c2fSTheodore Ts'o i = sbi->s_group_info_size; 3478df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3479df3da4eaSSuraj Jitindar Singh group_info = rcu_dereference(sbi->s_group_info); 3480f1fa3342SRoel Kluin while (i-- > 0) 3481df3da4eaSSuraj Jitindar Singh kfree(group_info[i]); 3482df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3483c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 3484c9de560dSAlex Tomas err_freesgi: 3485df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3486df3da4eaSSuraj Jitindar Singh kvfree(rcu_dereference(sbi->s_group_info)); 3487df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3488c9de560dSAlex Tomas return -ENOMEM; 3489c9de560dSAlex Tomas } 3490c9de560dSAlex Tomas 34912892c15dSEric Sandeen static void ext4_groupinfo_destroy_slabs(void) 34922892c15dSEric Sandeen { 34932892c15dSEric Sandeen int i; 34942892c15dSEric Sandeen 34952892c15dSEric Sandeen for (i = 0; i < NR_GRPINFO_CACHES; i++) { 34962892c15dSEric Sandeen kmem_cache_destroy(ext4_groupinfo_caches[i]); 34972892c15dSEric Sandeen ext4_groupinfo_caches[i] = NULL; 34982892c15dSEric Sandeen } 34992892c15dSEric Sandeen } 35002892c15dSEric Sandeen 35012892c15dSEric Sandeen static int ext4_groupinfo_create_slab(size_t size) 35022892c15dSEric Sandeen { 35032892c15dSEric Sandeen static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); 35042892c15dSEric Sandeen int slab_size; 35052892c15dSEric Sandeen int blocksize_bits = order_base_2(size); 35062892c15dSEric Sandeen int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 35072892c15dSEric Sandeen struct kmem_cache *cachep; 35082892c15dSEric Sandeen 35092892c15dSEric Sandeen if (cache_index >= NR_GRPINFO_CACHES) 35102892c15dSEric Sandeen return -EINVAL; 35112892c15dSEric Sandeen 35122892c15dSEric Sandeen if (unlikely(cache_index < 0)) 35132892c15dSEric Sandeen cache_index = 0; 35142892c15dSEric Sandeen 35152892c15dSEric Sandeen mutex_lock(&ext4_grpinfo_slab_create_mutex); 35162892c15dSEric Sandeen if (ext4_groupinfo_caches[cache_index]) { 35172892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 35182892c15dSEric Sandeen return 0; /* Already created */ 35192892c15dSEric Sandeen } 35202892c15dSEric Sandeen 35212892c15dSEric Sandeen slab_size = offsetof(struct ext4_group_info, 35222892c15dSEric Sandeen bb_counters[blocksize_bits + 2]); 35232892c15dSEric Sandeen 35242892c15dSEric Sandeen cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], 35252892c15dSEric Sandeen slab_size, 0, SLAB_RECLAIM_ACCOUNT, 35262892c15dSEric Sandeen NULL); 35272892c15dSEric Sandeen 3528823ba01fSTao Ma ext4_groupinfo_caches[cache_index] = cachep; 3529823ba01fSTao Ma 35302892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 35312892c15dSEric Sandeen if (!cachep) { 35329d8b9ec4STheodore Ts'o printk(KERN_EMERG 35339d8b9ec4STheodore Ts'o "EXT4-fs: no memory for groupinfo slab cache\n"); 35342892c15dSEric Sandeen return -ENOMEM; 35352892c15dSEric Sandeen } 35362892c15dSEric Sandeen 35372892c15dSEric Sandeen return 0; 35382892c15dSEric Sandeen } 35392892c15dSEric Sandeen 354055cdd0afSWang Jianchao static void ext4_discard_work(struct work_struct *work) 354155cdd0afSWang Jianchao { 354255cdd0afSWang Jianchao struct ext4_sb_info *sbi = container_of(work, 354355cdd0afSWang Jianchao struct ext4_sb_info, s_discard_work); 354455cdd0afSWang Jianchao struct super_block *sb = sbi->s_sb; 354555cdd0afSWang Jianchao struct ext4_free_data *fd, *nfd; 354655cdd0afSWang Jianchao struct ext4_buddy e4b; 35470f6bc579SRuan Jinjie LIST_HEAD(discard_list); 354855cdd0afSWang Jianchao ext4_group_t grp, load_grp; 354955cdd0afSWang Jianchao int err = 0; 355055cdd0afSWang Jianchao 355155cdd0afSWang Jianchao spin_lock(&sbi->s_md_lock); 355255cdd0afSWang Jianchao list_splice_init(&sbi->s_discard_list, &discard_list); 355355cdd0afSWang Jianchao spin_unlock(&sbi->s_md_lock); 355455cdd0afSWang Jianchao 355555cdd0afSWang Jianchao load_grp = UINT_MAX; 355655cdd0afSWang Jianchao list_for_each_entry_safe(fd, nfd, &discard_list, efd_list) { 355755cdd0afSWang Jianchao /* 35585036ab8dSWang Jianchao * If filesystem is umounting or no memory or suffering 35595036ab8dSWang Jianchao * from no space, give up the discard 356055cdd0afSWang Jianchao */ 35615036ab8dSWang Jianchao if ((sb->s_flags & SB_ACTIVE) && !err && 35625036ab8dSWang Jianchao !atomic_read(&sbi->s_retry_alloc_pending)) { 356355cdd0afSWang Jianchao grp = fd->efd_group; 356455cdd0afSWang Jianchao if (grp != load_grp) { 356555cdd0afSWang Jianchao if (load_grp != UINT_MAX) 356655cdd0afSWang Jianchao ext4_mb_unload_buddy(&e4b); 356755cdd0afSWang Jianchao 356855cdd0afSWang Jianchao err = ext4_mb_load_buddy(sb, grp, &e4b); 356955cdd0afSWang Jianchao if (err) { 357055cdd0afSWang Jianchao kmem_cache_free(ext4_free_data_cachep, fd); 357155cdd0afSWang Jianchao load_grp = UINT_MAX; 357255cdd0afSWang Jianchao continue; 357355cdd0afSWang Jianchao } else { 357455cdd0afSWang Jianchao load_grp = grp; 357555cdd0afSWang Jianchao } 357655cdd0afSWang Jianchao } 357755cdd0afSWang Jianchao 357855cdd0afSWang Jianchao ext4_lock_group(sb, grp); 357955cdd0afSWang Jianchao ext4_try_to_trim_range(sb, &e4b, fd->efd_start_cluster, 358055cdd0afSWang Jianchao fd->efd_start_cluster + fd->efd_count - 1, 1); 358155cdd0afSWang Jianchao ext4_unlock_group(sb, grp); 358255cdd0afSWang Jianchao } 358355cdd0afSWang Jianchao kmem_cache_free(ext4_free_data_cachep, fd); 358455cdd0afSWang Jianchao } 358555cdd0afSWang Jianchao 358655cdd0afSWang Jianchao if (load_grp != UINT_MAX) 358755cdd0afSWang Jianchao ext4_mb_unload_buddy(&e4b); 358855cdd0afSWang Jianchao } 358955cdd0afSWang Jianchao 35909d99012fSAkira Fujita int ext4_mb_init(struct super_block *sb) 3591c9de560dSAlex Tomas { 3592c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 35936be2ded1SAneesh Kumar K.V unsigned i, j; 3594935244cdSNicolai Stange unsigned offset, offset_incr; 3595c9de560dSAlex Tomas unsigned max; 359674767c5aSShen Feng int ret; 3597c9de560dSAlex Tomas 35984b68f6dfSHarshad Shirwadkar i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_offsets); 3599c9de560dSAlex Tomas 3600c9de560dSAlex Tomas sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 3601c9de560dSAlex Tomas if (sbi->s_mb_offsets == NULL) { 3602fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 3603fb1813f4SCurt Wohlgemuth goto out; 3604c9de560dSAlex Tomas } 3605ff7ef329SYasunori Goto 36064b68f6dfSHarshad Shirwadkar i = MB_NUM_ORDERS(sb) * sizeof(*sbi->s_mb_maxs); 3607c9de560dSAlex Tomas sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 3608c9de560dSAlex Tomas if (sbi->s_mb_maxs == NULL) { 3609fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 3610fb1813f4SCurt Wohlgemuth goto out; 3611fb1813f4SCurt Wohlgemuth } 3612fb1813f4SCurt Wohlgemuth 36132892c15dSEric Sandeen ret = ext4_groupinfo_create_slab(sb->s_blocksize); 36142892c15dSEric Sandeen if (ret < 0) 3615fb1813f4SCurt Wohlgemuth goto out; 3616c9de560dSAlex Tomas 3617c9de560dSAlex Tomas /* order 0 is regular bitmap */ 3618c9de560dSAlex Tomas sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 3619c9de560dSAlex Tomas sbi->s_mb_offsets[0] = 0; 3620c9de560dSAlex Tomas 3621c9de560dSAlex Tomas i = 1; 3622c9de560dSAlex Tomas offset = 0; 3623935244cdSNicolai Stange offset_incr = 1 << (sb->s_blocksize_bits - 1); 3624c9de560dSAlex Tomas max = sb->s_blocksize << 2; 3625c9de560dSAlex Tomas do { 3626c9de560dSAlex Tomas sbi->s_mb_offsets[i] = offset; 3627c9de560dSAlex Tomas sbi->s_mb_maxs[i] = max; 3628935244cdSNicolai Stange offset += offset_incr; 3629935244cdSNicolai Stange offset_incr = offset_incr >> 1; 3630c9de560dSAlex Tomas max = max >> 1; 3631c9de560dSAlex Tomas i++; 36324b68f6dfSHarshad Shirwadkar } while (i < MB_NUM_ORDERS(sb)); 36334b68f6dfSHarshad Shirwadkar 363483e80a6eSJan Kara sbi->s_mb_avg_fragment_size = 363583e80a6eSJan Kara kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), 363683e80a6eSJan Kara GFP_KERNEL); 363783e80a6eSJan Kara if (!sbi->s_mb_avg_fragment_size) { 363883e80a6eSJan Kara ret = -ENOMEM; 363983e80a6eSJan Kara goto out; 364083e80a6eSJan Kara } 364183e80a6eSJan Kara sbi->s_mb_avg_fragment_size_locks = 364283e80a6eSJan Kara kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), 364383e80a6eSJan Kara GFP_KERNEL); 364483e80a6eSJan Kara if (!sbi->s_mb_avg_fragment_size_locks) { 364583e80a6eSJan Kara ret = -ENOMEM; 364683e80a6eSJan Kara goto out; 364783e80a6eSJan Kara } 364883e80a6eSJan Kara for (i = 0; i < MB_NUM_ORDERS(sb); i++) { 364983e80a6eSJan Kara INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]); 365083e80a6eSJan Kara rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]); 365183e80a6eSJan Kara } 3652196e402aSHarshad Shirwadkar sbi->s_mb_largest_free_orders = 3653196e402aSHarshad Shirwadkar kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), 3654196e402aSHarshad Shirwadkar GFP_KERNEL); 3655196e402aSHarshad Shirwadkar if (!sbi->s_mb_largest_free_orders) { 3656196e402aSHarshad Shirwadkar ret = -ENOMEM; 3657196e402aSHarshad Shirwadkar goto out; 3658196e402aSHarshad Shirwadkar } 3659196e402aSHarshad Shirwadkar sbi->s_mb_largest_free_orders_locks = 3660196e402aSHarshad Shirwadkar kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), 3661196e402aSHarshad Shirwadkar GFP_KERNEL); 3662196e402aSHarshad Shirwadkar if (!sbi->s_mb_largest_free_orders_locks) { 3663196e402aSHarshad Shirwadkar ret = -ENOMEM; 3664196e402aSHarshad Shirwadkar goto out; 3665196e402aSHarshad Shirwadkar } 3666196e402aSHarshad Shirwadkar for (i = 0; i < MB_NUM_ORDERS(sb); i++) { 3667196e402aSHarshad Shirwadkar INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); 3668196e402aSHarshad Shirwadkar rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); 3669196e402aSHarshad Shirwadkar } 3670c9de560dSAlex Tomas 3671c9de560dSAlex Tomas spin_lock_init(&sbi->s_md_lock); 3672d08854f5STheodore Ts'o sbi->s_mb_free_pending = 0; 3673ce774e53SJinke Han INIT_LIST_HEAD(&sbi->s_freed_data_list[0]); 3674ce774e53SJinke Han INIT_LIST_HEAD(&sbi->s_freed_data_list[1]); 367555cdd0afSWang Jianchao INIT_LIST_HEAD(&sbi->s_discard_list); 367655cdd0afSWang Jianchao INIT_WORK(&sbi->s_discard_work, ext4_discard_work); 36775036ab8dSWang Jianchao atomic_set(&sbi->s_retry_alloc_pending, 0); 3678c9de560dSAlex Tomas 3679c9de560dSAlex Tomas sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 3680c9de560dSAlex Tomas sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 3681c9de560dSAlex Tomas sbi->s_mb_stats = MB_DEFAULT_STATS; 3682c9de560dSAlex Tomas sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 3683c9de560dSAlex Tomas sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 3684f52f3d2bSOjaswin Mujoo sbi->s_mb_best_avail_max_trim_order = MB_DEFAULT_BEST_AVAIL_TRIM_ORDER; 36857e170922SOjaswin Mujoo 368627baebb8STheodore Ts'o /* 368727baebb8STheodore Ts'o * The default group preallocation is 512, which for 4k block 368827baebb8STheodore Ts'o * sizes translates to 2 megabytes. However for bigalloc file 368927baebb8STheodore Ts'o * systems, this is probably too big (i.e, if the cluster size 369027baebb8STheodore Ts'o * is 1 megabyte, then group preallocation size becomes half a 369127baebb8STheodore Ts'o * gigabyte!). As a default, we will keep a two megabyte 369227baebb8STheodore Ts'o * group pralloc size for cluster sizes up to 64k, and after 369327baebb8STheodore Ts'o * that, we will force a minimum group preallocation size of 369427baebb8STheodore Ts'o * 32 clusters. This translates to 8 megs when the cluster 369527baebb8STheodore Ts'o * size is 256k, and 32 megs when the cluster size is 1 meg, 369627baebb8STheodore Ts'o * which seems reasonable as a default. 369727baebb8STheodore Ts'o */ 369827baebb8STheodore Ts'o sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> 369927baebb8STheodore Ts'o sbi->s_cluster_bits, 32); 3700d7a1fee1SDan Ehrenberg /* 3701d7a1fee1SDan Ehrenberg * If there is a s_stripe > 1, then we set the s_mb_group_prealloc 3702d7a1fee1SDan Ehrenberg * to the lowest multiple of s_stripe which is bigger than 3703d7a1fee1SDan Ehrenberg * the s_mb_group_prealloc as determined above. We want 3704d7a1fee1SDan Ehrenberg * the preallocation size to be an exact multiple of the 3705d7a1fee1SDan Ehrenberg * RAID stripe size so that preallocations don't fragment 3706d7a1fee1SDan Ehrenberg * the stripes. 3707d7a1fee1SDan Ehrenberg */ 3708d7a1fee1SDan Ehrenberg if (sbi->s_stripe > 1) { 3709d7a1fee1SDan Ehrenberg sbi->s_mb_group_prealloc = roundup( 3710c3defd99SKemeng Shi sbi->s_mb_group_prealloc, EXT4_B2C(sbi, sbi->s_stripe)); 3711d7a1fee1SDan Ehrenberg } 3712c9de560dSAlex Tomas 3713730c213cSEric Sandeen sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 3714c9de560dSAlex Tomas if (sbi->s_locality_groups == NULL) { 3715fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 3716029b10c5SAndrey Tsyvarev goto out; 3717c9de560dSAlex Tomas } 3718730c213cSEric Sandeen for_each_possible_cpu(i) { 3719c9de560dSAlex Tomas struct ext4_locality_group *lg; 3720730c213cSEric Sandeen lg = per_cpu_ptr(sbi->s_locality_groups, i); 3721c9de560dSAlex Tomas mutex_init(&lg->lg_mutex); 37226be2ded1SAneesh Kumar K.V for (j = 0; j < PREALLOC_TB_SIZE; j++) 37236be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 3724c9de560dSAlex Tomas spin_lock_init(&lg->lg_prealloc_lock); 3725c9de560dSAlex Tomas } 3726c9de560dSAlex Tomas 372710f0d2a5SChristoph Hellwig if (bdev_nonrot(sb->s_bdev)) 3728196e402aSHarshad Shirwadkar sbi->s_mb_max_linear_groups = 0; 3729196e402aSHarshad Shirwadkar else 3730196e402aSHarshad Shirwadkar sbi->s_mb_max_linear_groups = MB_DEFAULT_LINEAR_LIMIT; 373179a77c5aSYu Jian /* init file for buddy data */ 373279a77c5aSYu Jian ret = ext4_mb_init_backend(sb); 37337aa0baeaSTao Ma if (ret != 0) 37347aa0baeaSTao Ma goto out_free_locality_groups; 373579a77c5aSYu Jian 37367aa0baeaSTao Ma return 0; 37377aa0baeaSTao Ma 37387aa0baeaSTao Ma out_free_locality_groups: 37397aa0baeaSTao Ma free_percpu(sbi->s_locality_groups); 37407aa0baeaSTao Ma sbi->s_locality_groups = NULL; 3741fb1813f4SCurt Wohlgemuth out: 374283e80a6eSJan Kara kfree(sbi->s_mb_avg_fragment_size); 374383e80a6eSJan Kara kfree(sbi->s_mb_avg_fragment_size_locks); 3744196e402aSHarshad Shirwadkar kfree(sbi->s_mb_largest_free_orders); 3745196e402aSHarshad Shirwadkar kfree(sbi->s_mb_largest_free_orders_locks); 3746fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_offsets); 37477aa0baeaSTao Ma sbi->s_mb_offsets = NULL; 3748fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_maxs); 37497aa0baeaSTao Ma sbi->s_mb_maxs = NULL; 3750fb1813f4SCurt Wohlgemuth return ret; 3751c9de560dSAlex Tomas } 3752c9de560dSAlex Tomas 3753955ce5f5SAneesh Kumar K.V /* need to called with the ext4 group lock held */ 3754d3df1453SRitesh Harjani static int ext4_mb_cleanup_pa(struct ext4_group_info *grp) 3755c9de560dSAlex Tomas { 3756c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3757c9de560dSAlex Tomas struct list_head *cur, *tmp; 3758c9de560dSAlex Tomas int count = 0; 3759c9de560dSAlex Tomas 3760c9de560dSAlex Tomas list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 3761c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3762c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3763c9de560dSAlex Tomas count++; 3764688f05a0SAneesh Kumar K.V kmem_cache_free(ext4_pspace_cachep, pa); 3765c9de560dSAlex Tomas } 3766d3df1453SRitesh Harjani return count; 3767c9de560dSAlex Tomas } 3768c9de560dSAlex Tomas 376990817717SKemeng Shi void ext4_mb_release(struct super_block *sb) 3770c9de560dSAlex Tomas { 37718df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 3772c9de560dSAlex Tomas ext4_group_t i; 3773c9de560dSAlex Tomas int num_meta_group_infos; 3774df3da4eaSSuraj Jitindar Singh struct ext4_group_info *grinfo, ***group_info; 3775c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3776fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 3777d3df1453SRitesh Harjani int count; 3778c9de560dSAlex Tomas 377955cdd0afSWang Jianchao if (test_opt(sb, DISCARD)) { 378055cdd0afSWang Jianchao /* 378155cdd0afSWang Jianchao * wait the discard work to drain all of ext4_free_data 378255cdd0afSWang Jianchao */ 378355cdd0afSWang Jianchao flush_work(&sbi->s_discard_work); 378455cdd0afSWang Jianchao WARN_ON_ONCE(!list_empty(&sbi->s_discard_list)); 378555cdd0afSWang Jianchao } 378655cdd0afSWang Jianchao 3787c9de560dSAlex Tomas if (sbi->s_group_info) { 37888df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 37894b99faa2SKhazhismel Kumykov cond_resched(); 3790c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, i); 37915354b2afSTheodore Ts'o if (!grinfo) 37925354b2afSTheodore Ts'o continue; 3793a3450215SRitesh Harjani mb_group_bb_bitmap_free(grinfo); 3794c9de560dSAlex Tomas ext4_lock_group(sb, i); 3795d3df1453SRitesh Harjani count = ext4_mb_cleanup_pa(grinfo); 3796d3df1453SRitesh Harjani if (count) 3797d3df1453SRitesh Harjani mb_debug(sb, "mballoc: %d PAs left\n", 3798d3df1453SRitesh Harjani count); 3799c9de560dSAlex Tomas ext4_unlock_group(sb, i); 3800fb1813f4SCurt Wohlgemuth kmem_cache_free(cachep, grinfo); 3801c9de560dSAlex Tomas } 38028df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + 3803c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK(sb) - 1) >> 3804c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK_BITS(sb); 3805df3da4eaSSuraj Jitindar Singh rcu_read_lock(); 3806df3da4eaSSuraj Jitindar Singh group_info = rcu_dereference(sbi->s_group_info); 3807c9de560dSAlex Tomas for (i = 0; i < num_meta_group_infos; i++) 3808df3da4eaSSuraj Jitindar Singh kfree(group_info[i]); 3809df3da4eaSSuraj Jitindar Singh kvfree(group_info); 3810df3da4eaSSuraj Jitindar Singh rcu_read_unlock(); 3811c9de560dSAlex Tomas } 381283e80a6eSJan Kara kfree(sbi->s_mb_avg_fragment_size); 381383e80a6eSJan Kara kfree(sbi->s_mb_avg_fragment_size_locks); 3814196e402aSHarshad Shirwadkar kfree(sbi->s_mb_largest_free_orders); 3815196e402aSHarshad Shirwadkar kfree(sbi->s_mb_largest_free_orders_locks); 3816c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 3817c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 3818c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 3819c9de560dSAlex Tomas if (sbi->s_mb_stats) { 38209d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 38219d8b9ec4STheodore Ts'o "mballoc: %u blocks %u reqs (%u success)", 3822c9de560dSAlex Tomas atomic_read(&sbi->s_bal_allocated), 3823c9de560dSAlex Tomas atomic_read(&sbi->s_bal_reqs), 3824c9de560dSAlex Tomas atomic_read(&sbi->s_bal_success)); 38259d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 3826a6c75eafSHarshad Shirwadkar "mballoc: %u extents scanned, %u groups scanned, %u goal hits, " 38279d8b9ec4STheodore Ts'o "%u 2^N hits, %u breaks, %u lost", 3828c9de560dSAlex Tomas atomic_read(&sbi->s_bal_ex_scanned), 3829a6c75eafSHarshad Shirwadkar atomic_read(&sbi->s_bal_groups_scanned), 3830c9de560dSAlex Tomas atomic_read(&sbi->s_bal_goals), 3831c9de560dSAlex Tomas atomic_read(&sbi->s_bal_2orders), 3832c9de560dSAlex Tomas atomic_read(&sbi->s_bal_breaks), 3833c9de560dSAlex Tomas atomic_read(&sbi->s_mb_lost_chunks)); 38349d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 383567d25186SHarshad Shirwadkar "mballoc: %u generated and it took %llu", 383667d25186SHarshad Shirwadkar atomic_read(&sbi->s_mb_buddies_generated), 383767d25186SHarshad Shirwadkar atomic64_read(&sbi->s_mb_generation_time)); 38389d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 38399d8b9ec4STheodore Ts'o "mballoc: %u preallocated, %u discarded", 3840c9de560dSAlex Tomas atomic_read(&sbi->s_mb_preallocated), 3841c9de560dSAlex Tomas atomic_read(&sbi->s_mb_discarded)); 3842c9de560dSAlex Tomas } 3843c9de560dSAlex Tomas 3844730c213cSEric Sandeen free_percpu(sbi->s_locality_groups); 3845c9de560dSAlex Tomas } 3846c9de560dSAlex Tomas 384777ca6cdfSLukas Czerner static inline int ext4_issue_discard(struct super_block *sb, 38480efcd739SWenchao Hao ext4_group_t block_group, ext4_grpblk_t cluster, int count) 38495c521830SJiaying Zhang { 38505c521830SJiaying Zhang ext4_fsblk_t discard_block; 38515c521830SJiaying Zhang 385284130193STheodore Ts'o discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + 385384130193STheodore Ts'o ext4_group_first_block_no(sb, block_group)); 385484130193STheodore Ts'o count = EXT4_C2B(EXT4_SB(sb), count); 38555c521830SJiaying Zhang trace_ext4_discard_blocks(sb, 38565c521830SJiaying Zhang (unsigned long long) discard_block, count); 38570efcd739SWenchao Hao 385893259636SLukas Czerner return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); 38595c521830SJiaying Zhang } 38605c521830SJiaying Zhang 3861a0154344SDaeho Jeong static void ext4_free_data_in_buddy(struct super_block *sb, 3862a0154344SDaeho Jeong struct ext4_free_data *entry) 3863c9de560dSAlex Tomas { 3864c9de560dSAlex Tomas struct ext4_buddy e4b; 3865c894058dSAneesh Kumar K.V struct ext4_group_info *db; 3866c7f2bafaSKemeng Shi int err, count = 0; 3867c9de560dSAlex Tomas 3868d3df1453SRitesh Harjani mb_debug(sb, "gonna free %u blocks in group %u (0x%p):", 386918aadd47SBobi Jam entry->efd_count, entry->efd_group, entry); 3870c9de560dSAlex Tomas 387118aadd47SBobi Jam err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); 3872c9de560dSAlex Tomas /* we expect to find existing buddy because it's pinned */ 3873c9de560dSAlex Tomas BUG_ON(err != 0); 3874c9de560dSAlex Tomas 3875d08854f5STheodore Ts'o spin_lock(&EXT4_SB(sb)->s_md_lock); 3876d08854f5STheodore Ts'o EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; 3877d08854f5STheodore Ts'o spin_unlock(&EXT4_SB(sb)->s_md_lock); 387818aadd47SBobi Jam 3879c894058dSAneesh Kumar K.V db = e4b.bd_info; 3880c9de560dSAlex Tomas /* there are blocks to put in buddy to make them really free */ 388118aadd47SBobi Jam count += entry->efd_count; 388218aadd47SBobi Jam ext4_lock_group(sb, entry->efd_group); 3883c894058dSAneesh Kumar K.V /* Take it out of per group rb tree */ 388418aadd47SBobi Jam rb_erase(&entry->efd_node, &(db->bb_free_root)); 388518aadd47SBobi Jam mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); 3886c9de560dSAlex Tomas 38873d56b8d2STao Ma /* 38883d56b8d2STao Ma * Clear the trimmed flag for the group so that the next 38893d56b8d2STao Ma * ext4_trim_fs can trim it. 38903d56b8d2STao Ma * If the volume is mounted with -o discard, online discard 38913d56b8d2STao Ma * is supported and the free blocks will be trimmed online. 38923d56b8d2STao Ma */ 38933d56b8d2STao Ma if (!test_opt(sb, DISCARD)) 38943d56b8d2STao Ma EXT4_MB_GRP_CLEAR_TRIMMED(db); 38953d56b8d2STao Ma 3896c894058dSAneesh Kumar K.V if (!db->bb_free_root.rb_node) { 3897c894058dSAneesh Kumar K.V /* No more items in the per group rb tree 3898c894058dSAneesh Kumar K.V * balance refcounts from ext4_mb_free_metadata() 3899c894058dSAneesh Kumar K.V */ 39005eea586bSMatthew Wilcox (Oracle) folio_put(e4b.bd_buddy_folio); 390199b150d8SMatthew Wilcox (Oracle) folio_put(e4b.bd_bitmap_folio); 3902c894058dSAneesh Kumar K.V } 390318aadd47SBobi Jam ext4_unlock_group(sb, entry->efd_group); 3904e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3905c9de560dSAlex Tomas 3906c7f2bafaSKemeng Shi mb_debug(sb, "freed %d blocks in 1 structures\n", count); 3907c9de560dSAlex Tomas } 3908c9de560dSAlex Tomas 3909a0154344SDaeho Jeong /* 3910a0154344SDaeho Jeong * This function is called by the jbd2 layer once the commit has finished, 3911a0154344SDaeho Jeong * so we know we can free the blocks that were released with that commit. 3912a0154344SDaeho Jeong */ 3913a0154344SDaeho Jeong void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid) 3914a0154344SDaeho Jeong { 3915a0154344SDaeho Jeong struct ext4_sb_info *sbi = EXT4_SB(sb); 3916a0154344SDaeho Jeong struct ext4_free_data *entry, *tmp; 39170f6bc579SRuan Jinjie LIST_HEAD(freed_data_list); 3918ce774e53SJinke Han struct list_head *s_freed_head = &sbi->s_freed_data_list[commit_tid & 1]; 391955cdd0afSWang Jianchao bool wake; 3920a0154344SDaeho Jeong 3921ce774e53SJinke Han list_replace_init(s_freed_head, &freed_data_list); 3922a0154344SDaeho Jeong 392355cdd0afSWang Jianchao list_for_each_entry(entry, &freed_data_list, efd_list) 3924a0154344SDaeho Jeong ext4_free_data_in_buddy(sb, entry); 392555cdd0afSWang Jianchao 392655cdd0afSWang Jianchao if (test_opt(sb, DISCARD)) { 392755cdd0afSWang Jianchao spin_lock(&sbi->s_md_lock); 392855cdd0afSWang Jianchao wake = list_empty(&sbi->s_discard_list); 392955cdd0afSWang Jianchao list_splice_tail(&freed_data_list, &sbi->s_discard_list); 393055cdd0afSWang Jianchao spin_unlock(&sbi->s_md_lock); 393155cdd0afSWang Jianchao if (wake) 393255cdd0afSWang Jianchao queue_work(system_unbound_wq, &sbi->s_discard_work); 393355cdd0afSWang Jianchao } else { 393455cdd0afSWang Jianchao list_for_each_entry_safe(entry, tmp, &freed_data_list, efd_list) 393555cdd0afSWang Jianchao kmem_cache_free(ext4_free_data_cachep, entry); 393655cdd0afSWang Jianchao } 3937a0154344SDaeho Jeong } 3938a0154344SDaeho Jeong 39395dabfc78STheodore Ts'o int __init ext4_init_mballoc(void) 3940c9de560dSAlex Tomas { 394116828088STheodore Ts'o ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, 394216828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 3943c9de560dSAlex Tomas if (ext4_pspace_cachep == NULL) 3944f283529aSRitesh Harjani goto out; 3945c9de560dSAlex Tomas 394616828088STheodore Ts'o ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, 394716828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 3948f283529aSRitesh Harjani if (ext4_ac_cachep == NULL) 3949f283529aSRitesh Harjani goto out_pa_free; 3950c894058dSAneesh Kumar K.V 395118aadd47SBobi Jam ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, 395216828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 3953f283529aSRitesh Harjani if (ext4_free_data_cachep == NULL) 3954f283529aSRitesh Harjani goto out_ac_free; 3955f283529aSRitesh Harjani 3956c9de560dSAlex Tomas return 0; 3957f283529aSRitesh Harjani 3958f283529aSRitesh Harjani out_ac_free: 3959f283529aSRitesh Harjani kmem_cache_destroy(ext4_ac_cachep); 3960f283529aSRitesh Harjani out_pa_free: 3961f283529aSRitesh Harjani kmem_cache_destroy(ext4_pspace_cachep); 3962f283529aSRitesh Harjani out: 3963f283529aSRitesh Harjani return -ENOMEM; 3964c9de560dSAlex Tomas } 3965c9de560dSAlex Tomas 39665dabfc78STheodore Ts'o void ext4_exit_mballoc(void) 3967c9de560dSAlex Tomas { 39683e03f9caSJesper Dangaard Brouer /* 39693e03f9caSJesper Dangaard Brouer * Wait for completion of call_rcu()'s on ext4_pspace_cachep 39703e03f9caSJesper Dangaard Brouer * before destroying the slab cache. 39713e03f9caSJesper Dangaard Brouer */ 39723e03f9caSJesper Dangaard Brouer rcu_barrier(); 3973c9de560dSAlex Tomas kmem_cache_destroy(ext4_pspace_cachep); 3974256bdb49SEric Sandeen kmem_cache_destroy(ext4_ac_cachep); 397518aadd47SBobi Jam kmem_cache_destroy(ext4_free_data_cachep); 39762892c15dSEric Sandeen ext4_groupinfo_destroy_slabs(); 3977c9de560dSAlex Tomas } 3978c9de560dSAlex Tomas 3979c431d386SKemeng Shi #define EXT4_MB_BITMAP_MARKED_CHECK 0x0001 3980c431d386SKemeng Shi #define EXT4_MB_SYNC_UPDATE 0x0002 3981f9e2d95aSKemeng Shi static int 3982c431d386SKemeng Shi ext4_mb_mark_context(handle_t *handle, struct super_block *sb, bool state, 3983c431d386SKemeng Shi ext4_group_t group, ext4_grpblk_t blkoff, 3984c431d386SKemeng Shi ext4_grpblk_t len, int flags, ext4_grpblk_t *ret_changed) 3985f9e2d95aSKemeng Shi { 3986f9e2d95aSKemeng Shi struct ext4_sb_info *sbi = EXT4_SB(sb); 3987f9e2d95aSKemeng Shi struct buffer_head *bitmap_bh = NULL; 3988f9e2d95aSKemeng Shi struct ext4_group_desc *gdp; 3989f9e2d95aSKemeng Shi struct buffer_head *gdp_bh; 3990f9e2d95aSKemeng Shi int err; 3991c431d386SKemeng Shi unsigned int i, already, changed = len; 3992f9e2d95aSKemeng Shi 3993bdefd689SKemeng Shi KUNIT_STATIC_STUB_REDIRECT(ext4_mb_mark_context, 3994bdefd689SKemeng Shi handle, sb, state, group, blkoff, len, 3995bdefd689SKemeng Shi flags, ret_changed); 3996bdefd689SKemeng Shi 3997c431d386SKemeng Shi if (ret_changed) 3998c431d386SKemeng Shi *ret_changed = 0; 3999f9e2d95aSKemeng Shi bitmap_bh = ext4_read_block_bitmap(sb, group); 4000f9e2d95aSKemeng Shi if (IS_ERR(bitmap_bh)) 4001f9e2d95aSKemeng Shi return PTR_ERR(bitmap_bh); 4002f9e2d95aSKemeng Shi 4003c431d386SKemeng Shi if (handle) { 4004c431d386SKemeng Shi BUFFER_TRACE(bitmap_bh, "getting write access"); 4005c431d386SKemeng Shi err = ext4_journal_get_write_access(handle, sb, bitmap_bh, 4006c431d386SKemeng Shi EXT4_JTR_NONE); 4007c431d386SKemeng Shi if (err) 4008c431d386SKemeng Shi goto out_err; 4009c431d386SKemeng Shi } 4010c431d386SKemeng Shi 4011f9e2d95aSKemeng Shi err = -EIO; 4012f9e2d95aSKemeng Shi gdp = ext4_get_group_desc(sb, group, &gdp_bh); 4013f9e2d95aSKemeng Shi if (!gdp) 4014f9e2d95aSKemeng Shi goto out_err; 4015f9e2d95aSKemeng Shi 4016c431d386SKemeng Shi if (handle) { 4017c431d386SKemeng Shi BUFFER_TRACE(gdp_bh, "get_write_access"); 4018c431d386SKemeng Shi err = ext4_journal_get_write_access(handle, sb, gdp_bh, 4019c431d386SKemeng Shi EXT4_JTR_NONE); 4020c431d386SKemeng Shi if (err) 4021c431d386SKemeng Shi goto out_err; 4022c431d386SKemeng Shi } 4023c431d386SKemeng Shi 4024f9e2d95aSKemeng Shi ext4_lock_group(sb, group); 4025f9e2d95aSKemeng Shi if (ext4_has_group_desc_csum(sb) && 4026f9e2d95aSKemeng Shi (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) { 4027f9e2d95aSKemeng Shi gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 4028f9e2d95aSKemeng Shi ext4_free_group_clusters_set(sb, gdp, 4029f9e2d95aSKemeng Shi ext4_free_clusters_after_init(sb, group, gdp)); 4030f9e2d95aSKemeng Shi } 4031f9e2d95aSKemeng Shi 4032c431d386SKemeng Shi if (flags & EXT4_MB_BITMAP_MARKED_CHECK) { 4033f9e2d95aSKemeng Shi already = 0; 4034f9e2d95aSKemeng Shi for (i = 0; i < len; i++) 4035f9e2d95aSKemeng Shi if (mb_test_bit(blkoff + i, bitmap_bh->b_data) == 4036f9e2d95aSKemeng Shi state) 4037f9e2d95aSKemeng Shi already++; 4038f9e2d95aSKemeng Shi changed = len - already; 4039c431d386SKemeng Shi } 4040f9e2d95aSKemeng Shi 4041f9e2d95aSKemeng Shi if (state) { 4042f9e2d95aSKemeng Shi mb_set_bits(bitmap_bh->b_data, blkoff, len); 4043f9e2d95aSKemeng Shi ext4_free_group_clusters_set(sb, gdp, 4044f9e2d95aSKemeng Shi ext4_free_group_clusters(sb, gdp) - changed); 4045f9e2d95aSKemeng Shi } else { 4046f9e2d95aSKemeng Shi mb_clear_bits(bitmap_bh->b_data, blkoff, len); 4047f9e2d95aSKemeng Shi ext4_free_group_clusters_set(sb, gdp, 4048f9e2d95aSKemeng Shi ext4_free_group_clusters(sb, gdp) + changed); 4049f9e2d95aSKemeng Shi } 4050f9e2d95aSKemeng Shi 4051f9e2d95aSKemeng Shi ext4_block_bitmap_csum_set(sb, gdp, bitmap_bh); 4052f9e2d95aSKemeng Shi ext4_group_desc_csum_set(sb, group, gdp); 4053f9e2d95aSKemeng Shi ext4_unlock_group(sb, group); 4054c431d386SKemeng Shi if (ret_changed) 4055c431d386SKemeng Shi *ret_changed = changed; 4056f9e2d95aSKemeng Shi 4057f9e2d95aSKemeng Shi if (sbi->s_log_groups_per_flex) { 4058f9e2d95aSKemeng Shi ext4_group_t flex_group = ext4_flex_group(sbi, group); 4059f9e2d95aSKemeng Shi struct flex_groups *fg = sbi_array_rcu_deref(sbi, 4060f9e2d95aSKemeng Shi s_flex_groups, flex_group); 4061f9e2d95aSKemeng Shi 4062f9e2d95aSKemeng Shi if (state) 4063f9e2d95aSKemeng Shi atomic64_sub(changed, &fg->free_clusters); 4064f9e2d95aSKemeng Shi else 4065f9e2d95aSKemeng Shi atomic64_add(changed, &fg->free_clusters); 4066f9e2d95aSKemeng Shi } 4067f9e2d95aSKemeng Shi 4068c431d386SKemeng Shi err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 4069f9e2d95aSKemeng Shi if (err) 4070f9e2d95aSKemeng Shi goto out_err; 4071c431d386SKemeng Shi err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 4072f9e2d95aSKemeng Shi if (err) 4073f9e2d95aSKemeng Shi goto out_err; 4074f9e2d95aSKemeng Shi 4075c431d386SKemeng Shi if (flags & EXT4_MB_SYNC_UPDATE) { 4076f9e2d95aSKemeng Shi sync_dirty_buffer(bitmap_bh); 4077f9e2d95aSKemeng Shi sync_dirty_buffer(gdp_bh); 4078c431d386SKemeng Shi } 4079f9e2d95aSKemeng Shi 4080f9e2d95aSKemeng Shi out_err: 4081f9e2d95aSKemeng Shi brelse(bitmap_bh); 4082f9e2d95aSKemeng Shi return err; 4083f9e2d95aSKemeng Shi } 4084c9de560dSAlex Tomas 4085c9de560dSAlex Tomas /* 408673b2c716SUwe Kleine-König * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps 4087c9de560dSAlex Tomas * Returns 0 if success or error code 4088c9de560dSAlex Tomas */ 40894ddfef7bSEric Sandeen static noinline_for_stack int 40904ddfef7bSEric Sandeen ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 409153accfa9STheodore Ts'o handle_t *handle, unsigned int reserv_clstrs) 4092c9de560dSAlex Tomas { 4093c9de560dSAlex Tomas struct ext4_group_desc *gdp; 4094c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4095c9de560dSAlex Tomas struct super_block *sb; 4096c9de560dSAlex Tomas ext4_fsblk_t block; 4097519deca0SAneesh Kumar K.V int err, len; 40982f94711bSKemeng Shi int flags = 0; 40992f94711bSKemeng Shi ext4_grpblk_t changed; 4100c9de560dSAlex Tomas 4101c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 4102c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len <= 0); 4103c9de560dSAlex Tomas 4104c9de560dSAlex Tomas sb = ac->ac_sb; 4105c9de560dSAlex Tomas sbi = EXT4_SB(sb); 4106c9de560dSAlex Tomas 41072f94711bSKemeng Shi gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, NULL); 4108c9de560dSAlex Tomas if (!gdp) 41092f94711bSKemeng Shi return -EIO; 4110a9df9a49STheodore Ts'o ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 4111021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, gdp)); 411203cddb80SAneesh Kumar K.V 4113bda00de7SAkinobu Mita block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 411453accfa9STheodore Ts'o len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 4115ce9f24ccSJan Kara if (!ext4_inode_block_valid(ac->ac_inode, block, len)) { 411612062dddSEric Sandeen ext4_error(sb, "Allocating blocks %llu-%llu which overlap " 41171084f252STheodore Ts'o "fs metadata", block, block+len); 4118519deca0SAneesh Kumar K.V /* File system mounted not to panic on error 4119554a5cccSVegard Nossum * Fix the bitmap and return EFSCORRUPTED 4120519deca0SAneesh Kumar K.V * We leak some of the blocks here. 4121519deca0SAneesh Kumar K.V */ 41222f94711bSKemeng Shi err = ext4_mb_mark_context(handle, sb, true, 41232f94711bSKemeng Shi ac->ac_b_ex.fe_group, 41242f94711bSKemeng Shi ac->ac_b_ex.fe_start, 41252f94711bSKemeng Shi ac->ac_b_ex.fe_len, 41262f94711bSKemeng Shi 0, NULL); 4127519deca0SAneesh Kumar K.V if (!err) 4128554a5cccSVegard Nossum err = -EFSCORRUPTED; 41292f94711bSKemeng Shi return err; 4130c9de560dSAlex Tomas } 4131955ce5f5SAneesh Kumar K.V 4132c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 41332f94711bSKemeng Shi flags |= EXT4_MB_BITMAP_MARKED_CHECK; 4134c9de560dSAlex Tomas #endif 41352f94711bSKemeng Shi err = ext4_mb_mark_context(handle, sb, true, ac->ac_b_ex.fe_group, 41362f94711bSKemeng Shi ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len, 41372f94711bSKemeng Shi flags, &changed); 4138955ce5f5SAneesh Kumar K.V 41392f94711bSKemeng Shi if (err && changed == 0) 41402f94711bSKemeng Shi return err; 41412f94711bSKemeng Shi 41422f94711bSKemeng Shi #ifdef AGGRESSIVE_CHECK 41432f94711bSKemeng Shi BUG_ON(changed != ac->ac_b_ex.fe_len); 41442f94711bSKemeng Shi #endif 414557042651STheodore Ts'o percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); 4146d2a17637SMingming Cao /* 41476bc6e63fSAneesh Kumar K.V * Now reduce the dirty block count also. Should not go negative 4148d2a17637SMingming Cao */ 41496bc6e63fSAneesh Kumar K.V if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 41506bc6e63fSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 415157042651STheodore Ts'o percpu_counter_sub(&sbi->s_dirtyclusters_counter, 415257042651STheodore Ts'o reserv_clstrs); 4153c9de560dSAlex Tomas 4154c9de560dSAlex Tomas return err; 4155c9de560dSAlex Tomas } 4156c9de560dSAlex Tomas 4157c9de560dSAlex Tomas /* 41588016e29fSHarshad Shirwadkar * Idempotent helper for Ext4 fast commit replay path to set the state of 41598016e29fSHarshad Shirwadkar * blocks in bitmaps and update counters. 41608016e29fSHarshad Shirwadkar */ 41618016e29fSHarshad Shirwadkar void ext4_mb_mark_bb(struct super_block *sb, ext4_fsblk_t block, 4162d2f7cf40SKemeng Shi int len, bool state) 41638016e29fSHarshad Shirwadkar { 41648016e29fSHarshad Shirwadkar struct ext4_sb_info *sbi = EXT4_SB(sb); 41658016e29fSHarshad Shirwadkar ext4_group_t group; 41668016e29fSHarshad Shirwadkar ext4_grpblk_t blkoff; 4167f9e2d95aSKemeng Shi int err = 0; 4168f9e2d95aSKemeng Shi unsigned int clen, thisgrp_len; 41698016e29fSHarshad Shirwadkar 4170bfdc502aSRitesh Harjani while (len > 0) { 41718016e29fSHarshad Shirwadkar ext4_get_group_no_and_offset(sb, block, &group, &blkoff); 4172bfdc502aSRitesh Harjani 4173bfdc502aSRitesh Harjani /* 4174bfdc502aSRitesh Harjani * Check to see if we are freeing blocks across a group 4175bfdc502aSRitesh Harjani * boundary. 4176bfdc502aSRitesh Harjani * In case of flex_bg, this can happen that (block, len) may 4177bfdc502aSRitesh Harjani * span across more than one group. In that case we need to 4178bfdc502aSRitesh Harjani * get the corresponding group metadata to work with. 4179bfdc502aSRitesh Harjani * For this we have goto again loop. 4180bfdc502aSRitesh Harjani */ 4181bfdc502aSRitesh Harjani thisgrp_len = min_t(unsigned int, (unsigned int)len, 4182bfdc502aSRitesh Harjani EXT4_BLOCKS_PER_GROUP(sb) - EXT4_C2B(sbi, blkoff)); 4183bfdc502aSRitesh Harjani clen = EXT4_NUM_B2C(sbi, thisgrp_len); 4184bfdc502aSRitesh Harjani 41858c91c579SRitesh Harjani if (!ext4_sb_block_valid(sb, NULL, block, thisgrp_len)) { 41868c91c579SRitesh Harjani ext4_error(sb, "Marking blocks in system zone - " 41878c91c579SRitesh Harjani "Block = %llu, len = %u", 41888c91c579SRitesh Harjani block, thisgrp_len); 41898c91c579SRitesh Harjani break; 41908c91c579SRitesh Harjani } 41918c91c579SRitesh Harjani 4192c431d386SKemeng Shi err = ext4_mb_mark_context(NULL, sb, state, 4193c431d386SKemeng Shi group, blkoff, clen, 4194c431d386SKemeng Shi EXT4_MB_BITMAP_MARKED_CHECK | 4195c431d386SKemeng Shi EXT4_MB_SYNC_UPDATE, 4196c431d386SKemeng Shi NULL); 4197bfdc502aSRitesh Harjani if (err) 4198bfdc502aSRitesh Harjani break; 41998016e29fSHarshad Shirwadkar 4200bfdc502aSRitesh Harjani block += thisgrp_len; 4201bfdc502aSRitesh Harjani len -= thisgrp_len; 4202bfdc502aSRitesh Harjani BUG_ON(len < 0); 4203bfdc502aSRitesh Harjani } 42048016e29fSHarshad Shirwadkar } 42058016e29fSHarshad Shirwadkar 42068016e29fSHarshad Shirwadkar /* 4207c9de560dSAlex Tomas * here we normalize request for locality group 4208d7a1fee1SDan Ehrenberg * Group request are normalized to s_mb_group_prealloc, which goes to 4209d7a1fee1SDan Ehrenberg * s_strip if we set the same via mount option. 4210d7a1fee1SDan Ehrenberg * s_mb_group_prealloc can be configured via 4211b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_group_prealloc 4212c9de560dSAlex Tomas * 4213c9de560dSAlex Tomas * XXX: should we try to preallocate more than the group has now? 4214c9de560dSAlex Tomas */ 4215c9de560dSAlex Tomas static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 4216c9de560dSAlex Tomas { 4217c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 4218c9de560dSAlex Tomas struct ext4_locality_group *lg = ac->ac_lg; 4219c9de560dSAlex Tomas 4220c9de560dSAlex Tomas BUG_ON(lg == NULL); 4221c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 4222d3df1453SRitesh Harjani mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len); 4223c9de560dSAlex Tomas } 4224c9de560dSAlex Tomas 422538727786SOjaswin Mujoo /* 422638727786SOjaswin Mujoo * This function returns the next element to look at during inode 422738727786SOjaswin Mujoo * PA rbtree walk. We assume that we have held the inode PA rbtree lock 422838727786SOjaswin Mujoo * (ei->i_prealloc_lock) 422938727786SOjaswin Mujoo * 423038727786SOjaswin Mujoo * new_start The start of the range we want to compare 423138727786SOjaswin Mujoo * cur_start The existing start that we are comparing against 423238727786SOjaswin Mujoo * node The node of the rb_tree 423338727786SOjaswin Mujoo */ 423438727786SOjaswin Mujoo static inline struct rb_node* 423538727786SOjaswin Mujoo ext4_mb_pa_rb_next_iter(ext4_lblk_t new_start, ext4_lblk_t cur_start, struct rb_node *node) 423638727786SOjaswin Mujoo { 423738727786SOjaswin Mujoo if (new_start < cur_start) 423838727786SOjaswin Mujoo return node->rb_left; 423938727786SOjaswin Mujoo else 424038727786SOjaswin Mujoo return node->rb_right; 424138727786SOjaswin Mujoo } 424238727786SOjaswin Mujoo 42437692094aSOjaswin Mujoo static inline void 42447692094aSOjaswin Mujoo ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac, 4245bedc5d34SBaokun Li ext4_lblk_t start, loff_t end) 42467692094aSOjaswin Mujoo { 42477692094aSOjaswin Mujoo struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 42487692094aSOjaswin Mujoo struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 42497692094aSOjaswin Mujoo struct ext4_prealloc_space *tmp_pa; 4250bedc5d34SBaokun Li ext4_lblk_t tmp_pa_start; 4251bedc5d34SBaokun Li loff_t tmp_pa_end; 425238727786SOjaswin Mujoo struct rb_node *iter; 42537692094aSOjaswin Mujoo 425438727786SOjaswin Mujoo read_lock(&ei->i_prealloc_lock); 425538727786SOjaswin Mujoo for (iter = ei->i_prealloc_node.rb_node; iter; 425638727786SOjaswin Mujoo iter = ext4_mb_pa_rb_next_iter(start, tmp_pa_start, iter)) { 425738727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 425838727786SOjaswin Mujoo pa_node.inode_node); 42597692094aSOjaswin Mujoo tmp_pa_start = tmp_pa->pa_lstart; 4260bedc5d34SBaokun Li tmp_pa_end = pa_logical_end(sbi, tmp_pa); 42617692094aSOjaswin Mujoo 426238727786SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 426338727786SOjaswin Mujoo if (tmp_pa->pa_deleted == 0) 42647692094aSOjaswin Mujoo BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start)); 42657692094aSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 42667692094aSOjaswin Mujoo } 426738727786SOjaswin Mujoo read_unlock(&ei->i_prealloc_lock); 42687692094aSOjaswin Mujoo } 42697692094aSOjaswin Mujoo 4270c9de560dSAlex Tomas /* 42710830344cSOjaswin Mujoo * Given an allocation context "ac" and a range "start", "end", check 42720830344cSOjaswin Mujoo * and adjust boundaries if the range overlaps with any of the existing 42730830344cSOjaswin Mujoo * preallocatoins stored in the corresponding inode of the allocation context. 42740830344cSOjaswin Mujoo * 42750830344cSOjaswin Mujoo * Parameters: 42760830344cSOjaswin Mujoo * ac allocation context 42770830344cSOjaswin Mujoo * start start of the new range 42780830344cSOjaswin Mujoo * end end of the new range 42790830344cSOjaswin Mujoo */ 42800830344cSOjaswin Mujoo static inline void 42810830344cSOjaswin Mujoo ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac, 4282bedc5d34SBaokun Li ext4_lblk_t *start, loff_t *end) 42830830344cSOjaswin Mujoo { 42840830344cSOjaswin Mujoo struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 42850830344cSOjaswin Mujoo struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 428638727786SOjaswin Mujoo struct ext4_prealloc_space *tmp_pa = NULL, *left_pa = NULL, *right_pa = NULL; 428738727786SOjaswin Mujoo struct rb_node *iter; 4288bedc5d34SBaokun Li ext4_lblk_t new_start, tmp_pa_start, right_pa_start = -1; 4289bedc5d34SBaokun Li loff_t new_end, tmp_pa_end, left_pa_end = -1; 42900830344cSOjaswin Mujoo 42910830344cSOjaswin Mujoo new_start = *start; 42920830344cSOjaswin Mujoo new_end = *end; 42930830344cSOjaswin Mujoo 429438727786SOjaswin Mujoo /* 429538727786SOjaswin Mujoo * Adjust the normalized range so that it doesn't overlap with any 429638727786SOjaswin Mujoo * existing preallocated blocks(PAs). Make sure to hold the rbtree lock 429738727786SOjaswin Mujoo * so it doesn't change underneath us. 429838727786SOjaswin Mujoo */ 429938727786SOjaswin Mujoo read_lock(&ei->i_prealloc_lock); 43000830344cSOjaswin Mujoo 430138727786SOjaswin Mujoo /* Step 1: find any one immediate neighboring PA of the normalized range */ 430238727786SOjaswin Mujoo for (iter = ei->i_prealloc_node.rb_node; iter; 430338727786SOjaswin Mujoo iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, 430438727786SOjaswin Mujoo tmp_pa_start, iter)) { 430538727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 430638727786SOjaswin Mujoo pa_node.inode_node); 43070830344cSOjaswin Mujoo tmp_pa_start = tmp_pa->pa_lstart; 4308bedc5d34SBaokun Li tmp_pa_end = pa_logical_end(sbi, tmp_pa); 43090830344cSOjaswin Mujoo 43100830344cSOjaswin Mujoo /* PA must not overlap original request */ 431138727786SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 431238727786SOjaswin Mujoo if (tmp_pa->pa_deleted == 0) 43130830344cSOjaswin Mujoo BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end || 43140830344cSOjaswin Mujoo ac->ac_o_ex.fe_logical < tmp_pa_start)); 43150830344cSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 43160830344cSOjaswin Mujoo } 43170830344cSOjaswin Mujoo 431838727786SOjaswin Mujoo /* 431938727786SOjaswin Mujoo * Step 2: check if the found PA is left or right neighbor and 432038727786SOjaswin Mujoo * get the other neighbor 432138727786SOjaswin Mujoo */ 432238727786SOjaswin Mujoo if (tmp_pa) { 432338727786SOjaswin Mujoo if (tmp_pa->pa_lstart < ac->ac_o_ex.fe_logical) { 432438727786SOjaswin Mujoo struct rb_node *tmp; 432538727786SOjaswin Mujoo 432638727786SOjaswin Mujoo left_pa = tmp_pa; 432738727786SOjaswin Mujoo tmp = rb_next(&left_pa->pa_node.inode_node); 432838727786SOjaswin Mujoo if (tmp) { 432938727786SOjaswin Mujoo right_pa = rb_entry(tmp, 433038727786SOjaswin Mujoo struct ext4_prealloc_space, 433138727786SOjaswin Mujoo pa_node.inode_node); 433238727786SOjaswin Mujoo } 433338727786SOjaswin Mujoo } else { 433438727786SOjaswin Mujoo struct rb_node *tmp; 433538727786SOjaswin Mujoo 433638727786SOjaswin Mujoo right_pa = tmp_pa; 433738727786SOjaswin Mujoo tmp = rb_prev(&right_pa->pa_node.inode_node); 433838727786SOjaswin Mujoo if (tmp) { 433938727786SOjaswin Mujoo left_pa = rb_entry(tmp, 434038727786SOjaswin Mujoo struct ext4_prealloc_space, 434138727786SOjaswin Mujoo pa_node.inode_node); 434238727786SOjaswin Mujoo } 434338727786SOjaswin Mujoo } 434438727786SOjaswin Mujoo } 434538727786SOjaswin Mujoo 434638727786SOjaswin Mujoo /* Step 3: get the non deleted neighbors */ 434738727786SOjaswin Mujoo if (left_pa) { 434838727786SOjaswin Mujoo for (iter = &left_pa->pa_node.inode_node;; 434938727786SOjaswin Mujoo iter = rb_prev(iter)) { 435038727786SOjaswin Mujoo if (!iter) { 435138727786SOjaswin Mujoo left_pa = NULL; 435238727786SOjaswin Mujoo break; 435338727786SOjaswin Mujoo } 435438727786SOjaswin Mujoo 435538727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 435638727786SOjaswin Mujoo pa_node.inode_node); 435738727786SOjaswin Mujoo left_pa = tmp_pa; 435838727786SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 435938727786SOjaswin Mujoo if (tmp_pa->pa_deleted == 0) { 436038727786SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 436138727786SOjaswin Mujoo break; 43620830344cSOjaswin Mujoo } 43630830344cSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 43640830344cSOjaswin Mujoo } 436538727786SOjaswin Mujoo } 436638727786SOjaswin Mujoo 436738727786SOjaswin Mujoo if (right_pa) { 436838727786SOjaswin Mujoo for (iter = &right_pa->pa_node.inode_node;; 436938727786SOjaswin Mujoo iter = rb_next(iter)) { 437038727786SOjaswin Mujoo if (!iter) { 437138727786SOjaswin Mujoo right_pa = NULL; 437238727786SOjaswin Mujoo break; 437338727786SOjaswin Mujoo } 437438727786SOjaswin Mujoo 437538727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 437638727786SOjaswin Mujoo pa_node.inode_node); 437738727786SOjaswin Mujoo right_pa = tmp_pa; 437838727786SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 437938727786SOjaswin Mujoo if (tmp_pa->pa_deleted == 0) { 438038727786SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 438138727786SOjaswin Mujoo break; 438238727786SOjaswin Mujoo } 438338727786SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 438438727786SOjaswin Mujoo } 438538727786SOjaswin Mujoo } 438638727786SOjaswin Mujoo 438738727786SOjaswin Mujoo if (left_pa) { 4388bedc5d34SBaokun Li left_pa_end = pa_logical_end(sbi, left_pa); 438938727786SOjaswin Mujoo BUG_ON(left_pa_end > ac->ac_o_ex.fe_logical); 439038727786SOjaswin Mujoo } 439138727786SOjaswin Mujoo 439238727786SOjaswin Mujoo if (right_pa) { 439338727786SOjaswin Mujoo right_pa_start = right_pa->pa_lstart; 439438727786SOjaswin Mujoo BUG_ON(right_pa_start <= ac->ac_o_ex.fe_logical); 439538727786SOjaswin Mujoo } 439638727786SOjaswin Mujoo 439738727786SOjaswin Mujoo /* Step 4: trim our normalized range to not overlap with the neighbors */ 439838727786SOjaswin Mujoo if (left_pa) { 439938727786SOjaswin Mujoo if (left_pa_end > new_start) 440038727786SOjaswin Mujoo new_start = left_pa_end; 440138727786SOjaswin Mujoo } 440238727786SOjaswin Mujoo 440338727786SOjaswin Mujoo if (right_pa) { 440438727786SOjaswin Mujoo if (right_pa_start < new_end) 440538727786SOjaswin Mujoo new_end = right_pa_start; 440638727786SOjaswin Mujoo } 440738727786SOjaswin Mujoo read_unlock(&ei->i_prealloc_lock); 44080830344cSOjaswin Mujoo 44090830344cSOjaswin Mujoo /* XXX: extra loop to check we really don't overlap preallocations */ 44100830344cSOjaswin Mujoo ext4_mb_pa_assert_overlap(ac, new_start, new_end); 44110830344cSOjaswin Mujoo 44120830344cSOjaswin Mujoo *start = new_start; 44130830344cSOjaswin Mujoo *end = new_end; 44140830344cSOjaswin Mujoo } 44150830344cSOjaswin Mujoo 44160830344cSOjaswin Mujoo /* 4417c9de560dSAlex Tomas * Normalization means making request better in terms of 4418c9de560dSAlex Tomas * size and alignment 4419c9de560dSAlex Tomas */ 44204ddfef7bSEric Sandeen static noinline_for_stack void 44214ddfef7bSEric Sandeen ext4_mb_normalize_request(struct ext4_allocation_context *ac, 4422c9de560dSAlex Tomas struct ext4_allocation_request *ar) 4423c9de560dSAlex Tomas { 442453accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4425b07ffe69SKemeng Shi struct ext4_super_block *es = sbi->s_es; 4426c9de560dSAlex Tomas int bsbits, max; 4427bedc5d34SBaokun Li loff_t size, start_off, end; 44281592d2c5SCurt Wohlgemuth loff_t orig_size __maybe_unused; 44295a0790c2SAndi Kleen ext4_lblk_t start; 4430c9de560dSAlex Tomas 4431c9de560dSAlex Tomas /* do normalize only data requests, metadata requests 4432c9de560dSAlex Tomas do not need preallocation */ 4433c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 4434c9de560dSAlex Tomas return; 4435c9de560dSAlex Tomas 4436c9de560dSAlex Tomas /* sometime caller may want exact blocks */ 4437c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 4438c9de560dSAlex Tomas return; 4439c9de560dSAlex Tomas 4440c9de560dSAlex Tomas /* caller may indicate that preallocation isn't 4441c9de560dSAlex Tomas * required (it's a tail, for example) */ 4442c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 4443c9de560dSAlex Tomas return; 4444c9de560dSAlex Tomas 4445c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 4446c9de560dSAlex Tomas ext4_mb_normalize_group_request(ac); 4447c9de560dSAlex Tomas return ; 4448c9de560dSAlex Tomas } 4449c9de560dSAlex Tomas 4450c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 4451c9de560dSAlex Tomas 4452c9de560dSAlex Tomas /* first, let's learn actual file size 4453c9de560dSAlex Tomas * given current request is allocated */ 445443bbddc0SBaokun Li size = extent_logical_end(sbi, &ac->ac_o_ex); 4455c9de560dSAlex Tomas size = size << bsbits; 4456c9de560dSAlex Tomas if (size < i_size_read(ac->ac_inode)) 4457c9de560dSAlex Tomas size = i_size_read(ac->ac_inode); 44585a0790c2SAndi Kleen orig_size = size; 4459c9de560dSAlex Tomas 44601930479cSValerie Clement /* max size of free chunks */ 44611930479cSValerie Clement max = 2 << bsbits; 4462c9de560dSAlex Tomas 44631930479cSValerie Clement #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 44641930479cSValerie Clement (req <= (size) || max <= (chunk_size)) 4465c9de560dSAlex Tomas 4466c9de560dSAlex Tomas /* first, try to predict filesize */ 4467c9de560dSAlex Tomas /* XXX: should this table be tunable? */ 4468c9de560dSAlex Tomas start_off = 0; 4469c9de560dSAlex Tomas if (size <= 16 * 1024) { 4470c9de560dSAlex Tomas size = 16 * 1024; 4471c9de560dSAlex Tomas } else if (size <= 32 * 1024) { 4472c9de560dSAlex Tomas size = 32 * 1024; 4473c9de560dSAlex Tomas } else if (size <= 64 * 1024) { 4474c9de560dSAlex Tomas size = 64 * 1024; 4475c9de560dSAlex Tomas } else if (size <= 128 * 1024) { 4476c9de560dSAlex Tomas size = 128 * 1024; 4477c9de560dSAlex Tomas } else if (size <= 256 * 1024) { 4478c9de560dSAlex Tomas size = 256 * 1024; 4479c9de560dSAlex Tomas } else if (size <= 512 * 1024) { 4480c9de560dSAlex Tomas size = 512 * 1024; 4481c9de560dSAlex Tomas } else if (size <= 1024 * 1024) { 4482c9de560dSAlex Tomas size = 1024 * 1024; 44831930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 4484c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 44851930479cSValerie Clement (21 - bsbits)) << 21; 44861930479cSValerie Clement size = 2 * 1024 * 1024; 44871930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 4488c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 4489c9de560dSAlex Tomas (22 - bsbits)) << 22; 4490c9de560dSAlex Tomas size = 4 * 1024 * 1024; 4491b3916da0SKemeng Shi } else if (NRL_CHECK_SIZE(EXT4_C2B(sbi, ac->ac_o_ex.fe_len), 44921930479cSValerie Clement (8<<20)>>bsbits, max, 8 * 1024)) { 4493c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 4494c9de560dSAlex Tomas (23 - bsbits)) << 23; 4495c9de560dSAlex Tomas size = 8 * 1024 * 1024; 4496c9de560dSAlex Tomas } else { 4497c9de560dSAlex Tomas start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; 449891a48aafSKemeng Shi size = (loff_t) EXT4_C2B(sbi, 4499b27b1535SXiaoguang Wang ac->ac_o_ex.fe_len) << bsbits; 4500c9de560dSAlex Tomas } 45015a0790c2SAndi Kleen size = size >> bsbits; 45025a0790c2SAndi Kleen start = start_off >> bsbits; 4503c9de560dSAlex Tomas 4504a08f789dSBaokun Li /* 4505a08f789dSBaokun Li * For tiny groups (smaller than 8MB) the chosen allocation 4506a08f789dSBaokun Li * alignment may be larger than group size. Make sure the 4507a08f789dSBaokun Li * alignment does not move allocation to a different group which 4508a08f789dSBaokun Li * makes mballoc fail assertions later. 4509a08f789dSBaokun Li */ 4510a08f789dSBaokun Li start = max(start, rounddown(ac->ac_o_ex.fe_logical, 4511a08f789dSBaokun Li (ext4_lblk_t)EXT4_BLOCKS_PER_GROUP(ac->ac_sb))); 4512a08f789dSBaokun Li 45132dcf5fdeSBaokun Li /* avoid unnecessary preallocation that may trigger assertions */ 45142dcf5fdeSBaokun Li if (start + size > EXT_MAX_BLOCKS) 45152dcf5fdeSBaokun Li size = EXT_MAX_BLOCKS - start; 45162dcf5fdeSBaokun Li 4517c9de560dSAlex Tomas /* don't cover already allocated blocks in selected range */ 4518c9de560dSAlex Tomas if (ar->pleft && start <= ar->lleft) { 4519c9de560dSAlex Tomas size -= ar->lleft + 1 - start; 4520c9de560dSAlex Tomas start = ar->lleft + 1; 4521c9de560dSAlex Tomas } 4522c9de560dSAlex Tomas if (ar->pright && start + size - 1 >= ar->lright) 4523c9de560dSAlex Tomas size -= start + size - ar->lright; 4524c9de560dSAlex Tomas 4525cd648b8aSJan Kara /* 4526cd648b8aSJan Kara * Trim allocation request for filesystems with artificially small 4527cd648b8aSJan Kara * groups. 4528cd648b8aSJan Kara */ 4529cd648b8aSJan Kara if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) 4530cd648b8aSJan Kara size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb); 4531cd648b8aSJan Kara 4532c9de560dSAlex Tomas end = start + size; 4533c9de560dSAlex Tomas 45340830344cSOjaswin Mujoo ext4_mb_pa_adjust_overlap(ac, &start, &end); 4535c9de560dSAlex Tomas 4536c9de560dSAlex Tomas size = end - start; 4537c9de560dSAlex Tomas 4538cf4ff938SBaokun Li /* 4539cf4ff938SBaokun Li * In this function "start" and "size" are normalized for better 4540cf4ff938SBaokun Li * alignment and length such that we could preallocate more blocks. 4541cf4ff938SBaokun Li * This normalization is done such that original request of 4542cf4ff938SBaokun Li * ac->ac_o_ex.fe_logical & fe_len should always lie within "start" and 4543cf4ff938SBaokun Li * "size" boundaries. 4544cf4ff938SBaokun Li * (Note fe_len can be relaxed since FS block allocation API does not 4545cf4ff938SBaokun Li * provide gurantee on number of contiguous blocks allocation since that 4546cf4ff938SBaokun Li * depends upon free space left, etc). 4547cf4ff938SBaokun Li * In case of inode pa, later we use the allocated blocks 45481221b235SKemeng Shi * [pa_pstart + fe_logical - pa_lstart, fe_len/size] from the preallocated 4549cf4ff938SBaokun Li * range of goal/best blocks [start, size] to put it at the 4550cf4ff938SBaokun Li * ac_o_ex.fe_logical extent of this inode. 4551cf4ff938SBaokun Li * (See ext4_mb_use_inode_pa() for more details) 4552cf4ff938SBaokun Li */ 4553cf4ff938SBaokun Li if (start + size <= ac->ac_o_ex.fe_logical || 4554c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical) { 45559d8b9ec4STheodore Ts'o ext4_msg(ac->ac_sb, KERN_ERR, 45569d8b9ec4STheodore Ts'o "start %lu, size %lu, fe_logical %lu", 4557c9de560dSAlex Tomas (unsigned long) start, (unsigned long) size, 4558c9de560dSAlex Tomas (unsigned long) ac->ac_o_ex.fe_logical); 4559dfe076c1SDmitry Monakhov BUG(); 4560c9de560dSAlex Tomas } 4561b5b60778SMaurizio Lombardi BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 4562c9de560dSAlex Tomas 4563c9de560dSAlex Tomas /* now prepare goal request */ 4564c9de560dSAlex Tomas 4565c9de560dSAlex Tomas /* XXX: is it better to align blocks WRT to logical 4566c9de560dSAlex Tomas * placement or satisfy big request as is */ 4567c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = start; 456853accfa9STheodore Ts'o ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); 45697e170922SOjaswin Mujoo ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; 4570c9de560dSAlex Tomas 4571c9de560dSAlex Tomas /* define goal start in order to merge */ 4572b07ffe69SKemeng Shi if (ar->pright && (ar->lright == (start + size)) && 4573b07ffe69SKemeng Shi ar->pright >= size && 4574b07ffe69SKemeng Shi ar->pright - size >= le32_to_cpu(es->s_first_data_block)) { 4575c9de560dSAlex Tomas /* merge to the right */ 4576c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 4577b07ffe69SKemeng Shi &ac->ac_g_ex.fe_group, 4578b07ffe69SKemeng Shi &ac->ac_g_ex.fe_start); 4579c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 4580c9de560dSAlex Tomas } 4581b07ffe69SKemeng Shi if (ar->pleft && (ar->lleft + 1 == start) && 4582b07ffe69SKemeng Shi ar->pleft + 1 < ext4_blocks_count(es)) { 4583c9de560dSAlex Tomas /* merge to the left */ 4584c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 4585b07ffe69SKemeng Shi &ac->ac_g_ex.fe_group, 4586b07ffe69SKemeng Shi &ac->ac_g_ex.fe_start); 4587c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 4588c9de560dSAlex Tomas } 4589c9de560dSAlex Tomas 4590d3df1453SRitesh Harjani mb_debug(ac->ac_sb, "goal: %lld(was %lld) blocks at %u\n", size, 4591d3df1453SRitesh Harjani orig_size, start); 4592c9de560dSAlex Tomas } 4593c9de560dSAlex Tomas 4594c9de560dSAlex Tomas static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 4595c9de560dSAlex Tomas { 4596c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4597c9de560dSAlex Tomas 4598a6c75eafSHarshad Shirwadkar if (sbi->s_mb_stats && ac->ac_g_ex.fe_len >= 1) { 4599c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_reqs); 4600c9de560dSAlex Tomas atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 4601291dae47SCurt Wohlgemuth if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) 4602c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_success); 4603fdd9a009SOjaswin Mujoo 4604c9de560dSAlex Tomas atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 4605fdd9a009SOjaswin Mujoo for (int i=0; i<EXT4_MB_NUM_CRS; i++) { 4606fdd9a009SOjaswin Mujoo atomic_add(ac->ac_cX_found[i], &sbi->s_bal_cX_ex_scanned[i]); 4607fdd9a009SOjaswin Mujoo } 4608fdd9a009SOjaswin Mujoo 4609a6c75eafSHarshad Shirwadkar atomic_add(ac->ac_groups_scanned, &sbi->s_bal_groups_scanned); 4610c9de560dSAlex Tomas if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 4611c9de560dSAlex Tomas ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 4612c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_goals); 46137e170922SOjaswin Mujoo /* did we allocate as much as normalizer originally wanted? */ 46147e170922SOjaswin Mujoo if (ac->ac_f_ex.fe_len == ac->ac_orig_goal_len) 46153ef5d263SOjaswin Mujoo atomic_inc(&sbi->s_bal_len_goals); 46167e170922SOjaswin Mujoo 4617c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan) 4618c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_breaks); 4619c9de560dSAlex Tomas } 4620c9de560dSAlex Tomas 4621296c355cSTheodore Ts'o if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 4622296c355cSTheodore Ts'o trace_ext4_mballoc_alloc(ac); 4623296c355cSTheodore Ts'o else 4624296c355cSTheodore Ts'o trace_ext4_mballoc_prealloc(ac); 4625c9de560dSAlex Tomas } 4626c9de560dSAlex Tomas 4627c9de560dSAlex Tomas /* 4628b844167eSCurt Wohlgemuth * Called on failure; free up any blocks from the inode PA for this 4629b844167eSCurt Wohlgemuth * context. We don't need this for MB_GROUP_PA because we only change 4630b844167eSCurt Wohlgemuth * pa_free in ext4_mb_release_context(), but on failure, we've already 4631b844167eSCurt Wohlgemuth * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. 4632b844167eSCurt Wohlgemuth */ 4633b844167eSCurt Wohlgemuth static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) 4634b844167eSCurt Wohlgemuth { 4635b844167eSCurt Wohlgemuth struct ext4_prealloc_space *pa = ac->ac_pa; 463686f0afd4STheodore Ts'o struct ext4_buddy e4b; 463786f0afd4STheodore Ts'o int err; 4638b844167eSCurt Wohlgemuth 463986f0afd4STheodore Ts'o if (pa == NULL) { 4640c99d1e6eSTheodore Ts'o if (ac->ac_f_ex.fe_len == 0) 4641c99d1e6eSTheodore Ts'o return; 464286f0afd4STheodore Ts'o err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); 464319b8b035STheodore Ts'o if (WARN_RATELIMIT(err, 464419b8b035STheodore Ts'o "ext4: mb_load_buddy failed (%d)", err)) 464586f0afd4STheodore Ts'o /* 464686f0afd4STheodore Ts'o * This should never happen since we pin the 464786f0afd4STheodore Ts'o * pages in the ext4_allocation_context so 464886f0afd4STheodore Ts'o * ext4_mb_load_buddy() should never fail. 464986f0afd4STheodore Ts'o */ 465086f0afd4STheodore Ts'o return; 465186f0afd4STheodore Ts'o ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 465286f0afd4STheodore Ts'o mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, 465386f0afd4STheodore Ts'o ac->ac_f_ex.fe_len); 465486f0afd4STheodore Ts'o ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 4655c99d1e6eSTheodore Ts'o ext4_mb_unload_buddy(&e4b); 465686f0afd4STheodore Ts'o return; 465786f0afd4STheodore Ts'o } 465836cb0f52SKemeng Shi if (pa->pa_type == MB_INODE_PA) { 465936cb0f52SKemeng Shi spin_lock(&pa->pa_lock); 4660400db9d3SZheng Liu pa->pa_free += ac->ac_b_ex.fe_len; 466136cb0f52SKemeng Shi spin_unlock(&pa->pa_lock); 466236cb0f52SKemeng Shi } 4663b844167eSCurt Wohlgemuth } 4664b844167eSCurt Wohlgemuth 4665b844167eSCurt Wohlgemuth /* 4666c9de560dSAlex Tomas * use blocks preallocated to inode 4667c9de560dSAlex Tomas */ 4668c9de560dSAlex Tomas static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 4669c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 4670c9de560dSAlex Tomas { 467153accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4672c9de560dSAlex Tomas ext4_fsblk_t start; 4673c9de560dSAlex Tomas ext4_fsblk_t end; 4674c9de560dSAlex Tomas int len; 4675c9de560dSAlex Tomas 4676c9de560dSAlex Tomas /* found preallocated blocks, use them */ 4677c9de560dSAlex Tomas start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 467853accfa9STheodore Ts'o end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), 467953accfa9STheodore Ts'o start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); 468053accfa9STheodore Ts'o len = EXT4_NUM_B2C(sbi, end - start); 4681c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 4682c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 4683c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 4684c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 4685c9de560dSAlex Tomas ac->ac_pa = pa; 4686c9de560dSAlex Tomas 4687c9de560dSAlex Tomas BUG_ON(start < pa->pa_pstart); 468853accfa9STheodore Ts'o BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); 4689c9de560dSAlex Tomas BUG_ON(pa->pa_free < len); 469093cdf49fSOjaswin Mujoo BUG_ON(ac->ac_b_ex.fe_len <= 0); 4691c9de560dSAlex Tomas pa->pa_free -= len; 4692c9de560dSAlex Tomas 4693d3df1453SRitesh Harjani mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa); 4694c9de560dSAlex Tomas } 4695c9de560dSAlex Tomas 4696c9de560dSAlex Tomas /* 4697c9de560dSAlex Tomas * use blocks preallocated to locality group 4698c9de560dSAlex Tomas */ 4699c9de560dSAlex Tomas static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 4700c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 4701c9de560dSAlex Tomas { 470203cddb80SAneesh Kumar K.V unsigned int len = ac->ac_o_ex.fe_len; 47036be2ded1SAneesh Kumar K.V 4704c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 4705c9de560dSAlex Tomas &ac->ac_b_ex.fe_group, 4706c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 4707c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 4708c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 4709c9de560dSAlex Tomas ac->ac_pa = pa; 4710c9de560dSAlex Tomas 47111221b235SKemeng Shi /* we don't correct pa_pstart or pa_len here to avoid 471226346ff6SAneesh Kumar K.V * possible race when the group is being loaded concurrently 4713c9de560dSAlex Tomas * instead we correct pa later, after blocks are marked 471426346ff6SAneesh Kumar K.V * in on-disk bitmap -- see ext4_mb_release_context() 471526346ff6SAneesh Kumar K.V * Other CPUs are prevented from allocating from this pa by lg_mutex 4716c9de560dSAlex Tomas */ 4717d3df1453SRitesh Harjani mb_debug(ac->ac_sb, "use %u/%u from group pa %p\n", 47181afdc588SKemeng Shi pa->pa_lstart, len, pa); 4719c9de560dSAlex Tomas } 4720c9de560dSAlex Tomas 4721c9de560dSAlex Tomas /* 47225e745b04SAneesh Kumar K.V * Return the prealloc space that have minimal distance 47235e745b04SAneesh Kumar K.V * from the goal block. @cpa is the prealloc 47245e745b04SAneesh Kumar K.V * space that is having currently known minimal distance 47255e745b04SAneesh Kumar K.V * from the goal block. 47265e745b04SAneesh Kumar K.V */ 47275e745b04SAneesh Kumar K.V static struct ext4_prealloc_space * 47285e745b04SAneesh Kumar K.V ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 47295e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, 47305e745b04SAneesh Kumar K.V struct ext4_prealloc_space *cpa) 47315e745b04SAneesh Kumar K.V { 47325e745b04SAneesh Kumar K.V ext4_fsblk_t cur_distance, new_distance; 47335e745b04SAneesh Kumar K.V 47345e745b04SAneesh Kumar K.V if (cpa == NULL) { 47355e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 47365e745b04SAneesh Kumar K.V return pa; 47375e745b04SAneesh Kumar K.V } 473879211c8eSAndrew Morton cur_distance = abs(goal_block - cpa->pa_pstart); 473979211c8eSAndrew Morton new_distance = abs(goal_block - pa->pa_pstart); 47405e745b04SAneesh Kumar K.V 47415a54b2f1SColy Li if (cur_distance <= new_distance) 47425e745b04SAneesh Kumar K.V return cpa; 47435e745b04SAneesh Kumar K.V 47445e745b04SAneesh Kumar K.V /* drop the previous reference */ 47455e745b04SAneesh Kumar K.V atomic_dec(&cpa->pa_count); 47465e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 47475e745b04SAneesh Kumar K.V return pa; 47485e745b04SAneesh Kumar K.V } 47495e745b04SAneesh Kumar K.V 47505e745b04SAneesh Kumar K.V /* 47511eff5904SKemeng Shi * check if found pa meets EXT4_MB_HINT_GOAL_ONLY 47521eff5904SKemeng Shi */ 47531eff5904SKemeng Shi static bool 47541eff5904SKemeng Shi ext4_mb_pa_goal_check(struct ext4_allocation_context *ac, 47551eff5904SKemeng Shi struct ext4_prealloc_space *pa) 47561eff5904SKemeng Shi { 47571eff5904SKemeng Shi struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 47581eff5904SKemeng Shi ext4_fsblk_t start; 47591eff5904SKemeng Shi 47601eff5904SKemeng Shi if (likely(!(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))) 47611eff5904SKemeng Shi return true; 47621eff5904SKemeng Shi 47631eff5904SKemeng Shi /* 47641eff5904SKemeng Shi * If EXT4_MB_HINT_GOAL_ONLY is set, ac_g_ex will not be adjusted 47651eff5904SKemeng Shi * in ext4_mb_normalize_request and will keep same with ac_o_ex 47661eff5904SKemeng Shi * from ext4_mb_initialize_context. Choose ac_g_ex here to keep 47671eff5904SKemeng Shi * consistent with ext4_mb_find_by_goal. 47681eff5904SKemeng Shi */ 47691eff5904SKemeng Shi start = pa->pa_pstart + 47701eff5904SKemeng Shi (ac->ac_g_ex.fe_logical - pa->pa_lstart); 47711eff5904SKemeng Shi if (ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex) != start) 47721eff5904SKemeng Shi return false; 47731eff5904SKemeng Shi 47741eff5904SKemeng Shi if (ac->ac_g_ex.fe_len > pa->pa_len - 47751eff5904SKemeng Shi EXT4_B2C(sbi, ac->ac_g_ex.fe_logical - pa->pa_lstart)) 47761eff5904SKemeng Shi return false; 47771eff5904SKemeng Shi 47781eff5904SKemeng Shi return true; 47791eff5904SKemeng Shi } 47801eff5904SKemeng Shi 47811eff5904SKemeng Shi /* 4782c9de560dSAlex Tomas * search goal blocks in preallocated space 4783c9de560dSAlex Tomas */ 47844fca8f07SRitesh Harjani static noinline_for_stack bool 47854ddfef7bSEric Sandeen ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 4786c9de560dSAlex Tomas { 478753accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 47886be2ded1SAneesh Kumar K.V int order, i; 4789c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 4790c9de560dSAlex Tomas struct ext4_locality_group *lg; 47919d3de7eeSOjaswin Mujoo struct ext4_prealloc_space *tmp_pa = NULL, *cpa = NULL; 479238727786SOjaswin Mujoo struct rb_node *iter; 47935e745b04SAneesh Kumar K.V ext4_fsblk_t goal_block; 4794c9de560dSAlex Tomas 4795c9de560dSAlex Tomas /* only data can be preallocated */ 4796c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 47974fca8f07SRitesh Harjani return false; 4798c9de560dSAlex Tomas 47999d3de7eeSOjaswin Mujoo /* 48009d3de7eeSOjaswin Mujoo * first, try per-file preallocation by searching the inode pa rbtree. 48019d3de7eeSOjaswin Mujoo * 48029d3de7eeSOjaswin Mujoo * Here, we can't do a direct traversal of the tree because 48039d3de7eeSOjaswin Mujoo * ext4_mb_discard_group_preallocation() can paralelly mark the pa 48049d3de7eeSOjaswin Mujoo * deleted and that can cause direct traversal to skip some entries. 48059d3de7eeSOjaswin Mujoo */ 480638727786SOjaswin Mujoo read_lock(&ei->i_prealloc_lock); 48079d3de7eeSOjaswin Mujoo 48089d3de7eeSOjaswin Mujoo if (RB_EMPTY_ROOT(&ei->i_prealloc_node)) { 48099d3de7eeSOjaswin Mujoo goto try_group_pa; 48109d3de7eeSOjaswin Mujoo } 48119d3de7eeSOjaswin Mujoo 48129d3de7eeSOjaswin Mujoo /* 48139d3de7eeSOjaswin Mujoo * Step 1: Find a pa with logical start immediately adjacent to the 48149d3de7eeSOjaswin Mujoo * original logical start. This could be on the left or right. 48159d3de7eeSOjaswin Mujoo * 48169d3de7eeSOjaswin Mujoo * (tmp_pa->pa_lstart never changes so we can skip locking for it). 48179d3de7eeSOjaswin Mujoo */ 481838727786SOjaswin Mujoo for (iter = ei->i_prealloc_node.rb_node; iter; 481938727786SOjaswin Mujoo iter = ext4_mb_pa_rb_next_iter(ac->ac_o_ex.fe_logical, 48209d3de7eeSOjaswin Mujoo tmp_pa->pa_lstart, iter)) { 482138727786SOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 482238727786SOjaswin Mujoo pa_node.inode_node); 48239d3de7eeSOjaswin Mujoo } 4824c9de560dSAlex Tomas 48259d3de7eeSOjaswin Mujoo /* 48269d3de7eeSOjaswin Mujoo * Step 2: The adjacent pa might be to the right of logical start, find 48279d3de7eeSOjaswin Mujoo * the left adjacent pa. After this step we'd have a valid tmp_pa whose 48289d3de7eeSOjaswin Mujoo * logical start is towards the left of original request's logical start 48299d3de7eeSOjaswin Mujoo */ 48309d3de7eeSOjaswin Mujoo if (tmp_pa->pa_lstart > ac->ac_o_ex.fe_logical) { 48319d3de7eeSOjaswin Mujoo struct rb_node *tmp; 48329d3de7eeSOjaswin Mujoo tmp = rb_prev(&tmp_pa->pa_node.inode_node); 4833bcf43499SOjaswin Mujoo 48349d3de7eeSOjaswin Mujoo if (tmp) { 48359d3de7eeSOjaswin Mujoo tmp_pa = rb_entry(tmp, struct ext4_prealloc_space, 48369d3de7eeSOjaswin Mujoo pa_node.inode_node); 48379d3de7eeSOjaswin Mujoo } else { 48389d3de7eeSOjaswin Mujoo /* 48399d3de7eeSOjaswin Mujoo * If there is no adjacent pa to the left then finding 48409d3de7eeSOjaswin Mujoo * an overlapping pa is not possible hence stop searching 48419d3de7eeSOjaswin Mujoo * inode pa tree 48429d3de7eeSOjaswin Mujoo */ 48439d3de7eeSOjaswin Mujoo goto try_group_pa; 48449d3de7eeSOjaswin Mujoo } 48459d3de7eeSOjaswin Mujoo } 48469d3de7eeSOjaswin Mujoo 48479d3de7eeSOjaswin Mujoo BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); 48489d3de7eeSOjaswin Mujoo 48499d3de7eeSOjaswin Mujoo /* 48509d3de7eeSOjaswin Mujoo * Step 3: If the left adjacent pa is deleted, keep moving left to find 48519d3de7eeSOjaswin Mujoo * the first non deleted adjacent pa. After this step we should have a 48529d3de7eeSOjaswin Mujoo * valid tmp_pa which is guaranteed to be non deleted. 48539d3de7eeSOjaswin Mujoo */ 48549d3de7eeSOjaswin Mujoo for (iter = &tmp_pa->pa_node.inode_node;; iter = rb_prev(iter)) { 48559d3de7eeSOjaswin Mujoo if (!iter) { 48569d3de7eeSOjaswin Mujoo /* 48579d3de7eeSOjaswin Mujoo * no non deleted left adjacent pa, so stop searching 48589d3de7eeSOjaswin Mujoo * inode pa tree 48599d3de7eeSOjaswin Mujoo */ 48609d3de7eeSOjaswin Mujoo goto try_group_pa; 48619d3de7eeSOjaswin Mujoo } 48629d3de7eeSOjaswin Mujoo tmp_pa = rb_entry(iter, struct ext4_prealloc_space, 48639d3de7eeSOjaswin Mujoo pa_node.inode_node); 48649d3de7eeSOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 48659d3de7eeSOjaswin Mujoo if (tmp_pa->pa_deleted == 0) { 48669d3de7eeSOjaswin Mujoo /* 48679d3de7eeSOjaswin Mujoo * We will keep holding the pa_lock from 48689d3de7eeSOjaswin Mujoo * this point on because we don't want group discard 48699d3de7eeSOjaswin Mujoo * to delete this pa underneath us. Since group 48709d3de7eeSOjaswin Mujoo * discard is anyways an ENOSPC operation it 48719d3de7eeSOjaswin Mujoo * should be okay for it to wait a few more cycles. 48729d3de7eeSOjaswin Mujoo */ 48739d3de7eeSOjaswin Mujoo break; 48749d3de7eeSOjaswin Mujoo } else { 48759d3de7eeSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 48769d3de7eeSOjaswin Mujoo } 48779d3de7eeSOjaswin Mujoo } 48789d3de7eeSOjaswin Mujoo 48799d3de7eeSOjaswin Mujoo BUG_ON(!(tmp_pa && tmp_pa->pa_lstart <= ac->ac_o_ex.fe_logical)); 48809d3de7eeSOjaswin Mujoo BUG_ON(tmp_pa->pa_deleted == 1); 48819d3de7eeSOjaswin Mujoo 48829d3de7eeSOjaswin Mujoo /* 48839d3de7eeSOjaswin Mujoo * Step 4: We now have the non deleted left adjacent pa. Only this 48849d3de7eeSOjaswin Mujoo * pa can possibly satisfy the request hence check if it overlaps 48859d3de7eeSOjaswin Mujoo * original logical start and stop searching if it doesn't. 48869d3de7eeSOjaswin Mujoo */ 488743bbddc0SBaokun Li if (ac->ac_o_ex.fe_logical >= pa_logical_end(sbi, tmp_pa)) { 48889d3de7eeSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 48899d3de7eeSOjaswin Mujoo goto try_group_pa; 48909d3de7eeSOjaswin Mujoo } 4891c9de560dSAlex Tomas 4892fb0a387dSEric Sandeen /* non-extent files can't have physical blocks past 2^32 */ 489312e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && 4894bcf43499SOjaswin Mujoo (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) > 4895e86a7182SOjaswin Mujoo EXT4_MAX_BLOCK_FILE_PHYS)) { 4896e86a7182SOjaswin Mujoo /* 48979d3de7eeSOjaswin Mujoo * Since PAs don't overlap, we won't find any other PA to 48989d3de7eeSOjaswin Mujoo * satisfy this. 4899e86a7182SOjaswin Mujoo */ 49009d3de7eeSOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 49019d3de7eeSOjaswin Mujoo goto try_group_pa; 4902e86a7182SOjaswin Mujoo } 4903fb0a387dSEric Sandeen 49049d3de7eeSOjaswin Mujoo if (tmp_pa->pa_free && likely(ext4_mb_pa_goal_check(ac, tmp_pa))) { 4905bcf43499SOjaswin Mujoo atomic_inc(&tmp_pa->pa_count); 4906bcf43499SOjaswin Mujoo ext4_mb_use_inode_pa(ac, tmp_pa); 4907bcf43499SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 490838727786SOjaswin Mujoo read_unlock(&ei->i_prealloc_lock); 49094fca8f07SRitesh Harjani return true; 49109d3de7eeSOjaswin Mujoo } else { 49119d3de7eeSOjaswin Mujoo /* 49129d3de7eeSOjaswin Mujoo * We found a valid overlapping pa but couldn't use it because 49139d3de7eeSOjaswin Mujoo * it had no free blocks. This should ideally never happen 49149d3de7eeSOjaswin Mujoo * because: 49159d3de7eeSOjaswin Mujoo * 49169d3de7eeSOjaswin Mujoo * 1. When a new inode pa is added to rbtree it must have 49179d3de7eeSOjaswin Mujoo * pa_free > 0 since otherwise we won't actually need 49189d3de7eeSOjaswin Mujoo * preallocation. 49199d3de7eeSOjaswin Mujoo * 49209d3de7eeSOjaswin Mujoo * 2. An inode pa that is in the rbtree can only have it's 49219d3de7eeSOjaswin Mujoo * pa_free become zero when another thread calls: 49229d3de7eeSOjaswin Mujoo * ext4_mb_new_blocks 49239d3de7eeSOjaswin Mujoo * ext4_mb_use_preallocated 49249d3de7eeSOjaswin Mujoo * ext4_mb_use_inode_pa 49259d3de7eeSOjaswin Mujoo * 49269d3de7eeSOjaswin Mujoo * 3. Further, after the above calls make pa_free == 0, we will 49279d3de7eeSOjaswin Mujoo * immediately remove it from the rbtree in: 49289d3de7eeSOjaswin Mujoo * ext4_mb_new_blocks 49299d3de7eeSOjaswin Mujoo * ext4_mb_release_context 49309d3de7eeSOjaswin Mujoo * ext4_mb_put_pa 49319d3de7eeSOjaswin Mujoo * 49329d3de7eeSOjaswin Mujoo * 4. Since the pa_free becoming 0 and pa_free getting removed 49339d3de7eeSOjaswin Mujoo * from tree both happen in ext4_mb_new_blocks, which is always 49349d3de7eeSOjaswin Mujoo * called with i_data_sem held for data allocations, we can be 49359d3de7eeSOjaswin Mujoo * sure that another process will never see a pa in rbtree with 49369d3de7eeSOjaswin Mujoo * pa_free == 0. 49379d3de7eeSOjaswin Mujoo */ 49389d3de7eeSOjaswin Mujoo WARN_ON_ONCE(tmp_pa->pa_free == 0); 4939c9de560dSAlex Tomas } 4940bcf43499SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 49419d3de7eeSOjaswin Mujoo try_group_pa: 494238727786SOjaswin Mujoo read_unlock(&ei->i_prealloc_lock); 4943c9de560dSAlex Tomas 4944c9de560dSAlex Tomas /* can we use group allocation? */ 4945c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 49464fca8f07SRitesh Harjani return false; 4947c9de560dSAlex Tomas 4948c9de560dSAlex Tomas /* inode may have no locality group for some reason */ 4949c9de560dSAlex Tomas lg = ac->ac_lg; 4950c9de560dSAlex Tomas if (lg == NULL) 49514fca8f07SRitesh Harjani return false; 49526be2ded1SAneesh Kumar K.V order = fls(ac->ac_o_ex.fe_len) - 1; 49536be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 49546be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 49556be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 4956c9de560dSAlex Tomas 4957bda00de7SAkinobu Mita goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); 49585e745b04SAneesh Kumar K.V /* 49595e745b04SAneesh Kumar K.V * search for the prealloc space that is having 49605e745b04SAneesh Kumar K.V * minimal distance from the goal block. 49615e745b04SAneesh Kumar K.V */ 49626be2ded1SAneesh Kumar K.V for (i = order; i < PREALLOC_TB_SIZE; i++) { 4963c9de560dSAlex Tomas rcu_read_lock(); 4964bcf43499SOjaswin Mujoo list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i], 4965a8e38fd3SOjaswin Mujoo pa_node.lg_list) { 4966bcf43499SOjaswin Mujoo spin_lock(&tmp_pa->pa_lock); 4967bcf43499SOjaswin Mujoo if (tmp_pa->pa_deleted == 0 && 4968bcf43499SOjaswin Mujoo tmp_pa->pa_free >= ac->ac_o_ex.fe_len) { 49695e745b04SAneesh Kumar K.V 49705e745b04SAneesh Kumar K.V cpa = ext4_mb_check_group_pa(goal_block, 4971bcf43499SOjaswin Mujoo tmp_pa, cpa); 49725e745b04SAneesh Kumar K.V } 4973bcf43499SOjaswin Mujoo spin_unlock(&tmp_pa->pa_lock); 49745e745b04SAneesh Kumar K.V } 49755e745b04SAneesh Kumar K.V rcu_read_unlock(); 49765e745b04SAneesh Kumar K.V } 49775e745b04SAneesh Kumar K.V if (cpa) { 49785e745b04SAneesh Kumar K.V ext4_mb_use_group_pa(ac, cpa); 49794fca8f07SRitesh Harjani return true; 4980c9de560dSAlex Tomas } 49814fca8f07SRitesh Harjani return false; 4982c9de560dSAlex Tomas } 4983c9de560dSAlex Tomas 4984c9de560dSAlex Tomas /* 4985c9de560dSAlex Tomas * the function goes through all preallocation in this group and marks them 4986c9de560dSAlex Tomas * used in in-core bitmap. buddy must be generated from this bitmap 4987955ce5f5SAneesh Kumar K.V * Need to be called with ext4 group lock held 4988c9de560dSAlex Tomas */ 4989089ceeccSEric Sandeen static noinline_for_stack 4990089ceeccSEric Sandeen void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 4991c9de560dSAlex Tomas ext4_group_t group) 4992c9de560dSAlex Tomas { 4993c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 4994c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 4995c9de560dSAlex Tomas struct list_head *cur; 4996c9de560dSAlex Tomas ext4_group_t groupnr; 4997c9de560dSAlex Tomas ext4_grpblk_t start; 4998c9de560dSAlex Tomas int preallocated = 0; 4999c9de560dSAlex Tomas int len; 5000c9de560dSAlex Tomas 50015354b2afSTheodore Ts'o if (!grp) 50025354b2afSTheodore Ts'o return; 50035354b2afSTheodore Ts'o 5004c9de560dSAlex Tomas /* all form of preallocation discards first load group, 5005c9de560dSAlex Tomas * so the only competing code is preallocation use. 5006c9de560dSAlex Tomas * we don't need any locking here 5007c9de560dSAlex Tomas * notice we do NOT ignore preallocations with pa_deleted 5008c9de560dSAlex Tomas * otherwise we could leave used blocks available for 5009c9de560dSAlex Tomas * allocation in buddy when concurrent ext4_mb_put_pa() 5010c9de560dSAlex Tomas * is dropping preallocation 5011c9de560dSAlex Tomas */ 5012c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 5013c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 5014c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 5015c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 5016c9de560dSAlex Tomas &groupnr, &start); 5017c9de560dSAlex Tomas len = pa->pa_len; 5018c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5019c9de560dSAlex Tomas if (unlikely(len == 0)) 5020c9de560dSAlex Tomas continue; 5021c9de560dSAlex Tomas BUG_ON(groupnr != group); 5022123e3016SRitesh Harjani mb_set_bits(bitmap, start, len); 5023c9de560dSAlex Tomas preallocated += len; 5024c9de560dSAlex Tomas } 5025d3df1453SRitesh Harjani mb_debug(sb, "preallocated %d for group %u\n", preallocated, group); 5026c9de560dSAlex Tomas } 5027c9de560dSAlex Tomas 502827bc446eSbrookxu static void ext4_mb_mark_pa_deleted(struct super_block *sb, 502927bc446eSbrookxu struct ext4_prealloc_space *pa) 503027bc446eSbrookxu { 503127bc446eSbrookxu struct ext4_inode_info *ei; 503227bc446eSbrookxu 503327bc446eSbrookxu if (pa->pa_deleted) { 503427bc446eSbrookxu ext4_warning(sb, "deleted pa, type:%d, pblk:%llu, lblk:%u, len:%d\n", 503527bc446eSbrookxu pa->pa_type, pa->pa_pstart, pa->pa_lstart, 503627bc446eSbrookxu pa->pa_len); 503727bc446eSbrookxu return; 503827bc446eSbrookxu } 503927bc446eSbrookxu 504027bc446eSbrookxu pa->pa_deleted = 1; 504127bc446eSbrookxu 504227bc446eSbrookxu if (pa->pa_type == MB_INODE_PA) { 504327bc446eSbrookxu ei = EXT4_I(pa->pa_inode); 504427bc446eSbrookxu atomic_dec(&ei->i_prealloc_active); 504527bc446eSbrookxu } 504627bc446eSbrookxu } 504727bc446eSbrookxu 504882089725SOjaswin Mujoo static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa) 5049c9de560dSAlex Tomas { 505082089725SOjaswin Mujoo BUG_ON(!pa); 50514e8d2139SJunho Ryu BUG_ON(atomic_read(&pa->pa_count)); 50524e8d2139SJunho Ryu BUG_ON(pa->pa_deleted == 0); 5053c9de560dSAlex Tomas kmem_cache_free(ext4_pspace_cachep, pa); 5054c9de560dSAlex Tomas } 5055c9de560dSAlex Tomas 505682089725SOjaswin Mujoo static void ext4_mb_pa_callback(struct rcu_head *head) 505782089725SOjaswin Mujoo { 505882089725SOjaswin Mujoo struct ext4_prealloc_space *pa; 505982089725SOjaswin Mujoo 506082089725SOjaswin Mujoo pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 506182089725SOjaswin Mujoo ext4_mb_pa_free(pa); 506282089725SOjaswin Mujoo } 506382089725SOjaswin Mujoo 5064c9de560dSAlex Tomas /* 5065c9de560dSAlex Tomas * drops a reference to preallocated space descriptor 5066c9de560dSAlex Tomas * if this was the last reference and the space is consumed 5067c9de560dSAlex Tomas */ 5068c9de560dSAlex Tomas static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 5069c9de560dSAlex Tomas struct super_block *sb, struct ext4_prealloc_space *pa) 5070c9de560dSAlex Tomas { 5071a9df9a49STheodore Ts'o ext4_group_t grp; 5072d33a1976SEric Sandeen ext4_fsblk_t grp_blk; 507338727786SOjaswin Mujoo struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 5074c9de560dSAlex Tomas 5075c9de560dSAlex Tomas /* in this short window concurrent discard can set pa_deleted */ 5076c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 50774e8d2139SJunho Ryu if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { 50784e8d2139SJunho Ryu spin_unlock(&pa->pa_lock); 50794e8d2139SJunho Ryu return; 50804e8d2139SJunho Ryu } 50814e8d2139SJunho Ryu 5082c9de560dSAlex Tomas if (pa->pa_deleted == 1) { 5083c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5084c9de560dSAlex Tomas return; 5085c9de560dSAlex Tomas } 5086c9de560dSAlex Tomas 508727bc446eSbrookxu ext4_mb_mark_pa_deleted(sb, pa); 5088c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5089c9de560dSAlex Tomas 5090d33a1976SEric Sandeen grp_blk = pa->pa_pstart; 5091cc0fb9adSAneesh Kumar K.V /* 5092cc0fb9adSAneesh Kumar K.V * If doing group-based preallocation, pa_pstart may be in the 5093cc0fb9adSAneesh Kumar K.V * next group when pa is used up 5094cc0fb9adSAneesh Kumar K.V */ 5095cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 5096d33a1976SEric Sandeen grp_blk--; 5097d33a1976SEric Sandeen 5098bd86298eSLukas Czerner grp = ext4_get_group_number(sb, grp_blk); 5099c9de560dSAlex Tomas 5100c9de560dSAlex Tomas /* 5101c9de560dSAlex Tomas * possible race: 5102c9de560dSAlex Tomas * 5103c9de560dSAlex Tomas * P1 (buddy init) P2 (regular allocation) 5104c9de560dSAlex Tomas * find block B in PA 5105c9de560dSAlex Tomas * copy on-disk bitmap to buddy 5106c9de560dSAlex Tomas * mark B in on-disk bitmap 5107c9de560dSAlex Tomas * drop PA from group 5108c9de560dSAlex Tomas * mark all PAs in buddy 5109c9de560dSAlex Tomas * 5110c9de560dSAlex Tomas * thus, P1 initializes buddy with B available. to prevent this 5111c9de560dSAlex Tomas * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 5112c9de560dSAlex Tomas * against that pair 5113c9de560dSAlex Tomas */ 5114c9de560dSAlex Tomas ext4_lock_group(sb, grp); 5115c9de560dSAlex Tomas list_del(&pa->pa_group_list); 5116c9de560dSAlex Tomas ext4_unlock_group(sb, grp); 5117c9de560dSAlex Tomas 5118a8e38fd3SOjaswin Mujoo if (pa->pa_type == MB_INODE_PA) { 511938727786SOjaswin Mujoo write_lock(pa->pa_node_lock.inode_lock); 512038727786SOjaswin Mujoo rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 512138727786SOjaswin Mujoo write_unlock(pa->pa_node_lock.inode_lock); 512238727786SOjaswin Mujoo ext4_mb_pa_free(pa); 5123a8e38fd3SOjaswin Mujoo } else { 5124a8e38fd3SOjaswin Mujoo spin_lock(pa->pa_node_lock.lg_lock); 5125a8e38fd3SOjaswin Mujoo list_del_rcu(&pa->pa_node.lg_list); 5126a8e38fd3SOjaswin Mujoo spin_unlock(pa->pa_node_lock.lg_lock); 512738727786SOjaswin Mujoo call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 512838727786SOjaswin Mujoo } 5129a8e38fd3SOjaswin Mujoo } 5130c9de560dSAlex Tomas 513138727786SOjaswin Mujoo static void ext4_mb_pa_rb_insert(struct rb_root *root, struct rb_node *new) 513238727786SOjaswin Mujoo { 513338727786SOjaswin Mujoo struct rb_node **iter = &root->rb_node, *parent = NULL; 513438727786SOjaswin Mujoo struct ext4_prealloc_space *iter_pa, *new_pa; 513538727786SOjaswin Mujoo ext4_lblk_t iter_start, new_start; 513638727786SOjaswin Mujoo 513738727786SOjaswin Mujoo while (*iter) { 513838727786SOjaswin Mujoo iter_pa = rb_entry(*iter, struct ext4_prealloc_space, 513938727786SOjaswin Mujoo pa_node.inode_node); 514038727786SOjaswin Mujoo new_pa = rb_entry(new, struct ext4_prealloc_space, 514138727786SOjaswin Mujoo pa_node.inode_node); 514238727786SOjaswin Mujoo iter_start = iter_pa->pa_lstart; 514338727786SOjaswin Mujoo new_start = new_pa->pa_lstart; 514438727786SOjaswin Mujoo 514538727786SOjaswin Mujoo parent = *iter; 514638727786SOjaswin Mujoo if (new_start < iter_start) 514738727786SOjaswin Mujoo iter = &((*iter)->rb_left); 514838727786SOjaswin Mujoo else 514938727786SOjaswin Mujoo iter = &((*iter)->rb_right); 515038727786SOjaswin Mujoo } 515138727786SOjaswin Mujoo 515238727786SOjaswin Mujoo rb_link_node(new, parent, iter); 515338727786SOjaswin Mujoo rb_insert_color(new, root); 5154c9de560dSAlex Tomas } 5155c9de560dSAlex Tomas 5156c9de560dSAlex Tomas /* 5157c9de560dSAlex Tomas * creates new preallocated space for given inode 5158c9de560dSAlex Tomas */ 515953f86b17SRitesh Harjani static noinline_for_stack void 51604ddfef7bSEric Sandeen ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 5161c9de560dSAlex Tomas { 5162c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 516353accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(sb); 5164c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 5165c9de560dSAlex Tomas struct ext4_group_info *grp; 5166c9de560dSAlex Tomas struct ext4_inode_info *ei; 5167c9de560dSAlex Tomas 5168c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 5169c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 5170c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 5171c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 517253f86b17SRitesh Harjani BUG_ON(ac->ac_pa == NULL); 5173c9de560dSAlex Tomas 517453f86b17SRitesh Harjani pa = ac->ac_pa; 5175c9de560dSAlex Tomas 51767e170922SOjaswin Mujoo if (ac->ac_b_ex.fe_len < ac->ac_orig_goal_len) { 5177bc056e71SBaokun Li struct ext4_free_extent ex = { 5178bc056e71SBaokun Li .fe_logical = ac->ac_g_ex.fe_logical, 5179bc056e71SBaokun Li .fe_len = ac->ac_orig_goal_len, 5180bc056e71SBaokun Li }; 5181bc056e71SBaokun Li loff_t orig_goal_end = extent_logical_end(sbi, &ex); 51824fbf8bc7SBaokun Li loff_t o_ex_end = extent_logical_end(sbi, &ac->ac_o_ex); 5183c9de560dSAlex Tomas 51844fbf8bc7SBaokun Li /* 51854fbf8bc7SBaokun Li * We can't allocate as much as normalizer wants, so we try 51864fbf8bc7SBaokun Li * to get proper lstart to cover the original request, except 51874fbf8bc7SBaokun Li * when the goal doesn't cover the original request as below: 51884fbf8bc7SBaokun Li * 51894fbf8bc7SBaokun Li * orig_ex:2045/2055(10), isize:8417280 -> normalized:0/2048 51904fbf8bc7SBaokun Li * best_ex:0/200(200) -> adjusted: 1848/2048(200) 51914fbf8bc7SBaokun Li */ 5192c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 5193c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 5194c9de560dSAlex Tomas 519593cdf49fSOjaswin Mujoo /* 519693cdf49fSOjaswin Mujoo * Use the below logic for adjusting best extent as it keeps 519793cdf49fSOjaswin Mujoo * fragmentation in check while ensuring logical range of best 519893cdf49fSOjaswin Mujoo * extent doesn't overflow out of goal extent: 519993cdf49fSOjaswin Mujoo * 52007e170922SOjaswin Mujoo * 1. Check if best ex can be kept at end of goal (before 52017e170922SOjaswin Mujoo * cr_best_avail trimmed it) and still cover original start 520293cdf49fSOjaswin Mujoo * 2. Else, check if best ex can be kept at start of goal and 52034fbf8bc7SBaokun Li * still cover original end 520493cdf49fSOjaswin Mujoo * 3. Else, keep the best ex at start of original request. 520593cdf49fSOjaswin Mujoo */ 5206bc056e71SBaokun Li ex.fe_len = ac->ac_b_ex.fe_len; 5207bc056e71SBaokun Li 5208bc056e71SBaokun Li ex.fe_logical = orig_goal_end - EXT4_C2B(sbi, ex.fe_len); 5209bc056e71SBaokun Li if (ac->ac_o_ex.fe_logical >= ex.fe_logical) 521093cdf49fSOjaswin Mujoo goto adjust_bex; 5211c9de560dSAlex Tomas 5212bc056e71SBaokun Li ex.fe_logical = ac->ac_g_ex.fe_logical; 52134fbf8bc7SBaokun Li if (o_ex_end <= extent_logical_end(sbi, &ex)) 521493cdf49fSOjaswin Mujoo goto adjust_bex; 5215c9de560dSAlex Tomas 5216bc056e71SBaokun Li ex.fe_logical = ac->ac_o_ex.fe_logical; 521793cdf49fSOjaswin Mujoo adjust_bex: 5218bc056e71SBaokun Li ac->ac_b_ex.fe_logical = ex.fe_logical; 5219c9de560dSAlex Tomas 5220c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 5221bc056e71SBaokun Li BUG_ON(extent_logical_end(sbi, &ex) > orig_goal_end); 5222c9de560dSAlex Tomas } 5223c9de560dSAlex Tomas 5224c9de560dSAlex Tomas pa->pa_lstart = ac->ac_b_ex.fe_logical; 5225c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 5226c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 5227c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 5228c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 5229d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 5230c9de560dSAlex Tomas pa->pa_deleted = 0; 5231cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_INODE_PA; 5232c9de560dSAlex Tomas 5233d3df1453SRitesh Harjani mb_debug(sb, "new inode pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, 5234d3df1453SRitesh Harjani pa->pa_len, pa->pa_lstart); 52359bffad1eSTheodore Ts'o trace_ext4_mb_new_inode_pa(ac, pa); 5236c9de560dSAlex Tomas 523753accfa9STheodore Ts'o atomic_add(pa->pa_free, &sbi->s_mb_preallocated); 5238abc075d4SKemeng Shi ext4_mb_use_inode_pa(ac, pa); 5239c9de560dSAlex Tomas 5240c9de560dSAlex Tomas ei = EXT4_I(ac->ac_inode); 5241c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 52425354b2afSTheodore Ts'o if (!grp) 52435354b2afSTheodore Ts'o return; 5244c9de560dSAlex Tomas 5245a8e38fd3SOjaswin Mujoo pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock; 5246c9de560dSAlex Tomas pa->pa_inode = ac->ac_inode; 5247c9de560dSAlex Tomas 5248c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 5249c9de560dSAlex Tomas 525038727786SOjaswin Mujoo write_lock(pa->pa_node_lock.inode_lock); 525138727786SOjaswin Mujoo ext4_mb_pa_rb_insert(&ei->i_prealloc_node, &pa->pa_node.inode_node); 525238727786SOjaswin Mujoo write_unlock(pa->pa_node_lock.inode_lock); 525327bc446eSbrookxu atomic_inc(&ei->i_prealloc_active); 5254c9de560dSAlex Tomas } 5255c9de560dSAlex Tomas 5256c9de560dSAlex Tomas /* 5257c9de560dSAlex Tomas * creates new preallocated space for locality group inodes belongs to 5258c9de560dSAlex Tomas */ 525953f86b17SRitesh Harjani static noinline_for_stack void 52604ddfef7bSEric Sandeen ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 5261c9de560dSAlex Tomas { 5262c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 5263c9de560dSAlex Tomas struct ext4_locality_group *lg; 5264c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 5265c9de560dSAlex Tomas struct ext4_group_info *grp; 5266c9de560dSAlex Tomas 5267c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 5268c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 5269c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 5270c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 527153f86b17SRitesh Harjani BUG_ON(ac->ac_pa == NULL); 5272c9de560dSAlex Tomas 527353f86b17SRitesh Harjani pa = ac->ac_pa; 5274c9de560dSAlex Tomas 5275c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 5276c9de560dSAlex Tomas pa->pa_lstart = pa->pa_pstart; 5277c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 5278c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 5279c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 5280a8e38fd3SOjaswin Mujoo INIT_LIST_HEAD(&pa->pa_node.lg_list); 5281d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 5282c9de560dSAlex Tomas pa->pa_deleted = 0; 5283cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_GROUP_PA; 5284c9de560dSAlex Tomas 5285d3df1453SRitesh Harjani mb_debug(sb, "new group pa %p: %llu/%d for %u\n", pa, pa->pa_pstart, 5286d3df1453SRitesh Harjani pa->pa_len, pa->pa_lstart); 52879bffad1eSTheodore Ts'o trace_ext4_mb_new_group_pa(ac, pa); 5288c9de560dSAlex Tomas 5289c9de560dSAlex Tomas ext4_mb_use_group_pa(ac, pa); 5290c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 5291c9de560dSAlex Tomas 5292c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 52935354b2afSTheodore Ts'o if (!grp) 52945354b2afSTheodore Ts'o return; 5295c9de560dSAlex Tomas lg = ac->ac_lg; 5296c9de560dSAlex Tomas BUG_ON(lg == NULL); 5297c9de560dSAlex Tomas 5298a8e38fd3SOjaswin Mujoo pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock; 5299c9de560dSAlex Tomas pa->pa_inode = NULL; 5300c9de560dSAlex Tomas 5301c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 5302c9de560dSAlex Tomas 53036be2ded1SAneesh Kumar K.V /* 53046be2ded1SAneesh Kumar K.V * We will later add the new pa to the right bucket 53056be2ded1SAneesh Kumar K.V * after updating the pa_free in ext4_mb_release_context 53066be2ded1SAneesh Kumar K.V */ 5307c9de560dSAlex Tomas } 5308c9de560dSAlex Tomas 530953f86b17SRitesh Harjani static void ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 5310c9de560dSAlex Tomas { 5311c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 531253f86b17SRitesh Harjani ext4_mb_new_group_pa(ac); 5313c9de560dSAlex Tomas else 531453f86b17SRitesh Harjani ext4_mb_new_inode_pa(ac); 5315c9de560dSAlex Tomas } 5316c9de560dSAlex Tomas 5317c9de560dSAlex Tomas /* 5318c9de560dSAlex Tomas * finds all unused blocks in on-disk bitmap, frees them in 5319c9de560dSAlex Tomas * in-core bitmap and buddy. 5320c9de560dSAlex Tomas * @pa must be unlinked from inode and group lists, so that 5321c9de560dSAlex Tomas * nobody else can find/use it. 5322c9de560dSAlex Tomas * the caller MUST hold group/inode locks. 5323c9de560dSAlex Tomas * TODO: optimize the case when there are no in-core structures yet 5324c9de560dSAlex Tomas */ 5325820c2808SKemeng Shi static noinline_for_stack void 53264ddfef7bSEric Sandeen ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 53273e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 5328c9de560dSAlex Tomas { 5329c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 5330c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 5331498e5f24STheodore Ts'o unsigned int end; 5332498e5f24STheodore Ts'o unsigned int next; 5333c9de560dSAlex Tomas ext4_group_t group; 5334c9de560dSAlex Tomas ext4_grpblk_t bit; 5335ba80b101STheodore Ts'o unsigned long long grp_blk_start; 5336c9de560dSAlex Tomas int free = 0; 5337c9de560dSAlex Tomas 5338c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 5339c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 534053accfa9STheodore Ts'o grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); 5341c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 5342c9de560dSAlex Tomas end = bit + pa->pa_len; 5343c9de560dSAlex Tomas 5344c9de560dSAlex Tomas while (bit < end) { 5345ffad0a44SAneesh Kumar K.V bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 5346c9de560dSAlex Tomas if (bit >= end) 5347c9de560dSAlex Tomas break; 5348ffad0a44SAneesh Kumar K.V next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 5349d3df1453SRitesh Harjani mb_debug(sb, "free preallocated %u/%u in group %u\n", 53505a0790c2SAndi Kleen (unsigned) ext4_group_first_block_no(sb, group) + bit, 53515a0790c2SAndi Kleen (unsigned) next - bit, (unsigned) group); 5352c9de560dSAlex Tomas free += next - bit; 5353c9de560dSAlex Tomas 53543e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); 535553accfa9STheodore Ts'o trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + 535653accfa9STheodore Ts'o EXT4_C2B(sbi, bit)), 5357a9c667f8SLukas Czerner next - bit); 5358c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 5359c9de560dSAlex Tomas bit = next + 1; 5360c9de560dSAlex Tomas } 5361c9de560dSAlex Tomas if (free != pa->pa_free) { 53629d8b9ec4STheodore Ts'o ext4_msg(e4b->bd_sb, KERN_CRIT, 536336bad423SRitesh Harjani "pa %p: logic %lu, phys. %lu, len %d", 5364c9de560dSAlex Tomas pa, (unsigned long) pa->pa_lstart, 5365c9de560dSAlex Tomas (unsigned long) pa->pa_pstart, 536636bad423SRitesh Harjani pa->pa_len); 5367e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", 536826346ff6SAneesh Kumar K.V free, pa->pa_free); 5369e56eb659SAneesh Kumar K.V /* 5370e56eb659SAneesh Kumar K.V * pa is already deleted so we use the value obtained 5371e56eb659SAneesh Kumar K.V * from the bitmap and continue. 5372e56eb659SAneesh Kumar K.V */ 5373c9de560dSAlex Tomas } 5374c9de560dSAlex Tomas atomic_add(free, &sbi->s_mb_discarded); 5375c9de560dSAlex Tomas } 5376c9de560dSAlex Tomas 537720427949SKemeng Shi static noinline_for_stack void 53784ddfef7bSEric Sandeen ext4_mb_release_group_pa(struct ext4_buddy *e4b, 53793e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 5380c9de560dSAlex Tomas { 5381c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 5382c9de560dSAlex Tomas ext4_group_t group; 5383c9de560dSAlex Tomas ext4_grpblk_t bit; 5384c9de560dSAlex Tomas 538560e07cf5SYongqiang Yang trace_ext4_mb_release_group_pa(sb, pa); 5386c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 5387c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 5388463808f2STheodore Ts'o if (unlikely(group != e4b->bd_group && pa->pa_len != 0)) { 5389463808f2STheodore Ts'o ext4_warning(sb, "bad group: expected %u, group %u, pa_start %llu", 5390463808f2STheodore Ts'o e4b->bd_group, group, pa->pa_pstart); 539120427949SKemeng Shi return; 5392463808f2STheodore Ts'o } 5393c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 5394c9de560dSAlex Tomas atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 53953e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); 5396c9de560dSAlex Tomas } 5397c9de560dSAlex Tomas 5398c9de560dSAlex Tomas /* 5399c9de560dSAlex Tomas * releases all preallocations in given group 5400c9de560dSAlex Tomas * 5401c9de560dSAlex Tomas * first, we need to decide discard policy: 5402c9de560dSAlex Tomas * - when do we discard 5403c9de560dSAlex Tomas * 1) ENOSPC 5404c9de560dSAlex Tomas * - how many do we discard 5405c9de560dSAlex Tomas * 1) how many requested 5406c9de560dSAlex Tomas */ 54074ddfef7bSEric Sandeen static noinline_for_stack int 54084ddfef7bSEric Sandeen ext4_mb_discard_group_preallocations(struct super_block *sb, 54098c80fb31SChunguang Xu ext4_group_t group, int *busy) 5410c9de560dSAlex Tomas { 5411c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 5412c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 5413c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 54140f6bc579SRuan Jinjie LIST_HEAD(list); 5415c9de560dSAlex Tomas struct ext4_buddy e4b; 541638727786SOjaswin Mujoo struct ext4_inode_info *ei; 5417c9de560dSAlex Tomas int err; 54188c80fb31SChunguang Xu int free = 0; 5419c9de560dSAlex Tomas 54205354b2afSTheodore Ts'o if (!grp) 54215354b2afSTheodore Ts'o return 0; 5422d3df1453SRitesh Harjani mb_debug(sb, "discard preallocation for group %u\n", group); 5423c9de560dSAlex Tomas if (list_empty(&grp->bb_prealloc_list)) 5424bbc4ec77SRitesh Harjani goto out_dbg; 5425c9de560dSAlex Tomas 5426574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 54279008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 54289008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 542954d3adbcSTheodore Ts'o ext4_error_err(sb, -err, 543054d3adbcSTheodore Ts'o "Error %d reading block bitmap for %u", 54319008a58eSDarrick J. Wong err, group); 5432bbc4ec77SRitesh Harjani goto out_dbg; 5433c9de560dSAlex Tomas } 5434c9de560dSAlex Tomas 5435c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 5436ce89f46cSAneesh Kumar K.V if (err) { 54379651e6b2SKonstantin Khlebnikov ext4_warning(sb, "Error %d loading buddy information for %u", 54389651e6b2SKonstantin Khlebnikov err, group); 5439ce89f46cSAneesh Kumar K.V put_bh(bitmap_bh); 5440bbc4ec77SRitesh Harjani goto out_dbg; 5441ce89f46cSAneesh Kumar K.V } 5442c9de560dSAlex Tomas 5443c9de560dSAlex Tomas ext4_lock_group(sb, group); 5444c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, 5445c9de560dSAlex Tomas &grp->bb_prealloc_list, pa_group_list) { 5446c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 5447c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 5448c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 54498c80fb31SChunguang Xu *busy = 1; 5450c9de560dSAlex Tomas continue; 5451c9de560dSAlex Tomas } 5452c9de560dSAlex Tomas if (pa->pa_deleted) { 5453c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5454c9de560dSAlex Tomas continue; 5455c9de560dSAlex Tomas } 5456c9de560dSAlex Tomas 5457c9de560dSAlex Tomas /* seems this one can be freed ... */ 545827bc446eSbrookxu ext4_mb_mark_pa_deleted(sb, pa); 5459c9de560dSAlex Tomas 546070022da8SYe Bin if (!free) 546170022da8SYe Bin this_cpu_inc(discard_pa_seq); 546270022da8SYe Bin 5463c9de560dSAlex Tomas /* we can trust pa_free ... */ 5464c9de560dSAlex Tomas free += pa->pa_free; 5465c9de560dSAlex Tomas 5466c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5467c9de560dSAlex Tomas 5468c9de560dSAlex Tomas list_del(&pa->pa_group_list); 5469c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 5470c9de560dSAlex Tomas } 5471c9de560dSAlex Tomas 5472c9de560dSAlex Tomas /* now free all selected PAs */ 5473c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 5474c9de560dSAlex Tomas 5475c9de560dSAlex Tomas /* remove from object (inode or locality group) */ 5476a8e38fd3SOjaswin Mujoo if (pa->pa_type == MB_GROUP_PA) { 5477a8e38fd3SOjaswin Mujoo spin_lock(pa->pa_node_lock.lg_lock); 5478a8e38fd3SOjaswin Mujoo list_del_rcu(&pa->pa_node.lg_list); 5479a8e38fd3SOjaswin Mujoo spin_unlock(pa->pa_node_lock.lg_lock); 5480a8e38fd3SOjaswin Mujoo } else { 548138727786SOjaswin Mujoo write_lock(pa->pa_node_lock.inode_lock); 548238727786SOjaswin Mujoo ei = EXT4_I(pa->pa_inode); 548338727786SOjaswin Mujoo rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 548438727786SOjaswin Mujoo write_unlock(pa->pa_node_lock.inode_lock); 5485a8e38fd3SOjaswin Mujoo } 5486c9de560dSAlex Tomas 5487c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 548838727786SOjaswin Mujoo 548938727786SOjaswin Mujoo if (pa->pa_type == MB_GROUP_PA) { 549038727786SOjaswin Mujoo ext4_mb_release_group_pa(&e4b, pa); 5491c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 549238727786SOjaswin Mujoo } else { 549338727786SOjaswin Mujoo ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 549438727786SOjaswin Mujoo ext4_mb_pa_free(pa); 549538727786SOjaswin Mujoo } 5496c9de560dSAlex Tomas } 5497c9de560dSAlex Tomas 5498c9de560dSAlex Tomas ext4_unlock_group(sb, group); 5499e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 5500c9de560dSAlex Tomas put_bh(bitmap_bh); 5501bbc4ec77SRitesh Harjani out_dbg: 5502d3df1453SRitesh Harjani mb_debug(sb, "discarded (%d) blocks preallocated for group %u bb_free (%d)\n", 55038c80fb31SChunguang Xu free, group, grp->bb_free); 55048c80fb31SChunguang Xu return free; 5505c9de560dSAlex Tomas } 5506c9de560dSAlex Tomas 5507c9de560dSAlex Tomas /* 5508c9de560dSAlex Tomas * releases all non-used preallocated blocks for given inode 5509c9de560dSAlex Tomas * 5510c9de560dSAlex Tomas * It's important to discard preallocations under i_data_sem 5511c9de560dSAlex Tomas * We don't want another block to be served from the prealloc 5512c9de560dSAlex Tomas * space when we are discarding the inode prealloc space. 5513c9de560dSAlex Tomas * 5514c9de560dSAlex Tomas * FIXME!! Make sure it is valid at all the call sites 5515c9de560dSAlex Tomas */ 55162ffd2a6aSKemeng Shi void ext4_discard_preallocations(struct inode *inode) 5517c9de560dSAlex Tomas { 5518c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(inode); 5519c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 5520c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 5521c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 5522c9de560dSAlex Tomas ext4_group_t group = 0; 55230f6bc579SRuan Jinjie LIST_HEAD(list); 5524c9de560dSAlex Tomas struct ext4_buddy e4b; 552538727786SOjaswin Mujoo struct rb_node *iter; 5526c9de560dSAlex Tomas int err; 5527c9de560dSAlex Tomas 5528f0e54b60SKemeng Shi if (!S_ISREG(inode->i_mode)) 5529c9de560dSAlex Tomas return; 5530c9de560dSAlex Tomas 55318016e29fSHarshad Shirwadkar if (EXT4_SB(sb)->s_mount_state & EXT4_FC_REPLAY) 55328016e29fSHarshad Shirwadkar return; 55338016e29fSHarshad Shirwadkar 5534d3df1453SRitesh Harjani mb_debug(sb, "discard preallocation for inode %lu\n", 5535d3df1453SRitesh Harjani inode->i_ino); 553627bc446eSbrookxu trace_ext4_discard_preallocations(inode, 5537f0e54b60SKemeng Shi atomic_read(&ei->i_prealloc_active)); 553827bc446eSbrookxu 5539c9de560dSAlex Tomas repeat: 5540c9de560dSAlex Tomas /* first, collect all pa's in the inode */ 554138727786SOjaswin Mujoo write_lock(&ei->i_prealloc_lock); 55422ffd2a6aSKemeng Shi for (iter = rb_first(&ei->i_prealloc_node); iter; 554338727786SOjaswin Mujoo iter = rb_next(iter)) { 554438727786SOjaswin Mujoo pa = rb_entry(iter, struct ext4_prealloc_space, 554538727786SOjaswin Mujoo pa_node.inode_node); 5546a8e38fd3SOjaswin Mujoo BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock); 554738727786SOjaswin Mujoo 5548c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 5549c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 5550c9de560dSAlex Tomas /* this shouldn't happen often - nobody should 5551c9de560dSAlex Tomas * use preallocation while we're discarding it */ 5552c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 555338727786SOjaswin Mujoo write_unlock(&ei->i_prealloc_lock); 55549d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, 55559d8b9ec4STheodore Ts'o "uh-oh! used pa while discarding"); 5556c9de560dSAlex Tomas WARN_ON(1); 5557c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 5558c9de560dSAlex Tomas goto repeat; 5559c9de560dSAlex Tomas 5560c9de560dSAlex Tomas } 5561c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 556227bc446eSbrookxu ext4_mb_mark_pa_deleted(sb, pa); 5563c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 556438727786SOjaswin Mujoo rb_erase(&pa->pa_node.inode_node, &ei->i_prealloc_node); 5565c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 5566c9de560dSAlex Tomas continue; 5567c9de560dSAlex Tomas } 5568c9de560dSAlex Tomas 5569c9de560dSAlex Tomas /* someone is deleting pa right now */ 5570c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 557138727786SOjaswin Mujoo write_unlock(&ei->i_prealloc_lock); 5572c9de560dSAlex Tomas 5573c9de560dSAlex Tomas /* we have to wait here because pa_deleted 5574c9de560dSAlex Tomas * doesn't mean pa is already unlinked from 5575c9de560dSAlex Tomas * the list. as we might be called from 5576c9de560dSAlex Tomas * ->clear_inode() the inode will get freed 5577c9de560dSAlex Tomas * and concurrent thread which is unlinking 5578c9de560dSAlex Tomas * pa from inode's list may access already 5579c9de560dSAlex Tomas * freed memory, bad-bad-bad */ 5580c9de560dSAlex Tomas 5581c9de560dSAlex Tomas /* XXX: if this happens too often, we can 5582c9de560dSAlex Tomas * add a flag to force wait only in case 5583c9de560dSAlex Tomas * of ->clear_inode(), but not in case of 5584c9de560dSAlex Tomas * regular truncate */ 5585c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 5586c9de560dSAlex Tomas goto repeat; 5587c9de560dSAlex Tomas } 558838727786SOjaswin Mujoo write_unlock(&ei->i_prealloc_lock); 5589c9de560dSAlex Tomas 5590c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 5591cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_INODE_PA); 5592bd86298eSLukas Czerner group = ext4_get_group_number(sb, pa->pa_pstart); 5593c9de560dSAlex Tomas 55949651e6b2SKonstantin Khlebnikov err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 55959651e6b2SKonstantin Khlebnikov GFP_NOFS|__GFP_NOFAIL); 5596ce89f46cSAneesh Kumar K.V if (err) { 559754d3adbcSTheodore Ts'o ext4_error_err(sb, -err, "Error %d loading buddy information for %u", 55989651e6b2SKonstantin Khlebnikov err, group); 5599ce89f46cSAneesh Kumar K.V continue; 5600ce89f46cSAneesh Kumar K.V } 5601c9de560dSAlex Tomas 5602574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 56039008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 56049008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 560554d3adbcSTheodore Ts'o ext4_error_err(sb, -err, "Error %d reading block bitmap for %u", 56069008a58eSDarrick J. Wong err, group); 5607e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 5608ce89f46cSAneesh Kumar K.V continue; 5609c9de560dSAlex Tomas } 5610c9de560dSAlex Tomas 5611c9de560dSAlex Tomas ext4_lock_group(sb, group); 5612c9de560dSAlex Tomas list_del(&pa->pa_group_list); 56133e1e5f50SEric Sandeen ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 5614c9de560dSAlex Tomas ext4_unlock_group(sb, group); 5615c9de560dSAlex Tomas 5616e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 5617c9de560dSAlex Tomas put_bh(bitmap_bh); 5618c9de560dSAlex Tomas 5619c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 562038727786SOjaswin Mujoo ext4_mb_pa_free(pa); 5621c9de560dSAlex Tomas } 5622c9de560dSAlex Tomas } 5623c9de560dSAlex Tomas 562453f86b17SRitesh Harjani static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac) 562553f86b17SRitesh Harjani { 562653f86b17SRitesh Harjani struct ext4_prealloc_space *pa; 562753f86b17SRitesh Harjani 562853f86b17SRitesh Harjani BUG_ON(ext4_pspace_cachep == NULL); 562953f86b17SRitesh Harjani pa = kmem_cache_zalloc(ext4_pspace_cachep, GFP_NOFS); 563053f86b17SRitesh Harjani if (!pa) 563153f86b17SRitesh Harjani return -ENOMEM; 563253f86b17SRitesh Harjani atomic_set(&pa->pa_count, 1); 563353f86b17SRitesh Harjani ac->ac_pa = pa; 563453f86b17SRitesh Harjani return 0; 563553f86b17SRitesh Harjani } 563653f86b17SRitesh Harjani 563782089725SOjaswin Mujoo static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac) 563853f86b17SRitesh Harjani { 563953f86b17SRitesh Harjani struct ext4_prealloc_space *pa = ac->ac_pa; 564053f86b17SRitesh Harjani 564153f86b17SRitesh Harjani BUG_ON(!pa); 564253f86b17SRitesh Harjani ac->ac_pa = NULL; 564353f86b17SRitesh Harjani WARN_ON(!atomic_dec_and_test(&pa->pa_count)); 564482089725SOjaswin Mujoo /* 564582089725SOjaswin Mujoo * current function is only called due to an error or due to 564682089725SOjaswin Mujoo * len of found blocks < len of requested blocks hence the PA has not 564782089725SOjaswin Mujoo * been added to grp->bb_prealloc_list. So we don't need to lock it 564882089725SOjaswin Mujoo */ 564982089725SOjaswin Mujoo pa->pa_deleted = 1; 565082089725SOjaswin Mujoo ext4_mb_pa_free(pa); 565153f86b17SRitesh Harjani } 565253f86b17SRitesh Harjani 56536ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 5654e68cf40cSRitesh Harjani static inline void ext4_mb_show_pa(struct super_block *sb) 5655c9de560dSAlex Tomas { 5656e68cf40cSRitesh Harjani ext4_group_t i, ngroups; 5657c9de560dSAlex Tomas 565895257987SJan Kara if (ext4_forced_shutdown(sb)) 5659e3570639SEric Sandeen return; 5660e3570639SEric Sandeen 56618df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 5662d3df1453SRitesh Harjani mb_debug(sb, "groups: "); 56638df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 5664c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, i); 5665c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 5666c9de560dSAlex Tomas ext4_grpblk_t start; 5667c9de560dSAlex Tomas struct list_head *cur; 56685354b2afSTheodore Ts'o 56695354b2afSTheodore Ts'o if (!grp) 56705354b2afSTheodore Ts'o continue; 5671c9de560dSAlex Tomas ext4_lock_group(sb, i); 5672c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 5673c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, 5674c9de560dSAlex Tomas pa_group_list); 5675c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 5676c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 5677c9de560dSAlex Tomas NULL, &start); 5678c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 5679d3df1453SRitesh Harjani mb_debug(sb, "PA:%u:%d:%d\n", i, start, 5680d3df1453SRitesh Harjani pa->pa_len); 5681c9de560dSAlex Tomas } 568260bd63d1SSolofo Ramangalahy ext4_unlock_group(sb, i); 5683d3df1453SRitesh Harjani mb_debug(sb, "%u: %d/%d\n", i, grp->bb_free, 5684d3df1453SRitesh Harjani grp->bb_fragments); 5685c9de560dSAlex Tomas } 5686c9de560dSAlex Tomas } 5687e68cf40cSRitesh Harjani 5688e68cf40cSRitesh Harjani static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 5689e68cf40cSRitesh Harjani { 5690e68cf40cSRitesh Harjani struct super_block *sb = ac->ac_sb; 5691e68cf40cSRitesh Harjani 569295257987SJan Kara if (ext4_forced_shutdown(sb)) 5693e68cf40cSRitesh Harjani return; 5694e68cf40cSRitesh Harjani 5695d3df1453SRitesh Harjani mb_debug(sb, "Can't allocate:" 5696e68cf40cSRitesh Harjani " Allocation context details:"); 5697d3df1453SRitesh Harjani mb_debug(sb, "status %u flags 0x%x", 5698e68cf40cSRitesh Harjani ac->ac_status, ac->ac_flags); 5699d3df1453SRitesh Harjani mb_debug(sb, "orig %lu/%lu/%lu@%lu, " 5700e68cf40cSRitesh Harjani "goal %lu/%lu/%lu@%lu, " 5701e68cf40cSRitesh Harjani "best %lu/%lu/%lu@%lu cr %d", 5702e68cf40cSRitesh Harjani (unsigned long)ac->ac_o_ex.fe_group, 5703e68cf40cSRitesh Harjani (unsigned long)ac->ac_o_ex.fe_start, 5704e68cf40cSRitesh Harjani (unsigned long)ac->ac_o_ex.fe_len, 5705e68cf40cSRitesh Harjani (unsigned long)ac->ac_o_ex.fe_logical, 5706e68cf40cSRitesh Harjani (unsigned long)ac->ac_g_ex.fe_group, 5707e68cf40cSRitesh Harjani (unsigned long)ac->ac_g_ex.fe_start, 5708e68cf40cSRitesh Harjani (unsigned long)ac->ac_g_ex.fe_len, 5709e68cf40cSRitesh Harjani (unsigned long)ac->ac_g_ex.fe_logical, 5710e68cf40cSRitesh Harjani (unsigned long)ac->ac_b_ex.fe_group, 5711e68cf40cSRitesh Harjani (unsigned long)ac->ac_b_ex.fe_start, 5712e68cf40cSRitesh Harjani (unsigned long)ac->ac_b_ex.fe_len, 5713e68cf40cSRitesh Harjani (unsigned long)ac->ac_b_ex.fe_logical, 5714e68cf40cSRitesh Harjani (int)ac->ac_criteria); 5715d3df1453SRitesh Harjani mb_debug(sb, "%u found", ac->ac_found); 5716569f196fSRitesh Harjani mb_debug(sb, "used pa: %s, ", ac->ac_pa ? "yes" : "no"); 5717569f196fSRitesh Harjani if (ac->ac_pa) 5718569f196fSRitesh Harjani mb_debug(sb, "pa_type %s\n", ac->ac_pa->pa_type == MB_GROUP_PA ? 5719569f196fSRitesh Harjani "group pa" : "inode pa"); 5720e68cf40cSRitesh Harjani ext4_mb_show_pa(sb); 5721e68cf40cSRitesh Harjani } 5722c9de560dSAlex Tomas #else 5723e68cf40cSRitesh Harjani static inline void ext4_mb_show_pa(struct super_block *sb) 5724e68cf40cSRitesh Harjani { 5725e68cf40cSRitesh Harjani } 5726c9de560dSAlex Tomas static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 5727c9de560dSAlex Tomas { 5728e68cf40cSRitesh Harjani ext4_mb_show_pa(ac->ac_sb); 5729c9de560dSAlex Tomas } 5730c9de560dSAlex Tomas #endif 5731c9de560dSAlex Tomas 5732c9de560dSAlex Tomas /* 5733c9de560dSAlex Tomas * We use locality group preallocation for small size file. The size of the 5734c9de560dSAlex Tomas * file is determined by the current size or the resulting size after 5735c9de560dSAlex Tomas * allocation which ever is larger 5736c9de560dSAlex Tomas * 5737b713a5ecSTheodore Ts'o * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 5738c9de560dSAlex Tomas */ 5739c9de560dSAlex Tomas static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 5740c9de560dSAlex Tomas { 5741c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 5742c9de560dSAlex Tomas int bsbits = ac->ac_sb->s_blocksize_bits; 5743c9de560dSAlex Tomas loff_t size, isize; 5744a9f2a293SJan Kara bool inode_pa_eligible, group_pa_eligible; 5745c9de560dSAlex Tomas 5746c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 5747c9de560dSAlex Tomas return; 5748c9de560dSAlex Tomas 57494ba74d00STheodore Ts'o if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 57504ba74d00STheodore Ts'o return; 57514ba74d00STheodore Ts'o 5752a9f2a293SJan Kara group_pa_eligible = sbi->s_mb_group_prealloc > 0; 5753a9f2a293SJan Kara inode_pa_eligible = true; 575443bbddc0SBaokun Li size = extent_logical_end(sbi, &ac->ac_o_ex); 575550797481STheodore Ts'o isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 575650797481STheodore Ts'o >> bsbits; 5757c9de560dSAlex Tomas 5758a9f2a293SJan Kara /* No point in using inode preallocation for closed files */ 575982dd124cSNikolay Borisov if ((size == isize) && !ext4_fs_is_busy(sbi) && 5760a9f2a293SJan Kara !inode_is_open_for_write(ac->ac_inode)) 5761a9f2a293SJan Kara inode_pa_eligible = false; 576250797481STheodore Ts'o 576371780577STheodore Ts'o size = max(size, isize); 5764a9f2a293SJan Kara /* Don't use group allocation for large files */ 5765a9f2a293SJan Kara if (size > sbi->s_mb_stream_request) 5766a9f2a293SJan Kara group_pa_eligible = false; 5767a9f2a293SJan Kara 5768a9f2a293SJan Kara if (!group_pa_eligible) { 5769a9f2a293SJan Kara if (inode_pa_eligible) 57704ba74d00STheodore Ts'o ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 5771a9f2a293SJan Kara else 5772a9f2a293SJan Kara ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 5773c9de560dSAlex Tomas return; 57744ba74d00STheodore Ts'o } 5775c9de560dSAlex Tomas 5776c9de560dSAlex Tomas BUG_ON(ac->ac_lg != NULL); 5777c9de560dSAlex Tomas /* 5778c9de560dSAlex Tomas * locality group prealloc space are per cpu. The reason for having 5779c9de560dSAlex Tomas * per cpu locality group is to reduce the contention between block 5780c9de560dSAlex Tomas * request from multiple CPUs. 5781c9de560dSAlex Tomas */ 5782a0b6bc63SChristoph Lameter ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); 5783c9de560dSAlex Tomas 5784c9de560dSAlex Tomas /* we're going to use group allocation */ 5785c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 5786c9de560dSAlex Tomas 5787c9de560dSAlex Tomas /* serialize all allocations in the group */ 5788c9de560dSAlex Tomas mutex_lock(&ac->ac_lg->lg_mutex); 5789c9de560dSAlex Tomas } 5790c9de560dSAlex Tomas 5791d73eff68SGuoqing Jiang static noinline_for_stack void 57924ddfef7bSEric Sandeen ext4_mb_initialize_context(struct ext4_allocation_context *ac, 5793c9de560dSAlex Tomas struct ext4_allocation_request *ar) 5794c9de560dSAlex Tomas { 5795c9de560dSAlex Tomas struct super_block *sb = ar->inode->i_sb; 5796c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 5797c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 5798c9de560dSAlex Tomas ext4_group_t group; 5799498e5f24STheodore Ts'o unsigned int len; 5800498e5f24STheodore Ts'o ext4_fsblk_t goal; 5801c9de560dSAlex Tomas ext4_grpblk_t block; 5802c9de560dSAlex Tomas 5803c9de560dSAlex Tomas /* we can't allocate > group size */ 5804c9de560dSAlex Tomas len = ar->len; 5805c9de560dSAlex Tomas 5806c9de560dSAlex Tomas /* just a dirty hack to filter too big requests */ 580740ae3487STheodore Ts'o if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) 580840ae3487STheodore Ts'o len = EXT4_CLUSTERS_PER_GROUP(sb); 5809c9de560dSAlex Tomas 5810c9de560dSAlex Tomas /* start searching from the goal */ 5811c9de560dSAlex Tomas goal = ar->goal; 5812c9de560dSAlex Tomas if (goal < le32_to_cpu(es->s_first_data_block) || 5813c9de560dSAlex Tomas goal >= ext4_blocks_count(es)) 5814c9de560dSAlex Tomas goal = le32_to_cpu(es->s_first_data_block); 5815c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, goal, &group, &block); 5816c9de560dSAlex Tomas 5817c9de560dSAlex Tomas /* set up allocation goals */ 5818f5a44db5STheodore Ts'o ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); 5819c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 5820c9de560dSAlex Tomas ac->ac_sb = sb; 5821c9de560dSAlex Tomas ac->ac_inode = ar->inode; 582253accfa9STheodore Ts'o ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; 5823c9de560dSAlex Tomas ac->ac_o_ex.fe_group = group; 5824c9de560dSAlex Tomas ac->ac_o_ex.fe_start = block; 5825c9de560dSAlex Tomas ac->ac_o_ex.fe_len = len; 582653accfa9STheodore Ts'o ac->ac_g_ex = ac->ac_o_ex; 58277e170922SOjaswin Mujoo ac->ac_orig_goal_len = ac->ac_g_ex.fe_len; 5828c9de560dSAlex Tomas ac->ac_flags = ar->flags; 5829c9de560dSAlex Tomas 58303cb77bd2Sbrookxu /* we have to define context: we'll work with a file or 5831c9de560dSAlex Tomas * locality group. this is a policy, actually */ 5832c9de560dSAlex Tomas ext4_mb_group_or_file(ac); 5833c9de560dSAlex Tomas 5834d3df1453SRitesh Harjani mb_debug(sb, "init ac: %u blocks @ %u, goal %u, flags 0x%x, 2^%d, " 5835c9de560dSAlex Tomas "left: %u/%u, right %u/%u to %swritable\n", 5836c9de560dSAlex Tomas (unsigned) ar->len, (unsigned) ar->logical, 5837c9de560dSAlex Tomas (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 5838c9de560dSAlex Tomas (unsigned) ar->lleft, (unsigned) ar->pleft, 5839c9de560dSAlex Tomas (unsigned) ar->lright, (unsigned) ar->pright, 584082dd124cSNikolay Borisov inode_is_open_for_write(ar->inode) ? "" : "non-"); 5841c9de560dSAlex Tomas } 5842c9de560dSAlex Tomas 58436be2ded1SAneesh Kumar K.V static noinline_for_stack void 58446be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(struct super_block *sb, 58456be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg, 58466be2ded1SAneesh Kumar K.V int order, int total_entries) 58476be2ded1SAneesh Kumar K.V { 58486be2ded1SAneesh Kumar K.V ext4_group_t group = 0; 58496be2ded1SAneesh Kumar K.V struct ext4_buddy e4b; 58500f6bc579SRuan Jinjie LIST_HEAD(discard_list); 58516be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa, *tmp; 58526be2ded1SAneesh Kumar K.V 5853d3df1453SRitesh Harjani mb_debug(sb, "discard locality group preallocation\n"); 58546be2ded1SAneesh Kumar K.V 58556be2ded1SAneesh Kumar K.V spin_lock(&lg->lg_prealloc_lock); 58566be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 5857a8e38fd3SOjaswin Mujoo pa_node.lg_list, 585892e9c58cSMadhuparna Bhowmik lockdep_is_held(&lg->lg_prealloc_lock)) { 58596be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 58606be2ded1SAneesh Kumar K.V if (atomic_read(&pa->pa_count)) { 58616be2ded1SAneesh Kumar K.V /* 58626be2ded1SAneesh Kumar K.V * This is the pa that we just used 58636be2ded1SAneesh Kumar K.V * for block allocation. So don't 58646be2ded1SAneesh Kumar K.V * free that 58656be2ded1SAneesh Kumar K.V */ 58666be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 58676be2ded1SAneesh Kumar K.V continue; 58686be2ded1SAneesh Kumar K.V } 58696be2ded1SAneesh Kumar K.V if (pa->pa_deleted) { 58706be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 58716be2ded1SAneesh Kumar K.V continue; 58726be2ded1SAneesh Kumar K.V } 58736be2ded1SAneesh Kumar K.V /* only lg prealloc space */ 5874cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_GROUP_PA); 58756be2ded1SAneesh Kumar K.V 58766be2ded1SAneesh Kumar K.V /* seems this one can be freed ... */ 587727bc446eSbrookxu ext4_mb_mark_pa_deleted(sb, pa); 58786be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 58796be2ded1SAneesh Kumar K.V 5880a8e38fd3SOjaswin Mujoo list_del_rcu(&pa->pa_node.lg_list); 58816be2ded1SAneesh Kumar K.V list_add(&pa->u.pa_tmp_list, &discard_list); 58826be2ded1SAneesh Kumar K.V 58836be2ded1SAneesh Kumar K.V total_entries--; 58846be2ded1SAneesh Kumar K.V if (total_entries <= 5) { 58856be2ded1SAneesh Kumar K.V /* 58866be2ded1SAneesh Kumar K.V * we want to keep only 5 entries 58876be2ded1SAneesh Kumar K.V * allowing it to grow to 8. This 58886be2ded1SAneesh Kumar K.V * mak sure we don't call discard 58896be2ded1SAneesh Kumar K.V * soon for this list. 58906be2ded1SAneesh Kumar K.V */ 58916be2ded1SAneesh Kumar K.V break; 58926be2ded1SAneesh Kumar K.V } 58936be2ded1SAneesh Kumar K.V } 58946be2ded1SAneesh Kumar K.V spin_unlock(&lg->lg_prealloc_lock); 58956be2ded1SAneesh Kumar K.V 58966be2ded1SAneesh Kumar K.V list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 58979651e6b2SKonstantin Khlebnikov int err; 58986be2ded1SAneesh Kumar K.V 5899bd86298eSLukas Czerner group = ext4_get_group_number(sb, pa->pa_pstart); 59009651e6b2SKonstantin Khlebnikov err = ext4_mb_load_buddy_gfp(sb, group, &e4b, 59019651e6b2SKonstantin Khlebnikov GFP_NOFS|__GFP_NOFAIL); 59029651e6b2SKonstantin Khlebnikov if (err) { 590354d3adbcSTheodore Ts'o ext4_error_err(sb, -err, "Error %d loading buddy information for %u", 59049651e6b2SKonstantin Khlebnikov err, group); 59056be2ded1SAneesh Kumar K.V continue; 59066be2ded1SAneesh Kumar K.V } 59076be2ded1SAneesh Kumar K.V ext4_lock_group(sb, group); 59086be2ded1SAneesh Kumar K.V list_del(&pa->pa_group_list); 59093e1e5f50SEric Sandeen ext4_mb_release_group_pa(&e4b, pa); 59106be2ded1SAneesh Kumar K.V ext4_unlock_group(sb, group); 59116be2ded1SAneesh Kumar K.V 5912e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 59136be2ded1SAneesh Kumar K.V list_del(&pa->u.pa_tmp_list); 59146be2ded1SAneesh Kumar K.V call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 59156be2ded1SAneesh Kumar K.V } 59166be2ded1SAneesh Kumar K.V } 59176be2ded1SAneesh Kumar K.V 59186be2ded1SAneesh Kumar K.V /* 59196be2ded1SAneesh Kumar K.V * We have incremented pa_count. So it cannot be freed at this 59206be2ded1SAneesh Kumar K.V * point. Also we hold lg_mutex. So no parallel allocation is 59216be2ded1SAneesh Kumar K.V * possible from this lg. That means pa_free cannot be updated. 59226be2ded1SAneesh Kumar K.V * 59236be2ded1SAneesh Kumar K.V * A parallel ext4_mb_discard_group_preallocations is possible. 59246be2ded1SAneesh Kumar K.V * which can cause the lg_prealloc_list to be updated. 59256be2ded1SAneesh Kumar K.V */ 59266be2ded1SAneesh Kumar K.V 59276be2ded1SAneesh Kumar K.V static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 59286be2ded1SAneesh Kumar K.V { 59296be2ded1SAneesh Kumar K.V int order, added = 0, lg_prealloc_count = 1; 59306be2ded1SAneesh Kumar K.V struct super_block *sb = ac->ac_sb; 59316be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg = ac->ac_lg; 59326be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 59336be2ded1SAneesh Kumar K.V 59346be2ded1SAneesh Kumar K.V order = fls(pa->pa_free) - 1; 59356be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 59366be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 59376be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 59386be2ded1SAneesh Kumar K.V /* Add the prealloc space to lg */ 5939f1167009SNiu Yawei spin_lock(&lg->lg_prealloc_lock); 59406be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 5941a8e38fd3SOjaswin Mujoo pa_node.lg_list, 594292e9c58cSMadhuparna Bhowmik lockdep_is_held(&lg->lg_prealloc_lock)) { 59436be2ded1SAneesh Kumar K.V spin_lock(&tmp_pa->pa_lock); 59446be2ded1SAneesh Kumar K.V if (tmp_pa->pa_deleted) { 5945e7c9e3e9STheodore Ts'o spin_unlock(&tmp_pa->pa_lock); 59466be2ded1SAneesh Kumar K.V continue; 59476be2ded1SAneesh Kumar K.V } 59486be2ded1SAneesh Kumar K.V if (!added && pa->pa_free < tmp_pa->pa_free) { 59496be2ded1SAneesh Kumar K.V /* Add to the tail of the previous entry */ 5950a8e38fd3SOjaswin Mujoo list_add_tail_rcu(&pa->pa_node.lg_list, 5951a8e38fd3SOjaswin Mujoo &tmp_pa->pa_node.lg_list); 59526be2ded1SAneesh Kumar K.V added = 1; 59536be2ded1SAneesh Kumar K.V /* 59546be2ded1SAneesh Kumar K.V * we want to count the total 59556be2ded1SAneesh Kumar K.V * number of entries in the list 59566be2ded1SAneesh Kumar K.V */ 59576be2ded1SAneesh Kumar K.V } 59586be2ded1SAneesh Kumar K.V spin_unlock(&tmp_pa->pa_lock); 59596be2ded1SAneesh Kumar K.V lg_prealloc_count++; 59606be2ded1SAneesh Kumar K.V } 59616be2ded1SAneesh Kumar K.V if (!added) 5962a8e38fd3SOjaswin Mujoo list_add_tail_rcu(&pa->pa_node.lg_list, 59636be2ded1SAneesh Kumar K.V &lg->lg_prealloc_list[order]); 5964f1167009SNiu Yawei spin_unlock(&lg->lg_prealloc_lock); 59656be2ded1SAneesh Kumar K.V 59666be2ded1SAneesh Kumar K.V /* Now trim the list to be not more than 8 elements */ 5967ad635507SKemeng Shi if (lg_prealloc_count > 8) 59686be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(sb, lg, 59696be2ded1SAneesh Kumar K.V order, lg_prealloc_count); 59706be2ded1SAneesh Kumar K.V } 59716be2ded1SAneesh Kumar K.V 5972c9de560dSAlex Tomas /* 5973c9de560dSAlex Tomas * release all resource we used in allocation 5974c9de560dSAlex Tomas */ 597511fd1a5dSKemeng Shi static void ext4_mb_release_context(struct ext4_allocation_context *ac) 5976c9de560dSAlex Tomas { 597753accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 59786be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa = ac->ac_pa; 59796be2ded1SAneesh Kumar K.V if (pa) { 5980cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) { 5981c9de560dSAlex Tomas /* see comment in ext4_mb_use_group_pa() */ 59826be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 598353accfa9STheodore Ts'o pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 598453accfa9STheodore Ts'o pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 59856be2ded1SAneesh Kumar K.V pa->pa_free -= ac->ac_b_ex.fe_len; 59866be2ded1SAneesh Kumar K.V pa->pa_len -= ac->ac_b_ex.fe_len; 59876be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 598866d5e027Sbrookxu 59896be2ded1SAneesh Kumar K.V /* 59906be2ded1SAneesh Kumar K.V * We want to add the pa to the right bucket. 59916be2ded1SAneesh Kumar K.V * Remove it from the list and while adding 59926be2ded1SAneesh Kumar K.V * make sure the list to which we are adding 599344183d42SAmir Goldstein * doesn't grow big. 59946be2ded1SAneesh Kumar K.V */ 599566d5e027Sbrookxu if (likely(pa->pa_free)) { 5996a8e38fd3SOjaswin Mujoo spin_lock(pa->pa_node_lock.lg_lock); 5997a8e38fd3SOjaswin Mujoo list_del_rcu(&pa->pa_node.lg_list); 5998a8e38fd3SOjaswin Mujoo spin_unlock(pa->pa_node_lock.lg_lock); 59996be2ded1SAneesh Kumar K.V ext4_mb_add_n_trim(ac); 6000c9de560dSAlex Tomas } 600166d5e027Sbrookxu } 600227bc446eSbrookxu 60036be2ded1SAneesh Kumar K.V ext4_mb_put_pa(ac, ac->ac_sb, pa); 6004c9de560dSAlex Tomas } 6005ccedf35bSMatthew Wilcox (Oracle) if (ac->ac_bitmap_folio) 6006ccedf35bSMatthew Wilcox (Oracle) folio_put(ac->ac_bitmap_folio); 6007c84f1510SMatthew Wilcox (Oracle) if (ac->ac_buddy_folio) 6008c84f1510SMatthew Wilcox (Oracle) folio_put(ac->ac_buddy_folio); 6009c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 6010c9de560dSAlex Tomas mutex_unlock(&ac->ac_lg->lg_mutex); 6011c9de560dSAlex Tomas ext4_mb_collect_stats(ac); 6012c9de560dSAlex Tomas } 6013c9de560dSAlex Tomas 6014c9de560dSAlex Tomas static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 6015c9de560dSAlex Tomas { 60168df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 6017c9de560dSAlex Tomas int ret; 60188c80fb31SChunguang Xu int freed = 0, busy = 0; 60198c80fb31SChunguang Xu int retry = 0; 6020c9de560dSAlex Tomas 60219bffad1eSTheodore Ts'o trace_ext4_mb_discard_preallocations(sb, needed); 60228c80fb31SChunguang Xu 60238c80fb31SChunguang Xu if (needed == 0) 60248c80fb31SChunguang Xu needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; 60258c80fb31SChunguang Xu repeat: 60268df9675fSTheodore Ts'o for (i = 0; i < ngroups && needed > 0; i++) { 60278c80fb31SChunguang Xu ret = ext4_mb_discard_group_preallocations(sb, i, &busy); 6028c9de560dSAlex Tomas freed += ret; 6029c9de560dSAlex Tomas needed -= ret; 60308c80fb31SChunguang Xu cond_resched(); 60318c80fb31SChunguang Xu } 60328c80fb31SChunguang Xu 60338c80fb31SChunguang Xu if (needed > 0 && busy && ++retry < 3) { 60348c80fb31SChunguang Xu busy = 0; 60358c80fb31SChunguang Xu goto repeat; 6036c9de560dSAlex Tomas } 6037c9de560dSAlex Tomas 6038c9de560dSAlex Tomas return freed; 6039c9de560dSAlex Tomas } 6040c9de560dSAlex Tomas 6041cf5e2ca6SRitesh Harjani static bool ext4_mb_discard_preallocations_should_retry(struct super_block *sb, 604207b5b8e1SRitesh Harjani struct ext4_allocation_context *ac, u64 *seq) 6043cf5e2ca6SRitesh Harjani { 6044cf5e2ca6SRitesh Harjani int freed; 604507b5b8e1SRitesh Harjani u64 seq_retry = 0; 604607b5b8e1SRitesh Harjani bool ret = false; 6047cf5e2ca6SRitesh Harjani 6048cf5e2ca6SRitesh Harjani freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 604907b5b8e1SRitesh Harjani if (freed) { 605007b5b8e1SRitesh Harjani ret = true; 605107b5b8e1SRitesh Harjani goto out_dbg; 605207b5b8e1SRitesh Harjani } 605307b5b8e1SRitesh Harjani seq_retry = ext4_get_discard_pa_seq_sum(); 605499377830SRitesh Harjani if (!(ac->ac_flags & EXT4_MB_STRICT_CHECK) || seq_retry != *seq) { 605599377830SRitesh Harjani ac->ac_flags |= EXT4_MB_STRICT_CHECK; 605607b5b8e1SRitesh Harjani *seq = seq_retry; 605707b5b8e1SRitesh Harjani ret = true; 605807b5b8e1SRitesh Harjani } 605907b5b8e1SRitesh Harjani 606007b5b8e1SRitesh Harjani out_dbg: 606107b5b8e1SRitesh Harjani mb_debug(sb, "freed %d, retry ? %s\n", freed, ret ? "yes" : "no"); 606207b5b8e1SRitesh Harjani return ret; 6063cf5e2ca6SRitesh Harjani } 6064cf5e2ca6SRitesh Harjani 6065ad78b5efSKemeng Shi /* 6066ad78b5efSKemeng Shi * Simple allocator for Ext4 fast commit replay path. It searches for blocks 6067ad78b5efSKemeng Shi * linearly starting at the goal block and also excludes the blocks which 6068ad78b5efSKemeng Shi * are going to be in use after fast commit replay. 6069ad78b5efSKemeng Shi */ 6070ad78b5efSKemeng Shi static ext4_fsblk_t 6071ad78b5efSKemeng Shi ext4_mb_new_blocks_simple(struct ext4_allocation_request *ar, int *errp) 6072ad78b5efSKemeng Shi { 6073ad78b5efSKemeng Shi struct buffer_head *bitmap_bh; 6074ad78b5efSKemeng Shi struct super_block *sb = ar->inode->i_sb; 6075ad78b5efSKemeng Shi struct ext4_sb_info *sbi = EXT4_SB(sb); 6076ad78b5efSKemeng Shi ext4_group_t group, nr; 6077ad78b5efSKemeng Shi ext4_grpblk_t blkoff; 6078ad78b5efSKemeng Shi ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 6079ad78b5efSKemeng Shi ext4_grpblk_t i = 0; 6080ad78b5efSKemeng Shi ext4_fsblk_t goal, block; 608179ebf48cSLu Hongfei struct ext4_super_block *es = sbi->s_es; 6082ad78b5efSKemeng Shi 6083ad78b5efSKemeng Shi goal = ar->goal; 6084ad78b5efSKemeng Shi if (goal < le32_to_cpu(es->s_first_data_block) || 6085ad78b5efSKemeng Shi goal >= ext4_blocks_count(es)) 6086ad78b5efSKemeng Shi goal = le32_to_cpu(es->s_first_data_block); 6087ad78b5efSKemeng Shi 6088ad78b5efSKemeng Shi ar->len = 0; 6089ad78b5efSKemeng Shi ext4_get_group_no_and_offset(sb, goal, &group, &blkoff); 6090ad78b5efSKemeng Shi for (nr = ext4_get_groups_count(sb); nr > 0; nr--) { 6091ad78b5efSKemeng Shi bitmap_bh = ext4_read_block_bitmap(sb, group); 6092ad78b5efSKemeng Shi if (IS_ERR(bitmap_bh)) { 6093ad78b5efSKemeng Shi *errp = PTR_ERR(bitmap_bh); 6094ad78b5efSKemeng Shi pr_warn("Failed to read block bitmap\n"); 6095ad78b5efSKemeng Shi return 0; 6096ad78b5efSKemeng Shi } 6097ad78b5efSKemeng Shi 6098ad78b5efSKemeng Shi while (1) { 6099ad78b5efSKemeng Shi i = mb_find_next_zero_bit(bitmap_bh->b_data, max, 6100ad78b5efSKemeng Shi blkoff); 6101ad78b5efSKemeng Shi if (i >= max) 6102ad78b5efSKemeng Shi break; 6103ad78b5efSKemeng Shi if (ext4_fc_replay_check_excluded(sb, 6104ad78b5efSKemeng Shi ext4_group_first_block_no(sb, group) + 6105ad78b5efSKemeng Shi EXT4_C2B(sbi, i))) { 6106ad78b5efSKemeng Shi blkoff = i + 1; 6107ad78b5efSKemeng Shi } else 6108ad78b5efSKemeng Shi break; 6109ad78b5efSKemeng Shi } 6110ad78b5efSKemeng Shi brelse(bitmap_bh); 6111ad78b5efSKemeng Shi if (i < max) 6112ad78b5efSKemeng Shi break; 6113ad78b5efSKemeng Shi 6114ad78b5efSKemeng Shi if (++group >= ext4_get_groups_count(sb)) 6115ad78b5efSKemeng Shi group = 0; 6116ad78b5efSKemeng Shi 6117ad78b5efSKemeng Shi blkoff = 0; 6118ad78b5efSKemeng Shi } 6119ad78b5efSKemeng Shi 6120ad78b5efSKemeng Shi if (i >= max) { 6121ad78b5efSKemeng Shi *errp = -ENOSPC; 6122ad78b5efSKemeng Shi return 0; 6123ad78b5efSKemeng Shi } 6124ad78b5efSKemeng Shi 6125ad78b5efSKemeng Shi block = ext4_group_first_block_no(sb, group) + EXT4_C2B(sbi, i); 6126d2f7cf40SKemeng Shi ext4_mb_mark_bb(sb, block, 1, true); 6127ad78b5efSKemeng Shi ar->len = 1; 6128ad78b5efSKemeng Shi 61293f4830abSDan Carpenter *errp = 0; 6130ad78b5efSKemeng Shi return block; 6131ad78b5efSKemeng Shi } 61328016e29fSHarshad Shirwadkar 6133c9de560dSAlex Tomas /* 6134c9de560dSAlex Tomas * Main entry point into mballoc to allocate blocks 6135c9de560dSAlex Tomas * it tries to use preallocation first, then falls back 6136c9de560dSAlex Tomas * to usual allocation 6137c9de560dSAlex Tomas */ 6138c9de560dSAlex Tomas ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 6139c9de560dSAlex Tomas struct ext4_allocation_request *ar, int *errp) 6140c9de560dSAlex Tomas { 6141256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 6142c9de560dSAlex Tomas struct ext4_sb_info *sbi; 6143c9de560dSAlex Tomas struct super_block *sb; 6144c9de560dSAlex Tomas ext4_fsblk_t block = 0; 614560e58e0fSMingming Cao unsigned int inquota = 0; 614653accfa9STheodore Ts'o unsigned int reserv_clstrs = 0; 614780fa46d6STheodore Ts'o int retries = 0; 614807b5b8e1SRitesh Harjani u64 seq; 6149c9de560dSAlex Tomas 6150b10a44c3STheodore Ts'o might_sleep(); 6151c9de560dSAlex Tomas sb = ar->inode->i_sb; 6152c9de560dSAlex Tomas sbi = EXT4_SB(sb); 6153c9de560dSAlex Tomas 61549bffad1eSTheodore Ts'o trace_ext4_request_blocks(ar); 61558016e29fSHarshad Shirwadkar if (sbi->s_mount_state & EXT4_FC_REPLAY) 6156ad78b5efSKemeng Shi return ext4_mb_new_blocks_simple(ar, errp); 6157ba80b101STheodore Ts'o 615845dc63e7SDmitry Monakhov /* Allow to use superuser reservation for quota file */ 615902749a4cSTahsin Erdogan if (ext4_is_quota_file(ar->inode)) 616045dc63e7SDmitry Monakhov ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; 616145dc63e7SDmitry Monakhov 6162e3cf5d5dSTheodore Ts'o if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { 616360e58e0fSMingming Cao /* Without delayed allocation we need to verify 616460e58e0fSMingming Cao * there is enough free blocks to do block allocation 616560e58e0fSMingming Cao * and verify allocation doesn't exceed the quota limits. 6166d2a17637SMingming Cao */ 616755f020dbSAllison Henderson while (ar->len && 6168e7d5f315STheodore Ts'o ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { 616955f020dbSAllison Henderson 6170030ba6bcSAneesh Kumar K.V /* let others to free the space */ 6171bb8b20edSLukas Czerner cond_resched(); 6172030ba6bcSAneesh Kumar K.V ar->len = ar->len >> 1; 6173030ba6bcSAneesh Kumar K.V } 6174030ba6bcSAneesh Kumar K.V if (!ar->len) { 6175bbc4ec77SRitesh Harjani ext4_mb_show_pa(sb); 617607031431SMingming Cao *errp = -ENOSPC; 617707031431SMingming Cao return 0; 617807031431SMingming Cao } 617953accfa9STheodore Ts'o reserv_clstrs = ar->len; 618055f020dbSAllison Henderson if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { 618153accfa9STheodore Ts'o dquot_alloc_block_nofail(ar->inode, 618253accfa9STheodore Ts'o EXT4_C2B(sbi, ar->len)); 618355f020dbSAllison Henderson } else { 618455f020dbSAllison Henderson while (ar->len && 618553accfa9STheodore Ts'o dquot_alloc_block(ar->inode, 618653accfa9STheodore Ts'o EXT4_C2B(sbi, ar->len))) { 618755f020dbSAllison Henderson 6188c9de560dSAlex Tomas ar->flags |= EXT4_MB_HINT_NOPREALLOC; 6189c9de560dSAlex Tomas ar->len--; 6190c9de560dSAlex Tomas } 619155f020dbSAllison Henderson } 619260e58e0fSMingming Cao inquota = ar->len; 6193c9de560dSAlex Tomas if (ar->len == 0) { 6194c9de560dSAlex Tomas *errp = -EDQUOT; 61956c7a120aSAditya Kali goto out; 6196c9de560dSAlex Tomas } 619760e58e0fSMingming Cao } 6198d2a17637SMingming Cao 619985556c9aSWei Yongjun ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); 6200833576b3STheodore Ts'o if (!ac) { 6201363d4251SShen Feng ar->len = 0; 6202256bdb49SEric Sandeen *errp = -ENOMEM; 62036c7a120aSAditya Kali goto out; 6204256bdb49SEric Sandeen } 6205256bdb49SEric Sandeen 6206d73eff68SGuoqing Jiang ext4_mb_initialize_context(ac, ar); 6207c9de560dSAlex Tomas 6208256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 620981198536SRitesh Harjani seq = this_cpu_read(discard_pa_seq); 6210256bdb49SEric Sandeen if (!ext4_mb_use_preallocated(ac)) { 6211256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_ALLOC; 6212256bdb49SEric Sandeen ext4_mb_normalize_request(ac, ar); 621353f86b17SRitesh Harjani 621453f86b17SRitesh Harjani *errp = ext4_mb_pa_alloc(ac); 621553f86b17SRitesh Harjani if (*errp) 621653f86b17SRitesh Harjani goto errout; 6217c9de560dSAlex Tomas repeat: 6218c9de560dSAlex Tomas /* allocate space in core */ 62196c7a120aSAditya Kali *errp = ext4_mb_regular_allocator(ac); 622053f86b17SRitesh Harjani /* 622153f86b17SRitesh Harjani * pa allocated above is added to grp->bb_prealloc_list only 622253f86b17SRitesh Harjani * when we were able to allocate some block i.e. when 622353f86b17SRitesh Harjani * ac->ac_status == AC_STATUS_FOUND. 622453f86b17SRitesh Harjani * And error from above mean ac->ac_status != AC_STATUS_FOUND 622553f86b17SRitesh Harjani * So we have to free this pa here itself. 622653f86b17SRitesh Harjani */ 62272c00ef3eSAlexey Khoroshilov if (*errp) { 622882089725SOjaswin Mujoo ext4_mb_pa_put_free(ac); 62292c00ef3eSAlexey Khoroshilov ext4_discard_allocated_blocks(ac); 62302c00ef3eSAlexey Khoroshilov goto errout; 62312c00ef3eSAlexey Khoroshilov } 623253f86b17SRitesh Harjani if (ac->ac_status == AC_STATUS_FOUND && 623353f86b17SRitesh Harjani ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len) 623482089725SOjaswin Mujoo ext4_mb_pa_put_free(ac); 6235c9de560dSAlex Tomas } 6236256bdb49SEric Sandeen if (likely(ac->ac_status == AC_STATUS_FOUND)) { 623753accfa9STheodore Ts'o *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); 6238554a5cccSVegard Nossum if (*errp) { 6239b844167eSCurt Wohlgemuth ext4_discard_allocated_blocks(ac); 62406d138cedSEric Sandeen goto errout; 62416d138cedSEric Sandeen } else { 6242256bdb49SEric Sandeen block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 6243256bdb49SEric Sandeen ar->len = ac->ac_b_ex.fe_len; 6244519deca0SAneesh Kumar K.V } 6245c9de560dSAlex Tomas } else { 624680fa46d6STheodore Ts'o if (++retries < 3 && 624780fa46d6STheodore Ts'o ext4_mb_discard_preallocations_should_retry(sb, ac, &seq)) 6248c9de560dSAlex Tomas goto repeat; 624953f86b17SRitesh Harjani /* 625053f86b17SRitesh Harjani * If block allocation fails then the pa allocated above 625153f86b17SRitesh Harjani * needs to be freed here itself. 625253f86b17SRitesh Harjani */ 625382089725SOjaswin Mujoo ext4_mb_pa_put_free(ac); 6254c9de560dSAlex Tomas *errp = -ENOSPC; 62556c7a120aSAditya Kali } 62566c7a120aSAditya Kali 62576c7a120aSAditya Kali if (*errp) { 6258aaae558dSKemeng Shi errout: 6259256bdb49SEric Sandeen ac->ac_b_ex.fe_len = 0; 6260c9de560dSAlex Tomas ar->len = 0; 6261256bdb49SEric Sandeen ext4_mb_show_ac(ac); 6262c9de560dSAlex Tomas } 6263256bdb49SEric Sandeen ext4_mb_release_context(ac); 6264363d4251SShen Feng kmem_cache_free(ext4_ac_cachep, ac); 6265aaae558dSKemeng Shi out: 626660e58e0fSMingming Cao if (inquota && ar->len < inquota) 626753accfa9STheodore Ts'o dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); 62680087d9fbSAneesh Kumar K.V if (!ar->len) { 6269e3cf5d5dSTheodore Ts'o if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) 62700087d9fbSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 627157042651STheodore Ts'o percpu_counter_sub(&sbi->s_dirtyclusters_counter, 627253accfa9STheodore Ts'o reserv_clstrs); 62730087d9fbSAneesh Kumar K.V } 6274c9de560dSAlex Tomas 62759bffad1eSTheodore Ts'o trace_ext4_allocate_blocks(ar, (unsigned long long)block); 6276ba80b101STheodore Ts'o 6277c9de560dSAlex Tomas return block; 6278c9de560dSAlex Tomas } 6279c9de560dSAlex Tomas 6280c894058dSAneesh Kumar K.V /* 6281c894058dSAneesh Kumar K.V * We can merge two free data extents only if the physical blocks 6282c894058dSAneesh Kumar K.V * are contiguous, AND the extents were freed by the same transaction, 6283c894058dSAneesh Kumar K.V * AND the blocks are associated with the same group. 6284c894058dSAneesh Kumar K.V */ 6285a0154344SDaeho Jeong static void ext4_try_merge_freed_extent(struct ext4_sb_info *sbi, 6286a0154344SDaeho Jeong struct ext4_free_data *entry, 6287a0154344SDaeho Jeong struct ext4_free_data *new_entry, 6288a0154344SDaeho Jeong struct rb_root *entry_rb_root) 6289c894058dSAneesh Kumar K.V { 6290a0154344SDaeho Jeong if ((entry->efd_tid != new_entry->efd_tid) || 6291a0154344SDaeho Jeong (entry->efd_group != new_entry->efd_group)) 6292a0154344SDaeho Jeong return; 6293a0154344SDaeho Jeong if (entry->efd_start_cluster + entry->efd_count == 6294a0154344SDaeho Jeong new_entry->efd_start_cluster) { 6295a0154344SDaeho Jeong new_entry->efd_start_cluster = entry->efd_start_cluster; 6296a0154344SDaeho Jeong new_entry->efd_count += entry->efd_count; 6297a0154344SDaeho Jeong } else if (new_entry->efd_start_cluster + new_entry->efd_count == 6298a0154344SDaeho Jeong entry->efd_start_cluster) { 6299a0154344SDaeho Jeong new_entry->efd_count += entry->efd_count; 6300a0154344SDaeho Jeong } else 6301a0154344SDaeho Jeong return; 6302a0154344SDaeho Jeong spin_lock(&sbi->s_md_lock); 6303a0154344SDaeho Jeong list_del(&entry->efd_list); 6304a0154344SDaeho Jeong spin_unlock(&sbi->s_md_lock); 6305a0154344SDaeho Jeong rb_erase(&entry->efd_node, entry_rb_root); 6306a0154344SDaeho Jeong kmem_cache_free(ext4_free_data_cachep, entry); 6307c894058dSAneesh Kumar K.V } 6308c894058dSAneesh Kumar K.V 630985b67ffbSKemeng Shi static noinline_for_stack void 63104ddfef7bSEric Sandeen ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 63117a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry) 6312c9de560dSAlex Tomas { 6313e29136f8STheodore Ts'o ext4_group_t group = e4b->bd_group; 631484130193STheodore Ts'o ext4_grpblk_t cluster; 6315d08854f5STheodore Ts'o ext4_grpblk_t clusters = new_entry->efd_count; 63167a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 6317c9de560dSAlex Tomas struct ext4_group_info *db = e4b->bd_info; 6318c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 6319c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 6320c894058dSAneesh Kumar K.V struct rb_node **n = &db->bb_free_root.rb_node, *node; 6321c894058dSAneesh Kumar K.V struct rb_node *parent = NULL, *new_node; 6322c894058dSAneesh Kumar K.V 63230390131bSFrank Mayhar BUG_ON(!ext4_handle_valid(handle)); 632499b150d8SMatthew Wilcox (Oracle) BUG_ON(e4b->bd_bitmap_folio == NULL); 63255eea586bSMatthew Wilcox (Oracle) BUG_ON(e4b->bd_buddy_folio == NULL); 6326c9de560dSAlex Tomas 632718aadd47SBobi Jam new_node = &new_entry->efd_node; 632818aadd47SBobi Jam cluster = new_entry->efd_start_cluster; 6329c9de560dSAlex Tomas 6330c894058dSAneesh Kumar K.V if (!*n) { 6331c894058dSAneesh Kumar K.V /* first free block exent. We need to 6332c894058dSAneesh Kumar K.V protect buddy cache from being freed, 6333c9de560dSAlex Tomas * otherwise we'll refresh it from 6334c9de560dSAlex Tomas * on-disk bitmap and lose not-yet-available 6335c9de560dSAlex Tomas * blocks */ 63365eea586bSMatthew Wilcox (Oracle) folio_get(e4b->bd_buddy_folio); 633799b150d8SMatthew Wilcox (Oracle) folio_get(e4b->bd_bitmap_folio); 6338c894058dSAneesh Kumar K.V } 6339c894058dSAneesh Kumar K.V while (*n) { 6340c894058dSAneesh Kumar K.V parent = *n; 634118aadd47SBobi Jam entry = rb_entry(parent, struct ext4_free_data, efd_node); 634218aadd47SBobi Jam if (cluster < entry->efd_start_cluster) 6343c894058dSAneesh Kumar K.V n = &(*n)->rb_left; 634418aadd47SBobi Jam else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) 6345c894058dSAneesh Kumar K.V n = &(*n)->rb_right; 6346c894058dSAneesh Kumar K.V else { 6347e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 634884130193STheodore Ts'o ext4_group_first_block_no(sb, group) + 634984130193STheodore Ts'o EXT4_C2B(sbi, cluster), 6350e29136f8STheodore Ts'o "Block already on to-be-freed list"); 6351cca41553SChunguang Xu kmem_cache_free(ext4_free_data_cachep, new_entry); 635285b67ffbSKemeng Shi return; 6353c9de560dSAlex Tomas } 6354c9de560dSAlex Tomas } 6355c9de560dSAlex Tomas 6356c894058dSAneesh Kumar K.V rb_link_node(new_node, parent, n); 6357c894058dSAneesh Kumar K.V rb_insert_color(new_node, &db->bb_free_root); 6358c894058dSAneesh Kumar K.V 6359c894058dSAneesh Kumar K.V /* Now try to see the extent can be merged to left and right */ 6360c894058dSAneesh Kumar K.V node = rb_prev(new_node); 6361c894058dSAneesh Kumar K.V if (node) { 636218aadd47SBobi Jam entry = rb_entry(node, struct ext4_free_data, efd_node); 6363a0154344SDaeho Jeong ext4_try_merge_freed_extent(sbi, entry, new_entry, 6364a0154344SDaeho Jeong &(db->bb_free_root)); 6365c9de560dSAlex Tomas } 6366c894058dSAneesh Kumar K.V 6367c894058dSAneesh Kumar K.V node = rb_next(new_node); 6368c894058dSAneesh Kumar K.V if (node) { 636918aadd47SBobi Jam entry = rb_entry(node, struct ext4_free_data, efd_node); 6370a0154344SDaeho Jeong ext4_try_merge_freed_extent(sbi, entry, new_entry, 6371a0154344SDaeho Jeong &(db->bb_free_root)); 6372c894058dSAneesh Kumar K.V } 6373a0154344SDaeho Jeong 6374d08854f5STheodore Ts'o spin_lock(&sbi->s_md_lock); 6375ce774e53SJinke Han list_add_tail(&new_entry->efd_list, &sbi->s_freed_data_list[new_entry->efd_tid & 1]); 6376d08854f5STheodore Ts'o sbi->s_mb_free_pending += clusters; 6377d08854f5STheodore Ts'o spin_unlock(&sbi->s_md_lock); 6378c9de560dSAlex Tomas } 6379c9de560dSAlex Tomas 63808016e29fSHarshad Shirwadkar static void ext4_free_blocks_simple(struct inode *inode, ext4_fsblk_t block, 63818016e29fSHarshad Shirwadkar unsigned long count) 63828016e29fSHarshad Shirwadkar { 63838016e29fSHarshad Shirwadkar struct super_block *sb = inode->i_sb; 63848016e29fSHarshad Shirwadkar ext4_group_t group; 63858016e29fSHarshad Shirwadkar ext4_grpblk_t blkoff; 63868016e29fSHarshad Shirwadkar 63878016e29fSHarshad Shirwadkar ext4_get_group_no_and_offset(sb, block, &group, &blkoff); 6388c431d386SKemeng Shi ext4_mb_mark_context(NULL, sb, false, group, blkoff, count, 6389c431d386SKemeng Shi EXT4_MB_BITMAP_MARKED_CHECK | 6390c431d386SKemeng Shi EXT4_MB_SYNC_UPDATE, 6391c431d386SKemeng Shi NULL); 63928016e29fSHarshad Shirwadkar } 63938016e29fSHarshad Shirwadkar 639444338711STheodore Ts'o /** 63958ac3939dSRitesh Harjani * ext4_mb_clear_bb() -- helper function for freeing blocks. 63968ac3939dSRitesh Harjani * Used by ext4_free_blocks() 639744338711STheodore Ts'o * @handle: handle for this transaction 639844338711STheodore Ts'o * @inode: inode 6399c60990b3STheodore Ts'o * @block: starting physical block to be freed 6400c60990b3STheodore Ts'o * @count: number of blocks to be freed 64015def1360SYongqiang Yang * @flags: flags used by ext4_free_blocks 6402c9de560dSAlex Tomas */ 64038ac3939dSRitesh Harjani static void ext4_mb_clear_bb(handle_t *handle, struct inode *inode, 64048ac3939dSRitesh Harjani ext4_fsblk_t block, unsigned long count, 64058ac3939dSRitesh Harjani int flags) 6406c9de560dSAlex Tomas { 6407c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 64085354b2afSTheodore Ts'o struct ext4_group_info *grp; 6409498e5f24STheodore Ts'o unsigned int overflow; 6410c9de560dSAlex Tomas ext4_grpblk_t bit; 6411c9de560dSAlex Tomas ext4_group_t block_group; 6412c9de560dSAlex Tomas struct ext4_sb_info *sbi; 6413c9de560dSAlex Tomas struct ext4_buddy e4b; 641484130193STheodore Ts'o unsigned int count_clusters; 6415c9de560dSAlex Tomas int err = 0; 641638b8f70cSKemeng Shi int mark_flags = 0; 641738b8f70cSKemeng Shi ext4_grpblk_t changed; 6418c9de560dSAlex Tomas 64198016e29fSHarshad Shirwadkar sbi = EXT4_SB(sb); 64208016e29fSHarshad Shirwadkar 64211e1c2b86SLukas Czerner if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 64221e1c2b86SLukas Czerner !ext4_inode_block_valid(inode, block, count)) { 64231e1c2b86SLukas Czerner ext4_error(sb, "Freeing blocks in system zone - " 64241e1c2b86SLukas Czerner "Block = %llu, count = %lu", block, count); 64251e1c2b86SLukas Czerner /* err = 0. ext4_std_error should be a no op */ 642633e728c6SKemeng Shi goto error_out; 64271e1c2b86SLukas Czerner } 64281e1c2b86SLukas Czerner flags |= EXT4_FREE_BLOCKS_VALIDATED; 64291e1c2b86SLukas Czerner 6430c9de560dSAlex Tomas do_more: 6431c9de560dSAlex Tomas overflow = 0; 6432c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 6433c9de560dSAlex Tomas 64345354b2afSTheodore Ts'o grp = ext4_get_group_info(sb, block_group); 64355354b2afSTheodore Ts'o if (unlikely(!grp || EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 6436163a203dSDarrick J. Wong return; 6437163a203dSDarrick J. Wong 6438c9de560dSAlex Tomas /* 6439c9de560dSAlex Tomas * Check to see if we are freeing blocks across a group 6440c9de560dSAlex Tomas * boundary. 6441c9de560dSAlex Tomas */ 644284130193STheodore Ts'o if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { 644384130193STheodore Ts'o overflow = EXT4_C2B(sbi, bit) + count - 644484130193STheodore Ts'o EXT4_BLOCKS_PER_GROUP(sb); 6445c9de560dSAlex Tomas count -= overflow; 64461e1c2b86SLukas Czerner /* The range changed so it's no longer validated */ 64471e1c2b86SLukas Czerner flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 6448c9de560dSAlex Tomas } 6449810da240SLukas Czerner count_clusters = EXT4_NUM_B2C(sbi, count); 645033e728c6SKemeng Shi trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); 645133e728c6SKemeng Shi 645233e728c6SKemeng Shi /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ 645333e728c6SKemeng Shi err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, 645433e728c6SKemeng Shi GFP_NOFS|__GFP_NOFAIL); 645533e728c6SKemeng Shi if (err) 645633e728c6SKemeng Shi goto error_out; 6457c9de560dSAlex Tomas 64581e1c2b86SLukas Czerner if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 64591e1c2b86SLukas Czerner !ext4_inode_block_valid(inode, block, count)) { 646012062dddSEric Sandeen ext4_error(sb, "Freeing blocks in system zone - " 64610610b6e9STheodore Ts'o "Block = %llu, count = %lu", block, count); 6462519deca0SAneesh Kumar K.V /* err = 0. ext4_std_error should be a no op */ 646333e728c6SKemeng Shi goto error_clean; 646433e728c6SKemeng Shi } 646533e728c6SKemeng Shi 6466c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 646738b8f70cSKemeng Shi mark_flags |= EXT4_MB_BITMAP_MARKED_CHECK; 6468c9de560dSAlex Tomas #endif 646938b8f70cSKemeng Shi err = ext4_mb_mark_context(handle, sb, false, block_group, bit, 647038b8f70cSKemeng Shi count_clusters, mark_flags, &changed); 6471c9de560dSAlex Tomas 647233e728c6SKemeng Shi 647338b8f70cSKemeng Shi if (err && changed == 0) 647438b8f70cSKemeng Shi goto error_clean; 647533e728c6SKemeng Shi 647638b8f70cSKemeng Shi #ifdef AGGRESSIVE_CHECK 647738b8f70cSKemeng Shi BUG_ON(changed != count_clusters); 647838b8f70cSKemeng Shi #endif 6479e6362609STheodore Ts'o 6480f96c450dSDaeho Jeong /* 6481f96c450dSDaeho Jeong * We need to make sure we don't reuse the freed block until after the 6482f96c450dSDaeho Jeong * transaction is committed. We make an exception if the inode is to be 6483f96c450dSDaeho Jeong * written in writeback mode since writeback mode has weak data 6484f96c450dSDaeho Jeong * consistency guarantees. 6485f96c450dSDaeho Jeong */ 6486f96c450dSDaeho Jeong if (ext4_handle_valid(handle) && 6487f96c450dSDaeho Jeong ((flags & EXT4_FREE_BLOCKS_METADATA) || 6488f96c450dSDaeho Jeong !ext4_should_writeback_data(inode))) { 64897a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry; 64907a2fcbf7SAneesh Kumar K.V /* 64917444a072SMichal Hocko * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed 64927444a072SMichal Hocko * to fail. 64937a2fcbf7SAneesh Kumar K.V */ 64947444a072SMichal Hocko new_entry = kmem_cache_alloc(ext4_free_data_cachep, 64957444a072SMichal Hocko GFP_NOFS|__GFP_NOFAIL); 649618aadd47SBobi Jam new_entry->efd_start_cluster = bit; 649718aadd47SBobi Jam new_entry->efd_group = block_group; 649818aadd47SBobi Jam new_entry->efd_count = count_clusters; 649918aadd47SBobi Jam new_entry->efd_tid = handle->h_transaction->t_tid; 6500955ce5f5SAneesh Kumar K.V 65017a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, block_group); 65027a2fcbf7SAneesh Kumar K.V ext4_mb_free_metadata(handle, &e4b, new_entry); 6503c9de560dSAlex Tomas } else { 6504d71c1ae2SLukas Czerner if (test_opt(sb, DISCARD)) { 6505247c3d21SKemeng Shi err = ext4_issue_discard(sb, block_group, bit, 65060efcd739SWenchao Hao count_clusters); 6507fa606293SJan Kara /* 6508fa606293SJan Kara * Ignore EOPNOTSUPP error. This is consistent with 6509fa606293SJan Kara * what happens when using journal. 6510fa606293SJan Kara */ 6511fa606293SJan Kara if (err == -EOPNOTSUPP) 6512fa606293SJan Kara err = 0; 6513fa606293SJan Kara if (err) 6514d71c1ae2SLukas Czerner ext4_msg(sb, KERN_WARNING, "discard request in" 6515a00b482bSRitesh Harjani " group:%u block:%d count:%lu failed" 6516d71c1ae2SLukas Czerner " with %d", block_group, bit, count, 6517d71c1ae2SLukas Czerner err); 65188f9ff189SLukas Czerner } else 65198f9ff189SLukas Czerner EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); 6520d71c1ae2SLukas Czerner 6521955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 652284130193STheodore Ts'o mb_free_blocks(inode, &e4b, bit, count_clusters); 6523c9de560dSAlex Tomas } 6524c9de560dSAlex Tomas 6525955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 6526c9de560dSAlex Tomas 65279fe67149SEric Whitney /* 65289fe67149SEric Whitney * on a bigalloc file system, defer the s_freeclusters_counter 65299fe67149SEric Whitney * update to the caller (ext4_remove_space and friends) so they 65309fe67149SEric Whitney * can determine if a cluster freed here should be rereserved 65319fe67149SEric Whitney */ 65329fe67149SEric Whitney if (!(flags & EXT4_FREE_BLOCKS_RERESERVE_CLUSTER)) { 65337b415bf6SAditya Kali if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) 65347b415bf6SAditya Kali dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); 65359fe67149SEric Whitney percpu_counter_add(&sbi->s_freeclusters_counter, 65369fe67149SEric Whitney count_clusters); 65379fe67149SEric Whitney } 65387d734532SJan Kara 6539c9de560dSAlex Tomas if (overflow && !err) { 6540c9de560dSAlex Tomas block += count; 6541c9de560dSAlex Tomas count = overflow; 654233e728c6SKemeng Shi ext4_mb_unload_buddy(&e4b); 65431e1c2b86SLukas Czerner /* The range changed so it's no longer validated */ 65441e1c2b86SLukas Czerner flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 6545c9de560dSAlex Tomas goto do_more; 6546c9de560dSAlex Tomas } 654733e728c6SKemeng Shi 654833e728c6SKemeng Shi error_clean: 654933e728c6SKemeng Shi ext4_mb_unload_buddy(&e4b); 655033e728c6SKemeng Shi error_out: 6551c9de560dSAlex Tomas ext4_std_error(sb, err); 6552c9de560dSAlex Tomas } 65537360d173SLukas Czerner 65547360d173SLukas Czerner /** 65558ac3939dSRitesh Harjani * ext4_free_blocks() -- Free given blocks and update quota 65568ac3939dSRitesh Harjani * @handle: handle for this transaction 65578ac3939dSRitesh Harjani * @inode: inode 65588ac3939dSRitesh Harjani * @bh: optional buffer of the block to be freed 65598ac3939dSRitesh Harjani * @block: starting physical block to be freed 65608ac3939dSRitesh Harjani * @count: number of blocks to be freed 65618ac3939dSRitesh Harjani * @flags: flags used by ext4_free_blocks 65628ac3939dSRitesh Harjani */ 65638ac3939dSRitesh Harjani void ext4_free_blocks(handle_t *handle, struct inode *inode, 65648ac3939dSRitesh Harjani struct buffer_head *bh, ext4_fsblk_t block, 65658ac3939dSRitesh Harjani unsigned long count, int flags) 65668ac3939dSRitesh Harjani { 65678ac3939dSRitesh Harjani struct super_block *sb = inode->i_sb; 65688ac3939dSRitesh Harjani unsigned int overflow; 65698ac3939dSRitesh Harjani struct ext4_sb_info *sbi; 65708ac3939dSRitesh Harjani 65718ac3939dSRitesh Harjani sbi = EXT4_SB(sb); 65728ac3939dSRitesh Harjani 65738ac3939dSRitesh Harjani if (bh) { 65748ac3939dSRitesh Harjani if (block) 65758ac3939dSRitesh Harjani BUG_ON(block != bh->b_blocknr); 65768ac3939dSRitesh Harjani else 65778ac3939dSRitesh Harjani block = bh->b_blocknr; 65788ac3939dSRitesh Harjani } 65798ac3939dSRitesh Harjani 658011b6890bSKemeng Shi if (sbi->s_mount_state & EXT4_FC_REPLAY) { 65812ec6d0a5SKemeng Shi ext4_free_blocks_simple(inode, block, EXT4_NUM_B2C(sbi, count)); 658211b6890bSKemeng Shi return; 658311b6890bSKemeng Shi } 658411b6890bSKemeng Shi 658511b6890bSKemeng Shi might_sleep(); 658611b6890bSKemeng Shi 65878ac3939dSRitesh Harjani if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 65888ac3939dSRitesh Harjani !ext4_inode_block_valid(inode, block, count)) { 65898ac3939dSRitesh Harjani ext4_error(sb, "Freeing blocks not in datazone - " 65908ac3939dSRitesh Harjani "block = %llu, count = %lu", block, count); 65918ac3939dSRitesh Harjani return; 65928ac3939dSRitesh Harjani } 65931e1c2b86SLukas Czerner flags |= EXT4_FREE_BLOCKS_VALIDATED; 65948ac3939dSRitesh Harjani 65958ac3939dSRitesh Harjani ext4_debug("freeing block %llu\n", block); 65968ac3939dSRitesh Harjani trace_ext4_free_blocks(inode, block, count, flags); 65978ac3939dSRitesh Harjani 65988ac3939dSRitesh Harjani if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 65998ac3939dSRitesh Harjani BUG_ON(count > 1); 66008ac3939dSRitesh Harjani 66018ac3939dSRitesh Harjani ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, 66028ac3939dSRitesh Harjani inode, bh, block); 66038ac3939dSRitesh Harjani } 66048ac3939dSRitesh Harjani 66058ac3939dSRitesh Harjani /* 66068ac3939dSRitesh Harjani * If the extent to be freed does not begin on a cluster 66078ac3939dSRitesh Harjani * boundary, we need to deal with partial clusters at the 66088ac3939dSRitesh Harjani * beginning and end of the extent. Normally we will free 66098ac3939dSRitesh Harjani * blocks at the beginning or the end unless we are explicitly 66108ac3939dSRitesh Harjani * requested to avoid doing so. 66118ac3939dSRitesh Harjani */ 66128ac3939dSRitesh Harjani overflow = EXT4_PBLK_COFF(sbi, block); 66138ac3939dSRitesh Harjani if (overflow) { 66148ac3939dSRitesh Harjani if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { 66158ac3939dSRitesh Harjani overflow = sbi->s_cluster_ratio - overflow; 66168ac3939dSRitesh Harjani block += overflow; 66178ac3939dSRitesh Harjani if (count > overflow) 66188ac3939dSRitesh Harjani count -= overflow; 66198ac3939dSRitesh Harjani else 66208ac3939dSRitesh Harjani return; 66218ac3939dSRitesh Harjani } else { 66228ac3939dSRitesh Harjani block -= overflow; 66238ac3939dSRitesh Harjani count += overflow; 66248ac3939dSRitesh Harjani } 66251e1c2b86SLukas Czerner /* The range changed so it's no longer validated */ 66261e1c2b86SLukas Czerner flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 66278ac3939dSRitesh Harjani } 66288ac3939dSRitesh Harjani overflow = EXT4_LBLK_COFF(sbi, count); 66298ac3939dSRitesh Harjani if (overflow) { 66308ac3939dSRitesh Harjani if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { 66318ac3939dSRitesh Harjani if (count > overflow) 66328ac3939dSRitesh Harjani count -= overflow; 66338ac3939dSRitesh Harjani else 66348ac3939dSRitesh Harjani return; 66358ac3939dSRitesh Harjani } else 66368ac3939dSRitesh Harjani count += sbi->s_cluster_ratio - overflow; 66371e1c2b86SLukas Czerner /* The range changed so it's no longer validated */ 66381e1c2b86SLukas Czerner flags &= ~EXT4_FREE_BLOCKS_VALIDATED; 66398ac3939dSRitesh Harjani } 66408ac3939dSRitesh Harjani 66418ac3939dSRitesh Harjani if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 66428ac3939dSRitesh Harjani int i; 66438ac3939dSRitesh Harjani int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; 66448ac3939dSRitesh Harjani 66458ac3939dSRitesh Harjani for (i = 0; i < count; i++) { 66468ac3939dSRitesh Harjani cond_resched(); 66478ac3939dSRitesh Harjani if (is_metadata) 66488ac3939dSRitesh Harjani bh = sb_find_get_block(inode->i_sb, block + i); 66498ac3939dSRitesh Harjani ext4_forget(handle, is_metadata, inode, bh, block + i); 66508ac3939dSRitesh Harjani } 66518ac3939dSRitesh Harjani } 66528ac3939dSRitesh Harjani 66538ac3939dSRitesh Harjani ext4_mb_clear_bb(handle, inode, block, count, flags); 66548ac3939dSRitesh Harjani } 66558ac3939dSRitesh Harjani 66568ac3939dSRitesh Harjani /** 66570529155eSYongqiang Yang * ext4_group_add_blocks() -- Add given blocks to an existing group 66582846e820SAmir Goldstein * @handle: handle to this transaction 66592846e820SAmir Goldstein * @sb: super block 66604907cb7bSAnatol Pomozov * @block: start physical block to add to the block group 66612846e820SAmir Goldstein * @count: number of blocks to free 66622846e820SAmir Goldstein * 6663e73a347bSAmir Goldstein * This marks the blocks as free in the bitmap and buddy. 66642846e820SAmir Goldstein */ 6665cc7365dfSYongqiang Yang int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, 66662846e820SAmir Goldstein ext4_fsblk_t block, unsigned long count) 66672846e820SAmir Goldstein { 66682846e820SAmir Goldstein ext4_group_t block_group; 66692846e820SAmir Goldstein ext4_grpblk_t bit; 66702846e820SAmir Goldstein struct ext4_sb_info *sbi = EXT4_SB(sb); 6671e73a347bSAmir Goldstein struct ext4_buddy e4b; 66725c657db4SKemeng Shi int err = 0; 6673d77147ffSharshads ext4_fsblk_t first_cluster = EXT4_B2C(sbi, block); 6674d77147ffSharshads ext4_fsblk_t last_cluster = EXT4_B2C(sbi, block + count - 1); 6675d77147ffSharshads unsigned long cluster_count = last_cluster - first_cluster + 1; 66765c657db4SKemeng Shi ext4_grpblk_t changed; 66772846e820SAmir Goldstein 66782846e820SAmir Goldstein ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); 66792846e820SAmir Goldstein 66805c657db4SKemeng Shi if (cluster_count == 0) 66814740b830SYongqiang Yang return 0; 66824740b830SYongqiang Yang 66832846e820SAmir Goldstein ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 66842846e820SAmir Goldstein /* 66852846e820SAmir Goldstein * Check to see if we are freeing blocks across a group 66862846e820SAmir Goldstein * boundary. 66872846e820SAmir Goldstein */ 6688d77147ffSharshads if (bit + cluster_count > EXT4_CLUSTERS_PER_GROUP(sb)) { 6689d77147ffSharshads ext4_warning(sb, "too many blocks added to group %u", 6690cc7365dfSYongqiang Yang block_group); 6691cc7365dfSYongqiang Yang err = -EINVAL; 669203c7fc39SKemeng Shi goto error_out; 6693cc7365dfSYongqiang Yang } 66942cd05cc3STheodore Ts'o 669503c7fc39SKemeng Shi err = ext4_mb_load_buddy(sb, block_group, &e4b); 669603c7fc39SKemeng Shi if (err) 669703c7fc39SKemeng Shi goto error_out; 66982846e820SAmir Goldstein 6699a00b482bSRitesh Harjani if (!ext4_sb_block_valid(sb, NULL, block, count)) { 67002846e820SAmir Goldstein ext4_error(sb, "Adding blocks in system zones - " 67012846e820SAmir Goldstein "Block = %llu, count = %lu", 67022846e820SAmir Goldstein block, count); 6703cc7365dfSYongqiang Yang err = -EINVAL; 670403c7fc39SKemeng Shi goto error_clean; 670503c7fc39SKemeng Shi } 670603c7fc39SKemeng Shi 67075c657db4SKemeng Shi err = ext4_mb_mark_context(handle, sb, false, block_group, bit, 67085c657db4SKemeng Shi cluster_count, EXT4_MB_BITMAP_MARKED_CHECK, 67095c657db4SKemeng Shi &changed); 67105c657db4SKemeng Shi if (err && changed == 0) 671103c7fc39SKemeng Shi goto error_clean; 67122846e820SAmir Goldstein 67135c657db4SKemeng Shi if (changed != cluster_count) 67145c657db4SKemeng Shi ext4_error(sb, "bit already cleared in group %u", block_group); 67152846e820SAmir Goldstein 671603c7fc39SKemeng Shi ext4_lock_group(sb, block_group); 671703c7fc39SKemeng Shi mb_free_blocks(NULL, &e4b, bit, cluster_count); 671803c7fc39SKemeng Shi ext4_unlock_group(sb, block_group); 671903c7fc39SKemeng Shi percpu_counter_add(&sbi->s_freeclusters_counter, 67205c657db4SKemeng Shi changed); 672103c7fc39SKemeng Shi 672203c7fc39SKemeng Shi error_clean: 672303c7fc39SKemeng Shi ext4_mb_unload_buddy(&e4b); 672403c7fc39SKemeng Shi error_out: 67252846e820SAmir Goldstein ext4_std_error(sb, err); 6726cc7365dfSYongqiang Yang return err; 67272846e820SAmir Goldstein } 67282846e820SAmir Goldstein 67292846e820SAmir Goldstein /** 67307360d173SLukas Czerner * ext4_trim_extent -- function to TRIM one single free extent in the group 67317360d173SLukas Czerner * @sb: super block for the file system 67327360d173SLukas Czerner * @start: starting block of the free extent in the alloc. group 67337360d173SLukas Czerner * @count: number of blocks to TRIM 67347360d173SLukas Czerner * @e4b: ext4 buddy for the group 67357360d173SLukas Czerner * 67367360d173SLukas Czerner * Trim "count" blocks starting at "start" in the "group". To assure that no 67377360d173SLukas Czerner * one will allocate those blocks, mark it as used in buddy bitmap. This must 67387360d173SLukas Czerner * be called with under the group lock. 67397360d173SLukas Czerner */ 6740bd2eea8dSWang Jianchao static int ext4_trim_extent(struct super_block *sb, 6741bd2eea8dSWang Jianchao int start, int count, struct ext4_buddy *e4b) 6742e2cbd587Sjon ernst __releases(bitlock) 6743e2cbd587Sjon ernst __acquires(bitlock) 67447360d173SLukas Czerner { 67457360d173SLukas Czerner struct ext4_free_extent ex; 6746bd2eea8dSWang Jianchao ext4_group_t group = e4b->bd_group; 6747d71c1ae2SLukas Czerner int ret = 0; 67487360d173SLukas Czerner 6749b3d4c2b1STao Ma trace_ext4_trim_extent(sb, group, start, count); 6750b3d4c2b1STao Ma 67517360d173SLukas Czerner assert_spin_locked(ext4_group_lock_ptr(sb, group)); 67527360d173SLukas Czerner 67537360d173SLukas Czerner ex.fe_start = start; 67547360d173SLukas Czerner ex.fe_group = group; 67557360d173SLukas Czerner ex.fe_len = count; 67567360d173SLukas Czerner 67577360d173SLukas Czerner /* 67587360d173SLukas Czerner * Mark blocks used, so no one can reuse them while 67597360d173SLukas Czerner * being trimmed. 67607360d173SLukas Czerner */ 67617360d173SLukas Czerner mb_mark_used(e4b, &ex); 67627360d173SLukas Czerner ext4_unlock_group(sb, group); 67630efcd739SWenchao Hao ret = ext4_issue_discard(sb, group, start, count); 67647360d173SLukas Czerner ext4_lock_group(sb, group); 67657360d173SLukas Czerner mb_free_blocks(NULL, e4b, start, ex.fe_len); 6766d71c1ae2SLukas Czerner return ret; 67677360d173SLukas Czerner } 67687360d173SLukas Czerner 676945e4ab32SJan Kara static ext4_grpblk_t ext4_last_grp_cluster(struct super_block *sb, 677045e4ab32SJan Kara ext4_group_t grp) 677145e4ab32SJan Kara { 67727c784d62SSuraj Jitindar Singh unsigned long nr_clusters_in_group; 67737c784d62SSuraj Jitindar Singh 67747c784d62SSuraj Jitindar Singh if (grp < (ext4_get_groups_count(sb) - 1)) 67757c784d62SSuraj Jitindar Singh nr_clusters_in_group = EXT4_CLUSTERS_PER_GROUP(sb); 67767c784d62SSuraj Jitindar Singh else 67777c784d62SSuraj Jitindar Singh nr_clusters_in_group = (ext4_blocks_count(EXT4_SB(sb)->s_es) - 67787c784d62SSuraj Jitindar Singh ext4_group_first_block_no(sb, grp)) 67797c784d62SSuraj Jitindar Singh >> EXT4_CLUSTER_BITS(sb); 67807c784d62SSuraj Jitindar Singh 67817c784d62SSuraj Jitindar Singh return nr_clusters_in_group - 1; 678245e4ab32SJan Kara } 678345e4ab32SJan Kara 67845229a658SJan Kara static bool ext4_trim_interrupted(void) 67855229a658SJan Kara { 67865229a658SJan Kara return fatal_signal_pending(current) || freezing(current); 67875229a658SJan Kara } 67885229a658SJan Kara 67896920b391SWang Jianchao static int ext4_try_to_trim_range(struct super_block *sb, 67906920b391SWang Jianchao struct ext4_buddy *e4b, ext4_grpblk_t start, 67916920b391SWang Jianchao ext4_grpblk_t max, ext4_grpblk_t minblocks) 6792a5fda113STheodore Ts'o __acquires(ext4_group_lock_ptr(sb, e4b->bd_group)) 6793a5fda113STheodore Ts'o __releases(ext4_group_lock_ptr(sb, e4b->bd_group)) 67946920b391SWang Jianchao { 679568da4c44SYe Bin ext4_grpblk_t next, count, free_count, last, origin_start; 679645e4ab32SJan Kara bool set_trimmed = false; 67976920b391SWang Jianchao void *bitmap; 67986920b391SWang Jianchao 679917220215SBaokun Li if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 680017220215SBaokun Li return 0; 680117220215SBaokun Li 680268da4c44SYe Bin last = ext4_last_grp_cluster(sb, e4b->bd_group); 68036920b391SWang Jianchao bitmap = e4b->bd_bitmap; 680468da4c44SYe Bin if (start == 0 && max >= last) 680545e4ab32SJan Kara set_trimmed = true; 680668da4c44SYe Bin origin_start = start; 6807de8bf0e5SKemeng Shi start = max(e4b->bd_info->bb_first_free, start); 68086920b391SWang Jianchao count = 0; 68096920b391SWang Jianchao free_count = 0; 68106920b391SWang Jianchao 68116920b391SWang Jianchao while (start <= max) { 68126920b391SWang Jianchao start = mb_find_next_zero_bit(bitmap, max + 1, start); 68136920b391SWang Jianchao if (start > max) 68146920b391SWang Jianchao break; 681568da4c44SYe Bin 681668da4c44SYe Bin next = mb_find_next_bit(bitmap, last + 1, start); 681768da4c44SYe Bin if (origin_start == 0 && next >= last) 681868da4c44SYe Bin set_trimmed = true; 68196920b391SWang Jianchao 68206920b391SWang Jianchao if ((next - start) >= minblocks) { 6821afcc4e32SLukas Bulwahn int ret = ext4_trim_extent(sb, start, next - start, e4b); 6822afcc4e32SLukas Bulwahn 68236920b391SWang Jianchao if (ret && ret != -EOPNOTSUPP) 682445e4ab32SJan Kara return count; 68256920b391SWang Jianchao count += next - start; 68266920b391SWang Jianchao } 68276920b391SWang Jianchao free_count += next - start; 68286920b391SWang Jianchao start = next + 1; 68296920b391SWang Jianchao 68305229a658SJan Kara if (ext4_trim_interrupted()) 68315229a658SJan Kara return count; 68326920b391SWang Jianchao 68336920b391SWang Jianchao if (need_resched()) { 68346920b391SWang Jianchao ext4_unlock_group(sb, e4b->bd_group); 68356920b391SWang Jianchao cond_resched(); 68366920b391SWang Jianchao ext4_lock_group(sb, e4b->bd_group); 68376920b391SWang Jianchao } 68386920b391SWang Jianchao 68396920b391SWang Jianchao if ((e4b->bd_info->bb_free - free_count) < minblocks) 68406920b391SWang Jianchao break; 68416920b391SWang Jianchao } 68426920b391SWang Jianchao 684345e4ab32SJan Kara if (set_trimmed) 684445e4ab32SJan Kara EXT4_MB_GRP_SET_TRIMMED(e4b->bd_info); 684545e4ab32SJan Kara 68466920b391SWang Jianchao return count; 68476920b391SWang Jianchao } 68486920b391SWang Jianchao 68497360d173SLukas Czerner /** 68507360d173SLukas Czerner * ext4_trim_all_free -- function to trim all free space in alloc. group 68517360d173SLukas Czerner * @sb: super block for file system 685222612283STao Ma * @group: group to be trimmed 68537360d173SLukas Czerner * @start: first group block to examine 68547360d173SLukas Czerner * @max: last group block to examine 68557360d173SLukas Czerner * @minblocks: minimum extent block count 68567360d173SLukas Czerner * 68577360d173SLukas Czerner * ext4_trim_all_free walks through group's block bitmap searching for free 68587360d173SLukas Czerner * extents. When the free extent is found, mark it as used in group buddy 68597360d173SLukas Czerner * bitmap. Then issue a TRIM command on this extent and free the extent in 6860b6f5558cSWang Jianchao * the group buddy bitmap. 68617360d173SLukas Czerner */ 68620b75a840SLukas Czerner static ext4_grpblk_t 686378944086SLukas Czerner ext4_trim_all_free(struct super_block *sb, ext4_group_t group, 686478944086SLukas Czerner ext4_grpblk_t start, ext4_grpblk_t max, 686545e4ab32SJan Kara ext4_grpblk_t minblocks) 68667360d173SLukas Czerner { 686778944086SLukas Czerner struct ext4_buddy e4b; 68686920b391SWang Jianchao int ret; 68697360d173SLukas Czerner 6870b3d4c2b1STao Ma trace_ext4_trim_all_free(sb, group, start, max); 6871b3d4c2b1STao Ma 687278944086SLukas Czerner ret = ext4_mb_load_buddy(sb, group, &e4b); 687378944086SLukas Czerner if (ret) { 68749651e6b2SKonstantin Khlebnikov ext4_warning(sb, "Error %d loading buddy information for %u", 68759651e6b2SKonstantin Khlebnikov ret, group); 687678944086SLukas Czerner return ret; 687778944086SLukas Czerner } 687828739eeaSLukas Czerner 687928739eeaSLukas Czerner ext4_lock_group(sb, group); 68803d56b8d2STao Ma 68816920b391SWang Jianchao if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) || 688245e4ab32SJan Kara minblocks < EXT4_SB(sb)->s_last_trim_minblks) 68836920b391SWang Jianchao ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks); 688445e4ab32SJan Kara else 68856920b391SWang Jianchao ret = 0; 68866920b391SWang Jianchao 68877360d173SLukas Czerner ext4_unlock_group(sb, group); 688878944086SLukas Czerner ext4_mb_unload_buddy(&e4b); 68897360d173SLukas Czerner 68907360d173SLukas Czerner ext4_debug("trimmed %d blocks in the group %d\n", 68916920b391SWang Jianchao ret, group); 68927360d173SLukas Czerner 6893d71c1ae2SLukas Czerner return ret; 68947360d173SLukas Czerner } 68957360d173SLukas Czerner 68967360d173SLukas Czerner /** 68977360d173SLukas Czerner * ext4_trim_fs() -- trim ioctl handle function 68987360d173SLukas Czerner * @sb: superblock for filesystem 68997360d173SLukas Czerner * @range: fstrim_range structure 69007360d173SLukas Czerner * 69017360d173SLukas Czerner * start: First Byte to trim 69027360d173SLukas Czerner * len: number of Bytes to trim from start 69037360d173SLukas Czerner * minlen: minimum extent length in Bytes 69047360d173SLukas Czerner * ext4_trim_fs goes through all allocation groups containing Bytes from 69057360d173SLukas Czerner * start to start+len. For each such a group ext4_trim_all_free function 69067360d173SLukas Czerner * is invoked to trim all free space. 69077360d173SLukas Czerner */ 69087360d173SLukas Czerner int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) 69097360d173SLukas Czerner { 69107b47ef52SChristoph Hellwig unsigned int discard_granularity = bdev_discard_granularity(sb->s_bdev); 691178944086SLukas Czerner struct ext4_group_info *grp; 6912913eed83SLukas Czerner ext4_group_t group, first_group, last_group; 69137137d7a4STheodore Ts'o ext4_grpblk_t cnt = 0, first_cluster, last_cluster; 6914913eed83SLukas Czerner uint64_t start, end, minlen, trimmed = 0; 69150f0a25bfSJan Kara ext4_fsblk_t first_data_blk = 69160f0a25bfSJan Kara le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 6917913eed83SLukas Czerner ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); 69187360d173SLukas Czerner int ret = 0; 69197360d173SLukas Czerner 69207360d173SLukas Czerner start = range->start >> sb->s_blocksize_bits; 6921913eed83SLukas Czerner end = start + (range->len >> sb->s_blocksize_bits) - 1; 6922aaf7d73eSLukas Czerner minlen = EXT4_NUM_B2C(EXT4_SB(sb), 6923aaf7d73eSLukas Czerner range->minlen >> sb->s_blocksize_bits); 69247360d173SLukas Czerner 69255de35e8dSLukas Czerner if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || 69265de35e8dSLukas Czerner start >= max_blks || 69275de35e8dSLukas Czerner range->len < sb->s_blocksize) 69287360d173SLukas Czerner return -EINVAL; 6929173b6e38SJan Kara /* No point to try to trim less than discard granularity */ 69307b47ef52SChristoph Hellwig if (range->minlen < discard_granularity) { 6931173b6e38SJan Kara minlen = EXT4_NUM_B2C(EXT4_SB(sb), 69327b47ef52SChristoph Hellwig discard_granularity >> sb->s_blocksize_bits); 6933173b6e38SJan Kara if (minlen > EXT4_CLUSTERS_PER_GROUP(sb)) 6934173b6e38SJan Kara goto out; 6935173b6e38SJan Kara } 693645e4ab32SJan Kara if (end >= max_blks - 1) 6937913eed83SLukas Czerner end = max_blks - 1; 6938913eed83SLukas Czerner if (end <= first_data_blk) 693922f10457STao Ma goto out; 6940913eed83SLukas Czerner if (start < first_data_blk) 69410f0a25bfSJan Kara start = first_data_blk; 69427360d173SLukas Czerner 6943913eed83SLukas Czerner /* Determine first and last group to examine based on start and end */ 69447360d173SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, 69457137d7a4STheodore Ts'o &first_group, &first_cluster); 6946913eed83SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, 69477137d7a4STheodore Ts'o &last_group, &last_cluster); 69487360d173SLukas Czerner 6949913eed83SLukas Czerner /* end now represents the last cluster to discard in this group */ 6950913eed83SLukas Czerner end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 69517360d173SLukas Czerner 69527360d173SLukas Czerner for (group = first_group; group <= last_group; group++) { 69535229a658SJan Kara if (ext4_trim_interrupted()) 69545229a658SJan Kara break; 695578944086SLukas Czerner grp = ext4_get_group_info(sb, group); 69565354b2afSTheodore Ts'o if (!grp) 69575354b2afSTheodore Ts'o continue; 695878944086SLukas Czerner /* We only do this if the grp has never been initialized */ 695978944086SLukas Czerner if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 6960adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_group(sb, group, GFP_NOFS); 696178944086SLukas Czerner if (ret) 69627360d173SLukas Czerner break; 69637360d173SLukas Czerner } 69647360d173SLukas Czerner 69650ba08517STao Ma /* 6966913eed83SLukas Czerner * For all the groups except the last one, last cluster will 6967913eed83SLukas Czerner * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to 6968913eed83SLukas Czerner * change it for the last group, note that last_cluster is 6969913eed83SLukas Czerner * already computed earlier by ext4_get_group_no_and_offset() 69700ba08517STao Ma */ 697145e4ab32SJan Kara if (group == last_group) 6972913eed83SLukas Czerner end = last_cluster; 697378944086SLukas Czerner if (grp->bb_free >= minlen) { 69747137d7a4STheodore Ts'o cnt = ext4_trim_all_free(sb, group, first_cluster, 697545e4ab32SJan Kara end, minlen); 69767360d173SLukas Czerner if (cnt < 0) { 69777360d173SLukas Czerner ret = cnt; 69787360d173SLukas Czerner break; 69797360d173SLukas Czerner } 69807360d173SLukas Czerner trimmed += cnt; 698121e7fd22SLukas Czerner } 6982913eed83SLukas Czerner 6983913eed83SLukas Czerner /* 6984913eed83SLukas Czerner * For every group except the first one, we are sure 6985913eed83SLukas Czerner * that the first cluster to discard will be cluster #0. 6986913eed83SLukas Czerner */ 69877137d7a4STheodore Ts'o first_cluster = 0; 69887360d173SLukas Czerner } 69897360d173SLukas Czerner 69903d56b8d2STao Ma if (!ret) 69912327fb2eSLukas Czerner EXT4_SB(sb)->s_last_trim_minblks = minlen; 69923d56b8d2STao Ma 699322f10457STao Ma out: 6994aaf7d73eSLukas Czerner range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; 69957360d173SLukas Czerner return ret; 69967360d173SLukas Czerner } 69970c9ec4beSDarrick J. Wong 69980c9ec4beSDarrick J. Wong /* Iterate all the free extents in the group. */ 69990c9ec4beSDarrick J. Wong int 70000c9ec4beSDarrick J. Wong ext4_mballoc_query_range( 70010c9ec4beSDarrick J. Wong struct super_block *sb, 70020c9ec4beSDarrick J. Wong ext4_group_t group, 70030c9ec4beSDarrick J. Wong ext4_grpblk_t start, 70040c9ec4beSDarrick J. Wong ext4_grpblk_t end, 70050c9ec4beSDarrick J. Wong ext4_mballoc_query_range_fn formatter, 70060c9ec4beSDarrick J. Wong void *priv) 70070c9ec4beSDarrick J. Wong { 70080c9ec4beSDarrick J. Wong void *bitmap; 70090c9ec4beSDarrick J. Wong ext4_grpblk_t next; 70100c9ec4beSDarrick J. Wong struct ext4_buddy e4b; 70110c9ec4beSDarrick J. Wong int error; 70120c9ec4beSDarrick J. Wong 70130c9ec4beSDarrick J. Wong error = ext4_mb_load_buddy(sb, group, &e4b); 70140c9ec4beSDarrick J. Wong if (error) 70150c9ec4beSDarrick J. Wong return error; 70160c9ec4beSDarrick J. Wong bitmap = e4b.bd_bitmap; 70170c9ec4beSDarrick J. Wong 70180c9ec4beSDarrick J. Wong ext4_lock_group(sb, group); 70190c9ec4beSDarrick J. Wong 7020de8bf0e5SKemeng Shi start = max(e4b.bd_info->bb_first_free, start); 70210c9ec4beSDarrick J. Wong if (end >= EXT4_CLUSTERS_PER_GROUP(sb)) 70220c9ec4beSDarrick J. Wong end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 70230c9ec4beSDarrick J. Wong 70240c9ec4beSDarrick J. Wong while (start <= end) { 70250c9ec4beSDarrick J. Wong start = mb_find_next_zero_bit(bitmap, end + 1, start); 70260c9ec4beSDarrick J. Wong if (start > end) 70270c9ec4beSDarrick J. Wong break; 70280c9ec4beSDarrick J. Wong next = mb_find_next_bit(bitmap, end + 1, start); 70290c9ec4beSDarrick J. Wong 70300c9ec4beSDarrick J. Wong ext4_unlock_group(sb, group); 70310c9ec4beSDarrick J. Wong error = formatter(sb, group, start, next - start, priv); 70320c9ec4beSDarrick J. Wong if (error) 70330c9ec4beSDarrick J. Wong goto out_unload; 70340c9ec4beSDarrick J. Wong ext4_lock_group(sb, group); 70350c9ec4beSDarrick J. Wong 70360c9ec4beSDarrick J. Wong start = next + 1; 70370c9ec4beSDarrick J. Wong } 70380c9ec4beSDarrick J. Wong 70390c9ec4beSDarrick J. Wong ext4_unlock_group(sb, group); 70400c9ec4beSDarrick J. Wong out_unload: 70410c9ec4beSDarrick J. Wong ext4_mb_unload_buddy(&e4b); 70420c9ec4beSDarrick J. Wong 70430c9ec4beSDarrick J. Wong return error; 70440c9ec4beSDarrick J. Wong } 70457c9fa399SKemeng Shi 70467c9fa399SKemeng Shi #ifdef CONFIG_EXT4_KUNIT_TESTS 70477c9fa399SKemeng Shi #include "mballoc-test.c" 70487c9fa399SKemeng Shi #endif 7049