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 * 14b8a07463SAdam Buchbinder * You should have received a copy of the GNU General Public License 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 2418aadd47SBobi Jam #include "ext4_jbd2.h" 258f6e39a7SMingming Cao #include "mballoc.h" 2628623c2fSTheodore Ts'o #include <linux/log2.h> 27a0b30c12STheodore Ts'o #include <linux/module.h> 285a0e3ad6STejun Heo #include <linux/slab.h> 2966114cadSTejun Heo #include <linux/backing-dev.h> 309bffad1eSTheodore Ts'o #include <trace/events/ext4.h> 319bffad1eSTheodore Ts'o 32a0b30c12STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 33a0b30c12STheodore Ts'o ushort ext4_mballoc_debug __read_mostly; 34a0b30c12STheodore Ts'o 35a0b30c12STheodore Ts'o module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644); 36a0b30c12STheodore Ts'o MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc"); 37a0b30c12STheodore Ts'o #endif 38a0b30c12STheodore Ts'o 39c9de560dSAlex Tomas /* 40c9de560dSAlex Tomas * MUSTDO: 41c9de560dSAlex Tomas * - test ext4_ext_search_left() and ext4_ext_search_right() 42c9de560dSAlex Tomas * - search for metadata in few groups 43c9de560dSAlex Tomas * 44c9de560dSAlex Tomas * TODO v4: 45c9de560dSAlex Tomas * - normalization should take into account whether file is still open 46c9de560dSAlex Tomas * - discard preallocations if no free space left (policy?) 47c9de560dSAlex Tomas * - don't normalize tails 48c9de560dSAlex Tomas * - quota 49c9de560dSAlex Tomas * - reservation for superuser 50c9de560dSAlex Tomas * 51c9de560dSAlex Tomas * TODO v3: 52c9de560dSAlex Tomas * - bitmap read-ahead (proposed by Oleg Drokin aka green) 53c9de560dSAlex Tomas * - track min/max extents in each group for better group selection 54c9de560dSAlex Tomas * - mb_mark_used() may allocate chunk right after splitting buddy 55c9de560dSAlex Tomas * - tree of groups sorted by number of free blocks 56c9de560dSAlex Tomas * - error handling 57c9de560dSAlex Tomas */ 58c9de560dSAlex Tomas 59c9de560dSAlex Tomas /* 60c9de560dSAlex Tomas * The allocation request involve request for multiple number of blocks 61c9de560dSAlex Tomas * near to the goal(block) value specified. 62c9de560dSAlex Tomas * 63b713a5ecSTheodore Ts'o * During initialization phase of the allocator we decide to use the 64b713a5ecSTheodore Ts'o * group preallocation or inode preallocation depending on the size of 65b713a5ecSTheodore Ts'o * the file. The size of the file could be the resulting file size we 66b713a5ecSTheodore Ts'o * would have after allocation, or the current file size, which ever 67b713a5ecSTheodore Ts'o * is larger. If the size is less than sbi->s_mb_stream_request we 68b713a5ecSTheodore Ts'o * select to use the group preallocation. The default value of 69b713a5ecSTheodore Ts'o * s_mb_stream_request is 16 blocks. This can also be tuned via 70b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 71b713a5ecSTheodore Ts'o * terms of number of blocks. 72c9de560dSAlex Tomas * 73c9de560dSAlex Tomas * The main motivation for having small file use group preallocation is to 74b713a5ecSTheodore Ts'o * ensure that we have small files closer together on the disk. 75c9de560dSAlex Tomas * 76b713a5ecSTheodore Ts'o * First stage the allocator looks at the inode prealloc list, 77b713a5ecSTheodore Ts'o * ext4_inode_info->i_prealloc_list, which contains list of prealloc 78b713a5ecSTheodore Ts'o * spaces for this particular inode. The inode prealloc space is 79b713a5ecSTheodore Ts'o * represented as: 80c9de560dSAlex Tomas * 81c9de560dSAlex Tomas * pa_lstart -> the logical start block for this prealloc space 82c9de560dSAlex Tomas * pa_pstart -> the physical start block for this prealloc space 8353accfa9STheodore Ts'o * pa_len -> length for this prealloc space (in clusters) 8453accfa9STheodore Ts'o * pa_free -> free space available in this prealloc space (in clusters) 85c9de560dSAlex Tomas * 86c9de560dSAlex Tomas * The inode preallocation space is used looking at the _logical_ start 87c9de560dSAlex Tomas * block. If only the logical file block falls within the range of prealloc 88caaf7a29STao Ma * space we will consume the particular prealloc space. This makes sure that 89caaf7a29STao Ma * we have contiguous physical blocks representing the file blocks 90c9de560dSAlex Tomas * 91c9de560dSAlex Tomas * The important thing to be noted in case of inode prealloc space is that 92c9de560dSAlex Tomas * we don't modify the values associated to inode prealloc space except 93c9de560dSAlex Tomas * pa_free. 94c9de560dSAlex Tomas * 95c9de560dSAlex Tomas * If we are not able to find blocks in the inode prealloc space and if we 96c9de560dSAlex Tomas * have the group allocation flag set then we look at the locality group 97caaf7a29STao Ma * prealloc space. These are per CPU prealloc list represented as 98c9de560dSAlex Tomas * 99c9de560dSAlex Tomas * ext4_sb_info.s_locality_groups[smp_processor_id()] 100c9de560dSAlex Tomas * 101c9de560dSAlex Tomas * The reason for having a per cpu locality group is to reduce the contention 102c9de560dSAlex Tomas * between CPUs. It is possible to get scheduled at this point. 103c9de560dSAlex Tomas * 104c9de560dSAlex Tomas * The locality group prealloc space is used looking at whether we have 10525985edcSLucas De Marchi * enough free space (pa_free) within the prealloc space. 106c9de560dSAlex Tomas * 107c9de560dSAlex Tomas * If we can't allocate blocks via inode prealloc or/and locality group 108c9de560dSAlex Tomas * prealloc then we look at the buddy cache. The buddy cache is represented 109c9de560dSAlex Tomas * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 110c9de560dSAlex Tomas * mapped to the buddy and bitmap information regarding different 111c9de560dSAlex Tomas * groups. The buddy information is attached to buddy cache inode so that 112c9de560dSAlex Tomas * we can access them through the page cache. The information regarding 113c9de560dSAlex Tomas * each group is loaded via ext4_mb_load_buddy. The information involve 114c9de560dSAlex Tomas * block bitmap and buddy information. The information are stored in the 115c9de560dSAlex Tomas * inode as: 116c9de560dSAlex Tomas * 117c9de560dSAlex Tomas * { page } 118c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 119c9de560dSAlex Tomas * 120c9de560dSAlex Tomas * 121c9de560dSAlex Tomas * one block each for bitmap and buddy information. So for each group we 122ea1754a0SKirill A. Shutemov * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE / 123c9de560dSAlex Tomas * blocksize) blocks. So it can have information regarding groups_per_page 124c9de560dSAlex Tomas * which is blocks_per_page/2 125c9de560dSAlex Tomas * 126c9de560dSAlex Tomas * The buddy cache inode is not stored on disk. The inode is thrown 127c9de560dSAlex Tomas * away when the filesystem is unmounted. 128c9de560dSAlex Tomas * 129c9de560dSAlex Tomas * We look for count number of blocks in the buddy cache. If we were able 130c9de560dSAlex Tomas * to locate that many free blocks we return with additional information 131c9de560dSAlex Tomas * regarding rest of the contiguous physical block available 132c9de560dSAlex Tomas * 133c9de560dSAlex Tomas * Before allocating blocks via buddy cache we normalize the request 134c9de560dSAlex Tomas * blocks. This ensure we ask for more blocks that we needed. The extra 135c9de560dSAlex Tomas * blocks that we get after allocation is added to the respective prealloc 136c9de560dSAlex Tomas * list. In case of inode preallocation we follow a list of heuristics 137c9de560dSAlex Tomas * based on file size. This can be found in ext4_mb_normalize_request. If 138c9de560dSAlex Tomas * we are doing a group prealloc we try to normalize the request to 13927baebb8STheodore Ts'o * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is 14027baebb8STheodore Ts'o * dependent on the cluster size; for non-bigalloc file systems, it is 141c9de560dSAlex Tomas * 512 blocks. This can be tuned via 142d7a1fee1SDan Ehrenberg * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in 143c9de560dSAlex Tomas * terms of number of blocks. If we have mounted the file system with -O 144c9de560dSAlex Tomas * stripe=<value> option the group prealloc request is normalized to the 145d7a1fee1SDan Ehrenberg * the smallest multiple of the stripe value (sbi->s_stripe) which is 146d7a1fee1SDan Ehrenberg * greater than the default mb_group_prealloc. 147c9de560dSAlex Tomas * 148d7a1fee1SDan Ehrenberg * The regular allocator (using the buddy cache) supports a few tunables. 149c9de560dSAlex Tomas * 150b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_min_to_scan 151b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_max_to_scan 152b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req 153c9de560dSAlex Tomas * 154b713a5ecSTheodore Ts'o * The regular allocator uses buddy scan only if the request len is power of 155c9de560dSAlex Tomas * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 156c9de560dSAlex Tomas * value of s_mb_order2_reqs can be tuned via 157b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 158af901ca1SAndré Goddard Rosa * stripe size (sbi->s_stripe), we try to search for contiguous block in 159b713a5ecSTheodore Ts'o * stripe size. This should result in better allocation on RAID setups. If 160b713a5ecSTheodore Ts'o * not, we search in the specific group using bitmap for best extents. The 161b713a5ecSTheodore Ts'o * tunable min_to_scan and max_to_scan control the behaviour here. 162c9de560dSAlex Tomas * min_to_scan indicate how long the mballoc __must__ look for a best 163b713a5ecSTheodore Ts'o * extent and max_to_scan indicates how long the mballoc __can__ look for a 164c9de560dSAlex Tomas * best extent in the found extents. Searching for the blocks starts with 165c9de560dSAlex Tomas * the group specified as the goal value in allocation context via 166c9de560dSAlex Tomas * ac_g_ex. Each group is first checked based on the criteria whether it 167caaf7a29STao Ma * can be used for allocation. ext4_mb_good_group explains how the groups are 168c9de560dSAlex Tomas * checked. 169c9de560dSAlex Tomas * 170c9de560dSAlex Tomas * Both the prealloc space are getting populated as above. So for the first 171c9de560dSAlex Tomas * request we will hit the buddy cache which will result in this prealloc 172c9de560dSAlex Tomas * space getting filled. The prealloc space is then later used for the 173c9de560dSAlex Tomas * subsequent request. 174c9de560dSAlex Tomas */ 175c9de560dSAlex Tomas 176c9de560dSAlex Tomas /* 177c9de560dSAlex Tomas * mballoc operates on the following data: 178c9de560dSAlex Tomas * - on-disk bitmap 179c9de560dSAlex Tomas * - in-core buddy (actually includes buddy and bitmap) 180c9de560dSAlex Tomas * - preallocation descriptors (PAs) 181c9de560dSAlex Tomas * 182c9de560dSAlex Tomas * there are two types of preallocations: 183c9de560dSAlex Tomas * - inode 184c9de560dSAlex Tomas * assiged to specific inode and can be used for this inode only. 185c9de560dSAlex Tomas * it describes part of inode's space preallocated to specific 186c9de560dSAlex Tomas * physical blocks. any block from that preallocated can be used 187c9de560dSAlex Tomas * independent. the descriptor just tracks number of blocks left 188c9de560dSAlex Tomas * unused. so, before taking some block from descriptor, one must 189c9de560dSAlex Tomas * make sure corresponded logical block isn't allocated yet. this 190c9de560dSAlex Tomas * also means that freeing any block within descriptor's range 191c9de560dSAlex Tomas * must discard all preallocated blocks. 192c9de560dSAlex Tomas * - locality group 193c9de560dSAlex Tomas * assigned to specific locality group which does not translate to 194c9de560dSAlex Tomas * permanent set of inodes: inode can join and leave group. space 195c9de560dSAlex Tomas * from this type of preallocation can be used for any inode. thus 196c9de560dSAlex Tomas * it's consumed from the beginning to the end. 197c9de560dSAlex Tomas * 198c9de560dSAlex Tomas * relation between them can be expressed as: 199c9de560dSAlex Tomas * in-core buddy = on-disk bitmap + preallocation descriptors 200c9de560dSAlex Tomas * 201c9de560dSAlex Tomas * this mean blocks mballoc considers used are: 202c9de560dSAlex Tomas * - allocated blocks (persistent) 203c9de560dSAlex Tomas * - preallocated blocks (non-persistent) 204c9de560dSAlex Tomas * 205c9de560dSAlex Tomas * consistency in mballoc world means that at any time a block is either 206c9de560dSAlex Tomas * free or used in ALL structures. notice: "any time" should not be read 207c9de560dSAlex Tomas * literally -- time is discrete and delimited by locks. 208c9de560dSAlex Tomas * 209c9de560dSAlex Tomas * to keep it simple, we don't use block numbers, instead we count number of 210c9de560dSAlex Tomas * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 211c9de560dSAlex Tomas * 212c9de560dSAlex Tomas * all operations can be expressed as: 213c9de560dSAlex Tomas * - init buddy: buddy = on-disk + PAs 214c9de560dSAlex Tomas * - new PA: buddy += N; PA = N 215c9de560dSAlex Tomas * - use inode PA: on-disk += N; PA -= N 216c9de560dSAlex Tomas * - discard inode PA buddy -= on-disk - PA; PA = 0 217c9de560dSAlex Tomas * - use locality group PA on-disk += N; PA -= N 218c9de560dSAlex Tomas * - discard locality group PA buddy -= PA; PA = 0 219c9de560dSAlex Tomas * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 220c9de560dSAlex Tomas * is used in real operation because we can't know actual used 221c9de560dSAlex Tomas * bits from PA, only from on-disk bitmap 222c9de560dSAlex Tomas * 223c9de560dSAlex Tomas * if we follow this strict logic, then all operations above should be atomic. 224c9de560dSAlex Tomas * given some of them can block, we'd have to use something like semaphores 225c9de560dSAlex Tomas * killing performance on high-end SMP hardware. let's try to relax it using 226c9de560dSAlex Tomas * the following knowledge: 227c9de560dSAlex Tomas * 1) if buddy is referenced, it's already initialized 228c9de560dSAlex Tomas * 2) while block is used in buddy and the buddy is referenced, 229c9de560dSAlex Tomas * nobody can re-allocate that block 230c9de560dSAlex Tomas * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 231c9de560dSAlex Tomas * bit set and PA claims same block, it's OK. IOW, one can set bit in 232c9de560dSAlex Tomas * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 233c9de560dSAlex Tomas * block 234c9de560dSAlex Tomas * 235c9de560dSAlex Tomas * so, now we're building a concurrency table: 236c9de560dSAlex Tomas * - init buddy vs. 237c9de560dSAlex Tomas * - new PA 238c9de560dSAlex Tomas * blocks for PA are allocated in the buddy, buddy must be referenced 239c9de560dSAlex Tomas * until PA is linked to allocation group to avoid concurrent buddy init 240c9de560dSAlex Tomas * - use inode PA 241c9de560dSAlex Tomas * we need to make sure that either on-disk bitmap or PA has uptodate data 242c9de560dSAlex Tomas * given (3) we care that PA-=N operation doesn't interfere with init 243c9de560dSAlex Tomas * - discard inode PA 244c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 245c9de560dSAlex Tomas * - use locality group PA 246c9de560dSAlex Tomas * again PA-=N must be serialized with init 247c9de560dSAlex Tomas * - discard locality group PA 248c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 249c9de560dSAlex Tomas * - new PA vs. 250c9de560dSAlex Tomas * - use inode PA 251c9de560dSAlex Tomas * i_data_sem serializes them 252c9de560dSAlex Tomas * - discard inode PA 253c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 254c9de560dSAlex Tomas * - use locality group PA 255c9de560dSAlex Tomas * some mutex should serialize them 256c9de560dSAlex Tomas * - discard locality group PA 257c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 258c9de560dSAlex Tomas * - use inode PA 259c9de560dSAlex Tomas * - use inode PA 260c9de560dSAlex Tomas * i_data_sem or another mutex should serializes them 261c9de560dSAlex Tomas * - discard inode PA 262c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 263c9de560dSAlex Tomas * - use locality group PA 264c9de560dSAlex Tomas * nothing wrong here -- they're different PAs covering different blocks 265c9de560dSAlex Tomas * - discard locality group PA 266c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 267c9de560dSAlex Tomas * 268c9de560dSAlex Tomas * now we're ready to make few consequences: 269c9de560dSAlex Tomas * - PA is referenced and while it is no discard is possible 270c9de560dSAlex Tomas * - PA is referenced until block isn't marked in on-disk bitmap 271c9de560dSAlex Tomas * - PA changes only after on-disk bitmap 272c9de560dSAlex Tomas * - discard must not compete with init. either init is done before 273c9de560dSAlex Tomas * any discard or they're serialized somehow 274c9de560dSAlex Tomas * - buddy init as sum of on-disk bitmap and PAs is done atomically 275c9de560dSAlex Tomas * 276c9de560dSAlex Tomas * a special case when we've used PA to emptiness. no need to modify buddy 277c9de560dSAlex Tomas * in this case, but we should care about concurrent init 278c9de560dSAlex Tomas * 279c9de560dSAlex Tomas */ 280c9de560dSAlex Tomas 281c9de560dSAlex Tomas /* 282c9de560dSAlex Tomas * Logic in few words: 283c9de560dSAlex Tomas * 284c9de560dSAlex Tomas * - allocation: 285c9de560dSAlex Tomas * load group 286c9de560dSAlex Tomas * find blocks 287c9de560dSAlex Tomas * mark bits in on-disk bitmap 288c9de560dSAlex Tomas * release group 289c9de560dSAlex Tomas * 290c9de560dSAlex Tomas * - use preallocation: 291c9de560dSAlex Tomas * find proper PA (per-inode or group) 292c9de560dSAlex Tomas * load group 293c9de560dSAlex Tomas * mark bits in on-disk bitmap 294c9de560dSAlex Tomas * release group 295c9de560dSAlex Tomas * release PA 296c9de560dSAlex Tomas * 297c9de560dSAlex Tomas * - free: 298c9de560dSAlex Tomas * load group 299c9de560dSAlex Tomas * mark bits in on-disk bitmap 300c9de560dSAlex Tomas * release group 301c9de560dSAlex Tomas * 302c9de560dSAlex Tomas * - discard preallocations in group: 303c9de560dSAlex Tomas * mark PAs deleted 304c9de560dSAlex Tomas * move them onto local list 305c9de560dSAlex Tomas * load on-disk bitmap 306c9de560dSAlex Tomas * load group 307c9de560dSAlex Tomas * remove PA from object (inode or locality group) 308c9de560dSAlex Tomas * mark free blocks in-core 309c9de560dSAlex Tomas * 310c9de560dSAlex Tomas * - discard inode's preallocations: 311c9de560dSAlex Tomas */ 312c9de560dSAlex Tomas 313c9de560dSAlex Tomas /* 314c9de560dSAlex Tomas * Locking rules 315c9de560dSAlex Tomas * 316c9de560dSAlex Tomas * Locks: 317c9de560dSAlex Tomas * - bitlock on a group (group) 318c9de560dSAlex Tomas * - object (inode/locality) (object) 319c9de560dSAlex Tomas * - per-pa lock (pa) 320c9de560dSAlex Tomas * 321c9de560dSAlex Tomas * Paths: 322c9de560dSAlex Tomas * - new pa 323c9de560dSAlex Tomas * object 324c9de560dSAlex Tomas * group 325c9de560dSAlex Tomas * 326c9de560dSAlex Tomas * - find and use pa: 327c9de560dSAlex Tomas * pa 328c9de560dSAlex Tomas * 329c9de560dSAlex Tomas * - release consumed pa: 330c9de560dSAlex Tomas * pa 331c9de560dSAlex Tomas * group 332c9de560dSAlex Tomas * object 333c9de560dSAlex Tomas * 334c9de560dSAlex Tomas * - generate in-core bitmap: 335c9de560dSAlex Tomas * group 336c9de560dSAlex Tomas * pa 337c9de560dSAlex Tomas * 338c9de560dSAlex Tomas * - discard all for given object (inode, locality group): 339c9de560dSAlex Tomas * object 340c9de560dSAlex Tomas * pa 341c9de560dSAlex Tomas * group 342c9de560dSAlex Tomas * 343c9de560dSAlex Tomas * - discard all for given group: 344c9de560dSAlex Tomas * group 345c9de560dSAlex Tomas * pa 346c9de560dSAlex Tomas * group 347c9de560dSAlex Tomas * object 348c9de560dSAlex Tomas * 349c9de560dSAlex Tomas */ 350c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_pspace_cachep; 351c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_ac_cachep; 35218aadd47SBobi Jam static struct kmem_cache *ext4_free_data_cachep; 353fb1813f4SCurt Wohlgemuth 354fb1813f4SCurt Wohlgemuth /* We create slab caches for groupinfo data structures based on the 355fb1813f4SCurt Wohlgemuth * superblock block size. There will be one per mounted filesystem for 356fb1813f4SCurt Wohlgemuth * each unique s_blocksize_bits */ 3572892c15dSEric Sandeen #define NR_GRPINFO_CACHES 8 358fb1813f4SCurt Wohlgemuth static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; 359fb1813f4SCurt Wohlgemuth 3602892c15dSEric Sandeen static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { 3612892c15dSEric Sandeen "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", 3622892c15dSEric Sandeen "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", 3632892c15dSEric Sandeen "ext4_groupinfo_64k", "ext4_groupinfo_128k" 3642892c15dSEric Sandeen }; 3652892c15dSEric Sandeen 366c3a326a6SAneesh Kumar K.V static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 367c3a326a6SAneesh Kumar K.V ext4_group_t group); 3687a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 3697a2fcbf7SAneesh Kumar K.V ext4_group_t group); 37018aadd47SBobi Jam static void ext4_free_data_callback(struct super_block *sb, 37118aadd47SBobi Jam struct ext4_journal_cb_entry *jce, int rc); 372c3a326a6SAneesh Kumar K.V 373ffad0a44SAneesh Kumar K.V static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 374ffad0a44SAneesh Kumar K.V { 375c9de560dSAlex Tomas #if BITS_PER_LONG == 64 376ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 7UL) << 3; 377ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~7UL); 378c9de560dSAlex Tomas #elif BITS_PER_LONG == 32 379ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 3UL) << 3; 380ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~3UL); 381c9de560dSAlex Tomas #else 382c9de560dSAlex Tomas #error "how many bits you are?!" 383c9de560dSAlex Tomas #endif 384ffad0a44SAneesh Kumar K.V return addr; 385ffad0a44SAneesh Kumar K.V } 386c9de560dSAlex Tomas 387c9de560dSAlex Tomas static inline int mb_test_bit(int bit, void *addr) 388c9de560dSAlex Tomas { 389c9de560dSAlex Tomas /* 390c9de560dSAlex Tomas * ext4_test_bit on architecture like powerpc 391c9de560dSAlex Tomas * needs unsigned long aligned address 392c9de560dSAlex Tomas */ 393ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 394c9de560dSAlex Tomas return ext4_test_bit(bit, addr); 395c9de560dSAlex Tomas } 396c9de560dSAlex Tomas 397c9de560dSAlex Tomas static inline void mb_set_bit(int bit, void *addr) 398c9de560dSAlex Tomas { 399ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 400c9de560dSAlex Tomas ext4_set_bit(bit, addr); 401c9de560dSAlex Tomas } 402c9de560dSAlex Tomas 403c9de560dSAlex Tomas static inline void mb_clear_bit(int bit, void *addr) 404c9de560dSAlex Tomas { 405ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 406c9de560dSAlex Tomas ext4_clear_bit(bit, addr); 407c9de560dSAlex Tomas } 408c9de560dSAlex Tomas 409eabe0444SAndrey Sidorov static inline int mb_test_and_clear_bit(int bit, void *addr) 410eabe0444SAndrey Sidorov { 411eabe0444SAndrey Sidorov addr = mb_correct_addr_and_bit(&bit, addr); 412eabe0444SAndrey Sidorov return ext4_test_and_clear_bit(bit, addr); 413eabe0444SAndrey Sidorov } 414eabe0444SAndrey Sidorov 415ffad0a44SAneesh Kumar K.V static inline int mb_find_next_zero_bit(void *addr, int max, int start) 416ffad0a44SAneesh Kumar K.V { 417e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 418ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 419e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 420ffad0a44SAneesh Kumar K.V start += fix; 421ffad0a44SAneesh Kumar K.V 422e7dfb246SAneesh Kumar K.V ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 423e7dfb246SAneesh Kumar K.V if (ret > max) 424e7dfb246SAneesh Kumar K.V return max; 425e7dfb246SAneesh Kumar K.V return ret; 426ffad0a44SAneesh Kumar K.V } 427ffad0a44SAneesh Kumar K.V 428ffad0a44SAneesh Kumar K.V static inline int mb_find_next_bit(void *addr, int max, int start) 429ffad0a44SAneesh Kumar K.V { 430e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 431ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 432e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 433ffad0a44SAneesh Kumar K.V start += fix; 434ffad0a44SAneesh Kumar K.V 435e7dfb246SAneesh Kumar K.V ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 436e7dfb246SAneesh Kumar K.V if (ret > max) 437e7dfb246SAneesh Kumar K.V return max; 438e7dfb246SAneesh Kumar K.V return ret; 439ffad0a44SAneesh Kumar K.V } 440ffad0a44SAneesh Kumar K.V 441c9de560dSAlex Tomas static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 442c9de560dSAlex Tomas { 443c9de560dSAlex Tomas char *bb; 444c9de560dSAlex Tomas 445c5e8f3f3STheodore Ts'o BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 446c9de560dSAlex Tomas BUG_ON(max == NULL); 447c9de560dSAlex Tomas 448c9de560dSAlex Tomas if (order > e4b->bd_blkbits + 1) { 449c9de560dSAlex Tomas *max = 0; 450c9de560dSAlex Tomas return NULL; 451c9de560dSAlex Tomas } 452c9de560dSAlex Tomas 453c9de560dSAlex Tomas /* at order 0 we see each particular block */ 45484b775a3SColy Li if (order == 0) { 455c9de560dSAlex Tomas *max = 1 << (e4b->bd_blkbits + 3); 456c5e8f3f3STheodore Ts'o return e4b->bd_bitmap; 45784b775a3SColy Li } 458c9de560dSAlex Tomas 459c5e8f3f3STheodore Ts'o bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 460c9de560dSAlex Tomas *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 461c9de560dSAlex Tomas 462c9de560dSAlex Tomas return bb; 463c9de560dSAlex Tomas } 464c9de560dSAlex Tomas 465c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 466c9de560dSAlex Tomas static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 467c9de560dSAlex Tomas int first, int count) 468c9de560dSAlex Tomas { 469c9de560dSAlex Tomas int i; 470c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 471c9de560dSAlex Tomas 472c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 473c9de560dSAlex Tomas return; 474bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 475c9de560dSAlex Tomas for (i = 0; i < count; i++) { 476c9de560dSAlex Tomas if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 477c9de560dSAlex Tomas ext4_fsblk_t blocknr; 4785661bd68SAkinobu Mita 4795661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 48053accfa9STheodore Ts'o blocknr += EXT4_C2B(EXT4_SB(sb), first + i); 4815d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 482e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 483e29136f8STheodore Ts'o blocknr, 484e29136f8STheodore Ts'o "freeing block already freed " 485e29136f8STheodore Ts'o "(bit %u)", 486e29136f8STheodore Ts'o first + i); 487c9de560dSAlex Tomas } 488c9de560dSAlex Tomas mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 489c9de560dSAlex Tomas } 490c9de560dSAlex Tomas } 491c9de560dSAlex Tomas 492c9de560dSAlex Tomas static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 493c9de560dSAlex Tomas { 494c9de560dSAlex Tomas int i; 495c9de560dSAlex Tomas 496c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 497c9de560dSAlex Tomas return; 498bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 499c9de560dSAlex Tomas for (i = 0; i < count; i++) { 500c9de560dSAlex Tomas BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 501c9de560dSAlex Tomas mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 502c9de560dSAlex Tomas } 503c9de560dSAlex Tomas } 504c9de560dSAlex Tomas 505c9de560dSAlex Tomas static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 506c9de560dSAlex Tomas { 507c9de560dSAlex Tomas if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 508c9de560dSAlex Tomas unsigned char *b1, *b2; 509c9de560dSAlex Tomas int i; 510c9de560dSAlex Tomas b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 511c9de560dSAlex Tomas b2 = (unsigned char *) bitmap; 512c9de560dSAlex Tomas for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 513c9de560dSAlex Tomas if (b1[i] != b2[i]) { 5149d8b9ec4STheodore Ts'o ext4_msg(e4b->bd_sb, KERN_ERR, 5159d8b9ec4STheodore Ts'o "corruption in group %u " 5164776004fSTheodore Ts'o "at byte %u(%u): %x in copy != %x " 5179d8b9ec4STheodore Ts'o "on disk/prealloc", 518c9de560dSAlex Tomas e4b->bd_group, i, i * 8, b1[i], b2[i]); 519c9de560dSAlex Tomas BUG(); 520c9de560dSAlex Tomas } 521c9de560dSAlex Tomas } 522c9de560dSAlex Tomas } 523c9de560dSAlex Tomas } 524c9de560dSAlex Tomas 525c9de560dSAlex Tomas #else 526c9de560dSAlex Tomas static inline void mb_free_blocks_double(struct inode *inode, 527c9de560dSAlex Tomas struct ext4_buddy *e4b, int first, int count) 528c9de560dSAlex Tomas { 529c9de560dSAlex Tomas return; 530c9de560dSAlex Tomas } 531c9de560dSAlex Tomas static inline void mb_mark_used_double(struct ext4_buddy *e4b, 532c9de560dSAlex Tomas int first, int count) 533c9de560dSAlex Tomas { 534c9de560dSAlex Tomas return; 535c9de560dSAlex Tomas } 536c9de560dSAlex Tomas static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 537c9de560dSAlex Tomas { 538c9de560dSAlex Tomas return; 539c9de560dSAlex Tomas } 540c9de560dSAlex Tomas #endif 541c9de560dSAlex Tomas 542c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 543c9de560dSAlex Tomas 544c9de560dSAlex Tomas #define MB_CHECK_ASSERT(assert) \ 545c9de560dSAlex Tomas do { \ 546c9de560dSAlex Tomas if (!(assert)) { \ 547c9de560dSAlex Tomas printk(KERN_EMERG \ 548c9de560dSAlex Tomas "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 549c9de560dSAlex Tomas function, file, line, # assert); \ 550c9de560dSAlex Tomas BUG(); \ 551c9de560dSAlex Tomas } \ 552c9de560dSAlex Tomas } while (0) 553c9de560dSAlex Tomas 554c9de560dSAlex Tomas static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 555c9de560dSAlex Tomas const char *function, int line) 556c9de560dSAlex Tomas { 557c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 558c9de560dSAlex Tomas int order = e4b->bd_blkbits + 1; 559c9de560dSAlex Tomas int max; 560c9de560dSAlex Tomas int max2; 561c9de560dSAlex Tomas int i; 562c9de560dSAlex Tomas int j; 563c9de560dSAlex Tomas int k; 564c9de560dSAlex Tomas int count; 565c9de560dSAlex Tomas struct ext4_group_info *grp; 566c9de560dSAlex Tomas int fragments = 0; 567c9de560dSAlex Tomas int fstart; 568c9de560dSAlex Tomas struct list_head *cur; 569c9de560dSAlex Tomas void *buddy; 570c9de560dSAlex Tomas void *buddy2; 571c9de560dSAlex Tomas 572c9de560dSAlex Tomas { 573c9de560dSAlex Tomas static int mb_check_counter; 574c9de560dSAlex Tomas if (mb_check_counter++ % 100 != 0) 575c9de560dSAlex Tomas return 0; 576c9de560dSAlex Tomas } 577c9de560dSAlex Tomas 578c9de560dSAlex Tomas while (order > 1) { 579c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 580c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy); 581c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order - 1, &max2); 582c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy2); 583c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy != buddy2); 584c9de560dSAlex Tomas MB_CHECK_ASSERT(max * 2 == max2); 585c9de560dSAlex Tomas 586c9de560dSAlex Tomas count = 0; 587c9de560dSAlex Tomas for (i = 0; i < max; i++) { 588c9de560dSAlex Tomas 589c9de560dSAlex Tomas if (mb_test_bit(i, buddy)) { 590c9de560dSAlex Tomas /* only single bit in buddy2 may be 1 */ 591c9de560dSAlex Tomas if (!mb_test_bit(i << 1, buddy2)) { 592c9de560dSAlex Tomas MB_CHECK_ASSERT( 593c9de560dSAlex Tomas mb_test_bit((i<<1)+1, buddy2)); 594c9de560dSAlex Tomas } else if (!mb_test_bit((i << 1) + 1, buddy2)) { 595c9de560dSAlex Tomas MB_CHECK_ASSERT( 596c9de560dSAlex Tomas mb_test_bit(i << 1, buddy2)); 597c9de560dSAlex Tomas } 598c9de560dSAlex Tomas continue; 599c9de560dSAlex Tomas } 600c9de560dSAlex Tomas 6010a10da73SRobin Dong /* both bits in buddy2 must be 1 */ 602c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 603c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 604c9de560dSAlex Tomas 605c9de560dSAlex Tomas for (j = 0; j < (1 << order); j++) { 606c9de560dSAlex Tomas k = (i * (1 << order)) + j; 607c9de560dSAlex Tomas MB_CHECK_ASSERT( 608c5e8f3f3STheodore Ts'o !mb_test_bit(k, e4b->bd_bitmap)); 609c9de560dSAlex Tomas } 610c9de560dSAlex Tomas count++; 611c9de560dSAlex Tomas } 612c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 613c9de560dSAlex Tomas order--; 614c9de560dSAlex Tomas } 615c9de560dSAlex Tomas 616c9de560dSAlex Tomas fstart = -1; 617c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 618c9de560dSAlex Tomas for (i = 0; i < max; i++) { 619c9de560dSAlex Tomas if (!mb_test_bit(i, buddy)) { 620c9de560dSAlex Tomas MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 621c9de560dSAlex Tomas if (fstart == -1) { 622c9de560dSAlex Tomas fragments++; 623c9de560dSAlex Tomas fstart = i; 624c9de560dSAlex Tomas } 625c9de560dSAlex Tomas continue; 626c9de560dSAlex Tomas } 627c9de560dSAlex Tomas fstart = -1; 628c9de560dSAlex Tomas /* check used bits only */ 629c9de560dSAlex Tomas for (j = 0; j < e4b->bd_blkbits + 1; j++) { 630c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, j, &max2); 631c9de560dSAlex Tomas k = i >> j; 632c9de560dSAlex Tomas MB_CHECK_ASSERT(k < max2); 633c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 634c9de560dSAlex Tomas } 635c9de560dSAlex Tomas } 636c9de560dSAlex Tomas MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 637c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 638c9de560dSAlex Tomas 639c9de560dSAlex Tomas grp = ext4_get_group_info(sb, e4b->bd_group); 640c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 641c9de560dSAlex Tomas ext4_group_t groupnr; 642c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 64360bd63d1SSolofo Ramangalahy pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 64460bd63d1SSolofo Ramangalahy ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 645c9de560dSAlex Tomas MB_CHECK_ASSERT(groupnr == e4b->bd_group); 64660bd63d1SSolofo Ramangalahy for (i = 0; i < pa->pa_len; i++) 647c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 648c9de560dSAlex Tomas } 649c9de560dSAlex Tomas return 0; 650c9de560dSAlex Tomas } 651c9de560dSAlex Tomas #undef MB_CHECK_ASSERT 652c9de560dSAlex Tomas #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 65346e665e9SHarvey Harrison __FILE__, __func__, __LINE__) 654c9de560dSAlex Tomas #else 655c9de560dSAlex Tomas #define mb_check_buddy(e4b) 656c9de560dSAlex Tomas #endif 657c9de560dSAlex Tomas 6587c786059SColy Li /* 6597c786059SColy Li * Divide blocks started from @first with length @len into 6607c786059SColy Li * smaller chunks with power of 2 blocks. 6617c786059SColy Li * Clear the bits in bitmap which the blocks of the chunk(s) covered, 6627c786059SColy Li * then increase bb_counters[] for corresponded chunk size. 6637c786059SColy Li */ 664c9de560dSAlex Tomas static void ext4_mb_mark_free_simple(struct super_block *sb, 665a36b4498SEric Sandeen void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 666c9de560dSAlex Tomas struct ext4_group_info *grp) 667c9de560dSAlex Tomas { 668c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 669a36b4498SEric Sandeen ext4_grpblk_t min; 670a36b4498SEric Sandeen ext4_grpblk_t max; 671a36b4498SEric Sandeen ext4_grpblk_t chunk; 672c9de560dSAlex Tomas unsigned short border; 673c9de560dSAlex Tomas 6747137d7a4STheodore Ts'o BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); 675c9de560dSAlex Tomas 676c9de560dSAlex Tomas border = 2 << sb->s_blocksize_bits; 677c9de560dSAlex Tomas 678c9de560dSAlex Tomas while (len > 0) { 679c9de560dSAlex Tomas /* find how many blocks can be covered since this position */ 680c9de560dSAlex Tomas max = ffs(first | border) - 1; 681c9de560dSAlex Tomas 682c9de560dSAlex Tomas /* find how many blocks of power 2 we need to mark */ 683c9de560dSAlex Tomas min = fls(len) - 1; 684c9de560dSAlex Tomas 685c9de560dSAlex Tomas if (max < min) 686c9de560dSAlex Tomas min = max; 687c9de560dSAlex Tomas chunk = 1 << min; 688c9de560dSAlex Tomas 689c9de560dSAlex Tomas /* mark multiblock chunks only */ 690c9de560dSAlex Tomas grp->bb_counters[min]++; 691c9de560dSAlex Tomas if (min > 0) 692c9de560dSAlex Tomas mb_clear_bit(first >> min, 693c9de560dSAlex Tomas buddy + sbi->s_mb_offsets[min]); 694c9de560dSAlex Tomas 695c9de560dSAlex Tomas len -= chunk; 696c9de560dSAlex Tomas first += chunk; 697c9de560dSAlex Tomas } 698c9de560dSAlex Tomas } 699c9de560dSAlex Tomas 7008a57d9d6SCurt Wohlgemuth /* 7018a57d9d6SCurt Wohlgemuth * Cache the order of the largest free extent we have available in this block 7028a57d9d6SCurt Wohlgemuth * group. 7038a57d9d6SCurt Wohlgemuth */ 7048a57d9d6SCurt Wohlgemuth static void 7058a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) 7068a57d9d6SCurt Wohlgemuth { 7078a57d9d6SCurt Wohlgemuth int i; 7088a57d9d6SCurt Wohlgemuth int bits; 7098a57d9d6SCurt Wohlgemuth 7108a57d9d6SCurt Wohlgemuth grp->bb_largest_free_order = -1; /* uninit */ 7118a57d9d6SCurt Wohlgemuth 7128a57d9d6SCurt Wohlgemuth bits = sb->s_blocksize_bits + 1; 7138a57d9d6SCurt Wohlgemuth for (i = bits; i >= 0; i--) { 7148a57d9d6SCurt Wohlgemuth if (grp->bb_counters[i] > 0) { 7158a57d9d6SCurt Wohlgemuth grp->bb_largest_free_order = i; 7168a57d9d6SCurt Wohlgemuth break; 7178a57d9d6SCurt Wohlgemuth } 7188a57d9d6SCurt Wohlgemuth } 7198a57d9d6SCurt Wohlgemuth } 7208a57d9d6SCurt Wohlgemuth 721089ceeccSEric Sandeen static noinline_for_stack 722089ceeccSEric Sandeen void ext4_mb_generate_buddy(struct super_block *sb, 723c9de560dSAlex Tomas void *buddy, void *bitmap, ext4_group_t group) 724c9de560dSAlex Tomas { 725c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 726e43bb4e6SNamjae Jeon struct ext4_sb_info *sbi = EXT4_SB(sb); 7277137d7a4STheodore Ts'o ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 728a36b4498SEric Sandeen ext4_grpblk_t i = 0; 729a36b4498SEric Sandeen ext4_grpblk_t first; 730a36b4498SEric Sandeen ext4_grpblk_t len; 731c9de560dSAlex Tomas unsigned free = 0; 732c9de560dSAlex Tomas unsigned fragments = 0; 733c9de560dSAlex Tomas unsigned long long period = get_cycles(); 734c9de560dSAlex Tomas 735c9de560dSAlex Tomas /* initialize buddy from bitmap which is aggregation 736c9de560dSAlex Tomas * of on-disk bitmap and preallocations */ 737ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, 0); 738c9de560dSAlex Tomas grp->bb_first_free = i; 739c9de560dSAlex Tomas while (i < max) { 740c9de560dSAlex Tomas fragments++; 741c9de560dSAlex Tomas first = i; 742ffad0a44SAneesh Kumar K.V i = mb_find_next_bit(bitmap, max, i); 743c9de560dSAlex Tomas len = i - first; 744c9de560dSAlex Tomas free += len; 745c9de560dSAlex Tomas if (len > 1) 746c9de560dSAlex Tomas ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 747c9de560dSAlex Tomas else 748c9de560dSAlex Tomas grp->bb_counters[0]++; 749c9de560dSAlex Tomas if (i < max) 750ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, i); 751c9de560dSAlex Tomas } 752c9de560dSAlex Tomas grp->bb_fragments = fragments; 753c9de560dSAlex Tomas 754c9de560dSAlex Tomas if (free != grp->bb_free) { 755e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, 75694d4c066STheodore Ts'o "block bitmap and bg descriptor " 75794d4c066STheodore Ts'o "inconsistent: %u vs %u free clusters", 758e29136f8STheodore Ts'o free, grp->bb_free); 759e56eb659SAneesh Kumar K.V /* 760163a203dSDarrick J. Wong * If we intend to continue, we consider group descriptor 761e56eb659SAneesh Kumar K.V * corrupt and update bb_free using bitmap value 762e56eb659SAneesh Kumar K.V */ 763c9de560dSAlex Tomas grp->bb_free = free; 764e43bb4e6SNamjae Jeon if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp)) 765e43bb4e6SNamjae Jeon percpu_counter_sub(&sbi->s_freeclusters_counter, 766e43bb4e6SNamjae Jeon grp->bb_free); 767163a203dSDarrick J. Wong set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state); 768c9de560dSAlex Tomas } 7698a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, grp); 770c9de560dSAlex Tomas 771c9de560dSAlex Tomas clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 772c9de560dSAlex Tomas 773c9de560dSAlex Tomas period = get_cycles() - period; 774c9de560dSAlex Tomas spin_lock(&EXT4_SB(sb)->s_bal_lock); 775c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_buddies_generated++; 776c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_generation_time += period; 777c9de560dSAlex Tomas spin_unlock(&EXT4_SB(sb)->s_bal_lock); 778c9de560dSAlex Tomas } 779c9de560dSAlex Tomas 780eabe0444SAndrey Sidorov static void mb_regenerate_buddy(struct ext4_buddy *e4b) 781eabe0444SAndrey Sidorov { 782eabe0444SAndrey Sidorov int count; 783eabe0444SAndrey Sidorov int order = 1; 784eabe0444SAndrey Sidorov void *buddy; 785eabe0444SAndrey Sidorov 786eabe0444SAndrey Sidorov while ((buddy = mb_find_buddy(e4b, order++, &count))) { 787eabe0444SAndrey Sidorov ext4_set_bits(buddy, 0, count); 788eabe0444SAndrey Sidorov } 789eabe0444SAndrey Sidorov e4b->bd_info->bb_fragments = 0; 790eabe0444SAndrey Sidorov memset(e4b->bd_info->bb_counters, 0, 791eabe0444SAndrey Sidorov sizeof(*e4b->bd_info->bb_counters) * 792eabe0444SAndrey Sidorov (e4b->bd_sb->s_blocksize_bits + 2)); 793eabe0444SAndrey Sidorov 794eabe0444SAndrey Sidorov ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy, 795eabe0444SAndrey Sidorov e4b->bd_bitmap, e4b->bd_group); 796eabe0444SAndrey Sidorov } 797eabe0444SAndrey Sidorov 798c9de560dSAlex Tomas /* The buddy information is attached the buddy cache inode 799c9de560dSAlex Tomas * for convenience. The information regarding each group 800c9de560dSAlex Tomas * is loaded via ext4_mb_load_buddy. The information involve 801c9de560dSAlex Tomas * block bitmap and buddy information. The information are 802c9de560dSAlex Tomas * stored in the inode as 803c9de560dSAlex Tomas * 804c9de560dSAlex Tomas * { page } 805c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 806c9de560dSAlex Tomas * 807c9de560dSAlex Tomas * 808c9de560dSAlex Tomas * one block each for bitmap and buddy information. 809c9de560dSAlex Tomas * So for each group we take up 2 blocks. A page can 810ea1754a0SKirill A. Shutemov * contain blocks_per_page (PAGE_SIZE / blocksize) blocks. 811c9de560dSAlex Tomas * So it can have information regarding groups_per_page which 812c9de560dSAlex Tomas * is blocks_per_page/2 8138a57d9d6SCurt Wohlgemuth * 8148a57d9d6SCurt Wohlgemuth * Locking note: This routine takes the block group lock of all groups 8158a57d9d6SCurt Wohlgemuth * for this page; do not hold this lock when calling this routine! 816c9de560dSAlex Tomas */ 817c9de560dSAlex Tomas 818adb7ef60SKonstantin Khlebnikov static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp) 819c9de560dSAlex Tomas { 8208df9675fSTheodore Ts'o ext4_group_t ngroups; 821c9de560dSAlex Tomas int blocksize; 822c9de560dSAlex Tomas int blocks_per_page; 823c9de560dSAlex Tomas int groups_per_page; 824c9de560dSAlex Tomas int err = 0; 825c9de560dSAlex Tomas int i; 826813e5727STheodore Ts'o ext4_group_t first_group, group; 827c9de560dSAlex Tomas int first_block; 828c9de560dSAlex Tomas struct super_block *sb; 829c9de560dSAlex Tomas struct buffer_head *bhs; 830fa77dcfaSDarrick J. Wong struct buffer_head **bh = NULL; 831c9de560dSAlex Tomas struct inode *inode; 832c9de560dSAlex Tomas char *data; 833c9de560dSAlex Tomas char *bitmap; 8349b8b7d35SAmir Goldstein struct ext4_group_info *grinfo; 835c9de560dSAlex Tomas 8366ba495e9STheodore Ts'o mb_debug(1, "init page %lu\n", page->index); 837c9de560dSAlex Tomas 838c9de560dSAlex Tomas inode = page->mapping->host; 839c9de560dSAlex Tomas sb = inode->i_sb; 8408df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 841c9de560dSAlex Tomas blocksize = 1 << inode->i_blkbits; 84209cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / blocksize; 843c9de560dSAlex Tomas 844c9de560dSAlex Tomas groups_per_page = blocks_per_page >> 1; 845c9de560dSAlex Tomas if (groups_per_page == 0) 846c9de560dSAlex Tomas groups_per_page = 1; 847c9de560dSAlex Tomas 848c9de560dSAlex Tomas /* allocate buffer_heads to read bitmaps */ 849c9de560dSAlex Tomas if (groups_per_page > 1) { 850c9de560dSAlex Tomas i = sizeof(struct buffer_head *) * groups_per_page; 851adb7ef60SKonstantin Khlebnikov bh = kzalloc(i, gfp); 852813e5727STheodore Ts'o if (bh == NULL) { 853813e5727STheodore Ts'o err = -ENOMEM; 854c9de560dSAlex Tomas goto out; 855813e5727STheodore Ts'o } 856c9de560dSAlex Tomas } else 857c9de560dSAlex Tomas bh = &bhs; 858c9de560dSAlex Tomas 859c9de560dSAlex Tomas first_group = page->index * blocks_per_page / 2; 860c9de560dSAlex Tomas 861c9de560dSAlex Tomas /* read all groups the page covers into the cache */ 862813e5727STheodore Ts'o for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 863813e5727STheodore Ts'o if (group >= ngroups) 864c9de560dSAlex Tomas break; 865c9de560dSAlex Tomas 866813e5727STheodore Ts'o grinfo = ext4_get_group_info(sb, group); 8679b8b7d35SAmir Goldstein /* 8689b8b7d35SAmir Goldstein * If page is uptodate then we came here after online resize 8699b8b7d35SAmir Goldstein * which added some new uninitialized group info structs, so 8709b8b7d35SAmir Goldstein * we must skip all initialized uptodate buddies on the page, 8719b8b7d35SAmir Goldstein * which may be currently in use by an allocating task. 8729b8b7d35SAmir Goldstein */ 8739b8b7d35SAmir Goldstein if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { 8749b8b7d35SAmir Goldstein bh[i] = NULL; 8759b8b7d35SAmir Goldstein continue; 8769b8b7d35SAmir Goldstein } 8779008a58eSDarrick J. Wong bh[i] = ext4_read_block_bitmap_nowait(sb, group); 8789008a58eSDarrick J. Wong if (IS_ERR(bh[i])) { 8799008a58eSDarrick J. Wong err = PTR_ERR(bh[i]); 8809008a58eSDarrick J. Wong bh[i] = NULL; 881c9de560dSAlex Tomas goto out; 8822ccb5fb9SAneesh Kumar K.V } 883813e5727STheodore Ts'o mb_debug(1, "read bitmap for group %u\n", group); 884c9de560dSAlex Tomas } 885c9de560dSAlex Tomas 886c9de560dSAlex Tomas /* wait for I/O completion */ 887813e5727STheodore Ts'o for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 8889008a58eSDarrick J. Wong int err2; 8899008a58eSDarrick J. Wong 8909008a58eSDarrick J. Wong if (!bh[i]) 8919008a58eSDarrick J. Wong continue; 8929008a58eSDarrick J. Wong err2 = ext4_wait_block_bitmap(sb, group, bh[i]); 8939008a58eSDarrick J. Wong if (!err) 8949008a58eSDarrick J. Wong err = err2; 895813e5727STheodore Ts'o } 896c9de560dSAlex Tomas 897c9de560dSAlex Tomas first_block = page->index * blocks_per_page; 898c9de560dSAlex Tomas for (i = 0; i < blocks_per_page; i++) { 899c9de560dSAlex Tomas group = (first_block + i) >> 1; 9008df9675fSTheodore Ts'o if (group >= ngroups) 901c9de560dSAlex Tomas break; 902c9de560dSAlex Tomas 9039b8b7d35SAmir Goldstein if (!bh[group - first_group]) 9049b8b7d35SAmir Goldstein /* skip initialized uptodate buddy */ 9059b8b7d35SAmir Goldstein continue; 9069b8b7d35SAmir Goldstein 907bbdc322fSLukas Czerner if (!buffer_verified(bh[group - first_group])) 908bbdc322fSLukas Czerner /* Skip faulty bitmaps */ 909bbdc322fSLukas Czerner continue; 910bbdc322fSLukas Czerner err = 0; 911bbdc322fSLukas Czerner 912c9de560dSAlex Tomas /* 913c9de560dSAlex Tomas * data carry information regarding this 914c9de560dSAlex Tomas * particular group in the format specified 915c9de560dSAlex Tomas * above 916c9de560dSAlex Tomas * 917c9de560dSAlex Tomas */ 918c9de560dSAlex Tomas data = page_address(page) + (i * blocksize); 919c9de560dSAlex Tomas bitmap = bh[group - first_group]->b_data; 920c9de560dSAlex Tomas 921c9de560dSAlex Tomas /* 922c9de560dSAlex Tomas * We place the buddy block and bitmap block 923c9de560dSAlex Tomas * close together 924c9de560dSAlex Tomas */ 925c9de560dSAlex Tomas if ((first_block + i) & 1) { 926c9de560dSAlex Tomas /* this is block of buddy */ 927c9de560dSAlex Tomas BUG_ON(incore == NULL); 9286ba495e9STheodore Ts'o mb_debug(1, "put buddy for group %u in page %lu/%x\n", 929c9de560dSAlex Tomas group, page->index, i * blocksize); 930f307333eSTheodore Ts'o trace_ext4_mb_buddy_bitmap_load(sb, group); 931c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, group); 932c9de560dSAlex Tomas grinfo->bb_fragments = 0; 933c9de560dSAlex Tomas memset(grinfo->bb_counters, 0, 9341927805eSEric Sandeen sizeof(*grinfo->bb_counters) * 9351927805eSEric Sandeen (sb->s_blocksize_bits+2)); 936c9de560dSAlex Tomas /* 937c9de560dSAlex Tomas * incore got set to the group block bitmap below 938c9de560dSAlex Tomas */ 9397a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, group); 9409b8b7d35SAmir Goldstein /* init the buddy */ 9419b8b7d35SAmir Goldstein memset(data, 0xff, blocksize); 942c9de560dSAlex Tomas ext4_mb_generate_buddy(sb, data, incore, group); 9437a2fcbf7SAneesh Kumar K.V ext4_unlock_group(sb, group); 944c9de560dSAlex Tomas incore = NULL; 945c9de560dSAlex Tomas } else { 946c9de560dSAlex Tomas /* this is block of bitmap */ 947c9de560dSAlex Tomas BUG_ON(incore != NULL); 9486ba495e9STheodore Ts'o mb_debug(1, "put bitmap for group %u in page %lu/%x\n", 949c9de560dSAlex Tomas group, page->index, i * blocksize); 950f307333eSTheodore Ts'o trace_ext4_mb_bitmap_load(sb, group); 951c9de560dSAlex Tomas 952c9de560dSAlex Tomas /* see comments in ext4_mb_put_pa() */ 953c9de560dSAlex Tomas ext4_lock_group(sb, group); 954c9de560dSAlex Tomas memcpy(data, bitmap, blocksize); 955c9de560dSAlex Tomas 956c9de560dSAlex Tomas /* mark all preallocated blks used in in-core bitmap */ 957c9de560dSAlex Tomas ext4_mb_generate_from_pa(sb, data, group); 9587a2fcbf7SAneesh Kumar K.V ext4_mb_generate_from_freelist(sb, data, group); 959c9de560dSAlex Tomas ext4_unlock_group(sb, group); 960c9de560dSAlex Tomas 961c9de560dSAlex Tomas /* set incore so that the buddy information can be 962c9de560dSAlex Tomas * generated using this 963c9de560dSAlex Tomas */ 964c9de560dSAlex Tomas incore = data; 965c9de560dSAlex Tomas } 966c9de560dSAlex Tomas } 967c9de560dSAlex Tomas SetPageUptodate(page); 968c9de560dSAlex Tomas 969c9de560dSAlex Tomas out: 970c9de560dSAlex Tomas if (bh) { 9719b8b7d35SAmir Goldstein for (i = 0; i < groups_per_page; i++) 972c9de560dSAlex Tomas brelse(bh[i]); 973c9de560dSAlex Tomas if (bh != &bhs) 974c9de560dSAlex Tomas kfree(bh); 975c9de560dSAlex Tomas } 976c9de560dSAlex Tomas return err; 977c9de560dSAlex Tomas } 978c9de560dSAlex Tomas 9798a57d9d6SCurt Wohlgemuth /* 9802de8807bSAmir Goldstein * Lock the buddy and bitmap pages. This make sure other parallel init_group 9812de8807bSAmir Goldstein * on the same buddy page doesn't happen whild holding the buddy page lock. 9822de8807bSAmir Goldstein * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap 9832de8807bSAmir Goldstein * are on the same page e4b->bd_buddy_page is NULL and return value is 0. 984eee4adc7SEric Sandeen */ 9852de8807bSAmir Goldstein static int ext4_mb_get_buddy_page_lock(struct super_block *sb, 986adb7ef60SKonstantin Khlebnikov ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp) 987eee4adc7SEric Sandeen { 9882de8807bSAmir Goldstein struct inode *inode = EXT4_SB(sb)->s_buddy_cache; 9892de8807bSAmir Goldstein int block, pnum, poff; 990eee4adc7SEric Sandeen int blocks_per_page; 9912de8807bSAmir Goldstein struct page *page; 9922de8807bSAmir Goldstein 9932de8807bSAmir Goldstein e4b->bd_buddy_page = NULL; 9942de8807bSAmir Goldstein e4b->bd_bitmap_page = NULL; 995eee4adc7SEric Sandeen 99609cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / sb->s_blocksize; 997eee4adc7SEric Sandeen /* 998eee4adc7SEric Sandeen * the buddy cache inode stores the block bitmap 999eee4adc7SEric Sandeen * and buddy information in consecutive blocks. 1000eee4adc7SEric Sandeen * So for each group we need two blocks. 1001eee4adc7SEric Sandeen */ 1002eee4adc7SEric Sandeen block = group * 2; 1003eee4adc7SEric Sandeen pnum = block / blocks_per_page; 10042de8807bSAmir Goldstein poff = block % blocks_per_page; 1005adb7ef60SKonstantin Khlebnikov page = find_or_create_page(inode->i_mapping, pnum, gfp); 10062de8807bSAmir Goldstein if (!page) 1007c57ab39bSYounger Liu return -ENOMEM; 10082de8807bSAmir Goldstein BUG_ON(page->mapping != inode->i_mapping); 10092de8807bSAmir Goldstein e4b->bd_bitmap_page = page; 10102de8807bSAmir Goldstein e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1011eee4adc7SEric Sandeen 10122de8807bSAmir Goldstein if (blocks_per_page >= 2) { 10132de8807bSAmir Goldstein /* buddy and bitmap are on the same page */ 10142de8807bSAmir Goldstein return 0; 1015eee4adc7SEric Sandeen } 1016eee4adc7SEric Sandeen 10172de8807bSAmir Goldstein block++; 1018eee4adc7SEric Sandeen pnum = block / blocks_per_page; 1019adb7ef60SKonstantin Khlebnikov page = find_or_create_page(inode->i_mapping, pnum, gfp); 10202de8807bSAmir Goldstein if (!page) 1021c57ab39bSYounger Liu return -ENOMEM; 10222de8807bSAmir Goldstein BUG_ON(page->mapping != inode->i_mapping); 10232de8807bSAmir Goldstein e4b->bd_buddy_page = page; 10242de8807bSAmir Goldstein return 0; 1025eee4adc7SEric Sandeen } 1026eee4adc7SEric Sandeen 10272de8807bSAmir Goldstein static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) 10282de8807bSAmir Goldstein { 10292de8807bSAmir Goldstein if (e4b->bd_bitmap_page) { 10302de8807bSAmir Goldstein unlock_page(e4b->bd_bitmap_page); 103109cbfeafSKirill A. Shutemov put_page(e4b->bd_bitmap_page); 10322de8807bSAmir Goldstein } 10332de8807bSAmir Goldstein if (e4b->bd_buddy_page) { 10342de8807bSAmir Goldstein unlock_page(e4b->bd_buddy_page); 103509cbfeafSKirill A. Shutemov put_page(e4b->bd_buddy_page); 10362de8807bSAmir Goldstein } 1037eee4adc7SEric Sandeen } 1038eee4adc7SEric Sandeen 1039eee4adc7SEric Sandeen /* 10408a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 10418a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 10428a57d9d6SCurt Wohlgemuth * calling this routine! 10438a57d9d6SCurt Wohlgemuth */ 1044b6a758ecSAneesh Kumar K.V static noinline_for_stack 1045adb7ef60SKonstantin Khlebnikov int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp) 1046b6a758ecSAneesh Kumar K.V { 1047b6a758ecSAneesh Kumar K.V 1048b6a758ecSAneesh Kumar K.V struct ext4_group_info *this_grp; 10492de8807bSAmir Goldstein struct ext4_buddy e4b; 10502de8807bSAmir Goldstein struct page *page; 10512de8807bSAmir Goldstein int ret = 0; 1052b6a758ecSAneesh Kumar K.V 1053b10a44c3STheodore Ts'o might_sleep(); 1054b6a758ecSAneesh Kumar K.V mb_debug(1, "init group %u\n", group); 1055b6a758ecSAneesh Kumar K.V this_grp = ext4_get_group_info(sb, group); 1056b6a758ecSAneesh Kumar K.V /* 105708c3a813SAneesh Kumar K.V * This ensures that we don't reinit the buddy cache 105808c3a813SAneesh Kumar K.V * page which map to the group from which we are already 105908c3a813SAneesh Kumar K.V * allocating. If we are looking at the buddy cache we would 106008c3a813SAneesh Kumar K.V * have taken a reference using ext4_mb_load_buddy and that 10612de8807bSAmir Goldstein * would have pinned buddy page to page cache. 10622457aec6SMel Gorman * The call to ext4_mb_get_buddy_page_lock will mark the 10632457aec6SMel Gorman * page accessed. 1064b6a758ecSAneesh Kumar K.V */ 1065adb7ef60SKonstantin Khlebnikov ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp); 10662de8807bSAmir Goldstein if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { 1067b6a758ecSAneesh Kumar K.V /* 1068b6a758ecSAneesh Kumar K.V * somebody initialized the group 1069b6a758ecSAneesh Kumar K.V * return without doing anything 1070b6a758ecSAneesh Kumar K.V */ 1071b6a758ecSAneesh Kumar K.V goto err; 1072b6a758ecSAneesh Kumar K.V } 10732de8807bSAmir Goldstein 10742de8807bSAmir Goldstein page = e4b.bd_bitmap_page; 1075adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_cache(page, NULL, gfp); 10762de8807bSAmir Goldstein if (ret) 1077b6a758ecSAneesh Kumar K.V goto err; 10782de8807bSAmir Goldstein if (!PageUptodate(page)) { 1079b6a758ecSAneesh Kumar K.V ret = -EIO; 1080b6a758ecSAneesh Kumar K.V goto err; 1081b6a758ecSAneesh Kumar K.V } 1082b6a758ecSAneesh Kumar K.V 10832de8807bSAmir Goldstein if (e4b.bd_buddy_page == NULL) { 1084b6a758ecSAneesh Kumar K.V /* 1085b6a758ecSAneesh Kumar K.V * If both the bitmap and buddy are in 1086b6a758ecSAneesh Kumar K.V * the same page we don't need to force 1087b6a758ecSAneesh Kumar K.V * init the buddy 1088b6a758ecSAneesh Kumar K.V */ 10892de8807bSAmir Goldstein ret = 0; 1090b6a758ecSAneesh Kumar K.V goto err; 1091b6a758ecSAneesh Kumar K.V } 10922de8807bSAmir Goldstein /* init buddy cache */ 10932de8807bSAmir Goldstein page = e4b.bd_buddy_page; 1094adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp); 10952de8807bSAmir Goldstein if (ret) 10962de8807bSAmir Goldstein goto err; 10972de8807bSAmir Goldstein if (!PageUptodate(page)) { 1098b6a758ecSAneesh Kumar K.V ret = -EIO; 1099b6a758ecSAneesh Kumar K.V goto err; 1100b6a758ecSAneesh Kumar K.V } 1101b6a758ecSAneesh Kumar K.V err: 11022de8807bSAmir Goldstein ext4_mb_put_buddy_page_lock(&e4b); 1103b6a758ecSAneesh Kumar K.V return ret; 1104b6a758ecSAneesh Kumar K.V } 1105b6a758ecSAneesh Kumar K.V 11068a57d9d6SCurt Wohlgemuth /* 11078a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 11088a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 11098a57d9d6SCurt Wohlgemuth * calling this routine! 11108a57d9d6SCurt Wohlgemuth */ 11114ddfef7bSEric Sandeen static noinline_for_stack int 1112adb7ef60SKonstantin Khlebnikov ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group, 1113adb7ef60SKonstantin Khlebnikov struct ext4_buddy *e4b, gfp_t gfp) 1114c9de560dSAlex Tomas { 1115c9de560dSAlex Tomas int blocks_per_page; 1116c9de560dSAlex Tomas int block; 1117c9de560dSAlex Tomas int pnum; 1118c9de560dSAlex Tomas int poff; 1119c9de560dSAlex Tomas struct page *page; 1120fdf6c7a7SShen Feng int ret; 1121920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1122920313a7SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 1123920313a7SAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 1124c9de560dSAlex Tomas 1125b10a44c3STheodore Ts'o might_sleep(); 11266ba495e9STheodore Ts'o mb_debug(1, "load group %u\n", group); 1127c9de560dSAlex Tomas 112809cbfeafSKirill A. Shutemov blocks_per_page = PAGE_SIZE / sb->s_blocksize; 1129920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 1130c9de560dSAlex Tomas 1131c9de560dSAlex Tomas e4b->bd_blkbits = sb->s_blocksize_bits; 1132529da704STao Ma e4b->bd_info = grp; 1133c9de560dSAlex Tomas e4b->bd_sb = sb; 1134c9de560dSAlex Tomas e4b->bd_group = group; 1135c9de560dSAlex Tomas e4b->bd_buddy_page = NULL; 1136c9de560dSAlex Tomas e4b->bd_bitmap_page = NULL; 1137c9de560dSAlex Tomas 1138f41c0750SAneesh Kumar K.V if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1139f41c0750SAneesh Kumar K.V /* 1140f41c0750SAneesh Kumar K.V * we need full data about the group 1141f41c0750SAneesh Kumar K.V * to make a good selection 1142f41c0750SAneesh Kumar K.V */ 1143adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_group(sb, group, gfp); 1144f41c0750SAneesh Kumar K.V if (ret) 1145f41c0750SAneesh Kumar K.V return ret; 1146f41c0750SAneesh Kumar K.V } 1147f41c0750SAneesh Kumar K.V 1148c9de560dSAlex Tomas /* 1149c9de560dSAlex Tomas * the buddy cache inode stores the block bitmap 1150c9de560dSAlex Tomas * and buddy information in consecutive blocks. 1151c9de560dSAlex Tomas * So for each group we need two blocks. 1152c9de560dSAlex Tomas */ 1153c9de560dSAlex Tomas block = group * 2; 1154c9de560dSAlex Tomas pnum = block / blocks_per_page; 1155c9de560dSAlex Tomas poff = block % blocks_per_page; 1156c9de560dSAlex Tomas 1157c9de560dSAlex Tomas /* we could use find_or_create_page(), but it locks page 1158c9de560dSAlex Tomas * what we'd like to avoid in fast path ... */ 11592457aec6SMel Gorman page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1160c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1161c9de560dSAlex Tomas if (page) 1162920313a7SAneesh Kumar K.V /* 1163920313a7SAneesh Kumar K.V * drop the page reference and try 1164920313a7SAneesh Kumar K.V * to get the page with lock. If we 1165920313a7SAneesh Kumar K.V * are not uptodate that implies 1166920313a7SAneesh Kumar K.V * somebody just created the page but 1167920313a7SAneesh Kumar K.V * is yet to initialize the same. So 1168920313a7SAneesh Kumar K.V * wait for it to initialize. 1169920313a7SAneesh Kumar K.V */ 117009cbfeafSKirill A. Shutemov put_page(page); 1171adb7ef60SKonstantin Khlebnikov page = find_or_create_page(inode->i_mapping, pnum, gfp); 1172c9de560dSAlex Tomas if (page) { 1173c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 1174c9de560dSAlex Tomas if (!PageUptodate(page)) { 1175adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_cache(page, NULL, gfp); 1176fdf6c7a7SShen Feng if (ret) { 1177fdf6c7a7SShen Feng unlock_page(page); 1178fdf6c7a7SShen Feng goto err; 1179fdf6c7a7SShen Feng } 1180c9de560dSAlex Tomas mb_cmp_bitmaps(e4b, page_address(page) + 1181c9de560dSAlex Tomas (poff * sb->s_blocksize)); 1182c9de560dSAlex Tomas } 1183c9de560dSAlex Tomas unlock_page(page); 1184c9de560dSAlex Tomas } 1185c9de560dSAlex Tomas } 1186c57ab39bSYounger Liu if (page == NULL) { 1187c57ab39bSYounger Liu ret = -ENOMEM; 1188c57ab39bSYounger Liu goto err; 1189c57ab39bSYounger Liu } 1190c57ab39bSYounger Liu if (!PageUptodate(page)) { 1191fdf6c7a7SShen Feng ret = -EIO; 1192c9de560dSAlex Tomas goto err; 1193fdf6c7a7SShen Feng } 11942457aec6SMel Gorman 11952457aec6SMel Gorman /* Pages marked accessed already */ 1196c9de560dSAlex Tomas e4b->bd_bitmap_page = page; 1197c9de560dSAlex Tomas e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1198c9de560dSAlex Tomas 1199c9de560dSAlex Tomas block++; 1200c9de560dSAlex Tomas pnum = block / blocks_per_page; 1201c9de560dSAlex Tomas poff = block % blocks_per_page; 1202c9de560dSAlex Tomas 12032457aec6SMel Gorman page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED); 1204c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1205c9de560dSAlex Tomas if (page) 120609cbfeafSKirill A. Shutemov put_page(page); 1207adb7ef60SKonstantin Khlebnikov page = find_or_create_page(inode->i_mapping, pnum, gfp); 1208c9de560dSAlex Tomas if (page) { 1209c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 1210fdf6c7a7SShen Feng if (!PageUptodate(page)) { 1211adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_cache(page, e4b->bd_bitmap, 1212adb7ef60SKonstantin Khlebnikov gfp); 1213fdf6c7a7SShen Feng if (ret) { 1214fdf6c7a7SShen Feng unlock_page(page); 1215fdf6c7a7SShen Feng goto err; 1216fdf6c7a7SShen Feng } 1217fdf6c7a7SShen Feng } 1218c9de560dSAlex Tomas unlock_page(page); 1219c9de560dSAlex Tomas } 1220c9de560dSAlex Tomas } 1221c57ab39bSYounger Liu if (page == NULL) { 1222c57ab39bSYounger Liu ret = -ENOMEM; 1223c57ab39bSYounger Liu goto err; 1224c57ab39bSYounger Liu } 1225c57ab39bSYounger Liu if (!PageUptodate(page)) { 1226fdf6c7a7SShen Feng ret = -EIO; 1227c9de560dSAlex Tomas goto err; 1228fdf6c7a7SShen Feng } 12292457aec6SMel Gorman 12302457aec6SMel Gorman /* Pages marked accessed already */ 1231c9de560dSAlex Tomas e4b->bd_buddy_page = page; 1232c9de560dSAlex Tomas e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 1233c9de560dSAlex Tomas 1234c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 1235c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 1236c9de560dSAlex Tomas 1237c9de560dSAlex Tomas return 0; 1238c9de560dSAlex Tomas 1239c9de560dSAlex Tomas err: 124026626f11SYang Ruirui if (page) 124109cbfeafSKirill A. Shutemov put_page(page); 1242c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 124309cbfeafSKirill A. Shutemov put_page(e4b->bd_bitmap_page); 1244c9de560dSAlex Tomas if (e4b->bd_buddy_page) 124509cbfeafSKirill A. Shutemov put_page(e4b->bd_buddy_page); 1246c9de560dSAlex Tomas e4b->bd_buddy = NULL; 1247c9de560dSAlex Tomas e4b->bd_bitmap = NULL; 1248fdf6c7a7SShen Feng return ret; 1249c9de560dSAlex Tomas } 1250c9de560dSAlex Tomas 1251adb7ef60SKonstantin Khlebnikov static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1252adb7ef60SKonstantin Khlebnikov struct ext4_buddy *e4b) 1253adb7ef60SKonstantin Khlebnikov { 1254adb7ef60SKonstantin Khlebnikov return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS); 1255adb7ef60SKonstantin Khlebnikov } 1256adb7ef60SKonstantin Khlebnikov 1257e39e07fdSJing Zhang static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) 1258c9de560dSAlex Tomas { 1259c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 126009cbfeafSKirill A. Shutemov put_page(e4b->bd_bitmap_page); 1261c9de560dSAlex Tomas if (e4b->bd_buddy_page) 126209cbfeafSKirill A. Shutemov put_page(e4b->bd_buddy_page); 1263c9de560dSAlex Tomas } 1264c9de560dSAlex Tomas 1265c9de560dSAlex Tomas 1266c9de560dSAlex Tomas static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1267c9de560dSAlex Tomas { 1268c9de560dSAlex Tomas int order = 1; 1269b5cb316cSNicolai Stange int bb_incr = 1 << (e4b->bd_blkbits - 1); 1270c9de560dSAlex Tomas void *bb; 1271c9de560dSAlex Tomas 1272c5e8f3f3STheodore Ts'o BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 1273c9de560dSAlex Tomas BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1274c9de560dSAlex Tomas 1275c5e8f3f3STheodore Ts'o bb = e4b->bd_buddy; 1276c9de560dSAlex Tomas while (order <= e4b->bd_blkbits + 1) { 1277c9de560dSAlex Tomas block = block >> 1; 1278c9de560dSAlex Tomas if (!mb_test_bit(block, bb)) { 1279c9de560dSAlex Tomas /* this block is part of buddy of order 'order' */ 1280c9de560dSAlex Tomas return order; 1281c9de560dSAlex Tomas } 1282b5cb316cSNicolai Stange bb += bb_incr; 1283b5cb316cSNicolai Stange bb_incr >>= 1; 1284c9de560dSAlex Tomas order++; 1285c9de560dSAlex Tomas } 1286c9de560dSAlex Tomas return 0; 1287c9de560dSAlex Tomas } 1288c9de560dSAlex Tomas 1289955ce5f5SAneesh Kumar K.V static void mb_clear_bits(void *bm, int cur, int len) 1290c9de560dSAlex Tomas { 1291c9de560dSAlex Tomas __u32 *addr; 1292c9de560dSAlex Tomas 1293c9de560dSAlex Tomas len = cur + len; 1294c9de560dSAlex Tomas while (cur < len) { 1295c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1296c9de560dSAlex Tomas /* fast path: clear whole word at once */ 1297c9de560dSAlex Tomas addr = bm + (cur >> 3); 1298c9de560dSAlex Tomas *addr = 0; 1299c9de560dSAlex Tomas cur += 32; 1300c9de560dSAlex Tomas continue; 1301c9de560dSAlex Tomas } 1302e8134b27SAneesh Kumar K.V mb_clear_bit(cur, bm); 1303c9de560dSAlex Tomas cur++; 1304c9de560dSAlex Tomas } 1305c9de560dSAlex Tomas } 1306c9de560dSAlex Tomas 1307eabe0444SAndrey Sidorov /* clear bits in given range 1308eabe0444SAndrey Sidorov * will return first found zero bit if any, -1 otherwise 1309eabe0444SAndrey Sidorov */ 1310eabe0444SAndrey Sidorov static int mb_test_and_clear_bits(void *bm, int cur, int len) 1311eabe0444SAndrey Sidorov { 1312eabe0444SAndrey Sidorov __u32 *addr; 1313eabe0444SAndrey Sidorov int zero_bit = -1; 1314eabe0444SAndrey Sidorov 1315eabe0444SAndrey Sidorov len = cur + len; 1316eabe0444SAndrey Sidorov while (cur < len) { 1317eabe0444SAndrey Sidorov if ((cur & 31) == 0 && (len - cur) >= 32) { 1318eabe0444SAndrey Sidorov /* fast path: clear whole word at once */ 1319eabe0444SAndrey Sidorov addr = bm + (cur >> 3); 1320eabe0444SAndrey Sidorov if (*addr != (__u32)(-1) && zero_bit == -1) 1321eabe0444SAndrey Sidorov zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0); 1322eabe0444SAndrey Sidorov *addr = 0; 1323eabe0444SAndrey Sidorov cur += 32; 1324eabe0444SAndrey Sidorov continue; 1325eabe0444SAndrey Sidorov } 1326eabe0444SAndrey Sidorov if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1) 1327eabe0444SAndrey Sidorov zero_bit = cur; 1328eabe0444SAndrey Sidorov cur++; 1329eabe0444SAndrey Sidorov } 1330eabe0444SAndrey Sidorov 1331eabe0444SAndrey Sidorov return zero_bit; 1332eabe0444SAndrey Sidorov } 1333eabe0444SAndrey Sidorov 1334c3e94d1dSYongqiang Yang void ext4_set_bits(void *bm, int cur, int len) 1335c9de560dSAlex Tomas { 1336c9de560dSAlex Tomas __u32 *addr; 1337c9de560dSAlex Tomas 1338c9de560dSAlex Tomas len = cur + len; 1339c9de560dSAlex Tomas while (cur < len) { 1340c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1341c9de560dSAlex Tomas /* fast path: set whole word at once */ 1342c9de560dSAlex Tomas addr = bm + (cur >> 3); 1343c9de560dSAlex Tomas *addr = 0xffffffff; 1344c9de560dSAlex Tomas cur += 32; 1345c9de560dSAlex Tomas continue; 1346c9de560dSAlex Tomas } 1347e8134b27SAneesh Kumar K.V mb_set_bit(cur, bm); 1348c9de560dSAlex Tomas cur++; 1349c9de560dSAlex Tomas } 1350c9de560dSAlex Tomas } 1351c9de560dSAlex Tomas 1352eabe0444SAndrey Sidorov /* 1353eabe0444SAndrey Sidorov * _________________________________________________________________ */ 1354eabe0444SAndrey Sidorov 1355eabe0444SAndrey Sidorov static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side) 1356eabe0444SAndrey Sidorov { 1357eabe0444SAndrey Sidorov if (mb_test_bit(*bit + side, bitmap)) { 1358eabe0444SAndrey Sidorov mb_clear_bit(*bit, bitmap); 1359eabe0444SAndrey Sidorov (*bit) -= side; 1360eabe0444SAndrey Sidorov return 1; 1361eabe0444SAndrey Sidorov } 1362eabe0444SAndrey Sidorov else { 1363eabe0444SAndrey Sidorov (*bit) += side; 1364eabe0444SAndrey Sidorov mb_set_bit(*bit, bitmap); 1365eabe0444SAndrey Sidorov return -1; 1366eabe0444SAndrey Sidorov } 1367eabe0444SAndrey Sidorov } 1368eabe0444SAndrey Sidorov 1369eabe0444SAndrey Sidorov static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last) 1370eabe0444SAndrey Sidorov { 1371eabe0444SAndrey Sidorov int max; 1372eabe0444SAndrey Sidorov int order = 1; 1373eabe0444SAndrey Sidorov void *buddy = mb_find_buddy(e4b, order, &max); 1374eabe0444SAndrey Sidorov 1375eabe0444SAndrey Sidorov while (buddy) { 1376eabe0444SAndrey Sidorov void *buddy2; 1377eabe0444SAndrey Sidorov 1378eabe0444SAndrey Sidorov /* Bits in range [first; last] are known to be set since 1379eabe0444SAndrey Sidorov * corresponding blocks were allocated. Bits in range 1380eabe0444SAndrey Sidorov * (first; last) will stay set because they form buddies on 1381eabe0444SAndrey Sidorov * upper layer. We just deal with borders if they don't 1382eabe0444SAndrey Sidorov * align with upper layer and then go up. 1383eabe0444SAndrey Sidorov * Releasing entire group is all about clearing 1384eabe0444SAndrey Sidorov * single bit of highest order buddy. 1385eabe0444SAndrey Sidorov */ 1386eabe0444SAndrey Sidorov 1387eabe0444SAndrey Sidorov /* Example: 1388eabe0444SAndrey Sidorov * --------------------------------- 1389eabe0444SAndrey Sidorov * | 1 | 1 | 1 | 1 | 1390eabe0444SAndrey Sidorov * --------------------------------- 1391eabe0444SAndrey Sidorov * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1392eabe0444SAndrey Sidorov * --------------------------------- 1393eabe0444SAndrey Sidorov * 0 1 2 3 4 5 6 7 1394eabe0444SAndrey Sidorov * \_____________________/ 1395eabe0444SAndrey Sidorov * 1396eabe0444SAndrey Sidorov * Neither [1] nor [6] is aligned to above layer. 1397eabe0444SAndrey Sidorov * Left neighbour [0] is free, so mark it busy, 1398eabe0444SAndrey Sidorov * decrease bb_counters and extend range to 1399eabe0444SAndrey Sidorov * [0; 6] 1400eabe0444SAndrey Sidorov * Right neighbour [7] is busy. It can't be coaleasced with [6], so 1401eabe0444SAndrey Sidorov * mark [6] free, increase bb_counters and shrink range to 1402eabe0444SAndrey Sidorov * [0; 5]. 1403eabe0444SAndrey Sidorov * Then shift range to [0; 2], go up and do the same. 1404eabe0444SAndrey Sidorov */ 1405eabe0444SAndrey Sidorov 1406eabe0444SAndrey Sidorov 1407eabe0444SAndrey Sidorov if (first & 1) 1408eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1); 1409eabe0444SAndrey Sidorov if (!(last & 1)) 1410eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1); 1411eabe0444SAndrey Sidorov if (first > last) 1412eabe0444SAndrey Sidorov break; 1413eabe0444SAndrey Sidorov order++; 1414eabe0444SAndrey Sidorov 1415eabe0444SAndrey Sidorov if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) { 1416eabe0444SAndrey Sidorov mb_clear_bits(buddy, first, last - first + 1); 1417eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[order - 1] += last - first + 1; 1418eabe0444SAndrey Sidorov break; 1419eabe0444SAndrey Sidorov } 1420eabe0444SAndrey Sidorov first >>= 1; 1421eabe0444SAndrey Sidorov last >>= 1; 1422eabe0444SAndrey Sidorov buddy = buddy2; 1423eabe0444SAndrey Sidorov } 1424eabe0444SAndrey Sidorov } 1425eabe0444SAndrey Sidorov 14267e5a8cddSShen Feng static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1427c9de560dSAlex Tomas int first, int count) 1428c9de560dSAlex Tomas { 1429eabe0444SAndrey Sidorov int left_is_free = 0; 1430eabe0444SAndrey Sidorov int right_is_free = 0; 1431eabe0444SAndrey Sidorov int block; 1432eabe0444SAndrey Sidorov int last = first + count - 1; 1433c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 1434c9de560dSAlex Tomas 1435c99d1e6eSTheodore Ts'o if (WARN_ON(count == 0)) 1436c99d1e6eSTheodore Ts'o return; 1437eabe0444SAndrey Sidorov BUG_ON(last >= (sb->s_blocksize << 3)); 1438bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1439163a203dSDarrick J. Wong /* Don't bother if the block group is corrupt. */ 1440163a203dSDarrick J. Wong if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) 1441163a203dSDarrick J. Wong return; 1442163a203dSDarrick J. Wong 1443c9de560dSAlex Tomas mb_check_buddy(e4b); 1444c9de560dSAlex Tomas mb_free_blocks_double(inode, e4b, first, count); 1445c9de560dSAlex Tomas 1446c9de560dSAlex Tomas e4b->bd_info->bb_free += count; 1447c9de560dSAlex Tomas if (first < e4b->bd_info->bb_first_free) 1448c9de560dSAlex Tomas e4b->bd_info->bb_first_free = first; 1449c9de560dSAlex Tomas 1450eabe0444SAndrey Sidorov /* access memory sequentially: check left neighbour, 1451eabe0444SAndrey Sidorov * clear range and then check right neighbour 1452eabe0444SAndrey Sidorov */ 1453c9de560dSAlex Tomas if (first != 0) 1454eabe0444SAndrey Sidorov left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap); 1455eabe0444SAndrey Sidorov block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count); 1456eabe0444SAndrey Sidorov if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0]) 1457eabe0444SAndrey Sidorov right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap); 1458c9de560dSAlex Tomas 1459eabe0444SAndrey Sidorov if (unlikely(block != -1)) { 1460e43bb4e6SNamjae Jeon struct ext4_sb_info *sbi = EXT4_SB(sb); 1461c9de560dSAlex Tomas ext4_fsblk_t blocknr; 14625661bd68SAkinobu Mita 14635661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 146453accfa9STheodore Ts'o blocknr += EXT4_C2B(EXT4_SB(sb), block); 14655d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 1466e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 1467e29136f8STheodore Ts'o blocknr, 1468e29136f8STheodore Ts'o "freeing already freed block " 1469163a203dSDarrick J. Wong "(bit %u); block bitmap corrupt.", 1470163a203dSDarrick J. Wong block); 1471e43bb4e6SNamjae Jeon if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)) 1472e43bb4e6SNamjae Jeon percpu_counter_sub(&sbi->s_freeclusters_counter, 1473e43bb4e6SNamjae Jeon e4b->bd_info->bb_free); 1474163a203dSDarrick J. Wong /* Mark the block group as corrupt. */ 1475163a203dSDarrick J. Wong set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, 1476163a203dSDarrick J. Wong &e4b->bd_info->bb_state); 1477eabe0444SAndrey Sidorov mb_regenerate_buddy(e4b); 1478eabe0444SAndrey Sidorov goto done; 1479c9de560dSAlex Tomas } 1480c9de560dSAlex Tomas 1481eabe0444SAndrey Sidorov /* let's maintain fragments counter */ 1482eabe0444SAndrey Sidorov if (left_is_free && right_is_free) 1483eabe0444SAndrey Sidorov e4b->bd_info->bb_fragments--; 1484eabe0444SAndrey Sidorov else if (!left_is_free && !right_is_free) 1485eabe0444SAndrey Sidorov e4b->bd_info->bb_fragments++; 1486c9de560dSAlex Tomas 1487eabe0444SAndrey Sidorov /* buddy[0] == bd_bitmap is a special case, so handle 1488eabe0444SAndrey Sidorov * it right away and let mb_buddy_mark_free stay free of 1489eabe0444SAndrey Sidorov * zero order checks. 1490eabe0444SAndrey Sidorov * Check if neighbours are to be coaleasced, 1491eabe0444SAndrey Sidorov * adjust bitmap bb_counters and borders appropriately. 1492eabe0444SAndrey Sidorov */ 1493eabe0444SAndrey Sidorov if (first & 1) { 1494eabe0444SAndrey Sidorov first += !left_is_free; 1495eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1; 1496c9de560dSAlex Tomas } 1497eabe0444SAndrey Sidorov if (!(last & 1)) { 1498eabe0444SAndrey Sidorov last -= !right_is_free; 1499eabe0444SAndrey Sidorov e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1; 1500c9de560dSAlex Tomas } 1501eabe0444SAndrey Sidorov 1502eabe0444SAndrey Sidorov if (first <= last) 1503eabe0444SAndrey Sidorov mb_buddy_mark_free(e4b, first >> 1, last >> 1); 1504eabe0444SAndrey Sidorov 1505eabe0444SAndrey Sidorov done: 15068a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, e4b->bd_info); 1507c9de560dSAlex Tomas mb_check_buddy(e4b); 1508c9de560dSAlex Tomas } 1509c9de560dSAlex Tomas 151015c006a2SRobin Dong static int mb_find_extent(struct ext4_buddy *e4b, int block, 1511c9de560dSAlex Tomas int needed, struct ext4_free_extent *ex) 1512c9de560dSAlex Tomas { 1513c9de560dSAlex Tomas int next = block; 151415c006a2SRobin Dong int max, order; 1515c9de560dSAlex Tomas void *buddy; 1516c9de560dSAlex Tomas 1517bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1518c9de560dSAlex Tomas BUG_ON(ex == NULL); 1519c9de560dSAlex Tomas 152015c006a2SRobin Dong buddy = mb_find_buddy(e4b, 0, &max); 1521c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1522c9de560dSAlex Tomas BUG_ON(block >= max); 1523c9de560dSAlex Tomas if (mb_test_bit(block, buddy)) { 1524c9de560dSAlex Tomas ex->fe_len = 0; 1525c9de560dSAlex Tomas ex->fe_start = 0; 1526c9de560dSAlex Tomas ex->fe_group = 0; 1527c9de560dSAlex Tomas return 0; 1528c9de560dSAlex Tomas } 1529c9de560dSAlex Tomas 1530c9de560dSAlex Tomas /* find actual order */ 1531c9de560dSAlex Tomas order = mb_find_order_for_block(e4b, block); 1532c9de560dSAlex Tomas block = block >> order; 1533c9de560dSAlex Tomas 1534c9de560dSAlex Tomas ex->fe_len = 1 << order; 1535c9de560dSAlex Tomas ex->fe_start = block << order; 1536c9de560dSAlex Tomas ex->fe_group = e4b->bd_group; 1537c9de560dSAlex Tomas 1538c9de560dSAlex Tomas /* calc difference from given start */ 1539c9de560dSAlex Tomas next = next - ex->fe_start; 1540c9de560dSAlex Tomas ex->fe_len -= next; 1541c9de560dSAlex Tomas ex->fe_start += next; 1542c9de560dSAlex Tomas 1543c9de560dSAlex Tomas while (needed > ex->fe_len && 1544d8ec0c39SAlan Cox mb_find_buddy(e4b, order, &max)) { 1545c9de560dSAlex Tomas 1546c9de560dSAlex Tomas if (block + 1 >= max) 1547c9de560dSAlex Tomas break; 1548c9de560dSAlex Tomas 1549c9de560dSAlex Tomas next = (block + 1) * (1 << order); 1550c5e8f3f3STheodore Ts'o if (mb_test_bit(next, e4b->bd_bitmap)) 1551c9de560dSAlex Tomas break; 1552c9de560dSAlex Tomas 1553b051d8dcSRobin Dong order = mb_find_order_for_block(e4b, next); 1554c9de560dSAlex Tomas 1555c9de560dSAlex Tomas block = next >> order; 1556c9de560dSAlex Tomas ex->fe_len += 1 << order; 1557c9de560dSAlex Tomas } 1558c9de560dSAlex Tomas 1559c9de560dSAlex Tomas BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); 1560c9de560dSAlex Tomas return ex->fe_len; 1561c9de560dSAlex Tomas } 1562c9de560dSAlex Tomas 1563c9de560dSAlex Tomas static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 1564c9de560dSAlex Tomas { 1565c9de560dSAlex Tomas int ord; 1566c9de560dSAlex Tomas int mlen = 0; 1567c9de560dSAlex Tomas int max = 0; 1568c9de560dSAlex Tomas int cur; 1569c9de560dSAlex Tomas int start = ex->fe_start; 1570c9de560dSAlex Tomas int len = ex->fe_len; 1571c9de560dSAlex Tomas unsigned ret = 0; 1572c9de560dSAlex Tomas int len0 = len; 1573c9de560dSAlex Tomas void *buddy; 1574c9de560dSAlex Tomas 1575c9de560dSAlex Tomas BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 1576c9de560dSAlex Tomas BUG_ON(e4b->bd_group != ex->fe_group); 1577bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1578c9de560dSAlex Tomas mb_check_buddy(e4b); 1579c9de560dSAlex Tomas mb_mark_used_double(e4b, start, len); 1580c9de560dSAlex Tomas 1581c9de560dSAlex Tomas e4b->bd_info->bb_free -= len; 1582c9de560dSAlex Tomas if (e4b->bd_info->bb_first_free == start) 1583c9de560dSAlex Tomas e4b->bd_info->bb_first_free += len; 1584c9de560dSAlex Tomas 1585c9de560dSAlex Tomas /* let's maintain fragments counter */ 1586c9de560dSAlex Tomas if (start != 0) 1587c5e8f3f3STheodore Ts'o mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); 1588c9de560dSAlex Tomas if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 1589c5e8f3f3STheodore Ts'o max = !mb_test_bit(start + len, e4b->bd_bitmap); 1590c9de560dSAlex Tomas if (mlen && max) 1591c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1592c9de560dSAlex Tomas else if (!mlen && !max) 1593c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1594c9de560dSAlex Tomas 1595c9de560dSAlex Tomas /* let's maintain buddy itself */ 1596c9de560dSAlex Tomas while (len) { 1597c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, start); 1598c9de560dSAlex Tomas 1599c9de560dSAlex Tomas if (((start >> ord) << ord) == start && len >= (1 << ord)) { 1600c9de560dSAlex Tomas /* the whole chunk may be allocated at once! */ 1601c9de560dSAlex Tomas mlen = 1 << ord; 1602c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1603c9de560dSAlex Tomas BUG_ON((start >> ord) >= max); 1604c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1605c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1606c9de560dSAlex Tomas start += mlen; 1607c9de560dSAlex Tomas len -= mlen; 1608c9de560dSAlex Tomas BUG_ON(len < 0); 1609c9de560dSAlex Tomas continue; 1610c9de560dSAlex Tomas } 1611c9de560dSAlex Tomas 1612c9de560dSAlex Tomas /* store for history */ 1613c9de560dSAlex Tomas if (ret == 0) 1614c9de560dSAlex Tomas ret = len | (ord << 16); 1615c9de560dSAlex Tomas 1616c9de560dSAlex Tomas /* we have to split large buddy */ 1617c9de560dSAlex Tomas BUG_ON(ord <= 0); 1618c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1619c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1620c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1621c9de560dSAlex Tomas 1622c9de560dSAlex Tomas ord--; 1623c9de560dSAlex Tomas cur = (start >> ord) & ~1U; 1624c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1625c9de560dSAlex Tomas mb_clear_bit(cur, buddy); 1626c9de560dSAlex Tomas mb_clear_bit(cur + 1, buddy); 1627c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1628c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1629c9de560dSAlex Tomas } 16308a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); 1631c9de560dSAlex Tomas 1632c5e8f3f3STheodore Ts'o ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); 1633c9de560dSAlex Tomas mb_check_buddy(e4b); 1634c9de560dSAlex Tomas 1635c9de560dSAlex Tomas return ret; 1636c9de560dSAlex Tomas } 1637c9de560dSAlex Tomas 1638c9de560dSAlex Tomas /* 1639c9de560dSAlex Tomas * Must be called under group lock! 1640c9de560dSAlex Tomas */ 1641c9de560dSAlex Tomas static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 1642c9de560dSAlex Tomas struct ext4_buddy *e4b) 1643c9de560dSAlex Tomas { 1644c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1645c9de560dSAlex Tomas int ret; 1646c9de560dSAlex Tomas 1647c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 1648c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1649c9de560dSAlex Tomas 1650c9de560dSAlex Tomas ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 1651c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 1652c9de560dSAlex Tomas ret = mb_mark_used(e4b, &ac->ac_b_ex); 1653c9de560dSAlex Tomas 1654c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 1655c9de560dSAlex Tomas * allocated blocks for history */ 1656c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 1657c9de560dSAlex Tomas 1658c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 1659c9de560dSAlex Tomas ac->ac_tail = ret & 0xffff; 1660c9de560dSAlex Tomas ac->ac_buddy = ret >> 16; 1661c9de560dSAlex Tomas 1662c3a326a6SAneesh Kumar K.V /* 1663c3a326a6SAneesh Kumar K.V * take the page reference. We want the page to be pinned 1664c3a326a6SAneesh Kumar K.V * so that we don't get a ext4_mb_init_cache_call for this 1665c3a326a6SAneesh Kumar K.V * group until we update the bitmap. That would mean we 1666c3a326a6SAneesh Kumar K.V * double allocate blocks. The reference is dropped 1667c3a326a6SAneesh Kumar K.V * in ext4_mb_release_context 1668c3a326a6SAneesh Kumar K.V */ 1669c9de560dSAlex Tomas ac->ac_bitmap_page = e4b->bd_bitmap_page; 1670c9de560dSAlex Tomas get_page(ac->ac_bitmap_page); 1671c9de560dSAlex Tomas ac->ac_buddy_page = e4b->bd_buddy_page; 1672c9de560dSAlex Tomas get_page(ac->ac_buddy_page); 1673c9de560dSAlex Tomas /* store last allocated for subsequent stream allocation */ 16744ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 1675c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 1676c9de560dSAlex Tomas sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 1677c9de560dSAlex Tomas sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 1678c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 1679c9de560dSAlex Tomas } 1680c9de560dSAlex Tomas } 1681c9de560dSAlex Tomas 1682c9de560dSAlex Tomas /* 1683c9de560dSAlex Tomas * regular allocator, for general purposes allocation 1684c9de560dSAlex Tomas */ 1685c9de560dSAlex Tomas 1686c9de560dSAlex Tomas static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 1687c9de560dSAlex Tomas struct ext4_buddy *e4b, 1688c9de560dSAlex Tomas int finish_group) 1689c9de560dSAlex Tomas { 1690c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1691c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1692c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1693c9de560dSAlex Tomas struct ext4_free_extent ex; 1694c9de560dSAlex Tomas int max; 1695c9de560dSAlex Tomas 1696032115fcSAneesh Kumar K.V if (ac->ac_status == AC_STATUS_FOUND) 1697032115fcSAneesh Kumar K.V return; 1698c9de560dSAlex Tomas /* 1699c9de560dSAlex Tomas * We don't want to scan for a whole year 1700c9de560dSAlex Tomas */ 1701c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan && 1702c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1703c9de560dSAlex Tomas ac->ac_status = AC_STATUS_BREAK; 1704c9de560dSAlex Tomas return; 1705c9de560dSAlex Tomas } 1706c9de560dSAlex Tomas 1707c9de560dSAlex Tomas /* 1708c9de560dSAlex Tomas * Haven't found good chunk so far, let's continue 1709c9de560dSAlex Tomas */ 1710c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) 1711c9de560dSAlex Tomas return; 1712c9de560dSAlex Tomas 1713c9de560dSAlex Tomas if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 1714c9de560dSAlex Tomas && bex->fe_group == e4b->bd_group) { 1715c9de560dSAlex Tomas /* recheck chunk's availability - we don't know 1716c9de560dSAlex Tomas * when it was found (within this lock-unlock 1717c9de560dSAlex Tomas * period or not) */ 171815c006a2SRobin Dong max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex); 1719c9de560dSAlex Tomas if (max >= gex->fe_len) { 1720c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1721c9de560dSAlex Tomas return; 1722c9de560dSAlex Tomas } 1723c9de560dSAlex Tomas } 1724c9de560dSAlex Tomas } 1725c9de560dSAlex Tomas 1726c9de560dSAlex Tomas /* 1727c9de560dSAlex Tomas * The routine checks whether found extent is good enough. If it is, 1728c9de560dSAlex Tomas * then the extent gets marked used and flag is set to the context 1729c9de560dSAlex Tomas * to stop scanning. Otherwise, the extent is compared with the 1730c9de560dSAlex Tomas * previous found extent and if new one is better, then it's stored 1731c9de560dSAlex Tomas * in the context. Later, the best found extent will be used, if 1732c9de560dSAlex Tomas * mballoc can't find good enough extent. 1733c9de560dSAlex Tomas * 1734c9de560dSAlex Tomas * FIXME: real allocation policy is to be designed yet! 1735c9de560dSAlex Tomas */ 1736c9de560dSAlex Tomas static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 1737c9de560dSAlex Tomas struct ext4_free_extent *ex, 1738c9de560dSAlex Tomas struct ext4_buddy *e4b) 1739c9de560dSAlex Tomas { 1740c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1741c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1742c9de560dSAlex Tomas 1743c9de560dSAlex Tomas BUG_ON(ex->fe_len <= 0); 17447137d7a4STheodore Ts'o BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 17457137d7a4STheodore Ts'o BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 1746c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 1747c9de560dSAlex Tomas 1748c9de560dSAlex Tomas ac->ac_found++; 1749c9de560dSAlex Tomas 1750c9de560dSAlex Tomas /* 1751c9de560dSAlex Tomas * The special case - take what you catch first 1752c9de560dSAlex Tomas */ 1753c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1754c9de560dSAlex Tomas *bex = *ex; 1755c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1756c9de560dSAlex Tomas return; 1757c9de560dSAlex Tomas } 1758c9de560dSAlex Tomas 1759c9de560dSAlex Tomas /* 1760c9de560dSAlex Tomas * Let's check whether the chuck is good enough 1761c9de560dSAlex Tomas */ 1762c9de560dSAlex Tomas if (ex->fe_len == gex->fe_len) { 1763c9de560dSAlex Tomas *bex = *ex; 1764c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1765c9de560dSAlex Tomas return; 1766c9de560dSAlex Tomas } 1767c9de560dSAlex Tomas 1768c9de560dSAlex Tomas /* 1769c9de560dSAlex Tomas * If this is first found extent, just store it in the context 1770c9de560dSAlex Tomas */ 1771c9de560dSAlex Tomas if (bex->fe_len == 0) { 1772c9de560dSAlex Tomas *bex = *ex; 1773c9de560dSAlex Tomas return; 1774c9de560dSAlex Tomas } 1775c9de560dSAlex Tomas 1776c9de560dSAlex Tomas /* 1777c9de560dSAlex Tomas * If new found extent is better, store it in the context 1778c9de560dSAlex Tomas */ 1779c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) { 1780c9de560dSAlex Tomas /* if the request isn't satisfied, any found extent 1781c9de560dSAlex Tomas * larger than previous best one is better */ 1782c9de560dSAlex Tomas if (ex->fe_len > bex->fe_len) 1783c9de560dSAlex Tomas *bex = *ex; 1784c9de560dSAlex Tomas } else if (ex->fe_len > gex->fe_len) { 1785c9de560dSAlex Tomas /* if the request is satisfied, then we try to find 1786c9de560dSAlex Tomas * an extent that still satisfy the request, but is 1787c9de560dSAlex Tomas * smaller than previous one */ 1788c9de560dSAlex Tomas if (ex->fe_len < bex->fe_len) 1789c9de560dSAlex Tomas *bex = *ex; 1790c9de560dSAlex Tomas } 1791c9de560dSAlex Tomas 1792c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 0); 1793c9de560dSAlex Tomas } 1794c9de560dSAlex Tomas 1795089ceeccSEric Sandeen static noinline_for_stack 1796089ceeccSEric Sandeen int ext4_mb_try_best_found(struct ext4_allocation_context *ac, 1797c9de560dSAlex Tomas struct ext4_buddy *e4b) 1798c9de560dSAlex Tomas { 1799c9de560dSAlex Tomas struct ext4_free_extent ex = ac->ac_b_ex; 1800c9de560dSAlex Tomas ext4_group_t group = ex.fe_group; 1801c9de560dSAlex Tomas int max; 1802c9de560dSAlex Tomas int err; 1803c9de560dSAlex Tomas 1804c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1805c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1806c9de560dSAlex Tomas if (err) 1807c9de560dSAlex Tomas return err; 1808c9de560dSAlex Tomas 1809c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 181015c006a2SRobin Dong max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex); 1811c9de560dSAlex Tomas 1812c9de560dSAlex Tomas if (max > 0) { 1813c9de560dSAlex Tomas ac->ac_b_ex = ex; 1814c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1815c9de560dSAlex Tomas } 1816c9de560dSAlex Tomas 1817c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1818e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 1819c9de560dSAlex Tomas 1820c9de560dSAlex Tomas return 0; 1821c9de560dSAlex Tomas } 1822c9de560dSAlex Tomas 1823089ceeccSEric Sandeen static noinline_for_stack 1824089ceeccSEric Sandeen int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 1825c9de560dSAlex Tomas struct ext4_buddy *e4b) 1826c9de560dSAlex Tomas { 1827c9de560dSAlex Tomas ext4_group_t group = ac->ac_g_ex.fe_group; 1828c9de560dSAlex Tomas int max; 1829c9de560dSAlex Tomas int err; 1830c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1831838cd0cfSYongqiang Yang struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 1832c9de560dSAlex Tomas struct ext4_free_extent ex; 1833c9de560dSAlex Tomas 1834c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) 1835c9de560dSAlex Tomas return 0; 1836838cd0cfSYongqiang Yang if (grp->bb_free == 0) 1837838cd0cfSYongqiang Yang return 0; 1838c9de560dSAlex Tomas 1839c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1840c9de560dSAlex Tomas if (err) 1841c9de560dSAlex Tomas return err; 1842c9de560dSAlex Tomas 1843163a203dSDarrick J. Wong if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) { 1844163a203dSDarrick J. Wong ext4_mb_unload_buddy(e4b); 1845163a203dSDarrick J. Wong return 0; 1846163a203dSDarrick J. Wong } 1847163a203dSDarrick J. Wong 1848c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 184915c006a2SRobin Dong max = mb_find_extent(e4b, ac->ac_g_ex.fe_start, 1850c9de560dSAlex Tomas ac->ac_g_ex.fe_len, &ex); 1851ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADFA11; /* debug value */ 1852c9de560dSAlex Tomas 1853c9de560dSAlex Tomas if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { 1854c9de560dSAlex Tomas ext4_fsblk_t start; 1855c9de560dSAlex Tomas 18565661bd68SAkinobu Mita start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + 18575661bd68SAkinobu Mita ex.fe_start; 1858c9de560dSAlex Tomas /* use do_div to get remainder (would be 64-bit modulo) */ 1859c9de560dSAlex Tomas if (do_div(start, sbi->s_stripe) == 0) { 1860c9de560dSAlex Tomas ac->ac_found++; 1861c9de560dSAlex Tomas ac->ac_b_ex = ex; 1862c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1863c9de560dSAlex Tomas } 1864c9de560dSAlex Tomas } else if (max >= ac->ac_g_ex.fe_len) { 1865c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1866c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1867c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1868c9de560dSAlex Tomas ac->ac_found++; 1869c9de560dSAlex Tomas ac->ac_b_ex = ex; 1870c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1871c9de560dSAlex Tomas } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 1872c9de560dSAlex Tomas /* Sometimes, caller may want to merge even small 1873c9de560dSAlex Tomas * number of blocks to an existing extent */ 1874c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1875c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1876c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1877c9de560dSAlex Tomas ac->ac_found++; 1878c9de560dSAlex Tomas ac->ac_b_ex = ex; 1879c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1880c9de560dSAlex Tomas } 1881c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1882e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 1883c9de560dSAlex Tomas 1884c9de560dSAlex Tomas return 0; 1885c9de560dSAlex Tomas } 1886c9de560dSAlex Tomas 1887c9de560dSAlex Tomas /* 1888c9de560dSAlex Tomas * The routine scans buddy structures (not bitmap!) from given order 1889c9de560dSAlex Tomas * to max order and tries to find big enough chunk to satisfy the req 1890c9de560dSAlex Tomas */ 1891089ceeccSEric Sandeen static noinline_for_stack 1892089ceeccSEric Sandeen void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 1893c9de560dSAlex Tomas struct ext4_buddy *e4b) 1894c9de560dSAlex Tomas { 1895c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1896c9de560dSAlex Tomas struct ext4_group_info *grp = e4b->bd_info; 1897c9de560dSAlex Tomas void *buddy; 1898c9de560dSAlex Tomas int i; 1899c9de560dSAlex Tomas int k; 1900c9de560dSAlex Tomas int max; 1901c9de560dSAlex Tomas 1902c9de560dSAlex Tomas BUG_ON(ac->ac_2order <= 0); 1903c9de560dSAlex Tomas for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { 1904c9de560dSAlex Tomas if (grp->bb_counters[i] == 0) 1905c9de560dSAlex Tomas continue; 1906c9de560dSAlex Tomas 1907c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, i, &max); 1908c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1909c9de560dSAlex Tomas 1910ffad0a44SAneesh Kumar K.V k = mb_find_next_zero_bit(buddy, max, 0); 1911c9de560dSAlex Tomas BUG_ON(k >= max); 1912c9de560dSAlex Tomas 1913c9de560dSAlex Tomas ac->ac_found++; 1914c9de560dSAlex Tomas 1915c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 1 << i; 1916c9de560dSAlex Tomas ac->ac_b_ex.fe_start = k << i; 1917c9de560dSAlex Tomas ac->ac_b_ex.fe_group = e4b->bd_group; 1918c9de560dSAlex Tomas 1919c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1920c9de560dSAlex Tomas 1921c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); 1922c9de560dSAlex Tomas 1923c9de560dSAlex Tomas if (EXT4_SB(sb)->s_mb_stats) 1924c9de560dSAlex Tomas atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 1925c9de560dSAlex Tomas 1926c9de560dSAlex Tomas break; 1927c9de560dSAlex Tomas } 1928c9de560dSAlex Tomas } 1929c9de560dSAlex Tomas 1930c9de560dSAlex Tomas /* 1931c9de560dSAlex Tomas * The routine scans the group and measures all found extents. 1932c9de560dSAlex Tomas * In order to optimize scanning, caller must pass number of 1933c9de560dSAlex Tomas * free blocks in the group, so the routine can know upper limit. 1934c9de560dSAlex Tomas */ 1935089ceeccSEric Sandeen static noinline_for_stack 1936089ceeccSEric Sandeen void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 1937c9de560dSAlex Tomas struct ext4_buddy *e4b) 1938c9de560dSAlex Tomas { 1939c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1940c5e8f3f3STheodore Ts'o void *bitmap = e4b->bd_bitmap; 1941c9de560dSAlex Tomas struct ext4_free_extent ex; 1942c9de560dSAlex Tomas int i; 1943c9de560dSAlex Tomas int free; 1944c9de560dSAlex Tomas 1945c9de560dSAlex Tomas free = e4b->bd_info->bb_free; 1946c9de560dSAlex Tomas BUG_ON(free <= 0); 1947c9de560dSAlex Tomas 1948c9de560dSAlex Tomas i = e4b->bd_info->bb_first_free; 1949c9de560dSAlex Tomas 1950c9de560dSAlex Tomas while (free && ac->ac_status == AC_STATUS_CONTINUE) { 1951ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, 19527137d7a4STheodore Ts'o EXT4_CLUSTERS_PER_GROUP(sb), i); 19537137d7a4STheodore Ts'o if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { 195426346ff6SAneesh Kumar K.V /* 1955e56eb659SAneesh Kumar K.V * IF we have corrupt bitmap, we won't find any 195626346ff6SAneesh Kumar K.V * free blocks even though group info says we 195726346ff6SAneesh Kumar K.V * we have free blocks 195826346ff6SAneesh Kumar K.V */ 1959e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 196053accfa9STheodore Ts'o "%d free clusters as per " 1961fde4d95aSTheodore Ts'o "group info. But bitmap says 0", 196226346ff6SAneesh Kumar K.V free); 1963c9de560dSAlex Tomas break; 1964c9de560dSAlex Tomas } 1965c9de560dSAlex Tomas 196615c006a2SRobin Dong mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex); 1967c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 196826346ff6SAneesh Kumar K.V if (free < ex.fe_len) { 1969e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 197053accfa9STheodore Ts'o "%d free clusters as per " 1971fde4d95aSTheodore Ts'o "group info. But got %d blocks", 197226346ff6SAneesh Kumar K.V free, ex.fe_len); 1973e56eb659SAneesh Kumar K.V /* 1974e56eb659SAneesh Kumar K.V * The number of free blocks differs. This mostly 1975e56eb659SAneesh Kumar K.V * indicate that the bitmap is corrupt. So exit 1976e56eb659SAneesh Kumar K.V * without claiming the space. 1977e56eb659SAneesh Kumar K.V */ 1978e56eb659SAneesh Kumar K.V break; 197926346ff6SAneesh Kumar K.V } 1980ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADC0DE; /* debug value */ 1981c9de560dSAlex Tomas ext4_mb_measure_extent(ac, &ex, e4b); 1982c9de560dSAlex Tomas 1983c9de560dSAlex Tomas i += ex.fe_len; 1984c9de560dSAlex Tomas free -= ex.fe_len; 1985c9de560dSAlex Tomas } 1986c9de560dSAlex Tomas 1987c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 1); 1988c9de560dSAlex Tomas } 1989c9de560dSAlex Tomas 1990c9de560dSAlex Tomas /* 1991c9de560dSAlex Tomas * This is a special case for storages like raid5 1992506bf2d8SEric Sandeen * we try to find stripe-aligned chunks for stripe-size-multiple requests 1993c9de560dSAlex Tomas */ 1994089ceeccSEric Sandeen static noinline_for_stack 1995089ceeccSEric Sandeen void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 1996c9de560dSAlex Tomas struct ext4_buddy *e4b) 1997c9de560dSAlex Tomas { 1998c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1999c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2000c5e8f3f3STheodore Ts'o void *bitmap = e4b->bd_bitmap; 2001c9de560dSAlex Tomas struct ext4_free_extent ex; 2002c9de560dSAlex Tomas ext4_fsblk_t first_group_block; 2003c9de560dSAlex Tomas ext4_fsblk_t a; 2004c9de560dSAlex Tomas ext4_grpblk_t i; 2005c9de560dSAlex Tomas int max; 2006c9de560dSAlex Tomas 2007c9de560dSAlex Tomas BUG_ON(sbi->s_stripe == 0); 2008c9de560dSAlex Tomas 2009c9de560dSAlex Tomas /* find first stripe-aligned block in group */ 20105661bd68SAkinobu Mita first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); 20115661bd68SAkinobu Mita 2012c9de560dSAlex Tomas a = first_group_block + sbi->s_stripe - 1; 2013c9de560dSAlex Tomas do_div(a, sbi->s_stripe); 2014c9de560dSAlex Tomas i = (a * sbi->s_stripe) - first_group_block; 2015c9de560dSAlex Tomas 20167137d7a4STheodore Ts'o while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { 2017c9de560dSAlex Tomas if (!mb_test_bit(i, bitmap)) { 201815c006a2SRobin Dong max = mb_find_extent(e4b, i, sbi->s_stripe, &ex); 2019c9de560dSAlex Tomas if (max >= sbi->s_stripe) { 2020c9de560dSAlex Tomas ac->ac_found++; 2021ab0c00fcSTheodore Ts'o ex.fe_logical = 0xDEADF00D; /* debug value */ 2022c9de560dSAlex Tomas ac->ac_b_ex = ex; 2023c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 2024c9de560dSAlex Tomas break; 2025c9de560dSAlex Tomas } 2026c9de560dSAlex Tomas } 2027c9de560dSAlex Tomas i += sbi->s_stripe; 2028c9de560dSAlex Tomas } 2029c9de560dSAlex Tomas } 2030c9de560dSAlex Tomas 203142ac1848SLukas Czerner /* 203242ac1848SLukas Czerner * This is now called BEFORE we load the buddy bitmap. 203342ac1848SLukas Czerner * Returns either 1 or 0 indicating that the group is either suitable 203442ac1848SLukas Czerner * for the allocation or not. In addition it can also return negative 203542ac1848SLukas Czerner * error code when something goes wrong. 203642ac1848SLukas Czerner */ 2037c9de560dSAlex Tomas static int ext4_mb_good_group(struct ext4_allocation_context *ac, 2038c9de560dSAlex Tomas ext4_group_t group, int cr) 2039c9de560dSAlex Tomas { 2040c9de560dSAlex Tomas unsigned free, fragments; 2041a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 2042c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 2043c9de560dSAlex Tomas 2044c9de560dSAlex Tomas BUG_ON(cr < 0 || cr >= 4); 20458a57d9d6SCurt Wohlgemuth 204601fc48e8STheodore Ts'o free = grp->bb_free; 204701fc48e8STheodore Ts'o if (free == 0) 204801fc48e8STheodore Ts'o return 0; 204901fc48e8STheodore Ts'o if (cr <= 2 && free < ac->ac_g_ex.fe_len) 205001fc48e8STheodore Ts'o return 0; 205101fc48e8STheodore Ts'o 2052163a203dSDarrick J. Wong if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp))) 2053163a203dSDarrick J. Wong return 0; 2054163a203dSDarrick J. Wong 20558a57d9d6SCurt Wohlgemuth /* We only do this if the grp has never been initialized */ 20568a57d9d6SCurt Wohlgemuth if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 2057adb7ef60SKonstantin Khlebnikov int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS); 20588a57d9d6SCurt Wohlgemuth if (ret) 205942ac1848SLukas Czerner return ret; 20608a57d9d6SCurt Wohlgemuth } 2061c9de560dSAlex Tomas 2062c9de560dSAlex Tomas fragments = grp->bb_fragments; 2063c9de560dSAlex Tomas if (fragments == 0) 2064c9de560dSAlex Tomas return 0; 2065c9de560dSAlex Tomas 2066c9de560dSAlex Tomas switch (cr) { 2067c9de560dSAlex Tomas case 0: 2068c9de560dSAlex Tomas BUG_ON(ac->ac_2order == 0); 2069c9de560dSAlex Tomas 2070a4912123STheodore Ts'o /* Avoid using the first bg of a flexgroup for data files */ 2071a4912123STheodore Ts'o if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 2072a4912123STheodore Ts'o (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 2073a4912123STheodore Ts'o ((group % flex_size) == 0)) 2074a4912123STheodore Ts'o return 0; 2075a4912123STheodore Ts'o 207640ae3487STheodore Ts'o if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) || 207740ae3487STheodore Ts'o (free / fragments) >= ac->ac_g_ex.fe_len) 207840ae3487STheodore Ts'o return 1; 207940ae3487STheodore Ts'o 208040ae3487STheodore Ts'o if (grp->bb_largest_free_order < ac->ac_2order) 208140ae3487STheodore Ts'o return 0; 208240ae3487STheodore Ts'o 2083c9de560dSAlex Tomas return 1; 2084c9de560dSAlex Tomas case 1: 2085c9de560dSAlex Tomas if ((free / fragments) >= ac->ac_g_ex.fe_len) 2086c9de560dSAlex Tomas return 1; 2087c9de560dSAlex Tomas break; 2088c9de560dSAlex Tomas case 2: 2089c9de560dSAlex Tomas if (free >= ac->ac_g_ex.fe_len) 2090c9de560dSAlex Tomas return 1; 2091c9de560dSAlex Tomas break; 2092c9de560dSAlex Tomas case 3: 2093c9de560dSAlex Tomas return 1; 2094c9de560dSAlex Tomas default: 2095c9de560dSAlex Tomas BUG(); 2096c9de560dSAlex Tomas } 2097c9de560dSAlex Tomas 2098c9de560dSAlex Tomas return 0; 2099c9de560dSAlex Tomas } 2100c9de560dSAlex Tomas 21014ddfef7bSEric Sandeen static noinline_for_stack int 21024ddfef7bSEric Sandeen ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 2103c9de560dSAlex Tomas { 21048df9675fSTheodore Ts'o ext4_group_t ngroups, group, i; 2105c9de560dSAlex Tomas int cr; 210642ac1848SLukas Czerner int err = 0, first_err = 0; 2107c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2108c9de560dSAlex Tomas struct super_block *sb; 2109c9de560dSAlex Tomas struct ext4_buddy e4b; 2110c9de560dSAlex Tomas 2111c9de560dSAlex Tomas sb = ac->ac_sb; 2112c9de560dSAlex Tomas sbi = EXT4_SB(sb); 21138df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 2114fb0a387dSEric Sandeen /* non-extent files are limited to low blocks/groups */ 211512e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) 2116fb0a387dSEric Sandeen ngroups = sbi->s_blockfile_groups; 2117fb0a387dSEric Sandeen 2118c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2119c9de560dSAlex Tomas 2120c9de560dSAlex Tomas /* first, try the goal */ 2121c9de560dSAlex Tomas err = ext4_mb_find_by_goal(ac, &e4b); 2122c9de560dSAlex Tomas if (err || ac->ac_status == AC_STATUS_FOUND) 2123c9de560dSAlex Tomas goto out; 2124c9de560dSAlex Tomas 2125c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2126c9de560dSAlex Tomas goto out; 2127c9de560dSAlex Tomas 2128c9de560dSAlex Tomas /* 2129c9de560dSAlex Tomas * ac->ac2_order is set only if the fe_len is a power of 2 2130c9de560dSAlex Tomas * if ac2_order is set we also set criteria to 0 so that we 2131c9de560dSAlex Tomas * try exact allocation using buddy. 2132c9de560dSAlex Tomas */ 2133c9de560dSAlex Tomas i = fls(ac->ac_g_ex.fe_len); 2134c9de560dSAlex Tomas ac->ac_2order = 0; 2135c9de560dSAlex Tomas /* 2136c9de560dSAlex Tomas * We search using buddy data only if the order of the request 2137c9de560dSAlex Tomas * is greater than equal to the sbi_s_mb_order2_reqs 2138b713a5ecSTheodore Ts'o * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 2139c9de560dSAlex Tomas */ 2140c9de560dSAlex Tomas if (i >= sbi->s_mb_order2_reqs) { 2141c9de560dSAlex Tomas /* 2142c9de560dSAlex Tomas * This should tell if fe_len is exactly power of 2 2143c9de560dSAlex Tomas */ 2144c9de560dSAlex Tomas if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) 2145c9de560dSAlex Tomas ac->ac_2order = i - 1; 2146c9de560dSAlex Tomas } 2147c9de560dSAlex Tomas 21484ba74d00STheodore Ts'o /* if stream allocation is enabled, use global goal */ 21494ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2150c9de560dSAlex Tomas /* TBD: may be hot point */ 2151c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 2152c9de560dSAlex Tomas ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 2153c9de560dSAlex Tomas ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 2154c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 2155c9de560dSAlex Tomas } 21564ba74d00STheodore Ts'o 2157c9de560dSAlex Tomas /* Let's just scan groups to find more-less suitable blocks */ 2158c9de560dSAlex Tomas cr = ac->ac_2order ? 0 : 1; 2159c9de560dSAlex Tomas /* 2160c9de560dSAlex Tomas * cr == 0 try to get exact allocation, 2161c9de560dSAlex Tomas * cr == 3 try to get anything 2162c9de560dSAlex Tomas */ 2163c9de560dSAlex Tomas repeat: 2164c9de560dSAlex Tomas for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 2165c9de560dSAlex Tomas ac->ac_criteria = cr; 2166ed8f9c75SAneesh Kumar K.V /* 2167ed8f9c75SAneesh Kumar K.V * searching for the right group start 2168ed8f9c75SAneesh Kumar K.V * from the goal value specified 2169ed8f9c75SAneesh Kumar K.V */ 2170ed8f9c75SAneesh Kumar K.V group = ac->ac_g_ex.fe_group; 2171ed8f9c75SAneesh Kumar K.V 21728df9675fSTheodore Ts'o for (i = 0; i < ngroups; group++, i++) { 217342ac1848SLukas Czerner int ret = 0; 21742ed5724dSTheodore Ts'o cond_resched(); 2175e6155736SLachlan McIlroy /* 2176e6155736SLachlan McIlroy * Artificially restricted ngroups for non-extent 2177e6155736SLachlan McIlroy * files makes group > ngroups possible on first loop. 2178e6155736SLachlan McIlroy */ 2179e6155736SLachlan McIlroy if (group >= ngroups) 2180c9de560dSAlex Tomas group = 0; 2181c9de560dSAlex Tomas 21828a57d9d6SCurt Wohlgemuth /* This now checks without needing the buddy page */ 218342ac1848SLukas Czerner ret = ext4_mb_good_group(ac, group, cr); 218442ac1848SLukas Czerner if (ret <= 0) { 218542ac1848SLukas Czerner if (!first_err) 218642ac1848SLukas Czerner first_err = ret; 2187c9de560dSAlex Tomas continue; 218842ac1848SLukas Czerner } 2189c9de560dSAlex Tomas 2190c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2191c9de560dSAlex Tomas if (err) 2192c9de560dSAlex Tomas goto out; 2193c9de560dSAlex Tomas 2194c9de560dSAlex Tomas ext4_lock_group(sb, group); 21958a57d9d6SCurt Wohlgemuth 21968a57d9d6SCurt Wohlgemuth /* 21978a57d9d6SCurt Wohlgemuth * We need to check again after locking the 21988a57d9d6SCurt Wohlgemuth * block group 21998a57d9d6SCurt Wohlgemuth */ 220042ac1848SLukas Czerner ret = ext4_mb_good_group(ac, group, cr); 220142ac1848SLukas Czerner if (ret <= 0) { 2202c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2203e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 220442ac1848SLukas Czerner if (!first_err) 220542ac1848SLukas Czerner first_err = ret; 2206c9de560dSAlex Tomas continue; 2207c9de560dSAlex Tomas } 2208c9de560dSAlex Tomas 2209c9de560dSAlex Tomas ac->ac_groups_scanned++; 221040ae3487STheodore Ts'o if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2) 2211c9de560dSAlex Tomas ext4_mb_simple_scan_group(ac, &e4b); 2212506bf2d8SEric Sandeen else if (cr == 1 && sbi->s_stripe && 2213506bf2d8SEric Sandeen !(ac->ac_g_ex.fe_len % sbi->s_stripe)) 2214c9de560dSAlex Tomas ext4_mb_scan_aligned(ac, &e4b); 2215c9de560dSAlex Tomas else 2216c9de560dSAlex Tomas ext4_mb_complex_scan_group(ac, &e4b); 2217c9de560dSAlex Tomas 2218c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2219e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2220c9de560dSAlex Tomas 2221c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_CONTINUE) 2222c9de560dSAlex Tomas break; 2223c9de560dSAlex Tomas } 2224c9de560dSAlex Tomas } 2225c9de560dSAlex Tomas 2226c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2227c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2228c9de560dSAlex Tomas /* 2229c9de560dSAlex Tomas * We've been searching too long. Let's try to allocate 2230c9de560dSAlex Tomas * the best chunk we've found so far 2231c9de560dSAlex Tomas */ 2232c9de560dSAlex Tomas 2233c9de560dSAlex Tomas ext4_mb_try_best_found(ac, &e4b); 2234c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_FOUND) { 2235c9de560dSAlex Tomas /* 2236c9de560dSAlex Tomas * Someone more lucky has already allocated it. 2237c9de560dSAlex Tomas * The only thing we can do is just take first 2238c9de560dSAlex Tomas * found block(s) 2239c9de560dSAlex Tomas printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); 2240c9de560dSAlex Tomas */ 2241c9de560dSAlex Tomas ac->ac_b_ex.fe_group = 0; 2242c9de560dSAlex Tomas ac->ac_b_ex.fe_start = 0; 2243c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 0; 2244c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 2245c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_FIRST; 2246c9de560dSAlex Tomas cr = 3; 2247c9de560dSAlex Tomas atomic_inc(&sbi->s_mb_lost_chunks); 2248c9de560dSAlex Tomas goto repeat; 2249c9de560dSAlex Tomas } 2250c9de560dSAlex Tomas } 2251c9de560dSAlex Tomas out: 225242ac1848SLukas Czerner if (!err && ac->ac_status != AC_STATUS_FOUND && first_err) 225342ac1848SLukas Czerner err = first_err; 2254c9de560dSAlex Tomas return err; 2255c9de560dSAlex Tomas } 2256c9de560dSAlex Tomas 2257c9de560dSAlex Tomas static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2258c9de560dSAlex Tomas { 2259c9de560dSAlex Tomas struct super_block *sb = seq->private; 2260c9de560dSAlex Tomas ext4_group_t group; 2261c9de560dSAlex Tomas 22628df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2263c9de560dSAlex Tomas return NULL; 2264c9de560dSAlex Tomas group = *pos + 1; 2265a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2266c9de560dSAlex Tomas } 2267c9de560dSAlex Tomas 2268c9de560dSAlex Tomas static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2269c9de560dSAlex Tomas { 2270c9de560dSAlex Tomas struct super_block *sb = seq->private; 2271c9de560dSAlex Tomas ext4_group_t group; 2272c9de560dSAlex Tomas 2273c9de560dSAlex Tomas ++*pos; 22748df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2275c9de560dSAlex Tomas return NULL; 2276c9de560dSAlex Tomas group = *pos + 1; 2277a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2278c9de560dSAlex Tomas } 2279c9de560dSAlex Tomas 2280c9de560dSAlex Tomas static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2281c9de560dSAlex Tomas { 2282c9de560dSAlex Tomas struct super_block *sb = seq->private; 2283a9df9a49STheodore Ts'o ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2284c9de560dSAlex Tomas int i; 22851c8457caSAditya Kali int err, buddy_loaded = 0; 2286c9de560dSAlex Tomas struct ext4_buddy e4b; 22871c8457caSAditya Kali struct ext4_group_info *grinfo; 2288c9de560dSAlex Tomas struct sg { 2289c9de560dSAlex Tomas struct ext4_group_info info; 2290a36b4498SEric Sandeen ext4_grpblk_t counters[16]; 2291c9de560dSAlex Tomas } sg; 2292c9de560dSAlex Tomas 2293c9de560dSAlex Tomas group--; 2294c9de560dSAlex Tomas if (group == 0) 229597b4af2fSRasmus Villemoes seq_puts(seq, "#group: free frags first [" 229697b4af2fSRasmus Villemoes " 2^0 2^1 2^2 2^3 2^4 2^5 2^6 " 2297802cf1f9SHuaitong Han " 2^7 2^8 2^9 2^10 2^11 2^12 2^13 ]\n"); 2298c9de560dSAlex Tomas 2299c9de560dSAlex Tomas i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 2300c9de560dSAlex Tomas sizeof(struct ext4_group_info); 23011c8457caSAditya Kali grinfo = ext4_get_group_info(sb, group); 23021c8457caSAditya Kali /* Load the group info in memory only if not already loaded. */ 23031c8457caSAditya Kali if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) { 2304c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2305c9de560dSAlex Tomas if (err) { 2306a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: I/O error\n", group); 2307c9de560dSAlex Tomas return 0; 2308c9de560dSAlex Tomas } 23091c8457caSAditya Kali buddy_loaded = 1; 23101c8457caSAditya Kali } 23111c8457caSAditya Kali 2312c9de560dSAlex Tomas memcpy(&sg, ext4_get_group_info(sb, group), i); 23131c8457caSAditya Kali 23141c8457caSAditya Kali if (buddy_loaded) 2315e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2316c9de560dSAlex Tomas 2317a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 2318c9de560dSAlex Tomas sg.info.bb_fragments, sg.info.bb_first_free); 2319c9de560dSAlex Tomas for (i = 0; i <= 13; i++) 2320c9de560dSAlex Tomas seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? 2321c9de560dSAlex Tomas sg.info.bb_counters[i] : 0); 2322c9de560dSAlex Tomas seq_printf(seq, " ]\n"); 2323c9de560dSAlex Tomas 2324c9de560dSAlex Tomas return 0; 2325c9de560dSAlex Tomas } 2326c9de560dSAlex Tomas 2327c9de560dSAlex Tomas static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 2328c9de560dSAlex Tomas { 2329c9de560dSAlex Tomas } 2330c9de560dSAlex Tomas 23317f1346a9STobias Klauser static const struct seq_operations ext4_mb_seq_groups_ops = { 2332c9de560dSAlex Tomas .start = ext4_mb_seq_groups_start, 2333c9de560dSAlex Tomas .next = ext4_mb_seq_groups_next, 2334c9de560dSAlex Tomas .stop = ext4_mb_seq_groups_stop, 2335c9de560dSAlex Tomas .show = ext4_mb_seq_groups_show, 2336c9de560dSAlex Tomas }; 2337c9de560dSAlex Tomas 2338c9de560dSAlex Tomas static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) 2339c9de560dSAlex Tomas { 2340d9dda78bSAl Viro struct super_block *sb = PDE_DATA(inode); 2341c9de560dSAlex Tomas int rc; 2342c9de560dSAlex Tomas 2343c9de560dSAlex Tomas rc = seq_open(file, &ext4_mb_seq_groups_ops); 2344c9de560dSAlex Tomas if (rc == 0) { 2345a271fe85SJoe Perches struct seq_file *m = file->private_data; 2346c9de560dSAlex Tomas m->private = sb; 2347c9de560dSAlex Tomas } 2348c9de560dSAlex Tomas return rc; 2349c9de560dSAlex Tomas 2350c9de560dSAlex Tomas } 2351c9de560dSAlex Tomas 2352ebd173beSTheodore Ts'o const struct file_operations ext4_seq_mb_groups_fops = { 2353c9de560dSAlex Tomas .owner = THIS_MODULE, 2354c9de560dSAlex Tomas .open = ext4_mb_seq_groups_open, 2355c9de560dSAlex Tomas .read = seq_read, 2356c9de560dSAlex Tomas .llseek = seq_lseek, 2357c9de560dSAlex Tomas .release = seq_release, 2358c9de560dSAlex Tomas }; 2359c9de560dSAlex Tomas 2360fb1813f4SCurt Wohlgemuth static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) 2361fb1813f4SCurt Wohlgemuth { 2362fb1813f4SCurt Wohlgemuth int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 2363fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; 2364fb1813f4SCurt Wohlgemuth 2365fb1813f4SCurt Wohlgemuth BUG_ON(!cachep); 2366fb1813f4SCurt Wohlgemuth return cachep; 2367fb1813f4SCurt Wohlgemuth } 23685f21b0e6SFrederic Bohe 236928623c2fSTheodore Ts'o /* 237028623c2fSTheodore Ts'o * Allocate the top-level s_group_info array for the specified number 237128623c2fSTheodore Ts'o * of groups 237228623c2fSTheodore Ts'o */ 237328623c2fSTheodore Ts'o int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups) 237428623c2fSTheodore Ts'o { 237528623c2fSTheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(sb); 237628623c2fSTheodore Ts'o unsigned size; 237728623c2fSTheodore Ts'o struct ext4_group_info ***new_groupinfo; 237828623c2fSTheodore Ts'o 237928623c2fSTheodore Ts'o size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >> 238028623c2fSTheodore Ts'o EXT4_DESC_PER_BLOCK_BITS(sb); 238128623c2fSTheodore Ts'o if (size <= sbi->s_group_info_size) 238228623c2fSTheodore Ts'o return 0; 238328623c2fSTheodore Ts'o 238428623c2fSTheodore Ts'o size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size); 238528623c2fSTheodore Ts'o new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL); 238628623c2fSTheodore Ts'o if (!new_groupinfo) { 238728623c2fSTheodore Ts'o ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); 238828623c2fSTheodore Ts'o return -ENOMEM; 238928623c2fSTheodore Ts'o } 239028623c2fSTheodore Ts'o if (sbi->s_group_info) { 239128623c2fSTheodore Ts'o memcpy(new_groupinfo, sbi->s_group_info, 239228623c2fSTheodore Ts'o sbi->s_group_info_size * sizeof(*sbi->s_group_info)); 2393b93b41d4SAl Viro kvfree(sbi->s_group_info); 239428623c2fSTheodore Ts'o } 239528623c2fSTheodore Ts'o sbi->s_group_info = new_groupinfo; 239628623c2fSTheodore Ts'o sbi->s_group_info_size = size / sizeof(*sbi->s_group_info); 239728623c2fSTheodore Ts'o ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 239828623c2fSTheodore Ts'o sbi->s_group_info_size); 239928623c2fSTheodore Ts'o return 0; 240028623c2fSTheodore Ts'o } 240128623c2fSTheodore Ts'o 24025f21b0e6SFrederic Bohe /* Create and initialize ext4_group_info data for the given group. */ 2403920313a7SAneesh Kumar K.V int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 24045f21b0e6SFrederic Bohe struct ext4_group_desc *desc) 24055f21b0e6SFrederic Bohe { 2406fb1813f4SCurt Wohlgemuth int i; 24075f21b0e6SFrederic Bohe int metalen = 0; 24085f21b0e6SFrederic Bohe struct ext4_sb_info *sbi = EXT4_SB(sb); 24095f21b0e6SFrederic Bohe struct ext4_group_info **meta_group_info; 2410fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 24115f21b0e6SFrederic Bohe 24125f21b0e6SFrederic Bohe /* 24135f21b0e6SFrederic Bohe * First check if this group is the first of a reserved block. 24145f21b0e6SFrederic Bohe * If it's true, we have to allocate a new table of pointers 24155f21b0e6SFrederic Bohe * to ext4_group_info structures 24165f21b0e6SFrederic Bohe */ 24175f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 24185f21b0e6SFrederic Bohe metalen = sizeof(*meta_group_info) << 24195f21b0e6SFrederic Bohe EXT4_DESC_PER_BLOCK_BITS(sb); 24204fdb5543SDmitry Monakhov meta_group_info = kmalloc(metalen, GFP_NOFS); 24215f21b0e6SFrederic Bohe if (meta_group_info == NULL) { 24227f6a11e7SJoe Perches ext4_msg(sb, KERN_ERR, "can't allocate mem " 24239d8b9ec4STheodore Ts'o "for a buddy group"); 24245f21b0e6SFrederic Bohe goto exit_meta_group_info; 24255f21b0e6SFrederic Bohe } 24265f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = 24275f21b0e6SFrederic Bohe meta_group_info; 24285f21b0e6SFrederic Bohe } 24295f21b0e6SFrederic Bohe 24305f21b0e6SFrederic Bohe meta_group_info = 24315f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; 24325f21b0e6SFrederic Bohe i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 24335f21b0e6SFrederic Bohe 24344fdb5543SDmitry Monakhov meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS); 24355f21b0e6SFrederic Bohe if (meta_group_info[i] == NULL) { 24367f6a11e7SJoe Perches ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); 24375f21b0e6SFrederic Bohe goto exit_group_info; 24385f21b0e6SFrederic Bohe } 24395f21b0e6SFrederic Bohe set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 24405f21b0e6SFrederic Bohe &(meta_group_info[i]->bb_state)); 24415f21b0e6SFrederic Bohe 24425f21b0e6SFrederic Bohe /* 24435f21b0e6SFrederic Bohe * initialize bb_free to be able to skip 24445f21b0e6SFrederic Bohe * empty groups without initialization 24455f21b0e6SFrederic Bohe */ 24465f21b0e6SFrederic Bohe if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 24475f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 2448cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, group, desc); 24495f21b0e6SFrederic Bohe } else { 24505f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 2451021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc); 24525f21b0e6SFrederic Bohe } 24535f21b0e6SFrederic Bohe 24545f21b0e6SFrederic Bohe INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 2455920313a7SAneesh Kumar K.V init_rwsem(&meta_group_info[i]->alloc_sem); 245664e290ecSVenkatesh Pallipadi meta_group_info[i]->bb_free_root = RB_ROOT; 24578a57d9d6SCurt Wohlgemuth meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ 24585f21b0e6SFrederic Bohe 24595f21b0e6SFrederic Bohe #ifdef DOUBLE_CHECK 24605f21b0e6SFrederic Bohe { 24615f21b0e6SFrederic Bohe struct buffer_head *bh; 24625f21b0e6SFrederic Bohe meta_group_info[i]->bb_bitmap = 24634fdb5543SDmitry Monakhov kmalloc(sb->s_blocksize, GFP_NOFS); 24645f21b0e6SFrederic Bohe BUG_ON(meta_group_info[i]->bb_bitmap == NULL); 24655f21b0e6SFrederic Bohe bh = ext4_read_block_bitmap(sb, group); 24669008a58eSDarrick J. Wong BUG_ON(IS_ERR_OR_NULL(bh)); 24675f21b0e6SFrederic Bohe memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, 24685f21b0e6SFrederic Bohe sb->s_blocksize); 24695f21b0e6SFrederic Bohe put_bh(bh); 24705f21b0e6SFrederic Bohe } 24715f21b0e6SFrederic Bohe #endif 24725f21b0e6SFrederic Bohe 24735f21b0e6SFrederic Bohe return 0; 24745f21b0e6SFrederic Bohe 24755f21b0e6SFrederic Bohe exit_group_info: 24765f21b0e6SFrederic Bohe /* If a meta_group_info table has been allocated, release it now */ 2477caaf7a29STao Ma if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 24785f21b0e6SFrederic Bohe kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); 2479caaf7a29STao Ma sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; 2480caaf7a29STao Ma } 24815f21b0e6SFrederic Bohe exit_meta_group_info: 24825f21b0e6SFrederic Bohe return -ENOMEM; 24835f21b0e6SFrederic Bohe } /* ext4_mb_add_groupinfo */ 24845f21b0e6SFrederic Bohe 2485c9de560dSAlex Tomas static int ext4_mb_init_backend(struct super_block *sb) 2486c9de560dSAlex Tomas { 24878df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 2488c9de560dSAlex Tomas ext4_group_t i; 2489c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 249028623c2fSTheodore Ts'o int err; 24915f21b0e6SFrederic Bohe struct ext4_group_desc *desc; 2492fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep; 2493c9de560dSAlex Tomas 249428623c2fSTheodore Ts'o err = ext4_mb_alloc_groupinfo(sb, ngroups); 249528623c2fSTheodore Ts'o if (err) 249628623c2fSTheodore Ts'o return err; 24975f21b0e6SFrederic Bohe 2498c9de560dSAlex Tomas sbi->s_buddy_cache = new_inode(sb); 2499c9de560dSAlex Tomas if (sbi->s_buddy_cache == NULL) { 25009d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't get new inode"); 2501c9de560dSAlex Tomas goto err_freesgi; 2502c9de560dSAlex Tomas } 250348e6061bSYu Jian /* To avoid potentially colliding with an valid on-disk inode number, 250448e6061bSYu Jian * use EXT4_BAD_INO for the buddy cache inode number. This inode is 250548e6061bSYu Jian * not in the inode hash, so it should never be found by iget(), but 250648e6061bSYu Jian * this will avoid confusion if it ever shows up during debugging. */ 250748e6061bSYu Jian sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; 2508c9de560dSAlex Tomas EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 25098df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 2510c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, i, NULL); 2511c9de560dSAlex Tomas if (desc == NULL) { 25129d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); 2513c9de560dSAlex Tomas goto err_freebuddy; 2514c9de560dSAlex Tomas } 25155f21b0e6SFrederic Bohe if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 25165f21b0e6SFrederic Bohe goto err_freebuddy; 2517c9de560dSAlex Tomas } 2518c9de560dSAlex Tomas 2519c9de560dSAlex Tomas return 0; 2520c9de560dSAlex Tomas 2521c9de560dSAlex Tomas err_freebuddy: 2522fb1813f4SCurt Wohlgemuth cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2523f1fa3342SRoel Kluin while (i-- > 0) 2524fb1813f4SCurt Wohlgemuth kmem_cache_free(cachep, ext4_get_group_info(sb, i)); 252528623c2fSTheodore Ts'o i = sbi->s_group_info_size; 2526f1fa3342SRoel Kluin while (i-- > 0) 2527c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2528c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2529c9de560dSAlex Tomas err_freesgi: 2530b93b41d4SAl Viro kvfree(sbi->s_group_info); 2531c9de560dSAlex Tomas return -ENOMEM; 2532c9de560dSAlex Tomas } 2533c9de560dSAlex Tomas 25342892c15dSEric Sandeen static void ext4_groupinfo_destroy_slabs(void) 25352892c15dSEric Sandeen { 25362892c15dSEric Sandeen int i; 25372892c15dSEric Sandeen 25382892c15dSEric Sandeen for (i = 0; i < NR_GRPINFO_CACHES; i++) { 25392892c15dSEric Sandeen if (ext4_groupinfo_caches[i]) 25402892c15dSEric Sandeen kmem_cache_destroy(ext4_groupinfo_caches[i]); 25412892c15dSEric Sandeen ext4_groupinfo_caches[i] = NULL; 25422892c15dSEric Sandeen } 25432892c15dSEric Sandeen } 25442892c15dSEric Sandeen 25452892c15dSEric Sandeen static int ext4_groupinfo_create_slab(size_t size) 25462892c15dSEric Sandeen { 25472892c15dSEric Sandeen static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); 25482892c15dSEric Sandeen int slab_size; 25492892c15dSEric Sandeen int blocksize_bits = order_base_2(size); 25502892c15dSEric Sandeen int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 25512892c15dSEric Sandeen struct kmem_cache *cachep; 25522892c15dSEric Sandeen 25532892c15dSEric Sandeen if (cache_index >= NR_GRPINFO_CACHES) 25542892c15dSEric Sandeen return -EINVAL; 25552892c15dSEric Sandeen 25562892c15dSEric Sandeen if (unlikely(cache_index < 0)) 25572892c15dSEric Sandeen cache_index = 0; 25582892c15dSEric Sandeen 25592892c15dSEric Sandeen mutex_lock(&ext4_grpinfo_slab_create_mutex); 25602892c15dSEric Sandeen if (ext4_groupinfo_caches[cache_index]) { 25612892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 25622892c15dSEric Sandeen return 0; /* Already created */ 25632892c15dSEric Sandeen } 25642892c15dSEric Sandeen 25652892c15dSEric Sandeen slab_size = offsetof(struct ext4_group_info, 25662892c15dSEric Sandeen bb_counters[blocksize_bits + 2]); 25672892c15dSEric Sandeen 25682892c15dSEric Sandeen cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], 25692892c15dSEric Sandeen slab_size, 0, SLAB_RECLAIM_ACCOUNT, 25702892c15dSEric Sandeen NULL); 25712892c15dSEric Sandeen 2572823ba01fSTao Ma ext4_groupinfo_caches[cache_index] = cachep; 2573823ba01fSTao Ma 25742892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 25752892c15dSEric Sandeen if (!cachep) { 25769d8b9ec4STheodore Ts'o printk(KERN_EMERG 25779d8b9ec4STheodore Ts'o "EXT4-fs: no memory for groupinfo slab cache\n"); 25782892c15dSEric Sandeen return -ENOMEM; 25792892c15dSEric Sandeen } 25802892c15dSEric Sandeen 25812892c15dSEric Sandeen return 0; 25822892c15dSEric Sandeen } 25832892c15dSEric Sandeen 25849d99012fSAkira Fujita int ext4_mb_init(struct super_block *sb) 2585c9de560dSAlex Tomas { 2586c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 25876be2ded1SAneesh Kumar K.V unsigned i, j; 2588935244cdSNicolai Stange unsigned offset, offset_incr; 2589c9de560dSAlex Tomas unsigned max; 259074767c5aSShen Feng int ret; 2591c9de560dSAlex Tomas 25921927805eSEric Sandeen i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); 2593c9de560dSAlex Tomas 2594c9de560dSAlex Tomas sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 2595c9de560dSAlex Tomas if (sbi->s_mb_offsets == NULL) { 2596fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 2597fb1813f4SCurt Wohlgemuth goto out; 2598c9de560dSAlex Tomas } 2599ff7ef329SYasunori Goto 26001927805eSEric Sandeen i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); 2601c9de560dSAlex Tomas sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 2602c9de560dSAlex Tomas if (sbi->s_mb_maxs == NULL) { 2603fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 2604fb1813f4SCurt Wohlgemuth goto out; 2605fb1813f4SCurt Wohlgemuth } 2606fb1813f4SCurt Wohlgemuth 26072892c15dSEric Sandeen ret = ext4_groupinfo_create_slab(sb->s_blocksize); 26082892c15dSEric Sandeen if (ret < 0) 2609fb1813f4SCurt Wohlgemuth goto out; 2610c9de560dSAlex Tomas 2611c9de560dSAlex Tomas /* order 0 is regular bitmap */ 2612c9de560dSAlex Tomas sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 2613c9de560dSAlex Tomas sbi->s_mb_offsets[0] = 0; 2614c9de560dSAlex Tomas 2615c9de560dSAlex Tomas i = 1; 2616c9de560dSAlex Tomas offset = 0; 2617935244cdSNicolai Stange offset_incr = 1 << (sb->s_blocksize_bits - 1); 2618c9de560dSAlex Tomas max = sb->s_blocksize << 2; 2619c9de560dSAlex Tomas do { 2620c9de560dSAlex Tomas sbi->s_mb_offsets[i] = offset; 2621c9de560dSAlex Tomas sbi->s_mb_maxs[i] = max; 2622935244cdSNicolai Stange offset += offset_incr; 2623935244cdSNicolai Stange offset_incr = offset_incr >> 1; 2624c9de560dSAlex Tomas max = max >> 1; 2625c9de560dSAlex Tomas i++; 2626c9de560dSAlex Tomas } while (i <= sb->s_blocksize_bits + 1); 2627c9de560dSAlex Tomas 2628c9de560dSAlex Tomas spin_lock_init(&sbi->s_md_lock); 2629c9de560dSAlex Tomas spin_lock_init(&sbi->s_bal_lock); 2630*d08854f5STheodore Ts'o sbi->s_mb_free_pending = 0; 2631c9de560dSAlex Tomas 2632c9de560dSAlex Tomas sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 2633c9de560dSAlex Tomas sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 2634c9de560dSAlex Tomas sbi->s_mb_stats = MB_DEFAULT_STATS; 2635c9de560dSAlex Tomas sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 2636c9de560dSAlex Tomas sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 263727baebb8STheodore Ts'o /* 263827baebb8STheodore Ts'o * The default group preallocation is 512, which for 4k block 263927baebb8STheodore Ts'o * sizes translates to 2 megabytes. However for bigalloc file 264027baebb8STheodore Ts'o * systems, this is probably too big (i.e, if the cluster size 264127baebb8STheodore Ts'o * is 1 megabyte, then group preallocation size becomes half a 264227baebb8STheodore Ts'o * gigabyte!). As a default, we will keep a two megabyte 264327baebb8STheodore Ts'o * group pralloc size for cluster sizes up to 64k, and after 264427baebb8STheodore Ts'o * that, we will force a minimum group preallocation size of 264527baebb8STheodore Ts'o * 32 clusters. This translates to 8 megs when the cluster 264627baebb8STheodore Ts'o * size is 256k, and 32 megs when the cluster size is 1 meg, 264727baebb8STheodore Ts'o * which seems reasonable as a default. 264827baebb8STheodore Ts'o */ 264927baebb8STheodore Ts'o sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> 265027baebb8STheodore Ts'o sbi->s_cluster_bits, 32); 2651d7a1fee1SDan Ehrenberg /* 2652d7a1fee1SDan Ehrenberg * If there is a s_stripe > 1, then we set the s_mb_group_prealloc 2653d7a1fee1SDan Ehrenberg * to the lowest multiple of s_stripe which is bigger than 2654d7a1fee1SDan Ehrenberg * the s_mb_group_prealloc as determined above. We want 2655d7a1fee1SDan Ehrenberg * the preallocation size to be an exact multiple of the 2656d7a1fee1SDan Ehrenberg * RAID stripe size so that preallocations don't fragment 2657d7a1fee1SDan Ehrenberg * the stripes. 2658d7a1fee1SDan Ehrenberg */ 2659d7a1fee1SDan Ehrenberg if (sbi->s_stripe > 1) { 2660d7a1fee1SDan Ehrenberg sbi->s_mb_group_prealloc = roundup( 2661d7a1fee1SDan Ehrenberg sbi->s_mb_group_prealloc, sbi->s_stripe); 2662d7a1fee1SDan Ehrenberg } 2663c9de560dSAlex Tomas 2664730c213cSEric Sandeen sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 2665c9de560dSAlex Tomas if (sbi->s_locality_groups == NULL) { 2666fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 2667029b10c5SAndrey Tsyvarev goto out; 2668c9de560dSAlex Tomas } 2669730c213cSEric Sandeen for_each_possible_cpu(i) { 2670c9de560dSAlex Tomas struct ext4_locality_group *lg; 2671730c213cSEric Sandeen lg = per_cpu_ptr(sbi->s_locality_groups, i); 2672c9de560dSAlex Tomas mutex_init(&lg->lg_mutex); 26736be2ded1SAneesh Kumar K.V for (j = 0; j < PREALLOC_TB_SIZE; j++) 26746be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 2675c9de560dSAlex Tomas spin_lock_init(&lg->lg_prealloc_lock); 2676c9de560dSAlex Tomas } 2677c9de560dSAlex Tomas 267879a77c5aSYu Jian /* init file for buddy data */ 267979a77c5aSYu Jian ret = ext4_mb_init_backend(sb); 26807aa0baeaSTao Ma if (ret != 0) 26817aa0baeaSTao Ma goto out_free_locality_groups; 268279a77c5aSYu Jian 26837aa0baeaSTao Ma return 0; 26847aa0baeaSTao Ma 26857aa0baeaSTao Ma out_free_locality_groups: 26867aa0baeaSTao Ma free_percpu(sbi->s_locality_groups); 26877aa0baeaSTao Ma sbi->s_locality_groups = NULL; 2688fb1813f4SCurt Wohlgemuth out: 2689fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_offsets); 26907aa0baeaSTao Ma sbi->s_mb_offsets = NULL; 2691fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_maxs); 26927aa0baeaSTao Ma sbi->s_mb_maxs = NULL; 2693fb1813f4SCurt Wohlgemuth return ret; 2694c9de560dSAlex Tomas } 2695c9de560dSAlex Tomas 2696955ce5f5SAneesh Kumar K.V /* need to called with the ext4 group lock held */ 2697c9de560dSAlex Tomas static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) 2698c9de560dSAlex Tomas { 2699c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 2700c9de560dSAlex Tomas struct list_head *cur, *tmp; 2701c9de560dSAlex Tomas int count = 0; 2702c9de560dSAlex Tomas 2703c9de560dSAlex Tomas list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 2704c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 2705c9de560dSAlex Tomas list_del(&pa->pa_group_list); 2706c9de560dSAlex Tomas count++; 2707688f05a0SAneesh Kumar K.V kmem_cache_free(ext4_pspace_cachep, pa); 2708c9de560dSAlex Tomas } 2709c9de560dSAlex Tomas if (count) 27106ba495e9STheodore Ts'o mb_debug(1, "mballoc: %u PAs left\n", count); 2711c9de560dSAlex Tomas 2712c9de560dSAlex Tomas } 2713c9de560dSAlex Tomas 2714c9de560dSAlex Tomas int ext4_mb_release(struct super_block *sb) 2715c9de560dSAlex Tomas { 27168df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 2717c9de560dSAlex Tomas ext4_group_t i; 2718c9de560dSAlex Tomas int num_meta_group_infos; 2719c9de560dSAlex Tomas struct ext4_group_info *grinfo; 2720c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2721fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2722c9de560dSAlex Tomas 2723c9de560dSAlex Tomas if (sbi->s_group_info) { 27248df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 2725c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, i); 2726c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 2727c9de560dSAlex Tomas kfree(grinfo->bb_bitmap); 2728c9de560dSAlex Tomas #endif 2729c9de560dSAlex Tomas ext4_lock_group(sb, i); 2730c9de560dSAlex Tomas ext4_mb_cleanup_pa(grinfo); 2731c9de560dSAlex Tomas ext4_unlock_group(sb, i); 2732fb1813f4SCurt Wohlgemuth kmem_cache_free(cachep, grinfo); 2733c9de560dSAlex Tomas } 27348df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + 2735c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK(sb) - 1) >> 2736c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK_BITS(sb); 2737c9de560dSAlex Tomas for (i = 0; i < num_meta_group_infos; i++) 2738c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2739b93b41d4SAl Viro kvfree(sbi->s_group_info); 2740c9de560dSAlex Tomas } 2741c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2742c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 2743c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2744c9de560dSAlex Tomas if (sbi->s_mb_stats) { 27459d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 27469d8b9ec4STheodore Ts'o "mballoc: %u blocks %u reqs (%u success)", 2747c9de560dSAlex Tomas atomic_read(&sbi->s_bal_allocated), 2748c9de560dSAlex Tomas atomic_read(&sbi->s_bal_reqs), 2749c9de560dSAlex Tomas atomic_read(&sbi->s_bal_success)); 27509d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 27519d8b9ec4STheodore Ts'o "mballoc: %u extents scanned, %u goal hits, " 27529d8b9ec4STheodore Ts'o "%u 2^N hits, %u breaks, %u lost", 2753c9de560dSAlex Tomas atomic_read(&sbi->s_bal_ex_scanned), 2754c9de560dSAlex Tomas atomic_read(&sbi->s_bal_goals), 2755c9de560dSAlex Tomas atomic_read(&sbi->s_bal_2orders), 2756c9de560dSAlex Tomas atomic_read(&sbi->s_bal_breaks), 2757c9de560dSAlex Tomas atomic_read(&sbi->s_mb_lost_chunks)); 27589d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 27599d8b9ec4STheodore Ts'o "mballoc: %lu generated and it took %Lu", 2760ced156e4STao Ma sbi->s_mb_buddies_generated, 2761c9de560dSAlex Tomas sbi->s_mb_generation_time); 27629d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 27639d8b9ec4STheodore Ts'o "mballoc: %u preallocated, %u discarded", 2764c9de560dSAlex Tomas atomic_read(&sbi->s_mb_preallocated), 2765c9de560dSAlex Tomas atomic_read(&sbi->s_mb_discarded)); 2766c9de560dSAlex Tomas } 2767c9de560dSAlex Tomas 2768730c213cSEric Sandeen free_percpu(sbi->s_locality_groups); 2769c9de560dSAlex Tomas 2770c9de560dSAlex Tomas return 0; 2771c9de560dSAlex Tomas } 2772c9de560dSAlex Tomas 277377ca6cdfSLukas Czerner static inline int ext4_issue_discard(struct super_block *sb, 277484130193STheodore Ts'o ext4_group_t block_group, ext4_grpblk_t cluster, int count) 27755c521830SJiaying Zhang { 27765c521830SJiaying Zhang ext4_fsblk_t discard_block; 27775c521830SJiaying Zhang 277884130193STheodore Ts'o discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + 277984130193STheodore Ts'o ext4_group_first_block_no(sb, block_group)); 278084130193STheodore Ts'o count = EXT4_C2B(EXT4_SB(sb), count); 27815c521830SJiaying Zhang trace_ext4_discard_blocks(sb, 27825c521830SJiaying Zhang (unsigned long long) discard_block, count); 278393259636SLukas Czerner return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); 27845c521830SJiaying Zhang } 27855c521830SJiaying Zhang 27863e624fc7STheodore Ts'o /* 27873e624fc7STheodore Ts'o * This function is called by the jbd2 layer once the commit has finished, 27883e624fc7STheodore Ts'o * so we know we can free the blocks that were released with that commit. 27893e624fc7STheodore Ts'o */ 279018aadd47SBobi Jam static void ext4_free_data_callback(struct super_block *sb, 279118aadd47SBobi Jam struct ext4_journal_cb_entry *jce, 279218aadd47SBobi Jam int rc) 2793c9de560dSAlex Tomas { 279418aadd47SBobi Jam struct ext4_free_data *entry = (struct ext4_free_data *)jce; 2795c9de560dSAlex Tomas struct ext4_buddy e4b; 2796c894058dSAneesh Kumar K.V struct ext4_group_info *db; 2797d9f34504STheodore Ts'o int err, count = 0, count2 = 0; 2798c9de560dSAlex Tomas 27996ba495e9STheodore Ts'o mb_debug(1, "gonna free %u blocks in group %u (0x%p):", 280018aadd47SBobi Jam entry->efd_count, entry->efd_group, entry); 2801c9de560dSAlex Tomas 2802d71c1ae2SLukas Czerner if (test_opt(sb, DISCARD)) { 2803d71c1ae2SLukas Czerner err = ext4_issue_discard(sb, entry->efd_group, 2804d71c1ae2SLukas Czerner entry->efd_start_cluster, 2805d71c1ae2SLukas Czerner entry->efd_count); 2806d71c1ae2SLukas Czerner if (err && err != -EOPNOTSUPP) 2807d71c1ae2SLukas Czerner ext4_msg(sb, KERN_WARNING, "discard request in" 2808d71c1ae2SLukas Czerner " group:%d block:%d count:%d failed" 2809d71c1ae2SLukas Czerner " with %d", entry->efd_group, 2810d71c1ae2SLukas Czerner entry->efd_start_cluster, 2811d71c1ae2SLukas Czerner entry->efd_count, err); 2812d71c1ae2SLukas Czerner } 2813b90f6870STheodore Ts'o 281418aadd47SBobi Jam err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); 2815c9de560dSAlex Tomas /* we expect to find existing buddy because it's pinned */ 2816c9de560dSAlex Tomas BUG_ON(err != 0); 2817c9de560dSAlex Tomas 2818*d08854f5STheodore Ts'o spin_lock(&EXT4_SB(sb)->s_md_lock); 2819*d08854f5STheodore Ts'o EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count; 2820*d08854f5STheodore Ts'o spin_unlock(&EXT4_SB(sb)->s_md_lock); 282118aadd47SBobi Jam 2822c894058dSAneesh Kumar K.V db = e4b.bd_info; 2823c9de560dSAlex Tomas /* there are blocks to put in buddy to make them really free */ 282418aadd47SBobi Jam count += entry->efd_count; 2825c9de560dSAlex Tomas count2++; 282618aadd47SBobi Jam ext4_lock_group(sb, entry->efd_group); 2827c894058dSAneesh Kumar K.V /* Take it out of per group rb tree */ 282818aadd47SBobi Jam rb_erase(&entry->efd_node, &(db->bb_free_root)); 282918aadd47SBobi Jam mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); 2830c9de560dSAlex Tomas 28313d56b8d2STao Ma /* 28323d56b8d2STao Ma * Clear the trimmed flag for the group so that the next 28333d56b8d2STao Ma * ext4_trim_fs can trim it. 28343d56b8d2STao Ma * If the volume is mounted with -o discard, online discard 28353d56b8d2STao Ma * is supported and the free blocks will be trimmed online. 28363d56b8d2STao Ma */ 28373d56b8d2STao Ma if (!test_opt(sb, DISCARD)) 28383d56b8d2STao Ma EXT4_MB_GRP_CLEAR_TRIMMED(db); 28393d56b8d2STao Ma 2840c894058dSAneesh Kumar K.V if (!db->bb_free_root.rb_node) { 2841c894058dSAneesh Kumar K.V /* No more items in the per group rb tree 2842c894058dSAneesh Kumar K.V * balance refcounts from ext4_mb_free_metadata() 2843c894058dSAneesh Kumar K.V */ 284409cbfeafSKirill A. Shutemov put_page(e4b.bd_buddy_page); 284509cbfeafSKirill A. Shutemov put_page(e4b.bd_bitmap_page); 2846c894058dSAneesh Kumar K.V } 284718aadd47SBobi Jam ext4_unlock_group(sb, entry->efd_group); 284818aadd47SBobi Jam kmem_cache_free(ext4_free_data_cachep, entry); 2849e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2850c9de560dSAlex Tomas 28516ba495e9STheodore Ts'o mb_debug(1, "freed %u blocks in %u structures\n", count, count2); 2852c9de560dSAlex Tomas } 2853c9de560dSAlex Tomas 28545dabfc78STheodore Ts'o int __init ext4_init_mballoc(void) 2855c9de560dSAlex Tomas { 285616828088STheodore Ts'o ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, 285716828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 2858c9de560dSAlex Tomas if (ext4_pspace_cachep == NULL) 2859c9de560dSAlex Tomas return -ENOMEM; 2860c9de560dSAlex Tomas 286116828088STheodore Ts'o ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, 286216828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 2863256bdb49SEric Sandeen if (ext4_ac_cachep == NULL) { 2864256bdb49SEric Sandeen kmem_cache_destroy(ext4_pspace_cachep); 2865256bdb49SEric Sandeen return -ENOMEM; 2866256bdb49SEric Sandeen } 2867c894058dSAneesh Kumar K.V 286818aadd47SBobi Jam ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, 286916828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 287018aadd47SBobi Jam if (ext4_free_data_cachep == NULL) { 2871c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_pspace_cachep); 2872c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_ac_cachep); 2873c894058dSAneesh Kumar K.V return -ENOMEM; 2874c894058dSAneesh Kumar K.V } 2875c9de560dSAlex Tomas return 0; 2876c9de560dSAlex Tomas } 2877c9de560dSAlex Tomas 28785dabfc78STheodore Ts'o void ext4_exit_mballoc(void) 2879c9de560dSAlex Tomas { 28803e03f9caSJesper Dangaard Brouer /* 28813e03f9caSJesper Dangaard Brouer * Wait for completion of call_rcu()'s on ext4_pspace_cachep 28823e03f9caSJesper Dangaard Brouer * before destroying the slab cache. 28833e03f9caSJesper Dangaard Brouer */ 28843e03f9caSJesper Dangaard Brouer rcu_barrier(); 2885c9de560dSAlex Tomas kmem_cache_destroy(ext4_pspace_cachep); 2886256bdb49SEric Sandeen kmem_cache_destroy(ext4_ac_cachep); 288718aadd47SBobi Jam kmem_cache_destroy(ext4_free_data_cachep); 28882892c15dSEric Sandeen ext4_groupinfo_destroy_slabs(); 2889c9de560dSAlex Tomas } 2890c9de560dSAlex Tomas 2891c9de560dSAlex Tomas 2892c9de560dSAlex Tomas /* 289373b2c716SUwe Kleine-König * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps 2894c9de560dSAlex Tomas * Returns 0 if success or error code 2895c9de560dSAlex Tomas */ 28964ddfef7bSEric Sandeen static noinline_for_stack int 28974ddfef7bSEric Sandeen ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 289853accfa9STheodore Ts'o handle_t *handle, unsigned int reserv_clstrs) 2899c9de560dSAlex Tomas { 2900c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 2901c9de560dSAlex Tomas struct ext4_group_desc *gdp; 2902c9de560dSAlex Tomas struct buffer_head *gdp_bh; 2903c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2904c9de560dSAlex Tomas struct super_block *sb; 2905c9de560dSAlex Tomas ext4_fsblk_t block; 2906519deca0SAneesh Kumar K.V int err, len; 2907c9de560dSAlex Tomas 2908c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 2909c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len <= 0); 2910c9de560dSAlex Tomas 2911c9de560dSAlex Tomas sb = ac->ac_sb; 2912c9de560dSAlex Tomas sbi = EXT4_SB(sb); 2913c9de560dSAlex Tomas 2914574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 29159008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 29169008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 29179008a58eSDarrick J. Wong bitmap_bh = NULL; 2918c9de560dSAlex Tomas goto out_err; 29199008a58eSDarrick J. Wong } 2920c9de560dSAlex Tomas 29215d601255Sliang xie BUFFER_TRACE(bitmap_bh, "getting write access"); 2922c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 2923c9de560dSAlex Tomas if (err) 2924c9de560dSAlex Tomas goto out_err; 2925c9de560dSAlex Tomas 2926c9de560dSAlex Tomas err = -EIO; 2927c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 2928c9de560dSAlex Tomas if (!gdp) 2929c9de560dSAlex Tomas goto out_err; 2930c9de560dSAlex Tomas 2931a9df9a49STheodore Ts'o ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 2932021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, gdp)); 293303cddb80SAneesh Kumar K.V 29345d601255Sliang xie BUFFER_TRACE(gdp_bh, "get_write_access"); 2935c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gdp_bh); 2936c9de560dSAlex Tomas if (err) 2937c9de560dSAlex Tomas goto out_err; 2938c9de560dSAlex Tomas 2939bda00de7SAkinobu Mita block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 2940c9de560dSAlex Tomas 294153accfa9STheodore Ts'o len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 29426fd058f7STheodore Ts'o if (!ext4_data_block_valid(sbi, block, len)) { 294312062dddSEric Sandeen ext4_error(sb, "Allocating blocks %llu-%llu which overlap " 29441084f252STheodore Ts'o "fs metadata", block, block+len); 2945519deca0SAneesh Kumar K.V /* File system mounted not to panic on error 2946519deca0SAneesh Kumar K.V * Fix the bitmap and repeat the block allocation 2947519deca0SAneesh Kumar K.V * We leak some of the blocks here. 2948519deca0SAneesh Kumar K.V */ 2949955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2950c3e94d1dSYongqiang Yang ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 2951519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len); 2952955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 29530390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 2954519deca0SAneesh Kumar K.V if (!err) 2955519deca0SAneesh Kumar K.V err = -EAGAIN; 2956519deca0SAneesh Kumar K.V goto out_err; 2957c9de560dSAlex Tomas } 2958955ce5f5SAneesh Kumar K.V 2959955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2960c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 2961c9de560dSAlex Tomas { 2962c9de560dSAlex Tomas int i; 2963c9de560dSAlex Tomas for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 2964c9de560dSAlex Tomas BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 2965c9de560dSAlex Tomas bitmap_bh->b_data)); 2966c9de560dSAlex Tomas } 2967c9de560dSAlex Tomas } 2968c9de560dSAlex Tomas #endif 2969c3e94d1dSYongqiang Yang ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 2970c3e94d1dSYongqiang Yang ac->ac_b_ex.fe_len); 2971c9de560dSAlex Tomas if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 2972c9de560dSAlex Tomas gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 2973021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, 2974cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, 2975560671a0SAneesh Kumar K.V ac->ac_b_ex.fe_group, gdp)); 2976c9de560dSAlex Tomas } 2977021b65bbSTheodore Ts'o len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; 2978021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, len); 297979f1ba49STao Ma ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh); 2980feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp); 2981955ce5f5SAneesh Kumar K.V 2982955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 298357042651STheodore Ts'o percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); 2984d2a17637SMingming Cao /* 29856bc6e63fSAneesh Kumar K.V * Now reduce the dirty block count also. Should not go negative 2986d2a17637SMingming Cao */ 29876bc6e63fSAneesh Kumar K.V if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 29886bc6e63fSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 298957042651STheodore Ts'o percpu_counter_sub(&sbi->s_dirtyclusters_counter, 299057042651STheodore Ts'o reserv_clstrs); 2991c9de560dSAlex Tomas 2992772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 2993772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, 2994772cb7c8SJose R. Santos ac->ac_b_ex.fe_group); 299590ba983fSTheodore Ts'o atomic64_sub(ac->ac_b_ex.fe_len, 299624aaa8efSTheodore Ts'o &sbi->s_flex_groups[flex_group].free_clusters); 2997772cb7c8SJose R. Santos } 2998772cb7c8SJose R. Santos 29990390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 3000c9de560dSAlex Tomas if (err) 3001c9de560dSAlex Tomas goto out_err; 30020390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 3003c9de560dSAlex Tomas 3004c9de560dSAlex Tomas out_err: 300542a10addSAneesh Kumar K.V brelse(bitmap_bh); 3006c9de560dSAlex Tomas return err; 3007c9de560dSAlex Tomas } 3008c9de560dSAlex Tomas 3009c9de560dSAlex Tomas /* 3010c9de560dSAlex Tomas * here we normalize request for locality group 3011d7a1fee1SDan Ehrenberg * Group request are normalized to s_mb_group_prealloc, which goes to 3012d7a1fee1SDan Ehrenberg * s_strip if we set the same via mount option. 3013d7a1fee1SDan Ehrenberg * s_mb_group_prealloc can be configured via 3014b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_group_prealloc 3015c9de560dSAlex Tomas * 3016c9de560dSAlex Tomas * XXX: should we try to preallocate more than the group has now? 3017c9de560dSAlex Tomas */ 3018c9de560dSAlex Tomas static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 3019c9de560dSAlex Tomas { 3020c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3021c9de560dSAlex Tomas struct ext4_locality_group *lg = ac->ac_lg; 3022c9de560dSAlex Tomas 3023c9de560dSAlex Tomas BUG_ON(lg == NULL); 3024c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 30256ba495e9STheodore Ts'o mb_debug(1, "#%u: goal %u blocks for locality group\n", 3026c9de560dSAlex Tomas current->pid, ac->ac_g_ex.fe_len); 3027c9de560dSAlex Tomas } 3028c9de560dSAlex Tomas 3029c9de560dSAlex Tomas /* 3030c9de560dSAlex Tomas * Normalization means making request better in terms of 3031c9de560dSAlex Tomas * size and alignment 3032c9de560dSAlex Tomas */ 30334ddfef7bSEric Sandeen static noinline_for_stack void 30344ddfef7bSEric Sandeen ext4_mb_normalize_request(struct ext4_allocation_context *ac, 3035c9de560dSAlex Tomas struct ext4_allocation_request *ar) 3036c9de560dSAlex Tomas { 303753accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3038c9de560dSAlex Tomas int bsbits, max; 3039c9de560dSAlex Tomas ext4_lblk_t end; 30401592d2c5SCurt Wohlgemuth loff_t size, start_off; 30411592d2c5SCurt Wohlgemuth loff_t orig_size __maybe_unused; 30425a0790c2SAndi Kleen ext4_lblk_t start; 3043c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 30449a0762c5SAneesh Kumar K.V struct ext4_prealloc_space *pa; 3045c9de560dSAlex Tomas 3046c9de560dSAlex Tomas /* do normalize only data requests, metadata requests 3047c9de560dSAlex Tomas do not need preallocation */ 3048c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3049c9de560dSAlex Tomas return; 3050c9de560dSAlex Tomas 3051c9de560dSAlex Tomas /* sometime caller may want exact blocks */ 3052c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 3053c9de560dSAlex Tomas return; 3054c9de560dSAlex Tomas 3055c9de560dSAlex Tomas /* caller may indicate that preallocation isn't 3056c9de560dSAlex Tomas * required (it's a tail, for example) */ 3057c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 3058c9de560dSAlex Tomas return; 3059c9de560dSAlex Tomas 3060c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 3061c9de560dSAlex Tomas ext4_mb_normalize_group_request(ac); 3062c9de560dSAlex Tomas return ; 3063c9de560dSAlex Tomas } 3064c9de560dSAlex Tomas 3065c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 3066c9de560dSAlex Tomas 3067c9de560dSAlex Tomas /* first, let's learn actual file size 3068c9de560dSAlex Tomas * given current request is allocated */ 306953accfa9STheodore Ts'o size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); 3070c9de560dSAlex Tomas size = size << bsbits; 3071c9de560dSAlex Tomas if (size < i_size_read(ac->ac_inode)) 3072c9de560dSAlex Tomas size = i_size_read(ac->ac_inode); 30735a0790c2SAndi Kleen orig_size = size; 3074c9de560dSAlex Tomas 30751930479cSValerie Clement /* max size of free chunks */ 30761930479cSValerie Clement max = 2 << bsbits; 3077c9de560dSAlex Tomas 30781930479cSValerie Clement #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 30791930479cSValerie Clement (req <= (size) || max <= (chunk_size)) 3080c9de560dSAlex Tomas 3081c9de560dSAlex Tomas /* first, try to predict filesize */ 3082c9de560dSAlex Tomas /* XXX: should this table be tunable? */ 3083c9de560dSAlex Tomas start_off = 0; 3084c9de560dSAlex Tomas if (size <= 16 * 1024) { 3085c9de560dSAlex Tomas size = 16 * 1024; 3086c9de560dSAlex Tomas } else if (size <= 32 * 1024) { 3087c9de560dSAlex Tomas size = 32 * 1024; 3088c9de560dSAlex Tomas } else if (size <= 64 * 1024) { 3089c9de560dSAlex Tomas size = 64 * 1024; 3090c9de560dSAlex Tomas } else if (size <= 128 * 1024) { 3091c9de560dSAlex Tomas size = 128 * 1024; 3092c9de560dSAlex Tomas } else if (size <= 256 * 1024) { 3093c9de560dSAlex Tomas size = 256 * 1024; 3094c9de560dSAlex Tomas } else if (size <= 512 * 1024) { 3095c9de560dSAlex Tomas size = 512 * 1024; 3096c9de560dSAlex Tomas } else if (size <= 1024 * 1024) { 3097c9de560dSAlex Tomas size = 1024 * 1024; 30981930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 3099c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 31001930479cSValerie Clement (21 - bsbits)) << 21; 31011930479cSValerie Clement size = 2 * 1024 * 1024; 31021930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 3103c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3104c9de560dSAlex Tomas (22 - bsbits)) << 22; 3105c9de560dSAlex Tomas size = 4 * 1024 * 1024; 3106c9de560dSAlex Tomas } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, 31071930479cSValerie Clement (8<<20)>>bsbits, max, 8 * 1024)) { 3108c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3109c9de560dSAlex Tomas (23 - bsbits)) << 23; 3110c9de560dSAlex Tomas size = 8 * 1024 * 1024; 3111c9de560dSAlex Tomas } else { 3112c9de560dSAlex Tomas start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits; 3113b27b1535SXiaoguang Wang size = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb), 3114b27b1535SXiaoguang Wang ac->ac_o_ex.fe_len) << bsbits; 3115c9de560dSAlex Tomas } 31165a0790c2SAndi Kleen size = size >> bsbits; 31175a0790c2SAndi Kleen start = start_off >> bsbits; 3118c9de560dSAlex Tomas 3119c9de560dSAlex Tomas /* don't cover already allocated blocks in selected range */ 3120c9de560dSAlex Tomas if (ar->pleft && start <= ar->lleft) { 3121c9de560dSAlex Tomas size -= ar->lleft + 1 - start; 3122c9de560dSAlex Tomas start = ar->lleft + 1; 3123c9de560dSAlex Tomas } 3124c9de560dSAlex Tomas if (ar->pright && start + size - 1 >= ar->lright) 3125c9de560dSAlex Tomas size -= start + size - ar->lright; 3126c9de560dSAlex Tomas 3127c9de560dSAlex Tomas end = start + size; 3128c9de560dSAlex Tomas 3129c9de560dSAlex Tomas /* check we don't cross already preallocated blocks */ 3130c9de560dSAlex Tomas rcu_read_lock(); 31319a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3132498e5f24STheodore Ts'o ext4_lblk_t pa_end; 3133c9de560dSAlex Tomas 3134c9de560dSAlex Tomas if (pa->pa_deleted) 3135c9de560dSAlex Tomas continue; 3136c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3137c9de560dSAlex Tomas if (pa->pa_deleted) { 3138c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3139c9de560dSAlex Tomas continue; 3140c9de560dSAlex Tomas } 3141c9de560dSAlex Tomas 314253accfa9STheodore Ts'o pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), 314353accfa9STheodore Ts'o pa->pa_len); 3144c9de560dSAlex Tomas 3145c9de560dSAlex Tomas /* PA must not overlap original request */ 3146c9de560dSAlex Tomas BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || 3147c9de560dSAlex Tomas ac->ac_o_ex.fe_logical < pa->pa_lstart)); 3148c9de560dSAlex Tomas 314938877f4eSEric Sandeen /* skip PAs this normalized request doesn't overlap with */ 315038877f4eSEric Sandeen if (pa->pa_lstart >= end || pa_end <= start) { 3151c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3152c9de560dSAlex Tomas continue; 3153c9de560dSAlex Tomas } 3154c9de560dSAlex Tomas BUG_ON(pa->pa_lstart <= start && pa_end >= end); 3155c9de560dSAlex Tomas 315638877f4eSEric Sandeen /* adjust start or end to be adjacent to this pa */ 3157c9de560dSAlex Tomas if (pa_end <= ac->ac_o_ex.fe_logical) { 3158c9de560dSAlex Tomas BUG_ON(pa_end < start); 3159c9de560dSAlex Tomas start = pa_end; 316038877f4eSEric Sandeen } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { 3161c9de560dSAlex Tomas BUG_ON(pa->pa_lstart > end); 3162c9de560dSAlex Tomas end = pa->pa_lstart; 3163c9de560dSAlex Tomas } 3164c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3165c9de560dSAlex Tomas } 3166c9de560dSAlex Tomas rcu_read_unlock(); 3167c9de560dSAlex Tomas size = end - start; 3168c9de560dSAlex Tomas 3169c9de560dSAlex Tomas /* XXX: extra loop to check we really don't overlap preallocations */ 3170c9de560dSAlex Tomas rcu_read_lock(); 31719a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3172498e5f24STheodore Ts'o ext4_lblk_t pa_end; 317353accfa9STheodore Ts'o 3174c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3175c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 317653accfa9STheodore Ts'o pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), 317753accfa9STheodore Ts'o pa->pa_len); 3178c9de560dSAlex Tomas BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); 3179c9de560dSAlex Tomas } 3180c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3181c9de560dSAlex Tomas } 3182c9de560dSAlex Tomas rcu_read_unlock(); 3183c9de560dSAlex Tomas 3184c9de560dSAlex Tomas if (start + size <= ac->ac_o_ex.fe_logical && 3185c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical) { 31869d8b9ec4STheodore Ts'o ext4_msg(ac->ac_sb, KERN_ERR, 31879d8b9ec4STheodore Ts'o "start %lu, size %lu, fe_logical %lu", 3188c9de560dSAlex Tomas (unsigned long) start, (unsigned long) size, 3189c9de560dSAlex Tomas (unsigned long) ac->ac_o_ex.fe_logical); 3190dfe076c1SDmitry Monakhov BUG(); 3191c9de560dSAlex Tomas } 3192b5b60778SMaurizio Lombardi BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 3193c9de560dSAlex Tomas 3194c9de560dSAlex Tomas /* now prepare goal request */ 3195c9de560dSAlex Tomas 3196c9de560dSAlex Tomas /* XXX: is it better to align blocks WRT to logical 3197c9de560dSAlex Tomas * placement or satisfy big request as is */ 3198c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = start; 319953accfa9STheodore Ts'o ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); 3200c9de560dSAlex Tomas 3201c9de560dSAlex Tomas /* define goal start in order to merge */ 3202c9de560dSAlex Tomas if (ar->pright && (ar->lright == (start + size))) { 3203c9de560dSAlex Tomas /* merge to the right */ 3204c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 3205c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 3206c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 3207c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3208c9de560dSAlex Tomas } 3209c9de560dSAlex Tomas if (ar->pleft && (ar->lleft + 1 == start)) { 3210c9de560dSAlex Tomas /* merge to the left */ 3211c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 3212c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 3213c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 3214c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3215c9de560dSAlex Tomas } 3216c9de560dSAlex Tomas 32176ba495e9STheodore Ts'o mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, 3218c9de560dSAlex Tomas (unsigned) orig_size, (unsigned) start); 3219c9de560dSAlex Tomas } 3220c9de560dSAlex Tomas 3221c9de560dSAlex Tomas static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 3222c9de560dSAlex Tomas { 3223c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3224c9de560dSAlex Tomas 3225c9de560dSAlex Tomas if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { 3226c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_reqs); 3227c9de560dSAlex Tomas atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 3228291dae47SCurt Wohlgemuth if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) 3229c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_success); 3230c9de560dSAlex Tomas atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 3231c9de560dSAlex Tomas if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 3232c9de560dSAlex Tomas ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 3233c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_goals); 3234c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan) 3235c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_breaks); 3236c9de560dSAlex Tomas } 3237c9de560dSAlex Tomas 3238296c355cSTheodore Ts'o if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 3239296c355cSTheodore Ts'o trace_ext4_mballoc_alloc(ac); 3240296c355cSTheodore Ts'o else 3241296c355cSTheodore Ts'o trace_ext4_mballoc_prealloc(ac); 3242c9de560dSAlex Tomas } 3243c9de560dSAlex Tomas 3244c9de560dSAlex Tomas /* 3245b844167eSCurt Wohlgemuth * Called on failure; free up any blocks from the inode PA for this 3246b844167eSCurt Wohlgemuth * context. We don't need this for MB_GROUP_PA because we only change 3247b844167eSCurt Wohlgemuth * pa_free in ext4_mb_release_context(), but on failure, we've already 3248b844167eSCurt Wohlgemuth * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. 3249b844167eSCurt Wohlgemuth */ 3250b844167eSCurt Wohlgemuth static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) 3251b844167eSCurt Wohlgemuth { 3252b844167eSCurt Wohlgemuth struct ext4_prealloc_space *pa = ac->ac_pa; 325386f0afd4STheodore Ts'o struct ext4_buddy e4b; 325486f0afd4STheodore Ts'o int err; 3255b844167eSCurt Wohlgemuth 325686f0afd4STheodore Ts'o if (pa == NULL) { 3257c99d1e6eSTheodore Ts'o if (ac->ac_f_ex.fe_len == 0) 3258c99d1e6eSTheodore Ts'o return; 325986f0afd4STheodore Ts'o err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b); 326086f0afd4STheodore Ts'o if (err) { 326186f0afd4STheodore Ts'o /* 326286f0afd4STheodore Ts'o * This should never happen since we pin the 326386f0afd4STheodore Ts'o * pages in the ext4_allocation_context so 326486f0afd4STheodore Ts'o * ext4_mb_load_buddy() should never fail. 326586f0afd4STheodore Ts'o */ 326686f0afd4STheodore Ts'o WARN(1, "mb_load_buddy failed (%d)", err); 326786f0afd4STheodore Ts'o return; 326886f0afd4STheodore Ts'o } 326986f0afd4STheodore Ts'o ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 327086f0afd4STheodore Ts'o mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start, 327186f0afd4STheodore Ts'o ac->ac_f_ex.fe_len); 327286f0afd4STheodore Ts'o ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group); 3273c99d1e6eSTheodore Ts'o ext4_mb_unload_buddy(&e4b); 327486f0afd4STheodore Ts'o return; 327586f0afd4STheodore Ts'o } 327686f0afd4STheodore Ts'o if (pa->pa_type == MB_INODE_PA) 3277400db9d3SZheng Liu pa->pa_free += ac->ac_b_ex.fe_len; 3278b844167eSCurt Wohlgemuth } 3279b844167eSCurt Wohlgemuth 3280b844167eSCurt Wohlgemuth /* 3281c9de560dSAlex Tomas * use blocks preallocated to inode 3282c9de560dSAlex Tomas */ 3283c9de560dSAlex Tomas static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 3284c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3285c9de560dSAlex Tomas { 328653accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3287c9de560dSAlex Tomas ext4_fsblk_t start; 3288c9de560dSAlex Tomas ext4_fsblk_t end; 3289c9de560dSAlex Tomas int len; 3290c9de560dSAlex Tomas 3291c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3292c9de560dSAlex Tomas start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 329353accfa9STheodore Ts'o end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), 329453accfa9STheodore Ts'o start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); 329553accfa9STheodore Ts'o len = EXT4_NUM_B2C(sbi, end - start); 3296c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 3297c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3298c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3299c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3300c9de560dSAlex Tomas ac->ac_pa = pa; 3301c9de560dSAlex Tomas 3302c9de560dSAlex Tomas BUG_ON(start < pa->pa_pstart); 330353accfa9STheodore Ts'o BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); 3304c9de560dSAlex Tomas BUG_ON(pa->pa_free < len); 3305c9de560dSAlex Tomas pa->pa_free -= len; 3306c9de560dSAlex Tomas 33076ba495e9STheodore Ts'o mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); 3308c9de560dSAlex Tomas } 3309c9de560dSAlex Tomas 3310c9de560dSAlex Tomas /* 3311c9de560dSAlex Tomas * use blocks preallocated to locality group 3312c9de560dSAlex Tomas */ 3313c9de560dSAlex Tomas static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 3314c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3315c9de560dSAlex Tomas { 331603cddb80SAneesh Kumar K.V unsigned int len = ac->ac_o_ex.fe_len; 33176be2ded1SAneesh Kumar K.V 3318c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 3319c9de560dSAlex Tomas &ac->ac_b_ex.fe_group, 3320c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3321c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3322c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3323c9de560dSAlex Tomas ac->ac_pa = pa; 3324c9de560dSAlex Tomas 3325c9de560dSAlex Tomas /* we don't correct pa_pstart or pa_plen here to avoid 332626346ff6SAneesh Kumar K.V * possible race when the group is being loaded concurrently 3327c9de560dSAlex Tomas * instead we correct pa later, after blocks are marked 332826346ff6SAneesh Kumar K.V * in on-disk bitmap -- see ext4_mb_release_context() 332926346ff6SAneesh Kumar K.V * Other CPUs are prevented from allocating from this pa by lg_mutex 3330c9de560dSAlex Tomas */ 33316ba495e9STheodore Ts'o mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); 3332c9de560dSAlex Tomas } 3333c9de560dSAlex Tomas 3334c9de560dSAlex Tomas /* 33355e745b04SAneesh Kumar K.V * Return the prealloc space that have minimal distance 33365e745b04SAneesh Kumar K.V * from the goal block. @cpa is the prealloc 33375e745b04SAneesh Kumar K.V * space that is having currently known minimal distance 33385e745b04SAneesh Kumar K.V * from the goal block. 33395e745b04SAneesh Kumar K.V */ 33405e745b04SAneesh Kumar K.V static struct ext4_prealloc_space * 33415e745b04SAneesh Kumar K.V ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 33425e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, 33435e745b04SAneesh Kumar K.V struct ext4_prealloc_space *cpa) 33445e745b04SAneesh Kumar K.V { 33455e745b04SAneesh Kumar K.V ext4_fsblk_t cur_distance, new_distance; 33465e745b04SAneesh Kumar K.V 33475e745b04SAneesh Kumar K.V if (cpa == NULL) { 33485e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 33495e745b04SAneesh Kumar K.V return pa; 33505e745b04SAneesh Kumar K.V } 335179211c8eSAndrew Morton cur_distance = abs(goal_block - cpa->pa_pstart); 335279211c8eSAndrew Morton new_distance = abs(goal_block - pa->pa_pstart); 33535e745b04SAneesh Kumar K.V 33545a54b2f1SColy Li if (cur_distance <= new_distance) 33555e745b04SAneesh Kumar K.V return cpa; 33565e745b04SAneesh Kumar K.V 33575e745b04SAneesh Kumar K.V /* drop the previous reference */ 33585e745b04SAneesh Kumar K.V atomic_dec(&cpa->pa_count); 33595e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 33605e745b04SAneesh Kumar K.V return pa; 33615e745b04SAneesh Kumar K.V } 33625e745b04SAneesh Kumar K.V 33635e745b04SAneesh Kumar K.V /* 3364c9de560dSAlex Tomas * search goal blocks in preallocated space 3365c9de560dSAlex Tomas */ 33664ddfef7bSEric Sandeen static noinline_for_stack int 33674ddfef7bSEric Sandeen ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 3368c9de560dSAlex Tomas { 336953accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 33706be2ded1SAneesh Kumar K.V int order, i; 3371c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3372c9de560dSAlex Tomas struct ext4_locality_group *lg; 33735e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, *cpa = NULL; 33745e745b04SAneesh Kumar K.V ext4_fsblk_t goal_block; 3375c9de560dSAlex Tomas 3376c9de560dSAlex Tomas /* only data can be preallocated */ 3377c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3378c9de560dSAlex Tomas return 0; 3379c9de560dSAlex Tomas 3380c9de560dSAlex Tomas /* first, try per-file preallocation */ 3381c9de560dSAlex Tomas rcu_read_lock(); 33829a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3383c9de560dSAlex Tomas 3384c9de560dSAlex Tomas /* all fields in this condition don't change, 3385c9de560dSAlex Tomas * so we can skip locking for them */ 3386c9de560dSAlex Tomas if (ac->ac_o_ex.fe_logical < pa->pa_lstart || 338753accfa9STheodore Ts'o ac->ac_o_ex.fe_logical >= (pa->pa_lstart + 338853accfa9STheodore Ts'o EXT4_C2B(sbi, pa->pa_len))) 3389c9de560dSAlex Tomas continue; 3390c9de560dSAlex Tomas 3391fb0a387dSEric Sandeen /* non-extent files can't have physical blocks past 2^32 */ 339212e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && 339353accfa9STheodore Ts'o (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > 339453accfa9STheodore Ts'o EXT4_MAX_BLOCK_FILE_PHYS)) 3395fb0a387dSEric Sandeen continue; 3396fb0a387dSEric Sandeen 3397c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3398c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3399c9de560dSAlex Tomas if (pa->pa_deleted == 0 && pa->pa_free) { 3400c9de560dSAlex Tomas atomic_inc(&pa->pa_count); 3401c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 3402c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3403c9de560dSAlex Tomas ac->ac_criteria = 10; 3404c9de560dSAlex Tomas rcu_read_unlock(); 3405c9de560dSAlex Tomas return 1; 3406c9de560dSAlex Tomas } 3407c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3408c9de560dSAlex Tomas } 3409c9de560dSAlex Tomas rcu_read_unlock(); 3410c9de560dSAlex Tomas 3411c9de560dSAlex Tomas /* can we use group allocation? */ 3412c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 3413c9de560dSAlex Tomas return 0; 3414c9de560dSAlex Tomas 3415c9de560dSAlex Tomas /* inode may have no locality group for some reason */ 3416c9de560dSAlex Tomas lg = ac->ac_lg; 3417c9de560dSAlex Tomas if (lg == NULL) 3418c9de560dSAlex Tomas return 0; 34196be2ded1SAneesh Kumar K.V order = fls(ac->ac_o_ex.fe_len) - 1; 34206be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 34216be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 34226be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 3423c9de560dSAlex Tomas 3424bda00de7SAkinobu Mita goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); 34255e745b04SAneesh Kumar K.V /* 34265e745b04SAneesh Kumar K.V * search for the prealloc space that is having 34275e745b04SAneesh Kumar K.V * minimal distance from the goal block. 34285e745b04SAneesh Kumar K.V */ 34296be2ded1SAneesh Kumar K.V for (i = order; i < PREALLOC_TB_SIZE; i++) { 3430c9de560dSAlex Tomas rcu_read_lock(); 34316be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], 34326be2ded1SAneesh Kumar K.V pa_inode_list) { 3433c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 34346be2ded1SAneesh Kumar K.V if (pa->pa_deleted == 0 && 34356be2ded1SAneesh Kumar K.V pa->pa_free >= ac->ac_o_ex.fe_len) { 34365e745b04SAneesh Kumar K.V 34375e745b04SAneesh Kumar K.V cpa = ext4_mb_check_group_pa(goal_block, 34385e745b04SAneesh Kumar K.V pa, cpa); 34395e745b04SAneesh Kumar K.V } 3440c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 34415e745b04SAneesh Kumar K.V } 34425e745b04SAneesh Kumar K.V rcu_read_unlock(); 34435e745b04SAneesh Kumar K.V } 34445e745b04SAneesh Kumar K.V if (cpa) { 34455e745b04SAneesh Kumar K.V ext4_mb_use_group_pa(ac, cpa); 3446c9de560dSAlex Tomas ac->ac_criteria = 20; 3447c9de560dSAlex Tomas return 1; 3448c9de560dSAlex Tomas } 3449c9de560dSAlex Tomas return 0; 3450c9de560dSAlex Tomas } 3451c9de560dSAlex Tomas 3452c9de560dSAlex Tomas /* 34537a2fcbf7SAneesh Kumar K.V * the function goes through all block freed in the group 34547a2fcbf7SAneesh Kumar K.V * but not yet committed and marks them used in in-core bitmap. 34557a2fcbf7SAneesh Kumar K.V * buddy must be generated from this bitmap 3456955ce5f5SAneesh Kumar K.V * Need to be called with the ext4 group lock held 34577a2fcbf7SAneesh Kumar K.V */ 34587a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 34597a2fcbf7SAneesh Kumar K.V ext4_group_t group) 34607a2fcbf7SAneesh Kumar K.V { 34617a2fcbf7SAneesh Kumar K.V struct rb_node *n; 34627a2fcbf7SAneesh Kumar K.V struct ext4_group_info *grp; 34637a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 34647a2fcbf7SAneesh Kumar K.V 34657a2fcbf7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 34667a2fcbf7SAneesh Kumar K.V n = rb_first(&(grp->bb_free_root)); 34677a2fcbf7SAneesh Kumar K.V 34687a2fcbf7SAneesh Kumar K.V while (n) { 346918aadd47SBobi Jam entry = rb_entry(n, struct ext4_free_data, efd_node); 347018aadd47SBobi Jam ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); 34717a2fcbf7SAneesh Kumar K.V n = rb_next(n); 34727a2fcbf7SAneesh Kumar K.V } 34737a2fcbf7SAneesh Kumar K.V return; 34747a2fcbf7SAneesh Kumar K.V } 34757a2fcbf7SAneesh Kumar K.V 34767a2fcbf7SAneesh Kumar K.V /* 3477c9de560dSAlex Tomas * the function goes through all preallocation in this group and marks them 3478c9de560dSAlex Tomas * used in in-core bitmap. buddy must be generated from this bitmap 3479955ce5f5SAneesh Kumar K.V * Need to be called with ext4 group lock held 3480c9de560dSAlex Tomas */ 3481089ceeccSEric Sandeen static noinline_for_stack 3482089ceeccSEric Sandeen void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 3483c9de560dSAlex Tomas ext4_group_t group) 3484c9de560dSAlex Tomas { 3485c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3486c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3487c9de560dSAlex Tomas struct list_head *cur; 3488c9de560dSAlex Tomas ext4_group_t groupnr; 3489c9de560dSAlex Tomas ext4_grpblk_t start; 3490c9de560dSAlex Tomas int preallocated = 0; 3491c9de560dSAlex Tomas int len; 3492c9de560dSAlex Tomas 3493c9de560dSAlex Tomas /* all form of preallocation discards first load group, 3494c9de560dSAlex Tomas * so the only competing code is preallocation use. 3495c9de560dSAlex Tomas * we don't need any locking here 3496c9de560dSAlex Tomas * notice we do NOT ignore preallocations with pa_deleted 3497c9de560dSAlex Tomas * otherwise we could leave used blocks available for 3498c9de560dSAlex Tomas * allocation in buddy when concurrent ext4_mb_put_pa() 3499c9de560dSAlex Tomas * is dropping preallocation 3500c9de560dSAlex Tomas */ 3501c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 3502c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3503c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3504c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3505c9de560dSAlex Tomas &groupnr, &start); 3506c9de560dSAlex Tomas len = pa->pa_len; 3507c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3508c9de560dSAlex Tomas if (unlikely(len == 0)) 3509c9de560dSAlex Tomas continue; 3510c9de560dSAlex Tomas BUG_ON(groupnr != group); 3511c3e94d1dSYongqiang Yang ext4_set_bits(bitmap, start, len); 3512c9de560dSAlex Tomas preallocated += len; 3513c9de560dSAlex Tomas } 35146ba495e9STheodore Ts'o mb_debug(1, "prellocated %u for group %u\n", preallocated, group); 3515c9de560dSAlex Tomas } 3516c9de560dSAlex Tomas 3517c9de560dSAlex Tomas static void ext4_mb_pa_callback(struct rcu_head *head) 3518c9de560dSAlex Tomas { 3519c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3520c9de560dSAlex Tomas pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 35214e8d2139SJunho Ryu 35224e8d2139SJunho Ryu BUG_ON(atomic_read(&pa->pa_count)); 35234e8d2139SJunho Ryu BUG_ON(pa->pa_deleted == 0); 3524c9de560dSAlex Tomas kmem_cache_free(ext4_pspace_cachep, pa); 3525c9de560dSAlex Tomas } 3526c9de560dSAlex Tomas 3527c9de560dSAlex Tomas /* 3528c9de560dSAlex Tomas * drops a reference to preallocated space descriptor 3529c9de560dSAlex Tomas * if this was the last reference and the space is consumed 3530c9de560dSAlex Tomas */ 3531c9de560dSAlex Tomas static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 3532c9de560dSAlex Tomas struct super_block *sb, struct ext4_prealloc_space *pa) 3533c9de560dSAlex Tomas { 3534a9df9a49STheodore Ts'o ext4_group_t grp; 3535d33a1976SEric Sandeen ext4_fsblk_t grp_blk; 3536c9de560dSAlex Tomas 3537c9de560dSAlex Tomas /* in this short window concurrent discard can set pa_deleted */ 3538c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 35394e8d2139SJunho Ryu if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) { 35404e8d2139SJunho Ryu spin_unlock(&pa->pa_lock); 35414e8d2139SJunho Ryu return; 35424e8d2139SJunho Ryu } 35434e8d2139SJunho Ryu 3544c9de560dSAlex Tomas if (pa->pa_deleted == 1) { 3545c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3546c9de560dSAlex Tomas return; 3547c9de560dSAlex Tomas } 3548c9de560dSAlex Tomas 3549c9de560dSAlex Tomas pa->pa_deleted = 1; 3550c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3551c9de560dSAlex Tomas 3552d33a1976SEric Sandeen grp_blk = pa->pa_pstart; 3553cc0fb9adSAneesh Kumar K.V /* 3554cc0fb9adSAneesh Kumar K.V * If doing group-based preallocation, pa_pstart may be in the 3555cc0fb9adSAneesh Kumar K.V * next group when pa is used up 3556cc0fb9adSAneesh Kumar K.V */ 3557cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 3558d33a1976SEric Sandeen grp_blk--; 3559d33a1976SEric Sandeen 3560bd86298eSLukas Czerner grp = ext4_get_group_number(sb, grp_blk); 3561c9de560dSAlex Tomas 3562c9de560dSAlex Tomas /* 3563c9de560dSAlex Tomas * possible race: 3564c9de560dSAlex Tomas * 3565c9de560dSAlex Tomas * P1 (buddy init) P2 (regular allocation) 3566c9de560dSAlex Tomas * find block B in PA 3567c9de560dSAlex Tomas * copy on-disk bitmap to buddy 3568c9de560dSAlex Tomas * mark B in on-disk bitmap 3569c9de560dSAlex Tomas * drop PA from group 3570c9de560dSAlex Tomas * mark all PAs in buddy 3571c9de560dSAlex Tomas * 3572c9de560dSAlex Tomas * thus, P1 initializes buddy with B available. to prevent this 3573c9de560dSAlex Tomas * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 3574c9de560dSAlex Tomas * against that pair 3575c9de560dSAlex Tomas */ 3576c9de560dSAlex Tomas ext4_lock_group(sb, grp); 3577c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3578c9de560dSAlex Tomas ext4_unlock_group(sb, grp); 3579c9de560dSAlex Tomas 3580c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3581c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3582c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3583c9de560dSAlex Tomas 3584c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3585c9de560dSAlex Tomas } 3586c9de560dSAlex Tomas 3587c9de560dSAlex Tomas /* 3588c9de560dSAlex Tomas * creates new preallocated space for given inode 3589c9de560dSAlex Tomas */ 35904ddfef7bSEric Sandeen static noinline_for_stack int 35914ddfef7bSEric Sandeen ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 3592c9de560dSAlex Tomas { 3593c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 359453accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(sb); 3595c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3596c9de560dSAlex Tomas struct ext4_group_info *grp; 3597c9de560dSAlex Tomas struct ext4_inode_info *ei; 3598c9de560dSAlex Tomas 3599c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3600c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3601c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3602c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3603c9de560dSAlex Tomas 3604c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3605c9de560dSAlex Tomas if (pa == NULL) 3606c9de560dSAlex Tomas return -ENOMEM; 3607c9de560dSAlex Tomas 3608c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { 3609c9de560dSAlex Tomas int winl; 3610c9de560dSAlex Tomas int wins; 3611c9de560dSAlex Tomas int win; 3612c9de560dSAlex Tomas int offs; 3613c9de560dSAlex Tomas 3614c9de560dSAlex Tomas /* we can't allocate as much as normalizer wants. 3615c9de560dSAlex Tomas * so, found space must get proper lstart 3616c9de560dSAlex Tomas * to cover original request */ 3617c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 3618c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 3619c9de560dSAlex Tomas 3620c9de560dSAlex Tomas /* we're limited by original request in that 3621c9de560dSAlex Tomas * logical block must be covered any way 3622c9de560dSAlex Tomas * winl is window we can move our chunk within */ 3623c9de560dSAlex Tomas winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; 3624c9de560dSAlex Tomas 3625c9de560dSAlex Tomas /* also, we should cover whole original request */ 362653accfa9STheodore Ts'o wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); 3627c9de560dSAlex Tomas 3628c9de560dSAlex Tomas /* the smallest one defines real window */ 3629c9de560dSAlex Tomas win = min(winl, wins); 3630c9de560dSAlex Tomas 363153accfa9STheodore Ts'o offs = ac->ac_o_ex.fe_logical % 363253accfa9STheodore Ts'o EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 3633c9de560dSAlex Tomas if (offs && offs < win) 3634c9de560dSAlex Tomas win = offs; 3635c9de560dSAlex Tomas 363653accfa9STheodore Ts'o ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - 3637810da240SLukas Czerner EXT4_NUM_B2C(sbi, win); 3638c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 3639c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 3640c9de560dSAlex Tomas } 3641c9de560dSAlex Tomas 3642c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3643c9de560dSAlex Tomas * allocated blocks for history */ 3644c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3645c9de560dSAlex Tomas 3646c9de560dSAlex Tomas pa->pa_lstart = ac->ac_b_ex.fe_logical; 3647c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3648c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3649c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3650c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3651c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 3652d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3653d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 3654c9de560dSAlex Tomas pa->pa_deleted = 0; 3655cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_INODE_PA; 3656c9de560dSAlex Tomas 36576ba495e9STheodore Ts'o mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, 3658c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 36599bffad1eSTheodore Ts'o trace_ext4_mb_new_inode_pa(ac, pa); 3660c9de560dSAlex Tomas 3661c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 366253accfa9STheodore Ts'o atomic_add(pa->pa_free, &sbi->s_mb_preallocated); 3663c9de560dSAlex Tomas 3664c9de560dSAlex Tomas ei = EXT4_I(ac->ac_inode); 3665c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3666c9de560dSAlex Tomas 3667c9de560dSAlex Tomas pa->pa_obj_lock = &ei->i_prealloc_lock; 3668c9de560dSAlex Tomas pa->pa_inode = ac->ac_inode; 3669c9de560dSAlex Tomas 3670c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3671c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3672c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3673c9de560dSAlex Tomas 3674c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3675c9de560dSAlex Tomas list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); 3676c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3677c9de560dSAlex Tomas 3678c9de560dSAlex Tomas return 0; 3679c9de560dSAlex Tomas } 3680c9de560dSAlex Tomas 3681c9de560dSAlex Tomas /* 3682c9de560dSAlex Tomas * creates new preallocated space for locality group inodes belongs to 3683c9de560dSAlex Tomas */ 36844ddfef7bSEric Sandeen static noinline_for_stack int 36854ddfef7bSEric Sandeen ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 3686c9de560dSAlex Tomas { 3687c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3688c9de560dSAlex Tomas struct ext4_locality_group *lg; 3689c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3690c9de560dSAlex Tomas struct ext4_group_info *grp; 3691c9de560dSAlex Tomas 3692c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3693c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3694c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3695c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3696c9de560dSAlex Tomas 3697c9de560dSAlex Tomas BUG_ON(ext4_pspace_cachep == NULL); 3698c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3699c9de560dSAlex Tomas if (pa == NULL) 3700c9de560dSAlex Tomas return -ENOMEM; 3701c9de560dSAlex Tomas 3702c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3703c9de560dSAlex Tomas * allocated blocks for history */ 3704c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3705c9de560dSAlex Tomas 3706c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3707c9de560dSAlex Tomas pa->pa_lstart = pa->pa_pstart; 3708c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3709c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3710c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3711c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 37126be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3713d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 3714c9de560dSAlex Tomas pa->pa_deleted = 0; 3715cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_GROUP_PA; 3716c9de560dSAlex Tomas 37176ba495e9STheodore Ts'o mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, 3718c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 37199bffad1eSTheodore Ts'o trace_ext4_mb_new_group_pa(ac, pa); 3720c9de560dSAlex Tomas 3721c9de560dSAlex Tomas ext4_mb_use_group_pa(ac, pa); 3722c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3723c9de560dSAlex Tomas 3724c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3725c9de560dSAlex Tomas lg = ac->ac_lg; 3726c9de560dSAlex Tomas BUG_ON(lg == NULL); 3727c9de560dSAlex Tomas 3728c9de560dSAlex Tomas pa->pa_obj_lock = &lg->lg_prealloc_lock; 3729c9de560dSAlex Tomas pa->pa_inode = NULL; 3730c9de560dSAlex Tomas 3731c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3732c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3733c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3734c9de560dSAlex Tomas 37356be2ded1SAneesh Kumar K.V /* 37366be2ded1SAneesh Kumar K.V * We will later add the new pa to the right bucket 37376be2ded1SAneesh Kumar K.V * after updating the pa_free in ext4_mb_release_context 37386be2ded1SAneesh Kumar K.V */ 3739c9de560dSAlex Tomas return 0; 3740c9de560dSAlex Tomas } 3741c9de560dSAlex Tomas 3742c9de560dSAlex Tomas static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 3743c9de560dSAlex Tomas { 3744c9de560dSAlex Tomas int err; 3745c9de560dSAlex Tomas 3746c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 3747c9de560dSAlex Tomas err = ext4_mb_new_group_pa(ac); 3748c9de560dSAlex Tomas else 3749c9de560dSAlex Tomas err = ext4_mb_new_inode_pa(ac); 3750c9de560dSAlex Tomas return err; 3751c9de560dSAlex Tomas } 3752c9de560dSAlex Tomas 3753c9de560dSAlex Tomas /* 3754c9de560dSAlex Tomas * finds all unused blocks in on-disk bitmap, frees them in 3755c9de560dSAlex Tomas * in-core bitmap and buddy. 3756c9de560dSAlex Tomas * @pa must be unlinked from inode and group lists, so that 3757c9de560dSAlex Tomas * nobody else can find/use it. 3758c9de560dSAlex Tomas * the caller MUST hold group/inode locks. 3759c9de560dSAlex Tomas * TODO: optimize the case when there are no in-core structures yet 3760c9de560dSAlex Tomas */ 37614ddfef7bSEric Sandeen static noinline_for_stack int 37624ddfef7bSEric Sandeen ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 37633e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 3764c9de560dSAlex Tomas { 3765c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3766c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3767498e5f24STheodore Ts'o unsigned int end; 3768498e5f24STheodore Ts'o unsigned int next; 3769c9de560dSAlex Tomas ext4_group_t group; 3770c9de560dSAlex Tomas ext4_grpblk_t bit; 3771ba80b101STheodore Ts'o unsigned long long grp_blk_start; 3772c9de560dSAlex Tomas int err = 0; 3773c9de560dSAlex Tomas int free = 0; 3774c9de560dSAlex Tomas 3775c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3776c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 377753accfa9STheodore Ts'o grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); 3778c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3779c9de560dSAlex Tomas end = bit + pa->pa_len; 3780c9de560dSAlex Tomas 3781c9de560dSAlex Tomas while (bit < end) { 3782ffad0a44SAneesh Kumar K.V bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 3783c9de560dSAlex Tomas if (bit >= end) 3784c9de560dSAlex Tomas break; 3785ffad0a44SAneesh Kumar K.V next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 37866ba495e9STheodore Ts'o mb_debug(1, " free preallocated %u/%u in group %u\n", 37875a0790c2SAndi Kleen (unsigned) ext4_group_first_block_no(sb, group) + bit, 37885a0790c2SAndi Kleen (unsigned) next - bit, (unsigned) group); 3789c9de560dSAlex Tomas free += next - bit; 3790c9de560dSAlex Tomas 37913e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); 379253accfa9STheodore Ts'o trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + 379353accfa9STheodore Ts'o EXT4_C2B(sbi, bit)), 3794a9c667f8SLukas Czerner next - bit); 3795c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 3796c9de560dSAlex Tomas bit = next + 1; 3797c9de560dSAlex Tomas } 3798c9de560dSAlex Tomas if (free != pa->pa_free) { 37999d8b9ec4STheodore Ts'o ext4_msg(e4b->bd_sb, KERN_CRIT, 38009d8b9ec4STheodore Ts'o "pa %p: logic %lu, phys. %lu, len %lu", 3801c9de560dSAlex Tomas pa, (unsigned long) pa->pa_lstart, 3802c9de560dSAlex Tomas (unsigned long) pa->pa_pstart, 3803c9de560dSAlex Tomas (unsigned long) pa->pa_len); 3804e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", 380526346ff6SAneesh Kumar K.V free, pa->pa_free); 3806e56eb659SAneesh Kumar K.V /* 3807e56eb659SAneesh Kumar K.V * pa is already deleted so we use the value obtained 3808e56eb659SAneesh Kumar K.V * from the bitmap and continue. 3809e56eb659SAneesh Kumar K.V */ 3810c9de560dSAlex Tomas } 3811c9de560dSAlex Tomas atomic_add(free, &sbi->s_mb_discarded); 3812c9de560dSAlex Tomas 3813c9de560dSAlex Tomas return err; 3814c9de560dSAlex Tomas } 3815c9de560dSAlex Tomas 38164ddfef7bSEric Sandeen static noinline_for_stack int 38174ddfef7bSEric Sandeen ext4_mb_release_group_pa(struct ext4_buddy *e4b, 38183e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 3819c9de560dSAlex Tomas { 3820c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3821c9de560dSAlex Tomas ext4_group_t group; 3822c9de560dSAlex Tomas ext4_grpblk_t bit; 3823c9de560dSAlex Tomas 382460e07cf5SYongqiang Yang trace_ext4_mb_release_group_pa(sb, pa); 3825c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3826c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3827c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3828c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 3829c9de560dSAlex Tomas atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 38303e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); 3831c9de560dSAlex Tomas 3832c9de560dSAlex Tomas return 0; 3833c9de560dSAlex Tomas } 3834c9de560dSAlex Tomas 3835c9de560dSAlex Tomas /* 3836c9de560dSAlex Tomas * releases all preallocations in given group 3837c9de560dSAlex Tomas * 3838c9de560dSAlex Tomas * first, we need to decide discard policy: 3839c9de560dSAlex Tomas * - when do we discard 3840c9de560dSAlex Tomas * 1) ENOSPC 3841c9de560dSAlex Tomas * - how many do we discard 3842c9de560dSAlex Tomas * 1) how many requested 3843c9de560dSAlex Tomas */ 38444ddfef7bSEric Sandeen static noinline_for_stack int 38454ddfef7bSEric Sandeen ext4_mb_discard_group_preallocations(struct super_block *sb, 3846c9de560dSAlex Tomas ext4_group_t group, int needed) 3847c9de560dSAlex Tomas { 3848c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3849c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3850c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3851c9de560dSAlex Tomas struct list_head list; 3852c9de560dSAlex Tomas struct ext4_buddy e4b; 3853c9de560dSAlex Tomas int err; 3854c9de560dSAlex Tomas int busy = 0; 3855c9de560dSAlex Tomas int free = 0; 3856c9de560dSAlex Tomas 38576ba495e9STheodore Ts'o mb_debug(1, "discard preallocation for group %u\n", group); 3858c9de560dSAlex Tomas 3859c9de560dSAlex Tomas if (list_empty(&grp->bb_prealloc_list)) 3860c9de560dSAlex Tomas return 0; 3861c9de560dSAlex Tomas 3862574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 38639008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 38649008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 38659008a58eSDarrick J. Wong ext4_error(sb, "Error %d reading block bitmap for %u", 38669008a58eSDarrick J. Wong err, group); 3867ce89f46cSAneesh Kumar K.V return 0; 3868c9de560dSAlex Tomas } 3869c9de560dSAlex Tomas 3870c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3871ce89f46cSAneesh Kumar K.V if (err) { 387212062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", group); 3873ce89f46cSAneesh Kumar K.V put_bh(bitmap_bh); 3874ce89f46cSAneesh Kumar K.V return 0; 3875ce89f46cSAneesh Kumar K.V } 3876c9de560dSAlex Tomas 3877c9de560dSAlex Tomas if (needed == 0) 38787137d7a4STheodore Ts'o needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; 3879c9de560dSAlex Tomas 3880c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3881c9de560dSAlex Tomas repeat: 3882c9de560dSAlex Tomas ext4_lock_group(sb, group); 3883c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, 3884c9de560dSAlex Tomas &grp->bb_prealloc_list, pa_group_list) { 3885c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3886c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3887c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3888c9de560dSAlex Tomas busy = 1; 3889c9de560dSAlex Tomas continue; 3890c9de560dSAlex Tomas } 3891c9de560dSAlex Tomas if (pa->pa_deleted) { 3892c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3893c9de560dSAlex Tomas continue; 3894c9de560dSAlex Tomas } 3895c9de560dSAlex Tomas 3896c9de560dSAlex Tomas /* seems this one can be freed ... */ 3897c9de560dSAlex Tomas pa->pa_deleted = 1; 3898c9de560dSAlex Tomas 3899c9de560dSAlex Tomas /* we can trust pa_free ... */ 3900c9de560dSAlex Tomas free += pa->pa_free; 3901c9de560dSAlex Tomas 3902c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3903c9de560dSAlex Tomas 3904c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3905c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3906c9de560dSAlex Tomas } 3907c9de560dSAlex Tomas 3908c9de560dSAlex Tomas /* if we still need more blocks and some PAs were used, try again */ 3909c9de560dSAlex Tomas if (free < needed && busy) { 3910c9de560dSAlex Tomas busy = 0; 3911c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3912bb8b20edSLukas Czerner cond_resched(); 3913c9de560dSAlex Tomas goto repeat; 3914c9de560dSAlex Tomas } 3915c9de560dSAlex Tomas 3916c9de560dSAlex Tomas /* found anything to free? */ 3917c9de560dSAlex Tomas if (list_empty(&list)) { 3918c9de560dSAlex Tomas BUG_ON(free != 0); 3919c9de560dSAlex Tomas goto out; 3920c9de560dSAlex Tomas } 3921c9de560dSAlex Tomas 3922c9de560dSAlex Tomas /* now free all selected PAs */ 3923c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3924c9de560dSAlex Tomas 3925c9de560dSAlex Tomas /* remove from object (inode or locality group) */ 3926c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3927c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3928c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3929c9de560dSAlex Tomas 3930cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 39313e1e5f50SEric Sandeen ext4_mb_release_group_pa(&e4b, pa); 3932c9de560dSAlex Tomas else 39333e1e5f50SEric Sandeen ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 3934c9de560dSAlex Tomas 3935c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 3936c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3937c9de560dSAlex Tomas } 3938c9de560dSAlex Tomas 3939c9de560dSAlex Tomas out: 3940c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3941e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3942c9de560dSAlex Tomas put_bh(bitmap_bh); 3943c9de560dSAlex Tomas return free; 3944c9de560dSAlex Tomas } 3945c9de560dSAlex Tomas 3946c9de560dSAlex Tomas /* 3947c9de560dSAlex Tomas * releases all non-used preallocated blocks for given inode 3948c9de560dSAlex Tomas * 3949c9de560dSAlex Tomas * It's important to discard preallocations under i_data_sem 3950c9de560dSAlex Tomas * We don't want another block to be served from the prealloc 3951c9de560dSAlex Tomas * space when we are discarding the inode prealloc space. 3952c9de560dSAlex Tomas * 3953c9de560dSAlex Tomas * FIXME!! Make sure it is valid at all the call sites 3954c9de560dSAlex Tomas */ 3955c2ea3fdeSTheodore Ts'o void ext4_discard_preallocations(struct inode *inode) 3956c9de560dSAlex Tomas { 3957c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(inode); 3958c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 3959c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3960c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3961c9de560dSAlex Tomas ext4_group_t group = 0; 3962c9de560dSAlex Tomas struct list_head list; 3963c9de560dSAlex Tomas struct ext4_buddy e4b; 3964c9de560dSAlex Tomas int err; 3965c9de560dSAlex Tomas 3966c2ea3fdeSTheodore Ts'o if (!S_ISREG(inode->i_mode)) { 3967c9de560dSAlex Tomas /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ 3968c9de560dSAlex Tomas return; 3969c9de560dSAlex Tomas } 3970c9de560dSAlex Tomas 39716ba495e9STheodore Ts'o mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); 39729bffad1eSTheodore Ts'o trace_ext4_discard_preallocations(inode); 3973c9de560dSAlex Tomas 3974c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3975c9de560dSAlex Tomas 3976c9de560dSAlex Tomas repeat: 3977c9de560dSAlex Tomas /* first, collect all pa's in the inode */ 3978c9de560dSAlex Tomas spin_lock(&ei->i_prealloc_lock); 3979c9de560dSAlex Tomas while (!list_empty(&ei->i_prealloc_list)) { 3980c9de560dSAlex Tomas pa = list_entry(ei->i_prealloc_list.next, 3981c9de560dSAlex Tomas struct ext4_prealloc_space, pa_inode_list); 3982c9de560dSAlex Tomas BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); 3983c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3984c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3985c9de560dSAlex Tomas /* this shouldn't happen often - nobody should 3986c9de560dSAlex Tomas * use preallocation while we're discarding it */ 3987c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3988c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 39899d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, 39909d8b9ec4STheodore Ts'o "uh-oh! used pa while discarding"); 3991c9de560dSAlex Tomas WARN_ON(1); 3992c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 3993c9de560dSAlex Tomas goto repeat; 3994c9de560dSAlex Tomas 3995c9de560dSAlex Tomas } 3996c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 3997c9de560dSAlex Tomas pa->pa_deleted = 1; 3998c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3999c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 4000c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 4001c9de560dSAlex Tomas continue; 4002c9de560dSAlex Tomas } 4003c9de560dSAlex Tomas 4004c9de560dSAlex Tomas /* someone is deleting pa right now */ 4005c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 4006c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 4007c9de560dSAlex Tomas 4008c9de560dSAlex Tomas /* we have to wait here because pa_deleted 4009c9de560dSAlex Tomas * doesn't mean pa is already unlinked from 4010c9de560dSAlex Tomas * the list. as we might be called from 4011c9de560dSAlex Tomas * ->clear_inode() the inode will get freed 4012c9de560dSAlex Tomas * and concurrent thread which is unlinking 4013c9de560dSAlex Tomas * pa from inode's list may access already 4014c9de560dSAlex Tomas * freed memory, bad-bad-bad */ 4015c9de560dSAlex Tomas 4016c9de560dSAlex Tomas /* XXX: if this happens too often, we can 4017c9de560dSAlex Tomas * add a flag to force wait only in case 4018c9de560dSAlex Tomas * of ->clear_inode(), but not in case of 4019c9de560dSAlex Tomas * regular truncate */ 4020c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 4021c9de560dSAlex Tomas goto repeat; 4022c9de560dSAlex Tomas } 4023c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 4024c9de560dSAlex Tomas 4025c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 4026cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_INODE_PA); 4027bd86298eSLukas Czerner group = ext4_get_group_number(sb, pa->pa_pstart); 4028c9de560dSAlex Tomas 4029c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 4030ce89f46cSAneesh Kumar K.V if (err) { 403112062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", 403212062dddSEric Sandeen group); 4033ce89f46cSAneesh Kumar K.V continue; 4034ce89f46cSAneesh Kumar K.V } 4035c9de560dSAlex Tomas 4036574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 40379008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 40389008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 40399008a58eSDarrick J. Wong ext4_error(sb, "Error %d reading block bitmap for %u", 40409008a58eSDarrick J. Wong err, group); 4041e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 4042ce89f46cSAneesh Kumar K.V continue; 4043c9de560dSAlex Tomas } 4044c9de560dSAlex Tomas 4045c9de560dSAlex Tomas ext4_lock_group(sb, group); 4046c9de560dSAlex Tomas list_del(&pa->pa_group_list); 40473e1e5f50SEric Sandeen ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 4048c9de560dSAlex Tomas ext4_unlock_group(sb, group); 4049c9de560dSAlex Tomas 4050e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 4051c9de560dSAlex Tomas put_bh(bitmap_bh); 4052c9de560dSAlex Tomas 4053c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 4054c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 4055c9de560dSAlex Tomas } 4056c9de560dSAlex Tomas } 4057c9de560dSAlex Tomas 40586ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 4059c9de560dSAlex Tomas static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 4060c9de560dSAlex Tomas { 4061c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 40628df9675fSTheodore Ts'o ext4_group_t ngroups, i; 4063c9de560dSAlex Tomas 4064a0b30c12STheodore Ts'o if (!ext4_mballoc_debug || 40654dd89fc6STheodore Ts'o (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) 4066e3570639SEric Sandeen return; 4067e3570639SEric Sandeen 40687f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" 40699d8b9ec4STheodore Ts'o " Allocation context details:"); 40707f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", 4071c9de560dSAlex Tomas ac->ac_status, ac->ac_flags); 40727f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " 40739d8b9ec4STheodore Ts'o "goal %lu/%lu/%lu@%lu, " 40749d8b9ec4STheodore Ts'o "best %lu/%lu/%lu@%lu cr %d", 4075c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_group, 4076c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_start, 4077c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_len, 4078c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_logical, 4079c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_group, 4080c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_start, 4081c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_len, 4082c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_logical, 4083c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_group, 4084c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_start, 4085c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_len, 4086c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_logical, 4087c9de560dSAlex Tomas (int)ac->ac_criteria); 4088dc9ddd98SEric Sandeen ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found); 40897f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); 40908df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 40918df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 4092c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, i); 4093c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 4094c9de560dSAlex Tomas ext4_grpblk_t start; 4095c9de560dSAlex Tomas struct list_head *cur; 4096c9de560dSAlex Tomas ext4_lock_group(sb, i); 4097c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 4098c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, 4099c9de560dSAlex Tomas pa_group_list); 4100c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 4101c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 4102c9de560dSAlex Tomas NULL, &start); 4103c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 41041c718505SAkira Fujita printk(KERN_ERR "PA:%u:%d:%u \n", i, 4105c9de560dSAlex Tomas start, pa->pa_len); 4106c9de560dSAlex Tomas } 410760bd63d1SSolofo Ramangalahy ext4_unlock_group(sb, i); 4108c9de560dSAlex Tomas 4109c9de560dSAlex Tomas if (grp->bb_free == 0) 4110c9de560dSAlex Tomas continue; 41111c718505SAkira Fujita printk(KERN_ERR "%u: %d/%d \n", 4112c9de560dSAlex Tomas i, grp->bb_free, grp->bb_fragments); 4113c9de560dSAlex Tomas } 4114c9de560dSAlex Tomas printk(KERN_ERR "\n"); 4115c9de560dSAlex Tomas } 4116c9de560dSAlex Tomas #else 4117c9de560dSAlex Tomas static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 4118c9de560dSAlex Tomas { 4119c9de560dSAlex Tomas return; 4120c9de560dSAlex Tomas } 4121c9de560dSAlex Tomas #endif 4122c9de560dSAlex Tomas 4123c9de560dSAlex Tomas /* 4124c9de560dSAlex Tomas * We use locality group preallocation for small size file. The size of the 4125c9de560dSAlex Tomas * file is determined by the current size or the resulting size after 4126c9de560dSAlex Tomas * allocation which ever is larger 4127c9de560dSAlex Tomas * 4128b713a5ecSTheodore Ts'o * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 4129c9de560dSAlex Tomas */ 4130c9de560dSAlex Tomas static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 4131c9de560dSAlex Tomas { 4132c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4133c9de560dSAlex Tomas int bsbits = ac->ac_sb->s_blocksize_bits; 4134c9de560dSAlex Tomas loff_t size, isize; 4135c9de560dSAlex Tomas 4136c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 4137c9de560dSAlex Tomas return; 4138c9de560dSAlex Tomas 41394ba74d00STheodore Ts'o if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 41404ba74d00STheodore Ts'o return; 41414ba74d00STheodore Ts'o 414253accfa9STheodore Ts'o size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); 414350797481STheodore Ts'o isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 414450797481STheodore Ts'o >> bsbits; 4145c9de560dSAlex Tomas 414650797481STheodore Ts'o if ((size == isize) && 414750797481STheodore Ts'o !ext4_fs_is_busy(sbi) && 414850797481STheodore Ts'o (atomic_read(&ac->ac_inode->i_writecount) == 0)) { 414950797481STheodore Ts'o ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 415050797481STheodore Ts'o return; 415150797481STheodore Ts'o } 415250797481STheodore Ts'o 4153ebbe0277SRobin Dong if (sbi->s_mb_group_prealloc <= 0) { 4154ebbe0277SRobin Dong ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 4155ebbe0277SRobin Dong return; 4156ebbe0277SRobin Dong } 4157ebbe0277SRobin Dong 4158c9de560dSAlex Tomas /* don't use group allocation for large files */ 415971780577STheodore Ts'o size = max(size, isize); 4160cc483f10STao Ma if (size > sbi->s_mb_stream_request) { 41614ba74d00STheodore Ts'o ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 4162c9de560dSAlex Tomas return; 41634ba74d00STheodore Ts'o } 4164c9de560dSAlex Tomas 4165c9de560dSAlex Tomas BUG_ON(ac->ac_lg != NULL); 4166c9de560dSAlex Tomas /* 4167c9de560dSAlex Tomas * locality group prealloc space are per cpu. The reason for having 4168c9de560dSAlex Tomas * per cpu locality group is to reduce the contention between block 4169c9de560dSAlex Tomas * request from multiple CPUs. 4170c9de560dSAlex Tomas */ 4171a0b6bc63SChristoph Lameter ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups); 4172c9de560dSAlex Tomas 4173c9de560dSAlex Tomas /* we're going to use group allocation */ 4174c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 4175c9de560dSAlex Tomas 4176c9de560dSAlex Tomas /* serialize all allocations in the group */ 4177c9de560dSAlex Tomas mutex_lock(&ac->ac_lg->lg_mutex); 4178c9de560dSAlex Tomas } 4179c9de560dSAlex Tomas 41804ddfef7bSEric Sandeen static noinline_for_stack int 41814ddfef7bSEric Sandeen ext4_mb_initialize_context(struct ext4_allocation_context *ac, 4182c9de560dSAlex Tomas struct ext4_allocation_request *ar) 4183c9de560dSAlex Tomas { 4184c9de560dSAlex Tomas struct super_block *sb = ar->inode->i_sb; 4185c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 4186c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 4187c9de560dSAlex Tomas ext4_group_t group; 4188498e5f24STheodore Ts'o unsigned int len; 4189498e5f24STheodore Ts'o ext4_fsblk_t goal; 4190c9de560dSAlex Tomas ext4_grpblk_t block; 4191c9de560dSAlex Tomas 4192c9de560dSAlex Tomas /* we can't allocate > group size */ 4193c9de560dSAlex Tomas len = ar->len; 4194c9de560dSAlex Tomas 4195c9de560dSAlex Tomas /* just a dirty hack to filter too big requests */ 419640ae3487STheodore Ts'o if (len >= EXT4_CLUSTERS_PER_GROUP(sb)) 419740ae3487STheodore Ts'o len = EXT4_CLUSTERS_PER_GROUP(sb); 4198c9de560dSAlex Tomas 4199c9de560dSAlex Tomas /* start searching from the goal */ 4200c9de560dSAlex Tomas goal = ar->goal; 4201c9de560dSAlex Tomas if (goal < le32_to_cpu(es->s_first_data_block) || 4202c9de560dSAlex Tomas goal >= ext4_blocks_count(es)) 4203c9de560dSAlex Tomas goal = le32_to_cpu(es->s_first_data_block); 4204c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, goal, &group, &block); 4205c9de560dSAlex Tomas 4206c9de560dSAlex Tomas /* set up allocation goals */ 4207f5a44db5STheodore Ts'o ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical); 4208c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 4209c9de560dSAlex Tomas ac->ac_sb = sb; 4210c9de560dSAlex Tomas ac->ac_inode = ar->inode; 421153accfa9STheodore Ts'o ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; 4212c9de560dSAlex Tomas ac->ac_o_ex.fe_group = group; 4213c9de560dSAlex Tomas ac->ac_o_ex.fe_start = block; 4214c9de560dSAlex Tomas ac->ac_o_ex.fe_len = len; 421553accfa9STheodore Ts'o ac->ac_g_ex = ac->ac_o_ex; 4216c9de560dSAlex Tomas ac->ac_flags = ar->flags; 4217c9de560dSAlex Tomas 4218c9de560dSAlex Tomas /* we have to define context: we'll we work with a file or 4219c9de560dSAlex Tomas * locality group. this is a policy, actually */ 4220c9de560dSAlex Tomas ext4_mb_group_or_file(ac); 4221c9de560dSAlex Tomas 42226ba495e9STheodore Ts'o mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " 4223c9de560dSAlex Tomas "left: %u/%u, right %u/%u to %swritable\n", 4224c9de560dSAlex Tomas (unsigned) ar->len, (unsigned) ar->logical, 4225c9de560dSAlex Tomas (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 4226c9de560dSAlex Tomas (unsigned) ar->lleft, (unsigned) ar->pleft, 4227c9de560dSAlex Tomas (unsigned) ar->lright, (unsigned) ar->pright, 4228c9de560dSAlex Tomas atomic_read(&ar->inode->i_writecount) ? "" : "non-"); 4229c9de560dSAlex Tomas return 0; 4230c9de560dSAlex Tomas 4231c9de560dSAlex Tomas } 4232c9de560dSAlex Tomas 42336be2ded1SAneesh Kumar K.V static noinline_for_stack void 42346be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(struct super_block *sb, 42356be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg, 42366be2ded1SAneesh Kumar K.V int order, int total_entries) 42376be2ded1SAneesh Kumar K.V { 42386be2ded1SAneesh Kumar K.V ext4_group_t group = 0; 42396be2ded1SAneesh Kumar K.V struct ext4_buddy e4b; 42406be2ded1SAneesh Kumar K.V struct list_head discard_list; 42416be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa, *tmp; 42426be2ded1SAneesh Kumar K.V 42436ba495e9STheodore Ts'o mb_debug(1, "discard locality group preallocation\n"); 42446be2ded1SAneesh Kumar K.V 42456be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&discard_list); 42466be2ded1SAneesh Kumar K.V 42476be2ded1SAneesh Kumar K.V spin_lock(&lg->lg_prealloc_lock); 42486be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 42496be2ded1SAneesh Kumar K.V pa_inode_list) { 42506be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 42516be2ded1SAneesh Kumar K.V if (atomic_read(&pa->pa_count)) { 42526be2ded1SAneesh Kumar K.V /* 42536be2ded1SAneesh Kumar K.V * This is the pa that we just used 42546be2ded1SAneesh Kumar K.V * for block allocation. So don't 42556be2ded1SAneesh Kumar K.V * free that 42566be2ded1SAneesh Kumar K.V */ 42576be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 42586be2ded1SAneesh Kumar K.V continue; 42596be2ded1SAneesh Kumar K.V } 42606be2ded1SAneesh Kumar K.V if (pa->pa_deleted) { 42616be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 42626be2ded1SAneesh Kumar K.V continue; 42636be2ded1SAneesh Kumar K.V } 42646be2ded1SAneesh Kumar K.V /* only lg prealloc space */ 4265cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_GROUP_PA); 42666be2ded1SAneesh Kumar K.V 42676be2ded1SAneesh Kumar K.V /* seems this one can be freed ... */ 42686be2ded1SAneesh Kumar K.V pa->pa_deleted = 1; 42696be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 42706be2ded1SAneesh Kumar K.V 42716be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 42726be2ded1SAneesh Kumar K.V list_add(&pa->u.pa_tmp_list, &discard_list); 42736be2ded1SAneesh Kumar K.V 42746be2ded1SAneesh Kumar K.V total_entries--; 42756be2ded1SAneesh Kumar K.V if (total_entries <= 5) { 42766be2ded1SAneesh Kumar K.V /* 42776be2ded1SAneesh Kumar K.V * we want to keep only 5 entries 42786be2ded1SAneesh Kumar K.V * allowing it to grow to 8. This 42796be2ded1SAneesh Kumar K.V * mak sure we don't call discard 42806be2ded1SAneesh Kumar K.V * soon for this list. 42816be2ded1SAneesh Kumar K.V */ 42826be2ded1SAneesh Kumar K.V break; 42836be2ded1SAneesh Kumar K.V } 42846be2ded1SAneesh Kumar K.V } 42856be2ded1SAneesh Kumar K.V spin_unlock(&lg->lg_prealloc_lock); 42866be2ded1SAneesh Kumar K.V 42876be2ded1SAneesh Kumar K.V list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 42886be2ded1SAneesh Kumar K.V 4289bd86298eSLukas Czerner group = ext4_get_group_number(sb, pa->pa_pstart); 42906be2ded1SAneesh Kumar K.V if (ext4_mb_load_buddy(sb, group, &e4b)) { 429112062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", 429212062dddSEric Sandeen group); 42936be2ded1SAneesh Kumar K.V continue; 42946be2ded1SAneesh Kumar K.V } 42956be2ded1SAneesh Kumar K.V ext4_lock_group(sb, group); 42966be2ded1SAneesh Kumar K.V list_del(&pa->pa_group_list); 42973e1e5f50SEric Sandeen ext4_mb_release_group_pa(&e4b, pa); 42986be2ded1SAneesh Kumar K.V ext4_unlock_group(sb, group); 42996be2ded1SAneesh Kumar K.V 4300e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 43016be2ded1SAneesh Kumar K.V list_del(&pa->u.pa_tmp_list); 43026be2ded1SAneesh Kumar K.V call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 43036be2ded1SAneesh Kumar K.V } 43046be2ded1SAneesh Kumar K.V } 43056be2ded1SAneesh Kumar K.V 43066be2ded1SAneesh Kumar K.V /* 43076be2ded1SAneesh Kumar K.V * We have incremented pa_count. So it cannot be freed at this 43086be2ded1SAneesh Kumar K.V * point. Also we hold lg_mutex. So no parallel allocation is 43096be2ded1SAneesh Kumar K.V * possible from this lg. That means pa_free cannot be updated. 43106be2ded1SAneesh Kumar K.V * 43116be2ded1SAneesh Kumar K.V * A parallel ext4_mb_discard_group_preallocations is possible. 43126be2ded1SAneesh Kumar K.V * which can cause the lg_prealloc_list to be updated. 43136be2ded1SAneesh Kumar K.V */ 43146be2ded1SAneesh Kumar K.V 43156be2ded1SAneesh Kumar K.V static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 43166be2ded1SAneesh Kumar K.V { 43176be2ded1SAneesh Kumar K.V int order, added = 0, lg_prealloc_count = 1; 43186be2ded1SAneesh Kumar K.V struct super_block *sb = ac->ac_sb; 43196be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg = ac->ac_lg; 43206be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 43216be2ded1SAneesh Kumar K.V 43226be2ded1SAneesh Kumar K.V order = fls(pa->pa_free) - 1; 43236be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 43246be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 43256be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 43266be2ded1SAneesh Kumar K.V /* Add the prealloc space to lg */ 4327f1167009SNiu Yawei spin_lock(&lg->lg_prealloc_lock); 43286be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 43296be2ded1SAneesh Kumar K.V pa_inode_list) { 43306be2ded1SAneesh Kumar K.V spin_lock(&tmp_pa->pa_lock); 43316be2ded1SAneesh Kumar K.V if (tmp_pa->pa_deleted) { 4332e7c9e3e9STheodore Ts'o spin_unlock(&tmp_pa->pa_lock); 43336be2ded1SAneesh Kumar K.V continue; 43346be2ded1SAneesh Kumar K.V } 43356be2ded1SAneesh Kumar K.V if (!added && pa->pa_free < tmp_pa->pa_free) { 43366be2ded1SAneesh Kumar K.V /* Add to the tail of the previous entry */ 43376be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 43386be2ded1SAneesh Kumar K.V &tmp_pa->pa_inode_list); 43396be2ded1SAneesh Kumar K.V added = 1; 43406be2ded1SAneesh Kumar K.V /* 43416be2ded1SAneesh Kumar K.V * we want to count the total 43426be2ded1SAneesh Kumar K.V * number of entries in the list 43436be2ded1SAneesh Kumar K.V */ 43446be2ded1SAneesh Kumar K.V } 43456be2ded1SAneesh Kumar K.V spin_unlock(&tmp_pa->pa_lock); 43466be2ded1SAneesh Kumar K.V lg_prealloc_count++; 43476be2ded1SAneesh Kumar K.V } 43486be2ded1SAneesh Kumar K.V if (!added) 43496be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 43506be2ded1SAneesh Kumar K.V &lg->lg_prealloc_list[order]); 4351f1167009SNiu Yawei spin_unlock(&lg->lg_prealloc_lock); 43526be2ded1SAneesh Kumar K.V 43536be2ded1SAneesh Kumar K.V /* Now trim the list to be not more than 8 elements */ 43546be2ded1SAneesh Kumar K.V if (lg_prealloc_count > 8) { 43556be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(sb, lg, 43566be2ded1SAneesh Kumar K.V order, lg_prealloc_count); 43576be2ded1SAneesh Kumar K.V return; 43586be2ded1SAneesh Kumar K.V } 43596be2ded1SAneesh Kumar K.V return ; 43606be2ded1SAneesh Kumar K.V } 43616be2ded1SAneesh Kumar K.V 4362c9de560dSAlex Tomas /* 4363c9de560dSAlex Tomas * release all resource we used in allocation 4364c9de560dSAlex Tomas */ 4365c9de560dSAlex Tomas static int ext4_mb_release_context(struct ext4_allocation_context *ac) 4366c9de560dSAlex Tomas { 436753accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 43686be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa = ac->ac_pa; 43696be2ded1SAneesh Kumar K.V if (pa) { 4370cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) { 4371c9de560dSAlex Tomas /* see comment in ext4_mb_use_group_pa() */ 43726be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 437353accfa9STheodore Ts'o pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 437453accfa9STheodore Ts'o pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 43756be2ded1SAneesh Kumar K.V pa->pa_free -= ac->ac_b_ex.fe_len; 43766be2ded1SAneesh Kumar K.V pa->pa_len -= ac->ac_b_ex.fe_len; 43776be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 4378ba443916SAneesh Kumar K.V } 4379ba443916SAneesh Kumar K.V } 4380ba443916SAneesh Kumar K.V if (pa) { 43816be2ded1SAneesh Kumar K.V /* 43826be2ded1SAneesh Kumar K.V * We want to add the pa to the right bucket. 43836be2ded1SAneesh Kumar K.V * Remove it from the list and while adding 43846be2ded1SAneesh Kumar K.V * make sure the list to which we are adding 438544183d42SAmir Goldstein * doesn't grow big. 43866be2ded1SAneesh Kumar K.V */ 4387cc0fb9adSAneesh Kumar K.V if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { 43886be2ded1SAneesh Kumar K.V spin_lock(pa->pa_obj_lock); 43896be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 43906be2ded1SAneesh Kumar K.V spin_unlock(pa->pa_obj_lock); 43916be2ded1SAneesh Kumar K.V ext4_mb_add_n_trim(ac); 4392c9de560dSAlex Tomas } 43936be2ded1SAneesh Kumar K.V ext4_mb_put_pa(ac, ac->ac_sb, pa); 4394c9de560dSAlex Tomas } 4395c9de560dSAlex Tomas if (ac->ac_bitmap_page) 439609cbfeafSKirill A. Shutemov put_page(ac->ac_bitmap_page); 4397c9de560dSAlex Tomas if (ac->ac_buddy_page) 439809cbfeafSKirill A. Shutemov put_page(ac->ac_buddy_page); 4399c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 4400c9de560dSAlex Tomas mutex_unlock(&ac->ac_lg->lg_mutex); 4401c9de560dSAlex Tomas ext4_mb_collect_stats(ac); 4402c9de560dSAlex Tomas return 0; 4403c9de560dSAlex Tomas } 4404c9de560dSAlex Tomas 4405c9de560dSAlex Tomas static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 4406c9de560dSAlex Tomas { 44078df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 4408c9de560dSAlex Tomas int ret; 4409c9de560dSAlex Tomas int freed = 0; 4410c9de560dSAlex Tomas 44119bffad1eSTheodore Ts'o trace_ext4_mb_discard_preallocations(sb, needed); 44128df9675fSTheodore Ts'o for (i = 0; i < ngroups && needed > 0; i++) { 4413c9de560dSAlex Tomas ret = ext4_mb_discard_group_preallocations(sb, i, needed); 4414c9de560dSAlex Tomas freed += ret; 4415c9de560dSAlex Tomas needed -= ret; 4416c9de560dSAlex Tomas } 4417c9de560dSAlex Tomas 4418c9de560dSAlex Tomas return freed; 4419c9de560dSAlex Tomas } 4420c9de560dSAlex Tomas 4421c9de560dSAlex Tomas /* 4422c9de560dSAlex Tomas * Main entry point into mballoc to allocate blocks 4423c9de560dSAlex Tomas * it tries to use preallocation first, then falls back 4424c9de560dSAlex Tomas * to usual allocation 4425c9de560dSAlex Tomas */ 4426c9de560dSAlex Tomas ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 4427c9de560dSAlex Tomas struct ext4_allocation_request *ar, int *errp) 4428c9de560dSAlex Tomas { 44296bc6e63fSAneesh Kumar K.V int freed; 4430256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 4431c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4432c9de560dSAlex Tomas struct super_block *sb; 4433c9de560dSAlex Tomas ext4_fsblk_t block = 0; 443460e58e0fSMingming Cao unsigned int inquota = 0; 443553accfa9STheodore Ts'o unsigned int reserv_clstrs = 0; 4436c9de560dSAlex Tomas 4437b10a44c3STheodore Ts'o might_sleep(); 4438c9de560dSAlex Tomas sb = ar->inode->i_sb; 4439c9de560dSAlex Tomas sbi = EXT4_SB(sb); 4440c9de560dSAlex Tomas 44419bffad1eSTheodore Ts'o trace_ext4_request_blocks(ar); 4442ba80b101STheodore Ts'o 444345dc63e7SDmitry Monakhov /* Allow to use superuser reservation for quota file */ 444445dc63e7SDmitry Monakhov if (IS_NOQUOTA(ar->inode)) 444545dc63e7SDmitry Monakhov ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; 444645dc63e7SDmitry Monakhov 4447e3cf5d5dSTheodore Ts'o if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) { 444860e58e0fSMingming Cao /* Without delayed allocation we need to verify 444960e58e0fSMingming Cao * there is enough free blocks to do block allocation 445060e58e0fSMingming Cao * and verify allocation doesn't exceed the quota limits. 4451d2a17637SMingming Cao */ 445255f020dbSAllison Henderson while (ar->len && 4453e7d5f315STheodore Ts'o ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { 445455f020dbSAllison Henderson 4455030ba6bcSAneesh Kumar K.V /* let others to free the space */ 4456bb8b20edSLukas Czerner cond_resched(); 4457030ba6bcSAneesh Kumar K.V ar->len = ar->len >> 1; 4458030ba6bcSAneesh Kumar K.V } 4459030ba6bcSAneesh Kumar K.V if (!ar->len) { 446007031431SMingming Cao *errp = -ENOSPC; 446107031431SMingming Cao return 0; 446207031431SMingming Cao } 446353accfa9STheodore Ts'o reserv_clstrs = ar->len; 446455f020dbSAllison Henderson if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { 446553accfa9STheodore Ts'o dquot_alloc_block_nofail(ar->inode, 446653accfa9STheodore Ts'o EXT4_C2B(sbi, ar->len)); 446755f020dbSAllison Henderson } else { 446855f020dbSAllison Henderson while (ar->len && 446953accfa9STheodore Ts'o dquot_alloc_block(ar->inode, 447053accfa9STheodore Ts'o EXT4_C2B(sbi, ar->len))) { 447155f020dbSAllison Henderson 4472c9de560dSAlex Tomas ar->flags |= EXT4_MB_HINT_NOPREALLOC; 4473c9de560dSAlex Tomas ar->len--; 4474c9de560dSAlex Tomas } 447555f020dbSAllison Henderson } 447660e58e0fSMingming Cao inquota = ar->len; 4477c9de560dSAlex Tomas if (ar->len == 0) { 4478c9de560dSAlex Tomas *errp = -EDQUOT; 44796c7a120aSAditya Kali goto out; 4480c9de560dSAlex Tomas } 448160e58e0fSMingming Cao } 4482d2a17637SMingming Cao 448385556c9aSWei Yongjun ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS); 4484833576b3STheodore Ts'o if (!ac) { 4485363d4251SShen Feng ar->len = 0; 4486256bdb49SEric Sandeen *errp = -ENOMEM; 44876c7a120aSAditya Kali goto out; 4488256bdb49SEric Sandeen } 4489256bdb49SEric Sandeen 4490256bdb49SEric Sandeen *errp = ext4_mb_initialize_context(ac, ar); 4491c9de560dSAlex Tomas if (*errp) { 4492c9de560dSAlex Tomas ar->len = 0; 44936c7a120aSAditya Kali goto out; 4494c9de560dSAlex Tomas } 4495c9de560dSAlex Tomas 4496256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 4497256bdb49SEric Sandeen if (!ext4_mb_use_preallocated(ac)) { 4498256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_ALLOC; 4499256bdb49SEric Sandeen ext4_mb_normalize_request(ac, ar); 4500c9de560dSAlex Tomas repeat: 4501c9de560dSAlex Tomas /* allocate space in core */ 45026c7a120aSAditya Kali *errp = ext4_mb_regular_allocator(ac); 45032c00ef3eSAlexey Khoroshilov if (*errp) 45042c00ef3eSAlexey Khoroshilov goto discard_and_exit; 4505c9de560dSAlex Tomas 4506c9de560dSAlex Tomas /* as we've just preallocated more space than 45072c00ef3eSAlexey Khoroshilov * user requested originally, we store allocated 4508c9de560dSAlex Tomas * space in a special descriptor */ 4509256bdb49SEric Sandeen if (ac->ac_status == AC_STATUS_FOUND && 4510256bdb49SEric Sandeen ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 45112c00ef3eSAlexey Khoroshilov *errp = ext4_mb_new_preallocation(ac); 45122c00ef3eSAlexey Khoroshilov if (*errp) { 45132c00ef3eSAlexey Khoroshilov discard_and_exit: 45142c00ef3eSAlexey Khoroshilov ext4_discard_allocated_blocks(ac); 45152c00ef3eSAlexey Khoroshilov goto errout; 45162c00ef3eSAlexey Khoroshilov } 4517c9de560dSAlex Tomas } 4518256bdb49SEric Sandeen if (likely(ac->ac_status == AC_STATUS_FOUND)) { 451953accfa9STheodore Ts'o *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); 4520519deca0SAneesh Kumar K.V if (*errp == -EAGAIN) { 45218556e8f3SAneesh Kumar K.V /* 45228556e8f3SAneesh Kumar K.V * drop the reference that we took 45238556e8f3SAneesh Kumar K.V * in ext4_mb_use_best_found 45248556e8f3SAneesh Kumar K.V */ 45258556e8f3SAneesh Kumar K.V ext4_mb_release_context(ac); 4526519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_group = 0; 4527519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_start = 0; 4528519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len = 0; 4529519deca0SAneesh Kumar K.V ac->ac_status = AC_STATUS_CONTINUE; 4530519deca0SAneesh Kumar K.V goto repeat; 45316d138cedSEric Sandeen } else if (*errp) { 4532b844167eSCurt Wohlgemuth ext4_discard_allocated_blocks(ac); 45336d138cedSEric Sandeen goto errout; 45346d138cedSEric Sandeen } else { 4535256bdb49SEric Sandeen block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 4536256bdb49SEric Sandeen ar->len = ac->ac_b_ex.fe_len; 4537519deca0SAneesh Kumar K.V } 4538c9de560dSAlex Tomas } else { 4539256bdb49SEric Sandeen freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 4540c9de560dSAlex Tomas if (freed) 4541c9de560dSAlex Tomas goto repeat; 4542c9de560dSAlex Tomas *errp = -ENOSPC; 45436c7a120aSAditya Kali } 45446c7a120aSAditya Kali 45456d138cedSEric Sandeen errout: 45466c7a120aSAditya Kali if (*errp) { 4547256bdb49SEric Sandeen ac->ac_b_ex.fe_len = 0; 4548c9de560dSAlex Tomas ar->len = 0; 4549256bdb49SEric Sandeen ext4_mb_show_ac(ac); 4550c9de560dSAlex Tomas } 4551256bdb49SEric Sandeen ext4_mb_release_context(ac); 45526c7a120aSAditya Kali out: 45536c7a120aSAditya Kali if (ac) 4554363d4251SShen Feng kmem_cache_free(ext4_ac_cachep, ac); 455560e58e0fSMingming Cao if (inquota && ar->len < inquota) 455653accfa9STheodore Ts'o dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); 45570087d9fbSAneesh Kumar K.V if (!ar->len) { 4558e3cf5d5dSTheodore Ts'o if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) 45590087d9fbSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 456057042651STheodore Ts'o percpu_counter_sub(&sbi->s_dirtyclusters_counter, 456153accfa9STheodore Ts'o reserv_clstrs); 45620087d9fbSAneesh Kumar K.V } 4563c9de560dSAlex Tomas 45649bffad1eSTheodore Ts'o trace_ext4_allocate_blocks(ar, (unsigned long long)block); 4565ba80b101STheodore Ts'o 4566c9de560dSAlex Tomas return block; 4567c9de560dSAlex Tomas } 4568c9de560dSAlex Tomas 4569c894058dSAneesh Kumar K.V /* 4570c894058dSAneesh Kumar K.V * We can merge two free data extents only if the physical blocks 4571c894058dSAneesh Kumar K.V * are contiguous, AND the extents were freed by the same transaction, 4572c894058dSAneesh Kumar K.V * AND the blocks are associated with the same group. 4573c894058dSAneesh Kumar K.V */ 4574c894058dSAneesh Kumar K.V static int can_merge(struct ext4_free_data *entry1, 4575c894058dSAneesh Kumar K.V struct ext4_free_data *entry2) 4576c894058dSAneesh Kumar K.V { 457718aadd47SBobi Jam if ((entry1->efd_tid == entry2->efd_tid) && 457818aadd47SBobi Jam (entry1->efd_group == entry2->efd_group) && 457918aadd47SBobi Jam ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster)) 4580c894058dSAneesh Kumar K.V return 1; 4581c894058dSAneesh Kumar K.V return 0; 4582c894058dSAneesh Kumar K.V } 4583c894058dSAneesh Kumar K.V 45844ddfef7bSEric Sandeen static noinline_for_stack int 45854ddfef7bSEric Sandeen ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 45867a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry) 4587c9de560dSAlex Tomas { 4588e29136f8STheodore Ts'o ext4_group_t group = e4b->bd_group; 458984130193STheodore Ts'o ext4_grpblk_t cluster; 4590*d08854f5STheodore Ts'o ext4_grpblk_t clusters = new_entry->efd_count; 45917a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 4592c9de560dSAlex Tomas struct ext4_group_info *db = e4b->bd_info; 4593c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 4594c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 4595c894058dSAneesh Kumar K.V struct rb_node **n = &db->bb_free_root.rb_node, *node; 4596c894058dSAneesh Kumar K.V struct rb_node *parent = NULL, *new_node; 4597c894058dSAneesh Kumar K.V 45980390131bSFrank Mayhar BUG_ON(!ext4_handle_valid(handle)); 4599c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 4600c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 4601c9de560dSAlex Tomas 460218aadd47SBobi Jam new_node = &new_entry->efd_node; 460318aadd47SBobi Jam cluster = new_entry->efd_start_cluster; 4604c9de560dSAlex Tomas 4605c894058dSAneesh Kumar K.V if (!*n) { 4606c894058dSAneesh Kumar K.V /* first free block exent. We need to 4607c894058dSAneesh Kumar K.V protect buddy cache from being freed, 4608c9de560dSAlex Tomas * otherwise we'll refresh it from 4609c9de560dSAlex Tomas * on-disk bitmap and lose not-yet-available 4610c9de560dSAlex Tomas * blocks */ 461109cbfeafSKirill A. Shutemov get_page(e4b->bd_buddy_page); 461209cbfeafSKirill A. Shutemov get_page(e4b->bd_bitmap_page); 4613c894058dSAneesh Kumar K.V } 4614c894058dSAneesh Kumar K.V while (*n) { 4615c894058dSAneesh Kumar K.V parent = *n; 461618aadd47SBobi Jam entry = rb_entry(parent, struct ext4_free_data, efd_node); 461718aadd47SBobi Jam if (cluster < entry->efd_start_cluster) 4618c894058dSAneesh Kumar K.V n = &(*n)->rb_left; 461918aadd47SBobi Jam else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) 4620c894058dSAneesh Kumar K.V n = &(*n)->rb_right; 4621c894058dSAneesh Kumar K.V else { 4622e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 462384130193STheodore Ts'o ext4_group_first_block_no(sb, group) + 462484130193STheodore Ts'o EXT4_C2B(sbi, cluster), 4625e29136f8STheodore Ts'o "Block already on to-be-freed list"); 4626c894058dSAneesh Kumar K.V return 0; 4627c9de560dSAlex Tomas } 4628c9de560dSAlex Tomas } 4629c9de560dSAlex Tomas 4630c894058dSAneesh Kumar K.V rb_link_node(new_node, parent, n); 4631c894058dSAneesh Kumar K.V rb_insert_color(new_node, &db->bb_free_root); 4632c894058dSAneesh Kumar K.V 4633c894058dSAneesh Kumar K.V /* Now try to see the extent can be merged to left and right */ 4634c894058dSAneesh Kumar K.V node = rb_prev(new_node); 4635c894058dSAneesh Kumar K.V if (node) { 463618aadd47SBobi Jam entry = rb_entry(node, struct ext4_free_data, efd_node); 46375d3ee208SDmitry Monakhov if (can_merge(entry, new_entry) && 46385d3ee208SDmitry Monakhov ext4_journal_callback_try_del(handle, &entry->efd_jce)) { 463918aadd47SBobi Jam new_entry->efd_start_cluster = entry->efd_start_cluster; 464018aadd47SBobi Jam new_entry->efd_count += entry->efd_count; 4641c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 464218aadd47SBobi Jam kmem_cache_free(ext4_free_data_cachep, entry); 4643c9de560dSAlex Tomas } 4644c9de560dSAlex Tomas } 4645c894058dSAneesh Kumar K.V 4646c894058dSAneesh Kumar K.V node = rb_next(new_node); 4647c894058dSAneesh Kumar K.V if (node) { 464818aadd47SBobi Jam entry = rb_entry(node, struct ext4_free_data, efd_node); 46495d3ee208SDmitry Monakhov if (can_merge(new_entry, entry) && 46505d3ee208SDmitry Monakhov ext4_journal_callback_try_del(handle, &entry->efd_jce)) { 465118aadd47SBobi Jam new_entry->efd_count += entry->efd_count; 4652c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 465318aadd47SBobi Jam kmem_cache_free(ext4_free_data_cachep, entry); 4654c894058dSAneesh Kumar K.V } 4655c894058dSAneesh Kumar K.V } 46563e624fc7STheodore Ts'o /* Add the extent to transaction's private list */ 4657*d08854f5STheodore Ts'o new_entry->efd_jce.jce_func = ext4_free_data_callback; 4658*d08854f5STheodore Ts'o spin_lock(&sbi->s_md_lock); 4659*d08854f5STheodore Ts'o _ext4_journal_callback_add(handle, &new_entry->efd_jce); 4660*d08854f5STheodore Ts'o sbi->s_mb_free_pending += clusters; 4661*d08854f5STheodore Ts'o spin_unlock(&sbi->s_md_lock); 4662c9de560dSAlex Tomas return 0; 4663c9de560dSAlex Tomas } 4664c9de560dSAlex Tomas 466544338711STheodore Ts'o /** 466644338711STheodore Ts'o * ext4_free_blocks() -- Free given blocks and update quota 466744338711STheodore Ts'o * @handle: handle for this transaction 466844338711STheodore Ts'o * @inode: inode 466944338711STheodore Ts'o * @block: start physical block to free 467044338711STheodore Ts'o * @count: number of blocks to count 46715def1360SYongqiang Yang * @flags: flags used by ext4_free_blocks 4672c9de560dSAlex Tomas */ 467344338711STheodore Ts'o void ext4_free_blocks(handle_t *handle, struct inode *inode, 4674e6362609STheodore Ts'o struct buffer_head *bh, ext4_fsblk_t block, 4675e6362609STheodore Ts'o unsigned long count, int flags) 4676c9de560dSAlex Tomas { 467726346ff6SAneesh Kumar K.V struct buffer_head *bitmap_bh = NULL; 4678c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 4679c9de560dSAlex Tomas struct ext4_group_desc *gdp; 4680498e5f24STheodore Ts'o unsigned int overflow; 4681c9de560dSAlex Tomas ext4_grpblk_t bit; 4682c9de560dSAlex Tomas struct buffer_head *gd_bh; 4683c9de560dSAlex Tomas ext4_group_t block_group; 4684c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4685c9de560dSAlex Tomas struct ext4_buddy e4b; 468684130193STheodore Ts'o unsigned int count_clusters; 4687c9de560dSAlex Tomas int err = 0; 4688c9de560dSAlex Tomas int ret; 4689c9de560dSAlex Tomas 4690b10a44c3STheodore Ts'o might_sleep(); 4691e6362609STheodore Ts'o if (bh) { 4692e6362609STheodore Ts'o if (block) 4693e6362609STheodore Ts'o BUG_ON(block != bh->b_blocknr); 4694e6362609STheodore Ts'o else 4695e6362609STheodore Ts'o block = bh->b_blocknr; 4696e6362609STheodore Ts'o } 4697c9de560dSAlex Tomas 4698c9de560dSAlex Tomas sbi = EXT4_SB(sb); 46991f2acb60STheodore Ts'o if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 47001f2acb60STheodore Ts'o !ext4_data_block_valid(sbi, block, count)) { 470112062dddSEric Sandeen ext4_error(sb, "Freeing blocks not in datazone - " 47020610b6e9STheodore Ts'o "block = %llu, count = %lu", block, count); 4703c9de560dSAlex Tomas goto error_return; 4704c9de560dSAlex Tomas } 4705c9de560dSAlex Tomas 47060610b6e9STheodore Ts'o ext4_debug("freeing block %llu\n", block); 4707e6362609STheodore Ts'o trace_ext4_free_blocks(inode, block, count, flags); 4708e6362609STheodore Ts'o 47099c02ac97SDaeho Jeong if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 47109c02ac97SDaeho Jeong BUG_ON(count > 1); 4711e6362609STheodore Ts'o 4712e6362609STheodore Ts'o ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, 47139c02ac97SDaeho Jeong inode, bh, block); 4714e6362609STheodore Ts'o } 4715e6362609STheodore Ts'o 4716e6362609STheodore Ts'o /* 471784130193STheodore Ts'o * If the extent to be freed does not begin on a cluster 471884130193STheodore Ts'o * boundary, we need to deal with partial clusters at the 471984130193STheodore Ts'o * beginning and end of the extent. Normally we will free 472084130193STheodore Ts'o * blocks at the beginning or the end unless we are explicitly 472184130193STheodore Ts'o * requested to avoid doing so. 472284130193STheodore Ts'o */ 4723f5a44db5STheodore Ts'o overflow = EXT4_PBLK_COFF(sbi, block); 472484130193STheodore Ts'o if (overflow) { 472584130193STheodore Ts'o if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { 472684130193STheodore Ts'o overflow = sbi->s_cluster_ratio - overflow; 472784130193STheodore Ts'o block += overflow; 472884130193STheodore Ts'o if (count > overflow) 472984130193STheodore Ts'o count -= overflow; 473084130193STheodore Ts'o else 473184130193STheodore Ts'o return; 473284130193STheodore Ts'o } else { 473384130193STheodore Ts'o block -= overflow; 473484130193STheodore Ts'o count += overflow; 473584130193STheodore Ts'o } 473684130193STheodore Ts'o } 4737f5a44db5STheodore Ts'o overflow = EXT4_LBLK_COFF(sbi, count); 473884130193STheodore Ts'o if (overflow) { 473984130193STheodore Ts'o if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { 474084130193STheodore Ts'o if (count > overflow) 474184130193STheodore Ts'o count -= overflow; 474284130193STheodore Ts'o else 474384130193STheodore Ts'o return; 474484130193STheodore Ts'o } else 474584130193STheodore Ts'o count += sbi->s_cluster_ratio - overflow; 474684130193STheodore Ts'o } 474784130193STheodore Ts'o 47489c02ac97SDaeho Jeong if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) { 47499c02ac97SDaeho Jeong int i; 4750f96c450dSDaeho Jeong int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA; 47519c02ac97SDaeho Jeong 47529c02ac97SDaeho Jeong for (i = 0; i < count; i++) { 47539c02ac97SDaeho Jeong cond_resched(); 4754f96c450dSDaeho Jeong if (is_metadata) 47559c02ac97SDaeho Jeong bh = sb_find_get_block(inode->i_sb, block + i); 4756f96c450dSDaeho Jeong ext4_forget(handle, is_metadata, inode, bh, block + i); 47579c02ac97SDaeho Jeong } 47589c02ac97SDaeho Jeong } 47599c02ac97SDaeho Jeong 4760c9de560dSAlex Tomas do_more: 4761c9de560dSAlex Tomas overflow = 0; 4762c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 4763c9de560dSAlex Tomas 4764163a203dSDarrick J. Wong if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT( 4765163a203dSDarrick J. Wong ext4_get_group_info(sb, block_group)))) 4766163a203dSDarrick J. Wong return; 4767163a203dSDarrick J. Wong 4768c9de560dSAlex Tomas /* 4769c9de560dSAlex Tomas * Check to see if we are freeing blocks across a group 4770c9de560dSAlex Tomas * boundary. 4771c9de560dSAlex Tomas */ 477284130193STheodore Ts'o if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { 477384130193STheodore Ts'o overflow = EXT4_C2B(sbi, bit) + count - 477484130193STheodore Ts'o EXT4_BLOCKS_PER_GROUP(sb); 4775c9de560dSAlex Tomas count -= overflow; 4776c9de560dSAlex Tomas } 4777810da240SLukas Czerner count_clusters = EXT4_NUM_B2C(sbi, count); 4778574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, block_group); 47799008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 47809008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 47819008a58eSDarrick J. Wong bitmap_bh = NULL; 4782c9de560dSAlex Tomas goto error_return; 4783ce89f46cSAneesh Kumar K.V } 4784c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 4785ce89f46cSAneesh Kumar K.V if (!gdp) { 4786ce89f46cSAneesh Kumar K.V err = -EIO; 4787c9de560dSAlex Tomas goto error_return; 4788ce89f46cSAneesh Kumar K.V } 4789c9de560dSAlex Tomas 4790c9de560dSAlex Tomas if (in_range(ext4_block_bitmap(sb, gdp), block, count) || 4791c9de560dSAlex Tomas in_range(ext4_inode_bitmap(sb, gdp), block, count) || 4792c9de560dSAlex Tomas in_range(block, ext4_inode_table(sb, gdp), 4793c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group) || 4794c9de560dSAlex Tomas in_range(block + count - 1, ext4_inode_table(sb, gdp), 4795c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group)) { 4796c9de560dSAlex Tomas 479712062dddSEric Sandeen ext4_error(sb, "Freeing blocks in system zone - " 47980610b6e9STheodore Ts'o "Block = %llu, count = %lu", block, count); 4799519deca0SAneesh Kumar K.V /* err = 0. ext4_std_error should be a no op */ 4800519deca0SAneesh Kumar K.V goto error_return; 4801c9de560dSAlex Tomas } 4802c9de560dSAlex Tomas 4803c9de560dSAlex Tomas BUFFER_TRACE(bitmap_bh, "getting write access"); 4804c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 4805c9de560dSAlex Tomas if (err) 4806c9de560dSAlex Tomas goto error_return; 4807c9de560dSAlex Tomas 4808c9de560dSAlex Tomas /* 4809c9de560dSAlex Tomas * We are about to modify some metadata. Call the journal APIs 4810c9de560dSAlex Tomas * to unshare ->b_data if a currently-committing transaction is 4811c9de560dSAlex Tomas * using it 4812c9de560dSAlex Tomas */ 4813c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "get_write_access"); 4814c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gd_bh); 4815c9de560dSAlex Tomas if (err) 4816c9de560dSAlex Tomas goto error_return; 4817c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 4818c9de560dSAlex Tomas { 4819c9de560dSAlex Tomas int i; 482084130193STheodore Ts'o for (i = 0; i < count_clusters; i++) 4821c9de560dSAlex Tomas BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 4822c9de560dSAlex Tomas } 4823c9de560dSAlex Tomas #endif 482484130193STheodore Ts'o trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); 4825c9de560dSAlex Tomas 4826adb7ef60SKonstantin Khlebnikov /* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */ 4827adb7ef60SKonstantin Khlebnikov err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b, 4828adb7ef60SKonstantin Khlebnikov GFP_NOFS|__GFP_NOFAIL); 4829920313a7SAneesh Kumar K.V if (err) 4830920313a7SAneesh Kumar K.V goto error_return; 4831e6362609STheodore Ts'o 4832f96c450dSDaeho Jeong /* 4833f96c450dSDaeho Jeong * We need to make sure we don't reuse the freed block until after the 4834f96c450dSDaeho Jeong * transaction is committed. We make an exception if the inode is to be 4835f96c450dSDaeho Jeong * written in writeback mode since writeback mode has weak data 4836f96c450dSDaeho Jeong * consistency guarantees. 4837f96c450dSDaeho Jeong */ 4838f96c450dSDaeho Jeong if (ext4_handle_valid(handle) && 4839f96c450dSDaeho Jeong ((flags & EXT4_FREE_BLOCKS_METADATA) || 4840f96c450dSDaeho Jeong !ext4_should_writeback_data(inode))) { 48417a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry; 48427a2fcbf7SAneesh Kumar K.V /* 48437444a072SMichal Hocko * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed 48447444a072SMichal Hocko * to fail. 48457a2fcbf7SAneesh Kumar K.V */ 48467444a072SMichal Hocko new_entry = kmem_cache_alloc(ext4_free_data_cachep, 48477444a072SMichal Hocko GFP_NOFS|__GFP_NOFAIL); 484818aadd47SBobi Jam new_entry->efd_start_cluster = bit; 484918aadd47SBobi Jam new_entry->efd_group = block_group; 485018aadd47SBobi Jam new_entry->efd_count = count_clusters; 485118aadd47SBobi Jam new_entry->efd_tid = handle->h_transaction->t_tid; 4852955ce5f5SAneesh Kumar K.V 48537a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, block_group); 485484130193STheodore Ts'o mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 48557a2fcbf7SAneesh Kumar K.V ext4_mb_free_metadata(handle, &e4b, new_entry); 4856c9de560dSAlex Tomas } else { 48577a2fcbf7SAneesh Kumar K.V /* need to update group_info->bb_free and bitmap 48587a2fcbf7SAneesh Kumar K.V * with group lock held. generate_buddy look at 48597a2fcbf7SAneesh Kumar K.V * them with group lock_held 48607a2fcbf7SAneesh Kumar K.V */ 4861d71c1ae2SLukas Czerner if (test_opt(sb, DISCARD)) { 4862d71c1ae2SLukas Czerner err = ext4_issue_discard(sb, block_group, bit, count); 4863d71c1ae2SLukas Czerner if (err && err != -EOPNOTSUPP) 4864d71c1ae2SLukas Czerner ext4_msg(sb, KERN_WARNING, "discard request in" 4865d71c1ae2SLukas Czerner " group:%d block:%d count:%lu failed" 4866d71c1ae2SLukas Czerner " with %d", block_group, bit, count, 4867d71c1ae2SLukas Czerner err); 48688f9ff189SLukas Czerner } else 48698f9ff189SLukas Czerner EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info); 4870d71c1ae2SLukas Czerner 4871955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 487284130193STheodore Ts'o mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 487384130193STheodore Ts'o mb_free_blocks(inode, &e4b, bit, count_clusters); 4874c9de560dSAlex Tomas } 4875c9de560dSAlex Tomas 4876021b65bbSTheodore Ts'o ret = ext4_free_group_clusters(sb, gdp) + count_clusters; 4877021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, ret); 487879f1ba49STao Ma ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh); 4879feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, block_group, gdp); 4880955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 4881c9de560dSAlex Tomas 4882772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 4883772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 488490ba983fSTheodore Ts'o atomic64_add(count_clusters, 488524aaa8efSTheodore Ts'o &sbi->s_flex_groups[flex_group].free_clusters); 4886772cb7c8SJose R. Santos } 4887772cb7c8SJose R. Santos 48887b415bf6SAditya Kali if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) 48897b415bf6SAditya Kali dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); 48907d734532SJan Kara percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters); 48917d734532SJan Kara 48927d734532SJan Kara ext4_mb_unload_buddy(&e4b); 48937b415bf6SAditya Kali 48947a2fcbf7SAneesh Kumar K.V /* We dirtied the bitmap block */ 48957a2fcbf7SAneesh Kumar K.V BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 48967a2fcbf7SAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 48977a2fcbf7SAneesh Kumar K.V 4898c9de560dSAlex Tomas /* And the group descriptor block */ 4899c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 49000390131bSFrank Mayhar ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 4901c9de560dSAlex Tomas if (!err) 4902c9de560dSAlex Tomas err = ret; 4903c9de560dSAlex Tomas 4904c9de560dSAlex Tomas if (overflow && !err) { 4905c9de560dSAlex Tomas block += count; 4906c9de560dSAlex Tomas count = overflow; 4907c9de560dSAlex Tomas put_bh(bitmap_bh); 4908c9de560dSAlex Tomas goto do_more; 4909c9de560dSAlex Tomas } 4910c9de560dSAlex Tomas error_return: 4911c9de560dSAlex Tomas brelse(bitmap_bh); 4912c9de560dSAlex Tomas ext4_std_error(sb, err); 4913c9de560dSAlex Tomas return; 4914c9de560dSAlex Tomas } 49157360d173SLukas Czerner 49167360d173SLukas Czerner /** 49170529155eSYongqiang Yang * ext4_group_add_blocks() -- Add given blocks to an existing group 49182846e820SAmir Goldstein * @handle: handle to this transaction 49192846e820SAmir Goldstein * @sb: super block 49204907cb7bSAnatol Pomozov * @block: start physical block to add to the block group 49212846e820SAmir Goldstein * @count: number of blocks to free 49222846e820SAmir Goldstein * 4923e73a347bSAmir Goldstein * This marks the blocks as free in the bitmap and buddy. 49242846e820SAmir Goldstein */ 4925cc7365dfSYongqiang Yang int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, 49262846e820SAmir Goldstein ext4_fsblk_t block, unsigned long count) 49272846e820SAmir Goldstein { 49282846e820SAmir Goldstein struct buffer_head *bitmap_bh = NULL; 49292846e820SAmir Goldstein struct buffer_head *gd_bh; 49302846e820SAmir Goldstein ext4_group_t block_group; 49312846e820SAmir Goldstein ext4_grpblk_t bit; 49322846e820SAmir Goldstein unsigned int i; 49332846e820SAmir Goldstein struct ext4_group_desc *desc; 49342846e820SAmir Goldstein struct ext4_sb_info *sbi = EXT4_SB(sb); 4935e73a347bSAmir Goldstein struct ext4_buddy e4b; 49362846e820SAmir Goldstein int err = 0, ret, blk_free_count; 49372846e820SAmir Goldstein ext4_grpblk_t blocks_freed; 49382846e820SAmir Goldstein 49392846e820SAmir Goldstein ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); 49402846e820SAmir Goldstein 49414740b830SYongqiang Yang if (count == 0) 49424740b830SYongqiang Yang return 0; 49434740b830SYongqiang Yang 49442846e820SAmir Goldstein ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 49452846e820SAmir Goldstein /* 49462846e820SAmir Goldstein * Check to see if we are freeing blocks across a group 49472846e820SAmir Goldstein * boundary. 49482846e820SAmir Goldstein */ 4949cc7365dfSYongqiang Yang if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { 49508d2ae1cbSJakub Wilk ext4_warning(sb, "too much blocks added to group %u", 4951cc7365dfSYongqiang Yang block_group); 4952cc7365dfSYongqiang Yang err = -EINVAL; 49532846e820SAmir Goldstein goto error_return; 4954cc7365dfSYongqiang Yang } 49552cd05cc3STheodore Ts'o 49562846e820SAmir Goldstein bitmap_bh = ext4_read_block_bitmap(sb, block_group); 49579008a58eSDarrick J. Wong if (IS_ERR(bitmap_bh)) { 49589008a58eSDarrick J. Wong err = PTR_ERR(bitmap_bh); 49599008a58eSDarrick J. Wong bitmap_bh = NULL; 49602846e820SAmir Goldstein goto error_return; 4961cc7365dfSYongqiang Yang } 4962cc7365dfSYongqiang Yang 49632846e820SAmir Goldstein desc = ext4_get_group_desc(sb, block_group, &gd_bh); 4964cc7365dfSYongqiang Yang if (!desc) { 4965cc7365dfSYongqiang Yang err = -EIO; 49662846e820SAmir Goldstein goto error_return; 4967cc7365dfSYongqiang Yang } 49682846e820SAmir Goldstein 49692846e820SAmir Goldstein if (in_range(ext4_block_bitmap(sb, desc), block, count) || 49702846e820SAmir Goldstein in_range(ext4_inode_bitmap(sb, desc), block, count) || 49712846e820SAmir Goldstein in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || 49722846e820SAmir Goldstein in_range(block + count - 1, ext4_inode_table(sb, desc), 49732846e820SAmir Goldstein sbi->s_itb_per_group)) { 49742846e820SAmir Goldstein ext4_error(sb, "Adding blocks in system zones - " 49752846e820SAmir Goldstein "Block = %llu, count = %lu", 49762846e820SAmir Goldstein block, count); 4977cc7365dfSYongqiang Yang err = -EINVAL; 49782846e820SAmir Goldstein goto error_return; 49792846e820SAmir Goldstein } 49802846e820SAmir Goldstein 49812cd05cc3STheodore Ts'o BUFFER_TRACE(bitmap_bh, "getting write access"); 49822cd05cc3STheodore Ts'o err = ext4_journal_get_write_access(handle, bitmap_bh); 49832846e820SAmir Goldstein if (err) 49842846e820SAmir Goldstein goto error_return; 49852846e820SAmir Goldstein 49862846e820SAmir Goldstein /* 49872846e820SAmir Goldstein * We are about to modify some metadata. Call the journal APIs 49882846e820SAmir Goldstein * to unshare ->b_data if a currently-committing transaction is 49892846e820SAmir Goldstein * using it 49902846e820SAmir Goldstein */ 49912846e820SAmir Goldstein BUFFER_TRACE(gd_bh, "get_write_access"); 49922846e820SAmir Goldstein err = ext4_journal_get_write_access(handle, gd_bh); 49932846e820SAmir Goldstein if (err) 49942846e820SAmir Goldstein goto error_return; 4995e73a347bSAmir Goldstein 49962846e820SAmir Goldstein for (i = 0, blocks_freed = 0; i < count; i++) { 49972846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "clear bit"); 4998e73a347bSAmir Goldstein if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { 49992846e820SAmir Goldstein ext4_error(sb, "bit already cleared for block %llu", 50002846e820SAmir Goldstein (ext4_fsblk_t)(block + i)); 50012846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "bit already cleared"); 50022846e820SAmir Goldstein } else { 50032846e820SAmir Goldstein blocks_freed++; 50042846e820SAmir Goldstein } 50052846e820SAmir Goldstein } 5006e73a347bSAmir Goldstein 5007e73a347bSAmir Goldstein err = ext4_mb_load_buddy(sb, block_group, &e4b); 5008e73a347bSAmir Goldstein if (err) 5009e73a347bSAmir Goldstein goto error_return; 5010e73a347bSAmir Goldstein 5011e73a347bSAmir Goldstein /* 5012e73a347bSAmir Goldstein * need to update group_info->bb_free and bitmap 5013e73a347bSAmir Goldstein * with group lock held. generate_buddy look at 5014e73a347bSAmir Goldstein * them with group lock_held 5015e73a347bSAmir Goldstein */ 50162846e820SAmir Goldstein ext4_lock_group(sb, block_group); 5017e73a347bSAmir Goldstein mb_clear_bits(bitmap_bh->b_data, bit, count); 5018e73a347bSAmir Goldstein mb_free_blocks(NULL, &e4b, bit, count); 5019021b65bbSTheodore Ts'o blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc); 5020021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, desc, blk_free_count); 502179f1ba49STao Ma ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh); 5022feb0ab32SDarrick J. Wong ext4_group_desc_csum_set(sb, block_group, desc); 50232846e820SAmir Goldstein ext4_unlock_group(sb, block_group); 502457042651STheodore Ts'o percpu_counter_add(&sbi->s_freeclusters_counter, 5025810da240SLukas Czerner EXT4_NUM_B2C(sbi, blocks_freed)); 50262846e820SAmir Goldstein 50272846e820SAmir Goldstein if (sbi->s_log_groups_per_flex) { 50282846e820SAmir Goldstein ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 502990ba983fSTheodore Ts'o atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed), 503024aaa8efSTheodore Ts'o &sbi->s_flex_groups[flex_group].free_clusters); 50312846e820SAmir Goldstein } 5032e73a347bSAmir Goldstein 5033e73a347bSAmir Goldstein ext4_mb_unload_buddy(&e4b); 50342846e820SAmir Goldstein 50352846e820SAmir Goldstein /* We dirtied the bitmap block */ 50362846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 50372846e820SAmir Goldstein err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 50382846e820SAmir Goldstein 50392846e820SAmir Goldstein /* And the group descriptor block */ 50402846e820SAmir Goldstein BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 50412846e820SAmir Goldstein ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 50422846e820SAmir Goldstein if (!err) 50432846e820SAmir Goldstein err = ret; 50442846e820SAmir Goldstein 50452846e820SAmir Goldstein error_return: 50462846e820SAmir Goldstein brelse(bitmap_bh); 50472846e820SAmir Goldstein ext4_std_error(sb, err); 5048cc7365dfSYongqiang Yang return err; 50492846e820SAmir Goldstein } 50502846e820SAmir Goldstein 50512846e820SAmir Goldstein /** 50527360d173SLukas Czerner * ext4_trim_extent -- function to TRIM one single free extent in the group 50537360d173SLukas Czerner * @sb: super block for the file system 50547360d173SLukas Czerner * @start: starting block of the free extent in the alloc. group 50557360d173SLukas Czerner * @count: number of blocks to TRIM 50567360d173SLukas Czerner * @group: alloc. group we are working with 50577360d173SLukas Czerner * @e4b: ext4 buddy for the group 50587360d173SLukas Czerner * 50597360d173SLukas Czerner * Trim "count" blocks starting at "start" in the "group". To assure that no 50607360d173SLukas Czerner * one will allocate those blocks, mark it as used in buddy bitmap. This must 50617360d173SLukas Czerner * be called with under the group lock. 50627360d173SLukas Czerner */ 5063d71c1ae2SLukas Czerner static int ext4_trim_extent(struct super_block *sb, int start, int count, 50647360d173SLukas Czerner ext4_group_t group, struct ext4_buddy *e4b) 5065e2cbd587Sjon ernst __releases(bitlock) 5066e2cbd587Sjon ernst __acquires(bitlock) 50677360d173SLukas Czerner { 50687360d173SLukas Czerner struct ext4_free_extent ex; 5069d71c1ae2SLukas Czerner int ret = 0; 50707360d173SLukas Czerner 5071b3d4c2b1STao Ma trace_ext4_trim_extent(sb, group, start, count); 5072b3d4c2b1STao Ma 50737360d173SLukas Czerner assert_spin_locked(ext4_group_lock_ptr(sb, group)); 50747360d173SLukas Czerner 50757360d173SLukas Czerner ex.fe_start = start; 50767360d173SLukas Czerner ex.fe_group = group; 50777360d173SLukas Czerner ex.fe_len = count; 50787360d173SLukas Czerner 50797360d173SLukas Czerner /* 50807360d173SLukas Czerner * Mark blocks used, so no one can reuse them while 50817360d173SLukas Czerner * being trimmed. 50827360d173SLukas Czerner */ 50837360d173SLukas Czerner mb_mark_used(e4b, &ex); 50847360d173SLukas Czerner ext4_unlock_group(sb, group); 5085d71c1ae2SLukas Czerner ret = ext4_issue_discard(sb, group, start, count); 50867360d173SLukas Czerner ext4_lock_group(sb, group); 50877360d173SLukas Czerner mb_free_blocks(NULL, e4b, start, ex.fe_len); 5088d71c1ae2SLukas Czerner return ret; 50897360d173SLukas Czerner } 50907360d173SLukas Czerner 50917360d173SLukas Czerner /** 50927360d173SLukas Czerner * ext4_trim_all_free -- function to trim all free space in alloc. group 50937360d173SLukas Czerner * @sb: super block for file system 509422612283STao Ma * @group: group to be trimmed 50957360d173SLukas Czerner * @start: first group block to examine 50967360d173SLukas Czerner * @max: last group block to examine 50977360d173SLukas Czerner * @minblocks: minimum extent block count 50987360d173SLukas Czerner * 50997360d173SLukas Czerner * ext4_trim_all_free walks through group's buddy bitmap searching for free 51007360d173SLukas Czerner * extents. When the free block is found, ext4_trim_extent is called to TRIM 51017360d173SLukas Czerner * the extent. 51027360d173SLukas Czerner * 51037360d173SLukas Czerner * 51047360d173SLukas Czerner * ext4_trim_all_free walks through group's block bitmap searching for free 51057360d173SLukas Czerner * extents. When the free extent is found, mark it as used in group buddy 51067360d173SLukas Czerner * bitmap. Then issue a TRIM command on this extent and free the extent in 51077360d173SLukas Czerner * the group buddy bitmap. This is done until whole group is scanned. 51087360d173SLukas Czerner */ 51090b75a840SLukas Czerner static ext4_grpblk_t 511078944086SLukas Czerner ext4_trim_all_free(struct super_block *sb, ext4_group_t group, 511178944086SLukas Czerner ext4_grpblk_t start, ext4_grpblk_t max, 511278944086SLukas Czerner ext4_grpblk_t minblocks) 51137360d173SLukas Czerner { 51147360d173SLukas Czerner void *bitmap; 5115169ddc3eSTao Ma ext4_grpblk_t next, count = 0, free_count = 0; 511678944086SLukas Czerner struct ext4_buddy e4b; 5117d71c1ae2SLukas Czerner int ret = 0; 51187360d173SLukas Czerner 5119b3d4c2b1STao Ma trace_ext4_trim_all_free(sb, group, start, max); 5120b3d4c2b1STao Ma 512178944086SLukas Czerner ret = ext4_mb_load_buddy(sb, group, &e4b); 512278944086SLukas Czerner if (ret) { 512378944086SLukas Czerner ext4_error(sb, "Error in loading buddy " 512478944086SLukas Czerner "information for %u", group); 512578944086SLukas Czerner return ret; 512678944086SLukas Czerner } 512778944086SLukas Czerner bitmap = e4b.bd_bitmap; 512828739eeaSLukas Czerner 512928739eeaSLukas Czerner ext4_lock_group(sb, group); 51303d56b8d2STao Ma if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && 51313d56b8d2STao Ma minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) 51323d56b8d2STao Ma goto out; 51333d56b8d2STao Ma 513478944086SLukas Czerner start = (e4b.bd_info->bb_first_free > start) ? 513578944086SLukas Czerner e4b.bd_info->bb_first_free : start; 51367360d173SLukas Czerner 5137913eed83SLukas Czerner while (start <= max) { 5138913eed83SLukas Czerner start = mb_find_next_zero_bit(bitmap, max + 1, start); 5139913eed83SLukas Czerner if (start > max) 51407360d173SLukas Czerner break; 5141913eed83SLukas Czerner next = mb_find_next_bit(bitmap, max + 1, start); 51427360d173SLukas Czerner 51437360d173SLukas Czerner if ((next - start) >= minblocks) { 5144d71c1ae2SLukas Czerner ret = ext4_trim_extent(sb, start, 514578944086SLukas Czerner next - start, group, &e4b); 5146d71c1ae2SLukas Czerner if (ret && ret != -EOPNOTSUPP) 5147d71c1ae2SLukas Czerner break; 5148d71c1ae2SLukas Czerner ret = 0; 51497360d173SLukas Czerner count += next - start; 51507360d173SLukas Czerner } 5151169ddc3eSTao Ma free_count += next - start; 51527360d173SLukas Czerner start = next + 1; 51537360d173SLukas Czerner 51547360d173SLukas Czerner if (fatal_signal_pending(current)) { 51557360d173SLukas Czerner count = -ERESTARTSYS; 51567360d173SLukas Czerner break; 51577360d173SLukas Czerner } 51587360d173SLukas Czerner 51597360d173SLukas Czerner if (need_resched()) { 51607360d173SLukas Czerner ext4_unlock_group(sb, group); 51617360d173SLukas Czerner cond_resched(); 51627360d173SLukas Czerner ext4_lock_group(sb, group); 51637360d173SLukas Czerner } 51647360d173SLukas Czerner 5165169ddc3eSTao Ma if ((e4b.bd_info->bb_free - free_count) < minblocks) 51667360d173SLukas Czerner break; 51677360d173SLukas Czerner } 51683d56b8d2STao Ma 5169d71c1ae2SLukas Czerner if (!ret) { 5170d71c1ae2SLukas Czerner ret = count; 51713d56b8d2STao Ma EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); 5172d71c1ae2SLukas Czerner } 51733d56b8d2STao Ma out: 51747360d173SLukas Czerner ext4_unlock_group(sb, group); 517578944086SLukas Czerner ext4_mb_unload_buddy(&e4b); 51767360d173SLukas Czerner 51777360d173SLukas Czerner ext4_debug("trimmed %d blocks in the group %d\n", 51787360d173SLukas Czerner count, group); 51797360d173SLukas Czerner 5180d71c1ae2SLukas Czerner return ret; 51817360d173SLukas Czerner } 51827360d173SLukas Czerner 51837360d173SLukas Czerner /** 51847360d173SLukas Czerner * ext4_trim_fs() -- trim ioctl handle function 51857360d173SLukas Czerner * @sb: superblock for filesystem 51867360d173SLukas Czerner * @range: fstrim_range structure 51877360d173SLukas Czerner * 51887360d173SLukas Czerner * start: First Byte to trim 51897360d173SLukas Czerner * len: number of Bytes to trim from start 51907360d173SLukas Czerner * minlen: minimum extent length in Bytes 51917360d173SLukas Czerner * ext4_trim_fs goes through all allocation groups containing Bytes from 51927360d173SLukas Czerner * start to start+len. For each such a group ext4_trim_all_free function 51937360d173SLukas Czerner * is invoked to trim all free space. 51947360d173SLukas Czerner */ 51957360d173SLukas Czerner int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) 51967360d173SLukas Czerner { 519778944086SLukas Czerner struct ext4_group_info *grp; 5198913eed83SLukas Czerner ext4_group_t group, first_group, last_group; 51997137d7a4STheodore Ts'o ext4_grpblk_t cnt = 0, first_cluster, last_cluster; 5200913eed83SLukas Czerner uint64_t start, end, minlen, trimmed = 0; 52010f0a25bfSJan Kara ext4_fsblk_t first_data_blk = 52020f0a25bfSJan Kara le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 5203913eed83SLukas Czerner ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); 52047360d173SLukas Czerner int ret = 0; 52057360d173SLukas Czerner 52067360d173SLukas Czerner start = range->start >> sb->s_blocksize_bits; 5207913eed83SLukas Czerner end = start + (range->len >> sb->s_blocksize_bits) - 1; 5208aaf7d73eSLukas Czerner minlen = EXT4_NUM_B2C(EXT4_SB(sb), 5209aaf7d73eSLukas Czerner range->minlen >> sb->s_blocksize_bits); 52107360d173SLukas Czerner 52115de35e8dSLukas Czerner if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) || 52125de35e8dSLukas Czerner start >= max_blks || 52135de35e8dSLukas Czerner range->len < sb->s_blocksize) 52147360d173SLukas Czerner return -EINVAL; 5215913eed83SLukas Czerner if (end >= max_blks) 5216913eed83SLukas Czerner end = max_blks - 1; 5217913eed83SLukas Czerner if (end <= first_data_blk) 521822f10457STao Ma goto out; 5219913eed83SLukas Czerner if (start < first_data_blk) 52200f0a25bfSJan Kara start = first_data_blk; 52217360d173SLukas Czerner 5222913eed83SLukas Czerner /* Determine first and last group to examine based on start and end */ 52237360d173SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, 52247137d7a4STheodore Ts'o &first_group, &first_cluster); 5225913eed83SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, 52267137d7a4STheodore Ts'o &last_group, &last_cluster); 52277360d173SLukas Czerner 5228913eed83SLukas Czerner /* end now represents the last cluster to discard in this group */ 5229913eed83SLukas Czerner end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 52307360d173SLukas Czerner 52317360d173SLukas Czerner for (group = first_group; group <= last_group; group++) { 523278944086SLukas Czerner grp = ext4_get_group_info(sb, group); 523378944086SLukas Czerner /* We only do this if the grp has never been initialized */ 523478944086SLukas Czerner if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 5235adb7ef60SKonstantin Khlebnikov ret = ext4_mb_init_group(sb, group, GFP_NOFS); 523678944086SLukas Czerner if (ret) 52377360d173SLukas Czerner break; 52387360d173SLukas Czerner } 52397360d173SLukas Czerner 52400ba08517STao Ma /* 5241913eed83SLukas Czerner * For all the groups except the last one, last cluster will 5242913eed83SLukas Czerner * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to 5243913eed83SLukas Czerner * change it for the last group, note that last_cluster is 5244913eed83SLukas Czerner * already computed earlier by ext4_get_group_no_and_offset() 52450ba08517STao Ma */ 5246913eed83SLukas Czerner if (group == last_group) 5247913eed83SLukas Czerner end = last_cluster; 52487360d173SLukas Czerner 524978944086SLukas Czerner if (grp->bb_free >= minlen) { 52507137d7a4STheodore Ts'o cnt = ext4_trim_all_free(sb, group, first_cluster, 5251913eed83SLukas Czerner end, minlen); 52527360d173SLukas Czerner if (cnt < 0) { 52537360d173SLukas Czerner ret = cnt; 52547360d173SLukas Czerner break; 52557360d173SLukas Czerner } 52567360d173SLukas Czerner trimmed += cnt; 525721e7fd22SLukas Czerner } 5258913eed83SLukas Czerner 5259913eed83SLukas Czerner /* 5260913eed83SLukas Czerner * For every group except the first one, we are sure 5261913eed83SLukas Czerner * that the first cluster to discard will be cluster #0. 5262913eed83SLukas Czerner */ 52637137d7a4STheodore Ts'o first_cluster = 0; 52647360d173SLukas Czerner } 52657360d173SLukas Czerner 52663d56b8d2STao Ma if (!ret) 52673d56b8d2STao Ma atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); 52683d56b8d2STao Ma 526922f10457STao Ma out: 5270aaf7d73eSLukas Czerner range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits; 52717360d173SLukas Czerner return ret; 52727360d173SLukas Czerner } 5273