1c9de560dSAlex Tomas /* 2c9de560dSAlex Tomas * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com 3c9de560dSAlex Tomas * Written by Alex Tomas <alex@clusterfs.com> 4c9de560dSAlex Tomas * 5c9de560dSAlex Tomas * This program is free software; you can redistribute it and/or modify 6c9de560dSAlex Tomas * it under the terms of the GNU General Public License version 2 as 7c9de560dSAlex Tomas * published by the Free Software Foundation. 8c9de560dSAlex Tomas * 9c9de560dSAlex Tomas * This program is distributed in the hope that it will be useful, 10c9de560dSAlex Tomas * but WITHOUT ANY WARRANTY; without even the implied warranty of 11c9de560dSAlex Tomas * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12c9de560dSAlex Tomas * GNU General Public License for more details. 13c9de560dSAlex Tomas * 14c9de560dSAlex Tomas * You should have received a copy of the GNU General Public Licens 15c9de560dSAlex Tomas * along with this program; if not, write to the Free Software 16c9de560dSAlex Tomas * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- 17c9de560dSAlex Tomas */ 18c9de560dSAlex Tomas 19c9de560dSAlex Tomas 20c9de560dSAlex Tomas /* 21c9de560dSAlex Tomas * mballoc.c contains the multiblocks allocation routines 22c9de560dSAlex Tomas */ 23c9de560dSAlex Tomas 248f6e39a7SMingming Cao #include "mballoc.h" 256ba495e9STheodore Ts'o #include <linux/debugfs.h> 269bffad1eSTheodore Ts'o #include <trace/events/ext4.h> 279bffad1eSTheodore Ts'o 28c9de560dSAlex Tomas /* 29c9de560dSAlex Tomas * MUSTDO: 30c9de560dSAlex Tomas * - test ext4_ext_search_left() and ext4_ext_search_right() 31c9de560dSAlex Tomas * - search for metadata in few groups 32c9de560dSAlex Tomas * 33c9de560dSAlex Tomas * TODO v4: 34c9de560dSAlex Tomas * - normalization should take into account whether file is still open 35c9de560dSAlex Tomas * - discard preallocations if no free space left (policy?) 36c9de560dSAlex Tomas * - don't normalize tails 37c9de560dSAlex Tomas * - quota 38c9de560dSAlex Tomas * - reservation for superuser 39c9de560dSAlex Tomas * 40c9de560dSAlex Tomas * TODO v3: 41c9de560dSAlex Tomas * - bitmap read-ahead (proposed by Oleg Drokin aka green) 42c9de560dSAlex Tomas * - track min/max extents in each group for better group selection 43c9de560dSAlex Tomas * - mb_mark_used() may allocate chunk right after splitting buddy 44c9de560dSAlex Tomas * - tree of groups sorted by number of free blocks 45c9de560dSAlex Tomas * - error handling 46c9de560dSAlex Tomas */ 47c9de560dSAlex Tomas 48c9de560dSAlex Tomas /* 49c9de560dSAlex Tomas * The allocation request involve request for multiple number of blocks 50c9de560dSAlex Tomas * near to the goal(block) value specified. 51c9de560dSAlex Tomas * 52b713a5ecSTheodore Ts'o * During initialization phase of the allocator we decide to use the 53b713a5ecSTheodore Ts'o * group preallocation or inode preallocation depending on the size of 54b713a5ecSTheodore Ts'o * the file. The size of the file could be the resulting file size we 55b713a5ecSTheodore Ts'o * would have after allocation, or the current file size, which ever 56b713a5ecSTheodore Ts'o * is larger. If the size is less than sbi->s_mb_stream_request we 57b713a5ecSTheodore Ts'o * select to use the group preallocation. The default value of 58b713a5ecSTheodore Ts'o * s_mb_stream_request is 16 blocks. This can also be tuned via 59b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 60b713a5ecSTheodore Ts'o * terms of number of blocks. 61c9de560dSAlex Tomas * 62c9de560dSAlex Tomas * The main motivation for having small file use group preallocation is to 63b713a5ecSTheodore Ts'o * ensure that we have small files closer together on the disk. 64c9de560dSAlex Tomas * 65b713a5ecSTheodore Ts'o * First stage the allocator looks at the inode prealloc list, 66b713a5ecSTheodore Ts'o * ext4_inode_info->i_prealloc_list, which contains list of prealloc 67b713a5ecSTheodore Ts'o * spaces for this particular inode. The inode prealloc space is 68b713a5ecSTheodore Ts'o * represented as: 69c9de560dSAlex Tomas * 70c9de560dSAlex Tomas * pa_lstart -> the logical start block for this prealloc space 71c9de560dSAlex Tomas * pa_pstart -> the physical start block for this prealloc space 72c9de560dSAlex Tomas * pa_len -> lenght for this prealloc space 73c9de560dSAlex Tomas * pa_free -> free space available in this prealloc space 74c9de560dSAlex Tomas * 75c9de560dSAlex Tomas * The inode preallocation space is used looking at the _logical_ start 76c9de560dSAlex Tomas * block. If only the logical file block falls within the range of prealloc 77c9de560dSAlex Tomas * space we will consume the particular prealloc space. This make sure that 78c9de560dSAlex Tomas * that the we have contiguous physical blocks representing the file blocks 79c9de560dSAlex Tomas * 80c9de560dSAlex Tomas * The important thing to be noted in case of inode prealloc space is that 81c9de560dSAlex Tomas * we don't modify the values associated to inode prealloc space except 82c9de560dSAlex Tomas * pa_free. 83c9de560dSAlex Tomas * 84c9de560dSAlex Tomas * If we are not able to find blocks in the inode prealloc space and if we 85c9de560dSAlex Tomas * have the group allocation flag set then we look at the locality group 86c9de560dSAlex Tomas * prealloc space. These are per CPU prealloc list repreasented as 87c9de560dSAlex Tomas * 88c9de560dSAlex Tomas * ext4_sb_info.s_locality_groups[smp_processor_id()] 89c9de560dSAlex Tomas * 90c9de560dSAlex Tomas * The reason for having a per cpu locality group is to reduce the contention 91c9de560dSAlex Tomas * between CPUs. It is possible to get scheduled at this point. 92c9de560dSAlex Tomas * 93c9de560dSAlex Tomas * The locality group prealloc space is used looking at whether we have 94c9de560dSAlex Tomas * enough free space (pa_free) withing the prealloc space. 95c9de560dSAlex Tomas * 96c9de560dSAlex Tomas * If we can't allocate blocks via inode prealloc or/and locality group 97c9de560dSAlex Tomas * prealloc then we look at the buddy cache. The buddy cache is represented 98c9de560dSAlex Tomas * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 99c9de560dSAlex Tomas * mapped to the buddy and bitmap information regarding different 100c9de560dSAlex Tomas * groups. The buddy information is attached to buddy cache inode so that 101c9de560dSAlex Tomas * we can access them through the page cache. The information regarding 102c9de560dSAlex Tomas * each group is loaded via ext4_mb_load_buddy. The information involve 103c9de560dSAlex Tomas * block bitmap and buddy information. The information are stored in the 104c9de560dSAlex Tomas * inode as: 105c9de560dSAlex Tomas * 106c9de560dSAlex Tomas * { page } 107c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 108c9de560dSAlex Tomas * 109c9de560dSAlex Tomas * 110c9de560dSAlex Tomas * one block each for bitmap and buddy information. So for each group we 111c9de560dSAlex Tomas * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / 112c9de560dSAlex Tomas * blocksize) blocks. So it can have information regarding groups_per_page 113c9de560dSAlex Tomas * which is blocks_per_page/2 114c9de560dSAlex Tomas * 115c9de560dSAlex Tomas * The buddy cache inode is not stored on disk. The inode is thrown 116c9de560dSAlex Tomas * away when the filesystem is unmounted. 117c9de560dSAlex Tomas * 118c9de560dSAlex Tomas * We look for count number of blocks in the buddy cache. If we were able 119c9de560dSAlex Tomas * to locate that many free blocks we return with additional information 120c9de560dSAlex Tomas * regarding rest of the contiguous physical block available 121c9de560dSAlex Tomas * 122c9de560dSAlex Tomas * Before allocating blocks via buddy cache we normalize the request 123c9de560dSAlex Tomas * blocks. This ensure we ask for more blocks that we needed. The extra 124c9de560dSAlex Tomas * blocks that we get after allocation is added to the respective prealloc 125c9de560dSAlex Tomas * list. In case of inode preallocation we follow a list of heuristics 126c9de560dSAlex Tomas * based on file size. This can be found in ext4_mb_normalize_request. If 127c9de560dSAlex Tomas * we are doing a group prealloc we try to normalize the request to 128b713a5ecSTheodore Ts'o * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is 129c9de560dSAlex Tomas * 512 blocks. This can be tuned via 130b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in 131c9de560dSAlex Tomas * terms of number of blocks. If we have mounted the file system with -O 132c9de560dSAlex Tomas * stripe=<value> option the group prealloc request is normalized to the 133c9de560dSAlex Tomas * stripe value (sbi->s_stripe) 134c9de560dSAlex Tomas * 135b713a5ecSTheodore Ts'o * The regular allocator(using the buddy cache) supports few tunables. 136c9de560dSAlex Tomas * 137b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_min_to_scan 138b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_max_to_scan 139b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req 140c9de560dSAlex Tomas * 141b713a5ecSTheodore Ts'o * The regular allocator uses buddy scan only if the request len is power of 142c9de560dSAlex Tomas * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 143c9de560dSAlex Tomas * value of s_mb_order2_reqs can be tuned via 144b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 145c9de560dSAlex Tomas * stripe size (sbi->s_stripe), we try to search for contigous block in 146b713a5ecSTheodore Ts'o * stripe size. This should result in better allocation on RAID setups. If 147b713a5ecSTheodore Ts'o * not, we search in the specific group using bitmap for best extents. The 148b713a5ecSTheodore Ts'o * tunable min_to_scan and max_to_scan control the behaviour here. 149c9de560dSAlex Tomas * min_to_scan indicate how long the mballoc __must__ look for a best 150b713a5ecSTheodore Ts'o * extent and max_to_scan indicates how long the mballoc __can__ look for a 151c9de560dSAlex Tomas * best extent in the found extents. Searching for the blocks starts with 152c9de560dSAlex Tomas * the group specified as the goal value in allocation context via 153c9de560dSAlex Tomas * ac_g_ex. Each group is first checked based on the criteria whether it 154c9de560dSAlex Tomas * can used for allocation. ext4_mb_good_group explains how the groups are 155c9de560dSAlex Tomas * checked. 156c9de560dSAlex Tomas * 157c9de560dSAlex Tomas * Both the prealloc space are getting populated as above. So for the first 158c9de560dSAlex Tomas * request we will hit the buddy cache which will result in this prealloc 159c9de560dSAlex Tomas * space getting filled. The prealloc space is then later used for the 160c9de560dSAlex Tomas * subsequent request. 161c9de560dSAlex Tomas */ 162c9de560dSAlex Tomas 163c9de560dSAlex Tomas /* 164c9de560dSAlex Tomas * mballoc operates on the following data: 165c9de560dSAlex Tomas * - on-disk bitmap 166c9de560dSAlex Tomas * - in-core buddy (actually includes buddy and bitmap) 167c9de560dSAlex Tomas * - preallocation descriptors (PAs) 168c9de560dSAlex Tomas * 169c9de560dSAlex Tomas * there are two types of preallocations: 170c9de560dSAlex Tomas * - inode 171c9de560dSAlex Tomas * assiged to specific inode and can be used for this inode only. 172c9de560dSAlex Tomas * it describes part of inode's space preallocated to specific 173c9de560dSAlex Tomas * physical blocks. any block from that preallocated can be used 174c9de560dSAlex Tomas * independent. the descriptor just tracks number of blocks left 175c9de560dSAlex Tomas * unused. so, before taking some block from descriptor, one must 176c9de560dSAlex Tomas * make sure corresponded logical block isn't allocated yet. this 177c9de560dSAlex Tomas * also means that freeing any block within descriptor's range 178c9de560dSAlex Tomas * must discard all preallocated blocks. 179c9de560dSAlex Tomas * - locality group 180c9de560dSAlex Tomas * assigned to specific locality group which does not translate to 181c9de560dSAlex Tomas * permanent set of inodes: inode can join and leave group. space 182c9de560dSAlex Tomas * from this type of preallocation can be used for any inode. thus 183c9de560dSAlex Tomas * it's consumed from the beginning to the end. 184c9de560dSAlex Tomas * 185c9de560dSAlex Tomas * relation between them can be expressed as: 186c9de560dSAlex Tomas * in-core buddy = on-disk bitmap + preallocation descriptors 187c9de560dSAlex Tomas * 188c9de560dSAlex Tomas * this mean blocks mballoc considers used are: 189c9de560dSAlex Tomas * - allocated blocks (persistent) 190c9de560dSAlex Tomas * - preallocated blocks (non-persistent) 191c9de560dSAlex Tomas * 192c9de560dSAlex Tomas * consistency in mballoc world means that at any time a block is either 193c9de560dSAlex Tomas * free or used in ALL structures. notice: "any time" should not be read 194c9de560dSAlex Tomas * literally -- time is discrete and delimited by locks. 195c9de560dSAlex Tomas * 196c9de560dSAlex Tomas * to keep it simple, we don't use block numbers, instead we count number of 197c9de560dSAlex Tomas * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 198c9de560dSAlex Tomas * 199c9de560dSAlex Tomas * all operations can be expressed as: 200c9de560dSAlex Tomas * - init buddy: buddy = on-disk + PAs 201c9de560dSAlex Tomas * - new PA: buddy += N; PA = N 202c9de560dSAlex Tomas * - use inode PA: on-disk += N; PA -= N 203c9de560dSAlex Tomas * - discard inode PA buddy -= on-disk - PA; PA = 0 204c9de560dSAlex Tomas * - use locality group PA on-disk += N; PA -= N 205c9de560dSAlex Tomas * - discard locality group PA buddy -= PA; PA = 0 206c9de560dSAlex Tomas * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 207c9de560dSAlex Tomas * is used in real operation because we can't know actual used 208c9de560dSAlex Tomas * bits from PA, only from on-disk bitmap 209c9de560dSAlex Tomas * 210c9de560dSAlex Tomas * if we follow this strict logic, then all operations above should be atomic. 211c9de560dSAlex Tomas * given some of them can block, we'd have to use something like semaphores 212c9de560dSAlex Tomas * killing performance on high-end SMP hardware. let's try to relax it using 213c9de560dSAlex Tomas * the following knowledge: 214c9de560dSAlex Tomas * 1) if buddy is referenced, it's already initialized 215c9de560dSAlex Tomas * 2) while block is used in buddy and the buddy is referenced, 216c9de560dSAlex Tomas * nobody can re-allocate that block 217c9de560dSAlex Tomas * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 218c9de560dSAlex Tomas * bit set and PA claims same block, it's OK. IOW, one can set bit in 219c9de560dSAlex Tomas * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 220c9de560dSAlex Tomas * block 221c9de560dSAlex Tomas * 222c9de560dSAlex Tomas * so, now we're building a concurrency table: 223c9de560dSAlex Tomas * - init buddy vs. 224c9de560dSAlex Tomas * - new PA 225c9de560dSAlex Tomas * blocks for PA are allocated in the buddy, buddy must be referenced 226c9de560dSAlex Tomas * until PA is linked to allocation group to avoid concurrent buddy init 227c9de560dSAlex Tomas * - use inode PA 228c9de560dSAlex Tomas * we need to make sure that either on-disk bitmap or PA has uptodate data 229c9de560dSAlex Tomas * given (3) we care that PA-=N operation doesn't interfere with init 230c9de560dSAlex Tomas * - discard inode PA 231c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 232c9de560dSAlex Tomas * - use locality group PA 233c9de560dSAlex Tomas * again PA-=N must be serialized with init 234c9de560dSAlex Tomas * - discard locality group PA 235c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 236c9de560dSAlex Tomas * - new PA vs. 237c9de560dSAlex Tomas * - use inode PA 238c9de560dSAlex Tomas * i_data_sem serializes them 239c9de560dSAlex Tomas * - discard inode PA 240c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 241c9de560dSAlex Tomas * - use locality group PA 242c9de560dSAlex Tomas * some mutex should serialize them 243c9de560dSAlex Tomas * - discard locality group PA 244c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 245c9de560dSAlex Tomas * - use inode PA 246c9de560dSAlex Tomas * - use inode PA 247c9de560dSAlex Tomas * i_data_sem or another mutex should serializes them 248c9de560dSAlex Tomas * - discard inode PA 249c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 250c9de560dSAlex Tomas * - use locality group PA 251c9de560dSAlex Tomas * nothing wrong here -- they're different PAs covering different blocks 252c9de560dSAlex Tomas * - discard locality group PA 253c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 254c9de560dSAlex Tomas * 255c9de560dSAlex Tomas * now we're ready to make few consequences: 256c9de560dSAlex Tomas * - PA is referenced and while it is no discard is possible 257c9de560dSAlex Tomas * - PA is referenced until block isn't marked in on-disk bitmap 258c9de560dSAlex Tomas * - PA changes only after on-disk bitmap 259c9de560dSAlex Tomas * - discard must not compete with init. either init is done before 260c9de560dSAlex Tomas * any discard or they're serialized somehow 261c9de560dSAlex Tomas * - buddy init as sum of on-disk bitmap and PAs is done atomically 262c9de560dSAlex Tomas * 263c9de560dSAlex Tomas * a special case when we've used PA to emptiness. no need to modify buddy 264c9de560dSAlex Tomas * in this case, but we should care about concurrent init 265c9de560dSAlex Tomas * 266c9de560dSAlex Tomas */ 267c9de560dSAlex Tomas 268c9de560dSAlex Tomas /* 269c9de560dSAlex Tomas * Logic in few words: 270c9de560dSAlex Tomas * 271c9de560dSAlex Tomas * - allocation: 272c9de560dSAlex Tomas * load group 273c9de560dSAlex Tomas * find blocks 274c9de560dSAlex Tomas * mark bits in on-disk bitmap 275c9de560dSAlex Tomas * release group 276c9de560dSAlex Tomas * 277c9de560dSAlex Tomas * - use preallocation: 278c9de560dSAlex Tomas * find proper PA (per-inode or group) 279c9de560dSAlex Tomas * load group 280c9de560dSAlex Tomas * mark bits in on-disk bitmap 281c9de560dSAlex Tomas * release group 282c9de560dSAlex Tomas * release PA 283c9de560dSAlex Tomas * 284c9de560dSAlex Tomas * - free: 285c9de560dSAlex Tomas * load group 286c9de560dSAlex Tomas * mark bits in on-disk bitmap 287c9de560dSAlex Tomas * release group 288c9de560dSAlex Tomas * 289c9de560dSAlex Tomas * - discard preallocations in group: 290c9de560dSAlex Tomas * mark PAs deleted 291c9de560dSAlex Tomas * move them onto local list 292c9de560dSAlex Tomas * load on-disk bitmap 293c9de560dSAlex Tomas * load group 294c9de560dSAlex Tomas * remove PA from object (inode or locality group) 295c9de560dSAlex Tomas * mark free blocks in-core 296c9de560dSAlex Tomas * 297c9de560dSAlex Tomas * - discard inode's preallocations: 298c9de560dSAlex Tomas */ 299c9de560dSAlex Tomas 300c9de560dSAlex Tomas /* 301c9de560dSAlex Tomas * Locking rules 302c9de560dSAlex Tomas * 303c9de560dSAlex Tomas * Locks: 304c9de560dSAlex Tomas * - bitlock on a group (group) 305c9de560dSAlex Tomas * - object (inode/locality) (object) 306c9de560dSAlex Tomas * - per-pa lock (pa) 307c9de560dSAlex Tomas * 308c9de560dSAlex Tomas * Paths: 309c9de560dSAlex Tomas * - new pa 310c9de560dSAlex Tomas * object 311c9de560dSAlex Tomas * group 312c9de560dSAlex Tomas * 313c9de560dSAlex Tomas * - find and use pa: 314c9de560dSAlex Tomas * pa 315c9de560dSAlex Tomas * 316c9de560dSAlex Tomas * - release consumed pa: 317c9de560dSAlex Tomas * pa 318c9de560dSAlex Tomas * group 319c9de560dSAlex Tomas * object 320c9de560dSAlex Tomas * 321c9de560dSAlex Tomas * - generate in-core bitmap: 322c9de560dSAlex Tomas * group 323c9de560dSAlex Tomas * pa 324c9de560dSAlex Tomas * 325c9de560dSAlex Tomas * - discard all for given object (inode, locality group): 326c9de560dSAlex Tomas * object 327c9de560dSAlex Tomas * pa 328c9de560dSAlex Tomas * group 329c9de560dSAlex Tomas * 330c9de560dSAlex Tomas * - discard all for given group: 331c9de560dSAlex Tomas * group 332c9de560dSAlex Tomas * pa 333c9de560dSAlex Tomas * group 334c9de560dSAlex Tomas * object 335c9de560dSAlex Tomas * 336c9de560dSAlex Tomas */ 337c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_pspace_cachep; 338c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_ac_cachep; 339c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_free_ext_cachep; 340c3a326a6SAneesh Kumar K.V static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 341c3a326a6SAneesh Kumar K.V ext4_group_t group); 3427a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 3437a2fcbf7SAneesh Kumar K.V ext4_group_t group); 344c3a326a6SAneesh Kumar K.V static void release_blocks_on_commit(journal_t *journal, transaction_t *txn); 345c3a326a6SAneesh Kumar K.V 346ffad0a44SAneesh Kumar K.V static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 347ffad0a44SAneesh Kumar K.V { 348c9de560dSAlex Tomas #if BITS_PER_LONG == 64 349ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 7UL) << 3; 350ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~7UL); 351c9de560dSAlex Tomas #elif BITS_PER_LONG == 32 352ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 3UL) << 3; 353ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~3UL); 354c9de560dSAlex Tomas #else 355c9de560dSAlex Tomas #error "how many bits you are?!" 356c9de560dSAlex Tomas #endif 357ffad0a44SAneesh Kumar K.V return addr; 358ffad0a44SAneesh Kumar K.V } 359c9de560dSAlex Tomas 360c9de560dSAlex Tomas static inline int mb_test_bit(int bit, void *addr) 361c9de560dSAlex Tomas { 362c9de560dSAlex Tomas /* 363c9de560dSAlex Tomas * ext4_test_bit on architecture like powerpc 364c9de560dSAlex Tomas * needs unsigned long aligned address 365c9de560dSAlex Tomas */ 366ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 367c9de560dSAlex Tomas return ext4_test_bit(bit, addr); 368c9de560dSAlex Tomas } 369c9de560dSAlex Tomas 370c9de560dSAlex Tomas static inline void mb_set_bit(int bit, void *addr) 371c9de560dSAlex Tomas { 372ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 373c9de560dSAlex Tomas ext4_set_bit(bit, addr); 374c9de560dSAlex Tomas } 375c9de560dSAlex Tomas 376c9de560dSAlex Tomas static inline void mb_clear_bit(int bit, void *addr) 377c9de560dSAlex Tomas { 378ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 379c9de560dSAlex Tomas ext4_clear_bit(bit, addr); 380c9de560dSAlex Tomas } 381c9de560dSAlex Tomas 382ffad0a44SAneesh Kumar K.V static inline int mb_find_next_zero_bit(void *addr, int max, int start) 383ffad0a44SAneesh Kumar K.V { 384e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 385ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 386e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 387ffad0a44SAneesh Kumar K.V start += fix; 388ffad0a44SAneesh Kumar K.V 389e7dfb246SAneesh Kumar K.V ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 390e7dfb246SAneesh Kumar K.V if (ret > max) 391e7dfb246SAneesh Kumar K.V return max; 392e7dfb246SAneesh Kumar K.V return ret; 393ffad0a44SAneesh Kumar K.V } 394ffad0a44SAneesh Kumar K.V 395ffad0a44SAneesh Kumar K.V static inline int mb_find_next_bit(void *addr, int max, int start) 396ffad0a44SAneesh Kumar K.V { 397e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 398ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 399e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 400ffad0a44SAneesh Kumar K.V start += fix; 401ffad0a44SAneesh Kumar K.V 402e7dfb246SAneesh Kumar K.V ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 403e7dfb246SAneesh Kumar K.V if (ret > max) 404e7dfb246SAneesh Kumar K.V return max; 405e7dfb246SAneesh Kumar K.V return ret; 406ffad0a44SAneesh Kumar K.V } 407ffad0a44SAneesh Kumar K.V 408c9de560dSAlex Tomas static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 409c9de560dSAlex Tomas { 410c9de560dSAlex Tomas char *bb; 411c9de560dSAlex Tomas 412c9de560dSAlex Tomas BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); 413c9de560dSAlex Tomas BUG_ON(max == NULL); 414c9de560dSAlex Tomas 415c9de560dSAlex Tomas if (order > e4b->bd_blkbits + 1) { 416c9de560dSAlex Tomas *max = 0; 417c9de560dSAlex Tomas return NULL; 418c9de560dSAlex Tomas } 419c9de560dSAlex Tomas 420c9de560dSAlex Tomas /* at order 0 we see each particular block */ 421c9de560dSAlex Tomas *max = 1 << (e4b->bd_blkbits + 3); 422c9de560dSAlex Tomas if (order == 0) 423c9de560dSAlex Tomas return EXT4_MB_BITMAP(e4b); 424c9de560dSAlex Tomas 425c9de560dSAlex Tomas bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 426c9de560dSAlex Tomas *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 427c9de560dSAlex Tomas 428c9de560dSAlex Tomas return bb; 429c9de560dSAlex Tomas } 430c9de560dSAlex Tomas 431c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 432c9de560dSAlex Tomas static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 433c9de560dSAlex Tomas int first, int count) 434c9de560dSAlex Tomas { 435c9de560dSAlex Tomas int i; 436c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 437c9de560dSAlex Tomas 438c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 439c9de560dSAlex Tomas return; 440bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 441c9de560dSAlex Tomas for (i = 0; i < count; i++) { 442c9de560dSAlex Tomas if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 443c9de560dSAlex Tomas ext4_fsblk_t blocknr; 444c9de560dSAlex Tomas blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); 445c9de560dSAlex Tomas blocknr += first + i; 446c9de560dSAlex Tomas blocknr += 447c9de560dSAlex Tomas le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 4485d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 4495d1b1b3fSAneesh Kumar K.V __func__, "double-free of inode" 450a9df9a49STheodore Ts'o " %lu's block %llu(bit %u in group %u)", 451c9de560dSAlex Tomas inode ? inode->i_ino : 0, blocknr, 452c9de560dSAlex Tomas first + i, e4b->bd_group); 453c9de560dSAlex Tomas } 454c9de560dSAlex Tomas mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 455c9de560dSAlex Tomas } 456c9de560dSAlex Tomas } 457c9de560dSAlex Tomas 458c9de560dSAlex Tomas static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 459c9de560dSAlex Tomas { 460c9de560dSAlex Tomas int i; 461c9de560dSAlex Tomas 462c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 463c9de560dSAlex Tomas return; 464bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 465c9de560dSAlex Tomas for (i = 0; i < count; i++) { 466c9de560dSAlex Tomas BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 467c9de560dSAlex Tomas mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 468c9de560dSAlex Tomas } 469c9de560dSAlex Tomas } 470c9de560dSAlex Tomas 471c9de560dSAlex Tomas static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 472c9de560dSAlex Tomas { 473c9de560dSAlex Tomas if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 474c9de560dSAlex Tomas unsigned char *b1, *b2; 475c9de560dSAlex Tomas int i; 476c9de560dSAlex Tomas b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 477c9de560dSAlex Tomas b2 = (unsigned char *) bitmap; 478c9de560dSAlex Tomas for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 479c9de560dSAlex Tomas if (b1[i] != b2[i]) { 480a9df9a49STheodore Ts'o printk(KERN_ERR "corruption in group %u " 4814776004fSTheodore Ts'o "at byte %u(%u): %x in copy != %x " 4824776004fSTheodore Ts'o "on disk/prealloc\n", 483c9de560dSAlex Tomas e4b->bd_group, i, i * 8, b1[i], b2[i]); 484c9de560dSAlex Tomas BUG(); 485c9de560dSAlex Tomas } 486c9de560dSAlex Tomas } 487c9de560dSAlex Tomas } 488c9de560dSAlex Tomas } 489c9de560dSAlex Tomas 490c9de560dSAlex Tomas #else 491c9de560dSAlex Tomas static inline void mb_free_blocks_double(struct inode *inode, 492c9de560dSAlex Tomas struct ext4_buddy *e4b, int first, int count) 493c9de560dSAlex Tomas { 494c9de560dSAlex Tomas return; 495c9de560dSAlex Tomas } 496c9de560dSAlex Tomas static inline void mb_mark_used_double(struct ext4_buddy *e4b, 497c9de560dSAlex Tomas int first, int count) 498c9de560dSAlex Tomas { 499c9de560dSAlex Tomas return; 500c9de560dSAlex Tomas } 501c9de560dSAlex Tomas static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 502c9de560dSAlex Tomas { 503c9de560dSAlex Tomas return; 504c9de560dSAlex Tomas } 505c9de560dSAlex Tomas #endif 506c9de560dSAlex Tomas 507c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 508c9de560dSAlex Tomas 509c9de560dSAlex Tomas #define MB_CHECK_ASSERT(assert) \ 510c9de560dSAlex Tomas do { \ 511c9de560dSAlex Tomas if (!(assert)) { \ 512c9de560dSAlex Tomas printk(KERN_EMERG \ 513c9de560dSAlex Tomas "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 514c9de560dSAlex Tomas function, file, line, # assert); \ 515c9de560dSAlex Tomas BUG(); \ 516c9de560dSAlex Tomas } \ 517c9de560dSAlex Tomas } while (0) 518c9de560dSAlex Tomas 519c9de560dSAlex Tomas static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 520c9de560dSAlex Tomas const char *function, int line) 521c9de560dSAlex Tomas { 522c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 523c9de560dSAlex Tomas int order = e4b->bd_blkbits + 1; 524c9de560dSAlex Tomas int max; 525c9de560dSAlex Tomas int max2; 526c9de560dSAlex Tomas int i; 527c9de560dSAlex Tomas int j; 528c9de560dSAlex Tomas int k; 529c9de560dSAlex Tomas int count; 530c9de560dSAlex Tomas struct ext4_group_info *grp; 531c9de560dSAlex Tomas int fragments = 0; 532c9de560dSAlex Tomas int fstart; 533c9de560dSAlex Tomas struct list_head *cur; 534c9de560dSAlex Tomas void *buddy; 535c9de560dSAlex Tomas void *buddy2; 536c9de560dSAlex Tomas 537c9de560dSAlex Tomas { 538c9de560dSAlex Tomas static int mb_check_counter; 539c9de560dSAlex Tomas if (mb_check_counter++ % 100 != 0) 540c9de560dSAlex Tomas return 0; 541c9de560dSAlex Tomas } 542c9de560dSAlex Tomas 543c9de560dSAlex Tomas while (order > 1) { 544c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 545c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy); 546c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order - 1, &max2); 547c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy2); 548c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy != buddy2); 549c9de560dSAlex Tomas MB_CHECK_ASSERT(max * 2 == max2); 550c9de560dSAlex Tomas 551c9de560dSAlex Tomas count = 0; 552c9de560dSAlex Tomas for (i = 0; i < max; i++) { 553c9de560dSAlex Tomas 554c9de560dSAlex Tomas if (mb_test_bit(i, buddy)) { 555c9de560dSAlex Tomas /* only single bit in buddy2 may be 1 */ 556c9de560dSAlex Tomas if (!mb_test_bit(i << 1, buddy2)) { 557c9de560dSAlex Tomas MB_CHECK_ASSERT( 558c9de560dSAlex Tomas mb_test_bit((i<<1)+1, buddy2)); 559c9de560dSAlex Tomas } else if (!mb_test_bit((i << 1) + 1, buddy2)) { 560c9de560dSAlex Tomas MB_CHECK_ASSERT( 561c9de560dSAlex Tomas mb_test_bit(i << 1, buddy2)); 562c9de560dSAlex Tomas } 563c9de560dSAlex Tomas continue; 564c9de560dSAlex Tomas } 565c9de560dSAlex Tomas 566c9de560dSAlex Tomas /* both bits in buddy2 must be 0 */ 567c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 568c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 569c9de560dSAlex Tomas 570c9de560dSAlex Tomas for (j = 0; j < (1 << order); j++) { 571c9de560dSAlex Tomas k = (i * (1 << order)) + j; 572c9de560dSAlex Tomas MB_CHECK_ASSERT( 573c9de560dSAlex Tomas !mb_test_bit(k, EXT4_MB_BITMAP(e4b))); 574c9de560dSAlex Tomas } 575c9de560dSAlex Tomas count++; 576c9de560dSAlex Tomas } 577c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 578c9de560dSAlex Tomas order--; 579c9de560dSAlex Tomas } 580c9de560dSAlex Tomas 581c9de560dSAlex Tomas fstart = -1; 582c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 583c9de560dSAlex Tomas for (i = 0; i < max; i++) { 584c9de560dSAlex Tomas if (!mb_test_bit(i, buddy)) { 585c9de560dSAlex Tomas MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 586c9de560dSAlex Tomas if (fstart == -1) { 587c9de560dSAlex Tomas fragments++; 588c9de560dSAlex Tomas fstart = i; 589c9de560dSAlex Tomas } 590c9de560dSAlex Tomas continue; 591c9de560dSAlex Tomas } 592c9de560dSAlex Tomas fstart = -1; 593c9de560dSAlex Tomas /* check used bits only */ 594c9de560dSAlex Tomas for (j = 0; j < e4b->bd_blkbits + 1; j++) { 595c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, j, &max2); 596c9de560dSAlex Tomas k = i >> j; 597c9de560dSAlex Tomas MB_CHECK_ASSERT(k < max2); 598c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 599c9de560dSAlex Tomas } 600c9de560dSAlex Tomas } 601c9de560dSAlex Tomas MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 602c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 603c9de560dSAlex Tomas 604c9de560dSAlex Tomas grp = ext4_get_group_info(sb, e4b->bd_group); 605c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 606c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 607c9de560dSAlex Tomas ext4_group_t groupnr; 608c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 60960bd63d1SSolofo Ramangalahy pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 61060bd63d1SSolofo Ramangalahy ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 611c9de560dSAlex Tomas MB_CHECK_ASSERT(groupnr == e4b->bd_group); 61260bd63d1SSolofo Ramangalahy for (i = 0; i < pa->pa_len; i++) 613c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 614c9de560dSAlex Tomas } 615c9de560dSAlex Tomas return 0; 616c9de560dSAlex Tomas } 617c9de560dSAlex Tomas #undef MB_CHECK_ASSERT 618c9de560dSAlex Tomas #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 61946e665e9SHarvey Harrison __FILE__, __func__, __LINE__) 620c9de560dSAlex Tomas #else 621c9de560dSAlex Tomas #define mb_check_buddy(e4b) 622c9de560dSAlex Tomas #endif 623c9de560dSAlex Tomas 624c9de560dSAlex Tomas /* FIXME!! need more doc */ 625c9de560dSAlex Tomas static void ext4_mb_mark_free_simple(struct super_block *sb, 626a36b4498SEric Sandeen void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 627c9de560dSAlex Tomas struct ext4_group_info *grp) 628c9de560dSAlex Tomas { 629c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 630a36b4498SEric Sandeen ext4_grpblk_t min; 631a36b4498SEric Sandeen ext4_grpblk_t max; 632a36b4498SEric Sandeen ext4_grpblk_t chunk; 633c9de560dSAlex Tomas unsigned short border; 634c9de560dSAlex Tomas 635b73fce69SValerie Clement BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb)); 636c9de560dSAlex Tomas 637c9de560dSAlex Tomas border = 2 << sb->s_blocksize_bits; 638c9de560dSAlex Tomas 639c9de560dSAlex Tomas while (len > 0) { 640c9de560dSAlex Tomas /* find how many blocks can be covered since this position */ 641c9de560dSAlex Tomas max = ffs(first | border) - 1; 642c9de560dSAlex Tomas 643c9de560dSAlex Tomas /* find how many blocks of power 2 we need to mark */ 644c9de560dSAlex Tomas min = fls(len) - 1; 645c9de560dSAlex Tomas 646c9de560dSAlex Tomas if (max < min) 647c9de560dSAlex Tomas min = max; 648c9de560dSAlex Tomas chunk = 1 << min; 649c9de560dSAlex Tomas 650c9de560dSAlex Tomas /* mark multiblock chunks only */ 651c9de560dSAlex Tomas grp->bb_counters[min]++; 652c9de560dSAlex Tomas if (min > 0) 653c9de560dSAlex Tomas mb_clear_bit(first >> min, 654c9de560dSAlex Tomas buddy + sbi->s_mb_offsets[min]); 655c9de560dSAlex Tomas 656c9de560dSAlex Tomas len -= chunk; 657c9de560dSAlex Tomas first += chunk; 658c9de560dSAlex Tomas } 659c9de560dSAlex Tomas } 660c9de560dSAlex Tomas 661089ceeccSEric Sandeen static noinline_for_stack 662089ceeccSEric Sandeen void ext4_mb_generate_buddy(struct super_block *sb, 663c9de560dSAlex Tomas void *buddy, void *bitmap, ext4_group_t group) 664c9de560dSAlex Tomas { 665c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 666a36b4498SEric Sandeen ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb); 667a36b4498SEric Sandeen ext4_grpblk_t i = 0; 668a36b4498SEric Sandeen ext4_grpblk_t first; 669a36b4498SEric Sandeen ext4_grpblk_t len; 670c9de560dSAlex Tomas unsigned free = 0; 671c9de560dSAlex Tomas unsigned fragments = 0; 672c9de560dSAlex Tomas unsigned long long period = get_cycles(); 673c9de560dSAlex Tomas 674c9de560dSAlex Tomas /* initialize buddy from bitmap which is aggregation 675c9de560dSAlex Tomas * of on-disk bitmap and preallocations */ 676ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, 0); 677c9de560dSAlex Tomas grp->bb_first_free = i; 678c9de560dSAlex Tomas while (i < max) { 679c9de560dSAlex Tomas fragments++; 680c9de560dSAlex Tomas first = i; 681ffad0a44SAneesh Kumar K.V i = mb_find_next_bit(bitmap, max, i); 682c9de560dSAlex Tomas len = i - first; 683c9de560dSAlex Tomas free += len; 684c9de560dSAlex Tomas if (len > 1) 685c9de560dSAlex Tomas ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 686c9de560dSAlex Tomas else 687c9de560dSAlex Tomas grp->bb_counters[0]++; 688c9de560dSAlex Tomas if (i < max) 689ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, i); 690c9de560dSAlex Tomas } 691c9de560dSAlex Tomas grp->bb_fragments = fragments; 692c9de560dSAlex Tomas 693c9de560dSAlex Tomas if (free != grp->bb_free) { 6945d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, group, __func__, 695a9df9a49STheodore Ts'o "EXT4-fs: group %u: %u blocks in bitmap, %u in gd", 696c9de560dSAlex Tomas group, free, grp->bb_free); 697e56eb659SAneesh Kumar K.V /* 698e56eb659SAneesh Kumar K.V * If we intent to continue, we consider group descritor 699e56eb659SAneesh Kumar K.V * corrupt and update bb_free using bitmap value 700e56eb659SAneesh Kumar K.V */ 701c9de560dSAlex Tomas grp->bb_free = free; 702c9de560dSAlex Tomas } 703c9de560dSAlex Tomas 704c9de560dSAlex Tomas clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 705c9de560dSAlex Tomas 706c9de560dSAlex Tomas period = get_cycles() - period; 707c9de560dSAlex Tomas spin_lock(&EXT4_SB(sb)->s_bal_lock); 708c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_buddies_generated++; 709c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_generation_time += period; 710c9de560dSAlex Tomas spin_unlock(&EXT4_SB(sb)->s_bal_lock); 711c9de560dSAlex Tomas } 712c9de560dSAlex Tomas 713c9de560dSAlex Tomas /* The buddy information is attached the buddy cache inode 714c9de560dSAlex Tomas * for convenience. The information regarding each group 715c9de560dSAlex Tomas * is loaded via ext4_mb_load_buddy. The information involve 716c9de560dSAlex Tomas * block bitmap and buddy information. The information are 717c9de560dSAlex Tomas * stored in the inode as 718c9de560dSAlex Tomas * 719c9de560dSAlex Tomas * { page } 720c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 721c9de560dSAlex Tomas * 722c9de560dSAlex Tomas * 723c9de560dSAlex Tomas * one block each for bitmap and buddy information. 724c9de560dSAlex Tomas * So for each group we take up 2 blocks. A page can 725c9de560dSAlex Tomas * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. 726c9de560dSAlex Tomas * So it can have information regarding groups_per_page which 727c9de560dSAlex Tomas * is blocks_per_page/2 728c9de560dSAlex Tomas */ 729c9de560dSAlex Tomas 730c9de560dSAlex Tomas static int ext4_mb_init_cache(struct page *page, char *incore) 731c9de560dSAlex Tomas { 7328df9675fSTheodore Ts'o ext4_group_t ngroups; 733c9de560dSAlex Tomas int blocksize; 734c9de560dSAlex Tomas int blocks_per_page; 735c9de560dSAlex Tomas int groups_per_page; 736c9de560dSAlex Tomas int err = 0; 737c9de560dSAlex Tomas int i; 738c9de560dSAlex Tomas ext4_group_t first_group; 739c9de560dSAlex Tomas int first_block; 740c9de560dSAlex Tomas struct super_block *sb; 741c9de560dSAlex Tomas struct buffer_head *bhs; 742c9de560dSAlex Tomas struct buffer_head **bh; 743c9de560dSAlex Tomas struct inode *inode; 744c9de560dSAlex Tomas char *data; 745c9de560dSAlex Tomas char *bitmap; 746c9de560dSAlex Tomas 7476ba495e9STheodore Ts'o mb_debug(1, "init page %lu\n", page->index); 748c9de560dSAlex Tomas 749c9de560dSAlex Tomas inode = page->mapping->host; 750c9de560dSAlex Tomas sb = inode->i_sb; 7518df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 752c9de560dSAlex Tomas blocksize = 1 << inode->i_blkbits; 753c9de560dSAlex Tomas blocks_per_page = PAGE_CACHE_SIZE / blocksize; 754c9de560dSAlex Tomas 755c9de560dSAlex Tomas groups_per_page = blocks_per_page >> 1; 756c9de560dSAlex Tomas if (groups_per_page == 0) 757c9de560dSAlex Tomas groups_per_page = 1; 758c9de560dSAlex Tomas 759c9de560dSAlex Tomas /* allocate buffer_heads to read bitmaps */ 760c9de560dSAlex Tomas if (groups_per_page > 1) { 761c9de560dSAlex Tomas err = -ENOMEM; 762c9de560dSAlex Tomas i = sizeof(struct buffer_head *) * groups_per_page; 763c9de560dSAlex Tomas bh = kzalloc(i, GFP_NOFS); 764c9de560dSAlex Tomas if (bh == NULL) 765c9de560dSAlex Tomas goto out; 766c9de560dSAlex Tomas } else 767c9de560dSAlex Tomas bh = &bhs; 768c9de560dSAlex Tomas 769c9de560dSAlex Tomas first_group = page->index * blocks_per_page / 2; 770c9de560dSAlex Tomas 771c9de560dSAlex Tomas /* read all groups the page covers into the cache */ 772c9de560dSAlex Tomas for (i = 0; i < groups_per_page; i++) { 773c9de560dSAlex Tomas struct ext4_group_desc *desc; 774c9de560dSAlex Tomas 7758df9675fSTheodore Ts'o if (first_group + i >= ngroups) 776c9de560dSAlex Tomas break; 777c9de560dSAlex Tomas 778c9de560dSAlex Tomas err = -EIO; 779c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, first_group + i, NULL); 780c9de560dSAlex Tomas if (desc == NULL) 781c9de560dSAlex Tomas goto out; 782c9de560dSAlex Tomas 783c9de560dSAlex Tomas err = -ENOMEM; 784c9de560dSAlex Tomas bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); 785c9de560dSAlex Tomas if (bh[i] == NULL) 786c9de560dSAlex Tomas goto out; 787c9de560dSAlex Tomas 7882ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh[i])) 789c9de560dSAlex Tomas continue; 790c9de560dSAlex Tomas 791c806e68fSFrederic Bohe lock_buffer(bh[i]); 7922ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh[i])) { 7932ccb5fb9SAneesh Kumar K.V unlock_buffer(bh[i]); 7942ccb5fb9SAneesh Kumar K.V continue; 7952ccb5fb9SAneesh Kumar K.V } 796955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, first_group + i); 797c9de560dSAlex Tomas if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 798c9de560dSAlex Tomas ext4_init_block_bitmap(sb, bh[i], 799c9de560dSAlex Tomas first_group + i, desc); 8002ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh[i]); 801c9de560dSAlex Tomas set_buffer_uptodate(bh[i]); 802955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, first_group + i); 8033300bedaSAneesh Kumar K.V unlock_buffer(bh[i]); 804c9de560dSAlex Tomas continue; 805c9de560dSAlex Tomas } 806955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, first_group + i); 8072ccb5fb9SAneesh Kumar K.V if (buffer_uptodate(bh[i])) { 8082ccb5fb9SAneesh Kumar K.V /* 8092ccb5fb9SAneesh Kumar K.V * if not uninit if bh is uptodate, 8102ccb5fb9SAneesh Kumar K.V * bitmap is also uptodate 8112ccb5fb9SAneesh Kumar K.V */ 8122ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh[i]); 8132ccb5fb9SAneesh Kumar K.V unlock_buffer(bh[i]); 8142ccb5fb9SAneesh Kumar K.V continue; 8152ccb5fb9SAneesh Kumar K.V } 816c9de560dSAlex Tomas get_bh(bh[i]); 8172ccb5fb9SAneesh Kumar K.V /* 8182ccb5fb9SAneesh Kumar K.V * submit the buffer_head for read. We can 8192ccb5fb9SAneesh Kumar K.V * safely mark the bitmap as uptodate now. 8202ccb5fb9SAneesh Kumar K.V * We do it here so the bitmap uptodate bit 8212ccb5fb9SAneesh Kumar K.V * get set with buffer lock held. 8222ccb5fb9SAneesh Kumar K.V */ 8232ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh[i]); 824c9de560dSAlex Tomas bh[i]->b_end_io = end_buffer_read_sync; 825c9de560dSAlex Tomas submit_bh(READ, bh[i]); 8266ba495e9STheodore Ts'o mb_debug(1, "read bitmap for group %u\n", first_group + i); 827c9de560dSAlex Tomas } 828c9de560dSAlex Tomas 829c9de560dSAlex Tomas /* wait for I/O completion */ 830c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 831c9de560dSAlex Tomas wait_on_buffer(bh[i]); 832c9de560dSAlex Tomas 833c9de560dSAlex Tomas err = -EIO; 834c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 835c9de560dSAlex Tomas if (!buffer_uptodate(bh[i])) 836c9de560dSAlex Tomas goto out; 837c9de560dSAlex Tomas 83831b481dcSMingming Cao err = 0; 839c9de560dSAlex Tomas first_block = page->index * blocks_per_page; 84029eaf024SAneesh Kumar K.V /* init the page */ 84129eaf024SAneesh Kumar K.V memset(page_address(page), 0xff, PAGE_CACHE_SIZE); 842c9de560dSAlex Tomas for (i = 0; i < blocks_per_page; i++) { 843c9de560dSAlex Tomas int group; 844c9de560dSAlex Tomas struct ext4_group_info *grinfo; 845c9de560dSAlex Tomas 846c9de560dSAlex Tomas group = (first_block + i) >> 1; 8478df9675fSTheodore Ts'o if (group >= ngroups) 848c9de560dSAlex Tomas break; 849c9de560dSAlex Tomas 850c9de560dSAlex Tomas /* 851c9de560dSAlex Tomas * data carry information regarding this 852c9de560dSAlex Tomas * particular group in the format specified 853c9de560dSAlex Tomas * above 854c9de560dSAlex Tomas * 855c9de560dSAlex Tomas */ 856c9de560dSAlex Tomas data = page_address(page) + (i * blocksize); 857c9de560dSAlex Tomas bitmap = bh[group - first_group]->b_data; 858c9de560dSAlex Tomas 859c9de560dSAlex Tomas /* 860c9de560dSAlex Tomas * We place the buddy block and bitmap block 861c9de560dSAlex Tomas * close together 862c9de560dSAlex Tomas */ 863c9de560dSAlex Tomas if ((first_block + i) & 1) { 864c9de560dSAlex Tomas /* this is block of buddy */ 865c9de560dSAlex Tomas BUG_ON(incore == NULL); 8666ba495e9STheodore Ts'o mb_debug(1, "put buddy for group %u in page %lu/%x\n", 867c9de560dSAlex Tomas group, page->index, i * blocksize); 868c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, group); 869c9de560dSAlex Tomas grinfo->bb_fragments = 0; 870c9de560dSAlex Tomas memset(grinfo->bb_counters, 0, 8711927805eSEric Sandeen sizeof(*grinfo->bb_counters) * 8721927805eSEric Sandeen (sb->s_blocksize_bits+2)); 873c9de560dSAlex Tomas /* 874c9de560dSAlex Tomas * incore got set to the group block bitmap below 875c9de560dSAlex Tomas */ 8767a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, group); 877c9de560dSAlex Tomas ext4_mb_generate_buddy(sb, data, incore, group); 8787a2fcbf7SAneesh Kumar K.V ext4_unlock_group(sb, group); 879c9de560dSAlex Tomas incore = NULL; 880c9de560dSAlex Tomas } else { 881c9de560dSAlex Tomas /* this is block of bitmap */ 882c9de560dSAlex Tomas BUG_ON(incore != NULL); 8836ba495e9STheodore Ts'o mb_debug(1, "put bitmap for group %u in page %lu/%x\n", 884c9de560dSAlex Tomas group, page->index, i * blocksize); 885c9de560dSAlex Tomas 886c9de560dSAlex Tomas /* see comments in ext4_mb_put_pa() */ 887c9de560dSAlex Tomas ext4_lock_group(sb, group); 888c9de560dSAlex Tomas memcpy(data, bitmap, blocksize); 889c9de560dSAlex Tomas 890c9de560dSAlex Tomas /* mark all preallocated blks used in in-core bitmap */ 891c9de560dSAlex Tomas ext4_mb_generate_from_pa(sb, data, group); 8927a2fcbf7SAneesh Kumar K.V ext4_mb_generate_from_freelist(sb, data, group); 893c9de560dSAlex Tomas ext4_unlock_group(sb, group); 894c9de560dSAlex Tomas 895c9de560dSAlex Tomas /* set incore so that the buddy information can be 896c9de560dSAlex Tomas * generated using this 897c9de560dSAlex Tomas */ 898c9de560dSAlex Tomas incore = data; 899c9de560dSAlex Tomas } 900c9de560dSAlex Tomas } 901c9de560dSAlex Tomas SetPageUptodate(page); 902c9de560dSAlex Tomas 903c9de560dSAlex Tomas out: 904c9de560dSAlex Tomas if (bh) { 905c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 906c9de560dSAlex Tomas brelse(bh[i]); 907c9de560dSAlex Tomas if (bh != &bhs) 908c9de560dSAlex Tomas kfree(bh); 909c9de560dSAlex Tomas } 910c9de560dSAlex Tomas return err; 911c9de560dSAlex Tomas } 912c9de560dSAlex Tomas 913b6a758ecSAneesh Kumar K.V static noinline_for_stack 914b6a758ecSAneesh Kumar K.V int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) 915b6a758ecSAneesh Kumar K.V { 916b6a758ecSAneesh Kumar K.V 917b6a758ecSAneesh Kumar K.V int ret = 0; 918b6a758ecSAneesh Kumar K.V void *bitmap; 919b6a758ecSAneesh Kumar K.V int blocks_per_page; 920b6a758ecSAneesh Kumar K.V int block, pnum, poff; 921b6a758ecSAneesh Kumar K.V int num_grp_locked = 0; 922b6a758ecSAneesh Kumar K.V struct ext4_group_info *this_grp; 923b6a758ecSAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 924b6a758ecSAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 925b6a758ecSAneesh Kumar K.V struct page *page = NULL, *bitmap_page = NULL; 926b6a758ecSAneesh Kumar K.V 927b6a758ecSAneesh Kumar K.V mb_debug(1, "init group %u\n", group); 928b6a758ecSAneesh Kumar K.V blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 929b6a758ecSAneesh Kumar K.V this_grp = ext4_get_group_info(sb, group); 930b6a758ecSAneesh Kumar K.V /* 93108c3a813SAneesh Kumar K.V * This ensures that we don't reinit the buddy cache 93208c3a813SAneesh Kumar K.V * page which map to the group from which we are already 93308c3a813SAneesh Kumar K.V * allocating. If we are looking at the buddy cache we would 93408c3a813SAneesh Kumar K.V * have taken a reference using ext4_mb_load_buddy and that 93508c3a813SAneesh Kumar K.V * would have taken the alloc_sem lock. 936b6a758ecSAneesh Kumar K.V */ 937b6a758ecSAneesh Kumar K.V num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group); 938b6a758ecSAneesh Kumar K.V if (!EXT4_MB_GRP_NEED_INIT(this_grp)) { 939b6a758ecSAneesh Kumar K.V /* 940b6a758ecSAneesh Kumar K.V * somebody initialized the group 941b6a758ecSAneesh Kumar K.V * return without doing anything 942b6a758ecSAneesh Kumar K.V */ 943b6a758ecSAneesh Kumar K.V ret = 0; 944b6a758ecSAneesh Kumar K.V goto err; 945b6a758ecSAneesh Kumar K.V } 946b6a758ecSAneesh Kumar K.V /* 947b6a758ecSAneesh Kumar K.V * the buddy cache inode stores the block bitmap 948b6a758ecSAneesh Kumar K.V * and buddy information in consecutive blocks. 949b6a758ecSAneesh Kumar K.V * So for each group we need two blocks. 950b6a758ecSAneesh Kumar K.V */ 951b6a758ecSAneesh Kumar K.V block = group * 2; 952b6a758ecSAneesh Kumar K.V pnum = block / blocks_per_page; 953b6a758ecSAneesh Kumar K.V poff = block % blocks_per_page; 954b6a758ecSAneesh Kumar K.V page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 955b6a758ecSAneesh Kumar K.V if (page) { 956b6a758ecSAneesh Kumar K.V BUG_ON(page->mapping != inode->i_mapping); 957b6a758ecSAneesh Kumar K.V ret = ext4_mb_init_cache(page, NULL); 958b6a758ecSAneesh Kumar K.V if (ret) { 959b6a758ecSAneesh Kumar K.V unlock_page(page); 960b6a758ecSAneesh Kumar K.V goto err; 961b6a758ecSAneesh Kumar K.V } 962b6a758ecSAneesh Kumar K.V unlock_page(page); 963b6a758ecSAneesh Kumar K.V } 964b6a758ecSAneesh Kumar K.V if (page == NULL || !PageUptodate(page)) { 965b6a758ecSAneesh Kumar K.V ret = -EIO; 966b6a758ecSAneesh Kumar K.V goto err; 967b6a758ecSAneesh Kumar K.V } 968b6a758ecSAneesh Kumar K.V mark_page_accessed(page); 969b6a758ecSAneesh Kumar K.V bitmap_page = page; 970b6a758ecSAneesh Kumar K.V bitmap = page_address(page) + (poff * sb->s_blocksize); 971b6a758ecSAneesh Kumar K.V 972b6a758ecSAneesh Kumar K.V /* init buddy cache */ 973b6a758ecSAneesh Kumar K.V block++; 974b6a758ecSAneesh Kumar K.V pnum = block / blocks_per_page; 975b6a758ecSAneesh Kumar K.V poff = block % blocks_per_page; 976b6a758ecSAneesh Kumar K.V page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 977b6a758ecSAneesh Kumar K.V if (page == bitmap_page) { 978b6a758ecSAneesh Kumar K.V /* 979b6a758ecSAneesh Kumar K.V * If both the bitmap and buddy are in 980b6a758ecSAneesh Kumar K.V * the same page we don't need to force 981b6a758ecSAneesh Kumar K.V * init the buddy 982b6a758ecSAneesh Kumar K.V */ 983b6a758ecSAneesh Kumar K.V unlock_page(page); 984b6a758ecSAneesh Kumar K.V } else if (page) { 985b6a758ecSAneesh Kumar K.V BUG_ON(page->mapping != inode->i_mapping); 986b6a758ecSAneesh Kumar K.V ret = ext4_mb_init_cache(page, bitmap); 987b6a758ecSAneesh Kumar K.V if (ret) { 988b6a758ecSAneesh Kumar K.V unlock_page(page); 989b6a758ecSAneesh Kumar K.V goto err; 990b6a758ecSAneesh Kumar K.V } 991b6a758ecSAneesh Kumar K.V unlock_page(page); 992b6a758ecSAneesh Kumar K.V } 993b6a758ecSAneesh Kumar K.V if (page == NULL || !PageUptodate(page)) { 994b6a758ecSAneesh Kumar K.V ret = -EIO; 995b6a758ecSAneesh Kumar K.V goto err; 996b6a758ecSAneesh Kumar K.V } 997b6a758ecSAneesh Kumar K.V mark_page_accessed(page); 998b6a758ecSAneesh Kumar K.V err: 999b6a758ecSAneesh Kumar K.V ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked); 1000b6a758ecSAneesh Kumar K.V if (bitmap_page) 1001b6a758ecSAneesh Kumar K.V page_cache_release(bitmap_page); 1002b6a758ecSAneesh Kumar K.V if (page) 1003b6a758ecSAneesh Kumar K.V page_cache_release(page); 1004b6a758ecSAneesh Kumar K.V return ret; 1005b6a758ecSAneesh Kumar K.V } 1006b6a758ecSAneesh Kumar K.V 10074ddfef7bSEric Sandeen static noinline_for_stack int 10084ddfef7bSEric Sandeen ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1009c9de560dSAlex Tomas struct ext4_buddy *e4b) 1010c9de560dSAlex Tomas { 1011c9de560dSAlex Tomas int blocks_per_page; 1012c9de560dSAlex Tomas int block; 1013c9de560dSAlex Tomas int pnum; 1014c9de560dSAlex Tomas int poff; 1015c9de560dSAlex Tomas struct page *page; 1016fdf6c7a7SShen Feng int ret; 1017920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1018920313a7SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 1019920313a7SAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 1020c9de560dSAlex Tomas 10216ba495e9STheodore Ts'o mb_debug(1, "load group %u\n", group); 1022c9de560dSAlex Tomas 1023c9de560dSAlex Tomas blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1024920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 1025c9de560dSAlex Tomas 1026c9de560dSAlex Tomas e4b->bd_blkbits = sb->s_blocksize_bits; 1027c9de560dSAlex Tomas e4b->bd_info = ext4_get_group_info(sb, group); 1028c9de560dSAlex Tomas e4b->bd_sb = sb; 1029c9de560dSAlex Tomas e4b->bd_group = group; 1030c9de560dSAlex Tomas e4b->bd_buddy_page = NULL; 1031c9de560dSAlex Tomas e4b->bd_bitmap_page = NULL; 1032920313a7SAneesh Kumar K.V e4b->alloc_semp = &grp->alloc_sem; 1033920313a7SAneesh Kumar K.V 1034920313a7SAneesh Kumar K.V /* Take the read lock on the group alloc 1035920313a7SAneesh Kumar K.V * sem. This would make sure a parallel 1036920313a7SAneesh Kumar K.V * ext4_mb_init_group happening on other 1037920313a7SAneesh Kumar K.V * groups mapped by the page is blocked 1038920313a7SAneesh Kumar K.V * till we are done with allocation 1039920313a7SAneesh Kumar K.V */ 1040f41c0750SAneesh Kumar K.V repeat_load_buddy: 1041920313a7SAneesh Kumar K.V down_read(e4b->alloc_semp); 1042c9de560dSAlex Tomas 1043f41c0750SAneesh Kumar K.V if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1044f41c0750SAneesh Kumar K.V /* we need to check for group need init flag 1045f41c0750SAneesh Kumar K.V * with alloc_semp held so that we can be sure 1046f41c0750SAneesh Kumar K.V * that new blocks didn't get added to the group 1047f41c0750SAneesh Kumar K.V * when we are loading the buddy cache 1048f41c0750SAneesh Kumar K.V */ 1049f41c0750SAneesh Kumar K.V up_read(e4b->alloc_semp); 1050f41c0750SAneesh Kumar K.V /* 1051f41c0750SAneesh Kumar K.V * we need full data about the group 1052f41c0750SAneesh Kumar K.V * to make a good selection 1053f41c0750SAneesh Kumar K.V */ 1054f41c0750SAneesh Kumar K.V ret = ext4_mb_init_group(sb, group); 1055f41c0750SAneesh Kumar K.V if (ret) 1056f41c0750SAneesh Kumar K.V return ret; 1057f41c0750SAneesh Kumar K.V goto repeat_load_buddy; 1058f41c0750SAneesh Kumar K.V } 1059f41c0750SAneesh Kumar K.V 1060c9de560dSAlex Tomas /* 1061c9de560dSAlex Tomas * the buddy cache inode stores the block bitmap 1062c9de560dSAlex Tomas * and buddy information in consecutive blocks. 1063c9de560dSAlex Tomas * So for each group we need two blocks. 1064c9de560dSAlex Tomas */ 1065c9de560dSAlex Tomas block = group * 2; 1066c9de560dSAlex Tomas pnum = block / blocks_per_page; 1067c9de560dSAlex Tomas poff = block % blocks_per_page; 1068c9de560dSAlex Tomas 1069c9de560dSAlex Tomas /* we could use find_or_create_page(), but it locks page 1070c9de560dSAlex Tomas * what we'd like to avoid in fast path ... */ 1071c9de560dSAlex Tomas page = find_get_page(inode->i_mapping, pnum); 1072c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1073c9de560dSAlex Tomas if (page) 1074920313a7SAneesh Kumar K.V /* 1075920313a7SAneesh Kumar K.V * drop the page reference and try 1076920313a7SAneesh Kumar K.V * to get the page with lock. If we 1077920313a7SAneesh Kumar K.V * are not uptodate that implies 1078920313a7SAneesh Kumar K.V * somebody just created the page but 1079920313a7SAneesh Kumar K.V * is yet to initialize the same. So 1080920313a7SAneesh Kumar K.V * wait for it to initialize. 1081920313a7SAneesh Kumar K.V */ 1082c9de560dSAlex Tomas page_cache_release(page); 1083c9de560dSAlex Tomas page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1084c9de560dSAlex Tomas if (page) { 1085c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 1086c9de560dSAlex Tomas if (!PageUptodate(page)) { 1087fdf6c7a7SShen Feng ret = ext4_mb_init_cache(page, NULL); 1088fdf6c7a7SShen Feng if (ret) { 1089fdf6c7a7SShen Feng unlock_page(page); 1090fdf6c7a7SShen Feng goto err; 1091fdf6c7a7SShen Feng } 1092c9de560dSAlex Tomas mb_cmp_bitmaps(e4b, page_address(page) + 1093c9de560dSAlex Tomas (poff * sb->s_blocksize)); 1094c9de560dSAlex Tomas } 1095c9de560dSAlex Tomas unlock_page(page); 1096c9de560dSAlex Tomas } 1097c9de560dSAlex Tomas } 1098fdf6c7a7SShen Feng if (page == NULL || !PageUptodate(page)) { 1099fdf6c7a7SShen Feng ret = -EIO; 1100c9de560dSAlex Tomas goto err; 1101fdf6c7a7SShen Feng } 1102c9de560dSAlex Tomas e4b->bd_bitmap_page = page; 1103c9de560dSAlex Tomas e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1104c9de560dSAlex Tomas mark_page_accessed(page); 1105c9de560dSAlex Tomas 1106c9de560dSAlex Tomas block++; 1107c9de560dSAlex Tomas pnum = block / blocks_per_page; 1108c9de560dSAlex Tomas poff = block % blocks_per_page; 1109c9de560dSAlex Tomas 1110c9de560dSAlex Tomas page = find_get_page(inode->i_mapping, pnum); 1111c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1112c9de560dSAlex Tomas if (page) 1113c9de560dSAlex Tomas page_cache_release(page); 1114c9de560dSAlex Tomas page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1115c9de560dSAlex Tomas if (page) { 1116c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 1117fdf6c7a7SShen Feng if (!PageUptodate(page)) { 1118fdf6c7a7SShen Feng ret = ext4_mb_init_cache(page, e4b->bd_bitmap); 1119fdf6c7a7SShen Feng if (ret) { 1120fdf6c7a7SShen Feng unlock_page(page); 1121fdf6c7a7SShen Feng goto err; 1122fdf6c7a7SShen Feng } 1123fdf6c7a7SShen Feng } 1124c9de560dSAlex Tomas unlock_page(page); 1125c9de560dSAlex Tomas } 1126c9de560dSAlex Tomas } 1127fdf6c7a7SShen Feng if (page == NULL || !PageUptodate(page)) { 1128fdf6c7a7SShen Feng ret = -EIO; 1129c9de560dSAlex Tomas goto err; 1130fdf6c7a7SShen Feng } 1131c9de560dSAlex Tomas e4b->bd_buddy_page = page; 1132c9de560dSAlex Tomas e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 1133c9de560dSAlex Tomas mark_page_accessed(page); 1134c9de560dSAlex Tomas 1135c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 1136c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 1137c9de560dSAlex Tomas 1138c9de560dSAlex Tomas return 0; 1139c9de560dSAlex Tomas 1140c9de560dSAlex Tomas err: 1141c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 1142c9de560dSAlex Tomas page_cache_release(e4b->bd_bitmap_page); 1143c9de560dSAlex Tomas if (e4b->bd_buddy_page) 1144c9de560dSAlex Tomas page_cache_release(e4b->bd_buddy_page); 1145c9de560dSAlex Tomas e4b->bd_buddy = NULL; 1146c9de560dSAlex Tomas e4b->bd_bitmap = NULL; 1147920313a7SAneesh Kumar K.V 1148920313a7SAneesh Kumar K.V /* Done with the buddy cache */ 1149920313a7SAneesh Kumar K.V up_read(e4b->alloc_semp); 1150fdf6c7a7SShen Feng return ret; 1151c9de560dSAlex Tomas } 1152c9de560dSAlex Tomas 1153c9de560dSAlex Tomas static void ext4_mb_release_desc(struct ext4_buddy *e4b) 1154c9de560dSAlex Tomas { 1155c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 1156c9de560dSAlex Tomas page_cache_release(e4b->bd_bitmap_page); 1157c9de560dSAlex Tomas if (e4b->bd_buddy_page) 1158c9de560dSAlex Tomas page_cache_release(e4b->bd_buddy_page); 1159920313a7SAneesh Kumar K.V /* Done with the buddy cache */ 11608556e8f3SAneesh Kumar K.V if (e4b->alloc_semp) 1161920313a7SAneesh Kumar K.V up_read(e4b->alloc_semp); 1162c9de560dSAlex Tomas } 1163c9de560dSAlex Tomas 1164c9de560dSAlex Tomas 1165c9de560dSAlex Tomas static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1166c9de560dSAlex Tomas { 1167c9de560dSAlex Tomas int order = 1; 1168c9de560dSAlex Tomas void *bb; 1169c9de560dSAlex Tomas 1170c9de560dSAlex Tomas BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); 1171c9de560dSAlex Tomas BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1172c9de560dSAlex Tomas 1173c9de560dSAlex Tomas bb = EXT4_MB_BUDDY(e4b); 1174c9de560dSAlex Tomas while (order <= e4b->bd_blkbits + 1) { 1175c9de560dSAlex Tomas block = block >> 1; 1176c9de560dSAlex Tomas if (!mb_test_bit(block, bb)) { 1177c9de560dSAlex Tomas /* this block is part of buddy of order 'order' */ 1178c9de560dSAlex Tomas return order; 1179c9de560dSAlex Tomas } 1180c9de560dSAlex Tomas bb += 1 << (e4b->bd_blkbits - order); 1181c9de560dSAlex Tomas order++; 1182c9de560dSAlex Tomas } 1183c9de560dSAlex Tomas return 0; 1184c9de560dSAlex Tomas } 1185c9de560dSAlex Tomas 1186955ce5f5SAneesh Kumar K.V static void mb_clear_bits(void *bm, int cur, int len) 1187c9de560dSAlex Tomas { 1188c9de560dSAlex Tomas __u32 *addr; 1189c9de560dSAlex Tomas 1190c9de560dSAlex Tomas len = cur + len; 1191c9de560dSAlex Tomas while (cur < len) { 1192c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1193c9de560dSAlex Tomas /* fast path: clear whole word at once */ 1194c9de560dSAlex Tomas addr = bm + (cur >> 3); 1195c9de560dSAlex Tomas *addr = 0; 1196c9de560dSAlex Tomas cur += 32; 1197c9de560dSAlex Tomas continue; 1198c9de560dSAlex Tomas } 1199e8134b27SAneesh Kumar K.V mb_clear_bit(cur, bm); 1200c9de560dSAlex Tomas cur++; 1201c9de560dSAlex Tomas } 1202c9de560dSAlex Tomas } 1203c9de560dSAlex Tomas 1204955ce5f5SAneesh Kumar K.V static void mb_set_bits(void *bm, int cur, int len) 1205c9de560dSAlex Tomas { 1206c9de560dSAlex Tomas __u32 *addr; 1207c9de560dSAlex Tomas 1208c9de560dSAlex Tomas len = cur + len; 1209c9de560dSAlex Tomas while (cur < len) { 1210c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1211c9de560dSAlex Tomas /* fast path: set whole word at once */ 1212c9de560dSAlex Tomas addr = bm + (cur >> 3); 1213c9de560dSAlex Tomas *addr = 0xffffffff; 1214c9de560dSAlex Tomas cur += 32; 1215c9de560dSAlex Tomas continue; 1216c9de560dSAlex Tomas } 1217e8134b27SAneesh Kumar K.V mb_set_bit(cur, bm); 1218c9de560dSAlex Tomas cur++; 1219c9de560dSAlex Tomas } 1220c9de560dSAlex Tomas } 1221c9de560dSAlex Tomas 12227e5a8cddSShen Feng static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1223c9de560dSAlex Tomas int first, int count) 1224c9de560dSAlex Tomas { 1225c9de560dSAlex Tomas int block = 0; 1226c9de560dSAlex Tomas int max = 0; 1227c9de560dSAlex Tomas int order; 1228c9de560dSAlex Tomas void *buddy; 1229c9de560dSAlex Tomas void *buddy2; 1230c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 1231c9de560dSAlex Tomas 1232c9de560dSAlex Tomas BUG_ON(first + count > (sb->s_blocksize << 3)); 1233bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1234c9de560dSAlex Tomas mb_check_buddy(e4b); 1235c9de560dSAlex Tomas mb_free_blocks_double(inode, e4b, first, count); 1236c9de560dSAlex Tomas 1237c9de560dSAlex Tomas e4b->bd_info->bb_free += count; 1238c9de560dSAlex Tomas if (first < e4b->bd_info->bb_first_free) 1239c9de560dSAlex Tomas e4b->bd_info->bb_first_free = first; 1240c9de560dSAlex Tomas 1241c9de560dSAlex Tomas /* let's maintain fragments counter */ 1242c9de560dSAlex Tomas if (first != 0) 1243c9de560dSAlex Tomas block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); 1244c9de560dSAlex Tomas if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) 1245c9de560dSAlex Tomas max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); 1246c9de560dSAlex Tomas if (block && max) 1247c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1248c9de560dSAlex Tomas else if (!block && !max) 1249c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1250c9de560dSAlex Tomas 1251c9de560dSAlex Tomas /* let's maintain buddy itself */ 1252c9de560dSAlex Tomas while (count-- > 0) { 1253c9de560dSAlex Tomas block = first++; 1254c9de560dSAlex Tomas order = 0; 1255c9de560dSAlex Tomas 1256c9de560dSAlex Tomas if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { 1257c9de560dSAlex Tomas ext4_fsblk_t blocknr; 1258c9de560dSAlex Tomas blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); 1259c9de560dSAlex Tomas blocknr += block; 1260c9de560dSAlex Tomas blocknr += 1261c9de560dSAlex Tomas le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 12625d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 12635d1b1b3fSAneesh Kumar K.V __func__, "double-free of inode" 1264a9df9a49STheodore Ts'o " %lu's block %llu(bit %u in group %u)", 1265c9de560dSAlex Tomas inode ? inode->i_ino : 0, blocknr, block, 1266c9de560dSAlex Tomas e4b->bd_group); 1267c9de560dSAlex Tomas } 1268c9de560dSAlex Tomas mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); 1269c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]++; 1270c9de560dSAlex Tomas 1271c9de560dSAlex Tomas /* start of the buddy */ 1272c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 1273c9de560dSAlex Tomas 1274c9de560dSAlex Tomas do { 1275c9de560dSAlex Tomas block &= ~1UL; 1276c9de560dSAlex Tomas if (mb_test_bit(block, buddy) || 1277c9de560dSAlex Tomas mb_test_bit(block + 1, buddy)) 1278c9de560dSAlex Tomas break; 1279c9de560dSAlex Tomas 1280c9de560dSAlex Tomas /* both the buddies are free, try to coalesce them */ 1281c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order + 1, &max); 1282c9de560dSAlex Tomas 1283c9de560dSAlex Tomas if (!buddy2) 1284c9de560dSAlex Tomas break; 1285c9de560dSAlex Tomas 1286c9de560dSAlex Tomas if (order > 0) { 1287c9de560dSAlex Tomas /* for special purposes, we don't set 1288c9de560dSAlex Tomas * free bits in bitmap */ 1289c9de560dSAlex Tomas mb_set_bit(block, buddy); 1290c9de560dSAlex Tomas mb_set_bit(block + 1, buddy); 1291c9de560dSAlex Tomas } 1292c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]--; 1293c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]--; 1294c9de560dSAlex Tomas 1295c9de560dSAlex Tomas block = block >> 1; 1296c9de560dSAlex Tomas order++; 1297c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]++; 1298c9de560dSAlex Tomas 1299c9de560dSAlex Tomas mb_clear_bit(block, buddy2); 1300c9de560dSAlex Tomas buddy = buddy2; 1301c9de560dSAlex Tomas } while (1); 1302c9de560dSAlex Tomas } 1303c9de560dSAlex Tomas mb_check_buddy(e4b); 1304c9de560dSAlex Tomas } 1305c9de560dSAlex Tomas 1306c9de560dSAlex Tomas static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, 1307c9de560dSAlex Tomas int needed, struct ext4_free_extent *ex) 1308c9de560dSAlex Tomas { 1309c9de560dSAlex Tomas int next = block; 1310c9de560dSAlex Tomas int max; 1311c9de560dSAlex Tomas int ord; 1312c9de560dSAlex Tomas void *buddy; 1313c9de560dSAlex Tomas 1314bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1315c9de560dSAlex Tomas BUG_ON(ex == NULL); 1316c9de560dSAlex Tomas 1317c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 1318c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1319c9de560dSAlex Tomas BUG_ON(block >= max); 1320c9de560dSAlex Tomas if (mb_test_bit(block, buddy)) { 1321c9de560dSAlex Tomas ex->fe_len = 0; 1322c9de560dSAlex Tomas ex->fe_start = 0; 1323c9de560dSAlex Tomas ex->fe_group = 0; 1324c9de560dSAlex Tomas return 0; 1325c9de560dSAlex Tomas } 1326c9de560dSAlex Tomas 1327c9de560dSAlex Tomas /* FIXME dorp order completely ? */ 1328c9de560dSAlex Tomas if (likely(order == 0)) { 1329c9de560dSAlex Tomas /* find actual order */ 1330c9de560dSAlex Tomas order = mb_find_order_for_block(e4b, block); 1331c9de560dSAlex Tomas block = block >> order; 1332c9de560dSAlex Tomas } 1333c9de560dSAlex Tomas 1334c9de560dSAlex Tomas ex->fe_len = 1 << order; 1335c9de560dSAlex Tomas ex->fe_start = block << order; 1336c9de560dSAlex Tomas ex->fe_group = e4b->bd_group; 1337c9de560dSAlex Tomas 1338c9de560dSAlex Tomas /* calc difference from given start */ 1339c9de560dSAlex Tomas next = next - ex->fe_start; 1340c9de560dSAlex Tomas ex->fe_len -= next; 1341c9de560dSAlex Tomas ex->fe_start += next; 1342c9de560dSAlex Tomas 1343c9de560dSAlex Tomas while (needed > ex->fe_len && 1344c9de560dSAlex Tomas (buddy = mb_find_buddy(e4b, order, &max))) { 1345c9de560dSAlex Tomas 1346c9de560dSAlex Tomas if (block + 1 >= max) 1347c9de560dSAlex Tomas break; 1348c9de560dSAlex Tomas 1349c9de560dSAlex Tomas next = (block + 1) * (1 << order); 1350c9de560dSAlex Tomas if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) 1351c9de560dSAlex Tomas break; 1352c9de560dSAlex Tomas 1353c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, next); 1354c9de560dSAlex Tomas 1355c9de560dSAlex Tomas order = ord; 1356c9de560dSAlex Tomas block = next >> order; 1357c9de560dSAlex Tomas ex->fe_len += 1 << order; 1358c9de560dSAlex Tomas } 1359c9de560dSAlex Tomas 1360c9de560dSAlex Tomas BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); 1361c9de560dSAlex Tomas return ex->fe_len; 1362c9de560dSAlex Tomas } 1363c9de560dSAlex Tomas 1364c9de560dSAlex Tomas static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 1365c9de560dSAlex Tomas { 1366c9de560dSAlex Tomas int ord; 1367c9de560dSAlex Tomas int mlen = 0; 1368c9de560dSAlex Tomas int max = 0; 1369c9de560dSAlex Tomas int cur; 1370c9de560dSAlex Tomas int start = ex->fe_start; 1371c9de560dSAlex Tomas int len = ex->fe_len; 1372c9de560dSAlex Tomas unsigned ret = 0; 1373c9de560dSAlex Tomas int len0 = len; 1374c9de560dSAlex Tomas void *buddy; 1375c9de560dSAlex Tomas 1376c9de560dSAlex Tomas BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 1377c9de560dSAlex Tomas BUG_ON(e4b->bd_group != ex->fe_group); 1378bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1379c9de560dSAlex Tomas mb_check_buddy(e4b); 1380c9de560dSAlex Tomas mb_mark_used_double(e4b, start, len); 1381c9de560dSAlex Tomas 1382c9de560dSAlex Tomas e4b->bd_info->bb_free -= len; 1383c9de560dSAlex Tomas if (e4b->bd_info->bb_first_free == start) 1384c9de560dSAlex Tomas e4b->bd_info->bb_first_free += len; 1385c9de560dSAlex Tomas 1386c9de560dSAlex Tomas /* let's maintain fragments counter */ 1387c9de560dSAlex Tomas if (start != 0) 1388c9de560dSAlex Tomas mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); 1389c9de560dSAlex Tomas if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 1390c9de560dSAlex Tomas max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); 1391c9de560dSAlex Tomas if (mlen && max) 1392c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1393c9de560dSAlex Tomas else if (!mlen && !max) 1394c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1395c9de560dSAlex Tomas 1396c9de560dSAlex Tomas /* let's maintain buddy itself */ 1397c9de560dSAlex Tomas while (len) { 1398c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, start); 1399c9de560dSAlex Tomas 1400c9de560dSAlex Tomas if (((start >> ord) << ord) == start && len >= (1 << ord)) { 1401c9de560dSAlex Tomas /* the whole chunk may be allocated at once! */ 1402c9de560dSAlex Tomas mlen = 1 << ord; 1403c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1404c9de560dSAlex Tomas BUG_ON((start >> ord) >= max); 1405c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1406c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1407c9de560dSAlex Tomas start += mlen; 1408c9de560dSAlex Tomas len -= mlen; 1409c9de560dSAlex Tomas BUG_ON(len < 0); 1410c9de560dSAlex Tomas continue; 1411c9de560dSAlex Tomas } 1412c9de560dSAlex Tomas 1413c9de560dSAlex Tomas /* store for history */ 1414c9de560dSAlex Tomas if (ret == 0) 1415c9de560dSAlex Tomas ret = len | (ord << 16); 1416c9de560dSAlex Tomas 1417c9de560dSAlex Tomas /* we have to split large buddy */ 1418c9de560dSAlex Tomas BUG_ON(ord <= 0); 1419c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1420c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1421c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1422c9de560dSAlex Tomas 1423c9de560dSAlex Tomas ord--; 1424c9de560dSAlex Tomas cur = (start >> ord) & ~1U; 1425c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1426c9de560dSAlex Tomas mb_clear_bit(cur, buddy); 1427c9de560dSAlex Tomas mb_clear_bit(cur + 1, buddy); 1428c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1429c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1430c9de560dSAlex Tomas } 1431c9de560dSAlex Tomas 1432955ce5f5SAneesh Kumar K.V mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0); 1433c9de560dSAlex Tomas mb_check_buddy(e4b); 1434c9de560dSAlex Tomas 1435c9de560dSAlex Tomas return ret; 1436c9de560dSAlex Tomas } 1437c9de560dSAlex Tomas 1438c9de560dSAlex Tomas /* 1439c9de560dSAlex Tomas * Must be called under group lock! 1440c9de560dSAlex Tomas */ 1441c9de560dSAlex Tomas static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 1442c9de560dSAlex Tomas struct ext4_buddy *e4b) 1443c9de560dSAlex Tomas { 1444c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1445c9de560dSAlex Tomas int ret; 1446c9de560dSAlex Tomas 1447c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 1448c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1449c9de560dSAlex Tomas 1450c9de560dSAlex Tomas ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 1451c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 1452c9de560dSAlex Tomas ret = mb_mark_used(e4b, &ac->ac_b_ex); 1453c9de560dSAlex Tomas 1454c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 1455c9de560dSAlex Tomas * allocated blocks for history */ 1456c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 1457c9de560dSAlex Tomas 1458c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 1459c9de560dSAlex Tomas ac->ac_tail = ret & 0xffff; 1460c9de560dSAlex Tomas ac->ac_buddy = ret >> 16; 1461c9de560dSAlex Tomas 1462c3a326a6SAneesh Kumar K.V /* 1463c3a326a6SAneesh Kumar K.V * take the page reference. We want the page to be pinned 1464c3a326a6SAneesh Kumar K.V * so that we don't get a ext4_mb_init_cache_call for this 1465c3a326a6SAneesh Kumar K.V * group until we update the bitmap. That would mean we 1466c3a326a6SAneesh Kumar K.V * double allocate blocks. The reference is dropped 1467c3a326a6SAneesh Kumar K.V * in ext4_mb_release_context 1468c3a326a6SAneesh Kumar K.V */ 1469c9de560dSAlex Tomas ac->ac_bitmap_page = e4b->bd_bitmap_page; 1470c9de560dSAlex Tomas get_page(ac->ac_bitmap_page); 1471c9de560dSAlex Tomas ac->ac_buddy_page = e4b->bd_buddy_page; 1472c9de560dSAlex Tomas get_page(ac->ac_buddy_page); 14738556e8f3SAneesh Kumar K.V /* on allocation we use ac to track the held semaphore */ 14748556e8f3SAneesh Kumar K.V ac->alloc_semp = e4b->alloc_semp; 14758556e8f3SAneesh Kumar K.V e4b->alloc_semp = NULL; 1476c9de560dSAlex Tomas /* store last allocated for subsequent stream allocation */ 14774ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 1478c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 1479c9de560dSAlex Tomas sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 1480c9de560dSAlex Tomas sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 1481c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 1482c9de560dSAlex Tomas } 1483c9de560dSAlex Tomas } 1484c9de560dSAlex Tomas 1485c9de560dSAlex Tomas /* 1486c9de560dSAlex Tomas * regular allocator, for general purposes allocation 1487c9de560dSAlex Tomas */ 1488c9de560dSAlex Tomas 1489c9de560dSAlex Tomas static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 1490c9de560dSAlex Tomas struct ext4_buddy *e4b, 1491c9de560dSAlex Tomas int finish_group) 1492c9de560dSAlex Tomas { 1493c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1494c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1495c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1496c9de560dSAlex Tomas struct ext4_free_extent ex; 1497c9de560dSAlex Tomas int max; 1498c9de560dSAlex Tomas 1499032115fcSAneesh Kumar K.V if (ac->ac_status == AC_STATUS_FOUND) 1500032115fcSAneesh Kumar K.V return; 1501c9de560dSAlex Tomas /* 1502c9de560dSAlex Tomas * We don't want to scan for a whole year 1503c9de560dSAlex Tomas */ 1504c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan && 1505c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1506c9de560dSAlex Tomas ac->ac_status = AC_STATUS_BREAK; 1507c9de560dSAlex Tomas return; 1508c9de560dSAlex Tomas } 1509c9de560dSAlex Tomas 1510c9de560dSAlex Tomas /* 1511c9de560dSAlex Tomas * Haven't found good chunk so far, let's continue 1512c9de560dSAlex Tomas */ 1513c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) 1514c9de560dSAlex Tomas return; 1515c9de560dSAlex Tomas 1516c9de560dSAlex Tomas if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 1517c9de560dSAlex Tomas && bex->fe_group == e4b->bd_group) { 1518c9de560dSAlex Tomas /* recheck chunk's availability - we don't know 1519c9de560dSAlex Tomas * when it was found (within this lock-unlock 1520c9de560dSAlex Tomas * period or not) */ 1521c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); 1522c9de560dSAlex Tomas if (max >= gex->fe_len) { 1523c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1524c9de560dSAlex Tomas return; 1525c9de560dSAlex Tomas } 1526c9de560dSAlex Tomas } 1527c9de560dSAlex Tomas } 1528c9de560dSAlex Tomas 1529c9de560dSAlex Tomas /* 1530c9de560dSAlex Tomas * The routine checks whether found extent is good enough. If it is, 1531c9de560dSAlex Tomas * then the extent gets marked used and flag is set to the context 1532c9de560dSAlex Tomas * to stop scanning. Otherwise, the extent is compared with the 1533c9de560dSAlex Tomas * previous found extent and if new one is better, then it's stored 1534c9de560dSAlex Tomas * in the context. Later, the best found extent will be used, if 1535c9de560dSAlex Tomas * mballoc can't find good enough extent. 1536c9de560dSAlex Tomas * 1537c9de560dSAlex Tomas * FIXME: real allocation policy is to be designed yet! 1538c9de560dSAlex Tomas */ 1539c9de560dSAlex Tomas static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 1540c9de560dSAlex Tomas struct ext4_free_extent *ex, 1541c9de560dSAlex Tomas struct ext4_buddy *e4b) 1542c9de560dSAlex Tomas { 1543c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1544c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1545c9de560dSAlex Tomas 1546c9de560dSAlex Tomas BUG_ON(ex->fe_len <= 0); 15478d03c7a0SEric Sandeen BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 1548c9de560dSAlex Tomas BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 1549c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 1550c9de560dSAlex Tomas 1551c9de560dSAlex Tomas ac->ac_found++; 1552c9de560dSAlex Tomas 1553c9de560dSAlex Tomas /* 1554c9de560dSAlex Tomas * The special case - take what you catch first 1555c9de560dSAlex Tomas */ 1556c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1557c9de560dSAlex Tomas *bex = *ex; 1558c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1559c9de560dSAlex Tomas return; 1560c9de560dSAlex Tomas } 1561c9de560dSAlex Tomas 1562c9de560dSAlex Tomas /* 1563c9de560dSAlex Tomas * Let's check whether the chuck is good enough 1564c9de560dSAlex Tomas */ 1565c9de560dSAlex Tomas if (ex->fe_len == gex->fe_len) { 1566c9de560dSAlex Tomas *bex = *ex; 1567c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1568c9de560dSAlex Tomas return; 1569c9de560dSAlex Tomas } 1570c9de560dSAlex Tomas 1571c9de560dSAlex Tomas /* 1572c9de560dSAlex Tomas * If this is first found extent, just store it in the context 1573c9de560dSAlex Tomas */ 1574c9de560dSAlex Tomas if (bex->fe_len == 0) { 1575c9de560dSAlex Tomas *bex = *ex; 1576c9de560dSAlex Tomas return; 1577c9de560dSAlex Tomas } 1578c9de560dSAlex Tomas 1579c9de560dSAlex Tomas /* 1580c9de560dSAlex Tomas * If new found extent is better, store it in the context 1581c9de560dSAlex Tomas */ 1582c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) { 1583c9de560dSAlex Tomas /* if the request isn't satisfied, any found extent 1584c9de560dSAlex Tomas * larger than previous best one is better */ 1585c9de560dSAlex Tomas if (ex->fe_len > bex->fe_len) 1586c9de560dSAlex Tomas *bex = *ex; 1587c9de560dSAlex Tomas } else if (ex->fe_len > gex->fe_len) { 1588c9de560dSAlex Tomas /* if the request is satisfied, then we try to find 1589c9de560dSAlex Tomas * an extent that still satisfy the request, but is 1590c9de560dSAlex Tomas * smaller than previous one */ 1591c9de560dSAlex Tomas if (ex->fe_len < bex->fe_len) 1592c9de560dSAlex Tomas *bex = *ex; 1593c9de560dSAlex Tomas } 1594c9de560dSAlex Tomas 1595c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 0); 1596c9de560dSAlex Tomas } 1597c9de560dSAlex Tomas 1598089ceeccSEric Sandeen static noinline_for_stack 1599089ceeccSEric Sandeen int ext4_mb_try_best_found(struct ext4_allocation_context *ac, 1600c9de560dSAlex Tomas struct ext4_buddy *e4b) 1601c9de560dSAlex Tomas { 1602c9de560dSAlex Tomas struct ext4_free_extent ex = ac->ac_b_ex; 1603c9de560dSAlex Tomas ext4_group_t group = ex.fe_group; 1604c9de560dSAlex Tomas int max; 1605c9de560dSAlex Tomas int err; 1606c9de560dSAlex Tomas 1607c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1608c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1609c9de560dSAlex Tomas if (err) 1610c9de560dSAlex Tomas return err; 1611c9de560dSAlex Tomas 1612c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 1613c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); 1614c9de560dSAlex Tomas 1615c9de560dSAlex Tomas if (max > 0) { 1616c9de560dSAlex Tomas ac->ac_b_ex = ex; 1617c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1618c9de560dSAlex Tomas } 1619c9de560dSAlex Tomas 1620c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1621c9de560dSAlex Tomas ext4_mb_release_desc(e4b); 1622c9de560dSAlex Tomas 1623c9de560dSAlex Tomas return 0; 1624c9de560dSAlex Tomas } 1625c9de560dSAlex Tomas 1626089ceeccSEric Sandeen static noinline_for_stack 1627089ceeccSEric Sandeen int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 1628c9de560dSAlex Tomas struct ext4_buddy *e4b) 1629c9de560dSAlex Tomas { 1630c9de560dSAlex Tomas ext4_group_t group = ac->ac_g_ex.fe_group; 1631c9de560dSAlex Tomas int max; 1632c9de560dSAlex Tomas int err; 1633c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1634c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 1635c9de560dSAlex Tomas struct ext4_free_extent ex; 1636c9de560dSAlex Tomas 1637c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) 1638c9de560dSAlex Tomas return 0; 1639c9de560dSAlex Tomas 1640c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1641c9de560dSAlex Tomas if (err) 1642c9de560dSAlex Tomas return err; 1643c9de560dSAlex Tomas 1644c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 1645c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, 1646c9de560dSAlex Tomas ac->ac_g_ex.fe_len, &ex); 1647c9de560dSAlex Tomas 1648c9de560dSAlex Tomas if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { 1649c9de560dSAlex Tomas ext4_fsblk_t start; 1650c9de560dSAlex Tomas 1651c9de560dSAlex Tomas start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) + 1652c9de560dSAlex Tomas ex.fe_start + le32_to_cpu(es->s_first_data_block); 1653c9de560dSAlex Tomas /* use do_div to get remainder (would be 64-bit modulo) */ 1654c9de560dSAlex Tomas if (do_div(start, sbi->s_stripe) == 0) { 1655c9de560dSAlex Tomas ac->ac_found++; 1656c9de560dSAlex Tomas ac->ac_b_ex = ex; 1657c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1658c9de560dSAlex Tomas } 1659c9de560dSAlex Tomas } else if (max >= ac->ac_g_ex.fe_len) { 1660c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1661c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1662c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1663c9de560dSAlex Tomas ac->ac_found++; 1664c9de560dSAlex Tomas ac->ac_b_ex = ex; 1665c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1666c9de560dSAlex Tomas } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 1667c9de560dSAlex Tomas /* Sometimes, caller may want to merge even small 1668c9de560dSAlex Tomas * number of blocks to an existing extent */ 1669c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1670c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1671c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1672c9de560dSAlex Tomas ac->ac_found++; 1673c9de560dSAlex Tomas ac->ac_b_ex = ex; 1674c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1675c9de560dSAlex Tomas } 1676c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1677c9de560dSAlex Tomas ext4_mb_release_desc(e4b); 1678c9de560dSAlex Tomas 1679c9de560dSAlex Tomas return 0; 1680c9de560dSAlex Tomas } 1681c9de560dSAlex Tomas 1682c9de560dSAlex Tomas /* 1683c9de560dSAlex Tomas * The routine scans buddy structures (not bitmap!) from given order 1684c9de560dSAlex Tomas * to max order and tries to find big enough chunk to satisfy the req 1685c9de560dSAlex Tomas */ 1686089ceeccSEric Sandeen static noinline_for_stack 1687089ceeccSEric Sandeen void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 1688c9de560dSAlex Tomas struct ext4_buddy *e4b) 1689c9de560dSAlex Tomas { 1690c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1691c9de560dSAlex Tomas struct ext4_group_info *grp = e4b->bd_info; 1692c9de560dSAlex Tomas void *buddy; 1693c9de560dSAlex Tomas int i; 1694c9de560dSAlex Tomas int k; 1695c9de560dSAlex Tomas int max; 1696c9de560dSAlex Tomas 1697c9de560dSAlex Tomas BUG_ON(ac->ac_2order <= 0); 1698c9de560dSAlex Tomas for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { 1699c9de560dSAlex Tomas if (grp->bb_counters[i] == 0) 1700c9de560dSAlex Tomas continue; 1701c9de560dSAlex Tomas 1702c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, i, &max); 1703c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1704c9de560dSAlex Tomas 1705ffad0a44SAneesh Kumar K.V k = mb_find_next_zero_bit(buddy, max, 0); 1706c9de560dSAlex Tomas BUG_ON(k >= max); 1707c9de560dSAlex Tomas 1708c9de560dSAlex Tomas ac->ac_found++; 1709c9de560dSAlex Tomas 1710c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 1 << i; 1711c9de560dSAlex Tomas ac->ac_b_ex.fe_start = k << i; 1712c9de560dSAlex Tomas ac->ac_b_ex.fe_group = e4b->bd_group; 1713c9de560dSAlex Tomas 1714c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1715c9de560dSAlex Tomas 1716c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); 1717c9de560dSAlex Tomas 1718c9de560dSAlex Tomas if (EXT4_SB(sb)->s_mb_stats) 1719c9de560dSAlex Tomas atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 1720c9de560dSAlex Tomas 1721c9de560dSAlex Tomas break; 1722c9de560dSAlex Tomas } 1723c9de560dSAlex Tomas } 1724c9de560dSAlex Tomas 1725c9de560dSAlex Tomas /* 1726c9de560dSAlex Tomas * The routine scans the group and measures all found extents. 1727c9de560dSAlex Tomas * In order to optimize scanning, caller must pass number of 1728c9de560dSAlex Tomas * free blocks in the group, so the routine can know upper limit. 1729c9de560dSAlex Tomas */ 1730089ceeccSEric Sandeen static noinline_for_stack 1731089ceeccSEric Sandeen void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 1732c9de560dSAlex Tomas struct ext4_buddy *e4b) 1733c9de560dSAlex Tomas { 1734c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1735c9de560dSAlex Tomas void *bitmap = EXT4_MB_BITMAP(e4b); 1736c9de560dSAlex Tomas struct ext4_free_extent ex; 1737c9de560dSAlex Tomas int i; 1738c9de560dSAlex Tomas int free; 1739c9de560dSAlex Tomas 1740c9de560dSAlex Tomas free = e4b->bd_info->bb_free; 1741c9de560dSAlex Tomas BUG_ON(free <= 0); 1742c9de560dSAlex Tomas 1743c9de560dSAlex Tomas i = e4b->bd_info->bb_first_free; 1744c9de560dSAlex Tomas 1745c9de560dSAlex Tomas while (free && ac->ac_status == AC_STATUS_CONTINUE) { 1746ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, 1747c9de560dSAlex Tomas EXT4_BLOCKS_PER_GROUP(sb), i); 1748c9de560dSAlex Tomas if (i >= EXT4_BLOCKS_PER_GROUP(sb)) { 174926346ff6SAneesh Kumar K.V /* 1750e56eb659SAneesh Kumar K.V * IF we have corrupt bitmap, we won't find any 175126346ff6SAneesh Kumar K.V * free blocks even though group info says we 175226346ff6SAneesh Kumar K.V * we have free blocks 175326346ff6SAneesh Kumar K.V */ 17545d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 17555d1b1b3fSAneesh Kumar K.V __func__, "%d free blocks as per " 1756fde4d95aSTheodore Ts'o "group info. But bitmap says 0", 175726346ff6SAneesh Kumar K.V free); 1758c9de560dSAlex Tomas break; 1759c9de560dSAlex Tomas } 1760c9de560dSAlex Tomas 1761c9de560dSAlex Tomas mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); 1762c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 176326346ff6SAneesh Kumar K.V if (free < ex.fe_len) { 17645d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 17655d1b1b3fSAneesh Kumar K.V __func__, "%d free blocks as per " 1766fde4d95aSTheodore Ts'o "group info. But got %d blocks", 176726346ff6SAneesh Kumar K.V free, ex.fe_len); 1768e56eb659SAneesh Kumar K.V /* 1769e56eb659SAneesh Kumar K.V * The number of free blocks differs. This mostly 1770e56eb659SAneesh Kumar K.V * indicate that the bitmap is corrupt. So exit 1771e56eb659SAneesh Kumar K.V * without claiming the space. 1772e56eb659SAneesh Kumar K.V */ 1773e56eb659SAneesh Kumar K.V break; 177426346ff6SAneesh Kumar K.V } 1775c9de560dSAlex Tomas 1776c9de560dSAlex Tomas ext4_mb_measure_extent(ac, &ex, e4b); 1777c9de560dSAlex Tomas 1778c9de560dSAlex Tomas i += ex.fe_len; 1779c9de560dSAlex Tomas free -= ex.fe_len; 1780c9de560dSAlex Tomas } 1781c9de560dSAlex Tomas 1782c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 1); 1783c9de560dSAlex Tomas } 1784c9de560dSAlex Tomas 1785c9de560dSAlex Tomas /* 1786c9de560dSAlex Tomas * This is a special case for storages like raid5 1787c9de560dSAlex Tomas * we try to find stripe-aligned chunks for stripe-size requests 1788c9de560dSAlex Tomas * XXX should do so at least for multiples of stripe size as well 1789c9de560dSAlex Tomas */ 1790089ceeccSEric Sandeen static noinline_for_stack 1791089ceeccSEric Sandeen void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 1792c9de560dSAlex Tomas struct ext4_buddy *e4b) 1793c9de560dSAlex Tomas { 1794c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1795c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 1796c9de560dSAlex Tomas void *bitmap = EXT4_MB_BITMAP(e4b); 1797c9de560dSAlex Tomas struct ext4_free_extent ex; 1798c9de560dSAlex Tomas ext4_fsblk_t first_group_block; 1799c9de560dSAlex Tomas ext4_fsblk_t a; 1800c9de560dSAlex Tomas ext4_grpblk_t i; 1801c9de560dSAlex Tomas int max; 1802c9de560dSAlex Tomas 1803c9de560dSAlex Tomas BUG_ON(sbi->s_stripe == 0); 1804c9de560dSAlex Tomas 1805c9de560dSAlex Tomas /* find first stripe-aligned block in group */ 1806c9de560dSAlex Tomas first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb) 1807c9de560dSAlex Tomas + le32_to_cpu(sbi->s_es->s_first_data_block); 1808c9de560dSAlex Tomas a = first_group_block + sbi->s_stripe - 1; 1809c9de560dSAlex Tomas do_div(a, sbi->s_stripe); 1810c9de560dSAlex Tomas i = (a * sbi->s_stripe) - first_group_block; 1811c9de560dSAlex Tomas 1812c9de560dSAlex Tomas while (i < EXT4_BLOCKS_PER_GROUP(sb)) { 1813c9de560dSAlex Tomas if (!mb_test_bit(i, bitmap)) { 1814c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); 1815c9de560dSAlex Tomas if (max >= sbi->s_stripe) { 1816c9de560dSAlex Tomas ac->ac_found++; 1817c9de560dSAlex Tomas ac->ac_b_ex = ex; 1818c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1819c9de560dSAlex Tomas break; 1820c9de560dSAlex Tomas } 1821c9de560dSAlex Tomas } 1822c9de560dSAlex Tomas i += sbi->s_stripe; 1823c9de560dSAlex Tomas } 1824c9de560dSAlex Tomas } 1825c9de560dSAlex Tomas 1826c9de560dSAlex Tomas static int ext4_mb_good_group(struct ext4_allocation_context *ac, 1827c9de560dSAlex Tomas ext4_group_t group, int cr) 1828c9de560dSAlex Tomas { 1829c9de560dSAlex Tomas unsigned free, fragments; 1830c9de560dSAlex Tomas unsigned i, bits; 1831a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 1832c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 1833c9de560dSAlex Tomas 1834c9de560dSAlex Tomas BUG_ON(cr < 0 || cr >= 4); 1835c9de560dSAlex Tomas BUG_ON(EXT4_MB_GRP_NEED_INIT(grp)); 1836c9de560dSAlex Tomas 1837c9de560dSAlex Tomas free = grp->bb_free; 1838c9de560dSAlex Tomas fragments = grp->bb_fragments; 1839c9de560dSAlex Tomas if (free == 0) 1840c9de560dSAlex Tomas return 0; 1841c9de560dSAlex Tomas if (fragments == 0) 1842c9de560dSAlex Tomas return 0; 1843c9de560dSAlex Tomas 1844c9de560dSAlex Tomas switch (cr) { 1845c9de560dSAlex Tomas case 0: 1846c9de560dSAlex Tomas BUG_ON(ac->ac_2order == 0); 1847c9de560dSAlex Tomas 1848a4912123STheodore Ts'o /* Avoid using the first bg of a flexgroup for data files */ 1849a4912123STheodore Ts'o if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 1850a4912123STheodore Ts'o (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 1851a4912123STheodore Ts'o ((group % flex_size) == 0)) 1852a4912123STheodore Ts'o return 0; 1853a4912123STheodore Ts'o 1854c9de560dSAlex Tomas bits = ac->ac_sb->s_blocksize_bits + 1; 1855c9de560dSAlex Tomas for (i = ac->ac_2order; i <= bits; i++) 1856c9de560dSAlex Tomas if (grp->bb_counters[i] > 0) 1857c9de560dSAlex Tomas return 1; 1858c9de560dSAlex Tomas break; 1859c9de560dSAlex Tomas case 1: 1860c9de560dSAlex Tomas if ((free / fragments) >= ac->ac_g_ex.fe_len) 1861c9de560dSAlex Tomas return 1; 1862c9de560dSAlex Tomas break; 1863c9de560dSAlex Tomas case 2: 1864c9de560dSAlex Tomas if (free >= ac->ac_g_ex.fe_len) 1865c9de560dSAlex Tomas return 1; 1866c9de560dSAlex Tomas break; 1867c9de560dSAlex Tomas case 3: 1868c9de560dSAlex Tomas return 1; 1869c9de560dSAlex Tomas default: 1870c9de560dSAlex Tomas BUG(); 1871c9de560dSAlex Tomas } 1872c9de560dSAlex Tomas 1873c9de560dSAlex Tomas return 0; 1874c9de560dSAlex Tomas } 1875c9de560dSAlex Tomas 1876920313a7SAneesh Kumar K.V /* 1877920313a7SAneesh Kumar K.V * lock the group_info alloc_sem of all the groups 1878920313a7SAneesh Kumar K.V * belonging to the same buddy cache page. This 1879920313a7SAneesh Kumar K.V * make sure other parallel operation on the buddy 1880920313a7SAneesh Kumar K.V * cache doesn't happen whild holding the buddy cache 1881920313a7SAneesh Kumar K.V * lock 1882920313a7SAneesh Kumar K.V */ 1883920313a7SAneesh Kumar K.V int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group) 1884920313a7SAneesh Kumar K.V { 1885920313a7SAneesh Kumar K.V int i; 1886920313a7SAneesh Kumar K.V int block, pnum; 1887920313a7SAneesh Kumar K.V int blocks_per_page; 1888920313a7SAneesh Kumar K.V int groups_per_page; 18898df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 1890920313a7SAneesh Kumar K.V ext4_group_t first_group; 1891920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1892920313a7SAneesh Kumar K.V 1893920313a7SAneesh Kumar K.V blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1894920313a7SAneesh Kumar K.V /* 1895920313a7SAneesh Kumar K.V * the buddy cache inode stores the block bitmap 1896920313a7SAneesh Kumar K.V * and buddy information in consecutive blocks. 1897920313a7SAneesh Kumar K.V * So for each group we need two blocks. 1898920313a7SAneesh Kumar K.V */ 1899920313a7SAneesh Kumar K.V block = group * 2; 1900920313a7SAneesh Kumar K.V pnum = block / blocks_per_page; 1901920313a7SAneesh Kumar K.V first_group = pnum * blocks_per_page / 2; 1902920313a7SAneesh Kumar K.V 1903920313a7SAneesh Kumar K.V groups_per_page = blocks_per_page >> 1; 1904920313a7SAneesh Kumar K.V if (groups_per_page == 0) 1905920313a7SAneesh Kumar K.V groups_per_page = 1; 1906920313a7SAneesh Kumar K.V /* read all groups the page covers into the cache */ 1907920313a7SAneesh Kumar K.V for (i = 0; i < groups_per_page; i++) { 1908920313a7SAneesh Kumar K.V 19098df9675fSTheodore Ts'o if ((first_group + i) >= ngroups) 1910920313a7SAneesh Kumar K.V break; 1911920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, first_group + i); 1912920313a7SAneesh Kumar K.V /* take all groups write allocation 1913920313a7SAneesh Kumar K.V * semaphore. This make sure there is 1914920313a7SAneesh Kumar K.V * no block allocation going on in any 1915920313a7SAneesh Kumar K.V * of that groups 1916920313a7SAneesh Kumar K.V */ 1917b7be019eSAneesh Kumar K.V down_write_nested(&grp->alloc_sem, i); 1918920313a7SAneesh Kumar K.V } 1919920313a7SAneesh Kumar K.V return i; 1920920313a7SAneesh Kumar K.V } 1921920313a7SAneesh Kumar K.V 1922920313a7SAneesh Kumar K.V void ext4_mb_put_buddy_cache_lock(struct super_block *sb, 1923920313a7SAneesh Kumar K.V ext4_group_t group, int locked_group) 1924920313a7SAneesh Kumar K.V { 1925920313a7SAneesh Kumar K.V int i; 1926920313a7SAneesh Kumar K.V int block, pnum; 1927920313a7SAneesh Kumar K.V int blocks_per_page; 1928920313a7SAneesh Kumar K.V ext4_group_t first_group; 1929920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1930920313a7SAneesh Kumar K.V 1931920313a7SAneesh Kumar K.V blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1932920313a7SAneesh Kumar K.V /* 1933920313a7SAneesh Kumar K.V * the buddy cache inode stores the block bitmap 1934920313a7SAneesh Kumar K.V * and buddy information in consecutive blocks. 1935920313a7SAneesh Kumar K.V * So for each group we need two blocks. 1936920313a7SAneesh Kumar K.V */ 1937920313a7SAneesh Kumar K.V block = group * 2; 1938920313a7SAneesh Kumar K.V pnum = block / blocks_per_page; 1939920313a7SAneesh Kumar K.V first_group = pnum * blocks_per_page / 2; 1940920313a7SAneesh Kumar K.V /* release locks on all the groups */ 1941920313a7SAneesh Kumar K.V for (i = 0; i < locked_group; i++) { 1942920313a7SAneesh Kumar K.V 1943920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, first_group + i); 1944920313a7SAneesh Kumar K.V /* take all groups write allocation 1945920313a7SAneesh Kumar K.V * semaphore. This make sure there is 1946920313a7SAneesh Kumar K.V * no block allocation going on in any 1947920313a7SAneesh Kumar K.V * of that groups 1948920313a7SAneesh Kumar K.V */ 1949920313a7SAneesh Kumar K.V up_write(&grp->alloc_sem); 1950920313a7SAneesh Kumar K.V } 1951920313a7SAneesh Kumar K.V 1952920313a7SAneesh Kumar K.V } 1953920313a7SAneesh Kumar K.V 19544ddfef7bSEric Sandeen static noinline_for_stack int 19554ddfef7bSEric Sandeen ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 1956c9de560dSAlex Tomas { 19578df9675fSTheodore Ts'o ext4_group_t ngroups, group, i; 1958c9de560dSAlex Tomas int cr; 1959c9de560dSAlex Tomas int err = 0; 1960c9de560dSAlex Tomas int bsbits; 1961c9de560dSAlex Tomas struct ext4_sb_info *sbi; 1962c9de560dSAlex Tomas struct super_block *sb; 1963c9de560dSAlex Tomas struct ext4_buddy e4b; 1964c9de560dSAlex Tomas 1965c9de560dSAlex Tomas sb = ac->ac_sb; 1966c9de560dSAlex Tomas sbi = EXT4_SB(sb); 19678df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 1968fb0a387dSEric Sandeen /* non-extent files are limited to low blocks/groups */ 1969fb0a387dSEric Sandeen if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL)) 1970fb0a387dSEric Sandeen ngroups = sbi->s_blockfile_groups; 1971fb0a387dSEric Sandeen 1972c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1973c9de560dSAlex Tomas 1974c9de560dSAlex Tomas /* first, try the goal */ 1975c9de560dSAlex Tomas err = ext4_mb_find_by_goal(ac, &e4b); 1976c9de560dSAlex Tomas if (err || ac->ac_status == AC_STATUS_FOUND) 1977c9de560dSAlex Tomas goto out; 1978c9de560dSAlex Tomas 1979c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 1980c9de560dSAlex Tomas goto out; 1981c9de560dSAlex Tomas 1982c9de560dSAlex Tomas /* 1983c9de560dSAlex Tomas * ac->ac2_order is set only if the fe_len is a power of 2 1984c9de560dSAlex Tomas * if ac2_order is set we also set criteria to 0 so that we 1985c9de560dSAlex Tomas * try exact allocation using buddy. 1986c9de560dSAlex Tomas */ 1987c9de560dSAlex Tomas i = fls(ac->ac_g_ex.fe_len); 1988c9de560dSAlex Tomas ac->ac_2order = 0; 1989c9de560dSAlex Tomas /* 1990c9de560dSAlex Tomas * We search using buddy data only if the order of the request 1991c9de560dSAlex Tomas * is greater than equal to the sbi_s_mb_order2_reqs 1992b713a5ecSTheodore Ts'o * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 1993c9de560dSAlex Tomas */ 1994c9de560dSAlex Tomas if (i >= sbi->s_mb_order2_reqs) { 1995c9de560dSAlex Tomas /* 1996c9de560dSAlex Tomas * This should tell if fe_len is exactly power of 2 1997c9de560dSAlex Tomas */ 1998c9de560dSAlex Tomas if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) 1999c9de560dSAlex Tomas ac->ac_2order = i - 1; 2000c9de560dSAlex Tomas } 2001c9de560dSAlex Tomas 2002c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 2003c9de560dSAlex Tomas 20044ba74d00STheodore Ts'o /* if stream allocation is enabled, use global goal */ 20054ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2006c9de560dSAlex Tomas /* TBD: may be hot point */ 2007c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 2008c9de560dSAlex Tomas ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 2009c9de560dSAlex Tomas ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 2010c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 2011c9de560dSAlex Tomas } 20124ba74d00STheodore Ts'o 2013c9de560dSAlex Tomas /* Let's just scan groups to find more-less suitable blocks */ 2014c9de560dSAlex Tomas cr = ac->ac_2order ? 0 : 1; 2015c9de560dSAlex Tomas /* 2016c9de560dSAlex Tomas * cr == 0 try to get exact allocation, 2017c9de560dSAlex Tomas * cr == 3 try to get anything 2018c9de560dSAlex Tomas */ 2019c9de560dSAlex Tomas repeat: 2020c9de560dSAlex Tomas for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 2021c9de560dSAlex Tomas ac->ac_criteria = cr; 2022ed8f9c75SAneesh Kumar K.V /* 2023ed8f9c75SAneesh Kumar K.V * searching for the right group start 2024ed8f9c75SAneesh Kumar K.V * from the goal value specified 2025ed8f9c75SAneesh Kumar K.V */ 2026ed8f9c75SAneesh Kumar K.V group = ac->ac_g_ex.fe_group; 2027ed8f9c75SAneesh Kumar K.V 20288df9675fSTheodore Ts'o for (i = 0; i < ngroups; group++, i++) { 2029c9de560dSAlex Tomas struct ext4_group_info *grp; 2030c9de560dSAlex Tomas struct ext4_group_desc *desc; 2031c9de560dSAlex Tomas 20328df9675fSTheodore Ts'o if (group == ngroups) 2033c9de560dSAlex Tomas group = 0; 2034c9de560dSAlex Tomas 2035c9de560dSAlex Tomas /* quick check to skip empty groups */ 2036920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 2037c9de560dSAlex Tomas if (grp->bb_free == 0) 2038c9de560dSAlex Tomas continue; 2039c9de560dSAlex Tomas 2040c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2041c9de560dSAlex Tomas if (err) 2042c9de560dSAlex Tomas goto out; 2043c9de560dSAlex Tomas 2044c9de560dSAlex Tomas ext4_lock_group(sb, group); 2045c9de560dSAlex Tomas if (!ext4_mb_good_group(ac, group, cr)) { 2046c9de560dSAlex Tomas /* someone did allocation from this group */ 2047c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2048c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 2049c9de560dSAlex Tomas continue; 2050c9de560dSAlex Tomas } 2051c9de560dSAlex Tomas 2052c9de560dSAlex Tomas ac->ac_groups_scanned++; 2053c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, group, NULL); 205475507efbSTheodore Ts'o if (cr == 0) 2055c9de560dSAlex Tomas ext4_mb_simple_scan_group(ac, &e4b); 2056c9de560dSAlex Tomas else if (cr == 1 && 2057c9de560dSAlex Tomas ac->ac_g_ex.fe_len == sbi->s_stripe) 2058c9de560dSAlex Tomas ext4_mb_scan_aligned(ac, &e4b); 2059c9de560dSAlex Tomas else 2060c9de560dSAlex Tomas ext4_mb_complex_scan_group(ac, &e4b); 2061c9de560dSAlex Tomas 2062c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2063c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 2064c9de560dSAlex Tomas 2065c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_CONTINUE) 2066c9de560dSAlex Tomas break; 2067c9de560dSAlex Tomas } 2068c9de560dSAlex Tomas } 2069c9de560dSAlex Tomas 2070c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2071c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2072c9de560dSAlex Tomas /* 2073c9de560dSAlex Tomas * We've been searching too long. Let's try to allocate 2074c9de560dSAlex Tomas * the best chunk we've found so far 2075c9de560dSAlex Tomas */ 2076c9de560dSAlex Tomas 2077c9de560dSAlex Tomas ext4_mb_try_best_found(ac, &e4b); 2078c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_FOUND) { 2079c9de560dSAlex Tomas /* 2080c9de560dSAlex Tomas * Someone more lucky has already allocated it. 2081c9de560dSAlex Tomas * The only thing we can do is just take first 2082c9de560dSAlex Tomas * found block(s) 2083c9de560dSAlex Tomas printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); 2084c9de560dSAlex Tomas */ 2085c9de560dSAlex Tomas ac->ac_b_ex.fe_group = 0; 2086c9de560dSAlex Tomas ac->ac_b_ex.fe_start = 0; 2087c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 0; 2088c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 2089c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_FIRST; 2090c9de560dSAlex Tomas cr = 3; 2091c9de560dSAlex Tomas atomic_inc(&sbi->s_mb_lost_chunks); 2092c9de560dSAlex Tomas goto repeat; 2093c9de560dSAlex Tomas } 2094c9de560dSAlex Tomas } 2095c9de560dSAlex Tomas out: 2096c9de560dSAlex Tomas return err; 2097c9de560dSAlex Tomas } 2098c9de560dSAlex Tomas 2099c9de560dSAlex Tomas static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2100c9de560dSAlex Tomas { 2101c9de560dSAlex Tomas struct super_block *sb = seq->private; 2102c9de560dSAlex Tomas ext4_group_t group; 2103c9de560dSAlex Tomas 21048df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2105c9de560dSAlex Tomas return NULL; 2106c9de560dSAlex Tomas group = *pos + 1; 2107a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2108c9de560dSAlex Tomas } 2109c9de560dSAlex Tomas 2110c9de560dSAlex Tomas static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2111c9de560dSAlex Tomas { 2112c9de560dSAlex Tomas struct super_block *sb = seq->private; 2113c9de560dSAlex Tomas ext4_group_t group; 2114c9de560dSAlex Tomas 2115c9de560dSAlex Tomas ++*pos; 21168df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2117c9de560dSAlex Tomas return NULL; 2118c9de560dSAlex Tomas group = *pos + 1; 2119a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2120c9de560dSAlex Tomas } 2121c9de560dSAlex Tomas 2122c9de560dSAlex Tomas static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2123c9de560dSAlex Tomas { 2124c9de560dSAlex Tomas struct super_block *sb = seq->private; 2125a9df9a49STheodore Ts'o ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2126c9de560dSAlex Tomas int i; 2127c9de560dSAlex Tomas int err; 2128c9de560dSAlex Tomas struct ext4_buddy e4b; 2129c9de560dSAlex Tomas struct sg { 2130c9de560dSAlex Tomas struct ext4_group_info info; 2131a36b4498SEric Sandeen ext4_grpblk_t counters[16]; 2132c9de560dSAlex Tomas } sg; 2133c9de560dSAlex Tomas 2134c9de560dSAlex Tomas group--; 2135c9de560dSAlex Tomas if (group == 0) 2136c9de560dSAlex Tomas seq_printf(seq, "#%-5s: %-5s %-5s %-5s " 2137c9de560dSAlex Tomas "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " 2138c9de560dSAlex Tomas "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", 2139c9de560dSAlex Tomas "group", "free", "frags", "first", 2140c9de560dSAlex Tomas "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", 2141c9de560dSAlex Tomas "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); 2142c9de560dSAlex Tomas 2143c9de560dSAlex Tomas i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 2144c9de560dSAlex Tomas sizeof(struct ext4_group_info); 2145c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2146c9de560dSAlex Tomas if (err) { 2147a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: I/O error\n", group); 2148c9de560dSAlex Tomas return 0; 2149c9de560dSAlex Tomas } 2150c9de560dSAlex Tomas ext4_lock_group(sb, group); 2151c9de560dSAlex Tomas memcpy(&sg, ext4_get_group_info(sb, group), i); 2152c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2153c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 2154c9de560dSAlex Tomas 2155a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 2156c9de560dSAlex Tomas sg.info.bb_fragments, sg.info.bb_first_free); 2157c9de560dSAlex Tomas for (i = 0; i <= 13; i++) 2158c9de560dSAlex Tomas seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? 2159c9de560dSAlex Tomas sg.info.bb_counters[i] : 0); 2160c9de560dSAlex Tomas seq_printf(seq, " ]\n"); 2161c9de560dSAlex Tomas 2162c9de560dSAlex Tomas return 0; 2163c9de560dSAlex Tomas } 2164c9de560dSAlex Tomas 2165c9de560dSAlex Tomas static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 2166c9de560dSAlex Tomas { 2167c9de560dSAlex Tomas } 2168c9de560dSAlex Tomas 21697f1346a9STobias Klauser static const struct seq_operations ext4_mb_seq_groups_ops = { 2170c9de560dSAlex Tomas .start = ext4_mb_seq_groups_start, 2171c9de560dSAlex Tomas .next = ext4_mb_seq_groups_next, 2172c9de560dSAlex Tomas .stop = ext4_mb_seq_groups_stop, 2173c9de560dSAlex Tomas .show = ext4_mb_seq_groups_show, 2174c9de560dSAlex Tomas }; 2175c9de560dSAlex Tomas 2176c9de560dSAlex Tomas static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) 2177c9de560dSAlex Tomas { 2178c9de560dSAlex Tomas struct super_block *sb = PDE(inode)->data; 2179c9de560dSAlex Tomas int rc; 2180c9de560dSAlex Tomas 2181c9de560dSAlex Tomas rc = seq_open(file, &ext4_mb_seq_groups_ops); 2182c9de560dSAlex Tomas if (rc == 0) { 2183c9de560dSAlex Tomas struct seq_file *m = (struct seq_file *)file->private_data; 2184c9de560dSAlex Tomas m->private = sb; 2185c9de560dSAlex Tomas } 2186c9de560dSAlex Tomas return rc; 2187c9de560dSAlex Tomas 2188c9de560dSAlex Tomas } 2189c9de560dSAlex Tomas 21907f1346a9STobias Klauser static const struct file_operations ext4_mb_seq_groups_fops = { 2191c9de560dSAlex Tomas .owner = THIS_MODULE, 2192c9de560dSAlex Tomas .open = ext4_mb_seq_groups_open, 2193c9de560dSAlex Tomas .read = seq_read, 2194c9de560dSAlex Tomas .llseek = seq_lseek, 2195c9de560dSAlex Tomas .release = seq_release, 2196c9de560dSAlex Tomas }; 2197c9de560dSAlex Tomas 21985f21b0e6SFrederic Bohe 21995f21b0e6SFrederic Bohe /* Create and initialize ext4_group_info data for the given group. */ 2200920313a7SAneesh Kumar K.V int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 22015f21b0e6SFrederic Bohe struct ext4_group_desc *desc) 22025f21b0e6SFrederic Bohe { 22035f21b0e6SFrederic Bohe int i, len; 22045f21b0e6SFrederic Bohe int metalen = 0; 22055f21b0e6SFrederic Bohe struct ext4_sb_info *sbi = EXT4_SB(sb); 22065f21b0e6SFrederic Bohe struct ext4_group_info **meta_group_info; 22075f21b0e6SFrederic Bohe 22085f21b0e6SFrederic Bohe /* 22095f21b0e6SFrederic Bohe * First check if this group is the first of a reserved block. 22105f21b0e6SFrederic Bohe * If it's true, we have to allocate a new table of pointers 22115f21b0e6SFrederic Bohe * to ext4_group_info structures 22125f21b0e6SFrederic Bohe */ 22135f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 22145f21b0e6SFrederic Bohe metalen = sizeof(*meta_group_info) << 22155f21b0e6SFrederic Bohe EXT4_DESC_PER_BLOCK_BITS(sb); 22165f21b0e6SFrederic Bohe meta_group_info = kmalloc(metalen, GFP_KERNEL); 22175f21b0e6SFrederic Bohe if (meta_group_info == NULL) { 22185f21b0e6SFrederic Bohe printk(KERN_ERR "EXT4-fs: can't allocate mem for a " 22195f21b0e6SFrederic Bohe "buddy group\n"); 22205f21b0e6SFrederic Bohe goto exit_meta_group_info; 22215f21b0e6SFrederic Bohe } 22225f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = 22235f21b0e6SFrederic Bohe meta_group_info; 22245f21b0e6SFrederic Bohe } 22255f21b0e6SFrederic Bohe 22265f21b0e6SFrederic Bohe /* 22275f21b0e6SFrederic Bohe * calculate needed size. if change bb_counters size, 22285f21b0e6SFrederic Bohe * don't forget about ext4_mb_generate_buddy() 22295f21b0e6SFrederic Bohe */ 22305f21b0e6SFrederic Bohe len = offsetof(typeof(**meta_group_info), 22315f21b0e6SFrederic Bohe bb_counters[sb->s_blocksize_bits + 2]); 22325f21b0e6SFrederic Bohe 22335f21b0e6SFrederic Bohe meta_group_info = 22345f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; 22355f21b0e6SFrederic Bohe i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 22365f21b0e6SFrederic Bohe 22375f21b0e6SFrederic Bohe meta_group_info[i] = kzalloc(len, GFP_KERNEL); 22385f21b0e6SFrederic Bohe if (meta_group_info[i] == NULL) { 22395f21b0e6SFrederic Bohe printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n"); 22405f21b0e6SFrederic Bohe goto exit_group_info; 22415f21b0e6SFrederic Bohe } 22425f21b0e6SFrederic Bohe set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 22435f21b0e6SFrederic Bohe &(meta_group_info[i]->bb_state)); 22445f21b0e6SFrederic Bohe 22455f21b0e6SFrederic Bohe /* 22465f21b0e6SFrederic Bohe * initialize bb_free to be able to skip 22475f21b0e6SFrederic Bohe * empty groups without initialization 22485f21b0e6SFrederic Bohe */ 22495f21b0e6SFrederic Bohe if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 22505f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 22515f21b0e6SFrederic Bohe ext4_free_blocks_after_init(sb, group, desc); 22525f21b0e6SFrederic Bohe } else { 22535f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 2254560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc); 22555f21b0e6SFrederic Bohe } 22565f21b0e6SFrederic Bohe 22575f21b0e6SFrederic Bohe INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 2258920313a7SAneesh Kumar K.V init_rwsem(&meta_group_info[i]->alloc_sem); 22595a4a7989SJoe Perches meta_group_info[i]->bb_free_root.rb_node = NULL; 22605f21b0e6SFrederic Bohe 22615f21b0e6SFrederic Bohe #ifdef DOUBLE_CHECK 22625f21b0e6SFrederic Bohe { 22635f21b0e6SFrederic Bohe struct buffer_head *bh; 22645f21b0e6SFrederic Bohe meta_group_info[i]->bb_bitmap = 22655f21b0e6SFrederic Bohe kmalloc(sb->s_blocksize, GFP_KERNEL); 22665f21b0e6SFrederic Bohe BUG_ON(meta_group_info[i]->bb_bitmap == NULL); 22675f21b0e6SFrederic Bohe bh = ext4_read_block_bitmap(sb, group); 22685f21b0e6SFrederic Bohe BUG_ON(bh == NULL); 22695f21b0e6SFrederic Bohe memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, 22705f21b0e6SFrederic Bohe sb->s_blocksize); 22715f21b0e6SFrederic Bohe put_bh(bh); 22725f21b0e6SFrederic Bohe } 22735f21b0e6SFrederic Bohe #endif 22745f21b0e6SFrederic Bohe 22755f21b0e6SFrederic Bohe return 0; 22765f21b0e6SFrederic Bohe 22775f21b0e6SFrederic Bohe exit_group_info: 22785f21b0e6SFrederic Bohe /* If a meta_group_info table has been allocated, release it now */ 22795f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) 22805f21b0e6SFrederic Bohe kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); 22815f21b0e6SFrederic Bohe exit_meta_group_info: 22825f21b0e6SFrederic Bohe return -ENOMEM; 22835f21b0e6SFrederic Bohe } /* ext4_mb_add_groupinfo */ 22845f21b0e6SFrederic Bohe 2285c9de560dSAlex Tomas static int ext4_mb_init_backend(struct super_block *sb) 2286c9de560dSAlex Tomas { 22878df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 2288c9de560dSAlex Tomas ext4_group_t i; 2289c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 22905f21b0e6SFrederic Bohe struct ext4_super_block *es = sbi->s_es; 22915f21b0e6SFrederic Bohe int num_meta_group_infos; 22925f21b0e6SFrederic Bohe int num_meta_group_infos_max; 22935f21b0e6SFrederic Bohe int array_size; 22945f21b0e6SFrederic Bohe struct ext4_group_desc *desc; 2295c9de560dSAlex Tomas 22965f21b0e6SFrederic Bohe /* This is the number of blocks used by GDT */ 22978df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 22985f21b0e6SFrederic Bohe 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); 22995f21b0e6SFrederic Bohe 23005f21b0e6SFrederic Bohe /* 23015f21b0e6SFrederic Bohe * This is the total number of blocks used by GDT including 23025f21b0e6SFrederic Bohe * the number of reserved blocks for GDT. 23035f21b0e6SFrederic Bohe * The s_group_info array is allocated with this value 23045f21b0e6SFrederic Bohe * to allow a clean online resize without a complex 23055f21b0e6SFrederic Bohe * manipulation of pointer. 23065f21b0e6SFrederic Bohe * The drawback is the unused memory when no resize 23075f21b0e6SFrederic Bohe * occurs but it's very low in terms of pages 23085f21b0e6SFrederic Bohe * (see comments below) 23095f21b0e6SFrederic Bohe * Need to handle this properly when META_BG resizing is allowed 23105f21b0e6SFrederic Bohe */ 23115f21b0e6SFrederic Bohe num_meta_group_infos_max = num_meta_group_infos + 23125f21b0e6SFrederic Bohe le16_to_cpu(es->s_reserved_gdt_blocks); 23135f21b0e6SFrederic Bohe 23145f21b0e6SFrederic Bohe /* 23155f21b0e6SFrederic Bohe * array_size is the size of s_group_info array. We round it 23165f21b0e6SFrederic Bohe * to the next power of two because this approximation is done 23175f21b0e6SFrederic Bohe * internally by kmalloc so we can have some more memory 23185f21b0e6SFrederic Bohe * for free here (e.g. may be used for META_BG resize). 23195f21b0e6SFrederic Bohe */ 23205f21b0e6SFrederic Bohe array_size = 1; 23215f21b0e6SFrederic Bohe while (array_size < sizeof(*sbi->s_group_info) * 23225f21b0e6SFrederic Bohe num_meta_group_infos_max) 23235f21b0e6SFrederic Bohe array_size = array_size << 1; 2324c9de560dSAlex Tomas /* An 8TB filesystem with 64-bit pointers requires a 4096 byte 2325c9de560dSAlex Tomas * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. 2326c9de560dSAlex Tomas * So a two level scheme suffices for now. */ 23275f21b0e6SFrederic Bohe sbi->s_group_info = kmalloc(array_size, GFP_KERNEL); 2328c9de560dSAlex Tomas if (sbi->s_group_info == NULL) { 2329c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n"); 2330c9de560dSAlex Tomas return -ENOMEM; 2331c9de560dSAlex Tomas } 2332c9de560dSAlex Tomas sbi->s_buddy_cache = new_inode(sb); 2333c9de560dSAlex Tomas if (sbi->s_buddy_cache == NULL) { 2334c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: can't get new inode\n"); 2335c9de560dSAlex Tomas goto err_freesgi; 2336c9de560dSAlex Tomas } 2337c9de560dSAlex Tomas EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 23388df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 2339c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, i, NULL); 2340c9de560dSAlex Tomas if (desc == NULL) { 2341c9de560dSAlex Tomas printk(KERN_ERR 2342a9df9a49STheodore Ts'o "EXT4-fs: can't read descriptor %u\n", i); 2343c9de560dSAlex Tomas goto err_freebuddy; 2344c9de560dSAlex Tomas } 23455f21b0e6SFrederic Bohe if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 23465f21b0e6SFrederic Bohe goto err_freebuddy; 2347c9de560dSAlex Tomas } 2348c9de560dSAlex Tomas 2349c9de560dSAlex Tomas return 0; 2350c9de560dSAlex Tomas 2351c9de560dSAlex Tomas err_freebuddy: 2352f1fa3342SRoel Kluin while (i-- > 0) 2353c9de560dSAlex Tomas kfree(ext4_get_group_info(sb, i)); 2354c9de560dSAlex Tomas i = num_meta_group_infos; 2355f1fa3342SRoel Kluin while (i-- > 0) 2356c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2357c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2358c9de560dSAlex Tomas err_freesgi: 2359c9de560dSAlex Tomas kfree(sbi->s_group_info); 2360c9de560dSAlex Tomas return -ENOMEM; 2361c9de560dSAlex Tomas } 2362c9de560dSAlex Tomas 2363c9de560dSAlex Tomas int ext4_mb_init(struct super_block *sb, int needs_recovery) 2364c9de560dSAlex Tomas { 2365c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 23666be2ded1SAneesh Kumar K.V unsigned i, j; 2367c9de560dSAlex Tomas unsigned offset; 2368c9de560dSAlex Tomas unsigned max; 236974767c5aSShen Feng int ret; 2370c9de560dSAlex Tomas 23711927805eSEric Sandeen i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); 2372c9de560dSAlex Tomas 2373c9de560dSAlex Tomas sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 2374c9de560dSAlex Tomas if (sbi->s_mb_offsets == NULL) { 2375c9de560dSAlex Tomas return -ENOMEM; 2376c9de560dSAlex Tomas } 2377ff7ef329SYasunori Goto 23781927805eSEric Sandeen i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); 2379c9de560dSAlex Tomas sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 2380c9de560dSAlex Tomas if (sbi->s_mb_maxs == NULL) { 2381a7b19448SDan Carpenter kfree(sbi->s_mb_offsets); 2382c9de560dSAlex Tomas return -ENOMEM; 2383c9de560dSAlex Tomas } 2384c9de560dSAlex Tomas 2385c9de560dSAlex Tomas /* order 0 is regular bitmap */ 2386c9de560dSAlex Tomas sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 2387c9de560dSAlex Tomas sbi->s_mb_offsets[0] = 0; 2388c9de560dSAlex Tomas 2389c9de560dSAlex Tomas i = 1; 2390c9de560dSAlex Tomas offset = 0; 2391c9de560dSAlex Tomas max = sb->s_blocksize << 2; 2392c9de560dSAlex Tomas do { 2393c9de560dSAlex Tomas sbi->s_mb_offsets[i] = offset; 2394c9de560dSAlex Tomas sbi->s_mb_maxs[i] = max; 2395c9de560dSAlex Tomas offset += 1 << (sb->s_blocksize_bits - i); 2396c9de560dSAlex Tomas max = max >> 1; 2397c9de560dSAlex Tomas i++; 2398c9de560dSAlex Tomas } while (i <= sb->s_blocksize_bits + 1); 2399c9de560dSAlex Tomas 2400c9de560dSAlex Tomas /* init file for buddy data */ 240174767c5aSShen Feng ret = ext4_mb_init_backend(sb); 240274767c5aSShen Feng if (ret != 0) { 2403c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2404c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 240574767c5aSShen Feng return ret; 2406c9de560dSAlex Tomas } 2407c9de560dSAlex Tomas 2408c9de560dSAlex Tomas spin_lock_init(&sbi->s_md_lock); 2409c9de560dSAlex Tomas spin_lock_init(&sbi->s_bal_lock); 2410c9de560dSAlex Tomas 2411c9de560dSAlex Tomas sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 2412c9de560dSAlex Tomas sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 2413c9de560dSAlex Tomas sbi->s_mb_stats = MB_DEFAULT_STATS; 2414c9de560dSAlex Tomas sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 2415c9de560dSAlex Tomas sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 2416c9de560dSAlex Tomas sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; 2417c9de560dSAlex Tomas 2418730c213cSEric Sandeen sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 2419c9de560dSAlex Tomas if (sbi->s_locality_groups == NULL) { 2420c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2421c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 2422c9de560dSAlex Tomas return -ENOMEM; 2423c9de560dSAlex Tomas } 2424730c213cSEric Sandeen for_each_possible_cpu(i) { 2425c9de560dSAlex Tomas struct ext4_locality_group *lg; 2426730c213cSEric Sandeen lg = per_cpu_ptr(sbi->s_locality_groups, i); 2427c9de560dSAlex Tomas mutex_init(&lg->lg_mutex); 24286be2ded1SAneesh Kumar K.V for (j = 0; j < PREALLOC_TB_SIZE; j++) 24296be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 2430c9de560dSAlex Tomas spin_lock_init(&lg->lg_prealloc_lock); 2431c9de560dSAlex Tomas } 2432c9de560dSAlex Tomas 2433*296c355cSTheodore Ts'o if (sbi->s_proc) 2434*296c355cSTheodore Ts'o proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, 2435*296c355cSTheodore Ts'o &ext4_mb_seq_groups_fops, sb); 2436c9de560dSAlex Tomas 24370390131bSFrank Mayhar if (sbi->s_journal) 24383e624fc7STheodore Ts'o sbi->s_journal->j_commit_callback = release_blocks_on_commit; 2439c9de560dSAlex Tomas return 0; 2440c9de560dSAlex Tomas } 2441c9de560dSAlex Tomas 2442955ce5f5SAneesh Kumar K.V /* need to called with the ext4 group lock held */ 2443c9de560dSAlex Tomas static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) 2444c9de560dSAlex Tomas { 2445c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 2446c9de560dSAlex Tomas struct list_head *cur, *tmp; 2447c9de560dSAlex Tomas int count = 0; 2448c9de560dSAlex Tomas 2449c9de560dSAlex Tomas list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 2450c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 2451c9de560dSAlex Tomas list_del(&pa->pa_group_list); 2452c9de560dSAlex Tomas count++; 2453688f05a0SAneesh Kumar K.V kmem_cache_free(ext4_pspace_cachep, pa); 2454c9de560dSAlex Tomas } 2455c9de560dSAlex Tomas if (count) 24566ba495e9STheodore Ts'o mb_debug(1, "mballoc: %u PAs left\n", count); 2457c9de560dSAlex Tomas 2458c9de560dSAlex Tomas } 2459c9de560dSAlex Tomas 2460c9de560dSAlex Tomas int ext4_mb_release(struct super_block *sb) 2461c9de560dSAlex Tomas { 24628df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 2463c9de560dSAlex Tomas ext4_group_t i; 2464c9de560dSAlex Tomas int num_meta_group_infos; 2465c9de560dSAlex Tomas struct ext4_group_info *grinfo; 2466c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2467c9de560dSAlex Tomas 2468c9de560dSAlex Tomas if (sbi->s_group_info) { 24698df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 2470c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, i); 2471c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 2472c9de560dSAlex Tomas kfree(grinfo->bb_bitmap); 2473c9de560dSAlex Tomas #endif 2474c9de560dSAlex Tomas ext4_lock_group(sb, i); 2475c9de560dSAlex Tomas ext4_mb_cleanup_pa(grinfo); 2476c9de560dSAlex Tomas ext4_unlock_group(sb, i); 2477c9de560dSAlex Tomas kfree(grinfo); 2478c9de560dSAlex Tomas } 24798df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + 2480c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK(sb) - 1) >> 2481c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK_BITS(sb); 2482c9de560dSAlex Tomas for (i = 0; i < num_meta_group_infos; i++) 2483c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2484c9de560dSAlex Tomas kfree(sbi->s_group_info); 2485c9de560dSAlex Tomas } 2486c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2487c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 2488c9de560dSAlex Tomas if (sbi->s_buddy_cache) 2489c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2490c9de560dSAlex Tomas if (sbi->s_mb_stats) { 2491c9de560dSAlex Tomas printk(KERN_INFO 2492c9de560dSAlex Tomas "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n", 2493c9de560dSAlex Tomas atomic_read(&sbi->s_bal_allocated), 2494c9de560dSAlex Tomas atomic_read(&sbi->s_bal_reqs), 2495c9de560dSAlex Tomas atomic_read(&sbi->s_bal_success)); 2496c9de560dSAlex Tomas printk(KERN_INFO 2497c9de560dSAlex Tomas "EXT4-fs: mballoc: %u extents scanned, %u goal hits, " 2498c9de560dSAlex Tomas "%u 2^N hits, %u breaks, %u lost\n", 2499c9de560dSAlex Tomas atomic_read(&sbi->s_bal_ex_scanned), 2500c9de560dSAlex Tomas atomic_read(&sbi->s_bal_goals), 2501c9de560dSAlex Tomas atomic_read(&sbi->s_bal_2orders), 2502c9de560dSAlex Tomas atomic_read(&sbi->s_bal_breaks), 2503c9de560dSAlex Tomas atomic_read(&sbi->s_mb_lost_chunks)); 2504c9de560dSAlex Tomas printk(KERN_INFO 2505c9de560dSAlex Tomas "EXT4-fs: mballoc: %lu generated and it took %Lu\n", 2506c9de560dSAlex Tomas sbi->s_mb_buddies_generated++, 2507c9de560dSAlex Tomas sbi->s_mb_generation_time); 2508c9de560dSAlex Tomas printk(KERN_INFO 2509c9de560dSAlex Tomas "EXT4-fs: mballoc: %u preallocated, %u discarded\n", 2510c9de560dSAlex Tomas atomic_read(&sbi->s_mb_preallocated), 2511c9de560dSAlex Tomas atomic_read(&sbi->s_mb_discarded)); 2512c9de560dSAlex Tomas } 2513c9de560dSAlex Tomas 2514730c213cSEric Sandeen free_percpu(sbi->s_locality_groups); 2515*296c355cSTheodore Ts'o if (sbi->s_proc) 2516*296c355cSTheodore Ts'o remove_proc_entry("mb_groups", sbi->s_proc); 2517c9de560dSAlex Tomas 2518c9de560dSAlex Tomas return 0; 2519c9de560dSAlex Tomas } 2520c9de560dSAlex Tomas 25213e624fc7STheodore Ts'o /* 25223e624fc7STheodore Ts'o * This function is called by the jbd2 layer once the commit has finished, 25233e624fc7STheodore Ts'o * so we know we can free the blocks that were released with that commit. 25243e624fc7STheodore Ts'o */ 25253e624fc7STheodore Ts'o static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) 2526c9de560dSAlex Tomas { 25273e624fc7STheodore Ts'o struct super_block *sb = journal->j_private; 2528c9de560dSAlex Tomas struct ext4_buddy e4b; 2529c894058dSAneesh Kumar K.V struct ext4_group_info *db; 2530c894058dSAneesh Kumar K.V int err, count = 0, count2 = 0; 2531c894058dSAneesh Kumar K.V struct ext4_free_data *entry; 25328a0aba73STheodore Ts'o ext4_fsblk_t discard_block; 25333e624fc7STheodore Ts'o struct list_head *l, *ltmp; 2534c9de560dSAlex Tomas 25353e624fc7STheodore Ts'o list_for_each_safe(l, ltmp, &txn->t_private_list) { 25363e624fc7STheodore Ts'o entry = list_entry(l, struct ext4_free_data, list); 2537c9de560dSAlex Tomas 25386ba495e9STheodore Ts'o mb_debug(1, "gonna free %u blocks in group %u (0x%p):", 2539c894058dSAneesh Kumar K.V entry->count, entry->group, entry); 2540c9de560dSAlex Tomas 2541c894058dSAneesh Kumar K.V err = ext4_mb_load_buddy(sb, entry->group, &e4b); 2542c9de560dSAlex Tomas /* we expect to find existing buddy because it's pinned */ 2543c9de560dSAlex Tomas BUG_ON(err != 0); 2544c9de560dSAlex Tomas 2545c894058dSAneesh Kumar K.V db = e4b.bd_info; 2546c9de560dSAlex Tomas /* there are blocks to put in buddy to make them really free */ 2547c894058dSAneesh Kumar K.V count += entry->count; 2548c9de560dSAlex Tomas count2++; 2549c894058dSAneesh Kumar K.V ext4_lock_group(sb, entry->group); 2550c894058dSAneesh Kumar K.V /* Take it out of per group rb tree */ 2551c894058dSAneesh Kumar K.V rb_erase(&entry->node, &(db->bb_free_root)); 2552c894058dSAneesh Kumar K.V mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); 2553c9de560dSAlex Tomas 2554c894058dSAneesh Kumar K.V if (!db->bb_free_root.rb_node) { 2555c894058dSAneesh Kumar K.V /* No more items in the per group rb tree 2556c894058dSAneesh Kumar K.V * balance refcounts from ext4_mb_free_metadata() 2557c894058dSAneesh Kumar K.V */ 2558c9de560dSAlex Tomas page_cache_release(e4b.bd_buddy_page); 2559c9de560dSAlex Tomas page_cache_release(e4b.bd_bitmap_page); 2560c894058dSAneesh Kumar K.V } 2561c894058dSAneesh Kumar K.V ext4_unlock_group(sb, entry->group); 25628a0aba73STheodore Ts'o discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb) 25638a0aba73STheodore Ts'o + entry->start_blk 25648a0aba73STheodore Ts'o + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 25659bffad1eSTheodore Ts'o trace_ext4_discard_blocks(sb, (unsigned long long)discard_block, 2566ba80b101STheodore Ts'o entry->count); 25678a0aba73STheodore Ts'o sb_issue_discard(sb, discard_block, entry->count); 2568c9de560dSAlex Tomas 2569c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 2570c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 25713e624fc7STheodore Ts'o } 2572c9de560dSAlex Tomas 25736ba495e9STheodore Ts'o mb_debug(1, "freed %u blocks in %u structures\n", count, count2); 2574c9de560dSAlex Tomas } 2575c9de560dSAlex Tomas 25766ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 25776ba495e9STheodore Ts'o u8 mb_enable_debug __read_mostly; 25786ba495e9STheodore Ts'o 25796ba495e9STheodore Ts'o static struct dentry *debugfs_dir; 25806ba495e9STheodore Ts'o static struct dentry *debugfs_debug; 25816ba495e9STheodore Ts'o 25826ba495e9STheodore Ts'o static void __init ext4_create_debugfs_entry(void) 25836ba495e9STheodore Ts'o { 25846ba495e9STheodore Ts'o debugfs_dir = debugfs_create_dir("ext4", NULL); 25856ba495e9STheodore Ts'o if (debugfs_dir) 25866ba495e9STheodore Ts'o debugfs_debug = debugfs_create_u8("mballoc-debug", 25876ba495e9STheodore Ts'o S_IRUGO | S_IWUSR, 25886ba495e9STheodore Ts'o debugfs_dir, 25896ba495e9STheodore Ts'o &mb_enable_debug); 25906ba495e9STheodore Ts'o } 25916ba495e9STheodore Ts'o 25926ba495e9STheodore Ts'o static void ext4_remove_debugfs_entry(void) 25936ba495e9STheodore Ts'o { 25946ba495e9STheodore Ts'o debugfs_remove(debugfs_debug); 25956ba495e9STheodore Ts'o debugfs_remove(debugfs_dir); 25966ba495e9STheodore Ts'o } 25976ba495e9STheodore Ts'o 25986ba495e9STheodore Ts'o #else 25996ba495e9STheodore Ts'o 26006ba495e9STheodore Ts'o static void __init ext4_create_debugfs_entry(void) 26016ba495e9STheodore Ts'o { 26026ba495e9STheodore Ts'o } 26036ba495e9STheodore Ts'o 26046ba495e9STheodore Ts'o static void ext4_remove_debugfs_entry(void) 26056ba495e9STheodore Ts'o { 26066ba495e9STheodore Ts'o } 26076ba495e9STheodore Ts'o 26086ba495e9STheodore Ts'o #endif 26096ba495e9STheodore Ts'o 2610c9de560dSAlex Tomas int __init init_ext4_mballoc(void) 2611c9de560dSAlex Tomas { 2612c9de560dSAlex Tomas ext4_pspace_cachep = 2613c9de560dSAlex Tomas kmem_cache_create("ext4_prealloc_space", 2614c9de560dSAlex Tomas sizeof(struct ext4_prealloc_space), 2615c9de560dSAlex Tomas 0, SLAB_RECLAIM_ACCOUNT, NULL); 2616c9de560dSAlex Tomas if (ext4_pspace_cachep == NULL) 2617c9de560dSAlex Tomas return -ENOMEM; 2618c9de560dSAlex Tomas 2619256bdb49SEric Sandeen ext4_ac_cachep = 2620256bdb49SEric Sandeen kmem_cache_create("ext4_alloc_context", 2621256bdb49SEric Sandeen sizeof(struct ext4_allocation_context), 2622256bdb49SEric Sandeen 0, SLAB_RECLAIM_ACCOUNT, NULL); 2623256bdb49SEric Sandeen if (ext4_ac_cachep == NULL) { 2624256bdb49SEric Sandeen kmem_cache_destroy(ext4_pspace_cachep); 2625256bdb49SEric Sandeen return -ENOMEM; 2626256bdb49SEric Sandeen } 2627c894058dSAneesh Kumar K.V 2628c894058dSAneesh Kumar K.V ext4_free_ext_cachep = 2629c894058dSAneesh Kumar K.V kmem_cache_create("ext4_free_block_extents", 2630c894058dSAneesh Kumar K.V sizeof(struct ext4_free_data), 2631c894058dSAneesh Kumar K.V 0, SLAB_RECLAIM_ACCOUNT, NULL); 2632c894058dSAneesh Kumar K.V if (ext4_free_ext_cachep == NULL) { 2633c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_pspace_cachep); 2634c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_ac_cachep); 2635c894058dSAneesh Kumar K.V return -ENOMEM; 2636c894058dSAneesh Kumar K.V } 26376ba495e9STheodore Ts'o ext4_create_debugfs_entry(); 2638c9de560dSAlex Tomas return 0; 2639c9de560dSAlex Tomas } 2640c9de560dSAlex Tomas 2641c9de560dSAlex Tomas void exit_ext4_mballoc(void) 2642c9de560dSAlex Tomas { 26433e03f9caSJesper Dangaard Brouer /* 26443e03f9caSJesper Dangaard Brouer * Wait for completion of call_rcu()'s on ext4_pspace_cachep 26453e03f9caSJesper Dangaard Brouer * before destroying the slab cache. 26463e03f9caSJesper Dangaard Brouer */ 26473e03f9caSJesper Dangaard Brouer rcu_barrier(); 2648c9de560dSAlex Tomas kmem_cache_destroy(ext4_pspace_cachep); 2649256bdb49SEric Sandeen kmem_cache_destroy(ext4_ac_cachep); 2650c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_free_ext_cachep); 26516ba495e9STheodore Ts'o ext4_remove_debugfs_entry(); 2652c9de560dSAlex Tomas } 2653c9de560dSAlex Tomas 2654c9de560dSAlex Tomas 2655c9de560dSAlex Tomas /* 2656c9de560dSAlex Tomas * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps 2657c9de560dSAlex Tomas * Returns 0 if success or error code 2658c9de560dSAlex Tomas */ 26594ddfef7bSEric Sandeen static noinline_for_stack int 26604ddfef7bSEric Sandeen ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 2661498e5f24STheodore Ts'o handle_t *handle, unsigned int reserv_blks) 2662c9de560dSAlex Tomas { 2663c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 2664c9de560dSAlex Tomas struct ext4_super_block *es; 2665c9de560dSAlex Tomas struct ext4_group_desc *gdp; 2666c9de560dSAlex Tomas struct buffer_head *gdp_bh; 2667c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2668c9de560dSAlex Tomas struct super_block *sb; 2669c9de560dSAlex Tomas ext4_fsblk_t block; 2670519deca0SAneesh Kumar K.V int err, len; 2671c9de560dSAlex Tomas 2672c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 2673c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len <= 0); 2674c9de560dSAlex Tomas 2675c9de560dSAlex Tomas sb = ac->ac_sb; 2676c9de560dSAlex Tomas sbi = EXT4_SB(sb); 2677c9de560dSAlex Tomas es = sbi->s_es; 2678c9de560dSAlex Tomas 2679c9de560dSAlex Tomas 2680c9de560dSAlex Tomas err = -EIO; 2681574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 2682c9de560dSAlex Tomas if (!bitmap_bh) 2683c9de560dSAlex Tomas goto out_err; 2684c9de560dSAlex Tomas 2685c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 2686c9de560dSAlex Tomas if (err) 2687c9de560dSAlex Tomas goto out_err; 2688c9de560dSAlex Tomas 2689c9de560dSAlex Tomas err = -EIO; 2690c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 2691c9de560dSAlex Tomas if (!gdp) 2692c9de560dSAlex Tomas goto out_err; 2693c9de560dSAlex Tomas 2694a9df9a49STheodore Ts'o ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 26959fd9784cSThadeu Lima de Souza Cascardo ext4_free_blks_count(sb, gdp)); 269603cddb80SAneesh Kumar K.V 2697c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gdp_bh); 2698c9de560dSAlex Tomas if (err) 2699c9de560dSAlex Tomas goto out_err; 2700c9de560dSAlex Tomas 2701c9de560dSAlex Tomas block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb) 2702c9de560dSAlex Tomas + ac->ac_b_ex.fe_start 2703c9de560dSAlex Tomas + le32_to_cpu(es->s_first_data_block); 2704c9de560dSAlex Tomas 2705519deca0SAneesh Kumar K.V len = ac->ac_b_ex.fe_len; 27066fd058f7STheodore Ts'o if (!ext4_data_block_valid(sbi, block, len)) { 270746e665e9SHarvey Harrison ext4_error(sb, __func__, 27086fd058f7STheodore Ts'o "Allocating blocks %llu-%llu which overlap " 27096fd058f7STheodore Ts'o "fs metadata\n", block, block+len); 2710519deca0SAneesh Kumar K.V /* File system mounted not to panic on error 2711519deca0SAneesh Kumar K.V * Fix the bitmap and repeat the block allocation 2712519deca0SAneesh Kumar K.V * We leak some of the blocks here. 2713519deca0SAneesh Kumar K.V */ 2714955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2715955ce5f5SAneesh Kumar K.V mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 2716519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len); 2717955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 27180390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 2719519deca0SAneesh Kumar K.V if (!err) 2720519deca0SAneesh Kumar K.V err = -EAGAIN; 2721519deca0SAneesh Kumar K.V goto out_err; 2722c9de560dSAlex Tomas } 2723955ce5f5SAneesh Kumar K.V 2724955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2725c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 2726c9de560dSAlex Tomas { 2727c9de560dSAlex Tomas int i; 2728c9de560dSAlex Tomas for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 2729c9de560dSAlex Tomas BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 2730c9de560dSAlex Tomas bitmap_bh->b_data)); 2731c9de560dSAlex Tomas } 2732c9de560dSAlex Tomas } 2733c9de560dSAlex Tomas #endif 2734955ce5f5SAneesh Kumar K.V mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len); 2735c9de560dSAlex Tomas if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 2736c9de560dSAlex Tomas gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 2737560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, 2738560671a0SAneesh Kumar K.V ext4_free_blocks_after_init(sb, 2739560671a0SAneesh Kumar K.V ac->ac_b_ex.fe_group, gdp)); 2740c9de560dSAlex Tomas } 2741560671a0SAneesh Kumar K.V len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len; 2742560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, len); 2743c9de560dSAlex Tomas gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); 2744955ce5f5SAneesh Kumar K.V 2745955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 27466bc6e63fSAneesh Kumar K.V percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); 2747d2a17637SMingming Cao /* 27486bc6e63fSAneesh Kumar K.V * Now reduce the dirty block count also. Should not go negative 2749d2a17637SMingming Cao */ 27506bc6e63fSAneesh Kumar K.V if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 27516bc6e63fSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 27526bc6e63fSAneesh Kumar K.V percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks); 275360e58e0fSMingming Cao else { 27546bc6e63fSAneesh Kumar K.V percpu_counter_sub(&sbi->s_dirtyblocks_counter, 27556bc6e63fSAneesh Kumar K.V ac->ac_b_ex.fe_len); 275660e58e0fSMingming Cao /* convert reserved quota blocks to real quota blocks */ 275760e58e0fSMingming Cao vfs_dq_claim_block(ac->ac_inode, ac->ac_b_ex.fe_len); 275860e58e0fSMingming Cao } 2759c9de560dSAlex Tomas 2760772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 2761772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, 2762772cb7c8SJose R. Santos ac->ac_b_ex.fe_group); 27639f24e420STheodore Ts'o atomic_sub(ac->ac_b_ex.fe_len, 27649f24e420STheodore Ts'o &sbi->s_flex_groups[flex_group].free_blocks); 2765772cb7c8SJose R. Santos } 2766772cb7c8SJose R. Santos 27670390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 2768c9de560dSAlex Tomas if (err) 2769c9de560dSAlex Tomas goto out_err; 27700390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 2771c9de560dSAlex Tomas 2772c9de560dSAlex Tomas out_err: 2773c9de560dSAlex Tomas sb->s_dirt = 1; 277442a10addSAneesh Kumar K.V brelse(bitmap_bh); 2775c9de560dSAlex Tomas return err; 2776c9de560dSAlex Tomas } 2777c9de560dSAlex Tomas 2778c9de560dSAlex Tomas /* 2779c9de560dSAlex Tomas * here we normalize request for locality group 2780c9de560dSAlex Tomas * Group request are normalized to s_strip size if we set the same via mount 2781c9de560dSAlex Tomas * option. If not we set it to s_mb_group_prealloc which can be configured via 2782b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_group_prealloc 2783c9de560dSAlex Tomas * 2784c9de560dSAlex Tomas * XXX: should we try to preallocate more than the group has now? 2785c9de560dSAlex Tomas */ 2786c9de560dSAlex Tomas static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 2787c9de560dSAlex Tomas { 2788c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2789c9de560dSAlex Tomas struct ext4_locality_group *lg = ac->ac_lg; 2790c9de560dSAlex Tomas 2791c9de560dSAlex Tomas BUG_ON(lg == NULL); 2792c9de560dSAlex Tomas if (EXT4_SB(sb)->s_stripe) 2793c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe; 2794c9de560dSAlex Tomas else 2795c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 27966ba495e9STheodore Ts'o mb_debug(1, "#%u: goal %u blocks for locality group\n", 2797c9de560dSAlex Tomas current->pid, ac->ac_g_ex.fe_len); 2798c9de560dSAlex Tomas } 2799c9de560dSAlex Tomas 2800c9de560dSAlex Tomas /* 2801c9de560dSAlex Tomas * Normalization means making request better in terms of 2802c9de560dSAlex Tomas * size and alignment 2803c9de560dSAlex Tomas */ 28044ddfef7bSEric Sandeen static noinline_for_stack void 28054ddfef7bSEric Sandeen ext4_mb_normalize_request(struct ext4_allocation_context *ac, 2806c9de560dSAlex Tomas struct ext4_allocation_request *ar) 2807c9de560dSAlex Tomas { 2808c9de560dSAlex Tomas int bsbits, max; 2809c9de560dSAlex Tomas ext4_lblk_t end; 2810c9de560dSAlex Tomas loff_t size, orig_size, start_off; 2811c9de560dSAlex Tomas ext4_lblk_t start, orig_start; 2812c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 28139a0762c5SAneesh Kumar K.V struct ext4_prealloc_space *pa; 2814c9de560dSAlex Tomas 2815c9de560dSAlex Tomas /* do normalize only data requests, metadata requests 2816c9de560dSAlex Tomas do not need preallocation */ 2817c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 2818c9de560dSAlex Tomas return; 2819c9de560dSAlex Tomas 2820c9de560dSAlex Tomas /* sometime caller may want exact blocks */ 2821c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2822c9de560dSAlex Tomas return; 2823c9de560dSAlex Tomas 2824c9de560dSAlex Tomas /* caller may indicate that preallocation isn't 2825c9de560dSAlex Tomas * required (it's a tail, for example) */ 2826c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 2827c9de560dSAlex Tomas return; 2828c9de560dSAlex Tomas 2829c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 2830c9de560dSAlex Tomas ext4_mb_normalize_group_request(ac); 2831c9de560dSAlex Tomas return ; 2832c9de560dSAlex Tomas } 2833c9de560dSAlex Tomas 2834c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 2835c9de560dSAlex Tomas 2836c9de560dSAlex Tomas /* first, let's learn actual file size 2837c9de560dSAlex Tomas * given current request is allocated */ 2838c9de560dSAlex Tomas size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; 2839c9de560dSAlex Tomas size = size << bsbits; 2840c9de560dSAlex Tomas if (size < i_size_read(ac->ac_inode)) 2841c9de560dSAlex Tomas size = i_size_read(ac->ac_inode); 2842c9de560dSAlex Tomas 28431930479cSValerie Clement /* max size of free chunks */ 28441930479cSValerie Clement max = 2 << bsbits; 2845c9de560dSAlex Tomas 28461930479cSValerie Clement #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 28471930479cSValerie Clement (req <= (size) || max <= (chunk_size)) 2848c9de560dSAlex Tomas 2849c9de560dSAlex Tomas /* first, try to predict filesize */ 2850c9de560dSAlex Tomas /* XXX: should this table be tunable? */ 2851c9de560dSAlex Tomas start_off = 0; 2852c9de560dSAlex Tomas if (size <= 16 * 1024) { 2853c9de560dSAlex Tomas size = 16 * 1024; 2854c9de560dSAlex Tomas } else if (size <= 32 * 1024) { 2855c9de560dSAlex Tomas size = 32 * 1024; 2856c9de560dSAlex Tomas } else if (size <= 64 * 1024) { 2857c9de560dSAlex Tomas size = 64 * 1024; 2858c9de560dSAlex Tomas } else if (size <= 128 * 1024) { 2859c9de560dSAlex Tomas size = 128 * 1024; 2860c9de560dSAlex Tomas } else if (size <= 256 * 1024) { 2861c9de560dSAlex Tomas size = 256 * 1024; 2862c9de560dSAlex Tomas } else if (size <= 512 * 1024) { 2863c9de560dSAlex Tomas size = 512 * 1024; 2864c9de560dSAlex Tomas } else if (size <= 1024 * 1024) { 2865c9de560dSAlex Tomas size = 1024 * 1024; 28661930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 2867c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 28681930479cSValerie Clement (21 - bsbits)) << 21; 28691930479cSValerie Clement size = 2 * 1024 * 1024; 28701930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 2871c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 2872c9de560dSAlex Tomas (22 - bsbits)) << 22; 2873c9de560dSAlex Tomas size = 4 * 1024 * 1024; 2874c9de560dSAlex Tomas } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, 28751930479cSValerie Clement (8<<20)>>bsbits, max, 8 * 1024)) { 2876c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 2877c9de560dSAlex Tomas (23 - bsbits)) << 23; 2878c9de560dSAlex Tomas size = 8 * 1024 * 1024; 2879c9de560dSAlex Tomas } else { 2880c9de560dSAlex Tomas start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; 2881c9de560dSAlex Tomas size = ac->ac_o_ex.fe_len << bsbits; 2882c9de560dSAlex Tomas } 2883c9de560dSAlex Tomas orig_size = size = size >> bsbits; 2884c9de560dSAlex Tomas orig_start = start = start_off >> bsbits; 2885c9de560dSAlex Tomas 2886c9de560dSAlex Tomas /* don't cover already allocated blocks in selected range */ 2887c9de560dSAlex Tomas if (ar->pleft && start <= ar->lleft) { 2888c9de560dSAlex Tomas size -= ar->lleft + 1 - start; 2889c9de560dSAlex Tomas start = ar->lleft + 1; 2890c9de560dSAlex Tomas } 2891c9de560dSAlex Tomas if (ar->pright && start + size - 1 >= ar->lright) 2892c9de560dSAlex Tomas size -= start + size - ar->lright; 2893c9de560dSAlex Tomas 2894c9de560dSAlex Tomas end = start + size; 2895c9de560dSAlex Tomas 2896c9de560dSAlex Tomas /* check we don't cross already preallocated blocks */ 2897c9de560dSAlex Tomas rcu_read_lock(); 28989a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 2899498e5f24STheodore Ts'o ext4_lblk_t pa_end; 2900c9de560dSAlex Tomas 2901c9de560dSAlex Tomas if (pa->pa_deleted) 2902c9de560dSAlex Tomas continue; 2903c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 2904c9de560dSAlex Tomas if (pa->pa_deleted) { 2905c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 2906c9de560dSAlex Tomas continue; 2907c9de560dSAlex Tomas } 2908c9de560dSAlex Tomas 2909c9de560dSAlex Tomas pa_end = pa->pa_lstart + pa->pa_len; 2910c9de560dSAlex Tomas 2911c9de560dSAlex Tomas /* PA must not overlap original request */ 2912c9de560dSAlex Tomas BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || 2913c9de560dSAlex Tomas ac->ac_o_ex.fe_logical < pa->pa_lstart)); 2914c9de560dSAlex Tomas 291538877f4eSEric Sandeen /* skip PAs this normalized request doesn't overlap with */ 291638877f4eSEric Sandeen if (pa->pa_lstart >= end || pa_end <= start) { 2917c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 2918c9de560dSAlex Tomas continue; 2919c9de560dSAlex Tomas } 2920c9de560dSAlex Tomas BUG_ON(pa->pa_lstart <= start && pa_end >= end); 2921c9de560dSAlex Tomas 292238877f4eSEric Sandeen /* adjust start or end to be adjacent to this pa */ 2923c9de560dSAlex Tomas if (pa_end <= ac->ac_o_ex.fe_logical) { 2924c9de560dSAlex Tomas BUG_ON(pa_end < start); 2925c9de560dSAlex Tomas start = pa_end; 292638877f4eSEric Sandeen } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { 2927c9de560dSAlex Tomas BUG_ON(pa->pa_lstart > end); 2928c9de560dSAlex Tomas end = pa->pa_lstart; 2929c9de560dSAlex Tomas } 2930c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 2931c9de560dSAlex Tomas } 2932c9de560dSAlex Tomas rcu_read_unlock(); 2933c9de560dSAlex Tomas size = end - start; 2934c9de560dSAlex Tomas 2935c9de560dSAlex Tomas /* XXX: extra loop to check we really don't overlap preallocations */ 2936c9de560dSAlex Tomas rcu_read_lock(); 29379a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 2938498e5f24STheodore Ts'o ext4_lblk_t pa_end; 2939c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 2940c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 2941c9de560dSAlex Tomas pa_end = pa->pa_lstart + pa->pa_len; 2942c9de560dSAlex Tomas BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); 2943c9de560dSAlex Tomas } 2944c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 2945c9de560dSAlex Tomas } 2946c9de560dSAlex Tomas rcu_read_unlock(); 2947c9de560dSAlex Tomas 2948c9de560dSAlex Tomas if (start + size <= ac->ac_o_ex.fe_logical && 2949c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical) { 2950c9de560dSAlex Tomas printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n", 2951c9de560dSAlex Tomas (unsigned long) start, (unsigned long) size, 2952c9de560dSAlex Tomas (unsigned long) ac->ac_o_ex.fe_logical); 2953c9de560dSAlex Tomas } 2954c9de560dSAlex Tomas BUG_ON(start + size <= ac->ac_o_ex.fe_logical && 2955c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical); 29568d03c7a0SEric Sandeen BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 2957c9de560dSAlex Tomas 2958c9de560dSAlex Tomas /* now prepare goal request */ 2959c9de560dSAlex Tomas 2960c9de560dSAlex Tomas /* XXX: is it better to align blocks WRT to logical 2961c9de560dSAlex Tomas * placement or satisfy big request as is */ 2962c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = start; 2963c9de560dSAlex Tomas ac->ac_g_ex.fe_len = size; 2964c9de560dSAlex Tomas 2965c9de560dSAlex Tomas /* define goal start in order to merge */ 2966c9de560dSAlex Tomas if (ar->pright && (ar->lright == (start + size))) { 2967c9de560dSAlex Tomas /* merge to the right */ 2968c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 2969c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 2970c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 2971c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 2972c9de560dSAlex Tomas } 2973c9de560dSAlex Tomas if (ar->pleft && (ar->lleft + 1 == start)) { 2974c9de560dSAlex Tomas /* merge to the left */ 2975c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 2976c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 2977c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 2978c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 2979c9de560dSAlex Tomas } 2980c9de560dSAlex Tomas 29816ba495e9STheodore Ts'o mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, 2982c9de560dSAlex Tomas (unsigned) orig_size, (unsigned) start); 2983c9de560dSAlex Tomas } 2984c9de560dSAlex Tomas 2985c9de560dSAlex Tomas static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 2986c9de560dSAlex Tomas { 2987c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2988c9de560dSAlex Tomas 2989c9de560dSAlex Tomas if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { 2990c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_reqs); 2991c9de560dSAlex Tomas atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 2992c9de560dSAlex Tomas if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len) 2993c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_success); 2994c9de560dSAlex Tomas atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 2995c9de560dSAlex Tomas if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 2996c9de560dSAlex Tomas ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 2997c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_goals); 2998c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan) 2999c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_breaks); 3000c9de560dSAlex Tomas } 3001c9de560dSAlex Tomas 3002*296c355cSTheodore Ts'o if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 3003*296c355cSTheodore Ts'o trace_ext4_mballoc_alloc(ac); 3004*296c355cSTheodore Ts'o else 3005*296c355cSTheodore Ts'o trace_ext4_mballoc_prealloc(ac); 3006c9de560dSAlex Tomas } 3007c9de560dSAlex Tomas 3008c9de560dSAlex Tomas /* 3009c9de560dSAlex Tomas * use blocks preallocated to inode 3010c9de560dSAlex Tomas */ 3011c9de560dSAlex Tomas static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 3012c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3013c9de560dSAlex Tomas { 3014c9de560dSAlex Tomas ext4_fsblk_t start; 3015c9de560dSAlex Tomas ext4_fsblk_t end; 3016c9de560dSAlex Tomas int len; 3017c9de560dSAlex Tomas 3018c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3019c9de560dSAlex Tomas start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 3020c9de560dSAlex Tomas end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len); 3021c9de560dSAlex Tomas len = end - start; 3022c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 3023c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3024c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3025c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3026c9de560dSAlex Tomas ac->ac_pa = pa; 3027c9de560dSAlex Tomas 3028c9de560dSAlex Tomas BUG_ON(start < pa->pa_pstart); 3029c9de560dSAlex Tomas BUG_ON(start + len > pa->pa_pstart + pa->pa_len); 3030c9de560dSAlex Tomas BUG_ON(pa->pa_free < len); 3031c9de560dSAlex Tomas pa->pa_free -= len; 3032c9de560dSAlex Tomas 30336ba495e9STheodore Ts'o mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); 3034c9de560dSAlex Tomas } 3035c9de560dSAlex Tomas 3036c9de560dSAlex Tomas /* 3037c9de560dSAlex Tomas * use blocks preallocated to locality group 3038c9de560dSAlex Tomas */ 3039c9de560dSAlex Tomas static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 3040c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3041c9de560dSAlex Tomas { 304203cddb80SAneesh Kumar K.V unsigned int len = ac->ac_o_ex.fe_len; 30436be2ded1SAneesh Kumar K.V 3044c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 3045c9de560dSAlex Tomas &ac->ac_b_ex.fe_group, 3046c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3047c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3048c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3049c9de560dSAlex Tomas ac->ac_pa = pa; 3050c9de560dSAlex Tomas 3051c9de560dSAlex Tomas /* we don't correct pa_pstart or pa_plen here to avoid 305226346ff6SAneesh Kumar K.V * possible race when the group is being loaded concurrently 3053c9de560dSAlex Tomas * instead we correct pa later, after blocks are marked 305426346ff6SAneesh Kumar K.V * in on-disk bitmap -- see ext4_mb_release_context() 305526346ff6SAneesh Kumar K.V * Other CPUs are prevented from allocating from this pa by lg_mutex 3056c9de560dSAlex Tomas */ 30576ba495e9STheodore Ts'o mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); 3058c9de560dSAlex Tomas } 3059c9de560dSAlex Tomas 3060c9de560dSAlex Tomas /* 30615e745b04SAneesh Kumar K.V * Return the prealloc space that have minimal distance 30625e745b04SAneesh Kumar K.V * from the goal block. @cpa is the prealloc 30635e745b04SAneesh Kumar K.V * space that is having currently known minimal distance 30645e745b04SAneesh Kumar K.V * from the goal block. 30655e745b04SAneesh Kumar K.V */ 30665e745b04SAneesh Kumar K.V static struct ext4_prealloc_space * 30675e745b04SAneesh Kumar K.V ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 30685e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, 30695e745b04SAneesh Kumar K.V struct ext4_prealloc_space *cpa) 30705e745b04SAneesh Kumar K.V { 30715e745b04SAneesh Kumar K.V ext4_fsblk_t cur_distance, new_distance; 30725e745b04SAneesh Kumar K.V 30735e745b04SAneesh Kumar K.V if (cpa == NULL) { 30745e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 30755e745b04SAneesh Kumar K.V return pa; 30765e745b04SAneesh Kumar K.V } 30775e745b04SAneesh Kumar K.V cur_distance = abs(goal_block - cpa->pa_pstart); 30785e745b04SAneesh Kumar K.V new_distance = abs(goal_block - pa->pa_pstart); 30795e745b04SAneesh Kumar K.V 30805e745b04SAneesh Kumar K.V if (cur_distance < new_distance) 30815e745b04SAneesh Kumar K.V return cpa; 30825e745b04SAneesh Kumar K.V 30835e745b04SAneesh Kumar K.V /* drop the previous reference */ 30845e745b04SAneesh Kumar K.V atomic_dec(&cpa->pa_count); 30855e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 30865e745b04SAneesh Kumar K.V return pa; 30875e745b04SAneesh Kumar K.V } 30885e745b04SAneesh Kumar K.V 30895e745b04SAneesh Kumar K.V /* 3090c9de560dSAlex Tomas * search goal blocks in preallocated space 3091c9de560dSAlex Tomas */ 30924ddfef7bSEric Sandeen static noinline_for_stack int 30934ddfef7bSEric Sandeen ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 3094c9de560dSAlex Tomas { 30956be2ded1SAneesh Kumar K.V int order, i; 3096c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3097c9de560dSAlex Tomas struct ext4_locality_group *lg; 30985e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, *cpa = NULL; 30995e745b04SAneesh Kumar K.V ext4_fsblk_t goal_block; 3100c9de560dSAlex Tomas 3101c9de560dSAlex Tomas /* only data can be preallocated */ 3102c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3103c9de560dSAlex Tomas return 0; 3104c9de560dSAlex Tomas 3105c9de560dSAlex Tomas /* first, try per-file preallocation */ 3106c9de560dSAlex Tomas rcu_read_lock(); 31079a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3108c9de560dSAlex Tomas 3109c9de560dSAlex Tomas /* all fields in this condition don't change, 3110c9de560dSAlex Tomas * so we can skip locking for them */ 3111c9de560dSAlex Tomas if (ac->ac_o_ex.fe_logical < pa->pa_lstart || 3112c9de560dSAlex Tomas ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len) 3113c9de560dSAlex Tomas continue; 3114c9de560dSAlex Tomas 3115fb0a387dSEric Sandeen /* non-extent files can't have physical blocks past 2^32 */ 3116fb0a387dSEric Sandeen if (!(EXT4_I(ac->ac_inode)->i_flags & EXT4_EXTENTS_FL) && 3117fb0a387dSEric Sandeen pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS) 3118fb0a387dSEric Sandeen continue; 3119fb0a387dSEric Sandeen 3120c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3121c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3122c9de560dSAlex Tomas if (pa->pa_deleted == 0 && pa->pa_free) { 3123c9de560dSAlex Tomas atomic_inc(&pa->pa_count); 3124c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 3125c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3126c9de560dSAlex Tomas ac->ac_criteria = 10; 3127c9de560dSAlex Tomas rcu_read_unlock(); 3128c9de560dSAlex Tomas return 1; 3129c9de560dSAlex Tomas } 3130c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3131c9de560dSAlex Tomas } 3132c9de560dSAlex Tomas rcu_read_unlock(); 3133c9de560dSAlex Tomas 3134c9de560dSAlex Tomas /* can we use group allocation? */ 3135c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 3136c9de560dSAlex Tomas return 0; 3137c9de560dSAlex Tomas 3138c9de560dSAlex Tomas /* inode may have no locality group for some reason */ 3139c9de560dSAlex Tomas lg = ac->ac_lg; 3140c9de560dSAlex Tomas if (lg == NULL) 3141c9de560dSAlex Tomas return 0; 31426be2ded1SAneesh Kumar K.V order = fls(ac->ac_o_ex.fe_len) - 1; 31436be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 31446be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 31456be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 3146c9de560dSAlex Tomas 31475e745b04SAneesh Kumar K.V goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) + 31485e745b04SAneesh Kumar K.V ac->ac_g_ex.fe_start + 31495e745b04SAneesh Kumar K.V le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block); 31505e745b04SAneesh Kumar K.V /* 31515e745b04SAneesh Kumar K.V * search for the prealloc space that is having 31525e745b04SAneesh Kumar K.V * minimal distance from the goal block. 31535e745b04SAneesh Kumar K.V */ 31546be2ded1SAneesh Kumar K.V for (i = order; i < PREALLOC_TB_SIZE; i++) { 3155c9de560dSAlex Tomas rcu_read_lock(); 31566be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], 31576be2ded1SAneesh Kumar K.V pa_inode_list) { 3158c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 31596be2ded1SAneesh Kumar K.V if (pa->pa_deleted == 0 && 31606be2ded1SAneesh Kumar K.V pa->pa_free >= ac->ac_o_ex.fe_len) { 31615e745b04SAneesh Kumar K.V 31625e745b04SAneesh Kumar K.V cpa = ext4_mb_check_group_pa(goal_block, 31635e745b04SAneesh Kumar K.V pa, cpa); 31645e745b04SAneesh Kumar K.V } 3165c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 31665e745b04SAneesh Kumar K.V } 31675e745b04SAneesh Kumar K.V rcu_read_unlock(); 31685e745b04SAneesh Kumar K.V } 31695e745b04SAneesh Kumar K.V if (cpa) { 31705e745b04SAneesh Kumar K.V ext4_mb_use_group_pa(ac, cpa); 3171c9de560dSAlex Tomas ac->ac_criteria = 20; 3172c9de560dSAlex Tomas return 1; 3173c9de560dSAlex Tomas } 3174c9de560dSAlex Tomas return 0; 3175c9de560dSAlex Tomas } 3176c9de560dSAlex Tomas 3177c9de560dSAlex Tomas /* 31787a2fcbf7SAneesh Kumar K.V * the function goes through all block freed in the group 31797a2fcbf7SAneesh Kumar K.V * but not yet committed and marks them used in in-core bitmap. 31807a2fcbf7SAneesh Kumar K.V * buddy must be generated from this bitmap 3181955ce5f5SAneesh Kumar K.V * Need to be called with the ext4 group lock held 31827a2fcbf7SAneesh Kumar K.V */ 31837a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 31847a2fcbf7SAneesh Kumar K.V ext4_group_t group) 31857a2fcbf7SAneesh Kumar K.V { 31867a2fcbf7SAneesh Kumar K.V struct rb_node *n; 31877a2fcbf7SAneesh Kumar K.V struct ext4_group_info *grp; 31887a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 31897a2fcbf7SAneesh Kumar K.V 31907a2fcbf7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 31917a2fcbf7SAneesh Kumar K.V n = rb_first(&(grp->bb_free_root)); 31927a2fcbf7SAneesh Kumar K.V 31937a2fcbf7SAneesh Kumar K.V while (n) { 31947a2fcbf7SAneesh Kumar K.V entry = rb_entry(n, struct ext4_free_data, node); 3195955ce5f5SAneesh Kumar K.V mb_set_bits(bitmap, entry->start_blk, entry->count); 31967a2fcbf7SAneesh Kumar K.V n = rb_next(n); 31977a2fcbf7SAneesh Kumar K.V } 31987a2fcbf7SAneesh Kumar K.V return; 31997a2fcbf7SAneesh Kumar K.V } 32007a2fcbf7SAneesh Kumar K.V 32017a2fcbf7SAneesh Kumar K.V /* 3202c9de560dSAlex Tomas * the function goes through all preallocation in this group and marks them 3203c9de560dSAlex Tomas * used in in-core bitmap. buddy must be generated from this bitmap 3204955ce5f5SAneesh Kumar K.V * Need to be called with ext4 group lock held 3205c9de560dSAlex Tomas */ 3206089ceeccSEric Sandeen static noinline_for_stack 3207089ceeccSEric Sandeen void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 3208c9de560dSAlex Tomas ext4_group_t group) 3209c9de560dSAlex Tomas { 3210c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3211c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3212c9de560dSAlex Tomas struct list_head *cur; 3213c9de560dSAlex Tomas ext4_group_t groupnr; 3214c9de560dSAlex Tomas ext4_grpblk_t start; 3215c9de560dSAlex Tomas int preallocated = 0; 3216c9de560dSAlex Tomas int count = 0; 3217c9de560dSAlex Tomas int len; 3218c9de560dSAlex Tomas 3219c9de560dSAlex Tomas /* all form of preallocation discards first load group, 3220c9de560dSAlex Tomas * so the only competing code is preallocation use. 3221c9de560dSAlex Tomas * we don't need any locking here 3222c9de560dSAlex Tomas * notice we do NOT ignore preallocations with pa_deleted 3223c9de560dSAlex Tomas * otherwise we could leave used blocks available for 3224c9de560dSAlex Tomas * allocation in buddy when concurrent ext4_mb_put_pa() 3225c9de560dSAlex Tomas * is dropping preallocation 3226c9de560dSAlex Tomas */ 3227c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 3228c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3229c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3230c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3231c9de560dSAlex Tomas &groupnr, &start); 3232c9de560dSAlex Tomas len = pa->pa_len; 3233c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3234c9de560dSAlex Tomas if (unlikely(len == 0)) 3235c9de560dSAlex Tomas continue; 3236c9de560dSAlex Tomas BUG_ON(groupnr != group); 3237955ce5f5SAneesh Kumar K.V mb_set_bits(bitmap, start, len); 3238c9de560dSAlex Tomas preallocated += len; 3239c9de560dSAlex Tomas count++; 3240c9de560dSAlex Tomas } 32416ba495e9STheodore Ts'o mb_debug(1, "prellocated %u for group %u\n", preallocated, group); 3242c9de560dSAlex Tomas } 3243c9de560dSAlex Tomas 3244c9de560dSAlex Tomas static void ext4_mb_pa_callback(struct rcu_head *head) 3245c9de560dSAlex Tomas { 3246c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3247c9de560dSAlex Tomas pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 3248c9de560dSAlex Tomas kmem_cache_free(ext4_pspace_cachep, pa); 3249c9de560dSAlex Tomas } 3250c9de560dSAlex Tomas 3251c9de560dSAlex Tomas /* 3252c9de560dSAlex Tomas * drops a reference to preallocated space descriptor 3253c9de560dSAlex Tomas * if this was the last reference and the space is consumed 3254c9de560dSAlex Tomas */ 3255c9de560dSAlex Tomas static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 3256c9de560dSAlex Tomas struct super_block *sb, struct ext4_prealloc_space *pa) 3257c9de560dSAlex Tomas { 3258a9df9a49STheodore Ts'o ext4_group_t grp; 3259d33a1976SEric Sandeen ext4_fsblk_t grp_blk; 3260c9de560dSAlex Tomas 3261c9de560dSAlex Tomas if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) 3262c9de560dSAlex Tomas return; 3263c9de560dSAlex Tomas 3264c9de560dSAlex Tomas /* in this short window concurrent discard can set pa_deleted */ 3265c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3266c9de560dSAlex Tomas if (pa->pa_deleted == 1) { 3267c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3268c9de560dSAlex Tomas return; 3269c9de560dSAlex Tomas } 3270c9de560dSAlex Tomas 3271c9de560dSAlex Tomas pa->pa_deleted = 1; 3272c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3273c9de560dSAlex Tomas 3274d33a1976SEric Sandeen grp_blk = pa->pa_pstart; 3275cc0fb9adSAneesh Kumar K.V /* 3276cc0fb9adSAneesh Kumar K.V * If doing group-based preallocation, pa_pstart may be in the 3277cc0fb9adSAneesh Kumar K.V * next group when pa is used up 3278cc0fb9adSAneesh Kumar K.V */ 3279cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 3280d33a1976SEric Sandeen grp_blk--; 3281d33a1976SEric Sandeen 3282d33a1976SEric Sandeen ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); 3283c9de560dSAlex Tomas 3284c9de560dSAlex Tomas /* 3285c9de560dSAlex Tomas * possible race: 3286c9de560dSAlex Tomas * 3287c9de560dSAlex Tomas * P1 (buddy init) P2 (regular allocation) 3288c9de560dSAlex Tomas * find block B in PA 3289c9de560dSAlex Tomas * copy on-disk bitmap to buddy 3290c9de560dSAlex Tomas * mark B in on-disk bitmap 3291c9de560dSAlex Tomas * drop PA from group 3292c9de560dSAlex Tomas * mark all PAs in buddy 3293c9de560dSAlex Tomas * 3294c9de560dSAlex Tomas * thus, P1 initializes buddy with B available. to prevent this 3295c9de560dSAlex Tomas * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 3296c9de560dSAlex Tomas * against that pair 3297c9de560dSAlex Tomas */ 3298c9de560dSAlex Tomas ext4_lock_group(sb, grp); 3299c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3300c9de560dSAlex Tomas ext4_unlock_group(sb, grp); 3301c9de560dSAlex Tomas 3302c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3303c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3304c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3305c9de560dSAlex Tomas 3306c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3307c9de560dSAlex Tomas } 3308c9de560dSAlex Tomas 3309c9de560dSAlex Tomas /* 3310c9de560dSAlex Tomas * creates new preallocated space for given inode 3311c9de560dSAlex Tomas */ 33124ddfef7bSEric Sandeen static noinline_for_stack int 33134ddfef7bSEric Sandeen ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 3314c9de560dSAlex Tomas { 3315c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3316c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3317c9de560dSAlex Tomas struct ext4_group_info *grp; 3318c9de560dSAlex Tomas struct ext4_inode_info *ei; 3319c9de560dSAlex Tomas 3320c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3321c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3322c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3323c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3324c9de560dSAlex Tomas 3325c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3326c9de560dSAlex Tomas if (pa == NULL) 3327c9de560dSAlex Tomas return -ENOMEM; 3328c9de560dSAlex Tomas 3329c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { 3330c9de560dSAlex Tomas int winl; 3331c9de560dSAlex Tomas int wins; 3332c9de560dSAlex Tomas int win; 3333c9de560dSAlex Tomas int offs; 3334c9de560dSAlex Tomas 3335c9de560dSAlex Tomas /* we can't allocate as much as normalizer wants. 3336c9de560dSAlex Tomas * so, found space must get proper lstart 3337c9de560dSAlex Tomas * to cover original request */ 3338c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 3339c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 3340c9de560dSAlex Tomas 3341c9de560dSAlex Tomas /* we're limited by original request in that 3342c9de560dSAlex Tomas * logical block must be covered any way 3343c9de560dSAlex Tomas * winl is window we can move our chunk within */ 3344c9de560dSAlex Tomas winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; 3345c9de560dSAlex Tomas 3346c9de560dSAlex Tomas /* also, we should cover whole original request */ 3347c9de560dSAlex Tomas wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len; 3348c9de560dSAlex Tomas 3349c9de560dSAlex Tomas /* the smallest one defines real window */ 3350c9de560dSAlex Tomas win = min(winl, wins); 3351c9de560dSAlex Tomas 3352c9de560dSAlex Tomas offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len; 3353c9de560dSAlex Tomas if (offs && offs < win) 3354c9de560dSAlex Tomas win = offs; 3355c9de560dSAlex Tomas 3356c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win; 3357c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 3358c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 3359c9de560dSAlex Tomas } 3360c9de560dSAlex Tomas 3361c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3362c9de560dSAlex Tomas * allocated blocks for history */ 3363c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3364c9de560dSAlex Tomas 3365c9de560dSAlex Tomas pa->pa_lstart = ac->ac_b_ex.fe_logical; 3366c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3367c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3368c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3369c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3370c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 3371d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3372d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 3373c9de560dSAlex Tomas pa->pa_deleted = 0; 3374cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_INODE_PA; 3375c9de560dSAlex Tomas 33766ba495e9STheodore Ts'o mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, 3377c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 33789bffad1eSTheodore Ts'o trace_ext4_mb_new_inode_pa(ac, pa); 3379c9de560dSAlex Tomas 3380c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 3381c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3382c9de560dSAlex Tomas 3383c9de560dSAlex Tomas ei = EXT4_I(ac->ac_inode); 3384c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3385c9de560dSAlex Tomas 3386c9de560dSAlex Tomas pa->pa_obj_lock = &ei->i_prealloc_lock; 3387c9de560dSAlex Tomas pa->pa_inode = ac->ac_inode; 3388c9de560dSAlex Tomas 3389c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3390c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3391c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3392c9de560dSAlex Tomas 3393c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3394c9de560dSAlex Tomas list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); 3395c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3396c9de560dSAlex Tomas 3397c9de560dSAlex Tomas return 0; 3398c9de560dSAlex Tomas } 3399c9de560dSAlex Tomas 3400c9de560dSAlex Tomas /* 3401c9de560dSAlex Tomas * creates new preallocated space for locality group inodes belongs to 3402c9de560dSAlex Tomas */ 34034ddfef7bSEric Sandeen static noinline_for_stack int 34044ddfef7bSEric Sandeen ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 3405c9de560dSAlex Tomas { 3406c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3407c9de560dSAlex Tomas struct ext4_locality_group *lg; 3408c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3409c9de560dSAlex Tomas struct ext4_group_info *grp; 3410c9de560dSAlex Tomas 3411c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3412c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3413c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3414c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3415c9de560dSAlex Tomas 3416c9de560dSAlex Tomas BUG_ON(ext4_pspace_cachep == NULL); 3417c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3418c9de560dSAlex Tomas if (pa == NULL) 3419c9de560dSAlex Tomas return -ENOMEM; 3420c9de560dSAlex Tomas 3421c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3422c9de560dSAlex Tomas * allocated blocks for history */ 3423c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3424c9de560dSAlex Tomas 3425c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3426c9de560dSAlex Tomas pa->pa_lstart = pa->pa_pstart; 3427c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3428c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3429c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3430c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 34316be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3432d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 3433c9de560dSAlex Tomas pa->pa_deleted = 0; 3434cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_GROUP_PA; 3435c9de560dSAlex Tomas 34366ba495e9STheodore Ts'o mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, 3437c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 34389bffad1eSTheodore Ts'o trace_ext4_mb_new_group_pa(ac, pa); 3439c9de560dSAlex Tomas 3440c9de560dSAlex Tomas ext4_mb_use_group_pa(ac, pa); 3441c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3442c9de560dSAlex Tomas 3443c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3444c9de560dSAlex Tomas lg = ac->ac_lg; 3445c9de560dSAlex Tomas BUG_ON(lg == NULL); 3446c9de560dSAlex Tomas 3447c9de560dSAlex Tomas pa->pa_obj_lock = &lg->lg_prealloc_lock; 3448c9de560dSAlex Tomas pa->pa_inode = NULL; 3449c9de560dSAlex Tomas 3450c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3451c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3452c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3453c9de560dSAlex Tomas 34546be2ded1SAneesh Kumar K.V /* 34556be2ded1SAneesh Kumar K.V * We will later add the new pa to the right bucket 34566be2ded1SAneesh Kumar K.V * after updating the pa_free in ext4_mb_release_context 34576be2ded1SAneesh Kumar K.V */ 3458c9de560dSAlex Tomas return 0; 3459c9de560dSAlex Tomas } 3460c9de560dSAlex Tomas 3461c9de560dSAlex Tomas static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 3462c9de560dSAlex Tomas { 3463c9de560dSAlex Tomas int err; 3464c9de560dSAlex Tomas 3465c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 3466c9de560dSAlex Tomas err = ext4_mb_new_group_pa(ac); 3467c9de560dSAlex Tomas else 3468c9de560dSAlex Tomas err = ext4_mb_new_inode_pa(ac); 3469c9de560dSAlex Tomas return err; 3470c9de560dSAlex Tomas } 3471c9de560dSAlex Tomas 3472c9de560dSAlex Tomas /* 3473c9de560dSAlex Tomas * finds all unused blocks in on-disk bitmap, frees them in 3474c9de560dSAlex Tomas * in-core bitmap and buddy. 3475c9de560dSAlex Tomas * @pa must be unlinked from inode and group lists, so that 3476c9de560dSAlex Tomas * nobody else can find/use it. 3477c9de560dSAlex Tomas * the caller MUST hold group/inode locks. 3478c9de560dSAlex Tomas * TODO: optimize the case when there are no in-core structures yet 3479c9de560dSAlex Tomas */ 34804ddfef7bSEric Sandeen static noinline_for_stack int 34814ddfef7bSEric Sandeen ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 3482c83617dbSAneesh Kumar K.V struct ext4_prealloc_space *pa, 3483c83617dbSAneesh Kumar K.V struct ext4_allocation_context *ac) 3484c9de560dSAlex Tomas { 3485c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3486c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3487498e5f24STheodore Ts'o unsigned int end; 3488498e5f24STheodore Ts'o unsigned int next; 3489c9de560dSAlex Tomas ext4_group_t group; 3490c9de560dSAlex Tomas ext4_grpblk_t bit; 3491ba80b101STheodore Ts'o unsigned long long grp_blk_start; 3492c9de560dSAlex Tomas sector_t start; 3493c9de560dSAlex Tomas int err = 0; 3494c9de560dSAlex Tomas int free = 0; 3495c9de560dSAlex Tomas 3496c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3497c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3498ba80b101STheodore Ts'o grp_blk_start = pa->pa_pstart - bit; 3499c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3500c9de560dSAlex Tomas end = bit + pa->pa_len; 3501c9de560dSAlex Tomas 3502256bdb49SEric Sandeen if (ac) { 3503256bdb49SEric Sandeen ac->ac_sb = sb; 3504256bdb49SEric Sandeen ac->ac_inode = pa->pa_inode; 3505256bdb49SEric Sandeen } 3506c9de560dSAlex Tomas 3507c9de560dSAlex Tomas while (bit < end) { 3508ffad0a44SAneesh Kumar K.V bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 3509c9de560dSAlex Tomas if (bit >= end) 3510c9de560dSAlex Tomas break; 3511ffad0a44SAneesh Kumar K.V next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 3512c9de560dSAlex Tomas start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit + 3513c9de560dSAlex Tomas le32_to_cpu(sbi->s_es->s_first_data_block); 35146ba495e9STheodore Ts'o mb_debug(1, " free preallocated %u/%u in group %u\n", 3515c9de560dSAlex Tomas (unsigned) start, (unsigned) next - bit, 3516c9de560dSAlex Tomas (unsigned) group); 3517c9de560dSAlex Tomas free += next - bit; 3518c9de560dSAlex Tomas 3519256bdb49SEric Sandeen if (ac) { 3520256bdb49SEric Sandeen ac->ac_b_ex.fe_group = group; 3521256bdb49SEric Sandeen ac->ac_b_ex.fe_start = bit; 3522256bdb49SEric Sandeen ac->ac_b_ex.fe_len = next - bit; 3523256bdb49SEric Sandeen ac->ac_b_ex.fe_logical = 0; 3524*296c355cSTheodore Ts'o trace_ext4_mballoc_discard(ac); 3525256bdb49SEric Sandeen } 3526c9de560dSAlex Tomas 35279bffad1eSTheodore Ts'o trace_ext4_mb_release_inode_pa(ac, pa, grp_blk_start + bit, 3528ba80b101STheodore Ts'o next - bit); 3529c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 3530c9de560dSAlex Tomas bit = next + 1; 3531c9de560dSAlex Tomas } 3532c9de560dSAlex Tomas if (free != pa->pa_free) { 353326346ff6SAneesh Kumar K.V printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n", 3534c9de560dSAlex Tomas pa, (unsigned long) pa->pa_lstart, 3535c9de560dSAlex Tomas (unsigned long) pa->pa_pstart, 3536c9de560dSAlex Tomas (unsigned long) pa->pa_len); 35375d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, group, 35385d1b1b3fSAneesh Kumar K.V __func__, "free %u, pa_free %u", 353926346ff6SAneesh Kumar K.V free, pa->pa_free); 3540e56eb659SAneesh Kumar K.V /* 3541e56eb659SAneesh Kumar K.V * pa is already deleted so we use the value obtained 3542e56eb659SAneesh Kumar K.V * from the bitmap and continue. 3543e56eb659SAneesh Kumar K.V */ 3544c9de560dSAlex Tomas } 3545c9de560dSAlex Tomas atomic_add(free, &sbi->s_mb_discarded); 3546c9de560dSAlex Tomas 3547c9de560dSAlex Tomas return err; 3548c9de560dSAlex Tomas } 3549c9de560dSAlex Tomas 35504ddfef7bSEric Sandeen static noinline_for_stack int 35514ddfef7bSEric Sandeen ext4_mb_release_group_pa(struct ext4_buddy *e4b, 3552c83617dbSAneesh Kumar K.V struct ext4_prealloc_space *pa, 3553c83617dbSAneesh Kumar K.V struct ext4_allocation_context *ac) 3554c9de560dSAlex Tomas { 3555c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3556c9de560dSAlex Tomas ext4_group_t group; 3557c9de560dSAlex Tomas ext4_grpblk_t bit; 3558c9de560dSAlex Tomas 35599bffad1eSTheodore Ts'o trace_ext4_mb_release_group_pa(ac, pa); 3560c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3561c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3562c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3563c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 3564c9de560dSAlex Tomas atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 3565c9de560dSAlex Tomas 3566256bdb49SEric Sandeen if (ac) { 3567256bdb49SEric Sandeen ac->ac_sb = sb; 3568256bdb49SEric Sandeen ac->ac_inode = NULL; 3569256bdb49SEric Sandeen ac->ac_b_ex.fe_group = group; 3570256bdb49SEric Sandeen ac->ac_b_ex.fe_start = bit; 3571256bdb49SEric Sandeen ac->ac_b_ex.fe_len = pa->pa_len; 3572256bdb49SEric Sandeen ac->ac_b_ex.fe_logical = 0; 3573*296c355cSTheodore Ts'o trace_ext4_mballoc_discard(ac); 3574256bdb49SEric Sandeen } 3575c9de560dSAlex Tomas 3576c9de560dSAlex Tomas return 0; 3577c9de560dSAlex Tomas } 3578c9de560dSAlex Tomas 3579c9de560dSAlex Tomas /* 3580c9de560dSAlex Tomas * releases all preallocations in given group 3581c9de560dSAlex Tomas * 3582c9de560dSAlex Tomas * first, we need to decide discard policy: 3583c9de560dSAlex Tomas * - when do we discard 3584c9de560dSAlex Tomas * 1) ENOSPC 3585c9de560dSAlex Tomas * - how many do we discard 3586c9de560dSAlex Tomas * 1) how many requested 3587c9de560dSAlex Tomas */ 35884ddfef7bSEric Sandeen static noinline_for_stack int 35894ddfef7bSEric Sandeen ext4_mb_discard_group_preallocations(struct super_block *sb, 3590c9de560dSAlex Tomas ext4_group_t group, int needed) 3591c9de560dSAlex Tomas { 3592c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3593c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3594c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3595c83617dbSAneesh Kumar K.V struct ext4_allocation_context *ac; 3596c9de560dSAlex Tomas struct list_head list; 3597c9de560dSAlex Tomas struct ext4_buddy e4b; 3598c9de560dSAlex Tomas int err; 3599c9de560dSAlex Tomas int busy = 0; 3600c9de560dSAlex Tomas int free = 0; 3601c9de560dSAlex Tomas 36026ba495e9STheodore Ts'o mb_debug(1, "discard preallocation for group %u\n", group); 3603c9de560dSAlex Tomas 3604c9de560dSAlex Tomas if (list_empty(&grp->bb_prealloc_list)) 3605c9de560dSAlex Tomas return 0; 3606c9de560dSAlex Tomas 3607574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 3608c9de560dSAlex Tomas if (bitmap_bh == NULL) { 3609ce89f46cSAneesh Kumar K.V ext4_error(sb, __func__, "Error in reading block " 3610a9df9a49STheodore Ts'o "bitmap for %u", group); 3611ce89f46cSAneesh Kumar K.V return 0; 3612c9de560dSAlex Tomas } 3613c9de560dSAlex Tomas 3614c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3615ce89f46cSAneesh Kumar K.V if (err) { 3616ce89f46cSAneesh Kumar K.V ext4_error(sb, __func__, "Error in loading buddy " 3617a9df9a49STheodore Ts'o "information for %u", group); 3618ce89f46cSAneesh Kumar K.V put_bh(bitmap_bh); 3619ce89f46cSAneesh Kumar K.V return 0; 3620ce89f46cSAneesh Kumar K.V } 3621c9de560dSAlex Tomas 3622c9de560dSAlex Tomas if (needed == 0) 3623c9de560dSAlex Tomas needed = EXT4_BLOCKS_PER_GROUP(sb) + 1; 3624c9de560dSAlex Tomas 3625c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3626c83617dbSAneesh Kumar K.V ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 36279bffad1eSTheodore Ts'o if (ac) 36289bffad1eSTheodore Ts'o ac->ac_sb = sb; 3629c9de560dSAlex Tomas repeat: 3630c9de560dSAlex Tomas ext4_lock_group(sb, group); 3631c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, 3632c9de560dSAlex Tomas &grp->bb_prealloc_list, pa_group_list) { 3633c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3634c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3635c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3636c9de560dSAlex Tomas busy = 1; 3637c9de560dSAlex Tomas continue; 3638c9de560dSAlex Tomas } 3639c9de560dSAlex Tomas if (pa->pa_deleted) { 3640c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3641c9de560dSAlex Tomas continue; 3642c9de560dSAlex Tomas } 3643c9de560dSAlex Tomas 3644c9de560dSAlex Tomas /* seems this one can be freed ... */ 3645c9de560dSAlex Tomas pa->pa_deleted = 1; 3646c9de560dSAlex Tomas 3647c9de560dSAlex Tomas /* we can trust pa_free ... */ 3648c9de560dSAlex Tomas free += pa->pa_free; 3649c9de560dSAlex Tomas 3650c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3651c9de560dSAlex Tomas 3652c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3653c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3654c9de560dSAlex Tomas } 3655c9de560dSAlex Tomas 3656c9de560dSAlex Tomas /* if we still need more blocks and some PAs were used, try again */ 3657c9de560dSAlex Tomas if (free < needed && busy) { 3658c9de560dSAlex Tomas busy = 0; 3659c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3660c9de560dSAlex Tomas /* 3661c9de560dSAlex Tomas * Yield the CPU here so that we don't get soft lockup 3662c9de560dSAlex Tomas * in non preempt case. 3663c9de560dSAlex Tomas */ 3664c9de560dSAlex Tomas yield(); 3665c9de560dSAlex Tomas goto repeat; 3666c9de560dSAlex Tomas } 3667c9de560dSAlex Tomas 3668c9de560dSAlex Tomas /* found anything to free? */ 3669c9de560dSAlex Tomas if (list_empty(&list)) { 3670c9de560dSAlex Tomas BUG_ON(free != 0); 3671c9de560dSAlex Tomas goto out; 3672c9de560dSAlex Tomas } 3673c9de560dSAlex Tomas 3674c9de560dSAlex Tomas /* now free all selected PAs */ 3675c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3676c9de560dSAlex Tomas 3677c9de560dSAlex Tomas /* remove from object (inode or locality group) */ 3678c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3679c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3680c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3681c9de560dSAlex Tomas 3682cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 3683c83617dbSAneesh Kumar K.V ext4_mb_release_group_pa(&e4b, pa, ac); 3684c9de560dSAlex Tomas else 3685c83617dbSAneesh Kumar K.V ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); 3686c9de560dSAlex Tomas 3687c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 3688c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3689c9de560dSAlex Tomas } 3690c9de560dSAlex Tomas 3691c9de560dSAlex Tomas out: 3692c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3693c83617dbSAneesh Kumar K.V if (ac) 3694c83617dbSAneesh Kumar K.V kmem_cache_free(ext4_ac_cachep, ac); 3695c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 3696c9de560dSAlex Tomas put_bh(bitmap_bh); 3697c9de560dSAlex Tomas return free; 3698c9de560dSAlex Tomas } 3699c9de560dSAlex Tomas 3700c9de560dSAlex Tomas /* 3701c9de560dSAlex Tomas * releases all non-used preallocated blocks for given inode 3702c9de560dSAlex Tomas * 3703c9de560dSAlex Tomas * It's important to discard preallocations under i_data_sem 3704c9de560dSAlex Tomas * We don't want another block to be served from the prealloc 3705c9de560dSAlex Tomas * space when we are discarding the inode prealloc space. 3706c9de560dSAlex Tomas * 3707c9de560dSAlex Tomas * FIXME!! Make sure it is valid at all the call sites 3708c9de560dSAlex Tomas */ 3709c2ea3fdeSTheodore Ts'o void ext4_discard_preallocations(struct inode *inode) 3710c9de560dSAlex Tomas { 3711c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(inode); 3712c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 3713c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3714c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3715c83617dbSAneesh Kumar K.V struct ext4_allocation_context *ac; 3716c9de560dSAlex Tomas ext4_group_t group = 0; 3717c9de560dSAlex Tomas struct list_head list; 3718c9de560dSAlex Tomas struct ext4_buddy e4b; 3719c9de560dSAlex Tomas int err; 3720c9de560dSAlex Tomas 3721c2ea3fdeSTheodore Ts'o if (!S_ISREG(inode->i_mode)) { 3722c9de560dSAlex Tomas /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ 3723c9de560dSAlex Tomas return; 3724c9de560dSAlex Tomas } 3725c9de560dSAlex Tomas 37266ba495e9STheodore Ts'o mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); 37279bffad1eSTheodore Ts'o trace_ext4_discard_preallocations(inode); 3728c9de560dSAlex Tomas 3729c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3730c9de560dSAlex Tomas 3731c83617dbSAneesh Kumar K.V ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 37329bffad1eSTheodore Ts'o if (ac) { 37339bffad1eSTheodore Ts'o ac->ac_sb = sb; 37349bffad1eSTheodore Ts'o ac->ac_inode = inode; 37359bffad1eSTheodore Ts'o } 3736c9de560dSAlex Tomas repeat: 3737c9de560dSAlex Tomas /* first, collect all pa's in the inode */ 3738c9de560dSAlex Tomas spin_lock(&ei->i_prealloc_lock); 3739c9de560dSAlex Tomas while (!list_empty(&ei->i_prealloc_list)) { 3740c9de560dSAlex Tomas pa = list_entry(ei->i_prealloc_list.next, 3741c9de560dSAlex Tomas struct ext4_prealloc_space, pa_inode_list); 3742c9de560dSAlex Tomas BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); 3743c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3744c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3745c9de560dSAlex Tomas /* this shouldn't happen often - nobody should 3746c9de560dSAlex Tomas * use preallocation while we're discarding it */ 3747c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3748c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3749c9de560dSAlex Tomas printk(KERN_ERR "uh-oh! used pa while discarding\n"); 3750c9de560dSAlex Tomas WARN_ON(1); 3751c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 3752c9de560dSAlex Tomas goto repeat; 3753c9de560dSAlex Tomas 3754c9de560dSAlex Tomas } 3755c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 3756c9de560dSAlex Tomas pa->pa_deleted = 1; 3757c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3758c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3759c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3760c9de560dSAlex Tomas continue; 3761c9de560dSAlex Tomas } 3762c9de560dSAlex Tomas 3763c9de560dSAlex Tomas /* someone is deleting pa right now */ 3764c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3765c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3766c9de560dSAlex Tomas 3767c9de560dSAlex Tomas /* we have to wait here because pa_deleted 3768c9de560dSAlex Tomas * doesn't mean pa is already unlinked from 3769c9de560dSAlex Tomas * the list. as we might be called from 3770c9de560dSAlex Tomas * ->clear_inode() the inode will get freed 3771c9de560dSAlex Tomas * and concurrent thread which is unlinking 3772c9de560dSAlex Tomas * pa from inode's list may access already 3773c9de560dSAlex Tomas * freed memory, bad-bad-bad */ 3774c9de560dSAlex Tomas 3775c9de560dSAlex Tomas /* XXX: if this happens too often, we can 3776c9de560dSAlex Tomas * add a flag to force wait only in case 3777c9de560dSAlex Tomas * of ->clear_inode(), but not in case of 3778c9de560dSAlex Tomas * regular truncate */ 3779c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 3780c9de560dSAlex Tomas goto repeat; 3781c9de560dSAlex Tomas } 3782c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3783c9de560dSAlex Tomas 3784c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3785cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_INODE_PA); 3786c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); 3787c9de560dSAlex Tomas 3788c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3789ce89f46cSAneesh Kumar K.V if (err) { 3790ce89f46cSAneesh Kumar K.V ext4_error(sb, __func__, "Error in loading buddy " 3791a9df9a49STheodore Ts'o "information for %u", group); 3792ce89f46cSAneesh Kumar K.V continue; 3793ce89f46cSAneesh Kumar K.V } 3794c9de560dSAlex Tomas 3795574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 3796c9de560dSAlex Tomas if (bitmap_bh == NULL) { 3797ce89f46cSAneesh Kumar K.V ext4_error(sb, __func__, "Error in reading block " 3798a9df9a49STheodore Ts'o "bitmap for %u", group); 3799c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 3800ce89f46cSAneesh Kumar K.V continue; 3801c9de560dSAlex Tomas } 3802c9de560dSAlex Tomas 3803c9de560dSAlex Tomas ext4_lock_group(sb, group); 3804c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3805c83617dbSAneesh Kumar K.V ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); 3806c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3807c9de560dSAlex Tomas 3808c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 3809c9de560dSAlex Tomas put_bh(bitmap_bh); 3810c9de560dSAlex Tomas 3811c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 3812c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3813c9de560dSAlex Tomas } 3814c83617dbSAneesh Kumar K.V if (ac) 3815c83617dbSAneesh Kumar K.V kmem_cache_free(ext4_ac_cachep, ac); 3816c9de560dSAlex Tomas } 3817c9de560dSAlex Tomas 3818c9de560dSAlex Tomas /* 3819c9de560dSAlex Tomas * finds all preallocated spaces and return blocks being freed to them 3820c9de560dSAlex Tomas * if preallocated space becomes full (no block is used from the space) 3821c9de560dSAlex Tomas * then the function frees space in buddy 3822c9de560dSAlex Tomas * XXX: at the moment, truncate (which is the only way to free blocks) 3823c9de560dSAlex Tomas * discards all preallocations 3824c9de560dSAlex Tomas */ 3825c9de560dSAlex Tomas static void ext4_mb_return_to_preallocation(struct inode *inode, 3826c9de560dSAlex Tomas struct ext4_buddy *e4b, 3827c9de560dSAlex Tomas sector_t block, int count) 3828c9de560dSAlex Tomas { 3829c9de560dSAlex Tomas BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list)); 3830c9de560dSAlex Tomas } 38316ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 3832c9de560dSAlex Tomas static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 3833c9de560dSAlex Tomas { 3834c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 38358df9675fSTheodore Ts'o ext4_group_t ngroups, i; 3836c9de560dSAlex Tomas 3837c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: Can't allocate:" 3838c9de560dSAlex Tomas " Allocation context details:\n"); 3839c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: status %d flags %d\n", 3840c9de560dSAlex Tomas ac->ac_status, ac->ac_flags); 3841c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, " 3842c9de560dSAlex Tomas "best %lu/%lu/%lu@%lu cr %d\n", 3843c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_group, 3844c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_start, 3845c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_len, 3846c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_logical, 3847c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_group, 3848c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_start, 3849c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_len, 3850c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_logical, 3851c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_group, 3852c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_start, 3853c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_len, 3854c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_logical, 3855c9de560dSAlex Tomas (int)ac->ac_criteria); 3856c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned, 3857c9de560dSAlex Tomas ac->ac_found); 3858c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: groups: \n"); 38598df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 38608df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 3861c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, i); 3862c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3863c9de560dSAlex Tomas ext4_grpblk_t start; 3864c9de560dSAlex Tomas struct list_head *cur; 3865c9de560dSAlex Tomas ext4_lock_group(sb, i); 3866c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 3867c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, 3868c9de560dSAlex Tomas pa_group_list); 3869c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3870c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3871c9de560dSAlex Tomas NULL, &start); 3872c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 38731c718505SAkira Fujita printk(KERN_ERR "PA:%u:%d:%u \n", i, 3874c9de560dSAlex Tomas start, pa->pa_len); 3875c9de560dSAlex Tomas } 387660bd63d1SSolofo Ramangalahy ext4_unlock_group(sb, i); 3877c9de560dSAlex Tomas 3878c9de560dSAlex Tomas if (grp->bb_free == 0) 3879c9de560dSAlex Tomas continue; 38801c718505SAkira Fujita printk(KERN_ERR "%u: %d/%d \n", 3881c9de560dSAlex Tomas i, grp->bb_free, grp->bb_fragments); 3882c9de560dSAlex Tomas } 3883c9de560dSAlex Tomas printk(KERN_ERR "\n"); 3884c9de560dSAlex Tomas } 3885c9de560dSAlex Tomas #else 3886c9de560dSAlex Tomas static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 3887c9de560dSAlex Tomas { 3888c9de560dSAlex Tomas return; 3889c9de560dSAlex Tomas } 3890c9de560dSAlex Tomas #endif 3891c9de560dSAlex Tomas 3892c9de560dSAlex Tomas /* 3893c9de560dSAlex Tomas * We use locality group preallocation for small size file. The size of the 3894c9de560dSAlex Tomas * file is determined by the current size or the resulting size after 3895c9de560dSAlex Tomas * allocation which ever is larger 3896c9de560dSAlex Tomas * 3897b713a5ecSTheodore Ts'o * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 3898c9de560dSAlex Tomas */ 3899c9de560dSAlex Tomas static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 3900c9de560dSAlex Tomas { 3901c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3902c9de560dSAlex Tomas int bsbits = ac->ac_sb->s_blocksize_bits; 3903c9de560dSAlex Tomas loff_t size, isize; 3904c9de560dSAlex Tomas 3905c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3906c9de560dSAlex Tomas return; 3907c9de560dSAlex Tomas 39084ba74d00STheodore Ts'o if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 39094ba74d00STheodore Ts'o return; 39104ba74d00STheodore Ts'o 3911c9de560dSAlex Tomas size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; 391250797481STheodore Ts'o isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 391350797481STheodore Ts'o >> bsbits; 3914c9de560dSAlex Tomas 391550797481STheodore Ts'o if ((size == isize) && 391650797481STheodore Ts'o !ext4_fs_is_busy(sbi) && 391750797481STheodore Ts'o (atomic_read(&ac->ac_inode->i_writecount) == 0)) { 391850797481STheodore Ts'o ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 391950797481STheodore Ts'o return; 392050797481STheodore Ts'o } 392150797481STheodore Ts'o 3922c9de560dSAlex Tomas /* don't use group allocation for large files */ 392371780577STheodore Ts'o size = max(size, isize); 39244ba74d00STheodore Ts'o if (size >= sbi->s_mb_stream_request) { 39254ba74d00STheodore Ts'o ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 3926c9de560dSAlex Tomas return; 39274ba74d00STheodore Ts'o } 3928c9de560dSAlex Tomas 3929c9de560dSAlex Tomas BUG_ON(ac->ac_lg != NULL); 3930c9de560dSAlex Tomas /* 3931c9de560dSAlex Tomas * locality group prealloc space are per cpu. The reason for having 3932c9de560dSAlex Tomas * per cpu locality group is to reduce the contention between block 3933c9de560dSAlex Tomas * request from multiple CPUs. 3934c9de560dSAlex Tomas */ 3935730c213cSEric Sandeen ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id()); 3936c9de560dSAlex Tomas 3937c9de560dSAlex Tomas /* we're going to use group allocation */ 3938c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 3939c9de560dSAlex Tomas 3940c9de560dSAlex Tomas /* serialize all allocations in the group */ 3941c9de560dSAlex Tomas mutex_lock(&ac->ac_lg->lg_mutex); 3942c9de560dSAlex Tomas } 3943c9de560dSAlex Tomas 39444ddfef7bSEric Sandeen static noinline_for_stack int 39454ddfef7bSEric Sandeen ext4_mb_initialize_context(struct ext4_allocation_context *ac, 3946c9de560dSAlex Tomas struct ext4_allocation_request *ar) 3947c9de560dSAlex Tomas { 3948c9de560dSAlex Tomas struct super_block *sb = ar->inode->i_sb; 3949c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3950c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 3951c9de560dSAlex Tomas ext4_group_t group; 3952498e5f24STheodore Ts'o unsigned int len; 3953498e5f24STheodore Ts'o ext4_fsblk_t goal; 3954c9de560dSAlex Tomas ext4_grpblk_t block; 3955c9de560dSAlex Tomas 3956c9de560dSAlex Tomas /* we can't allocate > group size */ 3957c9de560dSAlex Tomas len = ar->len; 3958c9de560dSAlex Tomas 3959c9de560dSAlex Tomas /* just a dirty hack to filter too big requests */ 3960c9de560dSAlex Tomas if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10) 3961c9de560dSAlex Tomas len = EXT4_BLOCKS_PER_GROUP(sb) - 10; 3962c9de560dSAlex Tomas 3963c9de560dSAlex Tomas /* start searching from the goal */ 3964c9de560dSAlex Tomas goal = ar->goal; 3965c9de560dSAlex Tomas if (goal < le32_to_cpu(es->s_first_data_block) || 3966c9de560dSAlex Tomas goal >= ext4_blocks_count(es)) 3967c9de560dSAlex Tomas goal = le32_to_cpu(es->s_first_data_block); 3968c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, goal, &group, &block); 3969c9de560dSAlex Tomas 3970c9de560dSAlex Tomas /* set up allocation goals */ 3971833576b3STheodore Ts'o memset(ac, 0, sizeof(struct ext4_allocation_context)); 3972c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ar->logical; 3973c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 3974c9de560dSAlex Tomas ac->ac_sb = sb; 3975c9de560dSAlex Tomas ac->ac_inode = ar->inode; 3976c9de560dSAlex Tomas ac->ac_o_ex.fe_logical = ar->logical; 3977c9de560dSAlex Tomas ac->ac_o_ex.fe_group = group; 3978c9de560dSAlex Tomas ac->ac_o_ex.fe_start = block; 3979c9de560dSAlex Tomas ac->ac_o_ex.fe_len = len; 3980c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = ar->logical; 3981c9de560dSAlex Tomas ac->ac_g_ex.fe_group = group; 3982c9de560dSAlex Tomas ac->ac_g_ex.fe_start = block; 3983c9de560dSAlex Tomas ac->ac_g_ex.fe_len = len; 3984c9de560dSAlex Tomas ac->ac_flags = ar->flags; 3985c9de560dSAlex Tomas 3986c9de560dSAlex Tomas /* we have to define context: we'll we work with a file or 3987c9de560dSAlex Tomas * locality group. this is a policy, actually */ 3988c9de560dSAlex Tomas ext4_mb_group_or_file(ac); 3989c9de560dSAlex Tomas 39906ba495e9STheodore Ts'o mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " 3991c9de560dSAlex Tomas "left: %u/%u, right %u/%u to %swritable\n", 3992c9de560dSAlex Tomas (unsigned) ar->len, (unsigned) ar->logical, 3993c9de560dSAlex Tomas (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 3994c9de560dSAlex Tomas (unsigned) ar->lleft, (unsigned) ar->pleft, 3995c9de560dSAlex Tomas (unsigned) ar->lright, (unsigned) ar->pright, 3996c9de560dSAlex Tomas atomic_read(&ar->inode->i_writecount) ? "" : "non-"); 3997c9de560dSAlex Tomas return 0; 3998c9de560dSAlex Tomas 3999c9de560dSAlex Tomas } 4000c9de560dSAlex Tomas 40016be2ded1SAneesh Kumar K.V static noinline_for_stack void 40026be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(struct super_block *sb, 40036be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg, 40046be2ded1SAneesh Kumar K.V int order, int total_entries) 40056be2ded1SAneesh Kumar K.V { 40066be2ded1SAneesh Kumar K.V ext4_group_t group = 0; 40076be2ded1SAneesh Kumar K.V struct ext4_buddy e4b; 40086be2ded1SAneesh Kumar K.V struct list_head discard_list; 40096be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa, *tmp; 40106be2ded1SAneesh Kumar K.V struct ext4_allocation_context *ac; 40116be2ded1SAneesh Kumar K.V 40126ba495e9STheodore Ts'o mb_debug(1, "discard locality group preallocation\n"); 40136be2ded1SAneesh Kumar K.V 40146be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&discard_list); 40156be2ded1SAneesh Kumar K.V ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 40169bffad1eSTheodore Ts'o if (ac) 40179bffad1eSTheodore Ts'o ac->ac_sb = sb; 40186be2ded1SAneesh Kumar K.V 40196be2ded1SAneesh Kumar K.V spin_lock(&lg->lg_prealloc_lock); 40206be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 40216be2ded1SAneesh Kumar K.V pa_inode_list) { 40226be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 40236be2ded1SAneesh Kumar K.V if (atomic_read(&pa->pa_count)) { 40246be2ded1SAneesh Kumar K.V /* 40256be2ded1SAneesh Kumar K.V * This is the pa that we just used 40266be2ded1SAneesh Kumar K.V * for block allocation. So don't 40276be2ded1SAneesh Kumar K.V * free that 40286be2ded1SAneesh Kumar K.V */ 40296be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 40306be2ded1SAneesh Kumar K.V continue; 40316be2ded1SAneesh Kumar K.V } 40326be2ded1SAneesh Kumar K.V if (pa->pa_deleted) { 40336be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 40346be2ded1SAneesh Kumar K.V continue; 40356be2ded1SAneesh Kumar K.V } 40366be2ded1SAneesh Kumar K.V /* only lg prealloc space */ 4037cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_GROUP_PA); 40386be2ded1SAneesh Kumar K.V 40396be2ded1SAneesh Kumar K.V /* seems this one can be freed ... */ 40406be2ded1SAneesh Kumar K.V pa->pa_deleted = 1; 40416be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 40426be2ded1SAneesh Kumar K.V 40436be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 40446be2ded1SAneesh Kumar K.V list_add(&pa->u.pa_tmp_list, &discard_list); 40456be2ded1SAneesh Kumar K.V 40466be2ded1SAneesh Kumar K.V total_entries--; 40476be2ded1SAneesh Kumar K.V if (total_entries <= 5) { 40486be2ded1SAneesh Kumar K.V /* 40496be2ded1SAneesh Kumar K.V * we want to keep only 5 entries 40506be2ded1SAneesh Kumar K.V * allowing it to grow to 8. This 40516be2ded1SAneesh Kumar K.V * mak sure we don't call discard 40526be2ded1SAneesh Kumar K.V * soon for this list. 40536be2ded1SAneesh Kumar K.V */ 40546be2ded1SAneesh Kumar K.V break; 40556be2ded1SAneesh Kumar K.V } 40566be2ded1SAneesh Kumar K.V } 40576be2ded1SAneesh Kumar K.V spin_unlock(&lg->lg_prealloc_lock); 40586be2ded1SAneesh Kumar K.V 40596be2ded1SAneesh Kumar K.V list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 40606be2ded1SAneesh Kumar K.V 40616be2ded1SAneesh Kumar K.V ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); 40626be2ded1SAneesh Kumar K.V if (ext4_mb_load_buddy(sb, group, &e4b)) { 40636be2ded1SAneesh Kumar K.V ext4_error(sb, __func__, "Error in loading buddy " 4064a9df9a49STheodore Ts'o "information for %u", group); 40656be2ded1SAneesh Kumar K.V continue; 40666be2ded1SAneesh Kumar K.V } 40676be2ded1SAneesh Kumar K.V ext4_lock_group(sb, group); 40686be2ded1SAneesh Kumar K.V list_del(&pa->pa_group_list); 40696be2ded1SAneesh Kumar K.V ext4_mb_release_group_pa(&e4b, pa, ac); 40706be2ded1SAneesh Kumar K.V ext4_unlock_group(sb, group); 40716be2ded1SAneesh Kumar K.V 40726be2ded1SAneesh Kumar K.V ext4_mb_release_desc(&e4b); 40736be2ded1SAneesh Kumar K.V list_del(&pa->u.pa_tmp_list); 40746be2ded1SAneesh Kumar K.V call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 40756be2ded1SAneesh Kumar K.V } 40766be2ded1SAneesh Kumar K.V if (ac) 40776be2ded1SAneesh Kumar K.V kmem_cache_free(ext4_ac_cachep, ac); 40786be2ded1SAneesh Kumar K.V } 40796be2ded1SAneesh Kumar K.V 40806be2ded1SAneesh Kumar K.V /* 40816be2ded1SAneesh Kumar K.V * We have incremented pa_count. So it cannot be freed at this 40826be2ded1SAneesh Kumar K.V * point. Also we hold lg_mutex. So no parallel allocation is 40836be2ded1SAneesh Kumar K.V * possible from this lg. That means pa_free cannot be updated. 40846be2ded1SAneesh Kumar K.V * 40856be2ded1SAneesh Kumar K.V * A parallel ext4_mb_discard_group_preallocations is possible. 40866be2ded1SAneesh Kumar K.V * which can cause the lg_prealloc_list to be updated. 40876be2ded1SAneesh Kumar K.V */ 40886be2ded1SAneesh Kumar K.V 40896be2ded1SAneesh Kumar K.V static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 40906be2ded1SAneesh Kumar K.V { 40916be2ded1SAneesh Kumar K.V int order, added = 0, lg_prealloc_count = 1; 40926be2ded1SAneesh Kumar K.V struct super_block *sb = ac->ac_sb; 40936be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg = ac->ac_lg; 40946be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 40956be2ded1SAneesh Kumar K.V 40966be2ded1SAneesh Kumar K.V order = fls(pa->pa_free) - 1; 40976be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 40986be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 40996be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 41006be2ded1SAneesh Kumar K.V /* Add the prealloc space to lg */ 41016be2ded1SAneesh Kumar K.V rcu_read_lock(); 41026be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 41036be2ded1SAneesh Kumar K.V pa_inode_list) { 41046be2ded1SAneesh Kumar K.V spin_lock(&tmp_pa->pa_lock); 41056be2ded1SAneesh Kumar K.V if (tmp_pa->pa_deleted) { 4106e7c9e3e9STheodore Ts'o spin_unlock(&tmp_pa->pa_lock); 41076be2ded1SAneesh Kumar K.V continue; 41086be2ded1SAneesh Kumar K.V } 41096be2ded1SAneesh Kumar K.V if (!added && pa->pa_free < tmp_pa->pa_free) { 41106be2ded1SAneesh Kumar K.V /* Add to the tail of the previous entry */ 41116be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 41126be2ded1SAneesh Kumar K.V &tmp_pa->pa_inode_list); 41136be2ded1SAneesh Kumar K.V added = 1; 41146be2ded1SAneesh Kumar K.V /* 41156be2ded1SAneesh Kumar K.V * we want to count the total 41166be2ded1SAneesh Kumar K.V * number of entries in the list 41176be2ded1SAneesh Kumar K.V */ 41186be2ded1SAneesh Kumar K.V } 41196be2ded1SAneesh Kumar K.V spin_unlock(&tmp_pa->pa_lock); 41206be2ded1SAneesh Kumar K.V lg_prealloc_count++; 41216be2ded1SAneesh Kumar K.V } 41226be2ded1SAneesh Kumar K.V if (!added) 41236be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 41246be2ded1SAneesh Kumar K.V &lg->lg_prealloc_list[order]); 41256be2ded1SAneesh Kumar K.V rcu_read_unlock(); 41266be2ded1SAneesh Kumar K.V 41276be2ded1SAneesh Kumar K.V /* Now trim the list to be not more than 8 elements */ 41286be2ded1SAneesh Kumar K.V if (lg_prealloc_count > 8) { 41296be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(sb, lg, 41306be2ded1SAneesh Kumar K.V order, lg_prealloc_count); 41316be2ded1SAneesh Kumar K.V return; 41326be2ded1SAneesh Kumar K.V } 41336be2ded1SAneesh Kumar K.V return ; 41346be2ded1SAneesh Kumar K.V } 41356be2ded1SAneesh Kumar K.V 4136c9de560dSAlex Tomas /* 4137c9de560dSAlex Tomas * release all resource we used in allocation 4138c9de560dSAlex Tomas */ 4139c9de560dSAlex Tomas static int ext4_mb_release_context(struct ext4_allocation_context *ac) 4140c9de560dSAlex Tomas { 41416be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa = ac->ac_pa; 41426be2ded1SAneesh Kumar K.V if (pa) { 4143cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) { 4144c9de560dSAlex Tomas /* see comment in ext4_mb_use_group_pa() */ 41456be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 41466be2ded1SAneesh Kumar K.V pa->pa_pstart += ac->ac_b_ex.fe_len; 41476be2ded1SAneesh Kumar K.V pa->pa_lstart += ac->ac_b_ex.fe_len; 41486be2ded1SAneesh Kumar K.V pa->pa_free -= ac->ac_b_ex.fe_len; 41496be2ded1SAneesh Kumar K.V pa->pa_len -= ac->ac_b_ex.fe_len; 41506be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 4151ba443916SAneesh Kumar K.V } 4152ba443916SAneesh Kumar K.V } 4153ba443916SAneesh Kumar K.V if (ac->alloc_semp) 4154ba443916SAneesh Kumar K.V up_read(ac->alloc_semp); 4155ba443916SAneesh Kumar K.V if (pa) { 41566be2ded1SAneesh Kumar K.V /* 41576be2ded1SAneesh Kumar K.V * We want to add the pa to the right bucket. 41586be2ded1SAneesh Kumar K.V * Remove it from the list and while adding 41596be2ded1SAneesh Kumar K.V * make sure the list to which we are adding 4160ba443916SAneesh Kumar K.V * doesn't grow big. We need to release 4161ba443916SAneesh Kumar K.V * alloc_semp before calling ext4_mb_add_n_trim() 41626be2ded1SAneesh Kumar K.V */ 4163cc0fb9adSAneesh Kumar K.V if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { 41646be2ded1SAneesh Kumar K.V spin_lock(pa->pa_obj_lock); 41656be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 41666be2ded1SAneesh Kumar K.V spin_unlock(pa->pa_obj_lock); 41676be2ded1SAneesh Kumar K.V ext4_mb_add_n_trim(ac); 4168c9de560dSAlex Tomas } 41696be2ded1SAneesh Kumar K.V ext4_mb_put_pa(ac, ac->ac_sb, pa); 4170c9de560dSAlex Tomas } 4171c9de560dSAlex Tomas if (ac->ac_bitmap_page) 4172c9de560dSAlex Tomas page_cache_release(ac->ac_bitmap_page); 4173c9de560dSAlex Tomas if (ac->ac_buddy_page) 4174c9de560dSAlex Tomas page_cache_release(ac->ac_buddy_page); 4175c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 4176c9de560dSAlex Tomas mutex_unlock(&ac->ac_lg->lg_mutex); 4177c9de560dSAlex Tomas ext4_mb_collect_stats(ac); 4178c9de560dSAlex Tomas return 0; 4179c9de560dSAlex Tomas } 4180c9de560dSAlex Tomas 4181c9de560dSAlex Tomas static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 4182c9de560dSAlex Tomas { 41838df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 4184c9de560dSAlex Tomas int ret; 4185c9de560dSAlex Tomas int freed = 0; 4186c9de560dSAlex Tomas 41879bffad1eSTheodore Ts'o trace_ext4_mb_discard_preallocations(sb, needed); 41888df9675fSTheodore Ts'o for (i = 0; i < ngroups && needed > 0; i++) { 4189c9de560dSAlex Tomas ret = ext4_mb_discard_group_preallocations(sb, i, needed); 4190c9de560dSAlex Tomas freed += ret; 4191c9de560dSAlex Tomas needed -= ret; 4192c9de560dSAlex Tomas } 4193c9de560dSAlex Tomas 4194c9de560dSAlex Tomas return freed; 4195c9de560dSAlex Tomas } 4196c9de560dSAlex Tomas 4197c9de560dSAlex Tomas /* 4198c9de560dSAlex Tomas * Main entry point into mballoc to allocate blocks 4199c9de560dSAlex Tomas * it tries to use preallocation first, then falls back 4200c9de560dSAlex Tomas * to usual allocation 4201c9de560dSAlex Tomas */ 4202c9de560dSAlex Tomas ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 4203c9de560dSAlex Tomas struct ext4_allocation_request *ar, int *errp) 4204c9de560dSAlex Tomas { 42056bc6e63fSAneesh Kumar K.V int freed; 4206256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 4207c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4208c9de560dSAlex Tomas struct super_block *sb; 4209c9de560dSAlex Tomas ext4_fsblk_t block = 0; 421060e58e0fSMingming Cao unsigned int inquota = 0; 4211498e5f24STheodore Ts'o unsigned int reserv_blks = 0; 4212c9de560dSAlex Tomas 4213c9de560dSAlex Tomas sb = ar->inode->i_sb; 4214c9de560dSAlex Tomas sbi = EXT4_SB(sb); 4215c9de560dSAlex Tomas 42169bffad1eSTheodore Ts'o trace_ext4_request_blocks(ar); 4217ba80b101STheodore Ts'o 4218d2a17637SMingming Cao /* 421960e58e0fSMingming Cao * For delayed allocation, we could skip the ENOSPC and 422060e58e0fSMingming Cao * EDQUOT check, as blocks and quotas have been already 422160e58e0fSMingming Cao * reserved when data being copied into pagecache. 422260e58e0fSMingming Cao */ 422360e58e0fSMingming Cao if (EXT4_I(ar->inode)->i_delalloc_reserved_flag) 422460e58e0fSMingming Cao ar->flags |= EXT4_MB_DELALLOC_RESERVED; 422560e58e0fSMingming Cao else { 422660e58e0fSMingming Cao /* Without delayed allocation we need to verify 422760e58e0fSMingming Cao * there is enough free blocks to do block allocation 422860e58e0fSMingming Cao * and verify allocation doesn't exceed the quota limits. 4229d2a17637SMingming Cao */ 4230030ba6bcSAneesh Kumar K.V while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) { 4231030ba6bcSAneesh Kumar K.V /* let others to free the space */ 4232030ba6bcSAneesh Kumar K.V yield(); 4233030ba6bcSAneesh Kumar K.V ar->len = ar->len >> 1; 4234030ba6bcSAneesh Kumar K.V } 4235030ba6bcSAneesh Kumar K.V if (!ar->len) { 423607031431SMingming Cao *errp = -ENOSPC; 423707031431SMingming Cao return 0; 423807031431SMingming Cao } 42396bc6e63fSAneesh Kumar K.V reserv_blks = ar->len; 4240a269eb18SJan Kara while (ar->len && vfs_dq_alloc_block(ar->inode, ar->len)) { 4241c9de560dSAlex Tomas ar->flags |= EXT4_MB_HINT_NOPREALLOC; 4242c9de560dSAlex Tomas ar->len--; 4243c9de560dSAlex Tomas } 424460e58e0fSMingming Cao inquota = ar->len; 4245c9de560dSAlex Tomas if (ar->len == 0) { 4246c9de560dSAlex Tomas *errp = -EDQUOT; 42470087d9fbSAneesh Kumar K.V goto out3; 4248c9de560dSAlex Tomas } 424960e58e0fSMingming Cao } 4250d2a17637SMingming Cao 4251256bdb49SEric Sandeen ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 4252833576b3STheodore Ts'o if (!ac) { 4253363d4251SShen Feng ar->len = 0; 4254256bdb49SEric Sandeen *errp = -ENOMEM; 4255363d4251SShen Feng goto out1; 4256256bdb49SEric Sandeen } 4257256bdb49SEric Sandeen 4258256bdb49SEric Sandeen *errp = ext4_mb_initialize_context(ac, ar); 4259c9de560dSAlex Tomas if (*errp) { 4260c9de560dSAlex Tomas ar->len = 0; 4261363d4251SShen Feng goto out2; 4262c9de560dSAlex Tomas } 4263c9de560dSAlex Tomas 4264256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 4265256bdb49SEric Sandeen if (!ext4_mb_use_preallocated(ac)) { 4266256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_ALLOC; 4267256bdb49SEric Sandeen ext4_mb_normalize_request(ac, ar); 4268c9de560dSAlex Tomas repeat: 4269c9de560dSAlex Tomas /* allocate space in core */ 4270256bdb49SEric Sandeen ext4_mb_regular_allocator(ac); 4271c9de560dSAlex Tomas 4272c9de560dSAlex Tomas /* as we've just preallocated more space than 4273c9de560dSAlex Tomas * user requested orinally, we store allocated 4274c9de560dSAlex Tomas * space in a special descriptor */ 4275256bdb49SEric Sandeen if (ac->ac_status == AC_STATUS_FOUND && 4276256bdb49SEric Sandeen ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 4277256bdb49SEric Sandeen ext4_mb_new_preallocation(ac); 4278c9de560dSAlex Tomas } 4279256bdb49SEric Sandeen if (likely(ac->ac_status == AC_STATUS_FOUND)) { 42806bc6e63fSAneesh Kumar K.V *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks); 4281519deca0SAneesh Kumar K.V if (*errp == -EAGAIN) { 42828556e8f3SAneesh Kumar K.V /* 42838556e8f3SAneesh Kumar K.V * drop the reference that we took 42848556e8f3SAneesh Kumar K.V * in ext4_mb_use_best_found 42858556e8f3SAneesh Kumar K.V */ 42868556e8f3SAneesh Kumar K.V ext4_mb_release_context(ac); 4287519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_group = 0; 4288519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_start = 0; 4289519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len = 0; 4290519deca0SAneesh Kumar K.V ac->ac_status = AC_STATUS_CONTINUE; 4291519deca0SAneesh Kumar K.V goto repeat; 4292519deca0SAneesh Kumar K.V } else if (*errp) { 4293519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len = 0; 4294519deca0SAneesh Kumar K.V ar->len = 0; 4295519deca0SAneesh Kumar K.V ext4_mb_show_ac(ac); 4296519deca0SAneesh Kumar K.V } else { 4297256bdb49SEric Sandeen block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 4298256bdb49SEric Sandeen ar->len = ac->ac_b_ex.fe_len; 4299519deca0SAneesh Kumar K.V } 4300c9de560dSAlex Tomas } else { 4301256bdb49SEric Sandeen freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 4302c9de560dSAlex Tomas if (freed) 4303c9de560dSAlex Tomas goto repeat; 4304c9de560dSAlex Tomas *errp = -ENOSPC; 4305256bdb49SEric Sandeen ac->ac_b_ex.fe_len = 0; 4306c9de560dSAlex Tomas ar->len = 0; 4307256bdb49SEric Sandeen ext4_mb_show_ac(ac); 4308c9de560dSAlex Tomas } 4309c9de560dSAlex Tomas 4310256bdb49SEric Sandeen ext4_mb_release_context(ac); 4311c9de560dSAlex Tomas 4312363d4251SShen Feng out2: 4313363d4251SShen Feng kmem_cache_free(ext4_ac_cachep, ac); 4314363d4251SShen Feng out1: 431560e58e0fSMingming Cao if (inquota && ar->len < inquota) 4316a269eb18SJan Kara vfs_dq_free_block(ar->inode, inquota - ar->len); 43170087d9fbSAneesh Kumar K.V out3: 43180087d9fbSAneesh Kumar K.V if (!ar->len) { 43190087d9fbSAneesh Kumar K.V if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) 43200087d9fbSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 43210087d9fbSAneesh Kumar K.V percpu_counter_sub(&sbi->s_dirtyblocks_counter, 43220087d9fbSAneesh Kumar K.V reserv_blks); 43230087d9fbSAneesh Kumar K.V } 4324c9de560dSAlex Tomas 43259bffad1eSTheodore Ts'o trace_ext4_allocate_blocks(ar, (unsigned long long)block); 4326ba80b101STheodore Ts'o 4327c9de560dSAlex Tomas return block; 4328c9de560dSAlex Tomas } 4329c9de560dSAlex Tomas 4330c894058dSAneesh Kumar K.V /* 4331c894058dSAneesh Kumar K.V * We can merge two free data extents only if the physical blocks 4332c894058dSAneesh Kumar K.V * are contiguous, AND the extents were freed by the same transaction, 4333c894058dSAneesh Kumar K.V * AND the blocks are associated with the same group. 4334c894058dSAneesh Kumar K.V */ 4335c894058dSAneesh Kumar K.V static int can_merge(struct ext4_free_data *entry1, 4336c894058dSAneesh Kumar K.V struct ext4_free_data *entry2) 4337c894058dSAneesh Kumar K.V { 4338c894058dSAneesh Kumar K.V if ((entry1->t_tid == entry2->t_tid) && 4339c894058dSAneesh Kumar K.V (entry1->group == entry2->group) && 4340c894058dSAneesh Kumar K.V ((entry1->start_blk + entry1->count) == entry2->start_blk)) 4341c894058dSAneesh Kumar K.V return 1; 4342c894058dSAneesh Kumar K.V return 0; 4343c894058dSAneesh Kumar K.V } 4344c894058dSAneesh Kumar K.V 43454ddfef7bSEric Sandeen static noinline_for_stack int 43464ddfef7bSEric Sandeen ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 43477a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry) 4348c9de560dSAlex Tomas { 43497a2fcbf7SAneesh Kumar K.V ext4_grpblk_t block; 43507a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 4351c9de560dSAlex Tomas struct ext4_group_info *db = e4b->bd_info; 4352c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 4353c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 4354c894058dSAneesh Kumar K.V struct rb_node **n = &db->bb_free_root.rb_node, *node; 4355c894058dSAneesh Kumar K.V struct rb_node *parent = NULL, *new_node; 4356c894058dSAneesh Kumar K.V 43570390131bSFrank Mayhar BUG_ON(!ext4_handle_valid(handle)); 4358c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 4359c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 4360c9de560dSAlex Tomas 4361c894058dSAneesh Kumar K.V new_node = &new_entry->node; 43627a2fcbf7SAneesh Kumar K.V block = new_entry->start_blk; 4363c9de560dSAlex Tomas 4364c894058dSAneesh Kumar K.V if (!*n) { 4365c894058dSAneesh Kumar K.V /* first free block exent. We need to 4366c894058dSAneesh Kumar K.V protect buddy cache from being freed, 4367c9de560dSAlex Tomas * otherwise we'll refresh it from 4368c9de560dSAlex Tomas * on-disk bitmap and lose not-yet-available 4369c9de560dSAlex Tomas * blocks */ 4370c9de560dSAlex Tomas page_cache_get(e4b->bd_buddy_page); 4371c9de560dSAlex Tomas page_cache_get(e4b->bd_bitmap_page); 4372c894058dSAneesh Kumar K.V } 4373c894058dSAneesh Kumar K.V while (*n) { 4374c894058dSAneesh Kumar K.V parent = *n; 4375c894058dSAneesh Kumar K.V entry = rb_entry(parent, struct ext4_free_data, node); 4376c894058dSAneesh Kumar K.V if (block < entry->start_blk) 4377c894058dSAneesh Kumar K.V n = &(*n)->rb_left; 4378c894058dSAneesh Kumar K.V else if (block >= (entry->start_blk + entry->count)) 4379c894058dSAneesh Kumar K.V n = &(*n)->rb_right; 4380c894058dSAneesh Kumar K.V else { 43815d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, __func__, 4382fde4d95aSTheodore Ts'o "Double free of blocks %d (%d %d)", 4383c894058dSAneesh Kumar K.V block, entry->start_blk, entry->count); 4384c894058dSAneesh Kumar K.V return 0; 4385c9de560dSAlex Tomas } 4386c9de560dSAlex Tomas } 4387c9de560dSAlex Tomas 4388c894058dSAneesh Kumar K.V rb_link_node(new_node, parent, n); 4389c894058dSAneesh Kumar K.V rb_insert_color(new_node, &db->bb_free_root); 4390c894058dSAneesh Kumar K.V 4391c894058dSAneesh Kumar K.V /* Now try to see the extent can be merged to left and right */ 4392c894058dSAneesh Kumar K.V node = rb_prev(new_node); 4393c894058dSAneesh Kumar K.V if (node) { 4394c894058dSAneesh Kumar K.V entry = rb_entry(node, struct ext4_free_data, node); 4395c894058dSAneesh Kumar K.V if (can_merge(entry, new_entry)) { 4396c894058dSAneesh Kumar K.V new_entry->start_blk = entry->start_blk; 4397c894058dSAneesh Kumar K.V new_entry->count += entry->count; 4398c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 4399c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 4400c894058dSAneesh Kumar K.V list_del(&entry->list); 4401c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4402c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 4403c9de560dSAlex Tomas } 4404c9de560dSAlex Tomas } 4405c894058dSAneesh Kumar K.V 4406c894058dSAneesh Kumar K.V node = rb_next(new_node); 4407c894058dSAneesh Kumar K.V if (node) { 4408c894058dSAneesh Kumar K.V entry = rb_entry(node, struct ext4_free_data, node); 4409c894058dSAneesh Kumar K.V if (can_merge(new_entry, entry)) { 4410c894058dSAneesh Kumar K.V new_entry->count += entry->count; 4411c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 4412c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 4413c894058dSAneesh Kumar K.V list_del(&entry->list); 4414c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4415c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 4416c894058dSAneesh Kumar K.V } 4417c894058dSAneesh Kumar K.V } 44183e624fc7STheodore Ts'o /* Add the extent to transaction's private list */ 4419c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 44203e624fc7STheodore Ts'o list_add(&new_entry->list, &handle->h_transaction->t_private_list); 4421c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4422c9de560dSAlex Tomas return 0; 4423c9de560dSAlex Tomas } 4424c9de560dSAlex Tomas 4425c9de560dSAlex Tomas /* 4426c9de560dSAlex Tomas * Main entry point into mballoc to free blocks 4427c9de560dSAlex Tomas */ 4428c9de560dSAlex Tomas void ext4_mb_free_blocks(handle_t *handle, struct inode *inode, 44290610b6e9STheodore Ts'o ext4_fsblk_t block, unsigned long count, 4430c9de560dSAlex Tomas int metadata, unsigned long *freed) 4431c9de560dSAlex Tomas { 443226346ff6SAneesh Kumar K.V struct buffer_head *bitmap_bh = NULL; 4433c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 4434256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 4435c9de560dSAlex Tomas struct ext4_group_desc *gdp; 4436c9de560dSAlex Tomas struct ext4_super_block *es; 4437498e5f24STheodore Ts'o unsigned int overflow; 4438c9de560dSAlex Tomas ext4_grpblk_t bit; 4439c9de560dSAlex Tomas struct buffer_head *gd_bh; 4440c9de560dSAlex Tomas ext4_group_t block_group; 4441c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4442c9de560dSAlex Tomas struct ext4_buddy e4b; 4443c9de560dSAlex Tomas int err = 0; 4444c9de560dSAlex Tomas int ret; 4445c9de560dSAlex Tomas 4446c9de560dSAlex Tomas *freed = 0; 4447c9de560dSAlex Tomas 4448c9de560dSAlex Tomas sbi = EXT4_SB(sb); 4449c9de560dSAlex Tomas es = EXT4_SB(sb)->s_es; 4450c9de560dSAlex Tomas if (block < le32_to_cpu(es->s_first_data_block) || 4451c9de560dSAlex Tomas block + count < block || 4452c9de560dSAlex Tomas block + count > ext4_blocks_count(es)) { 445346e665e9SHarvey Harrison ext4_error(sb, __func__, 4454c9de560dSAlex Tomas "Freeing blocks not in datazone - " 44550610b6e9STheodore Ts'o "block = %llu, count = %lu", block, count); 4456c9de560dSAlex Tomas goto error_return; 4457c9de560dSAlex Tomas } 4458c9de560dSAlex Tomas 44590610b6e9STheodore Ts'o ext4_debug("freeing block %llu\n", block); 44609bffad1eSTheodore Ts'o trace_ext4_free_blocks(inode, block, count, metadata); 4461c9de560dSAlex Tomas 4462256bdb49SEric Sandeen ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 4463256bdb49SEric Sandeen if (ac) { 4464256bdb49SEric Sandeen ac->ac_inode = inode; 4465256bdb49SEric Sandeen ac->ac_sb = sb; 4466256bdb49SEric Sandeen } 4467c9de560dSAlex Tomas 4468c9de560dSAlex Tomas do_more: 4469c9de560dSAlex Tomas overflow = 0; 4470c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 4471c9de560dSAlex Tomas 4472c9de560dSAlex Tomas /* 4473c9de560dSAlex Tomas * Check to see if we are freeing blocks across a group 4474c9de560dSAlex Tomas * boundary. 4475c9de560dSAlex Tomas */ 4476c9de560dSAlex Tomas if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { 4477c9de560dSAlex Tomas overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb); 4478c9de560dSAlex Tomas count -= overflow; 4479c9de560dSAlex Tomas } 4480574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, block_group); 4481ce89f46cSAneesh Kumar K.V if (!bitmap_bh) { 4482ce89f46cSAneesh Kumar K.V err = -EIO; 4483c9de560dSAlex Tomas goto error_return; 4484ce89f46cSAneesh Kumar K.V } 4485c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 4486ce89f46cSAneesh Kumar K.V if (!gdp) { 4487ce89f46cSAneesh Kumar K.V err = -EIO; 4488c9de560dSAlex Tomas goto error_return; 4489ce89f46cSAneesh Kumar K.V } 4490c9de560dSAlex Tomas 4491c9de560dSAlex Tomas if (in_range(ext4_block_bitmap(sb, gdp), block, count) || 4492c9de560dSAlex Tomas in_range(ext4_inode_bitmap(sb, gdp), block, count) || 4493c9de560dSAlex Tomas in_range(block, ext4_inode_table(sb, gdp), 4494c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group) || 4495c9de560dSAlex Tomas in_range(block + count - 1, ext4_inode_table(sb, gdp), 4496c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group)) { 4497c9de560dSAlex Tomas 449846e665e9SHarvey Harrison ext4_error(sb, __func__, 4499c9de560dSAlex Tomas "Freeing blocks in system zone - " 45000610b6e9STheodore Ts'o "Block = %llu, count = %lu", block, count); 4501519deca0SAneesh Kumar K.V /* err = 0. ext4_std_error should be a no op */ 4502519deca0SAneesh Kumar K.V goto error_return; 4503c9de560dSAlex Tomas } 4504c9de560dSAlex Tomas 4505c9de560dSAlex Tomas BUFFER_TRACE(bitmap_bh, "getting write access"); 4506c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 4507c9de560dSAlex Tomas if (err) 4508c9de560dSAlex Tomas goto error_return; 4509c9de560dSAlex Tomas 4510c9de560dSAlex Tomas /* 4511c9de560dSAlex Tomas * We are about to modify some metadata. Call the journal APIs 4512c9de560dSAlex Tomas * to unshare ->b_data if a currently-committing transaction is 4513c9de560dSAlex Tomas * using it 4514c9de560dSAlex Tomas */ 4515c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "get_write_access"); 4516c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gd_bh); 4517c9de560dSAlex Tomas if (err) 4518c9de560dSAlex Tomas goto error_return; 4519c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 4520c9de560dSAlex Tomas { 4521c9de560dSAlex Tomas int i; 4522c9de560dSAlex Tomas for (i = 0; i < count; i++) 4523c9de560dSAlex Tomas BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 4524c9de560dSAlex Tomas } 4525c9de560dSAlex Tomas #endif 4526256bdb49SEric Sandeen if (ac) { 4527256bdb49SEric Sandeen ac->ac_b_ex.fe_group = block_group; 4528256bdb49SEric Sandeen ac->ac_b_ex.fe_start = bit; 4529256bdb49SEric Sandeen ac->ac_b_ex.fe_len = count; 4530*296c355cSTheodore Ts'o trace_ext4_mballoc_free(ac); 4531256bdb49SEric Sandeen } 4532c9de560dSAlex Tomas 4533920313a7SAneesh Kumar K.V err = ext4_mb_load_buddy(sb, block_group, &e4b); 4534920313a7SAneesh Kumar K.V if (err) 4535920313a7SAneesh Kumar K.V goto error_return; 45360390131bSFrank Mayhar if (metadata && ext4_handle_valid(handle)) { 45377a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry; 45387a2fcbf7SAneesh Kumar K.V /* 45397a2fcbf7SAneesh Kumar K.V * blocks being freed are metadata. these blocks shouldn't 45407a2fcbf7SAneesh Kumar K.V * be used until this transaction is committed 45417a2fcbf7SAneesh Kumar K.V */ 45427a2fcbf7SAneesh Kumar K.V new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); 45437a2fcbf7SAneesh Kumar K.V new_entry->start_blk = bit; 45447a2fcbf7SAneesh Kumar K.V new_entry->group = block_group; 45457a2fcbf7SAneesh Kumar K.V new_entry->count = count; 45467a2fcbf7SAneesh Kumar K.V new_entry->t_tid = handle->h_transaction->t_tid; 4547955ce5f5SAneesh Kumar K.V 45487a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, block_group); 4549955ce5f5SAneesh Kumar K.V mb_clear_bits(bitmap_bh->b_data, bit, count); 45507a2fcbf7SAneesh Kumar K.V ext4_mb_free_metadata(handle, &e4b, new_entry); 4551c9de560dSAlex Tomas } else { 45527a2fcbf7SAneesh Kumar K.V /* need to update group_info->bb_free and bitmap 45537a2fcbf7SAneesh Kumar K.V * with group lock held. generate_buddy look at 45547a2fcbf7SAneesh Kumar K.V * them with group lock_held 45557a2fcbf7SAneesh Kumar K.V */ 4556955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 4557955ce5f5SAneesh Kumar K.V mb_clear_bits(bitmap_bh->b_data, bit, count); 45587e5a8cddSShen Feng mb_free_blocks(inode, &e4b, bit, count); 4559c9de560dSAlex Tomas ext4_mb_return_to_preallocation(inode, &e4b, block, count); 4560c9de560dSAlex Tomas } 4561c9de560dSAlex Tomas 4562560671a0SAneesh Kumar K.V ret = ext4_free_blks_count(sb, gdp) + count; 4563560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, ret); 4564c9de560dSAlex Tomas gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); 4565955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 4566c9de560dSAlex Tomas percpu_counter_add(&sbi->s_freeblocks_counter, count); 4567c9de560dSAlex Tomas 4568772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 4569772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 45709f24e420STheodore Ts'o atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks); 4571772cb7c8SJose R. Santos } 4572772cb7c8SJose R. Santos 4573c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 4574c9de560dSAlex Tomas 4575c9de560dSAlex Tomas *freed += count; 4576c9de560dSAlex Tomas 45777a2fcbf7SAneesh Kumar K.V /* We dirtied the bitmap block */ 45787a2fcbf7SAneesh Kumar K.V BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 45797a2fcbf7SAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 45807a2fcbf7SAneesh Kumar K.V 4581c9de560dSAlex Tomas /* And the group descriptor block */ 4582c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 45830390131bSFrank Mayhar ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 4584c9de560dSAlex Tomas if (!err) 4585c9de560dSAlex Tomas err = ret; 4586c9de560dSAlex Tomas 4587c9de560dSAlex Tomas if (overflow && !err) { 4588c9de560dSAlex Tomas block += count; 4589c9de560dSAlex Tomas count = overflow; 4590c9de560dSAlex Tomas put_bh(bitmap_bh); 4591c9de560dSAlex Tomas goto do_more; 4592c9de560dSAlex Tomas } 4593c9de560dSAlex Tomas sb->s_dirt = 1; 4594c9de560dSAlex Tomas error_return: 4595c9de560dSAlex Tomas brelse(bitmap_bh); 4596c9de560dSAlex Tomas ext4_std_error(sb, err); 4597256bdb49SEric Sandeen if (ac) 4598256bdb49SEric Sandeen kmem_cache_free(ext4_ac_cachep, ac); 4599c9de560dSAlex Tomas return; 4600c9de560dSAlex Tomas } 4601