1c9de560dSAlex Tomas /* 2c9de560dSAlex Tomas * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com 3c9de560dSAlex Tomas * Written by Alex Tomas <alex@clusterfs.com> 4c9de560dSAlex Tomas * 5c9de560dSAlex Tomas * This program is free software; you can redistribute it and/or modify 6c9de560dSAlex Tomas * it under the terms of the GNU General Public License version 2 as 7c9de560dSAlex Tomas * published by the Free Software Foundation. 8c9de560dSAlex Tomas * 9c9de560dSAlex Tomas * This program is distributed in the hope that it will be useful, 10c9de560dSAlex Tomas * but WITHOUT ANY WARRANTY; without even the implied warranty of 11c9de560dSAlex Tomas * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12c9de560dSAlex Tomas * GNU General Public License for more details. 13c9de560dSAlex Tomas * 14c9de560dSAlex Tomas * You should have received a copy of the GNU General Public Licens 15c9de560dSAlex Tomas * along with this program; if not, write to the Free Software 16c9de560dSAlex Tomas * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- 17c9de560dSAlex Tomas */ 18c9de560dSAlex Tomas 19c9de560dSAlex Tomas 20c9de560dSAlex Tomas /* 21c9de560dSAlex Tomas * mballoc.c contains the multiblocks allocation routines 22c9de560dSAlex Tomas */ 23c9de560dSAlex Tomas 2418aadd47SBobi Jam #include "ext4_jbd2.h" 258f6e39a7SMingming Cao #include "mballoc.h" 266ba495e9STheodore Ts'o #include <linux/debugfs.h> 275a0e3ad6STejun Heo #include <linux/slab.h> 289bffad1eSTheodore Ts'o #include <trace/events/ext4.h> 299bffad1eSTheodore Ts'o 30c9de560dSAlex Tomas /* 31c9de560dSAlex Tomas * MUSTDO: 32c9de560dSAlex Tomas * - test ext4_ext_search_left() and ext4_ext_search_right() 33c9de560dSAlex Tomas * - search for metadata in few groups 34c9de560dSAlex Tomas * 35c9de560dSAlex Tomas * TODO v4: 36c9de560dSAlex Tomas * - normalization should take into account whether file is still open 37c9de560dSAlex Tomas * - discard preallocations if no free space left (policy?) 38c9de560dSAlex Tomas * - don't normalize tails 39c9de560dSAlex Tomas * - quota 40c9de560dSAlex Tomas * - reservation for superuser 41c9de560dSAlex Tomas * 42c9de560dSAlex Tomas * TODO v3: 43c9de560dSAlex Tomas * - bitmap read-ahead (proposed by Oleg Drokin aka green) 44c9de560dSAlex Tomas * - track min/max extents in each group for better group selection 45c9de560dSAlex Tomas * - mb_mark_used() may allocate chunk right after splitting buddy 46c9de560dSAlex Tomas * - tree of groups sorted by number of free blocks 47c9de560dSAlex Tomas * - error handling 48c9de560dSAlex Tomas */ 49c9de560dSAlex Tomas 50c9de560dSAlex Tomas /* 51c9de560dSAlex Tomas * The allocation request involve request for multiple number of blocks 52c9de560dSAlex Tomas * near to the goal(block) value specified. 53c9de560dSAlex Tomas * 54b713a5ecSTheodore Ts'o * During initialization phase of the allocator we decide to use the 55b713a5ecSTheodore Ts'o * group preallocation or inode preallocation depending on the size of 56b713a5ecSTheodore Ts'o * the file. The size of the file could be the resulting file size we 57b713a5ecSTheodore Ts'o * would have after allocation, or the current file size, which ever 58b713a5ecSTheodore Ts'o * is larger. If the size is less than sbi->s_mb_stream_request we 59b713a5ecSTheodore Ts'o * select to use the group preallocation. The default value of 60b713a5ecSTheodore Ts'o * s_mb_stream_request is 16 blocks. This can also be tuned via 61b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 62b713a5ecSTheodore Ts'o * terms of number of blocks. 63c9de560dSAlex Tomas * 64c9de560dSAlex Tomas * The main motivation for having small file use group preallocation is to 65b713a5ecSTheodore Ts'o * ensure that we have small files closer together on the disk. 66c9de560dSAlex Tomas * 67b713a5ecSTheodore Ts'o * First stage the allocator looks at the inode prealloc list, 68b713a5ecSTheodore Ts'o * ext4_inode_info->i_prealloc_list, which contains list of prealloc 69b713a5ecSTheodore Ts'o * spaces for this particular inode. The inode prealloc space is 70b713a5ecSTheodore Ts'o * represented as: 71c9de560dSAlex Tomas * 72c9de560dSAlex Tomas * pa_lstart -> the logical start block for this prealloc space 73c9de560dSAlex Tomas * pa_pstart -> the physical start block for this prealloc space 7453accfa9STheodore Ts'o * pa_len -> length for this prealloc space (in clusters) 7553accfa9STheodore Ts'o * pa_free -> free space available in this prealloc space (in clusters) 76c9de560dSAlex Tomas * 77c9de560dSAlex Tomas * The inode preallocation space is used looking at the _logical_ start 78c9de560dSAlex Tomas * block. If only the logical file block falls within the range of prealloc 79caaf7a29STao Ma * space we will consume the particular prealloc space. This makes sure that 80caaf7a29STao Ma * we have contiguous physical blocks representing the file blocks 81c9de560dSAlex Tomas * 82c9de560dSAlex Tomas * The important thing to be noted in case of inode prealloc space is that 83c9de560dSAlex Tomas * we don't modify the values associated to inode prealloc space except 84c9de560dSAlex Tomas * pa_free. 85c9de560dSAlex Tomas * 86c9de560dSAlex Tomas * If we are not able to find blocks in the inode prealloc space and if we 87c9de560dSAlex Tomas * have the group allocation flag set then we look at the locality group 88caaf7a29STao Ma * prealloc space. These are per CPU prealloc list represented as 89c9de560dSAlex Tomas * 90c9de560dSAlex Tomas * ext4_sb_info.s_locality_groups[smp_processor_id()] 91c9de560dSAlex Tomas * 92c9de560dSAlex Tomas * The reason for having a per cpu locality group is to reduce the contention 93c9de560dSAlex Tomas * between CPUs. It is possible to get scheduled at this point. 94c9de560dSAlex Tomas * 95c9de560dSAlex Tomas * The locality group prealloc space is used looking at whether we have 9625985edcSLucas De Marchi * enough free space (pa_free) within the prealloc space. 97c9de560dSAlex Tomas * 98c9de560dSAlex Tomas * If we can't allocate blocks via inode prealloc or/and locality group 99c9de560dSAlex Tomas * prealloc then we look at the buddy cache. The buddy cache is represented 100c9de560dSAlex Tomas * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 101c9de560dSAlex Tomas * mapped to the buddy and bitmap information regarding different 102c9de560dSAlex Tomas * groups. The buddy information is attached to buddy cache inode so that 103c9de560dSAlex Tomas * we can access them through the page cache. The information regarding 104c9de560dSAlex Tomas * each group is loaded via ext4_mb_load_buddy. The information involve 105c9de560dSAlex Tomas * block bitmap and buddy information. The information are stored in the 106c9de560dSAlex Tomas * inode as: 107c9de560dSAlex Tomas * 108c9de560dSAlex Tomas * { page } 109c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 110c9de560dSAlex Tomas * 111c9de560dSAlex Tomas * 112c9de560dSAlex Tomas * one block each for bitmap and buddy information. So for each group we 113c9de560dSAlex Tomas * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / 114c9de560dSAlex Tomas * blocksize) blocks. So it can have information regarding groups_per_page 115c9de560dSAlex Tomas * which is blocks_per_page/2 116c9de560dSAlex Tomas * 117c9de560dSAlex Tomas * The buddy cache inode is not stored on disk. The inode is thrown 118c9de560dSAlex Tomas * away when the filesystem is unmounted. 119c9de560dSAlex Tomas * 120c9de560dSAlex Tomas * We look for count number of blocks in the buddy cache. If we were able 121c9de560dSAlex Tomas * to locate that many free blocks we return with additional information 122c9de560dSAlex Tomas * regarding rest of the contiguous physical block available 123c9de560dSAlex Tomas * 124c9de560dSAlex Tomas * Before allocating blocks via buddy cache we normalize the request 125c9de560dSAlex Tomas * blocks. This ensure we ask for more blocks that we needed. The extra 126c9de560dSAlex Tomas * blocks that we get after allocation is added to the respective prealloc 127c9de560dSAlex Tomas * list. In case of inode preallocation we follow a list of heuristics 128c9de560dSAlex Tomas * based on file size. This can be found in ext4_mb_normalize_request. If 129c9de560dSAlex Tomas * we are doing a group prealloc we try to normalize the request to 13027baebb8STheodore Ts'o * sbi->s_mb_group_prealloc. The default value of s_mb_group_prealloc is 13127baebb8STheodore Ts'o * dependent on the cluster size; for non-bigalloc file systems, it is 132c9de560dSAlex Tomas * 512 blocks. This can be tuned via 133d7a1fee1SDan Ehrenberg * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in 134c9de560dSAlex Tomas * terms of number of blocks. If we have mounted the file system with -O 135c9de560dSAlex Tomas * stripe=<value> option the group prealloc request is normalized to the 136d7a1fee1SDan Ehrenberg * the smallest multiple of the stripe value (sbi->s_stripe) which is 137d7a1fee1SDan Ehrenberg * greater than the default mb_group_prealloc. 138c9de560dSAlex Tomas * 139d7a1fee1SDan Ehrenberg * The regular allocator (using the buddy cache) supports a few tunables. 140c9de560dSAlex Tomas * 141b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_min_to_scan 142b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_max_to_scan 143b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req 144c9de560dSAlex Tomas * 145b713a5ecSTheodore Ts'o * The regular allocator uses buddy scan only if the request len is power of 146c9de560dSAlex Tomas * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 147c9de560dSAlex Tomas * value of s_mb_order2_reqs can be tuned via 148b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 149af901ca1SAndré Goddard Rosa * stripe size (sbi->s_stripe), we try to search for contiguous block in 150b713a5ecSTheodore Ts'o * stripe size. This should result in better allocation on RAID setups. If 151b713a5ecSTheodore Ts'o * not, we search in the specific group using bitmap for best extents. The 152b713a5ecSTheodore Ts'o * tunable min_to_scan and max_to_scan control the behaviour here. 153c9de560dSAlex Tomas * min_to_scan indicate how long the mballoc __must__ look for a best 154b713a5ecSTheodore Ts'o * extent and max_to_scan indicates how long the mballoc __can__ look for a 155c9de560dSAlex Tomas * best extent in the found extents. Searching for the blocks starts with 156c9de560dSAlex Tomas * the group specified as the goal value in allocation context via 157c9de560dSAlex Tomas * ac_g_ex. Each group is first checked based on the criteria whether it 158caaf7a29STao Ma * can be used for allocation. ext4_mb_good_group explains how the groups are 159c9de560dSAlex Tomas * checked. 160c9de560dSAlex Tomas * 161c9de560dSAlex Tomas * Both the prealloc space are getting populated as above. So for the first 162c9de560dSAlex Tomas * request we will hit the buddy cache which will result in this prealloc 163c9de560dSAlex Tomas * space getting filled. The prealloc space is then later used for the 164c9de560dSAlex Tomas * subsequent request. 165c9de560dSAlex Tomas */ 166c9de560dSAlex Tomas 167c9de560dSAlex Tomas /* 168c9de560dSAlex Tomas * mballoc operates on the following data: 169c9de560dSAlex Tomas * - on-disk bitmap 170c9de560dSAlex Tomas * - in-core buddy (actually includes buddy and bitmap) 171c9de560dSAlex Tomas * - preallocation descriptors (PAs) 172c9de560dSAlex Tomas * 173c9de560dSAlex Tomas * there are two types of preallocations: 174c9de560dSAlex Tomas * - inode 175c9de560dSAlex Tomas * assiged to specific inode and can be used for this inode only. 176c9de560dSAlex Tomas * it describes part of inode's space preallocated to specific 177c9de560dSAlex Tomas * physical blocks. any block from that preallocated can be used 178c9de560dSAlex Tomas * independent. the descriptor just tracks number of blocks left 179c9de560dSAlex Tomas * unused. so, before taking some block from descriptor, one must 180c9de560dSAlex Tomas * make sure corresponded logical block isn't allocated yet. this 181c9de560dSAlex Tomas * also means that freeing any block within descriptor's range 182c9de560dSAlex Tomas * must discard all preallocated blocks. 183c9de560dSAlex Tomas * - locality group 184c9de560dSAlex Tomas * assigned to specific locality group which does not translate to 185c9de560dSAlex Tomas * permanent set of inodes: inode can join and leave group. space 186c9de560dSAlex Tomas * from this type of preallocation can be used for any inode. thus 187c9de560dSAlex Tomas * it's consumed from the beginning to the end. 188c9de560dSAlex Tomas * 189c9de560dSAlex Tomas * relation between them can be expressed as: 190c9de560dSAlex Tomas * in-core buddy = on-disk bitmap + preallocation descriptors 191c9de560dSAlex Tomas * 192c9de560dSAlex Tomas * this mean blocks mballoc considers used are: 193c9de560dSAlex Tomas * - allocated blocks (persistent) 194c9de560dSAlex Tomas * - preallocated blocks (non-persistent) 195c9de560dSAlex Tomas * 196c9de560dSAlex Tomas * consistency in mballoc world means that at any time a block is either 197c9de560dSAlex Tomas * free or used in ALL structures. notice: "any time" should not be read 198c9de560dSAlex Tomas * literally -- time is discrete and delimited by locks. 199c9de560dSAlex Tomas * 200c9de560dSAlex Tomas * to keep it simple, we don't use block numbers, instead we count number of 201c9de560dSAlex Tomas * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 202c9de560dSAlex Tomas * 203c9de560dSAlex Tomas * all operations can be expressed as: 204c9de560dSAlex Tomas * - init buddy: buddy = on-disk + PAs 205c9de560dSAlex Tomas * - new PA: buddy += N; PA = N 206c9de560dSAlex Tomas * - use inode PA: on-disk += N; PA -= N 207c9de560dSAlex Tomas * - discard inode PA buddy -= on-disk - PA; PA = 0 208c9de560dSAlex Tomas * - use locality group PA on-disk += N; PA -= N 209c9de560dSAlex Tomas * - discard locality group PA buddy -= PA; PA = 0 210c9de560dSAlex Tomas * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 211c9de560dSAlex Tomas * is used in real operation because we can't know actual used 212c9de560dSAlex Tomas * bits from PA, only from on-disk bitmap 213c9de560dSAlex Tomas * 214c9de560dSAlex Tomas * if we follow this strict logic, then all operations above should be atomic. 215c9de560dSAlex Tomas * given some of them can block, we'd have to use something like semaphores 216c9de560dSAlex Tomas * killing performance on high-end SMP hardware. let's try to relax it using 217c9de560dSAlex Tomas * the following knowledge: 218c9de560dSAlex Tomas * 1) if buddy is referenced, it's already initialized 219c9de560dSAlex Tomas * 2) while block is used in buddy and the buddy is referenced, 220c9de560dSAlex Tomas * nobody can re-allocate that block 221c9de560dSAlex Tomas * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 222c9de560dSAlex Tomas * bit set and PA claims same block, it's OK. IOW, one can set bit in 223c9de560dSAlex Tomas * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 224c9de560dSAlex Tomas * block 225c9de560dSAlex Tomas * 226c9de560dSAlex Tomas * so, now we're building a concurrency table: 227c9de560dSAlex Tomas * - init buddy vs. 228c9de560dSAlex Tomas * - new PA 229c9de560dSAlex Tomas * blocks for PA are allocated in the buddy, buddy must be referenced 230c9de560dSAlex Tomas * until PA is linked to allocation group to avoid concurrent buddy init 231c9de560dSAlex Tomas * - use inode PA 232c9de560dSAlex Tomas * we need to make sure that either on-disk bitmap or PA has uptodate data 233c9de560dSAlex Tomas * given (3) we care that PA-=N operation doesn't interfere with init 234c9de560dSAlex Tomas * - discard inode PA 235c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 236c9de560dSAlex Tomas * - use locality group PA 237c9de560dSAlex Tomas * again PA-=N must be serialized with init 238c9de560dSAlex Tomas * - discard locality group PA 239c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 240c9de560dSAlex Tomas * - new PA vs. 241c9de560dSAlex Tomas * - use inode PA 242c9de560dSAlex Tomas * i_data_sem serializes them 243c9de560dSAlex Tomas * - discard inode PA 244c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 245c9de560dSAlex Tomas * - use locality group PA 246c9de560dSAlex Tomas * some mutex should serialize them 247c9de560dSAlex Tomas * - discard locality group PA 248c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 249c9de560dSAlex Tomas * - use inode PA 250c9de560dSAlex Tomas * - use inode PA 251c9de560dSAlex Tomas * i_data_sem or another mutex should 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 * nothing wrong here -- they're different PAs covering different blocks 256c9de560dSAlex Tomas * - discard locality group PA 257c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 258c9de560dSAlex Tomas * 259c9de560dSAlex Tomas * now we're ready to make few consequences: 260c9de560dSAlex Tomas * - PA is referenced and while it is no discard is possible 261c9de560dSAlex Tomas * - PA is referenced until block isn't marked in on-disk bitmap 262c9de560dSAlex Tomas * - PA changes only after on-disk bitmap 263c9de560dSAlex Tomas * - discard must not compete with init. either init is done before 264c9de560dSAlex Tomas * any discard or they're serialized somehow 265c9de560dSAlex Tomas * - buddy init as sum of on-disk bitmap and PAs is done atomically 266c9de560dSAlex Tomas * 267c9de560dSAlex Tomas * a special case when we've used PA to emptiness. no need to modify buddy 268c9de560dSAlex Tomas * in this case, but we should care about concurrent init 269c9de560dSAlex Tomas * 270c9de560dSAlex Tomas */ 271c9de560dSAlex Tomas 272c9de560dSAlex Tomas /* 273c9de560dSAlex Tomas * Logic in few words: 274c9de560dSAlex Tomas * 275c9de560dSAlex Tomas * - allocation: 276c9de560dSAlex Tomas * load group 277c9de560dSAlex Tomas * find blocks 278c9de560dSAlex Tomas * mark bits in on-disk bitmap 279c9de560dSAlex Tomas * release group 280c9de560dSAlex Tomas * 281c9de560dSAlex Tomas * - use preallocation: 282c9de560dSAlex Tomas * find proper PA (per-inode or group) 283c9de560dSAlex Tomas * load group 284c9de560dSAlex Tomas * mark bits in on-disk bitmap 285c9de560dSAlex Tomas * release group 286c9de560dSAlex Tomas * release PA 287c9de560dSAlex Tomas * 288c9de560dSAlex Tomas * - free: 289c9de560dSAlex Tomas * load group 290c9de560dSAlex Tomas * mark bits in on-disk bitmap 291c9de560dSAlex Tomas * release group 292c9de560dSAlex Tomas * 293c9de560dSAlex Tomas * - discard preallocations in group: 294c9de560dSAlex Tomas * mark PAs deleted 295c9de560dSAlex Tomas * move them onto local list 296c9de560dSAlex Tomas * load on-disk bitmap 297c9de560dSAlex Tomas * load group 298c9de560dSAlex Tomas * remove PA from object (inode or locality group) 299c9de560dSAlex Tomas * mark free blocks in-core 300c9de560dSAlex Tomas * 301c9de560dSAlex Tomas * - discard inode's preallocations: 302c9de560dSAlex Tomas */ 303c9de560dSAlex Tomas 304c9de560dSAlex Tomas /* 305c9de560dSAlex Tomas * Locking rules 306c9de560dSAlex Tomas * 307c9de560dSAlex Tomas * Locks: 308c9de560dSAlex Tomas * - bitlock on a group (group) 309c9de560dSAlex Tomas * - object (inode/locality) (object) 310c9de560dSAlex Tomas * - per-pa lock (pa) 311c9de560dSAlex Tomas * 312c9de560dSAlex Tomas * Paths: 313c9de560dSAlex Tomas * - new pa 314c9de560dSAlex Tomas * object 315c9de560dSAlex Tomas * group 316c9de560dSAlex Tomas * 317c9de560dSAlex Tomas * - find and use pa: 318c9de560dSAlex Tomas * pa 319c9de560dSAlex Tomas * 320c9de560dSAlex Tomas * - release consumed pa: 321c9de560dSAlex Tomas * pa 322c9de560dSAlex Tomas * group 323c9de560dSAlex Tomas * object 324c9de560dSAlex Tomas * 325c9de560dSAlex Tomas * - generate in-core bitmap: 326c9de560dSAlex Tomas * group 327c9de560dSAlex Tomas * pa 328c9de560dSAlex Tomas * 329c9de560dSAlex Tomas * - discard all for given object (inode, locality group): 330c9de560dSAlex Tomas * object 331c9de560dSAlex Tomas * pa 332c9de560dSAlex Tomas * group 333c9de560dSAlex Tomas * 334c9de560dSAlex Tomas * - discard all for given group: 335c9de560dSAlex Tomas * group 336c9de560dSAlex Tomas * pa 337c9de560dSAlex Tomas * group 338c9de560dSAlex Tomas * object 339c9de560dSAlex Tomas * 340c9de560dSAlex Tomas */ 341c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_pspace_cachep; 342c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_ac_cachep; 34318aadd47SBobi Jam static struct kmem_cache *ext4_free_data_cachep; 344fb1813f4SCurt Wohlgemuth 345fb1813f4SCurt Wohlgemuth /* We create slab caches for groupinfo data structures based on the 346fb1813f4SCurt Wohlgemuth * superblock block size. There will be one per mounted filesystem for 347fb1813f4SCurt Wohlgemuth * each unique s_blocksize_bits */ 3482892c15dSEric Sandeen #define NR_GRPINFO_CACHES 8 349fb1813f4SCurt Wohlgemuth static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; 350fb1813f4SCurt Wohlgemuth 3512892c15dSEric Sandeen static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { 3522892c15dSEric Sandeen "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", 3532892c15dSEric Sandeen "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", 3542892c15dSEric Sandeen "ext4_groupinfo_64k", "ext4_groupinfo_128k" 3552892c15dSEric Sandeen }; 3562892c15dSEric Sandeen 357c3a326a6SAneesh Kumar K.V static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 358c3a326a6SAneesh Kumar K.V ext4_group_t group); 3597a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 3607a2fcbf7SAneesh Kumar K.V ext4_group_t group); 36118aadd47SBobi Jam static void ext4_free_data_callback(struct super_block *sb, 36218aadd47SBobi Jam struct ext4_journal_cb_entry *jce, int rc); 363c3a326a6SAneesh Kumar K.V 364ffad0a44SAneesh Kumar K.V static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 365ffad0a44SAneesh Kumar K.V { 366c9de560dSAlex Tomas #if BITS_PER_LONG == 64 367ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 7UL) << 3; 368ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~7UL); 369c9de560dSAlex Tomas #elif BITS_PER_LONG == 32 370ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 3UL) << 3; 371ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~3UL); 372c9de560dSAlex Tomas #else 373c9de560dSAlex Tomas #error "how many bits you are?!" 374c9de560dSAlex Tomas #endif 375ffad0a44SAneesh Kumar K.V return addr; 376ffad0a44SAneesh Kumar K.V } 377c9de560dSAlex Tomas 378c9de560dSAlex Tomas static inline int mb_test_bit(int bit, void *addr) 379c9de560dSAlex Tomas { 380c9de560dSAlex Tomas /* 381c9de560dSAlex Tomas * ext4_test_bit on architecture like powerpc 382c9de560dSAlex Tomas * needs unsigned long aligned address 383c9de560dSAlex Tomas */ 384ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 385c9de560dSAlex Tomas return ext4_test_bit(bit, addr); 386c9de560dSAlex Tomas } 387c9de560dSAlex Tomas 388c9de560dSAlex Tomas static inline void mb_set_bit(int bit, void *addr) 389c9de560dSAlex Tomas { 390ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 391c9de560dSAlex Tomas ext4_set_bit(bit, addr); 392c9de560dSAlex Tomas } 393c9de560dSAlex Tomas 394c9de560dSAlex Tomas static inline void mb_clear_bit(int bit, void *addr) 395c9de560dSAlex Tomas { 396ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 397c9de560dSAlex Tomas ext4_clear_bit(bit, addr); 398c9de560dSAlex Tomas } 399c9de560dSAlex Tomas 400ffad0a44SAneesh Kumar K.V static inline int mb_find_next_zero_bit(void *addr, int max, int start) 401ffad0a44SAneesh Kumar K.V { 402e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 403ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 404e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 405ffad0a44SAneesh Kumar K.V start += fix; 406ffad0a44SAneesh Kumar K.V 407e7dfb246SAneesh Kumar K.V ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 408e7dfb246SAneesh Kumar K.V if (ret > max) 409e7dfb246SAneesh Kumar K.V return max; 410e7dfb246SAneesh Kumar K.V return ret; 411ffad0a44SAneesh Kumar K.V } 412ffad0a44SAneesh Kumar K.V 413ffad0a44SAneesh Kumar K.V static inline int mb_find_next_bit(void *addr, int max, int start) 414ffad0a44SAneesh Kumar K.V { 415e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 416ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 417e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 418ffad0a44SAneesh Kumar K.V start += fix; 419ffad0a44SAneesh Kumar K.V 420e7dfb246SAneesh Kumar K.V ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 421e7dfb246SAneesh Kumar K.V if (ret > max) 422e7dfb246SAneesh Kumar K.V return max; 423e7dfb246SAneesh Kumar K.V return ret; 424ffad0a44SAneesh Kumar K.V } 425ffad0a44SAneesh Kumar K.V 426c9de560dSAlex Tomas static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 427c9de560dSAlex Tomas { 428c9de560dSAlex Tomas char *bb; 429c9de560dSAlex Tomas 430c5e8f3f3STheodore Ts'o BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 431c9de560dSAlex Tomas BUG_ON(max == NULL); 432c9de560dSAlex Tomas 433c9de560dSAlex Tomas if (order > e4b->bd_blkbits + 1) { 434c9de560dSAlex Tomas *max = 0; 435c9de560dSAlex Tomas return NULL; 436c9de560dSAlex Tomas } 437c9de560dSAlex Tomas 438c9de560dSAlex Tomas /* at order 0 we see each particular block */ 43984b775a3SColy Li if (order == 0) { 440c9de560dSAlex Tomas *max = 1 << (e4b->bd_blkbits + 3); 441c5e8f3f3STheodore Ts'o return e4b->bd_bitmap; 44284b775a3SColy Li } 443c9de560dSAlex Tomas 444c5e8f3f3STheodore Ts'o bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 445c9de560dSAlex Tomas *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 446c9de560dSAlex Tomas 447c9de560dSAlex Tomas return bb; 448c9de560dSAlex Tomas } 449c9de560dSAlex Tomas 450c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 451c9de560dSAlex Tomas static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 452c9de560dSAlex Tomas int first, int count) 453c9de560dSAlex Tomas { 454c9de560dSAlex Tomas int i; 455c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 456c9de560dSAlex Tomas 457c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 458c9de560dSAlex Tomas return; 459bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 460c9de560dSAlex Tomas for (i = 0; i < count; i++) { 461c9de560dSAlex Tomas if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 462c9de560dSAlex Tomas ext4_fsblk_t blocknr; 4635661bd68SAkinobu Mita 4645661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 46553accfa9STheodore Ts'o blocknr += EXT4_C2B(EXT4_SB(sb), first + i); 4665d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 467e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 468e29136f8STheodore Ts'o blocknr, 469e29136f8STheodore Ts'o "freeing block already freed " 470e29136f8STheodore Ts'o "(bit %u)", 471e29136f8STheodore Ts'o first + i); 472c9de560dSAlex Tomas } 473c9de560dSAlex Tomas mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 474c9de560dSAlex Tomas } 475c9de560dSAlex Tomas } 476c9de560dSAlex Tomas 477c9de560dSAlex Tomas static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 478c9de560dSAlex Tomas { 479c9de560dSAlex Tomas int i; 480c9de560dSAlex Tomas 481c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 482c9de560dSAlex Tomas return; 483bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 484c9de560dSAlex Tomas for (i = 0; i < count; i++) { 485c9de560dSAlex Tomas BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 486c9de560dSAlex Tomas mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 487c9de560dSAlex Tomas } 488c9de560dSAlex Tomas } 489c9de560dSAlex Tomas 490c9de560dSAlex Tomas static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 491c9de560dSAlex Tomas { 492c9de560dSAlex Tomas if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 493c9de560dSAlex Tomas unsigned char *b1, *b2; 494c9de560dSAlex Tomas int i; 495c9de560dSAlex Tomas b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 496c9de560dSAlex Tomas b2 = (unsigned char *) bitmap; 497c9de560dSAlex Tomas for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 498c9de560dSAlex Tomas if (b1[i] != b2[i]) { 4999d8b9ec4STheodore Ts'o ext4_msg(e4b->bd_sb, KERN_ERR, 5009d8b9ec4STheodore Ts'o "corruption in group %u " 5014776004fSTheodore Ts'o "at byte %u(%u): %x in copy != %x " 5029d8b9ec4STheodore Ts'o "on disk/prealloc", 503c9de560dSAlex Tomas e4b->bd_group, i, i * 8, b1[i], b2[i]); 504c9de560dSAlex Tomas BUG(); 505c9de560dSAlex Tomas } 506c9de560dSAlex Tomas } 507c9de560dSAlex Tomas } 508c9de560dSAlex Tomas } 509c9de560dSAlex Tomas 510c9de560dSAlex Tomas #else 511c9de560dSAlex Tomas static inline void mb_free_blocks_double(struct inode *inode, 512c9de560dSAlex Tomas struct ext4_buddy *e4b, int first, int count) 513c9de560dSAlex Tomas { 514c9de560dSAlex Tomas return; 515c9de560dSAlex Tomas } 516c9de560dSAlex Tomas static inline void mb_mark_used_double(struct ext4_buddy *e4b, 517c9de560dSAlex Tomas int first, int count) 518c9de560dSAlex Tomas { 519c9de560dSAlex Tomas return; 520c9de560dSAlex Tomas } 521c9de560dSAlex Tomas static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 522c9de560dSAlex Tomas { 523c9de560dSAlex Tomas return; 524c9de560dSAlex Tomas } 525c9de560dSAlex Tomas #endif 526c9de560dSAlex Tomas 527c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 528c9de560dSAlex Tomas 529c9de560dSAlex Tomas #define MB_CHECK_ASSERT(assert) \ 530c9de560dSAlex Tomas do { \ 531c9de560dSAlex Tomas if (!(assert)) { \ 532c9de560dSAlex Tomas printk(KERN_EMERG \ 533c9de560dSAlex Tomas "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 534c9de560dSAlex Tomas function, file, line, # assert); \ 535c9de560dSAlex Tomas BUG(); \ 536c9de560dSAlex Tomas } \ 537c9de560dSAlex Tomas } while (0) 538c9de560dSAlex Tomas 539c9de560dSAlex Tomas static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 540c9de560dSAlex Tomas const char *function, int line) 541c9de560dSAlex Tomas { 542c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 543c9de560dSAlex Tomas int order = e4b->bd_blkbits + 1; 544c9de560dSAlex Tomas int max; 545c9de560dSAlex Tomas int max2; 546c9de560dSAlex Tomas int i; 547c9de560dSAlex Tomas int j; 548c9de560dSAlex Tomas int k; 549c9de560dSAlex Tomas int count; 550c9de560dSAlex Tomas struct ext4_group_info *grp; 551c9de560dSAlex Tomas int fragments = 0; 552c9de560dSAlex Tomas int fstart; 553c9de560dSAlex Tomas struct list_head *cur; 554c9de560dSAlex Tomas void *buddy; 555c9de560dSAlex Tomas void *buddy2; 556c9de560dSAlex Tomas 557c9de560dSAlex Tomas { 558c9de560dSAlex Tomas static int mb_check_counter; 559c9de560dSAlex Tomas if (mb_check_counter++ % 100 != 0) 560c9de560dSAlex Tomas return 0; 561c9de560dSAlex Tomas } 562c9de560dSAlex Tomas 563c9de560dSAlex Tomas while (order > 1) { 564c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 565c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy); 566c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order - 1, &max2); 567c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy2); 568c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy != buddy2); 569c9de560dSAlex Tomas MB_CHECK_ASSERT(max * 2 == max2); 570c9de560dSAlex Tomas 571c9de560dSAlex Tomas count = 0; 572c9de560dSAlex Tomas for (i = 0; i < max; i++) { 573c9de560dSAlex Tomas 574c9de560dSAlex Tomas if (mb_test_bit(i, buddy)) { 575c9de560dSAlex Tomas /* only single bit in buddy2 may be 1 */ 576c9de560dSAlex Tomas if (!mb_test_bit(i << 1, buddy2)) { 577c9de560dSAlex Tomas MB_CHECK_ASSERT( 578c9de560dSAlex Tomas mb_test_bit((i<<1)+1, buddy2)); 579c9de560dSAlex Tomas } else if (!mb_test_bit((i << 1) + 1, buddy2)) { 580c9de560dSAlex Tomas MB_CHECK_ASSERT( 581c9de560dSAlex Tomas mb_test_bit(i << 1, buddy2)); 582c9de560dSAlex Tomas } 583c9de560dSAlex Tomas continue; 584c9de560dSAlex Tomas } 585c9de560dSAlex Tomas 5860a10da73SRobin Dong /* both bits in buddy2 must be 1 */ 587c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 588c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 589c9de560dSAlex Tomas 590c9de560dSAlex Tomas for (j = 0; j < (1 << order); j++) { 591c9de560dSAlex Tomas k = (i * (1 << order)) + j; 592c9de560dSAlex Tomas MB_CHECK_ASSERT( 593c5e8f3f3STheodore Ts'o !mb_test_bit(k, e4b->bd_bitmap)); 594c9de560dSAlex Tomas } 595c9de560dSAlex Tomas count++; 596c9de560dSAlex Tomas } 597c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 598c9de560dSAlex Tomas order--; 599c9de560dSAlex Tomas } 600c9de560dSAlex Tomas 601c9de560dSAlex Tomas fstart = -1; 602c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 603c9de560dSAlex Tomas for (i = 0; i < max; i++) { 604c9de560dSAlex Tomas if (!mb_test_bit(i, buddy)) { 605c9de560dSAlex Tomas MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 606c9de560dSAlex Tomas if (fstart == -1) { 607c9de560dSAlex Tomas fragments++; 608c9de560dSAlex Tomas fstart = i; 609c9de560dSAlex Tomas } 610c9de560dSAlex Tomas continue; 611c9de560dSAlex Tomas } 612c9de560dSAlex Tomas fstart = -1; 613c9de560dSAlex Tomas /* check used bits only */ 614c9de560dSAlex Tomas for (j = 0; j < e4b->bd_blkbits + 1; j++) { 615c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, j, &max2); 616c9de560dSAlex Tomas k = i >> j; 617c9de560dSAlex Tomas MB_CHECK_ASSERT(k < max2); 618c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 619c9de560dSAlex Tomas } 620c9de560dSAlex Tomas } 621c9de560dSAlex Tomas MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 622c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 623c9de560dSAlex Tomas 624c9de560dSAlex Tomas grp = ext4_get_group_info(sb, e4b->bd_group); 625c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 626c9de560dSAlex Tomas ext4_group_t groupnr; 627c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 62860bd63d1SSolofo Ramangalahy pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 62960bd63d1SSolofo Ramangalahy ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 630c9de560dSAlex Tomas MB_CHECK_ASSERT(groupnr == e4b->bd_group); 63160bd63d1SSolofo Ramangalahy for (i = 0; i < pa->pa_len; i++) 632c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 633c9de560dSAlex Tomas } 634c9de560dSAlex Tomas return 0; 635c9de560dSAlex Tomas } 636c9de560dSAlex Tomas #undef MB_CHECK_ASSERT 637c9de560dSAlex Tomas #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 63846e665e9SHarvey Harrison __FILE__, __func__, __LINE__) 639c9de560dSAlex Tomas #else 640c9de560dSAlex Tomas #define mb_check_buddy(e4b) 641c9de560dSAlex Tomas #endif 642c9de560dSAlex Tomas 6437c786059SColy Li /* 6447c786059SColy Li * Divide blocks started from @first with length @len into 6457c786059SColy Li * smaller chunks with power of 2 blocks. 6467c786059SColy Li * Clear the bits in bitmap which the blocks of the chunk(s) covered, 6477c786059SColy Li * then increase bb_counters[] for corresponded chunk size. 6487c786059SColy Li */ 649c9de560dSAlex Tomas static void ext4_mb_mark_free_simple(struct super_block *sb, 650a36b4498SEric Sandeen void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 651c9de560dSAlex Tomas struct ext4_group_info *grp) 652c9de560dSAlex Tomas { 653c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 654a36b4498SEric Sandeen ext4_grpblk_t min; 655a36b4498SEric Sandeen ext4_grpblk_t max; 656a36b4498SEric Sandeen ext4_grpblk_t chunk; 657c9de560dSAlex Tomas unsigned short border; 658c9de560dSAlex Tomas 6597137d7a4STheodore Ts'o BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb)); 660c9de560dSAlex Tomas 661c9de560dSAlex Tomas border = 2 << sb->s_blocksize_bits; 662c9de560dSAlex Tomas 663c9de560dSAlex Tomas while (len > 0) { 664c9de560dSAlex Tomas /* find how many blocks can be covered since this position */ 665c9de560dSAlex Tomas max = ffs(first | border) - 1; 666c9de560dSAlex Tomas 667c9de560dSAlex Tomas /* find how many blocks of power 2 we need to mark */ 668c9de560dSAlex Tomas min = fls(len) - 1; 669c9de560dSAlex Tomas 670c9de560dSAlex Tomas if (max < min) 671c9de560dSAlex Tomas min = max; 672c9de560dSAlex Tomas chunk = 1 << min; 673c9de560dSAlex Tomas 674c9de560dSAlex Tomas /* mark multiblock chunks only */ 675c9de560dSAlex Tomas grp->bb_counters[min]++; 676c9de560dSAlex Tomas if (min > 0) 677c9de560dSAlex Tomas mb_clear_bit(first >> min, 678c9de560dSAlex Tomas buddy + sbi->s_mb_offsets[min]); 679c9de560dSAlex Tomas 680c9de560dSAlex Tomas len -= chunk; 681c9de560dSAlex Tomas first += chunk; 682c9de560dSAlex Tomas } 683c9de560dSAlex Tomas } 684c9de560dSAlex Tomas 6858a57d9d6SCurt Wohlgemuth /* 6868a57d9d6SCurt Wohlgemuth * Cache the order of the largest free extent we have available in this block 6878a57d9d6SCurt Wohlgemuth * group. 6888a57d9d6SCurt Wohlgemuth */ 6898a57d9d6SCurt Wohlgemuth static void 6908a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) 6918a57d9d6SCurt Wohlgemuth { 6928a57d9d6SCurt Wohlgemuth int i; 6938a57d9d6SCurt Wohlgemuth int bits; 6948a57d9d6SCurt Wohlgemuth 6958a57d9d6SCurt Wohlgemuth grp->bb_largest_free_order = -1; /* uninit */ 6968a57d9d6SCurt Wohlgemuth 6978a57d9d6SCurt Wohlgemuth bits = sb->s_blocksize_bits + 1; 6988a57d9d6SCurt Wohlgemuth for (i = bits; i >= 0; i--) { 6998a57d9d6SCurt Wohlgemuth if (grp->bb_counters[i] > 0) { 7008a57d9d6SCurt Wohlgemuth grp->bb_largest_free_order = i; 7018a57d9d6SCurt Wohlgemuth break; 7028a57d9d6SCurt Wohlgemuth } 7038a57d9d6SCurt Wohlgemuth } 7048a57d9d6SCurt Wohlgemuth } 7058a57d9d6SCurt Wohlgemuth 706089ceeccSEric Sandeen static noinline_for_stack 707089ceeccSEric Sandeen void ext4_mb_generate_buddy(struct super_block *sb, 708c9de560dSAlex Tomas void *buddy, void *bitmap, ext4_group_t group) 709c9de560dSAlex Tomas { 710c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 7117137d7a4STheodore Ts'o ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb); 712a36b4498SEric Sandeen ext4_grpblk_t i = 0; 713a36b4498SEric Sandeen ext4_grpblk_t first; 714a36b4498SEric Sandeen ext4_grpblk_t len; 715c9de560dSAlex Tomas unsigned free = 0; 716c9de560dSAlex Tomas unsigned fragments = 0; 717c9de560dSAlex Tomas unsigned long long period = get_cycles(); 718c9de560dSAlex Tomas 719c9de560dSAlex Tomas /* initialize buddy from bitmap which is aggregation 720c9de560dSAlex Tomas * of on-disk bitmap and preallocations */ 721ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, 0); 722c9de560dSAlex Tomas grp->bb_first_free = i; 723c9de560dSAlex Tomas while (i < max) { 724c9de560dSAlex Tomas fragments++; 725c9de560dSAlex Tomas first = i; 726ffad0a44SAneesh Kumar K.V i = mb_find_next_bit(bitmap, max, i); 727c9de560dSAlex Tomas len = i - first; 728c9de560dSAlex Tomas free += len; 729c9de560dSAlex Tomas if (len > 1) 730c9de560dSAlex Tomas ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 731c9de560dSAlex Tomas else 732c9de560dSAlex Tomas grp->bb_counters[0]++; 733c9de560dSAlex Tomas if (i < max) 734ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, i); 735c9de560dSAlex Tomas } 736c9de560dSAlex Tomas grp->bb_fragments = fragments; 737c9de560dSAlex Tomas 738c9de560dSAlex Tomas if (free != grp->bb_free) { 739e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, 74053accfa9STheodore Ts'o "%u clusters in bitmap, %u in gd", 741e29136f8STheodore Ts'o free, grp->bb_free); 742e56eb659SAneesh Kumar K.V /* 743e56eb659SAneesh Kumar K.V * If we intent to continue, we consider group descritor 744e56eb659SAneesh Kumar K.V * corrupt and update bb_free using bitmap value 745e56eb659SAneesh Kumar K.V */ 746c9de560dSAlex Tomas grp->bb_free = free; 747c9de560dSAlex Tomas } 7488a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, grp); 749c9de560dSAlex Tomas 750c9de560dSAlex Tomas clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 751c9de560dSAlex Tomas 752c9de560dSAlex Tomas period = get_cycles() - period; 753c9de560dSAlex Tomas spin_lock(&EXT4_SB(sb)->s_bal_lock); 754c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_buddies_generated++; 755c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_generation_time += period; 756c9de560dSAlex Tomas spin_unlock(&EXT4_SB(sb)->s_bal_lock); 757c9de560dSAlex Tomas } 758c9de560dSAlex Tomas 759c9de560dSAlex Tomas /* The buddy information is attached the buddy cache inode 760c9de560dSAlex Tomas * for convenience. The information regarding each group 761c9de560dSAlex Tomas * is loaded via ext4_mb_load_buddy. The information involve 762c9de560dSAlex Tomas * block bitmap and buddy information. The information are 763c9de560dSAlex Tomas * stored in the inode as 764c9de560dSAlex Tomas * 765c9de560dSAlex Tomas * { page } 766c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 767c9de560dSAlex Tomas * 768c9de560dSAlex Tomas * 769c9de560dSAlex Tomas * one block each for bitmap and buddy information. 770c9de560dSAlex Tomas * So for each group we take up 2 blocks. A page can 771c9de560dSAlex Tomas * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. 772c9de560dSAlex Tomas * So it can have information regarding groups_per_page which 773c9de560dSAlex Tomas * is blocks_per_page/2 7748a57d9d6SCurt Wohlgemuth * 7758a57d9d6SCurt Wohlgemuth * Locking note: This routine takes the block group lock of all groups 7768a57d9d6SCurt Wohlgemuth * for this page; do not hold this lock when calling this routine! 777c9de560dSAlex Tomas */ 778c9de560dSAlex Tomas 779c9de560dSAlex Tomas static int ext4_mb_init_cache(struct page *page, char *incore) 780c9de560dSAlex Tomas { 7818df9675fSTheodore Ts'o ext4_group_t ngroups; 782c9de560dSAlex Tomas int blocksize; 783c9de560dSAlex Tomas int blocks_per_page; 784c9de560dSAlex Tomas int groups_per_page; 785c9de560dSAlex Tomas int err = 0; 786c9de560dSAlex Tomas int i; 787813e5727STheodore Ts'o ext4_group_t first_group, group; 788c9de560dSAlex Tomas int first_block; 789c9de560dSAlex Tomas struct super_block *sb; 790c9de560dSAlex Tomas struct buffer_head *bhs; 791c9de560dSAlex Tomas struct buffer_head **bh; 792c9de560dSAlex Tomas struct inode *inode; 793c9de560dSAlex Tomas char *data; 794c9de560dSAlex Tomas char *bitmap; 7959b8b7d35SAmir Goldstein struct ext4_group_info *grinfo; 796c9de560dSAlex Tomas 7976ba495e9STheodore Ts'o mb_debug(1, "init page %lu\n", page->index); 798c9de560dSAlex Tomas 799c9de560dSAlex Tomas inode = page->mapping->host; 800c9de560dSAlex Tomas sb = inode->i_sb; 8018df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 802c9de560dSAlex Tomas blocksize = 1 << inode->i_blkbits; 803c9de560dSAlex Tomas blocks_per_page = PAGE_CACHE_SIZE / blocksize; 804c9de560dSAlex Tomas 805c9de560dSAlex Tomas groups_per_page = blocks_per_page >> 1; 806c9de560dSAlex Tomas if (groups_per_page == 0) 807c9de560dSAlex Tomas groups_per_page = 1; 808c9de560dSAlex Tomas 809c9de560dSAlex Tomas /* allocate buffer_heads to read bitmaps */ 810c9de560dSAlex Tomas if (groups_per_page > 1) { 811c9de560dSAlex Tomas i = sizeof(struct buffer_head *) * groups_per_page; 812c9de560dSAlex Tomas bh = kzalloc(i, GFP_NOFS); 813813e5727STheodore Ts'o if (bh == NULL) { 814813e5727STheodore Ts'o err = -ENOMEM; 815c9de560dSAlex Tomas goto out; 816813e5727STheodore Ts'o } 817c9de560dSAlex Tomas } else 818c9de560dSAlex Tomas bh = &bhs; 819c9de560dSAlex Tomas 820c9de560dSAlex Tomas first_group = page->index * blocks_per_page / 2; 821c9de560dSAlex Tomas 822c9de560dSAlex Tomas /* read all groups the page covers into the cache */ 823813e5727STheodore Ts'o for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 824813e5727STheodore Ts'o if (group >= ngroups) 825c9de560dSAlex Tomas break; 826c9de560dSAlex Tomas 827813e5727STheodore Ts'o grinfo = ext4_get_group_info(sb, group); 8289b8b7d35SAmir Goldstein /* 8299b8b7d35SAmir Goldstein * If page is uptodate then we came here after online resize 8309b8b7d35SAmir Goldstein * which added some new uninitialized group info structs, so 8319b8b7d35SAmir Goldstein * we must skip all initialized uptodate buddies on the page, 8329b8b7d35SAmir Goldstein * which may be currently in use by an allocating task. 8339b8b7d35SAmir Goldstein */ 8349b8b7d35SAmir Goldstein if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) { 8359b8b7d35SAmir Goldstein bh[i] = NULL; 8369b8b7d35SAmir Goldstein continue; 8379b8b7d35SAmir Goldstein } 838813e5727STheodore Ts'o if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) { 839c9de560dSAlex Tomas err = -ENOMEM; 840c9de560dSAlex Tomas goto out; 8412ccb5fb9SAneesh Kumar K.V } 842813e5727STheodore Ts'o mb_debug(1, "read bitmap for group %u\n", group); 843c9de560dSAlex Tomas } 844c9de560dSAlex Tomas 845c9de560dSAlex Tomas /* wait for I/O completion */ 846813e5727STheodore Ts'o for (i = 0, group = first_group; i < groups_per_page; i++, group++) { 847813e5727STheodore Ts'o if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) { 848c9de560dSAlex Tomas err = -EIO; 849c9de560dSAlex Tomas goto out; 850813e5727STheodore Ts'o } 851813e5727STheodore Ts'o } 852c9de560dSAlex Tomas 853c9de560dSAlex Tomas first_block = page->index * blocks_per_page; 854c9de560dSAlex Tomas for (i = 0; i < blocks_per_page; i++) { 855c9de560dSAlex Tomas int group; 856c9de560dSAlex Tomas 857c9de560dSAlex Tomas group = (first_block + i) >> 1; 8588df9675fSTheodore Ts'o if (group >= ngroups) 859c9de560dSAlex Tomas break; 860c9de560dSAlex Tomas 8619b8b7d35SAmir Goldstein if (!bh[group - first_group]) 8629b8b7d35SAmir Goldstein /* skip initialized uptodate buddy */ 8639b8b7d35SAmir Goldstein continue; 8649b8b7d35SAmir Goldstein 865c9de560dSAlex Tomas /* 866c9de560dSAlex Tomas * data carry information regarding this 867c9de560dSAlex Tomas * particular group in the format specified 868c9de560dSAlex Tomas * above 869c9de560dSAlex Tomas * 870c9de560dSAlex Tomas */ 871c9de560dSAlex Tomas data = page_address(page) + (i * blocksize); 872c9de560dSAlex Tomas bitmap = bh[group - first_group]->b_data; 873c9de560dSAlex Tomas 874c9de560dSAlex Tomas /* 875c9de560dSAlex Tomas * We place the buddy block and bitmap block 876c9de560dSAlex Tomas * close together 877c9de560dSAlex Tomas */ 878c9de560dSAlex Tomas if ((first_block + i) & 1) { 879c9de560dSAlex Tomas /* this is block of buddy */ 880c9de560dSAlex Tomas BUG_ON(incore == NULL); 8816ba495e9STheodore Ts'o mb_debug(1, "put buddy for group %u in page %lu/%x\n", 882c9de560dSAlex Tomas group, page->index, i * blocksize); 883f307333eSTheodore Ts'o trace_ext4_mb_buddy_bitmap_load(sb, group); 884c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, group); 885c9de560dSAlex Tomas grinfo->bb_fragments = 0; 886c9de560dSAlex Tomas memset(grinfo->bb_counters, 0, 8871927805eSEric Sandeen sizeof(*grinfo->bb_counters) * 8881927805eSEric Sandeen (sb->s_blocksize_bits+2)); 889c9de560dSAlex Tomas /* 890c9de560dSAlex Tomas * incore got set to the group block bitmap below 891c9de560dSAlex Tomas */ 8927a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, group); 8939b8b7d35SAmir Goldstein /* init the buddy */ 8949b8b7d35SAmir Goldstein memset(data, 0xff, blocksize); 895c9de560dSAlex Tomas ext4_mb_generate_buddy(sb, data, incore, group); 8967a2fcbf7SAneesh Kumar K.V ext4_unlock_group(sb, group); 897c9de560dSAlex Tomas incore = NULL; 898c9de560dSAlex Tomas } else { 899c9de560dSAlex Tomas /* this is block of bitmap */ 900c9de560dSAlex Tomas BUG_ON(incore != NULL); 9016ba495e9STheodore Ts'o mb_debug(1, "put bitmap for group %u in page %lu/%x\n", 902c9de560dSAlex Tomas group, page->index, i * blocksize); 903f307333eSTheodore Ts'o trace_ext4_mb_bitmap_load(sb, group); 904c9de560dSAlex Tomas 905c9de560dSAlex Tomas /* see comments in ext4_mb_put_pa() */ 906c9de560dSAlex Tomas ext4_lock_group(sb, group); 907c9de560dSAlex Tomas memcpy(data, bitmap, blocksize); 908c9de560dSAlex Tomas 909c9de560dSAlex Tomas /* mark all preallocated blks used in in-core bitmap */ 910c9de560dSAlex Tomas ext4_mb_generate_from_pa(sb, data, group); 9117a2fcbf7SAneesh Kumar K.V ext4_mb_generate_from_freelist(sb, data, group); 912c9de560dSAlex Tomas ext4_unlock_group(sb, group); 913c9de560dSAlex Tomas 914c9de560dSAlex Tomas /* set incore so that the buddy information can be 915c9de560dSAlex Tomas * generated using this 916c9de560dSAlex Tomas */ 917c9de560dSAlex Tomas incore = data; 918c9de560dSAlex Tomas } 919c9de560dSAlex Tomas } 920c9de560dSAlex Tomas SetPageUptodate(page); 921c9de560dSAlex Tomas 922c9de560dSAlex Tomas out: 923c9de560dSAlex Tomas if (bh) { 9249b8b7d35SAmir Goldstein for (i = 0; i < groups_per_page; i++) 925c9de560dSAlex Tomas brelse(bh[i]); 926c9de560dSAlex Tomas if (bh != &bhs) 927c9de560dSAlex Tomas kfree(bh); 928c9de560dSAlex Tomas } 929c9de560dSAlex Tomas return err; 930c9de560dSAlex Tomas } 931c9de560dSAlex Tomas 9328a57d9d6SCurt Wohlgemuth /* 9332de8807bSAmir Goldstein * Lock the buddy and bitmap pages. This make sure other parallel init_group 9342de8807bSAmir Goldstein * on the same buddy page doesn't happen whild holding the buddy page lock. 9352de8807bSAmir Goldstein * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap 9362de8807bSAmir Goldstein * are on the same page e4b->bd_buddy_page is NULL and return value is 0. 937eee4adc7SEric Sandeen */ 9382de8807bSAmir Goldstein static int ext4_mb_get_buddy_page_lock(struct super_block *sb, 9392de8807bSAmir Goldstein ext4_group_t group, struct ext4_buddy *e4b) 940eee4adc7SEric Sandeen { 9412de8807bSAmir Goldstein struct inode *inode = EXT4_SB(sb)->s_buddy_cache; 9422de8807bSAmir Goldstein int block, pnum, poff; 943eee4adc7SEric Sandeen int blocks_per_page; 9442de8807bSAmir Goldstein struct page *page; 9452de8807bSAmir Goldstein 9462de8807bSAmir Goldstein e4b->bd_buddy_page = NULL; 9472de8807bSAmir Goldstein e4b->bd_bitmap_page = NULL; 948eee4adc7SEric Sandeen 949eee4adc7SEric Sandeen blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 950eee4adc7SEric Sandeen /* 951eee4adc7SEric Sandeen * the buddy cache inode stores the block bitmap 952eee4adc7SEric Sandeen * and buddy information in consecutive blocks. 953eee4adc7SEric Sandeen * So for each group we need two blocks. 954eee4adc7SEric Sandeen */ 955eee4adc7SEric Sandeen block = group * 2; 956eee4adc7SEric Sandeen pnum = block / blocks_per_page; 9572de8807bSAmir Goldstein poff = block % blocks_per_page; 9582de8807bSAmir Goldstein page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 9592de8807bSAmir Goldstein if (!page) 9602de8807bSAmir Goldstein return -EIO; 9612de8807bSAmir Goldstein BUG_ON(page->mapping != inode->i_mapping); 9622de8807bSAmir Goldstein e4b->bd_bitmap_page = page; 9632de8807bSAmir Goldstein e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 964eee4adc7SEric Sandeen 9652de8807bSAmir Goldstein if (blocks_per_page >= 2) { 9662de8807bSAmir Goldstein /* buddy and bitmap are on the same page */ 9672de8807bSAmir Goldstein return 0; 968eee4adc7SEric Sandeen } 969eee4adc7SEric Sandeen 9702de8807bSAmir Goldstein block++; 971eee4adc7SEric Sandeen pnum = block / blocks_per_page; 9722de8807bSAmir Goldstein poff = block % blocks_per_page; 9732de8807bSAmir Goldstein page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 9742de8807bSAmir Goldstein if (!page) 9752de8807bSAmir Goldstein return -EIO; 9762de8807bSAmir Goldstein BUG_ON(page->mapping != inode->i_mapping); 9772de8807bSAmir Goldstein e4b->bd_buddy_page = page; 9782de8807bSAmir Goldstein return 0; 979eee4adc7SEric Sandeen } 980eee4adc7SEric Sandeen 9812de8807bSAmir Goldstein static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b) 9822de8807bSAmir Goldstein { 9832de8807bSAmir Goldstein if (e4b->bd_bitmap_page) { 9842de8807bSAmir Goldstein unlock_page(e4b->bd_bitmap_page); 9852de8807bSAmir Goldstein page_cache_release(e4b->bd_bitmap_page); 9862de8807bSAmir Goldstein } 9872de8807bSAmir Goldstein if (e4b->bd_buddy_page) { 9882de8807bSAmir Goldstein unlock_page(e4b->bd_buddy_page); 9892de8807bSAmir Goldstein page_cache_release(e4b->bd_buddy_page); 9902de8807bSAmir Goldstein } 991eee4adc7SEric Sandeen } 992eee4adc7SEric Sandeen 993eee4adc7SEric Sandeen /* 9948a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 9958a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 9968a57d9d6SCurt Wohlgemuth * calling this routine! 9978a57d9d6SCurt Wohlgemuth */ 998b6a758ecSAneesh Kumar K.V static noinline_for_stack 999b6a758ecSAneesh Kumar K.V int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) 1000b6a758ecSAneesh Kumar K.V { 1001b6a758ecSAneesh Kumar K.V 1002b6a758ecSAneesh Kumar K.V struct ext4_group_info *this_grp; 10032de8807bSAmir Goldstein struct ext4_buddy e4b; 10042de8807bSAmir Goldstein struct page *page; 10052de8807bSAmir Goldstein int ret = 0; 1006b6a758ecSAneesh Kumar K.V 1007b6a758ecSAneesh Kumar K.V mb_debug(1, "init group %u\n", group); 1008b6a758ecSAneesh Kumar K.V this_grp = ext4_get_group_info(sb, group); 1009b6a758ecSAneesh Kumar K.V /* 101008c3a813SAneesh Kumar K.V * This ensures that we don't reinit the buddy cache 101108c3a813SAneesh Kumar K.V * page which map to the group from which we are already 101208c3a813SAneesh Kumar K.V * allocating. If we are looking at the buddy cache we would 101308c3a813SAneesh Kumar K.V * have taken a reference using ext4_mb_load_buddy and that 10142de8807bSAmir Goldstein * would have pinned buddy page to page cache. 1015b6a758ecSAneesh Kumar K.V */ 10162de8807bSAmir Goldstein ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b); 10172de8807bSAmir Goldstein if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) { 1018b6a758ecSAneesh Kumar K.V /* 1019b6a758ecSAneesh Kumar K.V * somebody initialized the group 1020b6a758ecSAneesh Kumar K.V * return without doing anything 1021b6a758ecSAneesh Kumar K.V */ 1022b6a758ecSAneesh Kumar K.V goto err; 1023b6a758ecSAneesh Kumar K.V } 10242de8807bSAmir Goldstein 10252de8807bSAmir Goldstein page = e4b.bd_bitmap_page; 1026b6a758ecSAneesh Kumar K.V ret = ext4_mb_init_cache(page, NULL); 10272de8807bSAmir Goldstein if (ret) 1028b6a758ecSAneesh Kumar K.V goto err; 10292de8807bSAmir Goldstein if (!PageUptodate(page)) { 1030b6a758ecSAneesh Kumar K.V ret = -EIO; 1031b6a758ecSAneesh Kumar K.V goto err; 1032b6a758ecSAneesh Kumar K.V } 1033b6a758ecSAneesh Kumar K.V mark_page_accessed(page); 1034b6a758ecSAneesh Kumar K.V 10352de8807bSAmir Goldstein if (e4b.bd_buddy_page == NULL) { 1036b6a758ecSAneesh Kumar K.V /* 1037b6a758ecSAneesh Kumar K.V * If both the bitmap and buddy are in 1038b6a758ecSAneesh Kumar K.V * the same page we don't need to force 1039b6a758ecSAneesh Kumar K.V * init the buddy 1040b6a758ecSAneesh Kumar K.V */ 10412de8807bSAmir Goldstein ret = 0; 1042b6a758ecSAneesh Kumar K.V goto err; 1043b6a758ecSAneesh Kumar K.V } 10442de8807bSAmir Goldstein /* init buddy cache */ 10452de8807bSAmir Goldstein page = e4b.bd_buddy_page; 10462de8807bSAmir Goldstein ret = ext4_mb_init_cache(page, e4b.bd_bitmap); 10472de8807bSAmir Goldstein if (ret) 10482de8807bSAmir Goldstein goto err; 10492de8807bSAmir Goldstein if (!PageUptodate(page)) { 1050b6a758ecSAneesh Kumar K.V ret = -EIO; 1051b6a758ecSAneesh Kumar K.V goto err; 1052b6a758ecSAneesh Kumar K.V } 1053b6a758ecSAneesh Kumar K.V mark_page_accessed(page); 1054b6a758ecSAneesh Kumar K.V err: 10552de8807bSAmir Goldstein ext4_mb_put_buddy_page_lock(&e4b); 1056b6a758ecSAneesh Kumar K.V return ret; 1057b6a758ecSAneesh Kumar K.V } 1058b6a758ecSAneesh Kumar K.V 10598a57d9d6SCurt Wohlgemuth /* 10608a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 10618a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 10628a57d9d6SCurt Wohlgemuth * calling this routine! 10638a57d9d6SCurt Wohlgemuth */ 10644ddfef7bSEric Sandeen static noinline_for_stack int 10654ddfef7bSEric Sandeen ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1066c9de560dSAlex Tomas struct ext4_buddy *e4b) 1067c9de560dSAlex Tomas { 1068c9de560dSAlex Tomas int blocks_per_page; 1069c9de560dSAlex Tomas int block; 1070c9de560dSAlex Tomas int pnum; 1071c9de560dSAlex Tomas int poff; 1072c9de560dSAlex Tomas struct page *page; 1073fdf6c7a7SShen Feng int ret; 1074920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1075920313a7SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 1076920313a7SAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 1077c9de560dSAlex Tomas 10786ba495e9STheodore Ts'o mb_debug(1, "load group %u\n", group); 1079c9de560dSAlex Tomas 1080c9de560dSAlex Tomas blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1081920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 1082c9de560dSAlex Tomas 1083c9de560dSAlex Tomas e4b->bd_blkbits = sb->s_blocksize_bits; 1084529da704STao Ma e4b->bd_info = grp; 1085c9de560dSAlex Tomas e4b->bd_sb = sb; 1086c9de560dSAlex Tomas e4b->bd_group = group; 1087c9de560dSAlex Tomas e4b->bd_buddy_page = NULL; 1088c9de560dSAlex Tomas e4b->bd_bitmap_page = NULL; 1089c9de560dSAlex Tomas 1090f41c0750SAneesh Kumar K.V if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1091f41c0750SAneesh Kumar K.V /* 1092f41c0750SAneesh Kumar K.V * we need full data about the group 1093f41c0750SAneesh Kumar K.V * to make a good selection 1094f41c0750SAneesh Kumar K.V */ 1095f41c0750SAneesh Kumar K.V ret = ext4_mb_init_group(sb, group); 1096f41c0750SAneesh Kumar K.V if (ret) 1097f41c0750SAneesh Kumar K.V return ret; 1098f41c0750SAneesh Kumar K.V } 1099f41c0750SAneesh Kumar K.V 1100c9de560dSAlex Tomas /* 1101c9de560dSAlex Tomas * the buddy cache inode stores the block bitmap 1102c9de560dSAlex Tomas * and buddy information in consecutive blocks. 1103c9de560dSAlex Tomas * So for each group we need two blocks. 1104c9de560dSAlex Tomas */ 1105c9de560dSAlex Tomas block = group * 2; 1106c9de560dSAlex Tomas pnum = block / blocks_per_page; 1107c9de560dSAlex Tomas poff = block % blocks_per_page; 1108c9de560dSAlex Tomas 1109c9de560dSAlex Tomas /* we could use find_or_create_page(), but it locks page 1110c9de560dSAlex Tomas * what we'd like to avoid in fast path ... */ 1111c9de560dSAlex Tomas page = find_get_page(inode->i_mapping, pnum); 1112c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1113c9de560dSAlex Tomas if (page) 1114920313a7SAneesh Kumar K.V /* 1115920313a7SAneesh Kumar K.V * drop the page reference and try 1116920313a7SAneesh Kumar K.V * to get the page with lock. If we 1117920313a7SAneesh Kumar K.V * are not uptodate that implies 1118920313a7SAneesh Kumar K.V * somebody just created the page but 1119920313a7SAneesh Kumar K.V * is yet to initialize the same. So 1120920313a7SAneesh Kumar K.V * wait for it to initialize. 1121920313a7SAneesh Kumar K.V */ 1122c9de560dSAlex Tomas page_cache_release(page); 1123c9de560dSAlex Tomas page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1124c9de560dSAlex Tomas if (page) { 1125c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 1126c9de560dSAlex Tomas if (!PageUptodate(page)) { 1127fdf6c7a7SShen Feng ret = ext4_mb_init_cache(page, NULL); 1128fdf6c7a7SShen Feng if (ret) { 1129fdf6c7a7SShen Feng unlock_page(page); 1130fdf6c7a7SShen Feng goto err; 1131fdf6c7a7SShen Feng } 1132c9de560dSAlex Tomas mb_cmp_bitmaps(e4b, page_address(page) + 1133c9de560dSAlex Tomas (poff * sb->s_blocksize)); 1134c9de560dSAlex Tomas } 1135c9de560dSAlex Tomas unlock_page(page); 1136c9de560dSAlex Tomas } 1137c9de560dSAlex Tomas } 1138fdf6c7a7SShen Feng if (page == NULL || !PageUptodate(page)) { 1139fdf6c7a7SShen Feng ret = -EIO; 1140c9de560dSAlex Tomas goto err; 1141fdf6c7a7SShen Feng } 1142c9de560dSAlex Tomas e4b->bd_bitmap_page = page; 1143c9de560dSAlex Tomas e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1144c9de560dSAlex Tomas mark_page_accessed(page); 1145c9de560dSAlex Tomas 1146c9de560dSAlex Tomas block++; 1147c9de560dSAlex Tomas pnum = block / blocks_per_page; 1148c9de560dSAlex Tomas poff = block % blocks_per_page; 1149c9de560dSAlex Tomas 1150c9de560dSAlex Tomas page = find_get_page(inode->i_mapping, pnum); 1151c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1152c9de560dSAlex Tomas if (page) 1153c9de560dSAlex Tomas page_cache_release(page); 1154c9de560dSAlex Tomas page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1155c9de560dSAlex Tomas if (page) { 1156c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 1157fdf6c7a7SShen Feng if (!PageUptodate(page)) { 1158fdf6c7a7SShen Feng ret = ext4_mb_init_cache(page, e4b->bd_bitmap); 1159fdf6c7a7SShen Feng if (ret) { 1160fdf6c7a7SShen Feng unlock_page(page); 1161fdf6c7a7SShen Feng goto err; 1162fdf6c7a7SShen Feng } 1163fdf6c7a7SShen Feng } 1164c9de560dSAlex Tomas unlock_page(page); 1165c9de560dSAlex Tomas } 1166c9de560dSAlex Tomas } 1167fdf6c7a7SShen Feng if (page == NULL || !PageUptodate(page)) { 1168fdf6c7a7SShen Feng ret = -EIO; 1169c9de560dSAlex Tomas goto err; 1170fdf6c7a7SShen Feng } 1171c9de560dSAlex Tomas e4b->bd_buddy_page = page; 1172c9de560dSAlex Tomas e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 1173c9de560dSAlex Tomas mark_page_accessed(page); 1174c9de560dSAlex Tomas 1175c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 1176c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 1177c9de560dSAlex Tomas 1178c9de560dSAlex Tomas return 0; 1179c9de560dSAlex Tomas 1180c9de560dSAlex Tomas err: 118126626f11SYang Ruirui if (page) 118226626f11SYang Ruirui page_cache_release(page); 1183c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 1184c9de560dSAlex Tomas page_cache_release(e4b->bd_bitmap_page); 1185c9de560dSAlex Tomas if (e4b->bd_buddy_page) 1186c9de560dSAlex Tomas page_cache_release(e4b->bd_buddy_page); 1187c9de560dSAlex Tomas e4b->bd_buddy = NULL; 1188c9de560dSAlex Tomas e4b->bd_bitmap = NULL; 1189fdf6c7a7SShen Feng return ret; 1190c9de560dSAlex Tomas } 1191c9de560dSAlex Tomas 1192e39e07fdSJing Zhang static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) 1193c9de560dSAlex Tomas { 1194c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 1195c9de560dSAlex Tomas page_cache_release(e4b->bd_bitmap_page); 1196c9de560dSAlex Tomas if (e4b->bd_buddy_page) 1197c9de560dSAlex Tomas page_cache_release(e4b->bd_buddy_page); 1198c9de560dSAlex Tomas } 1199c9de560dSAlex Tomas 1200c9de560dSAlex Tomas 1201c9de560dSAlex Tomas static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1202c9de560dSAlex Tomas { 1203c9de560dSAlex Tomas int order = 1; 1204c9de560dSAlex Tomas void *bb; 1205c9de560dSAlex Tomas 1206c5e8f3f3STheodore Ts'o BUG_ON(e4b->bd_bitmap == e4b->bd_buddy); 1207c9de560dSAlex Tomas BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1208c9de560dSAlex Tomas 1209c5e8f3f3STheodore Ts'o bb = e4b->bd_buddy; 1210c9de560dSAlex Tomas while (order <= e4b->bd_blkbits + 1) { 1211c9de560dSAlex Tomas block = block >> 1; 1212c9de560dSAlex Tomas if (!mb_test_bit(block, bb)) { 1213c9de560dSAlex Tomas /* this block is part of buddy of order 'order' */ 1214c9de560dSAlex Tomas return order; 1215c9de560dSAlex Tomas } 1216c9de560dSAlex Tomas bb += 1 << (e4b->bd_blkbits - order); 1217c9de560dSAlex Tomas order++; 1218c9de560dSAlex Tomas } 1219c9de560dSAlex Tomas return 0; 1220c9de560dSAlex Tomas } 1221c9de560dSAlex Tomas 1222955ce5f5SAneesh Kumar K.V static void mb_clear_bits(void *bm, int cur, int len) 1223c9de560dSAlex Tomas { 1224c9de560dSAlex Tomas __u32 *addr; 1225c9de560dSAlex Tomas 1226c9de560dSAlex Tomas len = cur + len; 1227c9de560dSAlex Tomas while (cur < len) { 1228c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1229c9de560dSAlex Tomas /* fast path: clear whole word at once */ 1230c9de560dSAlex Tomas addr = bm + (cur >> 3); 1231c9de560dSAlex Tomas *addr = 0; 1232c9de560dSAlex Tomas cur += 32; 1233c9de560dSAlex Tomas continue; 1234c9de560dSAlex Tomas } 1235e8134b27SAneesh Kumar K.V mb_clear_bit(cur, bm); 1236c9de560dSAlex Tomas cur++; 1237c9de560dSAlex Tomas } 1238c9de560dSAlex Tomas } 1239c9de560dSAlex Tomas 1240c3e94d1dSYongqiang Yang void ext4_set_bits(void *bm, int cur, int len) 1241c9de560dSAlex Tomas { 1242c9de560dSAlex Tomas __u32 *addr; 1243c9de560dSAlex Tomas 1244c9de560dSAlex Tomas len = cur + len; 1245c9de560dSAlex Tomas while (cur < len) { 1246c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1247c9de560dSAlex Tomas /* fast path: set whole word at once */ 1248c9de560dSAlex Tomas addr = bm + (cur >> 3); 1249c9de560dSAlex Tomas *addr = 0xffffffff; 1250c9de560dSAlex Tomas cur += 32; 1251c9de560dSAlex Tomas continue; 1252c9de560dSAlex Tomas } 1253e8134b27SAneesh Kumar K.V mb_set_bit(cur, bm); 1254c9de560dSAlex Tomas cur++; 1255c9de560dSAlex Tomas } 1256c9de560dSAlex Tomas } 1257c9de560dSAlex Tomas 12587e5a8cddSShen Feng static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1259c9de560dSAlex Tomas int first, int count) 1260c9de560dSAlex Tomas { 1261c9de560dSAlex Tomas int block = 0; 1262c9de560dSAlex Tomas int max = 0; 1263c9de560dSAlex Tomas int order; 1264c9de560dSAlex Tomas void *buddy; 1265c9de560dSAlex Tomas void *buddy2; 1266c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 1267c9de560dSAlex Tomas 1268c9de560dSAlex Tomas BUG_ON(first + count > (sb->s_blocksize << 3)); 1269bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1270c9de560dSAlex Tomas mb_check_buddy(e4b); 1271c9de560dSAlex Tomas mb_free_blocks_double(inode, e4b, first, count); 1272c9de560dSAlex Tomas 1273c9de560dSAlex Tomas e4b->bd_info->bb_free += count; 1274c9de560dSAlex Tomas if (first < e4b->bd_info->bb_first_free) 1275c9de560dSAlex Tomas e4b->bd_info->bb_first_free = first; 1276c9de560dSAlex Tomas 1277c9de560dSAlex Tomas /* let's maintain fragments counter */ 1278c9de560dSAlex Tomas if (first != 0) 1279c5e8f3f3STheodore Ts'o block = !mb_test_bit(first - 1, e4b->bd_bitmap); 1280c9de560dSAlex Tomas if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) 1281c5e8f3f3STheodore Ts'o max = !mb_test_bit(first + count, e4b->bd_bitmap); 1282c9de560dSAlex Tomas if (block && max) 1283c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1284c9de560dSAlex Tomas else if (!block && !max) 1285c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1286c9de560dSAlex Tomas 1287c9de560dSAlex Tomas /* let's maintain buddy itself */ 1288c9de560dSAlex Tomas while (count-- > 0) { 1289c9de560dSAlex Tomas block = first++; 1290c9de560dSAlex Tomas order = 0; 1291c9de560dSAlex Tomas 1292c5e8f3f3STheodore Ts'o if (!mb_test_bit(block, e4b->bd_bitmap)) { 1293c9de560dSAlex Tomas ext4_fsblk_t blocknr; 12945661bd68SAkinobu Mita 12955661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 129653accfa9STheodore Ts'o blocknr += EXT4_C2B(EXT4_SB(sb), block); 12975d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 1298e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 1299e29136f8STheodore Ts'o blocknr, 1300e29136f8STheodore Ts'o "freeing already freed block " 1301e29136f8STheodore Ts'o "(bit %u)", block); 1302c9de560dSAlex Tomas } 1303c5e8f3f3STheodore Ts'o mb_clear_bit(block, e4b->bd_bitmap); 1304c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]++; 1305c9de560dSAlex Tomas 1306c9de560dSAlex Tomas /* start of the buddy */ 1307c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 1308c9de560dSAlex Tomas 1309c9de560dSAlex Tomas do { 1310c9de560dSAlex Tomas block &= ~1UL; 1311c9de560dSAlex Tomas if (mb_test_bit(block, buddy) || 1312c9de560dSAlex Tomas mb_test_bit(block + 1, buddy)) 1313c9de560dSAlex Tomas break; 1314c9de560dSAlex Tomas 1315c9de560dSAlex Tomas /* both the buddies are free, try to coalesce them */ 1316c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order + 1, &max); 1317c9de560dSAlex Tomas 1318c9de560dSAlex Tomas if (!buddy2) 1319c9de560dSAlex Tomas break; 1320c9de560dSAlex Tomas 1321c9de560dSAlex Tomas if (order > 0) { 1322c9de560dSAlex Tomas /* for special purposes, we don't set 1323c9de560dSAlex Tomas * free bits in bitmap */ 1324c9de560dSAlex Tomas mb_set_bit(block, buddy); 1325c9de560dSAlex Tomas mb_set_bit(block + 1, buddy); 1326c9de560dSAlex Tomas } 1327c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]--; 1328c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]--; 1329c9de560dSAlex Tomas 1330c9de560dSAlex Tomas block = block >> 1; 1331c9de560dSAlex Tomas order++; 1332c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]++; 1333c9de560dSAlex Tomas 1334c9de560dSAlex Tomas mb_clear_bit(block, buddy2); 1335c9de560dSAlex Tomas buddy = buddy2; 1336c9de560dSAlex Tomas } while (1); 1337c9de560dSAlex Tomas } 13388a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, e4b->bd_info); 1339c9de560dSAlex Tomas mb_check_buddy(e4b); 1340c9de560dSAlex Tomas } 1341c9de560dSAlex Tomas 1342c9de560dSAlex Tomas static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, 1343c9de560dSAlex Tomas int needed, struct ext4_free_extent *ex) 1344c9de560dSAlex Tomas { 1345c9de560dSAlex Tomas int next = block; 1346c9de560dSAlex Tomas int max; 1347c9de560dSAlex Tomas void *buddy; 1348c9de560dSAlex Tomas 1349bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1350c9de560dSAlex Tomas BUG_ON(ex == NULL); 1351c9de560dSAlex Tomas 1352c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 1353c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1354c9de560dSAlex Tomas BUG_ON(block >= max); 1355c9de560dSAlex Tomas if (mb_test_bit(block, buddy)) { 1356c9de560dSAlex Tomas ex->fe_len = 0; 1357c9de560dSAlex Tomas ex->fe_start = 0; 1358c9de560dSAlex Tomas ex->fe_group = 0; 1359c9de560dSAlex Tomas return 0; 1360c9de560dSAlex Tomas } 1361c9de560dSAlex Tomas 1362c9de560dSAlex Tomas /* FIXME dorp order completely ? */ 1363c9de560dSAlex Tomas if (likely(order == 0)) { 1364c9de560dSAlex Tomas /* find actual order */ 1365c9de560dSAlex Tomas order = mb_find_order_for_block(e4b, block); 1366c9de560dSAlex Tomas block = block >> order; 1367c9de560dSAlex Tomas } 1368c9de560dSAlex Tomas 1369c9de560dSAlex Tomas ex->fe_len = 1 << order; 1370c9de560dSAlex Tomas ex->fe_start = block << order; 1371c9de560dSAlex Tomas ex->fe_group = e4b->bd_group; 1372c9de560dSAlex Tomas 1373c9de560dSAlex Tomas /* calc difference from given start */ 1374c9de560dSAlex Tomas next = next - ex->fe_start; 1375c9de560dSAlex Tomas ex->fe_len -= next; 1376c9de560dSAlex Tomas ex->fe_start += next; 1377c9de560dSAlex Tomas 1378c9de560dSAlex Tomas while (needed > ex->fe_len && 1379c9de560dSAlex Tomas (buddy = mb_find_buddy(e4b, order, &max))) { 1380c9de560dSAlex Tomas 1381c9de560dSAlex Tomas if (block + 1 >= max) 1382c9de560dSAlex Tomas break; 1383c9de560dSAlex Tomas 1384c9de560dSAlex Tomas next = (block + 1) * (1 << order); 1385c5e8f3f3STheodore Ts'o if (mb_test_bit(next, e4b->bd_bitmap)) 1386c9de560dSAlex Tomas break; 1387c9de560dSAlex Tomas 1388b051d8dcSRobin Dong order = mb_find_order_for_block(e4b, next); 1389c9de560dSAlex Tomas 1390c9de560dSAlex Tomas block = next >> order; 1391c9de560dSAlex Tomas ex->fe_len += 1 << order; 1392c9de560dSAlex Tomas } 1393c9de560dSAlex Tomas 1394c9de560dSAlex Tomas BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); 1395c9de560dSAlex Tomas return ex->fe_len; 1396c9de560dSAlex Tomas } 1397c9de560dSAlex Tomas 1398c9de560dSAlex Tomas static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 1399c9de560dSAlex Tomas { 1400c9de560dSAlex Tomas int ord; 1401c9de560dSAlex Tomas int mlen = 0; 1402c9de560dSAlex Tomas int max = 0; 1403c9de560dSAlex Tomas int cur; 1404c9de560dSAlex Tomas int start = ex->fe_start; 1405c9de560dSAlex Tomas int len = ex->fe_len; 1406c9de560dSAlex Tomas unsigned ret = 0; 1407c9de560dSAlex Tomas int len0 = len; 1408c9de560dSAlex Tomas void *buddy; 1409c9de560dSAlex Tomas 1410c9de560dSAlex Tomas BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 1411c9de560dSAlex Tomas BUG_ON(e4b->bd_group != ex->fe_group); 1412bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1413c9de560dSAlex Tomas mb_check_buddy(e4b); 1414c9de560dSAlex Tomas mb_mark_used_double(e4b, start, len); 1415c9de560dSAlex Tomas 1416c9de560dSAlex Tomas e4b->bd_info->bb_free -= len; 1417c9de560dSAlex Tomas if (e4b->bd_info->bb_first_free == start) 1418c9de560dSAlex Tomas e4b->bd_info->bb_first_free += len; 1419c9de560dSAlex Tomas 1420c9de560dSAlex Tomas /* let's maintain fragments counter */ 1421c9de560dSAlex Tomas if (start != 0) 1422c5e8f3f3STheodore Ts'o mlen = !mb_test_bit(start - 1, e4b->bd_bitmap); 1423c9de560dSAlex Tomas if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 1424c5e8f3f3STheodore Ts'o max = !mb_test_bit(start + len, e4b->bd_bitmap); 1425c9de560dSAlex Tomas if (mlen && max) 1426c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1427c9de560dSAlex Tomas else if (!mlen && !max) 1428c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1429c9de560dSAlex Tomas 1430c9de560dSAlex Tomas /* let's maintain buddy itself */ 1431c9de560dSAlex Tomas while (len) { 1432c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, start); 1433c9de560dSAlex Tomas 1434c9de560dSAlex Tomas if (((start >> ord) << ord) == start && len >= (1 << ord)) { 1435c9de560dSAlex Tomas /* the whole chunk may be allocated at once! */ 1436c9de560dSAlex Tomas mlen = 1 << ord; 1437c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1438c9de560dSAlex Tomas BUG_ON((start >> ord) >= max); 1439c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1440c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1441c9de560dSAlex Tomas start += mlen; 1442c9de560dSAlex Tomas len -= mlen; 1443c9de560dSAlex Tomas BUG_ON(len < 0); 1444c9de560dSAlex Tomas continue; 1445c9de560dSAlex Tomas } 1446c9de560dSAlex Tomas 1447c9de560dSAlex Tomas /* store for history */ 1448c9de560dSAlex Tomas if (ret == 0) 1449c9de560dSAlex Tomas ret = len | (ord << 16); 1450c9de560dSAlex Tomas 1451c9de560dSAlex Tomas /* we have to split large buddy */ 1452c9de560dSAlex Tomas BUG_ON(ord <= 0); 1453c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1454c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1455c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1456c9de560dSAlex Tomas 1457c9de560dSAlex Tomas ord--; 1458c9de560dSAlex Tomas cur = (start >> ord) & ~1U; 1459c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1460c9de560dSAlex Tomas mb_clear_bit(cur, buddy); 1461c9de560dSAlex Tomas mb_clear_bit(cur + 1, buddy); 1462c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1463c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1464c9de560dSAlex Tomas } 14658a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); 1466c9de560dSAlex Tomas 1467c5e8f3f3STheodore Ts'o ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0); 1468c9de560dSAlex Tomas mb_check_buddy(e4b); 1469c9de560dSAlex Tomas 1470c9de560dSAlex Tomas return ret; 1471c9de560dSAlex Tomas } 1472c9de560dSAlex Tomas 1473c9de560dSAlex Tomas /* 1474c9de560dSAlex Tomas * Must be called under group lock! 1475c9de560dSAlex Tomas */ 1476c9de560dSAlex Tomas static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 1477c9de560dSAlex Tomas struct ext4_buddy *e4b) 1478c9de560dSAlex Tomas { 1479c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1480c9de560dSAlex Tomas int ret; 1481c9de560dSAlex Tomas 1482c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 1483c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1484c9de560dSAlex Tomas 1485c9de560dSAlex Tomas ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 1486c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 1487c9de560dSAlex Tomas ret = mb_mark_used(e4b, &ac->ac_b_ex); 1488c9de560dSAlex Tomas 1489c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 1490c9de560dSAlex Tomas * allocated blocks for history */ 1491c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 1492c9de560dSAlex Tomas 1493c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 1494c9de560dSAlex Tomas ac->ac_tail = ret & 0xffff; 1495c9de560dSAlex Tomas ac->ac_buddy = ret >> 16; 1496c9de560dSAlex Tomas 1497c3a326a6SAneesh Kumar K.V /* 1498c3a326a6SAneesh Kumar K.V * take the page reference. We want the page to be pinned 1499c3a326a6SAneesh Kumar K.V * so that we don't get a ext4_mb_init_cache_call for this 1500c3a326a6SAneesh Kumar K.V * group until we update the bitmap. That would mean we 1501c3a326a6SAneesh Kumar K.V * double allocate blocks. The reference is dropped 1502c3a326a6SAneesh Kumar K.V * in ext4_mb_release_context 1503c3a326a6SAneesh Kumar K.V */ 1504c9de560dSAlex Tomas ac->ac_bitmap_page = e4b->bd_bitmap_page; 1505c9de560dSAlex Tomas get_page(ac->ac_bitmap_page); 1506c9de560dSAlex Tomas ac->ac_buddy_page = e4b->bd_buddy_page; 1507c9de560dSAlex Tomas get_page(ac->ac_buddy_page); 1508c9de560dSAlex Tomas /* store last allocated for subsequent stream allocation */ 15094ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 1510c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 1511c9de560dSAlex Tomas sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 1512c9de560dSAlex Tomas sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 1513c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 1514c9de560dSAlex Tomas } 1515c9de560dSAlex Tomas } 1516c9de560dSAlex Tomas 1517c9de560dSAlex Tomas /* 1518c9de560dSAlex Tomas * regular allocator, for general purposes allocation 1519c9de560dSAlex Tomas */ 1520c9de560dSAlex Tomas 1521c9de560dSAlex Tomas static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 1522c9de560dSAlex Tomas struct ext4_buddy *e4b, 1523c9de560dSAlex Tomas int finish_group) 1524c9de560dSAlex Tomas { 1525c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1526c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1527c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1528c9de560dSAlex Tomas struct ext4_free_extent ex; 1529c9de560dSAlex Tomas int max; 1530c9de560dSAlex Tomas 1531032115fcSAneesh Kumar K.V if (ac->ac_status == AC_STATUS_FOUND) 1532032115fcSAneesh Kumar K.V return; 1533c9de560dSAlex Tomas /* 1534c9de560dSAlex Tomas * We don't want to scan for a whole year 1535c9de560dSAlex Tomas */ 1536c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan && 1537c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1538c9de560dSAlex Tomas ac->ac_status = AC_STATUS_BREAK; 1539c9de560dSAlex Tomas return; 1540c9de560dSAlex Tomas } 1541c9de560dSAlex Tomas 1542c9de560dSAlex Tomas /* 1543c9de560dSAlex Tomas * Haven't found good chunk so far, let's continue 1544c9de560dSAlex Tomas */ 1545c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) 1546c9de560dSAlex Tomas return; 1547c9de560dSAlex Tomas 1548c9de560dSAlex Tomas if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 1549c9de560dSAlex Tomas && bex->fe_group == e4b->bd_group) { 1550c9de560dSAlex Tomas /* recheck chunk's availability - we don't know 1551c9de560dSAlex Tomas * when it was found (within this lock-unlock 1552c9de560dSAlex Tomas * period or not) */ 1553c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); 1554c9de560dSAlex Tomas if (max >= gex->fe_len) { 1555c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1556c9de560dSAlex Tomas return; 1557c9de560dSAlex Tomas } 1558c9de560dSAlex Tomas } 1559c9de560dSAlex Tomas } 1560c9de560dSAlex Tomas 1561c9de560dSAlex Tomas /* 1562c9de560dSAlex Tomas * The routine checks whether found extent is good enough. If it is, 1563c9de560dSAlex Tomas * then the extent gets marked used and flag is set to the context 1564c9de560dSAlex Tomas * to stop scanning. Otherwise, the extent is compared with the 1565c9de560dSAlex Tomas * previous found extent and if new one is better, then it's stored 1566c9de560dSAlex Tomas * in the context. Later, the best found extent will be used, if 1567c9de560dSAlex Tomas * mballoc can't find good enough extent. 1568c9de560dSAlex Tomas * 1569c9de560dSAlex Tomas * FIXME: real allocation policy is to be designed yet! 1570c9de560dSAlex Tomas */ 1571c9de560dSAlex Tomas static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 1572c9de560dSAlex Tomas struct ext4_free_extent *ex, 1573c9de560dSAlex Tomas struct ext4_buddy *e4b) 1574c9de560dSAlex Tomas { 1575c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1576c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1577c9de560dSAlex Tomas 1578c9de560dSAlex Tomas BUG_ON(ex->fe_len <= 0); 15797137d7a4STheodore Ts'o BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 15807137d7a4STheodore Ts'o BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 1581c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 1582c9de560dSAlex Tomas 1583c9de560dSAlex Tomas ac->ac_found++; 1584c9de560dSAlex Tomas 1585c9de560dSAlex Tomas /* 1586c9de560dSAlex Tomas * The special case - take what you catch first 1587c9de560dSAlex Tomas */ 1588c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1589c9de560dSAlex Tomas *bex = *ex; 1590c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1591c9de560dSAlex Tomas return; 1592c9de560dSAlex Tomas } 1593c9de560dSAlex Tomas 1594c9de560dSAlex Tomas /* 1595c9de560dSAlex Tomas * Let's check whether the chuck is good enough 1596c9de560dSAlex Tomas */ 1597c9de560dSAlex Tomas if (ex->fe_len == gex->fe_len) { 1598c9de560dSAlex Tomas *bex = *ex; 1599c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1600c9de560dSAlex Tomas return; 1601c9de560dSAlex Tomas } 1602c9de560dSAlex Tomas 1603c9de560dSAlex Tomas /* 1604c9de560dSAlex Tomas * If this is first found extent, just store it in the context 1605c9de560dSAlex Tomas */ 1606c9de560dSAlex Tomas if (bex->fe_len == 0) { 1607c9de560dSAlex Tomas *bex = *ex; 1608c9de560dSAlex Tomas return; 1609c9de560dSAlex Tomas } 1610c9de560dSAlex Tomas 1611c9de560dSAlex Tomas /* 1612c9de560dSAlex Tomas * If new found extent is better, store it in the context 1613c9de560dSAlex Tomas */ 1614c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) { 1615c9de560dSAlex Tomas /* if the request isn't satisfied, any found extent 1616c9de560dSAlex Tomas * larger than previous best one is better */ 1617c9de560dSAlex Tomas if (ex->fe_len > bex->fe_len) 1618c9de560dSAlex Tomas *bex = *ex; 1619c9de560dSAlex Tomas } else if (ex->fe_len > gex->fe_len) { 1620c9de560dSAlex Tomas /* if the request is satisfied, then we try to find 1621c9de560dSAlex Tomas * an extent that still satisfy the request, but is 1622c9de560dSAlex Tomas * smaller than previous one */ 1623c9de560dSAlex Tomas if (ex->fe_len < bex->fe_len) 1624c9de560dSAlex Tomas *bex = *ex; 1625c9de560dSAlex Tomas } 1626c9de560dSAlex Tomas 1627c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 0); 1628c9de560dSAlex Tomas } 1629c9de560dSAlex Tomas 1630089ceeccSEric Sandeen static noinline_for_stack 1631089ceeccSEric Sandeen int ext4_mb_try_best_found(struct ext4_allocation_context *ac, 1632c9de560dSAlex Tomas struct ext4_buddy *e4b) 1633c9de560dSAlex Tomas { 1634c9de560dSAlex Tomas struct ext4_free_extent ex = ac->ac_b_ex; 1635c9de560dSAlex Tomas ext4_group_t group = ex.fe_group; 1636c9de560dSAlex Tomas int max; 1637c9de560dSAlex Tomas int err; 1638c9de560dSAlex Tomas 1639c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1640c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1641c9de560dSAlex Tomas if (err) 1642c9de560dSAlex Tomas return err; 1643c9de560dSAlex Tomas 1644c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 1645c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); 1646c9de560dSAlex Tomas 1647c9de560dSAlex Tomas if (max > 0) { 1648c9de560dSAlex Tomas ac->ac_b_ex = ex; 1649c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1650c9de560dSAlex Tomas } 1651c9de560dSAlex Tomas 1652c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1653e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 1654c9de560dSAlex Tomas 1655c9de560dSAlex Tomas return 0; 1656c9de560dSAlex Tomas } 1657c9de560dSAlex Tomas 1658089ceeccSEric Sandeen static noinline_for_stack 1659089ceeccSEric Sandeen int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 1660c9de560dSAlex Tomas struct ext4_buddy *e4b) 1661c9de560dSAlex Tomas { 1662c9de560dSAlex Tomas ext4_group_t group = ac->ac_g_ex.fe_group; 1663c9de560dSAlex Tomas int max; 1664c9de560dSAlex Tomas int err; 1665c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1666c9de560dSAlex Tomas struct ext4_free_extent ex; 1667c9de560dSAlex Tomas 1668c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) 1669c9de560dSAlex Tomas return 0; 1670c9de560dSAlex Tomas 1671c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1672c9de560dSAlex Tomas if (err) 1673c9de560dSAlex Tomas return err; 1674c9de560dSAlex Tomas 1675c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 1676c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, 1677c9de560dSAlex Tomas ac->ac_g_ex.fe_len, &ex); 1678c9de560dSAlex Tomas 1679c9de560dSAlex Tomas if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { 1680c9de560dSAlex Tomas ext4_fsblk_t start; 1681c9de560dSAlex Tomas 16825661bd68SAkinobu Mita start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + 16835661bd68SAkinobu Mita ex.fe_start; 1684c9de560dSAlex Tomas /* use do_div to get remainder (would be 64-bit modulo) */ 1685c9de560dSAlex Tomas if (do_div(start, sbi->s_stripe) == 0) { 1686c9de560dSAlex Tomas ac->ac_found++; 1687c9de560dSAlex Tomas ac->ac_b_ex = ex; 1688c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1689c9de560dSAlex Tomas } 1690c9de560dSAlex Tomas } else if (max >= ac->ac_g_ex.fe_len) { 1691c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1692c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1693c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1694c9de560dSAlex Tomas ac->ac_found++; 1695c9de560dSAlex Tomas ac->ac_b_ex = ex; 1696c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1697c9de560dSAlex Tomas } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 1698c9de560dSAlex Tomas /* Sometimes, caller may want to merge even small 1699c9de560dSAlex Tomas * number of blocks to an existing extent */ 1700c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1701c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1702c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1703c9de560dSAlex Tomas ac->ac_found++; 1704c9de560dSAlex Tomas ac->ac_b_ex = ex; 1705c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1706c9de560dSAlex Tomas } 1707c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1708e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 1709c9de560dSAlex Tomas 1710c9de560dSAlex Tomas return 0; 1711c9de560dSAlex Tomas } 1712c9de560dSAlex Tomas 1713c9de560dSAlex Tomas /* 1714c9de560dSAlex Tomas * The routine scans buddy structures (not bitmap!) from given order 1715c9de560dSAlex Tomas * to max order and tries to find big enough chunk to satisfy the req 1716c9de560dSAlex Tomas */ 1717089ceeccSEric Sandeen static noinline_for_stack 1718089ceeccSEric Sandeen void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 1719c9de560dSAlex Tomas struct ext4_buddy *e4b) 1720c9de560dSAlex Tomas { 1721c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1722c9de560dSAlex Tomas struct ext4_group_info *grp = e4b->bd_info; 1723c9de560dSAlex Tomas void *buddy; 1724c9de560dSAlex Tomas int i; 1725c9de560dSAlex Tomas int k; 1726c9de560dSAlex Tomas int max; 1727c9de560dSAlex Tomas 1728c9de560dSAlex Tomas BUG_ON(ac->ac_2order <= 0); 1729c9de560dSAlex Tomas for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { 1730c9de560dSAlex Tomas if (grp->bb_counters[i] == 0) 1731c9de560dSAlex Tomas continue; 1732c9de560dSAlex Tomas 1733c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, i, &max); 1734c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1735c9de560dSAlex Tomas 1736ffad0a44SAneesh Kumar K.V k = mb_find_next_zero_bit(buddy, max, 0); 1737c9de560dSAlex Tomas BUG_ON(k >= max); 1738c9de560dSAlex Tomas 1739c9de560dSAlex Tomas ac->ac_found++; 1740c9de560dSAlex Tomas 1741c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 1 << i; 1742c9de560dSAlex Tomas ac->ac_b_ex.fe_start = k << i; 1743c9de560dSAlex Tomas ac->ac_b_ex.fe_group = e4b->bd_group; 1744c9de560dSAlex Tomas 1745c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1746c9de560dSAlex Tomas 1747c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); 1748c9de560dSAlex Tomas 1749c9de560dSAlex Tomas if (EXT4_SB(sb)->s_mb_stats) 1750c9de560dSAlex Tomas atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 1751c9de560dSAlex Tomas 1752c9de560dSAlex Tomas break; 1753c9de560dSAlex Tomas } 1754c9de560dSAlex Tomas } 1755c9de560dSAlex Tomas 1756c9de560dSAlex Tomas /* 1757c9de560dSAlex Tomas * The routine scans the group and measures all found extents. 1758c9de560dSAlex Tomas * In order to optimize scanning, caller must pass number of 1759c9de560dSAlex Tomas * free blocks in the group, so the routine can know upper limit. 1760c9de560dSAlex Tomas */ 1761089ceeccSEric Sandeen static noinline_for_stack 1762089ceeccSEric Sandeen void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 1763c9de560dSAlex Tomas struct ext4_buddy *e4b) 1764c9de560dSAlex Tomas { 1765c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1766c5e8f3f3STheodore Ts'o void *bitmap = e4b->bd_bitmap; 1767c9de560dSAlex Tomas struct ext4_free_extent ex; 1768c9de560dSAlex Tomas int i; 1769c9de560dSAlex Tomas int free; 1770c9de560dSAlex Tomas 1771c9de560dSAlex Tomas free = e4b->bd_info->bb_free; 1772c9de560dSAlex Tomas BUG_ON(free <= 0); 1773c9de560dSAlex Tomas 1774c9de560dSAlex Tomas i = e4b->bd_info->bb_first_free; 1775c9de560dSAlex Tomas 1776c9de560dSAlex Tomas while (free && ac->ac_status == AC_STATUS_CONTINUE) { 1777ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, 17787137d7a4STheodore Ts'o EXT4_CLUSTERS_PER_GROUP(sb), i); 17797137d7a4STheodore Ts'o if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) { 178026346ff6SAneesh Kumar K.V /* 1781e56eb659SAneesh Kumar K.V * IF we have corrupt bitmap, we won't find any 178226346ff6SAneesh Kumar K.V * free blocks even though group info says we 178326346ff6SAneesh Kumar K.V * we have free blocks 178426346ff6SAneesh Kumar K.V */ 1785e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 178653accfa9STheodore Ts'o "%d free clusters as per " 1787fde4d95aSTheodore Ts'o "group info. But bitmap says 0", 178826346ff6SAneesh Kumar K.V free); 1789c9de560dSAlex Tomas break; 1790c9de560dSAlex Tomas } 1791c9de560dSAlex Tomas 1792c9de560dSAlex Tomas mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); 1793c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 179426346ff6SAneesh Kumar K.V if (free < ex.fe_len) { 1795e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 179653accfa9STheodore Ts'o "%d free clusters as per " 1797fde4d95aSTheodore Ts'o "group info. But got %d blocks", 179826346ff6SAneesh Kumar K.V free, ex.fe_len); 1799e56eb659SAneesh Kumar K.V /* 1800e56eb659SAneesh Kumar K.V * The number of free blocks differs. This mostly 1801e56eb659SAneesh Kumar K.V * indicate that the bitmap is corrupt. So exit 1802e56eb659SAneesh Kumar K.V * without claiming the space. 1803e56eb659SAneesh Kumar K.V */ 1804e56eb659SAneesh Kumar K.V break; 180526346ff6SAneesh Kumar K.V } 1806c9de560dSAlex Tomas 1807c9de560dSAlex Tomas ext4_mb_measure_extent(ac, &ex, e4b); 1808c9de560dSAlex Tomas 1809c9de560dSAlex Tomas i += ex.fe_len; 1810c9de560dSAlex Tomas free -= ex.fe_len; 1811c9de560dSAlex Tomas } 1812c9de560dSAlex Tomas 1813c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 1); 1814c9de560dSAlex Tomas } 1815c9de560dSAlex Tomas 1816c9de560dSAlex Tomas /* 1817c9de560dSAlex Tomas * This is a special case for storages like raid5 1818506bf2d8SEric Sandeen * we try to find stripe-aligned chunks for stripe-size-multiple requests 1819c9de560dSAlex Tomas */ 1820089ceeccSEric Sandeen static noinline_for_stack 1821089ceeccSEric Sandeen void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 1822c9de560dSAlex Tomas struct ext4_buddy *e4b) 1823c9de560dSAlex Tomas { 1824c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1825c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 1826c5e8f3f3STheodore Ts'o void *bitmap = e4b->bd_bitmap; 1827c9de560dSAlex Tomas struct ext4_free_extent ex; 1828c9de560dSAlex Tomas ext4_fsblk_t first_group_block; 1829c9de560dSAlex Tomas ext4_fsblk_t a; 1830c9de560dSAlex Tomas ext4_grpblk_t i; 1831c9de560dSAlex Tomas int max; 1832c9de560dSAlex Tomas 1833c9de560dSAlex Tomas BUG_ON(sbi->s_stripe == 0); 1834c9de560dSAlex Tomas 1835c9de560dSAlex Tomas /* find first stripe-aligned block in group */ 18365661bd68SAkinobu Mita first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); 18375661bd68SAkinobu Mita 1838c9de560dSAlex Tomas a = first_group_block + sbi->s_stripe - 1; 1839c9de560dSAlex Tomas do_div(a, sbi->s_stripe); 1840c9de560dSAlex Tomas i = (a * sbi->s_stripe) - first_group_block; 1841c9de560dSAlex Tomas 18427137d7a4STheodore Ts'o while (i < EXT4_CLUSTERS_PER_GROUP(sb)) { 1843c9de560dSAlex Tomas if (!mb_test_bit(i, bitmap)) { 1844c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); 1845c9de560dSAlex Tomas if (max >= sbi->s_stripe) { 1846c9de560dSAlex Tomas ac->ac_found++; 1847c9de560dSAlex Tomas ac->ac_b_ex = ex; 1848c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1849c9de560dSAlex Tomas break; 1850c9de560dSAlex Tomas } 1851c9de560dSAlex Tomas } 1852c9de560dSAlex Tomas i += sbi->s_stripe; 1853c9de560dSAlex Tomas } 1854c9de560dSAlex Tomas } 1855c9de560dSAlex Tomas 18568a57d9d6SCurt Wohlgemuth /* This is now called BEFORE we load the buddy bitmap. */ 1857c9de560dSAlex Tomas static int ext4_mb_good_group(struct ext4_allocation_context *ac, 1858c9de560dSAlex Tomas ext4_group_t group, int cr) 1859c9de560dSAlex Tomas { 1860c9de560dSAlex Tomas unsigned free, fragments; 1861a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 1862c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 1863c9de560dSAlex Tomas 1864c9de560dSAlex Tomas BUG_ON(cr < 0 || cr >= 4); 18658a57d9d6SCurt Wohlgemuth 18668a57d9d6SCurt Wohlgemuth /* We only do this if the grp has never been initialized */ 18678a57d9d6SCurt Wohlgemuth if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 18688a57d9d6SCurt Wohlgemuth int ret = ext4_mb_init_group(ac->ac_sb, group); 18698a57d9d6SCurt Wohlgemuth if (ret) 18708a57d9d6SCurt Wohlgemuth return 0; 18718a57d9d6SCurt Wohlgemuth } 1872c9de560dSAlex Tomas 1873c9de560dSAlex Tomas free = grp->bb_free; 1874c9de560dSAlex Tomas fragments = grp->bb_fragments; 1875c9de560dSAlex Tomas if (free == 0) 1876c9de560dSAlex Tomas return 0; 1877c9de560dSAlex Tomas if (fragments == 0) 1878c9de560dSAlex Tomas return 0; 1879c9de560dSAlex Tomas 1880c9de560dSAlex Tomas switch (cr) { 1881c9de560dSAlex Tomas case 0: 1882c9de560dSAlex Tomas BUG_ON(ac->ac_2order == 0); 1883c9de560dSAlex Tomas 18848a57d9d6SCurt Wohlgemuth if (grp->bb_largest_free_order < ac->ac_2order) 18858a57d9d6SCurt Wohlgemuth return 0; 18868a57d9d6SCurt Wohlgemuth 1887a4912123STheodore Ts'o /* Avoid using the first bg of a flexgroup for data files */ 1888a4912123STheodore Ts'o if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 1889a4912123STheodore Ts'o (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 1890a4912123STheodore Ts'o ((group % flex_size) == 0)) 1891a4912123STheodore Ts'o return 0; 1892a4912123STheodore Ts'o 1893c9de560dSAlex Tomas return 1; 1894c9de560dSAlex Tomas case 1: 1895c9de560dSAlex Tomas if ((free / fragments) >= ac->ac_g_ex.fe_len) 1896c9de560dSAlex Tomas return 1; 1897c9de560dSAlex Tomas break; 1898c9de560dSAlex Tomas case 2: 1899c9de560dSAlex Tomas if (free >= ac->ac_g_ex.fe_len) 1900c9de560dSAlex Tomas return 1; 1901c9de560dSAlex Tomas break; 1902c9de560dSAlex Tomas case 3: 1903c9de560dSAlex Tomas return 1; 1904c9de560dSAlex Tomas default: 1905c9de560dSAlex Tomas BUG(); 1906c9de560dSAlex Tomas } 1907c9de560dSAlex Tomas 1908c9de560dSAlex Tomas return 0; 1909c9de560dSAlex Tomas } 1910c9de560dSAlex Tomas 19114ddfef7bSEric Sandeen static noinline_for_stack int 19124ddfef7bSEric Sandeen ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 1913c9de560dSAlex Tomas { 19148df9675fSTheodore Ts'o ext4_group_t ngroups, group, i; 1915c9de560dSAlex Tomas int cr; 1916c9de560dSAlex Tomas int err = 0; 1917c9de560dSAlex Tomas struct ext4_sb_info *sbi; 1918c9de560dSAlex Tomas struct super_block *sb; 1919c9de560dSAlex Tomas struct ext4_buddy e4b; 1920c9de560dSAlex Tomas 1921c9de560dSAlex Tomas sb = ac->ac_sb; 1922c9de560dSAlex Tomas sbi = EXT4_SB(sb); 19238df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 1924fb0a387dSEric Sandeen /* non-extent files are limited to low blocks/groups */ 192512e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) 1926fb0a387dSEric Sandeen ngroups = sbi->s_blockfile_groups; 1927fb0a387dSEric Sandeen 1928c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1929c9de560dSAlex Tomas 1930c9de560dSAlex Tomas /* first, try the goal */ 1931c9de560dSAlex Tomas err = ext4_mb_find_by_goal(ac, &e4b); 1932c9de560dSAlex Tomas if (err || ac->ac_status == AC_STATUS_FOUND) 1933c9de560dSAlex Tomas goto out; 1934c9de560dSAlex Tomas 1935c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 1936c9de560dSAlex Tomas goto out; 1937c9de560dSAlex Tomas 1938c9de560dSAlex Tomas /* 1939c9de560dSAlex Tomas * ac->ac2_order is set only if the fe_len is a power of 2 1940c9de560dSAlex Tomas * if ac2_order is set we also set criteria to 0 so that we 1941c9de560dSAlex Tomas * try exact allocation using buddy. 1942c9de560dSAlex Tomas */ 1943c9de560dSAlex Tomas i = fls(ac->ac_g_ex.fe_len); 1944c9de560dSAlex Tomas ac->ac_2order = 0; 1945c9de560dSAlex Tomas /* 1946c9de560dSAlex Tomas * We search using buddy data only if the order of the request 1947c9de560dSAlex Tomas * is greater than equal to the sbi_s_mb_order2_reqs 1948b713a5ecSTheodore Ts'o * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 1949c9de560dSAlex Tomas */ 1950c9de560dSAlex Tomas if (i >= sbi->s_mb_order2_reqs) { 1951c9de560dSAlex Tomas /* 1952c9de560dSAlex Tomas * This should tell if fe_len is exactly power of 2 1953c9de560dSAlex Tomas */ 1954c9de560dSAlex Tomas if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) 1955c9de560dSAlex Tomas ac->ac_2order = i - 1; 1956c9de560dSAlex Tomas } 1957c9de560dSAlex Tomas 19584ba74d00STheodore Ts'o /* if stream allocation is enabled, use global goal */ 19594ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 1960c9de560dSAlex Tomas /* TBD: may be hot point */ 1961c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 1962c9de560dSAlex Tomas ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 1963c9de560dSAlex Tomas ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 1964c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 1965c9de560dSAlex Tomas } 19664ba74d00STheodore Ts'o 1967c9de560dSAlex Tomas /* Let's just scan groups to find more-less suitable blocks */ 1968c9de560dSAlex Tomas cr = ac->ac_2order ? 0 : 1; 1969c9de560dSAlex Tomas /* 1970c9de560dSAlex Tomas * cr == 0 try to get exact allocation, 1971c9de560dSAlex Tomas * cr == 3 try to get anything 1972c9de560dSAlex Tomas */ 1973c9de560dSAlex Tomas repeat: 1974c9de560dSAlex Tomas for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 1975c9de560dSAlex Tomas ac->ac_criteria = cr; 1976ed8f9c75SAneesh Kumar K.V /* 1977ed8f9c75SAneesh Kumar K.V * searching for the right group start 1978ed8f9c75SAneesh Kumar K.V * from the goal value specified 1979ed8f9c75SAneesh Kumar K.V */ 1980ed8f9c75SAneesh Kumar K.V group = ac->ac_g_ex.fe_group; 1981ed8f9c75SAneesh Kumar K.V 19828df9675fSTheodore Ts'o for (i = 0; i < ngroups; group++, i++) { 19838df9675fSTheodore Ts'o if (group == ngroups) 1984c9de560dSAlex Tomas group = 0; 1985c9de560dSAlex Tomas 19868a57d9d6SCurt Wohlgemuth /* This now checks without needing the buddy page */ 19878a57d9d6SCurt Wohlgemuth if (!ext4_mb_good_group(ac, group, cr)) 1988c9de560dSAlex Tomas continue; 1989c9de560dSAlex Tomas 1990c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 1991c9de560dSAlex Tomas if (err) 1992c9de560dSAlex Tomas goto out; 1993c9de560dSAlex Tomas 1994c9de560dSAlex Tomas ext4_lock_group(sb, group); 19958a57d9d6SCurt Wohlgemuth 19968a57d9d6SCurt Wohlgemuth /* 19978a57d9d6SCurt Wohlgemuth * We need to check again after locking the 19988a57d9d6SCurt Wohlgemuth * block group 19998a57d9d6SCurt Wohlgemuth */ 2000c9de560dSAlex Tomas if (!ext4_mb_good_group(ac, group, cr)) { 2001c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2002e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2003c9de560dSAlex Tomas continue; 2004c9de560dSAlex Tomas } 2005c9de560dSAlex Tomas 2006c9de560dSAlex Tomas ac->ac_groups_scanned++; 200775507efbSTheodore Ts'o if (cr == 0) 2008c9de560dSAlex Tomas ext4_mb_simple_scan_group(ac, &e4b); 2009506bf2d8SEric Sandeen else if (cr == 1 && sbi->s_stripe && 2010506bf2d8SEric Sandeen !(ac->ac_g_ex.fe_len % sbi->s_stripe)) 2011c9de560dSAlex Tomas ext4_mb_scan_aligned(ac, &e4b); 2012c9de560dSAlex Tomas else 2013c9de560dSAlex Tomas ext4_mb_complex_scan_group(ac, &e4b); 2014c9de560dSAlex Tomas 2015c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2016e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2017c9de560dSAlex Tomas 2018c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_CONTINUE) 2019c9de560dSAlex Tomas break; 2020c9de560dSAlex Tomas } 2021c9de560dSAlex Tomas } 2022c9de560dSAlex Tomas 2023c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2024c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2025c9de560dSAlex Tomas /* 2026c9de560dSAlex Tomas * We've been searching too long. Let's try to allocate 2027c9de560dSAlex Tomas * the best chunk we've found so far 2028c9de560dSAlex Tomas */ 2029c9de560dSAlex Tomas 2030c9de560dSAlex Tomas ext4_mb_try_best_found(ac, &e4b); 2031c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_FOUND) { 2032c9de560dSAlex Tomas /* 2033c9de560dSAlex Tomas * Someone more lucky has already allocated it. 2034c9de560dSAlex Tomas * The only thing we can do is just take first 2035c9de560dSAlex Tomas * found block(s) 2036c9de560dSAlex Tomas printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); 2037c9de560dSAlex Tomas */ 2038c9de560dSAlex Tomas ac->ac_b_ex.fe_group = 0; 2039c9de560dSAlex Tomas ac->ac_b_ex.fe_start = 0; 2040c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 0; 2041c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 2042c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_FIRST; 2043c9de560dSAlex Tomas cr = 3; 2044c9de560dSAlex Tomas atomic_inc(&sbi->s_mb_lost_chunks); 2045c9de560dSAlex Tomas goto repeat; 2046c9de560dSAlex Tomas } 2047c9de560dSAlex Tomas } 2048c9de560dSAlex Tomas out: 2049c9de560dSAlex Tomas return err; 2050c9de560dSAlex Tomas } 2051c9de560dSAlex Tomas 2052c9de560dSAlex Tomas static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2053c9de560dSAlex Tomas { 2054c9de560dSAlex Tomas struct super_block *sb = seq->private; 2055c9de560dSAlex Tomas ext4_group_t group; 2056c9de560dSAlex Tomas 20578df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2058c9de560dSAlex Tomas return NULL; 2059c9de560dSAlex Tomas group = *pos + 1; 2060a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2061c9de560dSAlex Tomas } 2062c9de560dSAlex Tomas 2063c9de560dSAlex Tomas static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2064c9de560dSAlex Tomas { 2065c9de560dSAlex Tomas struct super_block *sb = seq->private; 2066c9de560dSAlex Tomas ext4_group_t group; 2067c9de560dSAlex Tomas 2068c9de560dSAlex Tomas ++*pos; 20698df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2070c9de560dSAlex Tomas return NULL; 2071c9de560dSAlex Tomas group = *pos + 1; 2072a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2073c9de560dSAlex Tomas } 2074c9de560dSAlex Tomas 2075c9de560dSAlex Tomas static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2076c9de560dSAlex Tomas { 2077c9de560dSAlex Tomas struct super_block *sb = seq->private; 2078a9df9a49STheodore Ts'o ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2079c9de560dSAlex Tomas int i; 2080c9de560dSAlex Tomas int err; 2081c9de560dSAlex Tomas struct ext4_buddy e4b; 2082c9de560dSAlex Tomas struct sg { 2083c9de560dSAlex Tomas struct ext4_group_info info; 2084a36b4498SEric Sandeen ext4_grpblk_t counters[16]; 2085c9de560dSAlex Tomas } sg; 2086c9de560dSAlex Tomas 2087c9de560dSAlex Tomas group--; 2088c9de560dSAlex Tomas if (group == 0) 2089c9de560dSAlex Tomas seq_printf(seq, "#%-5s: %-5s %-5s %-5s " 2090c9de560dSAlex Tomas "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " 2091c9de560dSAlex Tomas "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", 2092c9de560dSAlex Tomas "group", "free", "frags", "first", 2093c9de560dSAlex Tomas "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", 2094c9de560dSAlex Tomas "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); 2095c9de560dSAlex Tomas 2096c9de560dSAlex Tomas i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 2097c9de560dSAlex Tomas sizeof(struct ext4_group_info); 2098c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2099c9de560dSAlex Tomas if (err) { 2100a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: I/O error\n", group); 2101c9de560dSAlex Tomas return 0; 2102c9de560dSAlex Tomas } 2103c9de560dSAlex Tomas ext4_lock_group(sb, group); 2104c9de560dSAlex Tomas memcpy(&sg, ext4_get_group_info(sb, group), i); 2105c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2106e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2107c9de560dSAlex Tomas 2108a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 2109c9de560dSAlex Tomas sg.info.bb_fragments, sg.info.bb_first_free); 2110c9de560dSAlex Tomas for (i = 0; i <= 13; i++) 2111c9de560dSAlex Tomas seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? 2112c9de560dSAlex Tomas sg.info.bb_counters[i] : 0); 2113c9de560dSAlex Tomas seq_printf(seq, " ]\n"); 2114c9de560dSAlex Tomas 2115c9de560dSAlex Tomas return 0; 2116c9de560dSAlex Tomas } 2117c9de560dSAlex Tomas 2118c9de560dSAlex Tomas static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 2119c9de560dSAlex Tomas { 2120c9de560dSAlex Tomas } 2121c9de560dSAlex Tomas 21227f1346a9STobias Klauser static const struct seq_operations ext4_mb_seq_groups_ops = { 2123c9de560dSAlex Tomas .start = ext4_mb_seq_groups_start, 2124c9de560dSAlex Tomas .next = ext4_mb_seq_groups_next, 2125c9de560dSAlex Tomas .stop = ext4_mb_seq_groups_stop, 2126c9de560dSAlex Tomas .show = ext4_mb_seq_groups_show, 2127c9de560dSAlex Tomas }; 2128c9de560dSAlex Tomas 2129c9de560dSAlex Tomas static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) 2130c9de560dSAlex Tomas { 2131c9de560dSAlex Tomas struct super_block *sb = PDE(inode)->data; 2132c9de560dSAlex Tomas int rc; 2133c9de560dSAlex Tomas 2134c9de560dSAlex Tomas rc = seq_open(file, &ext4_mb_seq_groups_ops); 2135c9de560dSAlex Tomas if (rc == 0) { 2136a271fe85SJoe Perches struct seq_file *m = file->private_data; 2137c9de560dSAlex Tomas m->private = sb; 2138c9de560dSAlex Tomas } 2139c9de560dSAlex Tomas return rc; 2140c9de560dSAlex Tomas 2141c9de560dSAlex Tomas } 2142c9de560dSAlex Tomas 21437f1346a9STobias Klauser static const struct file_operations ext4_mb_seq_groups_fops = { 2144c9de560dSAlex Tomas .owner = THIS_MODULE, 2145c9de560dSAlex Tomas .open = ext4_mb_seq_groups_open, 2146c9de560dSAlex Tomas .read = seq_read, 2147c9de560dSAlex Tomas .llseek = seq_lseek, 2148c9de560dSAlex Tomas .release = seq_release, 2149c9de560dSAlex Tomas }; 2150c9de560dSAlex Tomas 2151fb1813f4SCurt Wohlgemuth static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) 2152fb1813f4SCurt Wohlgemuth { 2153fb1813f4SCurt Wohlgemuth int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 2154fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; 2155fb1813f4SCurt Wohlgemuth 2156fb1813f4SCurt Wohlgemuth BUG_ON(!cachep); 2157fb1813f4SCurt Wohlgemuth return cachep; 2158fb1813f4SCurt Wohlgemuth } 21595f21b0e6SFrederic Bohe 21605f21b0e6SFrederic Bohe /* Create and initialize ext4_group_info data for the given group. */ 2161920313a7SAneesh Kumar K.V int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 21625f21b0e6SFrederic Bohe struct ext4_group_desc *desc) 21635f21b0e6SFrederic Bohe { 2164fb1813f4SCurt Wohlgemuth int i; 21655f21b0e6SFrederic Bohe int metalen = 0; 21665f21b0e6SFrederic Bohe struct ext4_sb_info *sbi = EXT4_SB(sb); 21675f21b0e6SFrederic Bohe struct ext4_group_info **meta_group_info; 2168fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 21695f21b0e6SFrederic Bohe 21705f21b0e6SFrederic Bohe /* 21715f21b0e6SFrederic Bohe * First check if this group is the first of a reserved block. 21725f21b0e6SFrederic Bohe * If it's true, we have to allocate a new table of pointers 21735f21b0e6SFrederic Bohe * to ext4_group_info structures 21745f21b0e6SFrederic Bohe */ 21755f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 21765f21b0e6SFrederic Bohe metalen = sizeof(*meta_group_info) << 21775f21b0e6SFrederic Bohe EXT4_DESC_PER_BLOCK_BITS(sb); 21785f21b0e6SFrederic Bohe meta_group_info = kmalloc(metalen, GFP_KERNEL); 21795f21b0e6SFrederic Bohe if (meta_group_info == NULL) { 21807f6a11e7SJoe Perches ext4_msg(sb, KERN_ERR, "can't allocate mem " 21819d8b9ec4STheodore Ts'o "for a buddy group"); 21825f21b0e6SFrederic Bohe goto exit_meta_group_info; 21835f21b0e6SFrederic Bohe } 21845f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = 21855f21b0e6SFrederic Bohe meta_group_info; 21865f21b0e6SFrederic Bohe } 21875f21b0e6SFrederic Bohe 21885f21b0e6SFrederic Bohe meta_group_info = 21895f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; 21905f21b0e6SFrederic Bohe i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 21915f21b0e6SFrederic Bohe 2192fb1813f4SCurt Wohlgemuth meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL); 21935f21b0e6SFrederic Bohe if (meta_group_info[i] == NULL) { 21947f6a11e7SJoe Perches ext4_msg(sb, KERN_ERR, "can't allocate buddy mem"); 21955f21b0e6SFrederic Bohe goto exit_group_info; 21965f21b0e6SFrederic Bohe } 2197fb1813f4SCurt Wohlgemuth memset(meta_group_info[i], 0, kmem_cache_size(cachep)); 21985f21b0e6SFrederic Bohe set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 21995f21b0e6SFrederic Bohe &(meta_group_info[i]->bb_state)); 22005f21b0e6SFrederic Bohe 22015f21b0e6SFrederic Bohe /* 22025f21b0e6SFrederic Bohe * initialize bb_free to be able to skip 22035f21b0e6SFrederic Bohe * empty groups without initialization 22045f21b0e6SFrederic Bohe */ 22055f21b0e6SFrederic Bohe if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 22065f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 2207cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, group, desc); 22085f21b0e6SFrederic Bohe } else { 22095f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 2210021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, desc); 22115f21b0e6SFrederic Bohe } 22125f21b0e6SFrederic Bohe 22135f21b0e6SFrederic Bohe INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 2214920313a7SAneesh Kumar K.V init_rwsem(&meta_group_info[i]->alloc_sem); 221564e290ecSVenkatesh Pallipadi meta_group_info[i]->bb_free_root = RB_ROOT; 22168a57d9d6SCurt Wohlgemuth meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ 22175f21b0e6SFrederic Bohe 22185f21b0e6SFrederic Bohe #ifdef DOUBLE_CHECK 22195f21b0e6SFrederic Bohe { 22205f21b0e6SFrederic Bohe struct buffer_head *bh; 22215f21b0e6SFrederic Bohe meta_group_info[i]->bb_bitmap = 22225f21b0e6SFrederic Bohe kmalloc(sb->s_blocksize, GFP_KERNEL); 22235f21b0e6SFrederic Bohe BUG_ON(meta_group_info[i]->bb_bitmap == NULL); 22245f21b0e6SFrederic Bohe bh = ext4_read_block_bitmap(sb, group); 22255f21b0e6SFrederic Bohe BUG_ON(bh == NULL); 22265f21b0e6SFrederic Bohe memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, 22275f21b0e6SFrederic Bohe sb->s_blocksize); 22285f21b0e6SFrederic Bohe put_bh(bh); 22295f21b0e6SFrederic Bohe } 22305f21b0e6SFrederic Bohe #endif 22315f21b0e6SFrederic Bohe 22325f21b0e6SFrederic Bohe return 0; 22335f21b0e6SFrederic Bohe 22345f21b0e6SFrederic Bohe exit_group_info: 22355f21b0e6SFrederic Bohe /* If a meta_group_info table has been allocated, release it now */ 2236caaf7a29STao Ma if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 22375f21b0e6SFrederic Bohe kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); 2238caaf7a29STao Ma sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL; 2239caaf7a29STao Ma } 22405f21b0e6SFrederic Bohe exit_meta_group_info: 22415f21b0e6SFrederic Bohe return -ENOMEM; 22425f21b0e6SFrederic Bohe } /* ext4_mb_add_groupinfo */ 22435f21b0e6SFrederic Bohe 2244c9de560dSAlex Tomas static int ext4_mb_init_backend(struct super_block *sb) 2245c9de560dSAlex Tomas { 22468df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 2247c9de560dSAlex Tomas ext4_group_t i; 2248c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 22495f21b0e6SFrederic Bohe struct ext4_super_block *es = sbi->s_es; 22505f21b0e6SFrederic Bohe int num_meta_group_infos; 22515f21b0e6SFrederic Bohe int num_meta_group_infos_max; 22525f21b0e6SFrederic Bohe int array_size; 22535f21b0e6SFrederic Bohe struct ext4_group_desc *desc; 2254fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep; 2255c9de560dSAlex Tomas 22565f21b0e6SFrederic Bohe /* This is the number of blocks used by GDT */ 22578df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 22585f21b0e6SFrederic Bohe 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); 22595f21b0e6SFrederic Bohe 22605f21b0e6SFrederic Bohe /* 22615f21b0e6SFrederic Bohe * This is the total number of blocks used by GDT including 22625f21b0e6SFrederic Bohe * the number of reserved blocks for GDT. 22635f21b0e6SFrederic Bohe * The s_group_info array is allocated with this value 22645f21b0e6SFrederic Bohe * to allow a clean online resize without a complex 22655f21b0e6SFrederic Bohe * manipulation of pointer. 22665f21b0e6SFrederic Bohe * The drawback is the unused memory when no resize 22675f21b0e6SFrederic Bohe * occurs but it's very low in terms of pages 22685f21b0e6SFrederic Bohe * (see comments below) 22695f21b0e6SFrederic Bohe * Need to handle this properly when META_BG resizing is allowed 22705f21b0e6SFrederic Bohe */ 22715f21b0e6SFrederic Bohe num_meta_group_infos_max = num_meta_group_infos + 22725f21b0e6SFrederic Bohe le16_to_cpu(es->s_reserved_gdt_blocks); 22735f21b0e6SFrederic Bohe 22745f21b0e6SFrederic Bohe /* 22755f21b0e6SFrederic Bohe * array_size is the size of s_group_info array. We round it 22765f21b0e6SFrederic Bohe * to the next power of two because this approximation is done 22775f21b0e6SFrederic Bohe * internally by kmalloc so we can have some more memory 22785f21b0e6SFrederic Bohe * for free here (e.g. may be used for META_BG resize). 22795f21b0e6SFrederic Bohe */ 22805f21b0e6SFrederic Bohe array_size = 1; 22815f21b0e6SFrederic Bohe while (array_size < sizeof(*sbi->s_group_info) * 22825f21b0e6SFrederic Bohe num_meta_group_infos_max) 22835f21b0e6SFrederic Bohe array_size = array_size << 1; 2284c9de560dSAlex Tomas /* An 8TB filesystem with 64-bit pointers requires a 4096 byte 2285c9de560dSAlex Tomas * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. 2286c9de560dSAlex Tomas * So a two level scheme suffices for now. */ 2287f18a5f21STheodore Ts'o sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL); 2288c9de560dSAlex Tomas if (sbi->s_group_info == NULL) { 22899d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group"); 2290c9de560dSAlex Tomas return -ENOMEM; 2291c9de560dSAlex Tomas } 2292c9de560dSAlex Tomas sbi->s_buddy_cache = new_inode(sb); 2293c9de560dSAlex Tomas if (sbi->s_buddy_cache == NULL) { 22949d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't get new inode"); 2295c9de560dSAlex Tomas goto err_freesgi; 2296c9de560dSAlex Tomas } 229748e6061bSYu Jian /* To avoid potentially colliding with an valid on-disk inode number, 229848e6061bSYu Jian * use EXT4_BAD_INO for the buddy cache inode number. This inode is 229948e6061bSYu Jian * not in the inode hash, so it should never be found by iget(), but 230048e6061bSYu Jian * this will avoid confusion if it ever shows up during debugging. */ 230148e6061bSYu Jian sbi->s_buddy_cache->i_ino = EXT4_BAD_INO; 2302c9de560dSAlex Tomas EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 23038df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 2304c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, i, NULL); 2305c9de560dSAlex Tomas if (desc == NULL) { 23069d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i); 2307c9de560dSAlex Tomas goto err_freebuddy; 2308c9de560dSAlex Tomas } 23095f21b0e6SFrederic Bohe if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 23105f21b0e6SFrederic Bohe goto err_freebuddy; 2311c9de560dSAlex Tomas } 2312c9de560dSAlex Tomas 2313c9de560dSAlex Tomas return 0; 2314c9de560dSAlex Tomas 2315c9de560dSAlex Tomas err_freebuddy: 2316fb1813f4SCurt Wohlgemuth cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2317f1fa3342SRoel Kluin while (i-- > 0) 2318fb1813f4SCurt Wohlgemuth kmem_cache_free(cachep, ext4_get_group_info(sb, i)); 2319c9de560dSAlex Tomas i = num_meta_group_infos; 2320f1fa3342SRoel Kluin while (i-- > 0) 2321c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2322c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2323c9de560dSAlex Tomas err_freesgi: 2324f18a5f21STheodore Ts'o ext4_kvfree(sbi->s_group_info); 2325c9de560dSAlex Tomas return -ENOMEM; 2326c9de560dSAlex Tomas } 2327c9de560dSAlex Tomas 23282892c15dSEric Sandeen static void ext4_groupinfo_destroy_slabs(void) 23292892c15dSEric Sandeen { 23302892c15dSEric Sandeen int i; 23312892c15dSEric Sandeen 23322892c15dSEric Sandeen for (i = 0; i < NR_GRPINFO_CACHES; i++) { 23332892c15dSEric Sandeen if (ext4_groupinfo_caches[i]) 23342892c15dSEric Sandeen kmem_cache_destroy(ext4_groupinfo_caches[i]); 23352892c15dSEric Sandeen ext4_groupinfo_caches[i] = NULL; 23362892c15dSEric Sandeen } 23372892c15dSEric Sandeen } 23382892c15dSEric Sandeen 23392892c15dSEric Sandeen static int ext4_groupinfo_create_slab(size_t size) 23402892c15dSEric Sandeen { 23412892c15dSEric Sandeen static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); 23422892c15dSEric Sandeen int slab_size; 23432892c15dSEric Sandeen int blocksize_bits = order_base_2(size); 23442892c15dSEric Sandeen int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 23452892c15dSEric Sandeen struct kmem_cache *cachep; 23462892c15dSEric Sandeen 23472892c15dSEric Sandeen if (cache_index >= NR_GRPINFO_CACHES) 23482892c15dSEric Sandeen return -EINVAL; 23492892c15dSEric Sandeen 23502892c15dSEric Sandeen if (unlikely(cache_index < 0)) 23512892c15dSEric Sandeen cache_index = 0; 23522892c15dSEric Sandeen 23532892c15dSEric Sandeen mutex_lock(&ext4_grpinfo_slab_create_mutex); 23542892c15dSEric Sandeen if (ext4_groupinfo_caches[cache_index]) { 23552892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 23562892c15dSEric Sandeen return 0; /* Already created */ 23572892c15dSEric Sandeen } 23582892c15dSEric Sandeen 23592892c15dSEric Sandeen slab_size = offsetof(struct ext4_group_info, 23602892c15dSEric Sandeen bb_counters[blocksize_bits + 2]); 23612892c15dSEric Sandeen 23622892c15dSEric Sandeen cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], 23632892c15dSEric Sandeen slab_size, 0, SLAB_RECLAIM_ACCOUNT, 23642892c15dSEric Sandeen NULL); 23652892c15dSEric Sandeen 2366823ba01fSTao Ma ext4_groupinfo_caches[cache_index] = cachep; 2367823ba01fSTao Ma 23682892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 23692892c15dSEric Sandeen if (!cachep) { 23709d8b9ec4STheodore Ts'o printk(KERN_EMERG 23719d8b9ec4STheodore Ts'o "EXT4-fs: no memory for groupinfo slab cache\n"); 23722892c15dSEric Sandeen return -ENOMEM; 23732892c15dSEric Sandeen } 23742892c15dSEric Sandeen 23752892c15dSEric Sandeen return 0; 23762892c15dSEric Sandeen } 23772892c15dSEric Sandeen 2378c9de560dSAlex Tomas int ext4_mb_init(struct super_block *sb, int needs_recovery) 2379c9de560dSAlex Tomas { 2380c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 23816be2ded1SAneesh Kumar K.V unsigned i, j; 2382c9de560dSAlex Tomas unsigned offset; 2383c9de560dSAlex Tomas unsigned max; 238474767c5aSShen Feng int ret; 2385c9de560dSAlex Tomas 23861927805eSEric Sandeen i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); 2387c9de560dSAlex Tomas 2388c9de560dSAlex Tomas sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 2389c9de560dSAlex Tomas if (sbi->s_mb_offsets == NULL) { 2390fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 2391fb1813f4SCurt Wohlgemuth goto out; 2392c9de560dSAlex Tomas } 2393ff7ef329SYasunori Goto 23941927805eSEric Sandeen i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); 2395c9de560dSAlex Tomas sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 2396c9de560dSAlex Tomas if (sbi->s_mb_maxs == NULL) { 2397fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 2398fb1813f4SCurt Wohlgemuth goto out; 2399fb1813f4SCurt Wohlgemuth } 2400fb1813f4SCurt Wohlgemuth 24012892c15dSEric Sandeen ret = ext4_groupinfo_create_slab(sb->s_blocksize); 24022892c15dSEric Sandeen if (ret < 0) 2403fb1813f4SCurt Wohlgemuth goto out; 2404c9de560dSAlex Tomas 2405c9de560dSAlex Tomas /* order 0 is regular bitmap */ 2406c9de560dSAlex Tomas sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 2407c9de560dSAlex Tomas sbi->s_mb_offsets[0] = 0; 2408c9de560dSAlex Tomas 2409c9de560dSAlex Tomas i = 1; 2410c9de560dSAlex Tomas offset = 0; 2411c9de560dSAlex Tomas max = sb->s_blocksize << 2; 2412c9de560dSAlex Tomas do { 2413c9de560dSAlex Tomas sbi->s_mb_offsets[i] = offset; 2414c9de560dSAlex Tomas sbi->s_mb_maxs[i] = max; 2415c9de560dSAlex Tomas offset += 1 << (sb->s_blocksize_bits - i); 2416c9de560dSAlex Tomas max = max >> 1; 2417c9de560dSAlex Tomas i++; 2418c9de560dSAlex Tomas } while (i <= sb->s_blocksize_bits + 1); 2419c9de560dSAlex Tomas 2420c9de560dSAlex Tomas spin_lock_init(&sbi->s_md_lock); 2421c9de560dSAlex Tomas spin_lock_init(&sbi->s_bal_lock); 2422c9de560dSAlex Tomas 2423c9de560dSAlex Tomas sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 2424c9de560dSAlex Tomas sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 2425c9de560dSAlex Tomas sbi->s_mb_stats = MB_DEFAULT_STATS; 2426c9de560dSAlex Tomas sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 2427c9de560dSAlex Tomas sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 242827baebb8STheodore Ts'o /* 242927baebb8STheodore Ts'o * The default group preallocation is 512, which for 4k block 243027baebb8STheodore Ts'o * sizes translates to 2 megabytes. However for bigalloc file 243127baebb8STheodore Ts'o * systems, this is probably too big (i.e, if the cluster size 243227baebb8STheodore Ts'o * is 1 megabyte, then group preallocation size becomes half a 243327baebb8STheodore Ts'o * gigabyte!). As a default, we will keep a two megabyte 243427baebb8STheodore Ts'o * group pralloc size for cluster sizes up to 64k, and after 243527baebb8STheodore Ts'o * that, we will force a minimum group preallocation size of 243627baebb8STheodore Ts'o * 32 clusters. This translates to 8 megs when the cluster 243727baebb8STheodore Ts'o * size is 256k, and 32 megs when the cluster size is 1 meg, 243827baebb8STheodore Ts'o * which seems reasonable as a default. 243927baebb8STheodore Ts'o */ 244027baebb8STheodore Ts'o sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >> 244127baebb8STheodore Ts'o sbi->s_cluster_bits, 32); 2442d7a1fee1SDan Ehrenberg /* 2443d7a1fee1SDan Ehrenberg * If there is a s_stripe > 1, then we set the s_mb_group_prealloc 2444d7a1fee1SDan Ehrenberg * to the lowest multiple of s_stripe which is bigger than 2445d7a1fee1SDan Ehrenberg * the s_mb_group_prealloc as determined above. We want 2446d7a1fee1SDan Ehrenberg * the preallocation size to be an exact multiple of the 2447d7a1fee1SDan Ehrenberg * RAID stripe size so that preallocations don't fragment 2448d7a1fee1SDan Ehrenberg * the stripes. 2449d7a1fee1SDan Ehrenberg */ 2450d7a1fee1SDan Ehrenberg if (sbi->s_stripe > 1) { 2451d7a1fee1SDan Ehrenberg sbi->s_mb_group_prealloc = roundup( 2452d7a1fee1SDan Ehrenberg sbi->s_mb_group_prealloc, sbi->s_stripe); 2453d7a1fee1SDan Ehrenberg } 2454c9de560dSAlex Tomas 2455730c213cSEric Sandeen sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 2456c9de560dSAlex Tomas if (sbi->s_locality_groups == NULL) { 2457fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 24587aa0baeaSTao Ma goto out_free_groupinfo_slab; 2459c9de560dSAlex Tomas } 2460730c213cSEric Sandeen for_each_possible_cpu(i) { 2461c9de560dSAlex Tomas struct ext4_locality_group *lg; 2462730c213cSEric Sandeen lg = per_cpu_ptr(sbi->s_locality_groups, i); 2463c9de560dSAlex Tomas mutex_init(&lg->lg_mutex); 24646be2ded1SAneesh Kumar K.V for (j = 0; j < PREALLOC_TB_SIZE; j++) 24656be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 2466c9de560dSAlex Tomas spin_lock_init(&lg->lg_prealloc_lock); 2467c9de560dSAlex Tomas } 2468c9de560dSAlex Tomas 246979a77c5aSYu Jian /* init file for buddy data */ 247079a77c5aSYu Jian ret = ext4_mb_init_backend(sb); 24717aa0baeaSTao Ma if (ret != 0) 24727aa0baeaSTao Ma goto out_free_locality_groups; 247379a77c5aSYu Jian 2474296c355cSTheodore Ts'o if (sbi->s_proc) 2475296c355cSTheodore Ts'o proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, 2476296c355cSTheodore Ts'o &ext4_mb_seq_groups_fops, sb); 2477c9de560dSAlex Tomas 24787aa0baeaSTao Ma return 0; 24797aa0baeaSTao Ma 24807aa0baeaSTao Ma out_free_locality_groups: 24817aa0baeaSTao Ma free_percpu(sbi->s_locality_groups); 24827aa0baeaSTao Ma sbi->s_locality_groups = NULL; 24837aa0baeaSTao Ma out_free_groupinfo_slab: 24847aa0baeaSTao Ma ext4_groupinfo_destroy_slabs(); 2485fb1813f4SCurt Wohlgemuth out: 2486fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_offsets); 24877aa0baeaSTao Ma sbi->s_mb_offsets = NULL; 2488fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_maxs); 24897aa0baeaSTao Ma sbi->s_mb_maxs = NULL; 2490fb1813f4SCurt Wohlgemuth return ret; 2491c9de560dSAlex Tomas } 2492c9de560dSAlex Tomas 2493955ce5f5SAneesh Kumar K.V /* need to called with the ext4 group lock held */ 2494c9de560dSAlex Tomas static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) 2495c9de560dSAlex Tomas { 2496c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 2497c9de560dSAlex Tomas struct list_head *cur, *tmp; 2498c9de560dSAlex Tomas int count = 0; 2499c9de560dSAlex Tomas 2500c9de560dSAlex Tomas list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 2501c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 2502c9de560dSAlex Tomas list_del(&pa->pa_group_list); 2503c9de560dSAlex Tomas count++; 2504688f05a0SAneesh Kumar K.V kmem_cache_free(ext4_pspace_cachep, pa); 2505c9de560dSAlex Tomas } 2506c9de560dSAlex Tomas if (count) 25076ba495e9STheodore Ts'o mb_debug(1, "mballoc: %u PAs left\n", count); 2508c9de560dSAlex Tomas 2509c9de560dSAlex Tomas } 2510c9de560dSAlex Tomas 2511c9de560dSAlex Tomas int ext4_mb_release(struct super_block *sb) 2512c9de560dSAlex Tomas { 25138df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 2514c9de560dSAlex Tomas ext4_group_t i; 2515c9de560dSAlex Tomas int num_meta_group_infos; 2516c9de560dSAlex Tomas struct ext4_group_info *grinfo; 2517c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2518fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2519c9de560dSAlex Tomas 2520c9de560dSAlex Tomas if (sbi->s_group_info) { 25218df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 2522c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, i); 2523c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 2524c9de560dSAlex Tomas kfree(grinfo->bb_bitmap); 2525c9de560dSAlex Tomas #endif 2526c9de560dSAlex Tomas ext4_lock_group(sb, i); 2527c9de560dSAlex Tomas ext4_mb_cleanup_pa(grinfo); 2528c9de560dSAlex Tomas ext4_unlock_group(sb, i); 2529fb1813f4SCurt Wohlgemuth kmem_cache_free(cachep, grinfo); 2530c9de560dSAlex Tomas } 25318df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + 2532c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK(sb) - 1) >> 2533c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK_BITS(sb); 2534c9de560dSAlex Tomas for (i = 0; i < num_meta_group_infos; i++) 2535c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2536f18a5f21STheodore Ts'o ext4_kvfree(sbi->s_group_info); 2537c9de560dSAlex Tomas } 2538c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2539c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 2540c9de560dSAlex Tomas if (sbi->s_buddy_cache) 2541c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2542c9de560dSAlex Tomas if (sbi->s_mb_stats) { 25439d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 25449d8b9ec4STheodore Ts'o "mballoc: %u blocks %u reqs (%u success)", 2545c9de560dSAlex Tomas atomic_read(&sbi->s_bal_allocated), 2546c9de560dSAlex Tomas atomic_read(&sbi->s_bal_reqs), 2547c9de560dSAlex Tomas atomic_read(&sbi->s_bal_success)); 25489d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 25499d8b9ec4STheodore Ts'o "mballoc: %u extents scanned, %u goal hits, " 25509d8b9ec4STheodore Ts'o "%u 2^N hits, %u breaks, %u lost", 2551c9de560dSAlex Tomas atomic_read(&sbi->s_bal_ex_scanned), 2552c9de560dSAlex Tomas atomic_read(&sbi->s_bal_goals), 2553c9de560dSAlex Tomas atomic_read(&sbi->s_bal_2orders), 2554c9de560dSAlex Tomas atomic_read(&sbi->s_bal_breaks), 2555c9de560dSAlex Tomas atomic_read(&sbi->s_mb_lost_chunks)); 25569d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 25579d8b9ec4STheodore Ts'o "mballoc: %lu generated and it took %Lu", 2558ced156e4STao Ma sbi->s_mb_buddies_generated, 2559c9de560dSAlex Tomas sbi->s_mb_generation_time); 25609d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_INFO, 25619d8b9ec4STheodore Ts'o "mballoc: %u preallocated, %u discarded", 2562c9de560dSAlex Tomas atomic_read(&sbi->s_mb_preallocated), 2563c9de560dSAlex Tomas atomic_read(&sbi->s_mb_discarded)); 2564c9de560dSAlex Tomas } 2565c9de560dSAlex Tomas 2566730c213cSEric Sandeen free_percpu(sbi->s_locality_groups); 2567296c355cSTheodore Ts'o if (sbi->s_proc) 2568296c355cSTheodore Ts'o remove_proc_entry("mb_groups", sbi->s_proc); 2569c9de560dSAlex Tomas 2570c9de560dSAlex Tomas return 0; 2571c9de560dSAlex Tomas } 2572c9de560dSAlex Tomas 257377ca6cdfSLukas Czerner static inline int ext4_issue_discard(struct super_block *sb, 257484130193STheodore Ts'o ext4_group_t block_group, ext4_grpblk_t cluster, int count) 25755c521830SJiaying Zhang { 25765c521830SJiaying Zhang ext4_fsblk_t discard_block; 25775c521830SJiaying Zhang 257884130193STheodore Ts'o discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) + 257984130193STheodore Ts'o ext4_group_first_block_no(sb, block_group)); 258084130193STheodore Ts'o count = EXT4_C2B(EXT4_SB(sb), count); 25815c521830SJiaying Zhang trace_ext4_discard_blocks(sb, 25825c521830SJiaying Zhang (unsigned long long) discard_block, count); 258393259636SLukas Czerner return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); 25845c521830SJiaying Zhang } 25855c521830SJiaying Zhang 25863e624fc7STheodore Ts'o /* 25873e624fc7STheodore Ts'o * This function is called by the jbd2 layer once the commit has finished, 25883e624fc7STheodore Ts'o * so we know we can free the blocks that were released with that commit. 25893e624fc7STheodore Ts'o */ 259018aadd47SBobi Jam static void ext4_free_data_callback(struct super_block *sb, 259118aadd47SBobi Jam struct ext4_journal_cb_entry *jce, 259218aadd47SBobi Jam int rc) 2593c9de560dSAlex Tomas { 259418aadd47SBobi Jam struct ext4_free_data *entry = (struct ext4_free_data *)jce; 2595c9de560dSAlex Tomas struct ext4_buddy e4b; 2596c894058dSAneesh Kumar K.V struct ext4_group_info *db; 2597d9f34504STheodore Ts'o int err, count = 0, count2 = 0; 2598c9de560dSAlex Tomas 25996ba495e9STheodore Ts'o mb_debug(1, "gonna free %u blocks in group %u (0x%p):", 260018aadd47SBobi Jam entry->efd_count, entry->efd_group, entry); 2601c9de560dSAlex Tomas 2602d9f34504STheodore Ts'o if (test_opt(sb, DISCARD)) 260318aadd47SBobi Jam ext4_issue_discard(sb, entry->efd_group, 260418aadd47SBobi Jam entry->efd_start_cluster, entry->efd_count); 2605b90f6870STheodore Ts'o 260618aadd47SBobi Jam err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b); 2607c9de560dSAlex Tomas /* we expect to find existing buddy because it's pinned */ 2608c9de560dSAlex Tomas BUG_ON(err != 0); 2609c9de560dSAlex Tomas 261018aadd47SBobi Jam 2611c894058dSAneesh Kumar K.V db = e4b.bd_info; 2612c9de560dSAlex Tomas /* there are blocks to put in buddy to make them really free */ 261318aadd47SBobi Jam count += entry->efd_count; 2614c9de560dSAlex Tomas count2++; 261518aadd47SBobi Jam ext4_lock_group(sb, entry->efd_group); 2616c894058dSAneesh Kumar K.V /* Take it out of per group rb tree */ 261718aadd47SBobi Jam rb_erase(&entry->efd_node, &(db->bb_free_root)); 261818aadd47SBobi Jam mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count); 2619c9de560dSAlex Tomas 26203d56b8d2STao Ma /* 26213d56b8d2STao Ma * Clear the trimmed flag for the group so that the next 26223d56b8d2STao Ma * ext4_trim_fs can trim it. 26233d56b8d2STao Ma * If the volume is mounted with -o discard, online discard 26243d56b8d2STao Ma * is supported and the free blocks will be trimmed online. 26253d56b8d2STao Ma */ 26263d56b8d2STao Ma if (!test_opt(sb, DISCARD)) 26273d56b8d2STao Ma EXT4_MB_GRP_CLEAR_TRIMMED(db); 26283d56b8d2STao Ma 2629c894058dSAneesh Kumar K.V if (!db->bb_free_root.rb_node) { 2630c894058dSAneesh Kumar K.V /* No more items in the per group rb tree 2631c894058dSAneesh Kumar K.V * balance refcounts from ext4_mb_free_metadata() 2632c894058dSAneesh Kumar K.V */ 2633c9de560dSAlex Tomas page_cache_release(e4b.bd_buddy_page); 2634c9de560dSAlex Tomas page_cache_release(e4b.bd_bitmap_page); 2635c894058dSAneesh Kumar K.V } 263618aadd47SBobi Jam ext4_unlock_group(sb, entry->efd_group); 263718aadd47SBobi Jam kmem_cache_free(ext4_free_data_cachep, entry); 2638e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2639c9de560dSAlex Tomas 26406ba495e9STheodore Ts'o mb_debug(1, "freed %u blocks in %u structures\n", count, count2); 2641c9de560dSAlex Tomas } 2642c9de560dSAlex Tomas 26436ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 26446ba495e9STheodore Ts'o u8 mb_enable_debug __read_mostly; 26456ba495e9STheodore Ts'o 26466ba495e9STheodore Ts'o static struct dentry *debugfs_dir; 26476ba495e9STheodore Ts'o static struct dentry *debugfs_debug; 26486ba495e9STheodore Ts'o 26496ba495e9STheodore Ts'o static void __init ext4_create_debugfs_entry(void) 26506ba495e9STheodore Ts'o { 26516ba495e9STheodore Ts'o debugfs_dir = debugfs_create_dir("ext4", NULL); 26526ba495e9STheodore Ts'o if (debugfs_dir) 26536ba495e9STheodore Ts'o debugfs_debug = debugfs_create_u8("mballoc-debug", 26546ba495e9STheodore Ts'o S_IRUGO | S_IWUSR, 26556ba495e9STheodore Ts'o debugfs_dir, 26566ba495e9STheodore Ts'o &mb_enable_debug); 26576ba495e9STheodore Ts'o } 26586ba495e9STheodore Ts'o 26596ba495e9STheodore Ts'o static void ext4_remove_debugfs_entry(void) 26606ba495e9STheodore Ts'o { 26616ba495e9STheodore Ts'o debugfs_remove(debugfs_debug); 26626ba495e9STheodore Ts'o debugfs_remove(debugfs_dir); 26636ba495e9STheodore Ts'o } 26646ba495e9STheodore Ts'o 26656ba495e9STheodore Ts'o #else 26666ba495e9STheodore Ts'o 26676ba495e9STheodore Ts'o static void __init ext4_create_debugfs_entry(void) 26686ba495e9STheodore Ts'o { 26696ba495e9STheodore Ts'o } 26706ba495e9STheodore Ts'o 26716ba495e9STheodore Ts'o static void ext4_remove_debugfs_entry(void) 26726ba495e9STheodore Ts'o { 26736ba495e9STheodore Ts'o } 26746ba495e9STheodore Ts'o 26756ba495e9STheodore Ts'o #endif 26766ba495e9STheodore Ts'o 26775dabfc78STheodore Ts'o int __init ext4_init_mballoc(void) 2678c9de560dSAlex Tomas { 267916828088STheodore Ts'o ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, 268016828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 2681c9de560dSAlex Tomas if (ext4_pspace_cachep == NULL) 2682c9de560dSAlex Tomas return -ENOMEM; 2683c9de560dSAlex Tomas 268416828088STheodore Ts'o ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, 268516828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 2686256bdb49SEric Sandeen if (ext4_ac_cachep == NULL) { 2687256bdb49SEric Sandeen kmem_cache_destroy(ext4_pspace_cachep); 2688256bdb49SEric Sandeen return -ENOMEM; 2689256bdb49SEric Sandeen } 2690c894058dSAneesh Kumar K.V 269118aadd47SBobi Jam ext4_free_data_cachep = KMEM_CACHE(ext4_free_data, 269216828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 269318aadd47SBobi Jam if (ext4_free_data_cachep == NULL) { 2694c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_pspace_cachep); 2695c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_ac_cachep); 2696c894058dSAneesh Kumar K.V return -ENOMEM; 2697c894058dSAneesh Kumar K.V } 26986ba495e9STheodore Ts'o ext4_create_debugfs_entry(); 2699c9de560dSAlex Tomas return 0; 2700c9de560dSAlex Tomas } 2701c9de560dSAlex Tomas 27025dabfc78STheodore Ts'o void ext4_exit_mballoc(void) 2703c9de560dSAlex Tomas { 27043e03f9caSJesper Dangaard Brouer /* 27053e03f9caSJesper Dangaard Brouer * Wait for completion of call_rcu()'s on ext4_pspace_cachep 27063e03f9caSJesper Dangaard Brouer * before destroying the slab cache. 27073e03f9caSJesper Dangaard Brouer */ 27083e03f9caSJesper Dangaard Brouer rcu_barrier(); 2709c9de560dSAlex Tomas kmem_cache_destroy(ext4_pspace_cachep); 2710256bdb49SEric Sandeen kmem_cache_destroy(ext4_ac_cachep); 271118aadd47SBobi Jam kmem_cache_destroy(ext4_free_data_cachep); 27122892c15dSEric Sandeen ext4_groupinfo_destroy_slabs(); 27136ba495e9STheodore Ts'o ext4_remove_debugfs_entry(); 2714c9de560dSAlex Tomas } 2715c9de560dSAlex Tomas 2716c9de560dSAlex Tomas 2717c9de560dSAlex Tomas /* 271873b2c716SUwe Kleine-König * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps 2719c9de560dSAlex Tomas * Returns 0 if success or error code 2720c9de560dSAlex Tomas */ 27214ddfef7bSEric Sandeen static noinline_for_stack int 27224ddfef7bSEric Sandeen ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 272353accfa9STheodore Ts'o handle_t *handle, unsigned int reserv_clstrs) 2724c9de560dSAlex Tomas { 2725c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 2726c9de560dSAlex Tomas struct ext4_group_desc *gdp; 2727c9de560dSAlex Tomas struct buffer_head *gdp_bh; 2728c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2729c9de560dSAlex Tomas struct super_block *sb; 2730c9de560dSAlex Tomas ext4_fsblk_t block; 2731519deca0SAneesh Kumar K.V int err, len; 2732c9de560dSAlex Tomas 2733c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 2734c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len <= 0); 2735c9de560dSAlex Tomas 2736c9de560dSAlex Tomas sb = ac->ac_sb; 2737c9de560dSAlex Tomas sbi = EXT4_SB(sb); 2738c9de560dSAlex Tomas 2739c9de560dSAlex Tomas err = -EIO; 2740574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 2741c9de560dSAlex Tomas if (!bitmap_bh) 2742c9de560dSAlex Tomas goto out_err; 2743c9de560dSAlex Tomas 2744c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 2745c9de560dSAlex Tomas if (err) 2746c9de560dSAlex Tomas goto out_err; 2747c9de560dSAlex Tomas 2748c9de560dSAlex Tomas err = -EIO; 2749c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 2750c9de560dSAlex Tomas if (!gdp) 2751c9de560dSAlex Tomas goto out_err; 2752c9de560dSAlex Tomas 2753a9df9a49STheodore Ts'o ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 2754021b65bbSTheodore Ts'o ext4_free_group_clusters(sb, gdp)); 275503cddb80SAneesh Kumar K.V 2756c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gdp_bh); 2757c9de560dSAlex Tomas if (err) 2758c9de560dSAlex Tomas goto out_err; 2759c9de560dSAlex Tomas 2760bda00de7SAkinobu Mita block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 2761c9de560dSAlex Tomas 276253accfa9STheodore Ts'o len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 27636fd058f7STheodore Ts'o if (!ext4_data_block_valid(sbi, block, len)) { 276412062dddSEric Sandeen ext4_error(sb, "Allocating blocks %llu-%llu which overlap " 27651084f252STheodore Ts'o "fs metadata", block, block+len); 2766519deca0SAneesh Kumar K.V /* File system mounted not to panic on error 2767519deca0SAneesh Kumar K.V * Fix the bitmap and repeat the block allocation 2768519deca0SAneesh Kumar K.V * We leak some of the blocks here. 2769519deca0SAneesh Kumar K.V */ 2770955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2771c3e94d1dSYongqiang Yang ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 2772519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len); 2773955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 27740390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 2775519deca0SAneesh Kumar K.V if (!err) 2776519deca0SAneesh Kumar K.V err = -EAGAIN; 2777519deca0SAneesh Kumar K.V goto out_err; 2778c9de560dSAlex Tomas } 2779955ce5f5SAneesh Kumar K.V 2780955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2781c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 2782c9de560dSAlex Tomas { 2783c9de560dSAlex Tomas int i; 2784c9de560dSAlex Tomas for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 2785c9de560dSAlex Tomas BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 2786c9de560dSAlex Tomas bitmap_bh->b_data)); 2787c9de560dSAlex Tomas } 2788c9de560dSAlex Tomas } 2789c9de560dSAlex Tomas #endif 2790c3e94d1dSYongqiang Yang ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 2791c3e94d1dSYongqiang Yang ac->ac_b_ex.fe_len); 2792c9de560dSAlex Tomas if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 2793c9de560dSAlex Tomas gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 2794021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, 2795cff1dfd7STheodore Ts'o ext4_free_clusters_after_init(sb, 2796560671a0SAneesh Kumar K.V ac->ac_b_ex.fe_group, gdp)); 2797c9de560dSAlex Tomas } 2798021b65bbSTheodore Ts'o len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len; 2799021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, len); 2800c9de560dSAlex Tomas gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); 2801955ce5f5SAneesh Kumar K.V 2802955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 280357042651STheodore Ts'o percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len); 2804d2a17637SMingming Cao /* 28056bc6e63fSAneesh Kumar K.V * Now reduce the dirty block count also. Should not go negative 2806d2a17637SMingming Cao */ 28076bc6e63fSAneesh Kumar K.V if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 28086bc6e63fSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 280957042651STheodore Ts'o percpu_counter_sub(&sbi->s_dirtyclusters_counter, 281057042651STheodore Ts'o reserv_clstrs); 2811c9de560dSAlex Tomas 2812772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 2813772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, 2814772cb7c8SJose R. Santos ac->ac_b_ex.fe_group); 28159f24e420STheodore Ts'o atomic_sub(ac->ac_b_ex.fe_len, 281624aaa8efSTheodore Ts'o &sbi->s_flex_groups[flex_group].free_clusters); 2817772cb7c8SJose R. Santos } 2818772cb7c8SJose R. Santos 28190390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 2820c9de560dSAlex Tomas if (err) 2821c9de560dSAlex Tomas goto out_err; 28220390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 2823c9de560dSAlex Tomas 2824c9de560dSAlex Tomas out_err: 2825a0375156STheodore Ts'o ext4_mark_super_dirty(sb); 282642a10addSAneesh Kumar K.V brelse(bitmap_bh); 2827c9de560dSAlex Tomas return err; 2828c9de560dSAlex Tomas } 2829c9de560dSAlex Tomas 2830c9de560dSAlex Tomas /* 2831c9de560dSAlex Tomas * here we normalize request for locality group 2832d7a1fee1SDan Ehrenberg * Group request are normalized to s_mb_group_prealloc, which goes to 2833d7a1fee1SDan Ehrenberg * s_strip if we set the same via mount option. 2834d7a1fee1SDan Ehrenberg * s_mb_group_prealloc can be configured via 2835b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_group_prealloc 2836c9de560dSAlex Tomas * 2837c9de560dSAlex Tomas * XXX: should we try to preallocate more than the group has now? 2838c9de560dSAlex Tomas */ 2839c9de560dSAlex Tomas static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 2840c9de560dSAlex Tomas { 2841c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2842c9de560dSAlex Tomas struct ext4_locality_group *lg = ac->ac_lg; 2843c9de560dSAlex Tomas 2844c9de560dSAlex Tomas BUG_ON(lg == NULL); 2845c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 28466ba495e9STheodore Ts'o mb_debug(1, "#%u: goal %u blocks for locality group\n", 2847c9de560dSAlex Tomas current->pid, ac->ac_g_ex.fe_len); 2848c9de560dSAlex Tomas } 2849c9de560dSAlex Tomas 2850c9de560dSAlex Tomas /* 2851c9de560dSAlex Tomas * Normalization means making request better in terms of 2852c9de560dSAlex Tomas * size and alignment 2853c9de560dSAlex Tomas */ 28544ddfef7bSEric Sandeen static noinline_for_stack void 28554ddfef7bSEric Sandeen ext4_mb_normalize_request(struct ext4_allocation_context *ac, 2856c9de560dSAlex Tomas struct ext4_allocation_request *ar) 2857c9de560dSAlex Tomas { 285853accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2859c9de560dSAlex Tomas int bsbits, max; 2860c9de560dSAlex Tomas ext4_lblk_t end; 28611592d2c5SCurt Wohlgemuth loff_t size, start_off; 28621592d2c5SCurt Wohlgemuth loff_t orig_size __maybe_unused; 28635a0790c2SAndi Kleen ext4_lblk_t start; 2864c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 28659a0762c5SAneesh Kumar K.V struct ext4_prealloc_space *pa; 2866c9de560dSAlex Tomas 2867c9de560dSAlex Tomas /* do normalize only data requests, metadata requests 2868c9de560dSAlex Tomas do not need preallocation */ 2869c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 2870c9de560dSAlex Tomas return; 2871c9de560dSAlex Tomas 2872c9de560dSAlex Tomas /* sometime caller may want exact blocks */ 2873c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2874c9de560dSAlex Tomas return; 2875c9de560dSAlex Tomas 2876c9de560dSAlex Tomas /* caller may indicate that preallocation isn't 2877c9de560dSAlex Tomas * required (it's a tail, for example) */ 2878c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 2879c9de560dSAlex Tomas return; 2880c9de560dSAlex Tomas 2881c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 2882c9de560dSAlex Tomas ext4_mb_normalize_group_request(ac); 2883c9de560dSAlex Tomas return ; 2884c9de560dSAlex Tomas } 2885c9de560dSAlex Tomas 2886c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 2887c9de560dSAlex Tomas 2888c9de560dSAlex Tomas /* first, let's learn actual file size 2889c9de560dSAlex Tomas * given current request is allocated */ 289053accfa9STheodore Ts'o size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); 2891c9de560dSAlex Tomas size = size << bsbits; 2892c9de560dSAlex Tomas if (size < i_size_read(ac->ac_inode)) 2893c9de560dSAlex Tomas size = i_size_read(ac->ac_inode); 28945a0790c2SAndi Kleen orig_size = size; 2895c9de560dSAlex Tomas 28961930479cSValerie Clement /* max size of free chunks */ 28971930479cSValerie Clement max = 2 << bsbits; 2898c9de560dSAlex Tomas 28991930479cSValerie Clement #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 29001930479cSValerie Clement (req <= (size) || max <= (chunk_size)) 2901c9de560dSAlex Tomas 2902c9de560dSAlex Tomas /* first, try to predict filesize */ 2903c9de560dSAlex Tomas /* XXX: should this table be tunable? */ 2904c9de560dSAlex Tomas start_off = 0; 2905c9de560dSAlex Tomas if (size <= 16 * 1024) { 2906c9de560dSAlex Tomas size = 16 * 1024; 2907c9de560dSAlex Tomas } else if (size <= 32 * 1024) { 2908c9de560dSAlex Tomas size = 32 * 1024; 2909c9de560dSAlex Tomas } else if (size <= 64 * 1024) { 2910c9de560dSAlex Tomas size = 64 * 1024; 2911c9de560dSAlex Tomas } else if (size <= 128 * 1024) { 2912c9de560dSAlex Tomas size = 128 * 1024; 2913c9de560dSAlex Tomas } else if (size <= 256 * 1024) { 2914c9de560dSAlex Tomas size = 256 * 1024; 2915c9de560dSAlex Tomas } else if (size <= 512 * 1024) { 2916c9de560dSAlex Tomas size = 512 * 1024; 2917c9de560dSAlex Tomas } else if (size <= 1024 * 1024) { 2918c9de560dSAlex Tomas size = 1024 * 1024; 29191930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 2920c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 29211930479cSValerie Clement (21 - bsbits)) << 21; 29221930479cSValerie Clement size = 2 * 1024 * 1024; 29231930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 2924c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 2925c9de560dSAlex Tomas (22 - bsbits)) << 22; 2926c9de560dSAlex Tomas size = 4 * 1024 * 1024; 2927c9de560dSAlex Tomas } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, 29281930479cSValerie Clement (8<<20)>>bsbits, max, 8 * 1024)) { 2929c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 2930c9de560dSAlex Tomas (23 - bsbits)) << 23; 2931c9de560dSAlex Tomas size = 8 * 1024 * 1024; 2932c9de560dSAlex Tomas } else { 2933c9de560dSAlex Tomas start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; 2934c9de560dSAlex Tomas size = ac->ac_o_ex.fe_len << bsbits; 2935c9de560dSAlex Tomas } 29365a0790c2SAndi Kleen size = size >> bsbits; 29375a0790c2SAndi Kleen start = start_off >> bsbits; 2938c9de560dSAlex Tomas 2939c9de560dSAlex Tomas /* don't cover already allocated blocks in selected range */ 2940c9de560dSAlex Tomas if (ar->pleft && start <= ar->lleft) { 2941c9de560dSAlex Tomas size -= ar->lleft + 1 - start; 2942c9de560dSAlex Tomas start = ar->lleft + 1; 2943c9de560dSAlex Tomas } 2944c9de560dSAlex Tomas if (ar->pright && start + size - 1 >= ar->lright) 2945c9de560dSAlex Tomas size -= start + size - ar->lright; 2946c9de560dSAlex Tomas 2947c9de560dSAlex Tomas end = start + size; 2948c9de560dSAlex Tomas 2949c9de560dSAlex Tomas /* check we don't cross already preallocated blocks */ 2950c9de560dSAlex Tomas rcu_read_lock(); 29519a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 2952498e5f24STheodore Ts'o ext4_lblk_t pa_end; 2953c9de560dSAlex Tomas 2954c9de560dSAlex Tomas if (pa->pa_deleted) 2955c9de560dSAlex Tomas continue; 2956c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 2957c9de560dSAlex Tomas if (pa->pa_deleted) { 2958c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 2959c9de560dSAlex Tomas continue; 2960c9de560dSAlex Tomas } 2961c9de560dSAlex Tomas 296253accfa9STheodore Ts'o pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), 296353accfa9STheodore Ts'o pa->pa_len); 2964c9de560dSAlex Tomas 2965c9de560dSAlex Tomas /* PA must not overlap original request */ 2966c9de560dSAlex Tomas BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || 2967c9de560dSAlex Tomas ac->ac_o_ex.fe_logical < pa->pa_lstart)); 2968c9de560dSAlex Tomas 296938877f4eSEric Sandeen /* skip PAs this normalized request doesn't overlap with */ 297038877f4eSEric Sandeen if (pa->pa_lstart >= end || pa_end <= start) { 2971c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 2972c9de560dSAlex Tomas continue; 2973c9de560dSAlex Tomas } 2974c9de560dSAlex Tomas BUG_ON(pa->pa_lstart <= start && pa_end >= end); 2975c9de560dSAlex Tomas 297638877f4eSEric Sandeen /* adjust start or end to be adjacent to this pa */ 2977c9de560dSAlex Tomas if (pa_end <= ac->ac_o_ex.fe_logical) { 2978c9de560dSAlex Tomas BUG_ON(pa_end < start); 2979c9de560dSAlex Tomas start = pa_end; 298038877f4eSEric Sandeen } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { 2981c9de560dSAlex Tomas BUG_ON(pa->pa_lstart > end); 2982c9de560dSAlex Tomas end = pa->pa_lstart; 2983c9de560dSAlex Tomas } 2984c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 2985c9de560dSAlex Tomas } 2986c9de560dSAlex Tomas rcu_read_unlock(); 2987c9de560dSAlex Tomas size = end - start; 2988c9de560dSAlex Tomas 2989c9de560dSAlex Tomas /* XXX: extra loop to check we really don't overlap preallocations */ 2990c9de560dSAlex Tomas rcu_read_lock(); 29919a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 2992498e5f24STheodore Ts'o ext4_lblk_t pa_end; 299353accfa9STheodore Ts'o 2994c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 2995c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 299653accfa9STheodore Ts'o pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb), 299753accfa9STheodore Ts'o pa->pa_len); 2998c9de560dSAlex Tomas BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); 2999c9de560dSAlex Tomas } 3000c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3001c9de560dSAlex Tomas } 3002c9de560dSAlex Tomas rcu_read_unlock(); 3003c9de560dSAlex Tomas 3004c9de560dSAlex Tomas if (start + size <= ac->ac_o_ex.fe_logical && 3005c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical) { 30069d8b9ec4STheodore Ts'o ext4_msg(ac->ac_sb, KERN_ERR, 30079d8b9ec4STheodore Ts'o "start %lu, size %lu, fe_logical %lu", 3008c9de560dSAlex Tomas (unsigned long) start, (unsigned long) size, 3009c9de560dSAlex Tomas (unsigned long) ac->ac_o_ex.fe_logical); 3010c9de560dSAlex Tomas } 3011c9de560dSAlex Tomas BUG_ON(start + size <= ac->ac_o_ex.fe_logical && 3012c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical); 30137137d7a4STheodore Ts'o BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb)); 3014c9de560dSAlex Tomas 3015c9de560dSAlex Tomas /* now prepare goal request */ 3016c9de560dSAlex Tomas 3017c9de560dSAlex Tomas /* XXX: is it better to align blocks WRT to logical 3018c9de560dSAlex Tomas * placement or satisfy big request as is */ 3019c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = start; 302053accfa9STheodore Ts'o ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size); 3021c9de560dSAlex Tomas 3022c9de560dSAlex Tomas /* define goal start in order to merge */ 3023c9de560dSAlex Tomas if (ar->pright && (ar->lright == (start + size))) { 3024c9de560dSAlex Tomas /* merge to the right */ 3025c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 3026c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 3027c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 3028c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3029c9de560dSAlex Tomas } 3030c9de560dSAlex Tomas if (ar->pleft && (ar->lleft + 1 == start)) { 3031c9de560dSAlex Tomas /* merge to the left */ 3032c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 3033c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 3034c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 3035c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3036c9de560dSAlex Tomas } 3037c9de560dSAlex Tomas 30386ba495e9STheodore Ts'o mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, 3039c9de560dSAlex Tomas (unsigned) orig_size, (unsigned) start); 3040c9de560dSAlex Tomas } 3041c9de560dSAlex Tomas 3042c9de560dSAlex Tomas static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 3043c9de560dSAlex Tomas { 3044c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3045c9de560dSAlex Tomas 3046c9de560dSAlex Tomas if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { 3047c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_reqs); 3048c9de560dSAlex Tomas atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 3049291dae47SCurt Wohlgemuth if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) 3050c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_success); 3051c9de560dSAlex Tomas atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 3052c9de560dSAlex Tomas if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 3053c9de560dSAlex Tomas ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 3054c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_goals); 3055c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan) 3056c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_breaks); 3057c9de560dSAlex Tomas } 3058c9de560dSAlex Tomas 3059296c355cSTheodore Ts'o if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 3060296c355cSTheodore Ts'o trace_ext4_mballoc_alloc(ac); 3061296c355cSTheodore Ts'o else 3062296c355cSTheodore Ts'o trace_ext4_mballoc_prealloc(ac); 3063c9de560dSAlex Tomas } 3064c9de560dSAlex Tomas 3065c9de560dSAlex Tomas /* 3066b844167eSCurt Wohlgemuth * Called on failure; free up any blocks from the inode PA for this 3067b844167eSCurt Wohlgemuth * context. We don't need this for MB_GROUP_PA because we only change 3068b844167eSCurt Wohlgemuth * pa_free in ext4_mb_release_context(), but on failure, we've already 3069b844167eSCurt Wohlgemuth * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. 3070b844167eSCurt Wohlgemuth */ 3071b844167eSCurt Wohlgemuth static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) 3072b844167eSCurt Wohlgemuth { 3073b844167eSCurt Wohlgemuth struct ext4_prealloc_space *pa = ac->ac_pa; 3074b844167eSCurt Wohlgemuth int len; 3075b844167eSCurt Wohlgemuth 3076b844167eSCurt Wohlgemuth if (pa && pa->pa_type == MB_INODE_PA) { 3077b844167eSCurt Wohlgemuth len = ac->ac_b_ex.fe_len; 3078b844167eSCurt Wohlgemuth pa->pa_free += len; 3079b844167eSCurt Wohlgemuth } 3080b844167eSCurt Wohlgemuth 3081b844167eSCurt Wohlgemuth } 3082b844167eSCurt Wohlgemuth 3083b844167eSCurt Wohlgemuth /* 3084c9de560dSAlex Tomas * use blocks preallocated to inode 3085c9de560dSAlex Tomas */ 3086c9de560dSAlex Tomas static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 3087c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3088c9de560dSAlex Tomas { 308953accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3090c9de560dSAlex Tomas ext4_fsblk_t start; 3091c9de560dSAlex Tomas ext4_fsblk_t end; 3092c9de560dSAlex Tomas int len; 3093c9de560dSAlex Tomas 3094c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3095c9de560dSAlex Tomas start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 309653accfa9STheodore Ts'o end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len), 309753accfa9STheodore Ts'o start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len)); 309853accfa9STheodore Ts'o len = EXT4_NUM_B2C(sbi, end - start); 3099c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 3100c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3101c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3102c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3103c9de560dSAlex Tomas ac->ac_pa = pa; 3104c9de560dSAlex Tomas 3105c9de560dSAlex Tomas BUG_ON(start < pa->pa_pstart); 310653accfa9STheodore Ts'o BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len)); 3107c9de560dSAlex Tomas BUG_ON(pa->pa_free < len); 3108c9de560dSAlex Tomas pa->pa_free -= len; 3109c9de560dSAlex Tomas 31106ba495e9STheodore Ts'o mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); 3111c9de560dSAlex Tomas } 3112c9de560dSAlex Tomas 3113c9de560dSAlex Tomas /* 3114c9de560dSAlex Tomas * use blocks preallocated to locality group 3115c9de560dSAlex Tomas */ 3116c9de560dSAlex Tomas static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 3117c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3118c9de560dSAlex Tomas { 311903cddb80SAneesh Kumar K.V unsigned int len = ac->ac_o_ex.fe_len; 31206be2ded1SAneesh Kumar K.V 3121c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 3122c9de560dSAlex Tomas &ac->ac_b_ex.fe_group, 3123c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3124c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3125c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3126c9de560dSAlex Tomas ac->ac_pa = pa; 3127c9de560dSAlex Tomas 3128c9de560dSAlex Tomas /* we don't correct pa_pstart or pa_plen here to avoid 312926346ff6SAneesh Kumar K.V * possible race when the group is being loaded concurrently 3130c9de560dSAlex Tomas * instead we correct pa later, after blocks are marked 313126346ff6SAneesh Kumar K.V * in on-disk bitmap -- see ext4_mb_release_context() 313226346ff6SAneesh Kumar K.V * Other CPUs are prevented from allocating from this pa by lg_mutex 3133c9de560dSAlex Tomas */ 31346ba495e9STheodore Ts'o mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); 3135c9de560dSAlex Tomas } 3136c9de560dSAlex Tomas 3137c9de560dSAlex Tomas /* 31385e745b04SAneesh Kumar K.V * Return the prealloc space that have minimal distance 31395e745b04SAneesh Kumar K.V * from the goal block. @cpa is the prealloc 31405e745b04SAneesh Kumar K.V * space that is having currently known minimal distance 31415e745b04SAneesh Kumar K.V * from the goal block. 31425e745b04SAneesh Kumar K.V */ 31435e745b04SAneesh Kumar K.V static struct ext4_prealloc_space * 31445e745b04SAneesh Kumar K.V ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 31455e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, 31465e745b04SAneesh Kumar K.V struct ext4_prealloc_space *cpa) 31475e745b04SAneesh Kumar K.V { 31485e745b04SAneesh Kumar K.V ext4_fsblk_t cur_distance, new_distance; 31495e745b04SAneesh Kumar K.V 31505e745b04SAneesh Kumar K.V if (cpa == NULL) { 31515e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 31525e745b04SAneesh Kumar K.V return pa; 31535e745b04SAneesh Kumar K.V } 31545e745b04SAneesh Kumar K.V cur_distance = abs(goal_block - cpa->pa_pstart); 31555e745b04SAneesh Kumar K.V new_distance = abs(goal_block - pa->pa_pstart); 31565e745b04SAneesh Kumar K.V 31575a54b2f1SColy Li if (cur_distance <= new_distance) 31585e745b04SAneesh Kumar K.V return cpa; 31595e745b04SAneesh Kumar K.V 31605e745b04SAneesh Kumar K.V /* drop the previous reference */ 31615e745b04SAneesh Kumar K.V atomic_dec(&cpa->pa_count); 31625e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 31635e745b04SAneesh Kumar K.V return pa; 31645e745b04SAneesh Kumar K.V } 31655e745b04SAneesh Kumar K.V 31665e745b04SAneesh Kumar K.V /* 3167c9de560dSAlex Tomas * search goal blocks in preallocated space 3168c9de560dSAlex Tomas */ 31694ddfef7bSEric Sandeen static noinline_for_stack int 31704ddfef7bSEric Sandeen ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 3171c9de560dSAlex Tomas { 317253accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 31736be2ded1SAneesh Kumar K.V int order, i; 3174c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3175c9de560dSAlex Tomas struct ext4_locality_group *lg; 31765e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, *cpa = NULL; 31775e745b04SAneesh Kumar K.V ext4_fsblk_t goal_block; 3178c9de560dSAlex Tomas 3179c9de560dSAlex Tomas /* only data can be preallocated */ 3180c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3181c9de560dSAlex Tomas return 0; 3182c9de560dSAlex Tomas 3183c9de560dSAlex Tomas /* first, try per-file preallocation */ 3184c9de560dSAlex Tomas rcu_read_lock(); 31859a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3186c9de560dSAlex Tomas 3187c9de560dSAlex Tomas /* all fields in this condition don't change, 3188c9de560dSAlex Tomas * so we can skip locking for them */ 3189c9de560dSAlex Tomas if (ac->ac_o_ex.fe_logical < pa->pa_lstart || 319053accfa9STheodore Ts'o ac->ac_o_ex.fe_logical >= (pa->pa_lstart + 319153accfa9STheodore Ts'o EXT4_C2B(sbi, pa->pa_len))) 3192c9de560dSAlex Tomas continue; 3193c9de560dSAlex Tomas 3194fb0a387dSEric Sandeen /* non-extent files can't have physical blocks past 2^32 */ 319512e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && 319653accfa9STheodore Ts'o (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) > 319753accfa9STheodore Ts'o EXT4_MAX_BLOCK_FILE_PHYS)) 3198fb0a387dSEric Sandeen continue; 3199fb0a387dSEric Sandeen 3200c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3201c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3202c9de560dSAlex Tomas if (pa->pa_deleted == 0 && pa->pa_free) { 3203c9de560dSAlex Tomas atomic_inc(&pa->pa_count); 3204c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 3205c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3206c9de560dSAlex Tomas ac->ac_criteria = 10; 3207c9de560dSAlex Tomas rcu_read_unlock(); 3208c9de560dSAlex Tomas return 1; 3209c9de560dSAlex Tomas } 3210c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3211c9de560dSAlex Tomas } 3212c9de560dSAlex Tomas rcu_read_unlock(); 3213c9de560dSAlex Tomas 3214c9de560dSAlex Tomas /* can we use group allocation? */ 3215c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 3216c9de560dSAlex Tomas return 0; 3217c9de560dSAlex Tomas 3218c9de560dSAlex Tomas /* inode may have no locality group for some reason */ 3219c9de560dSAlex Tomas lg = ac->ac_lg; 3220c9de560dSAlex Tomas if (lg == NULL) 3221c9de560dSAlex Tomas return 0; 32226be2ded1SAneesh Kumar K.V order = fls(ac->ac_o_ex.fe_len) - 1; 32236be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 32246be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 32256be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 3226c9de560dSAlex Tomas 3227bda00de7SAkinobu Mita goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); 32285e745b04SAneesh Kumar K.V /* 32295e745b04SAneesh Kumar K.V * search for the prealloc space that is having 32305e745b04SAneesh Kumar K.V * minimal distance from the goal block. 32315e745b04SAneesh Kumar K.V */ 32326be2ded1SAneesh Kumar K.V for (i = order; i < PREALLOC_TB_SIZE; i++) { 3233c9de560dSAlex Tomas rcu_read_lock(); 32346be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], 32356be2ded1SAneesh Kumar K.V pa_inode_list) { 3236c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 32376be2ded1SAneesh Kumar K.V if (pa->pa_deleted == 0 && 32386be2ded1SAneesh Kumar K.V pa->pa_free >= ac->ac_o_ex.fe_len) { 32395e745b04SAneesh Kumar K.V 32405e745b04SAneesh Kumar K.V cpa = ext4_mb_check_group_pa(goal_block, 32415e745b04SAneesh Kumar K.V pa, cpa); 32425e745b04SAneesh Kumar K.V } 3243c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 32445e745b04SAneesh Kumar K.V } 32455e745b04SAneesh Kumar K.V rcu_read_unlock(); 32465e745b04SAneesh Kumar K.V } 32475e745b04SAneesh Kumar K.V if (cpa) { 32485e745b04SAneesh Kumar K.V ext4_mb_use_group_pa(ac, cpa); 3249c9de560dSAlex Tomas ac->ac_criteria = 20; 3250c9de560dSAlex Tomas return 1; 3251c9de560dSAlex Tomas } 3252c9de560dSAlex Tomas return 0; 3253c9de560dSAlex Tomas } 3254c9de560dSAlex Tomas 3255c9de560dSAlex Tomas /* 32567a2fcbf7SAneesh Kumar K.V * the function goes through all block freed in the group 32577a2fcbf7SAneesh Kumar K.V * but not yet committed and marks them used in in-core bitmap. 32587a2fcbf7SAneesh Kumar K.V * buddy must be generated from this bitmap 3259955ce5f5SAneesh Kumar K.V * Need to be called with the ext4 group lock held 32607a2fcbf7SAneesh Kumar K.V */ 32617a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 32627a2fcbf7SAneesh Kumar K.V ext4_group_t group) 32637a2fcbf7SAneesh Kumar K.V { 32647a2fcbf7SAneesh Kumar K.V struct rb_node *n; 32657a2fcbf7SAneesh Kumar K.V struct ext4_group_info *grp; 32667a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 32677a2fcbf7SAneesh Kumar K.V 32687a2fcbf7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 32697a2fcbf7SAneesh Kumar K.V n = rb_first(&(grp->bb_free_root)); 32707a2fcbf7SAneesh Kumar K.V 32717a2fcbf7SAneesh Kumar K.V while (n) { 327218aadd47SBobi Jam entry = rb_entry(n, struct ext4_free_data, efd_node); 327318aadd47SBobi Jam ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count); 32747a2fcbf7SAneesh Kumar K.V n = rb_next(n); 32757a2fcbf7SAneesh Kumar K.V } 32767a2fcbf7SAneesh Kumar K.V return; 32777a2fcbf7SAneesh Kumar K.V } 32787a2fcbf7SAneesh Kumar K.V 32797a2fcbf7SAneesh Kumar K.V /* 3280c9de560dSAlex Tomas * the function goes through all preallocation in this group and marks them 3281c9de560dSAlex Tomas * used in in-core bitmap. buddy must be generated from this bitmap 3282955ce5f5SAneesh Kumar K.V * Need to be called with ext4 group lock held 3283c9de560dSAlex Tomas */ 3284089ceeccSEric Sandeen static noinline_for_stack 3285089ceeccSEric Sandeen void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 3286c9de560dSAlex Tomas ext4_group_t group) 3287c9de560dSAlex Tomas { 3288c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3289c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3290c9de560dSAlex Tomas struct list_head *cur; 3291c9de560dSAlex Tomas ext4_group_t groupnr; 3292c9de560dSAlex Tomas ext4_grpblk_t start; 3293c9de560dSAlex Tomas int preallocated = 0; 3294c9de560dSAlex Tomas int len; 3295c9de560dSAlex Tomas 3296c9de560dSAlex Tomas /* all form of preallocation discards first load group, 3297c9de560dSAlex Tomas * so the only competing code is preallocation use. 3298c9de560dSAlex Tomas * we don't need any locking here 3299c9de560dSAlex Tomas * notice we do NOT ignore preallocations with pa_deleted 3300c9de560dSAlex Tomas * otherwise we could leave used blocks available for 3301c9de560dSAlex Tomas * allocation in buddy when concurrent ext4_mb_put_pa() 3302c9de560dSAlex Tomas * is dropping preallocation 3303c9de560dSAlex Tomas */ 3304c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 3305c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3306c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3307c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3308c9de560dSAlex Tomas &groupnr, &start); 3309c9de560dSAlex Tomas len = pa->pa_len; 3310c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3311c9de560dSAlex Tomas if (unlikely(len == 0)) 3312c9de560dSAlex Tomas continue; 3313c9de560dSAlex Tomas BUG_ON(groupnr != group); 3314c3e94d1dSYongqiang Yang ext4_set_bits(bitmap, start, len); 3315c9de560dSAlex Tomas preallocated += len; 3316c9de560dSAlex Tomas } 33176ba495e9STheodore Ts'o mb_debug(1, "prellocated %u for group %u\n", preallocated, group); 3318c9de560dSAlex Tomas } 3319c9de560dSAlex Tomas 3320c9de560dSAlex Tomas static void ext4_mb_pa_callback(struct rcu_head *head) 3321c9de560dSAlex Tomas { 3322c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3323c9de560dSAlex Tomas pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 3324c9de560dSAlex Tomas kmem_cache_free(ext4_pspace_cachep, pa); 3325c9de560dSAlex Tomas } 3326c9de560dSAlex Tomas 3327c9de560dSAlex Tomas /* 3328c9de560dSAlex Tomas * drops a reference to preallocated space descriptor 3329c9de560dSAlex Tomas * if this was the last reference and the space is consumed 3330c9de560dSAlex Tomas */ 3331c9de560dSAlex Tomas static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 3332c9de560dSAlex Tomas struct super_block *sb, struct ext4_prealloc_space *pa) 3333c9de560dSAlex Tomas { 3334a9df9a49STheodore Ts'o ext4_group_t grp; 3335d33a1976SEric Sandeen ext4_fsblk_t grp_blk; 3336c9de560dSAlex Tomas 3337c9de560dSAlex Tomas if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) 3338c9de560dSAlex Tomas return; 3339c9de560dSAlex Tomas 3340c9de560dSAlex Tomas /* in this short window concurrent discard can set pa_deleted */ 3341c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3342c9de560dSAlex Tomas if (pa->pa_deleted == 1) { 3343c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3344c9de560dSAlex Tomas return; 3345c9de560dSAlex Tomas } 3346c9de560dSAlex Tomas 3347c9de560dSAlex Tomas pa->pa_deleted = 1; 3348c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3349c9de560dSAlex Tomas 3350d33a1976SEric Sandeen grp_blk = pa->pa_pstart; 3351cc0fb9adSAneesh Kumar K.V /* 3352cc0fb9adSAneesh Kumar K.V * If doing group-based preallocation, pa_pstart may be in the 3353cc0fb9adSAneesh Kumar K.V * next group when pa is used up 3354cc0fb9adSAneesh Kumar K.V */ 3355cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 3356d33a1976SEric Sandeen grp_blk--; 3357d33a1976SEric Sandeen 3358d33a1976SEric Sandeen ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); 3359c9de560dSAlex Tomas 3360c9de560dSAlex Tomas /* 3361c9de560dSAlex Tomas * possible race: 3362c9de560dSAlex Tomas * 3363c9de560dSAlex Tomas * P1 (buddy init) P2 (regular allocation) 3364c9de560dSAlex Tomas * find block B in PA 3365c9de560dSAlex Tomas * copy on-disk bitmap to buddy 3366c9de560dSAlex Tomas * mark B in on-disk bitmap 3367c9de560dSAlex Tomas * drop PA from group 3368c9de560dSAlex Tomas * mark all PAs in buddy 3369c9de560dSAlex Tomas * 3370c9de560dSAlex Tomas * thus, P1 initializes buddy with B available. to prevent this 3371c9de560dSAlex Tomas * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 3372c9de560dSAlex Tomas * against that pair 3373c9de560dSAlex Tomas */ 3374c9de560dSAlex Tomas ext4_lock_group(sb, grp); 3375c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3376c9de560dSAlex Tomas ext4_unlock_group(sb, grp); 3377c9de560dSAlex Tomas 3378c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3379c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3380c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3381c9de560dSAlex Tomas 3382c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3383c9de560dSAlex Tomas } 3384c9de560dSAlex Tomas 3385c9de560dSAlex Tomas /* 3386c9de560dSAlex Tomas * creates new preallocated space for given inode 3387c9de560dSAlex Tomas */ 33884ddfef7bSEric Sandeen static noinline_for_stack int 33894ddfef7bSEric Sandeen ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 3390c9de560dSAlex Tomas { 3391c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 339253accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(sb); 3393c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3394c9de560dSAlex Tomas struct ext4_group_info *grp; 3395c9de560dSAlex Tomas struct ext4_inode_info *ei; 3396c9de560dSAlex Tomas 3397c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3398c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3399c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3400c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3401c9de560dSAlex Tomas 3402c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3403c9de560dSAlex Tomas if (pa == NULL) 3404c9de560dSAlex Tomas return -ENOMEM; 3405c9de560dSAlex Tomas 3406c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { 3407c9de560dSAlex Tomas int winl; 3408c9de560dSAlex Tomas int wins; 3409c9de560dSAlex Tomas int win; 3410c9de560dSAlex Tomas int offs; 3411c9de560dSAlex Tomas 3412c9de560dSAlex Tomas /* we can't allocate as much as normalizer wants. 3413c9de560dSAlex Tomas * so, found space must get proper lstart 3414c9de560dSAlex Tomas * to cover original request */ 3415c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 3416c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 3417c9de560dSAlex Tomas 3418c9de560dSAlex Tomas /* we're limited by original request in that 3419c9de560dSAlex Tomas * logical block must be covered any way 3420c9de560dSAlex Tomas * winl is window we can move our chunk within */ 3421c9de560dSAlex Tomas winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; 3422c9de560dSAlex Tomas 3423c9de560dSAlex Tomas /* also, we should cover whole original request */ 342453accfa9STheodore Ts'o wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len); 3425c9de560dSAlex Tomas 3426c9de560dSAlex Tomas /* the smallest one defines real window */ 3427c9de560dSAlex Tomas win = min(winl, wins); 3428c9de560dSAlex Tomas 342953accfa9STheodore Ts'o offs = ac->ac_o_ex.fe_logical % 343053accfa9STheodore Ts'o EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 3431c9de560dSAlex Tomas if (offs && offs < win) 3432c9de560dSAlex Tomas win = offs; 3433c9de560dSAlex Tomas 343453accfa9STheodore Ts'o ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - 343553accfa9STheodore Ts'o EXT4_B2C(sbi, win); 3436c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 3437c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 3438c9de560dSAlex Tomas } 3439c9de560dSAlex Tomas 3440c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3441c9de560dSAlex Tomas * allocated blocks for history */ 3442c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3443c9de560dSAlex Tomas 3444c9de560dSAlex Tomas pa->pa_lstart = ac->ac_b_ex.fe_logical; 3445c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3446c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3447c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3448c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3449c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 3450d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3451d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 3452c9de560dSAlex Tomas pa->pa_deleted = 0; 3453cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_INODE_PA; 3454c9de560dSAlex Tomas 34556ba495e9STheodore Ts'o mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, 3456c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 34579bffad1eSTheodore Ts'o trace_ext4_mb_new_inode_pa(ac, pa); 3458c9de560dSAlex Tomas 3459c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 346053accfa9STheodore Ts'o atomic_add(pa->pa_free, &sbi->s_mb_preallocated); 3461c9de560dSAlex Tomas 3462c9de560dSAlex Tomas ei = EXT4_I(ac->ac_inode); 3463c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3464c9de560dSAlex Tomas 3465c9de560dSAlex Tomas pa->pa_obj_lock = &ei->i_prealloc_lock; 3466c9de560dSAlex Tomas pa->pa_inode = ac->ac_inode; 3467c9de560dSAlex Tomas 3468c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3469c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3470c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3471c9de560dSAlex Tomas 3472c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3473c9de560dSAlex Tomas list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); 3474c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3475c9de560dSAlex Tomas 3476c9de560dSAlex Tomas return 0; 3477c9de560dSAlex Tomas } 3478c9de560dSAlex Tomas 3479c9de560dSAlex Tomas /* 3480c9de560dSAlex Tomas * creates new preallocated space for locality group inodes belongs to 3481c9de560dSAlex Tomas */ 34824ddfef7bSEric Sandeen static noinline_for_stack int 34834ddfef7bSEric Sandeen ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 3484c9de560dSAlex Tomas { 3485c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3486c9de560dSAlex Tomas struct ext4_locality_group *lg; 3487c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3488c9de560dSAlex Tomas struct ext4_group_info *grp; 3489c9de560dSAlex Tomas 3490c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3491c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3492c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3493c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3494c9de560dSAlex Tomas 3495c9de560dSAlex Tomas BUG_ON(ext4_pspace_cachep == NULL); 3496c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3497c9de560dSAlex Tomas if (pa == NULL) 3498c9de560dSAlex Tomas return -ENOMEM; 3499c9de560dSAlex Tomas 3500c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3501c9de560dSAlex Tomas * allocated blocks for history */ 3502c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3503c9de560dSAlex Tomas 3504c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3505c9de560dSAlex Tomas pa->pa_lstart = pa->pa_pstart; 3506c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3507c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3508c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3509c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 35106be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3511d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 3512c9de560dSAlex Tomas pa->pa_deleted = 0; 3513cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_GROUP_PA; 3514c9de560dSAlex Tomas 35156ba495e9STheodore Ts'o mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, 3516c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 35179bffad1eSTheodore Ts'o trace_ext4_mb_new_group_pa(ac, pa); 3518c9de560dSAlex Tomas 3519c9de560dSAlex Tomas ext4_mb_use_group_pa(ac, pa); 3520c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3521c9de560dSAlex Tomas 3522c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3523c9de560dSAlex Tomas lg = ac->ac_lg; 3524c9de560dSAlex Tomas BUG_ON(lg == NULL); 3525c9de560dSAlex Tomas 3526c9de560dSAlex Tomas pa->pa_obj_lock = &lg->lg_prealloc_lock; 3527c9de560dSAlex Tomas pa->pa_inode = NULL; 3528c9de560dSAlex Tomas 3529c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3530c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3531c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3532c9de560dSAlex Tomas 35336be2ded1SAneesh Kumar K.V /* 35346be2ded1SAneesh Kumar K.V * We will later add the new pa to the right bucket 35356be2ded1SAneesh Kumar K.V * after updating the pa_free in ext4_mb_release_context 35366be2ded1SAneesh Kumar K.V */ 3537c9de560dSAlex Tomas return 0; 3538c9de560dSAlex Tomas } 3539c9de560dSAlex Tomas 3540c9de560dSAlex Tomas static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 3541c9de560dSAlex Tomas { 3542c9de560dSAlex Tomas int err; 3543c9de560dSAlex Tomas 3544c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 3545c9de560dSAlex Tomas err = ext4_mb_new_group_pa(ac); 3546c9de560dSAlex Tomas else 3547c9de560dSAlex Tomas err = ext4_mb_new_inode_pa(ac); 3548c9de560dSAlex Tomas return err; 3549c9de560dSAlex Tomas } 3550c9de560dSAlex Tomas 3551c9de560dSAlex Tomas /* 3552c9de560dSAlex Tomas * finds all unused blocks in on-disk bitmap, frees them in 3553c9de560dSAlex Tomas * in-core bitmap and buddy. 3554c9de560dSAlex Tomas * @pa must be unlinked from inode and group lists, so that 3555c9de560dSAlex Tomas * nobody else can find/use it. 3556c9de560dSAlex Tomas * the caller MUST hold group/inode locks. 3557c9de560dSAlex Tomas * TODO: optimize the case when there are no in-core structures yet 3558c9de560dSAlex Tomas */ 35594ddfef7bSEric Sandeen static noinline_for_stack int 35604ddfef7bSEric Sandeen ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 35613e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 3562c9de560dSAlex Tomas { 3563c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3564c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3565498e5f24STheodore Ts'o unsigned int end; 3566498e5f24STheodore Ts'o unsigned int next; 3567c9de560dSAlex Tomas ext4_group_t group; 3568c9de560dSAlex Tomas ext4_grpblk_t bit; 3569ba80b101STheodore Ts'o unsigned long long grp_blk_start; 3570c9de560dSAlex Tomas int err = 0; 3571c9de560dSAlex Tomas int free = 0; 3572c9de560dSAlex Tomas 3573c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3574c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 357553accfa9STheodore Ts'o grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit); 3576c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3577c9de560dSAlex Tomas end = bit + pa->pa_len; 3578c9de560dSAlex Tomas 3579c9de560dSAlex Tomas while (bit < end) { 3580ffad0a44SAneesh Kumar K.V bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 3581c9de560dSAlex Tomas if (bit >= end) 3582c9de560dSAlex Tomas break; 3583ffad0a44SAneesh Kumar K.V next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 35846ba495e9STheodore Ts'o mb_debug(1, " free preallocated %u/%u in group %u\n", 35855a0790c2SAndi Kleen (unsigned) ext4_group_first_block_no(sb, group) + bit, 35865a0790c2SAndi Kleen (unsigned) next - bit, (unsigned) group); 3587c9de560dSAlex Tomas free += next - bit; 3588c9de560dSAlex Tomas 35893e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); 359053accfa9STheodore Ts'o trace_ext4_mb_release_inode_pa(pa, (grp_blk_start + 359153accfa9STheodore Ts'o EXT4_C2B(sbi, bit)), 3592a9c667f8SLukas Czerner next - bit); 3593c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 3594c9de560dSAlex Tomas bit = next + 1; 3595c9de560dSAlex Tomas } 3596c9de560dSAlex Tomas if (free != pa->pa_free) { 35979d8b9ec4STheodore Ts'o ext4_msg(e4b->bd_sb, KERN_CRIT, 35989d8b9ec4STheodore Ts'o "pa %p: logic %lu, phys. %lu, len %lu", 3599c9de560dSAlex Tomas pa, (unsigned long) pa->pa_lstart, 3600c9de560dSAlex Tomas (unsigned long) pa->pa_pstart, 3601c9de560dSAlex Tomas (unsigned long) pa->pa_len); 3602e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", 360326346ff6SAneesh Kumar K.V free, pa->pa_free); 3604e56eb659SAneesh Kumar K.V /* 3605e56eb659SAneesh Kumar K.V * pa is already deleted so we use the value obtained 3606e56eb659SAneesh Kumar K.V * from the bitmap and continue. 3607e56eb659SAneesh Kumar K.V */ 3608c9de560dSAlex Tomas } 3609c9de560dSAlex Tomas atomic_add(free, &sbi->s_mb_discarded); 3610c9de560dSAlex Tomas 3611c9de560dSAlex Tomas return err; 3612c9de560dSAlex Tomas } 3613c9de560dSAlex Tomas 36144ddfef7bSEric Sandeen static noinline_for_stack int 36154ddfef7bSEric Sandeen ext4_mb_release_group_pa(struct ext4_buddy *e4b, 36163e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 3617c9de560dSAlex Tomas { 3618c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3619c9de560dSAlex Tomas ext4_group_t group; 3620c9de560dSAlex Tomas ext4_grpblk_t bit; 3621c9de560dSAlex Tomas 362260e07cf5SYongqiang Yang trace_ext4_mb_release_group_pa(sb, pa); 3623c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3624c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3625c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3626c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 3627c9de560dSAlex Tomas atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 36283e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); 3629c9de560dSAlex Tomas 3630c9de560dSAlex Tomas return 0; 3631c9de560dSAlex Tomas } 3632c9de560dSAlex Tomas 3633c9de560dSAlex Tomas /* 3634c9de560dSAlex Tomas * releases all preallocations in given group 3635c9de560dSAlex Tomas * 3636c9de560dSAlex Tomas * first, we need to decide discard policy: 3637c9de560dSAlex Tomas * - when do we discard 3638c9de560dSAlex Tomas * 1) ENOSPC 3639c9de560dSAlex Tomas * - how many do we discard 3640c9de560dSAlex Tomas * 1) how many requested 3641c9de560dSAlex Tomas */ 36424ddfef7bSEric Sandeen static noinline_for_stack int 36434ddfef7bSEric Sandeen ext4_mb_discard_group_preallocations(struct super_block *sb, 3644c9de560dSAlex Tomas ext4_group_t group, int needed) 3645c9de560dSAlex Tomas { 3646c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3647c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3648c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3649c9de560dSAlex Tomas struct list_head list; 3650c9de560dSAlex Tomas struct ext4_buddy e4b; 3651c9de560dSAlex Tomas int err; 3652c9de560dSAlex Tomas int busy = 0; 3653c9de560dSAlex Tomas int free = 0; 3654c9de560dSAlex Tomas 36556ba495e9STheodore Ts'o mb_debug(1, "discard preallocation for group %u\n", group); 3656c9de560dSAlex Tomas 3657c9de560dSAlex Tomas if (list_empty(&grp->bb_prealloc_list)) 3658c9de560dSAlex Tomas return 0; 3659c9de560dSAlex Tomas 3660574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 3661c9de560dSAlex Tomas if (bitmap_bh == NULL) { 366212062dddSEric Sandeen ext4_error(sb, "Error reading block bitmap for %u", group); 3663ce89f46cSAneesh Kumar K.V return 0; 3664c9de560dSAlex Tomas } 3665c9de560dSAlex Tomas 3666c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3667ce89f46cSAneesh Kumar K.V if (err) { 366812062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", group); 3669ce89f46cSAneesh Kumar K.V put_bh(bitmap_bh); 3670ce89f46cSAneesh Kumar K.V return 0; 3671ce89f46cSAneesh Kumar K.V } 3672c9de560dSAlex Tomas 3673c9de560dSAlex Tomas if (needed == 0) 36747137d7a4STheodore Ts'o needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1; 3675c9de560dSAlex Tomas 3676c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3677c9de560dSAlex Tomas repeat: 3678c9de560dSAlex Tomas ext4_lock_group(sb, group); 3679c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, 3680c9de560dSAlex Tomas &grp->bb_prealloc_list, pa_group_list) { 3681c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3682c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3683c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3684c9de560dSAlex Tomas busy = 1; 3685c9de560dSAlex Tomas continue; 3686c9de560dSAlex Tomas } 3687c9de560dSAlex Tomas if (pa->pa_deleted) { 3688c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3689c9de560dSAlex Tomas continue; 3690c9de560dSAlex Tomas } 3691c9de560dSAlex Tomas 3692c9de560dSAlex Tomas /* seems this one can be freed ... */ 3693c9de560dSAlex Tomas pa->pa_deleted = 1; 3694c9de560dSAlex Tomas 3695c9de560dSAlex Tomas /* we can trust pa_free ... */ 3696c9de560dSAlex Tomas free += pa->pa_free; 3697c9de560dSAlex Tomas 3698c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3699c9de560dSAlex Tomas 3700c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3701c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3702c9de560dSAlex Tomas } 3703c9de560dSAlex Tomas 3704c9de560dSAlex Tomas /* if we still need more blocks and some PAs were used, try again */ 3705c9de560dSAlex Tomas if (free < needed && busy) { 3706c9de560dSAlex Tomas busy = 0; 3707c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3708c9de560dSAlex Tomas /* 3709c9de560dSAlex Tomas * Yield the CPU here so that we don't get soft lockup 3710c9de560dSAlex Tomas * in non preempt case. 3711c9de560dSAlex Tomas */ 3712c9de560dSAlex Tomas yield(); 3713c9de560dSAlex Tomas goto repeat; 3714c9de560dSAlex Tomas } 3715c9de560dSAlex Tomas 3716c9de560dSAlex Tomas /* found anything to free? */ 3717c9de560dSAlex Tomas if (list_empty(&list)) { 3718c9de560dSAlex Tomas BUG_ON(free != 0); 3719c9de560dSAlex Tomas goto out; 3720c9de560dSAlex Tomas } 3721c9de560dSAlex Tomas 3722c9de560dSAlex Tomas /* now free all selected PAs */ 3723c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3724c9de560dSAlex Tomas 3725c9de560dSAlex Tomas /* remove from object (inode or locality group) */ 3726c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3727c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3728c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3729c9de560dSAlex Tomas 3730cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 37313e1e5f50SEric Sandeen ext4_mb_release_group_pa(&e4b, pa); 3732c9de560dSAlex Tomas else 37333e1e5f50SEric Sandeen ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 3734c9de560dSAlex Tomas 3735c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 3736c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3737c9de560dSAlex Tomas } 3738c9de560dSAlex Tomas 3739c9de560dSAlex Tomas out: 3740c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3741e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3742c9de560dSAlex Tomas put_bh(bitmap_bh); 3743c9de560dSAlex Tomas return free; 3744c9de560dSAlex Tomas } 3745c9de560dSAlex Tomas 3746c9de560dSAlex Tomas /* 3747c9de560dSAlex Tomas * releases all non-used preallocated blocks for given inode 3748c9de560dSAlex Tomas * 3749c9de560dSAlex Tomas * It's important to discard preallocations under i_data_sem 3750c9de560dSAlex Tomas * We don't want another block to be served from the prealloc 3751c9de560dSAlex Tomas * space when we are discarding the inode prealloc space. 3752c9de560dSAlex Tomas * 3753c9de560dSAlex Tomas * FIXME!! Make sure it is valid at all the call sites 3754c9de560dSAlex Tomas */ 3755c2ea3fdeSTheodore Ts'o void ext4_discard_preallocations(struct inode *inode) 3756c9de560dSAlex Tomas { 3757c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(inode); 3758c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 3759c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3760c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3761c9de560dSAlex Tomas ext4_group_t group = 0; 3762c9de560dSAlex Tomas struct list_head list; 3763c9de560dSAlex Tomas struct ext4_buddy e4b; 3764c9de560dSAlex Tomas int err; 3765c9de560dSAlex Tomas 3766c2ea3fdeSTheodore Ts'o if (!S_ISREG(inode->i_mode)) { 3767c9de560dSAlex Tomas /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ 3768c9de560dSAlex Tomas return; 3769c9de560dSAlex Tomas } 3770c9de560dSAlex Tomas 37716ba495e9STheodore Ts'o mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); 37729bffad1eSTheodore Ts'o trace_ext4_discard_preallocations(inode); 3773c9de560dSAlex Tomas 3774c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3775c9de560dSAlex Tomas 3776c9de560dSAlex Tomas repeat: 3777c9de560dSAlex Tomas /* first, collect all pa's in the inode */ 3778c9de560dSAlex Tomas spin_lock(&ei->i_prealloc_lock); 3779c9de560dSAlex Tomas while (!list_empty(&ei->i_prealloc_list)) { 3780c9de560dSAlex Tomas pa = list_entry(ei->i_prealloc_list.next, 3781c9de560dSAlex Tomas struct ext4_prealloc_space, pa_inode_list); 3782c9de560dSAlex Tomas BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); 3783c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3784c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3785c9de560dSAlex Tomas /* this shouldn't happen often - nobody should 3786c9de560dSAlex Tomas * use preallocation while we're discarding it */ 3787c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3788c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 37899d8b9ec4STheodore Ts'o ext4_msg(sb, KERN_ERR, 37909d8b9ec4STheodore Ts'o "uh-oh! used pa while discarding"); 3791c9de560dSAlex Tomas WARN_ON(1); 3792c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 3793c9de560dSAlex Tomas goto repeat; 3794c9de560dSAlex Tomas 3795c9de560dSAlex Tomas } 3796c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 3797c9de560dSAlex Tomas pa->pa_deleted = 1; 3798c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3799c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3800c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3801c9de560dSAlex Tomas continue; 3802c9de560dSAlex Tomas } 3803c9de560dSAlex Tomas 3804c9de560dSAlex Tomas /* someone is deleting pa right now */ 3805c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3806c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3807c9de560dSAlex Tomas 3808c9de560dSAlex Tomas /* we have to wait here because pa_deleted 3809c9de560dSAlex Tomas * doesn't mean pa is already unlinked from 3810c9de560dSAlex Tomas * the list. as we might be called from 3811c9de560dSAlex Tomas * ->clear_inode() the inode will get freed 3812c9de560dSAlex Tomas * and concurrent thread which is unlinking 3813c9de560dSAlex Tomas * pa from inode's list may access already 3814c9de560dSAlex Tomas * freed memory, bad-bad-bad */ 3815c9de560dSAlex Tomas 3816c9de560dSAlex Tomas /* XXX: if this happens too often, we can 3817c9de560dSAlex Tomas * add a flag to force wait only in case 3818c9de560dSAlex Tomas * of ->clear_inode(), but not in case of 3819c9de560dSAlex Tomas * regular truncate */ 3820c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 3821c9de560dSAlex Tomas goto repeat; 3822c9de560dSAlex Tomas } 3823c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3824c9de560dSAlex Tomas 3825c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3826cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_INODE_PA); 3827c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); 3828c9de560dSAlex Tomas 3829c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3830ce89f46cSAneesh Kumar K.V if (err) { 383112062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", 383212062dddSEric Sandeen group); 3833ce89f46cSAneesh Kumar K.V continue; 3834ce89f46cSAneesh Kumar K.V } 3835c9de560dSAlex Tomas 3836574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 3837c9de560dSAlex Tomas if (bitmap_bh == NULL) { 383812062dddSEric Sandeen ext4_error(sb, "Error reading block bitmap for %u", 383912062dddSEric Sandeen group); 3840e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3841ce89f46cSAneesh Kumar K.V continue; 3842c9de560dSAlex Tomas } 3843c9de560dSAlex Tomas 3844c9de560dSAlex Tomas ext4_lock_group(sb, group); 3845c9de560dSAlex Tomas list_del(&pa->pa_group_list); 38463e1e5f50SEric Sandeen ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 3847c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3848c9de560dSAlex Tomas 3849e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3850c9de560dSAlex Tomas put_bh(bitmap_bh); 3851c9de560dSAlex Tomas 3852c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 3853c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3854c9de560dSAlex Tomas } 3855c9de560dSAlex Tomas } 3856c9de560dSAlex Tomas 38576ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 3858c9de560dSAlex Tomas static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 3859c9de560dSAlex Tomas { 3860c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 38618df9675fSTheodore Ts'o ext4_group_t ngroups, i; 3862c9de560dSAlex Tomas 38634dd89fc6STheodore Ts'o if (!mb_enable_debug || 38644dd89fc6STheodore Ts'o (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) 3865e3570639SEric Sandeen return; 3866e3570639SEric Sandeen 38677f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:" 38689d8b9ec4STheodore Ts'o " Allocation context details:"); 38697f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d", 3870c9de560dSAlex Tomas ac->ac_status, ac->ac_flags); 38717f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, " 38729d8b9ec4STheodore Ts'o "goal %lu/%lu/%lu@%lu, " 38739d8b9ec4STheodore Ts'o "best %lu/%lu/%lu@%lu cr %d", 3874c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_group, 3875c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_start, 3876c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_len, 3877c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_logical, 3878c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_group, 3879c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_start, 3880c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_len, 3881c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_logical, 3882c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_group, 3883c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_start, 3884c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_len, 3885c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_logical, 3886c9de560dSAlex Tomas (int)ac->ac_criteria); 38877f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found", 38889d8b9ec4STheodore Ts'o ac->ac_ex_scanned, ac->ac_found); 38897f6a11e7SJoe Perches ext4_msg(ac->ac_sb, KERN_ERR, "groups: "); 38908df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 38918df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 3892c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, i); 3893c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3894c9de560dSAlex Tomas ext4_grpblk_t start; 3895c9de560dSAlex Tomas struct list_head *cur; 3896c9de560dSAlex Tomas ext4_lock_group(sb, i); 3897c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 3898c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, 3899c9de560dSAlex Tomas pa_group_list); 3900c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3901c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3902c9de560dSAlex Tomas NULL, &start); 3903c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 39041c718505SAkira Fujita printk(KERN_ERR "PA:%u:%d:%u \n", i, 3905c9de560dSAlex Tomas start, pa->pa_len); 3906c9de560dSAlex Tomas } 390760bd63d1SSolofo Ramangalahy ext4_unlock_group(sb, i); 3908c9de560dSAlex Tomas 3909c9de560dSAlex Tomas if (grp->bb_free == 0) 3910c9de560dSAlex Tomas continue; 39111c718505SAkira Fujita printk(KERN_ERR "%u: %d/%d \n", 3912c9de560dSAlex Tomas i, grp->bb_free, grp->bb_fragments); 3913c9de560dSAlex Tomas } 3914c9de560dSAlex Tomas printk(KERN_ERR "\n"); 3915c9de560dSAlex Tomas } 3916c9de560dSAlex Tomas #else 3917c9de560dSAlex Tomas static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 3918c9de560dSAlex Tomas { 3919c9de560dSAlex Tomas return; 3920c9de560dSAlex Tomas } 3921c9de560dSAlex Tomas #endif 3922c9de560dSAlex Tomas 3923c9de560dSAlex Tomas /* 3924c9de560dSAlex Tomas * We use locality group preallocation for small size file. The size of the 3925c9de560dSAlex Tomas * file is determined by the current size or the resulting size after 3926c9de560dSAlex Tomas * allocation which ever is larger 3927c9de560dSAlex Tomas * 3928b713a5ecSTheodore Ts'o * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 3929c9de560dSAlex Tomas */ 3930c9de560dSAlex Tomas static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 3931c9de560dSAlex Tomas { 3932c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3933c9de560dSAlex Tomas int bsbits = ac->ac_sb->s_blocksize_bits; 3934c9de560dSAlex Tomas loff_t size, isize; 3935c9de560dSAlex Tomas 3936c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3937c9de560dSAlex Tomas return; 3938c9de560dSAlex Tomas 39394ba74d00STheodore Ts'o if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 39404ba74d00STheodore Ts'o return; 39414ba74d00STheodore Ts'o 394253accfa9STheodore Ts'o size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len); 394350797481STheodore Ts'o isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 394450797481STheodore Ts'o >> bsbits; 3945c9de560dSAlex Tomas 394650797481STheodore Ts'o if ((size == isize) && 394750797481STheodore Ts'o !ext4_fs_is_busy(sbi) && 394850797481STheodore Ts'o (atomic_read(&ac->ac_inode->i_writecount) == 0)) { 394950797481STheodore Ts'o ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 395050797481STheodore Ts'o return; 395150797481STheodore Ts'o } 395250797481STheodore Ts'o 3953ebbe0277SRobin Dong if (sbi->s_mb_group_prealloc <= 0) { 3954ebbe0277SRobin Dong ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 3955ebbe0277SRobin Dong return; 3956ebbe0277SRobin Dong } 3957ebbe0277SRobin Dong 3958c9de560dSAlex Tomas /* don't use group allocation for large files */ 395971780577STheodore Ts'o size = max(size, isize); 3960cc483f10STao Ma if (size > sbi->s_mb_stream_request) { 39614ba74d00STheodore Ts'o ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 3962c9de560dSAlex Tomas return; 39634ba74d00STheodore Ts'o } 3964c9de560dSAlex Tomas 3965c9de560dSAlex Tomas BUG_ON(ac->ac_lg != NULL); 3966c9de560dSAlex Tomas /* 3967c9de560dSAlex Tomas * locality group prealloc space are per cpu. The reason for having 3968c9de560dSAlex Tomas * per cpu locality group is to reduce the contention between block 3969c9de560dSAlex Tomas * request from multiple CPUs. 3970c9de560dSAlex Tomas */ 3971ca0c9584SChristoph Lameter ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups); 3972c9de560dSAlex Tomas 3973c9de560dSAlex Tomas /* we're going to use group allocation */ 3974c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 3975c9de560dSAlex Tomas 3976c9de560dSAlex Tomas /* serialize all allocations in the group */ 3977c9de560dSAlex Tomas mutex_lock(&ac->ac_lg->lg_mutex); 3978c9de560dSAlex Tomas } 3979c9de560dSAlex Tomas 39804ddfef7bSEric Sandeen static noinline_for_stack int 39814ddfef7bSEric Sandeen ext4_mb_initialize_context(struct ext4_allocation_context *ac, 3982c9de560dSAlex Tomas struct ext4_allocation_request *ar) 3983c9de560dSAlex Tomas { 3984c9de560dSAlex Tomas struct super_block *sb = ar->inode->i_sb; 3985c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3986c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 3987c9de560dSAlex Tomas ext4_group_t group; 3988498e5f24STheodore Ts'o unsigned int len; 3989498e5f24STheodore Ts'o ext4_fsblk_t goal; 3990c9de560dSAlex Tomas ext4_grpblk_t block; 3991c9de560dSAlex Tomas 3992c9de560dSAlex Tomas /* we can't allocate > group size */ 3993c9de560dSAlex Tomas len = ar->len; 3994c9de560dSAlex Tomas 3995c9de560dSAlex Tomas /* just a dirty hack to filter too big requests */ 39967137d7a4STheodore Ts'o if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10) 39977137d7a4STheodore Ts'o len = EXT4_CLUSTERS_PER_GROUP(sb) - 10; 3998c9de560dSAlex Tomas 3999c9de560dSAlex Tomas /* start searching from the goal */ 4000c9de560dSAlex Tomas goal = ar->goal; 4001c9de560dSAlex Tomas if (goal < le32_to_cpu(es->s_first_data_block) || 4002c9de560dSAlex Tomas goal >= ext4_blocks_count(es)) 4003c9de560dSAlex Tomas goal = le32_to_cpu(es->s_first_data_block); 4004c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, goal, &group, &block); 4005c9de560dSAlex Tomas 4006c9de560dSAlex Tomas /* set up allocation goals */ 4007833576b3STheodore Ts'o memset(ac, 0, sizeof(struct ext4_allocation_context)); 400853accfa9STheodore Ts'o ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1); 4009c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 4010c9de560dSAlex Tomas ac->ac_sb = sb; 4011c9de560dSAlex Tomas ac->ac_inode = ar->inode; 401253accfa9STheodore Ts'o ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical; 4013c9de560dSAlex Tomas ac->ac_o_ex.fe_group = group; 4014c9de560dSAlex Tomas ac->ac_o_ex.fe_start = block; 4015c9de560dSAlex Tomas ac->ac_o_ex.fe_len = len; 401653accfa9STheodore Ts'o ac->ac_g_ex = ac->ac_o_ex; 4017c9de560dSAlex Tomas ac->ac_flags = ar->flags; 4018c9de560dSAlex Tomas 4019c9de560dSAlex Tomas /* we have to define context: we'll we work with a file or 4020c9de560dSAlex Tomas * locality group. this is a policy, actually */ 4021c9de560dSAlex Tomas ext4_mb_group_or_file(ac); 4022c9de560dSAlex Tomas 40236ba495e9STheodore Ts'o mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " 4024c9de560dSAlex Tomas "left: %u/%u, right %u/%u to %swritable\n", 4025c9de560dSAlex Tomas (unsigned) ar->len, (unsigned) ar->logical, 4026c9de560dSAlex Tomas (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 4027c9de560dSAlex Tomas (unsigned) ar->lleft, (unsigned) ar->pleft, 4028c9de560dSAlex Tomas (unsigned) ar->lright, (unsigned) ar->pright, 4029c9de560dSAlex Tomas atomic_read(&ar->inode->i_writecount) ? "" : "non-"); 4030c9de560dSAlex Tomas return 0; 4031c9de560dSAlex Tomas 4032c9de560dSAlex Tomas } 4033c9de560dSAlex Tomas 40346be2ded1SAneesh Kumar K.V static noinline_for_stack void 40356be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(struct super_block *sb, 40366be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg, 40376be2ded1SAneesh Kumar K.V int order, int total_entries) 40386be2ded1SAneesh Kumar K.V { 40396be2ded1SAneesh Kumar K.V ext4_group_t group = 0; 40406be2ded1SAneesh Kumar K.V struct ext4_buddy e4b; 40416be2ded1SAneesh Kumar K.V struct list_head discard_list; 40426be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa, *tmp; 40436be2ded1SAneesh Kumar K.V 40446ba495e9STheodore Ts'o mb_debug(1, "discard locality group preallocation\n"); 40456be2ded1SAneesh Kumar K.V 40466be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&discard_list); 40476be2ded1SAneesh Kumar K.V 40486be2ded1SAneesh Kumar K.V spin_lock(&lg->lg_prealloc_lock); 40496be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 40506be2ded1SAneesh Kumar K.V pa_inode_list) { 40516be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 40526be2ded1SAneesh Kumar K.V if (atomic_read(&pa->pa_count)) { 40536be2ded1SAneesh Kumar K.V /* 40546be2ded1SAneesh Kumar K.V * This is the pa that we just used 40556be2ded1SAneesh Kumar K.V * for block allocation. So don't 40566be2ded1SAneesh Kumar K.V * free that 40576be2ded1SAneesh Kumar K.V */ 40586be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 40596be2ded1SAneesh Kumar K.V continue; 40606be2ded1SAneesh Kumar K.V } 40616be2ded1SAneesh Kumar K.V if (pa->pa_deleted) { 40626be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 40636be2ded1SAneesh Kumar K.V continue; 40646be2ded1SAneesh Kumar K.V } 40656be2ded1SAneesh Kumar K.V /* only lg prealloc space */ 4066cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_GROUP_PA); 40676be2ded1SAneesh Kumar K.V 40686be2ded1SAneesh Kumar K.V /* seems this one can be freed ... */ 40696be2ded1SAneesh Kumar K.V pa->pa_deleted = 1; 40706be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 40716be2ded1SAneesh Kumar K.V 40726be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 40736be2ded1SAneesh Kumar K.V list_add(&pa->u.pa_tmp_list, &discard_list); 40746be2ded1SAneesh Kumar K.V 40756be2ded1SAneesh Kumar K.V total_entries--; 40766be2ded1SAneesh Kumar K.V if (total_entries <= 5) { 40776be2ded1SAneesh Kumar K.V /* 40786be2ded1SAneesh Kumar K.V * we want to keep only 5 entries 40796be2ded1SAneesh Kumar K.V * allowing it to grow to 8. This 40806be2ded1SAneesh Kumar K.V * mak sure we don't call discard 40816be2ded1SAneesh Kumar K.V * soon for this list. 40826be2ded1SAneesh Kumar K.V */ 40836be2ded1SAneesh Kumar K.V break; 40846be2ded1SAneesh Kumar K.V } 40856be2ded1SAneesh Kumar K.V } 40866be2ded1SAneesh Kumar K.V spin_unlock(&lg->lg_prealloc_lock); 40876be2ded1SAneesh Kumar K.V 40886be2ded1SAneesh Kumar K.V list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 40896be2ded1SAneesh Kumar K.V 40906be2ded1SAneesh Kumar K.V ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); 40916be2ded1SAneesh Kumar K.V if (ext4_mb_load_buddy(sb, group, &e4b)) { 409212062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", 409312062dddSEric Sandeen group); 40946be2ded1SAneesh Kumar K.V continue; 40956be2ded1SAneesh Kumar K.V } 40966be2ded1SAneesh Kumar K.V ext4_lock_group(sb, group); 40976be2ded1SAneesh Kumar K.V list_del(&pa->pa_group_list); 40983e1e5f50SEric Sandeen ext4_mb_release_group_pa(&e4b, pa); 40996be2ded1SAneesh Kumar K.V ext4_unlock_group(sb, group); 41006be2ded1SAneesh Kumar K.V 4101e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 41026be2ded1SAneesh Kumar K.V list_del(&pa->u.pa_tmp_list); 41036be2ded1SAneesh Kumar K.V call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 41046be2ded1SAneesh Kumar K.V } 41056be2ded1SAneesh Kumar K.V } 41066be2ded1SAneesh Kumar K.V 41076be2ded1SAneesh Kumar K.V /* 41086be2ded1SAneesh Kumar K.V * We have incremented pa_count. So it cannot be freed at this 41096be2ded1SAneesh Kumar K.V * point. Also we hold lg_mutex. So no parallel allocation is 41106be2ded1SAneesh Kumar K.V * possible from this lg. That means pa_free cannot be updated. 41116be2ded1SAneesh Kumar K.V * 41126be2ded1SAneesh Kumar K.V * A parallel ext4_mb_discard_group_preallocations is possible. 41136be2ded1SAneesh Kumar K.V * which can cause the lg_prealloc_list to be updated. 41146be2ded1SAneesh Kumar K.V */ 41156be2ded1SAneesh Kumar K.V 41166be2ded1SAneesh Kumar K.V static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 41176be2ded1SAneesh Kumar K.V { 41186be2ded1SAneesh Kumar K.V int order, added = 0, lg_prealloc_count = 1; 41196be2ded1SAneesh Kumar K.V struct super_block *sb = ac->ac_sb; 41206be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg = ac->ac_lg; 41216be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 41226be2ded1SAneesh Kumar K.V 41236be2ded1SAneesh Kumar K.V order = fls(pa->pa_free) - 1; 41246be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 41256be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 41266be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 41276be2ded1SAneesh Kumar K.V /* Add the prealloc space to lg */ 41286be2ded1SAneesh Kumar K.V rcu_read_lock(); 41296be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 41306be2ded1SAneesh Kumar K.V pa_inode_list) { 41316be2ded1SAneesh Kumar K.V spin_lock(&tmp_pa->pa_lock); 41326be2ded1SAneesh Kumar K.V if (tmp_pa->pa_deleted) { 4133e7c9e3e9STheodore Ts'o spin_unlock(&tmp_pa->pa_lock); 41346be2ded1SAneesh Kumar K.V continue; 41356be2ded1SAneesh Kumar K.V } 41366be2ded1SAneesh Kumar K.V if (!added && pa->pa_free < tmp_pa->pa_free) { 41376be2ded1SAneesh Kumar K.V /* Add to the tail of the previous entry */ 41386be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 41396be2ded1SAneesh Kumar K.V &tmp_pa->pa_inode_list); 41406be2ded1SAneesh Kumar K.V added = 1; 41416be2ded1SAneesh Kumar K.V /* 41426be2ded1SAneesh Kumar K.V * we want to count the total 41436be2ded1SAneesh Kumar K.V * number of entries in the list 41446be2ded1SAneesh Kumar K.V */ 41456be2ded1SAneesh Kumar K.V } 41466be2ded1SAneesh Kumar K.V spin_unlock(&tmp_pa->pa_lock); 41476be2ded1SAneesh Kumar K.V lg_prealloc_count++; 41486be2ded1SAneesh Kumar K.V } 41496be2ded1SAneesh Kumar K.V if (!added) 41506be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 41516be2ded1SAneesh Kumar K.V &lg->lg_prealloc_list[order]); 41526be2ded1SAneesh Kumar K.V rcu_read_unlock(); 41536be2ded1SAneesh Kumar K.V 41546be2ded1SAneesh Kumar K.V /* Now trim the list to be not more than 8 elements */ 41556be2ded1SAneesh Kumar K.V if (lg_prealloc_count > 8) { 41566be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(sb, lg, 41576be2ded1SAneesh Kumar K.V order, lg_prealloc_count); 41586be2ded1SAneesh Kumar K.V return; 41596be2ded1SAneesh Kumar K.V } 41606be2ded1SAneesh Kumar K.V return ; 41616be2ded1SAneesh Kumar K.V } 41626be2ded1SAneesh Kumar K.V 4163c9de560dSAlex Tomas /* 4164c9de560dSAlex Tomas * release all resource we used in allocation 4165c9de560dSAlex Tomas */ 4166c9de560dSAlex Tomas static int ext4_mb_release_context(struct ext4_allocation_context *ac) 4167c9de560dSAlex Tomas { 416853accfa9STheodore Ts'o struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 41696be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa = ac->ac_pa; 41706be2ded1SAneesh Kumar K.V if (pa) { 4171cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) { 4172c9de560dSAlex Tomas /* see comment in ext4_mb_use_group_pa() */ 41736be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 417453accfa9STheodore Ts'o pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 417553accfa9STheodore Ts'o pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len); 41766be2ded1SAneesh Kumar K.V pa->pa_free -= ac->ac_b_ex.fe_len; 41776be2ded1SAneesh Kumar K.V pa->pa_len -= ac->ac_b_ex.fe_len; 41786be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 4179ba443916SAneesh Kumar K.V } 4180ba443916SAneesh Kumar K.V } 4181ba443916SAneesh Kumar K.V if (pa) { 41826be2ded1SAneesh Kumar K.V /* 41836be2ded1SAneesh Kumar K.V * We want to add the pa to the right bucket. 41846be2ded1SAneesh Kumar K.V * Remove it from the list and while adding 41856be2ded1SAneesh Kumar K.V * make sure the list to which we are adding 418644183d42SAmir Goldstein * doesn't grow big. 41876be2ded1SAneesh Kumar K.V */ 4188cc0fb9adSAneesh Kumar K.V if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { 41896be2ded1SAneesh Kumar K.V spin_lock(pa->pa_obj_lock); 41906be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 41916be2ded1SAneesh Kumar K.V spin_unlock(pa->pa_obj_lock); 41926be2ded1SAneesh Kumar K.V ext4_mb_add_n_trim(ac); 4193c9de560dSAlex Tomas } 41946be2ded1SAneesh Kumar K.V ext4_mb_put_pa(ac, ac->ac_sb, pa); 4195c9de560dSAlex Tomas } 4196c9de560dSAlex Tomas if (ac->ac_bitmap_page) 4197c9de560dSAlex Tomas page_cache_release(ac->ac_bitmap_page); 4198c9de560dSAlex Tomas if (ac->ac_buddy_page) 4199c9de560dSAlex Tomas page_cache_release(ac->ac_buddy_page); 4200c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 4201c9de560dSAlex Tomas mutex_unlock(&ac->ac_lg->lg_mutex); 4202c9de560dSAlex Tomas ext4_mb_collect_stats(ac); 4203c9de560dSAlex Tomas return 0; 4204c9de560dSAlex Tomas } 4205c9de560dSAlex Tomas 4206c9de560dSAlex Tomas static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 4207c9de560dSAlex Tomas { 42088df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 4209c9de560dSAlex Tomas int ret; 4210c9de560dSAlex Tomas int freed = 0; 4211c9de560dSAlex Tomas 42129bffad1eSTheodore Ts'o trace_ext4_mb_discard_preallocations(sb, needed); 42138df9675fSTheodore Ts'o for (i = 0; i < ngroups && needed > 0; i++) { 4214c9de560dSAlex Tomas ret = ext4_mb_discard_group_preallocations(sb, i, needed); 4215c9de560dSAlex Tomas freed += ret; 4216c9de560dSAlex Tomas needed -= ret; 4217c9de560dSAlex Tomas } 4218c9de560dSAlex Tomas 4219c9de560dSAlex Tomas return freed; 4220c9de560dSAlex Tomas } 4221c9de560dSAlex Tomas 4222c9de560dSAlex Tomas /* 4223c9de560dSAlex Tomas * Main entry point into mballoc to allocate blocks 4224c9de560dSAlex Tomas * it tries to use preallocation first, then falls back 4225c9de560dSAlex Tomas * to usual allocation 4226c9de560dSAlex Tomas */ 4227c9de560dSAlex Tomas ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 4228c9de560dSAlex Tomas struct ext4_allocation_request *ar, int *errp) 4229c9de560dSAlex Tomas { 42306bc6e63fSAneesh Kumar K.V int freed; 4231256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 4232c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4233c9de560dSAlex Tomas struct super_block *sb; 4234c9de560dSAlex Tomas ext4_fsblk_t block = 0; 423560e58e0fSMingming Cao unsigned int inquota = 0; 423653accfa9STheodore Ts'o unsigned int reserv_clstrs = 0; 4237c9de560dSAlex Tomas 4238c9de560dSAlex Tomas sb = ar->inode->i_sb; 4239c9de560dSAlex Tomas sbi = EXT4_SB(sb); 4240c9de560dSAlex Tomas 42419bffad1eSTheodore Ts'o trace_ext4_request_blocks(ar); 4242ba80b101STheodore Ts'o 424345dc63e7SDmitry Monakhov /* Allow to use superuser reservation for quota file */ 424445dc63e7SDmitry Monakhov if (IS_NOQUOTA(ar->inode)) 424545dc63e7SDmitry Monakhov ar->flags |= EXT4_MB_USE_ROOT_BLOCKS; 424645dc63e7SDmitry Monakhov 4247d2a17637SMingming Cao /* 424860e58e0fSMingming Cao * For delayed allocation, we could skip the ENOSPC and 424960e58e0fSMingming Cao * EDQUOT check, as blocks and quotas have been already 425060e58e0fSMingming Cao * reserved when data being copied into pagecache. 425160e58e0fSMingming Cao */ 4252f2321097STheodore Ts'o if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED)) 425360e58e0fSMingming Cao ar->flags |= EXT4_MB_DELALLOC_RESERVED; 425460e58e0fSMingming Cao else { 425560e58e0fSMingming Cao /* Without delayed allocation we need to verify 425660e58e0fSMingming Cao * there is enough free blocks to do block allocation 425760e58e0fSMingming Cao * and verify allocation doesn't exceed the quota limits. 4258d2a17637SMingming Cao */ 425955f020dbSAllison Henderson while (ar->len && 4260e7d5f315STheodore Ts'o ext4_claim_free_clusters(sbi, ar->len, ar->flags)) { 426155f020dbSAllison Henderson 4262030ba6bcSAneesh Kumar K.V /* let others to free the space */ 4263030ba6bcSAneesh Kumar K.V yield(); 4264030ba6bcSAneesh Kumar K.V ar->len = ar->len >> 1; 4265030ba6bcSAneesh Kumar K.V } 4266030ba6bcSAneesh Kumar K.V if (!ar->len) { 426707031431SMingming Cao *errp = -ENOSPC; 426807031431SMingming Cao return 0; 426907031431SMingming Cao } 427053accfa9STheodore Ts'o reserv_clstrs = ar->len; 427155f020dbSAllison Henderson if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) { 427253accfa9STheodore Ts'o dquot_alloc_block_nofail(ar->inode, 427353accfa9STheodore Ts'o EXT4_C2B(sbi, ar->len)); 427455f020dbSAllison Henderson } else { 427555f020dbSAllison Henderson while (ar->len && 427653accfa9STheodore Ts'o dquot_alloc_block(ar->inode, 427753accfa9STheodore Ts'o EXT4_C2B(sbi, ar->len))) { 427855f020dbSAllison Henderson 4279c9de560dSAlex Tomas ar->flags |= EXT4_MB_HINT_NOPREALLOC; 4280c9de560dSAlex Tomas ar->len--; 4281c9de560dSAlex Tomas } 428255f020dbSAllison Henderson } 428360e58e0fSMingming Cao inquota = ar->len; 4284c9de560dSAlex Tomas if (ar->len == 0) { 4285c9de560dSAlex Tomas *errp = -EDQUOT; 42866c7a120aSAditya Kali goto out; 4287c9de560dSAlex Tomas } 428860e58e0fSMingming Cao } 4289d2a17637SMingming Cao 4290256bdb49SEric Sandeen ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 4291833576b3STheodore Ts'o if (!ac) { 4292363d4251SShen Feng ar->len = 0; 4293256bdb49SEric Sandeen *errp = -ENOMEM; 42946c7a120aSAditya Kali goto out; 4295256bdb49SEric Sandeen } 4296256bdb49SEric Sandeen 4297256bdb49SEric Sandeen *errp = ext4_mb_initialize_context(ac, ar); 4298c9de560dSAlex Tomas if (*errp) { 4299c9de560dSAlex Tomas ar->len = 0; 43006c7a120aSAditya Kali goto out; 4301c9de560dSAlex Tomas } 4302c9de560dSAlex Tomas 4303256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 4304256bdb49SEric Sandeen if (!ext4_mb_use_preallocated(ac)) { 4305256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_ALLOC; 4306256bdb49SEric Sandeen ext4_mb_normalize_request(ac, ar); 4307c9de560dSAlex Tomas repeat: 4308c9de560dSAlex Tomas /* allocate space in core */ 43096c7a120aSAditya Kali *errp = ext4_mb_regular_allocator(ac); 43106c7a120aSAditya Kali if (*errp) 43116c7a120aSAditya Kali goto errout; 4312c9de560dSAlex Tomas 4313c9de560dSAlex Tomas /* as we've just preallocated more space than 4314c9de560dSAlex Tomas * user requested orinally, we store allocated 4315c9de560dSAlex Tomas * space in a special descriptor */ 4316256bdb49SEric Sandeen if (ac->ac_status == AC_STATUS_FOUND && 4317256bdb49SEric Sandeen ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 4318256bdb49SEric Sandeen ext4_mb_new_preallocation(ac); 4319c9de560dSAlex Tomas } 4320256bdb49SEric Sandeen if (likely(ac->ac_status == AC_STATUS_FOUND)) { 432153accfa9STheodore Ts'o *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs); 4322519deca0SAneesh Kumar K.V if (*errp == -EAGAIN) { 43238556e8f3SAneesh Kumar K.V /* 43248556e8f3SAneesh Kumar K.V * drop the reference that we took 43258556e8f3SAneesh Kumar K.V * in ext4_mb_use_best_found 43268556e8f3SAneesh Kumar K.V */ 43278556e8f3SAneesh Kumar K.V ext4_mb_release_context(ac); 4328519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_group = 0; 4329519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_start = 0; 4330519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len = 0; 4331519deca0SAneesh Kumar K.V ac->ac_status = AC_STATUS_CONTINUE; 4332519deca0SAneesh Kumar K.V goto repeat; 43336c7a120aSAditya Kali } else if (*errp) 43346c7a120aSAditya Kali errout: 4335b844167eSCurt Wohlgemuth ext4_discard_allocated_blocks(ac); 43366c7a120aSAditya Kali else { 4337256bdb49SEric Sandeen block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 4338256bdb49SEric Sandeen ar->len = ac->ac_b_ex.fe_len; 4339519deca0SAneesh Kumar K.V } 4340c9de560dSAlex Tomas } else { 4341256bdb49SEric Sandeen freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 4342c9de560dSAlex Tomas if (freed) 4343c9de560dSAlex Tomas goto repeat; 4344c9de560dSAlex Tomas *errp = -ENOSPC; 43456c7a120aSAditya Kali } 43466c7a120aSAditya Kali 43476c7a120aSAditya Kali if (*errp) { 4348256bdb49SEric Sandeen ac->ac_b_ex.fe_len = 0; 4349c9de560dSAlex Tomas ar->len = 0; 4350256bdb49SEric Sandeen ext4_mb_show_ac(ac); 4351c9de560dSAlex Tomas } 4352256bdb49SEric Sandeen ext4_mb_release_context(ac); 43536c7a120aSAditya Kali out: 43546c7a120aSAditya Kali if (ac) 4355363d4251SShen Feng kmem_cache_free(ext4_ac_cachep, ac); 435660e58e0fSMingming Cao if (inquota && ar->len < inquota) 435753accfa9STheodore Ts'o dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len)); 43580087d9fbSAneesh Kumar K.V if (!ar->len) { 4359f2321097STheodore Ts'o if (!ext4_test_inode_state(ar->inode, 4360f2321097STheodore Ts'o EXT4_STATE_DELALLOC_RESERVED)) 43610087d9fbSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 436257042651STheodore Ts'o percpu_counter_sub(&sbi->s_dirtyclusters_counter, 436353accfa9STheodore Ts'o reserv_clstrs); 43640087d9fbSAneesh Kumar K.V } 4365c9de560dSAlex Tomas 43669bffad1eSTheodore Ts'o trace_ext4_allocate_blocks(ar, (unsigned long long)block); 4367ba80b101STheodore Ts'o 4368c9de560dSAlex Tomas return block; 4369c9de560dSAlex Tomas } 4370c9de560dSAlex Tomas 4371c894058dSAneesh Kumar K.V /* 4372c894058dSAneesh Kumar K.V * We can merge two free data extents only if the physical blocks 4373c894058dSAneesh Kumar K.V * are contiguous, AND the extents were freed by the same transaction, 4374c894058dSAneesh Kumar K.V * AND the blocks are associated with the same group. 4375c894058dSAneesh Kumar K.V */ 4376c894058dSAneesh Kumar K.V static int can_merge(struct ext4_free_data *entry1, 4377c894058dSAneesh Kumar K.V struct ext4_free_data *entry2) 4378c894058dSAneesh Kumar K.V { 437918aadd47SBobi Jam if ((entry1->efd_tid == entry2->efd_tid) && 438018aadd47SBobi Jam (entry1->efd_group == entry2->efd_group) && 438118aadd47SBobi Jam ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster)) 4382c894058dSAneesh Kumar K.V return 1; 4383c894058dSAneesh Kumar K.V return 0; 4384c894058dSAneesh Kumar K.V } 4385c894058dSAneesh Kumar K.V 43864ddfef7bSEric Sandeen static noinline_for_stack int 43874ddfef7bSEric Sandeen ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 43887a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry) 4389c9de560dSAlex Tomas { 4390e29136f8STheodore Ts'o ext4_group_t group = e4b->bd_group; 439184130193STheodore Ts'o ext4_grpblk_t cluster; 43927a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 4393c9de560dSAlex Tomas struct ext4_group_info *db = e4b->bd_info; 4394c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 4395c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 4396c894058dSAneesh Kumar K.V struct rb_node **n = &db->bb_free_root.rb_node, *node; 4397c894058dSAneesh Kumar K.V struct rb_node *parent = NULL, *new_node; 4398c894058dSAneesh Kumar K.V 43990390131bSFrank Mayhar BUG_ON(!ext4_handle_valid(handle)); 4400c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 4401c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 4402c9de560dSAlex Tomas 440318aadd47SBobi Jam new_node = &new_entry->efd_node; 440418aadd47SBobi Jam cluster = new_entry->efd_start_cluster; 4405c9de560dSAlex Tomas 4406c894058dSAneesh Kumar K.V if (!*n) { 4407c894058dSAneesh Kumar K.V /* first free block exent. We need to 4408c894058dSAneesh Kumar K.V protect buddy cache from being freed, 4409c9de560dSAlex Tomas * otherwise we'll refresh it from 4410c9de560dSAlex Tomas * on-disk bitmap and lose not-yet-available 4411c9de560dSAlex Tomas * blocks */ 4412c9de560dSAlex Tomas page_cache_get(e4b->bd_buddy_page); 4413c9de560dSAlex Tomas page_cache_get(e4b->bd_bitmap_page); 4414c894058dSAneesh Kumar K.V } 4415c894058dSAneesh Kumar K.V while (*n) { 4416c894058dSAneesh Kumar K.V parent = *n; 441718aadd47SBobi Jam entry = rb_entry(parent, struct ext4_free_data, efd_node); 441818aadd47SBobi Jam if (cluster < entry->efd_start_cluster) 4419c894058dSAneesh Kumar K.V n = &(*n)->rb_left; 442018aadd47SBobi Jam else if (cluster >= (entry->efd_start_cluster + entry->efd_count)) 4421c894058dSAneesh Kumar K.V n = &(*n)->rb_right; 4422c894058dSAneesh Kumar K.V else { 4423e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 442484130193STheodore Ts'o ext4_group_first_block_no(sb, group) + 442584130193STheodore Ts'o EXT4_C2B(sbi, cluster), 4426e29136f8STheodore Ts'o "Block already on to-be-freed list"); 4427c894058dSAneesh Kumar K.V return 0; 4428c9de560dSAlex Tomas } 4429c9de560dSAlex Tomas } 4430c9de560dSAlex Tomas 4431c894058dSAneesh Kumar K.V rb_link_node(new_node, parent, n); 4432c894058dSAneesh Kumar K.V rb_insert_color(new_node, &db->bb_free_root); 4433c894058dSAneesh Kumar K.V 4434c894058dSAneesh Kumar K.V /* Now try to see the extent can be merged to left and right */ 4435c894058dSAneesh Kumar K.V node = rb_prev(new_node); 4436c894058dSAneesh Kumar K.V if (node) { 443718aadd47SBobi Jam entry = rb_entry(node, struct ext4_free_data, efd_node); 4438c894058dSAneesh Kumar K.V if (can_merge(entry, new_entry)) { 443918aadd47SBobi Jam new_entry->efd_start_cluster = entry->efd_start_cluster; 444018aadd47SBobi Jam new_entry->efd_count += entry->efd_count; 4441c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 444218aadd47SBobi Jam ext4_journal_callback_del(handle, &entry->efd_jce); 444318aadd47SBobi Jam kmem_cache_free(ext4_free_data_cachep, entry); 4444c9de560dSAlex Tomas } 4445c9de560dSAlex Tomas } 4446c894058dSAneesh Kumar K.V 4447c894058dSAneesh Kumar K.V node = rb_next(new_node); 4448c894058dSAneesh Kumar K.V if (node) { 444918aadd47SBobi Jam entry = rb_entry(node, struct ext4_free_data, efd_node); 4450c894058dSAneesh Kumar K.V if (can_merge(new_entry, entry)) { 445118aadd47SBobi Jam new_entry->efd_count += entry->efd_count; 4452c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 445318aadd47SBobi Jam ext4_journal_callback_del(handle, &entry->efd_jce); 445418aadd47SBobi Jam kmem_cache_free(ext4_free_data_cachep, entry); 4455c894058dSAneesh Kumar K.V } 4456c894058dSAneesh Kumar K.V } 44573e624fc7STheodore Ts'o /* Add the extent to transaction's private list */ 445818aadd47SBobi Jam ext4_journal_callback_add(handle, ext4_free_data_callback, 445918aadd47SBobi Jam &new_entry->efd_jce); 4460c9de560dSAlex Tomas return 0; 4461c9de560dSAlex Tomas } 4462c9de560dSAlex Tomas 446344338711STheodore Ts'o /** 446444338711STheodore Ts'o * ext4_free_blocks() -- Free given blocks and update quota 446544338711STheodore Ts'o * @handle: handle for this transaction 446644338711STheodore Ts'o * @inode: inode 446744338711STheodore Ts'o * @block: start physical block to free 446844338711STheodore Ts'o * @count: number of blocks to count 44695def1360SYongqiang Yang * @flags: flags used by ext4_free_blocks 4470c9de560dSAlex Tomas */ 447144338711STheodore Ts'o void ext4_free_blocks(handle_t *handle, struct inode *inode, 4472e6362609STheodore Ts'o struct buffer_head *bh, ext4_fsblk_t block, 4473e6362609STheodore Ts'o unsigned long count, int flags) 4474c9de560dSAlex Tomas { 447526346ff6SAneesh Kumar K.V struct buffer_head *bitmap_bh = NULL; 4476c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 4477c9de560dSAlex Tomas struct ext4_group_desc *gdp; 447844338711STheodore Ts'o unsigned long freed = 0; 4479498e5f24STheodore Ts'o unsigned int overflow; 4480c9de560dSAlex Tomas ext4_grpblk_t bit; 4481c9de560dSAlex Tomas struct buffer_head *gd_bh; 4482c9de560dSAlex Tomas ext4_group_t block_group; 4483c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4484c9de560dSAlex Tomas struct ext4_buddy e4b; 448584130193STheodore Ts'o unsigned int count_clusters; 4486c9de560dSAlex Tomas int err = 0; 4487c9de560dSAlex Tomas int ret; 4488c9de560dSAlex Tomas 4489e6362609STheodore Ts'o if (bh) { 4490e6362609STheodore Ts'o if (block) 4491e6362609STheodore Ts'o BUG_ON(block != bh->b_blocknr); 4492e6362609STheodore Ts'o else 4493e6362609STheodore Ts'o block = bh->b_blocknr; 4494e6362609STheodore Ts'o } 4495c9de560dSAlex Tomas 4496c9de560dSAlex Tomas sbi = EXT4_SB(sb); 44971f2acb60STheodore Ts'o if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 44981f2acb60STheodore Ts'o !ext4_data_block_valid(sbi, block, count)) { 449912062dddSEric Sandeen ext4_error(sb, "Freeing blocks not in datazone - " 45000610b6e9STheodore Ts'o "block = %llu, count = %lu", block, count); 4501c9de560dSAlex Tomas goto error_return; 4502c9de560dSAlex Tomas } 4503c9de560dSAlex Tomas 45040610b6e9STheodore Ts'o ext4_debug("freeing block %llu\n", block); 4505e6362609STheodore Ts'o trace_ext4_free_blocks(inode, block, count, flags); 4506e6362609STheodore Ts'o 4507e6362609STheodore Ts'o if (flags & EXT4_FREE_BLOCKS_FORGET) { 4508e6362609STheodore Ts'o struct buffer_head *tbh = bh; 4509e6362609STheodore Ts'o int i; 4510e6362609STheodore Ts'o 4511e6362609STheodore Ts'o BUG_ON(bh && (count > 1)); 4512e6362609STheodore Ts'o 4513e6362609STheodore Ts'o for (i = 0; i < count; i++) { 4514e6362609STheodore Ts'o if (!bh) 4515e6362609STheodore Ts'o tbh = sb_find_get_block(inode->i_sb, 4516e6362609STheodore Ts'o block + i); 451787783690SNamhyung Kim if (unlikely(!tbh)) 451887783690SNamhyung Kim continue; 4519e6362609STheodore Ts'o ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, 4520e6362609STheodore Ts'o inode, tbh, block + i); 4521e6362609STheodore Ts'o } 4522e6362609STheodore Ts'o } 4523e6362609STheodore Ts'o 4524e6362609STheodore Ts'o /* 4525e6362609STheodore Ts'o * We need to make sure we don't reuse the freed block until 4526e6362609STheodore Ts'o * after the transaction is committed, which we can do by 4527e6362609STheodore Ts'o * treating the block as metadata, below. We make an 4528e6362609STheodore Ts'o * exception if the inode is to be written in writeback mode 4529e6362609STheodore Ts'o * since writeback mode has weak data consistency guarantees. 4530e6362609STheodore Ts'o */ 4531e6362609STheodore Ts'o if (!ext4_should_writeback_data(inode)) 4532e6362609STheodore Ts'o flags |= EXT4_FREE_BLOCKS_METADATA; 4533c9de560dSAlex Tomas 453484130193STheodore Ts'o /* 453584130193STheodore Ts'o * If the extent to be freed does not begin on a cluster 453684130193STheodore Ts'o * boundary, we need to deal with partial clusters at the 453784130193STheodore Ts'o * beginning and end of the extent. Normally we will free 453884130193STheodore Ts'o * blocks at the beginning or the end unless we are explicitly 453984130193STheodore Ts'o * requested to avoid doing so. 454084130193STheodore Ts'o */ 454184130193STheodore Ts'o overflow = block & (sbi->s_cluster_ratio - 1); 454284130193STheodore Ts'o if (overflow) { 454384130193STheodore Ts'o if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) { 454484130193STheodore Ts'o overflow = sbi->s_cluster_ratio - overflow; 454584130193STheodore Ts'o block += overflow; 454684130193STheodore Ts'o if (count > overflow) 454784130193STheodore Ts'o count -= overflow; 454884130193STheodore Ts'o else 454984130193STheodore Ts'o return; 455084130193STheodore Ts'o } else { 455184130193STheodore Ts'o block -= overflow; 455284130193STheodore Ts'o count += overflow; 455384130193STheodore Ts'o } 455484130193STheodore Ts'o } 455584130193STheodore Ts'o overflow = count & (sbi->s_cluster_ratio - 1); 455684130193STheodore Ts'o if (overflow) { 455784130193STheodore Ts'o if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) { 455884130193STheodore Ts'o if (count > overflow) 455984130193STheodore Ts'o count -= overflow; 456084130193STheodore Ts'o else 456184130193STheodore Ts'o return; 456284130193STheodore Ts'o } else 456384130193STheodore Ts'o count += sbi->s_cluster_ratio - overflow; 456484130193STheodore Ts'o } 456584130193STheodore Ts'o 4566c9de560dSAlex Tomas do_more: 4567c9de560dSAlex Tomas overflow = 0; 4568c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 4569c9de560dSAlex Tomas 4570c9de560dSAlex Tomas /* 4571c9de560dSAlex Tomas * Check to see if we are freeing blocks across a group 4572c9de560dSAlex Tomas * boundary. 4573c9de560dSAlex Tomas */ 457484130193STheodore Ts'o if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) { 457584130193STheodore Ts'o overflow = EXT4_C2B(sbi, bit) + count - 457684130193STheodore Ts'o EXT4_BLOCKS_PER_GROUP(sb); 4577c9de560dSAlex Tomas count -= overflow; 4578c9de560dSAlex Tomas } 457984130193STheodore Ts'o count_clusters = EXT4_B2C(sbi, count); 4580574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, block_group); 4581ce89f46cSAneesh Kumar K.V if (!bitmap_bh) { 4582ce89f46cSAneesh Kumar K.V err = -EIO; 4583c9de560dSAlex Tomas goto error_return; 4584ce89f46cSAneesh Kumar K.V } 4585c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 4586ce89f46cSAneesh Kumar K.V if (!gdp) { 4587ce89f46cSAneesh Kumar K.V err = -EIO; 4588c9de560dSAlex Tomas goto error_return; 4589ce89f46cSAneesh Kumar K.V } 4590c9de560dSAlex Tomas 4591c9de560dSAlex Tomas if (in_range(ext4_block_bitmap(sb, gdp), block, count) || 4592c9de560dSAlex Tomas in_range(ext4_inode_bitmap(sb, gdp), block, count) || 4593c9de560dSAlex Tomas in_range(block, ext4_inode_table(sb, gdp), 4594c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group) || 4595c9de560dSAlex Tomas in_range(block + count - 1, ext4_inode_table(sb, gdp), 4596c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group)) { 4597c9de560dSAlex Tomas 459812062dddSEric Sandeen ext4_error(sb, "Freeing blocks in system zone - " 45990610b6e9STheodore Ts'o "Block = %llu, count = %lu", block, count); 4600519deca0SAneesh Kumar K.V /* err = 0. ext4_std_error should be a no op */ 4601519deca0SAneesh Kumar K.V goto error_return; 4602c9de560dSAlex Tomas } 4603c9de560dSAlex Tomas 4604c9de560dSAlex Tomas BUFFER_TRACE(bitmap_bh, "getting write access"); 4605c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 4606c9de560dSAlex Tomas if (err) 4607c9de560dSAlex Tomas goto error_return; 4608c9de560dSAlex Tomas 4609c9de560dSAlex Tomas /* 4610c9de560dSAlex Tomas * We are about to modify some metadata. Call the journal APIs 4611c9de560dSAlex Tomas * to unshare ->b_data if a currently-committing transaction is 4612c9de560dSAlex Tomas * using it 4613c9de560dSAlex Tomas */ 4614c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "get_write_access"); 4615c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gd_bh); 4616c9de560dSAlex Tomas if (err) 4617c9de560dSAlex Tomas goto error_return; 4618c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 4619c9de560dSAlex Tomas { 4620c9de560dSAlex Tomas int i; 462184130193STheodore Ts'o for (i = 0; i < count_clusters; i++) 4622c9de560dSAlex Tomas BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 4623c9de560dSAlex Tomas } 4624c9de560dSAlex Tomas #endif 462584130193STheodore Ts'o trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters); 4626c9de560dSAlex Tomas 4627920313a7SAneesh Kumar K.V err = ext4_mb_load_buddy(sb, block_group, &e4b); 4628920313a7SAneesh Kumar K.V if (err) 4629920313a7SAneesh Kumar K.V goto error_return; 4630e6362609STheodore Ts'o 4631e6362609STheodore Ts'o if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { 46327a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry; 46337a2fcbf7SAneesh Kumar K.V /* 46347a2fcbf7SAneesh Kumar K.V * blocks being freed are metadata. these blocks shouldn't 46357a2fcbf7SAneesh Kumar K.V * be used until this transaction is committed 46367a2fcbf7SAneesh Kumar K.V */ 463718aadd47SBobi Jam new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS); 4638b72143abSTheodore Ts'o if (!new_entry) { 4639b72143abSTheodore Ts'o err = -ENOMEM; 4640b72143abSTheodore Ts'o goto error_return; 4641b72143abSTheodore Ts'o } 464218aadd47SBobi Jam new_entry->efd_start_cluster = bit; 464318aadd47SBobi Jam new_entry->efd_group = block_group; 464418aadd47SBobi Jam new_entry->efd_count = count_clusters; 464518aadd47SBobi Jam new_entry->efd_tid = handle->h_transaction->t_tid; 4646955ce5f5SAneesh Kumar K.V 46477a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, block_group); 464884130193STheodore Ts'o mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 46497a2fcbf7SAneesh Kumar K.V ext4_mb_free_metadata(handle, &e4b, new_entry); 4650c9de560dSAlex Tomas } else { 46517a2fcbf7SAneesh Kumar K.V /* need to update group_info->bb_free and bitmap 46527a2fcbf7SAneesh Kumar K.V * with group lock held. generate_buddy look at 46537a2fcbf7SAneesh Kumar K.V * them with group lock_held 46547a2fcbf7SAneesh Kumar K.V */ 4655955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 465684130193STheodore Ts'o mb_clear_bits(bitmap_bh->b_data, bit, count_clusters); 465784130193STheodore Ts'o mb_free_blocks(inode, &e4b, bit, count_clusters); 4658c9de560dSAlex Tomas } 4659c9de560dSAlex Tomas 4660021b65bbSTheodore Ts'o ret = ext4_free_group_clusters(sb, gdp) + count_clusters; 4661021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, gdp, ret); 4662c9de560dSAlex Tomas gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); 4663955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 466457042651STheodore Ts'o percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters); 4665c9de560dSAlex Tomas 4666772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 4667772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 466824aaa8efSTheodore Ts'o atomic_add(count_clusters, 466924aaa8efSTheodore Ts'o &sbi->s_flex_groups[flex_group].free_clusters); 4670772cb7c8SJose R. Santos } 4671772cb7c8SJose R. Santos 4672e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 4673c9de560dSAlex Tomas 467444338711STheodore Ts'o freed += count; 4675c9de560dSAlex Tomas 46767b415bf6SAditya Kali if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE)) 46777b415bf6SAditya Kali dquot_free_block(inode, EXT4_C2B(sbi, count_clusters)); 46787b415bf6SAditya Kali 46797a2fcbf7SAneesh Kumar K.V /* We dirtied the bitmap block */ 46807a2fcbf7SAneesh Kumar K.V BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 46817a2fcbf7SAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 46827a2fcbf7SAneesh Kumar K.V 4683c9de560dSAlex Tomas /* And the group descriptor block */ 4684c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 46850390131bSFrank Mayhar ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 4686c9de560dSAlex Tomas if (!err) 4687c9de560dSAlex Tomas err = ret; 4688c9de560dSAlex Tomas 4689c9de560dSAlex Tomas if (overflow && !err) { 4690c9de560dSAlex Tomas block += count; 4691c9de560dSAlex Tomas count = overflow; 4692c9de560dSAlex Tomas put_bh(bitmap_bh); 4693c9de560dSAlex Tomas goto do_more; 4694c9de560dSAlex Tomas } 4695a0375156STheodore Ts'o ext4_mark_super_dirty(sb); 4696c9de560dSAlex Tomas error_return: 4697c9de560dSAlex Tomas brelse(bitmap_bh); 4698c9de560dSAlex Tomas ext4_std_error(sb, err); 4699c9de560dSAlex Tomas return; 4700c9de560dSAlex Tomas } 47017360d173SLukas Czerner 47027360d173SLukas Czerner /** 47030529155eSYongqiang Yang * ext4_group_add_blocks() -- Add given blocks to an existing group 47042846e820SAmir Goldstein * @handle: handle to this transaction 47052846e820SAmir Goldstein * @sb: super block 47062846e820SAmir Goldstein * @block: start physcial block to add to the block group 47072846e820SAmir Goldstein * @count: number of blocks to free 47082846e820SAmir Goldstein * 4709e73a347bSAmir Goldstein * This marks the blocks as free in the bitmap and buddy. 47102846e820SAmir Goldstein */ 4711cc7365dfSYongqiang Yang int ext4_group_add_blocks(handle_t *handle, struct super_block *sb, 47122846e820SAmir Goldstein ext4_fsblk_t block, unsigned long count) 47132846e820SAmir Goldstein { 47142846e820SAmir Goldstein struct buffer_head *bitmap_bh = NULL; 47152846e820SAmir Goldstein struct buffer_head *gd_bh; 47162846e820SAmir Goldstein ext4_group_t block_group; 47172846e820SAmir Goldstein ext4_grpblk_t bit; 47182846e820SAmir Goldstein unsigned int i; 47192846e820SAmir Goldstein struct ext4_group_desc *desc; 47202846e820SAmir Goldstein struct ext4_sb_info *sbi = EXT4_SB(sb); 4721e73a347bSAmir Goldstein struct ext4_buddy e4b; 47222846e820SAmir Goldstein int err = 0, ret, blk_free_count; 47232846e820SAmir Goldstein ext4_grpblk_t blocks_freed; 47242846e820SAmir Goldstein 47252846e820SAmir Goldstein ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1); 47262846e820SAmir Goldstein 47274740b830SYongqiang Yang if (count == 0) 47284740b830SYongqiang Yang return 0; 47294740b830SYongqiang Yang 47302846e820SAmir Goldstein ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 47312846e820SAmir Goldstein /* 47322846e820SAmir Goldstein * Check to see if we are freeing blocks across a group 47332846e820SAmir Goldstein * boundary. 47342846e820SAmir Goldstein */ 4735cc7365dfSYongqiang Yang if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { 4736cc7365dfSYongqiang Yang ext4_warning(sb, "too much blocks added to group %u\n", 4737cc7365dfSYongqiang Yang block_group); 4738cc7365dfSYongqiang Yang err = -EINVAL; 47392846e820SAmir Goldstein goto error_return; 4740cc7365dfSYongqiang Yang } 47412cd05cc3STheodore Ts'o 47422846e820SAmir Goldstein bitmap_bh = ext4_read_block_bitmap(sb, block_group); 4743cc7365dfSYongqiang Yang if (!bitmap_bh) { 4744cc7365dfSYongqiang Yang err = -EIO; 47452846e820SAmir Goldstein goto error_return; 4746cc7365dfSYongqiang Yang } 4747cc7365dfSYongqiang Yang 47482846e820SAmir Goldstein desc = ext4_get_group_desc(sb, block_group, &gd_bh); 4749cc7365dfSYongqiang Yang if (!desc) { 4750cc7365dfSYongqiang Yang err = -EIO; 47512846e820SAmir Goldstein goto error_return; 4752cc7365dfSYongqiang Yang } 47532846e820SAmir Goldstein 47542846e820SAmir Goldstein if (in_range(ext4_block_bitmap(sb, desc), block, count) || 47552846e820SAmir Goldstein in_range(ext4_inode_bitmap(sb, desc), block, count) || 47562846e820SAmir Goldstein in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) || 47572846e820SAmir Goldstein in_range(block + count - 1, ext4_inode_table(sb, desc), 47582846e820SAmir Goldstein sbi->s_itb_per_group)) { 47592846e820SAmir Goldstein ext4_error(sb, "Adding blocks in system zones - " 47602846e820SAmir Goldstein "Block = %llu, count = %lu", 47612846e820SAmir Goldstein block, count); 4762cc7365dfSYongqiang Yang err = -EINVAL; 47632846e820SAmir Goldstein goto error_return; 47642846e820SAmir Goldstein } 47652846e820SAmir Goldstein 47662cd05cc3STheodore Ts'o BUFFER_TRACE(bitmap_bh, "getting write access"); 47672cd05cc3STheodore Ts'o err = ext4_journal_get_write_access(handle, bitmap_bh); 47682846e820SAmir Goldstein if (err) 47692846e820SAmir Goldstein goto error_return; 47702846e820SAmir Goldstein 47712846e820SAmir Goldstein /* 47722846e820SAmir Goldstein * We are about to modify some metadata. Call the journal APIs 47732846e820SAmir Goldstein * to unshare ->b_data if a currently-committing transaction is 47742846e820SAmir Goldstein * using it 47752846e820SAmir Goldstein */ 47762846e820SAmir Goldstein BUFFER_TRACE(gd_bh, "get_write_access"); 47772846e820SAmir Goldstein err = ext4_journal_get_write_access(handle, gd_bh); 47782846e820SAmir Goldstein if (err) 47792846e820SAmir Goldstein goto error_return; 4780e73a347bSAmir Goldstein 47812846e820SAmir Goldstein for (i = 0, blocks_freed = 0; i < count; i++) { 47822846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "clear bit"); 4783e73a347bSAmir Goldstein if (!mb_test_bit(bit + i, bitmap_bh->b_data)) { 47842846e820SAmir Goldstein ext4_error(sb, "bit already cleared for block %llu", 47852846e820SAmir Goldstein (ext4_fsblk_t)(block + i)); 47862846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "bit already cleared"); 47872846e820SAmir Goldstein } else { 47882846e820SAmir Goldstein blocks_freed++; 47892846e820SAmir Goldstein } 47902846e820SAmir Goldstein } 4791e73a347bSAmir Goldstein 4792e73a347bSAmir Goldstein err = ext4_mb_load_buddy(sb, block_group, &e4b); 4793e73a347bSAmir Goldstein if (err) 4794e73a347bSAmir Goldstein goto error_return; 4795e73a347bSAmir Goldstein 4796e73a347bSAmir Goldstein /* 4797e73a347bSAmir Goldstein * need to update group_info->bb_free and bitmap 4798e73a347bSAmir Goldstein * with group lock held. generate_buddy look at 4799e73a347bSAmir Goldstein * them with group lock_held 4800e73a347bSAmir Goldstein */ 48012846e820SAmir Goldstein ext4_lock_group(sb, block_group); 4802e73a347bSAmir Goldstein mb_clear_bits(bitmap_bh->b_data, bit, count); 4803e73a347bSAmir Goldstein mb_free_blocks(NULL, &e4b, bit, count); 4804021b65bbSTheodore Ts'o blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc); 4805021b65bbSTheodore Ts'o ext4_free_group_clusters_set(sb, desc, blk_free_count); 48062846e820SAmir Goldstein desc->bg_checksum = ext4_group_desc_csum(sbi, block_group, desc); 48072846e820SAmir Goldstein ext4_unlock_group(sb, block_group); 480857042651STheodore Ts'o percpu_counter_add(&sbi->s_freeclusters_counter, 480957042651STheodore Ts'o EXT4_B2C(sbi, blocks_freed)); 48102846e820SAmir Goldstein 48112846e820SAmir Goldstein if (sbi->s_log_groups_per_flex) { 48122846e820SAmir Goldstein ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 481324aaa8efSTheodore Ts'o atomic_add(EXT4_B2C(sbi, blocks_freed), 481424aaa8efSTheodore Ts'o &sbi->s_flex_groups[flex_group].free_clusters); 48152846e820SAmir Goldstein } 4816e73a347bSAmir Goldstein 4817e73a347bSAmir Goldstein ext4_mb_unload_buddy(&e4b); 48182846e820SAmir Goldstein 48192846e820SAmir Goldstein /* We dirtied the bitmap block */ 48202846e820SAmir Goldstein BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 48212846e820SAmir Goldstein err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 48222846e820SAmir Goldstein 48232846e820SAmir Goldstein /* And the group descriptor block */ 48242846e820SAmir Goldstein BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 48252846e820SAmir Goldstein ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 48262846e820SAmir Goldstein if (!err) 48272846e820SAmir Goldstein err = ret; 48282846e820SAmir Goldstein 48292846e820SAmir Goldstein error_return: 48302846e820SAmir Goldstein brelse(bitmap_bh); 48312846e820SAmir Goldstein ext4_std_error(sb, err); 4832cc7365dfSYongqiang Yang return err; 48332846e820SAmir Goldstein } 48342846e820SAmir Goldstein 48352846e820SAmir Goldstein /** 48367360d173SLukas Czerner * ext4_trim_extent -- function to TRIM one single free extent in the group 48377360d173SLukas Czerner * @sb: super block for the file system 48387360d173SLukas Czerner * @start: starting block of the free extent in the alloc. group 48397360d173SLukas Czerner * @count: number of blocks to TRIM 48407360d173SLukas Czerner * @group: alloc. group we are working with 48417360d173SLukas Czerner * @e4b: ext4 buddy for the group 48427360d173SLukas Czerner * 48437360d173SLukas Czerner * Trim "count" blocks starting at "start" in the "group". To assure that no 48447360d173SLukas Czerner * one will allocate those blocks, mark it as used in buddy bitmap. This must 48457360d173SLukas Czerner * be called with under the group lock. 48467360d173SLukas Czerner */ 4847d9f34504STheodore Ts'o static void ext4_trim_extent(struct super_block *sb, int start, int count, 48487360d173SLukas Czerner ext4_group_t group, struct ext4_buddy *e4b) 48497360d173SLukas Czerner { 48507360d173SLukas Czerner struct ext4_free_extent ex; 48517360d173SLukas Czerner 4852b3d4c2b1STao Ma trace_ext4_trim_extent(sb, group, start, count); 4853b3d4c2b1STao Ma 48547360d173SLukas Czerner assert_spin_locked(ext4_group_lock_ptr(sb, group)); 48557360d173SLukas Czerner 48567360d173SLukas Czerner ex.fe_start = start; 48577360d173SLukas Czerner ex.fe_group = group; 48587360d173SLukas Czerner ex.fe_len = count; 48597360d173SLukas Czerner 48607360d173SLukas Czerner /* 48617360d173SLukas Czerner * Mark blocks used, so no one can reuse them while 48627360d173SLukas Czerner * being trimmed. 48637360d173SLukas Czerner */ 48647360d173SLukas Czerner mb_mark_used(e4b, &ex); 48657360d173SLukas Czerner ext4_unlock_group(sb, group); 4866d9f34504STheodore Ts'o ext4_issue_discard(sb, group, start, count); 48677360d173SLukas Czerner ext4_lock_group(sb, group); 48687360d173SLukas Czerner mb_free_blocks(NULL, e4b, start, ex.fe_len); 48697360d173SLukas Czerner } 48707360d173SLukas Czerner 48717360d173SLukas Czerner /** 48727360d173SLukas Czerner * ext4_trim_all_free -- function to trim all free space in alloc. group 48737360d173SLukas Czerner * @sb: super block for file system 487422612283STao Ma * @group: group to be trimmed 48757360d173SLukas Czerner * @start: first group block to examine 48767360d173SLukas Czerner * @max: last group block to examine 48777360d173SLukas Czerner * @minblocks: minimum extent block count 48787360d173SLukas Czerner * 48797360d173SLukas Czerner * ext4_trim_all_free walks through group's buddy bitmap searching for free 48807360d173SLukas Czerner * extents. When the free block is found, ext4_trim_extent is called to TRIM 48817360d173SLukas Czerner * the extent. 48827360d173SLukas Czerner * 48837360d173SLukas Czerner * 48847360d173SLukas Czerner * ext4_trim_all_free walks through group's block bitmap searching for free 48857360d173SLukas Czerner * extents. When the free extent is found, mark it as used in group buddy 48867360d173SLukas Czerner * bitmap. Then issue a TRIM command on this extent and free the extent in 48877360d173SLukas Czerner * the group buddy bitmap. This is done until whole group is scanned. 48887360d173SLukas Czerner */ 48890b75a840SLukas Czerner static ext4_grpblk_t 489078944086SLukas Czerner ext4_trim_all_free(struct super_block *sb, ext4_group_t group, 489178944086SLukas Czerner ext4_grpblk_t start, ext4_grpblk_t max, 489278944086SLukas Czerner ext4_grpblk_t minblocks) 48937360d173SLukas Czerner { 48947360d173SLukas Czerner void *bitmap; 4895169ddc3eSTao Ma ext4_grpblk_t next, count = 0, free_count = 0; 489678944086SLukas Czerner struct ext4_buddy e4b; 489778944086SLukas Czerner int ret; 48987360d173SLukas Czerner 4899b3d4c2b1STao Ma trace_ext4_trim_all_free(sb, group, start, max); 4900b3d4c2b1STao Ma 490178944086SLukas Czerner ret = ext4_mb_load_buddy(sb, group, &e4b); 490278944086SLukas Czerner if (ret) { 490378944086SLukas Czerner ext4_error(sb, "Error in loading buddy " 490478944086SLukas Czerner "information for %u", group); 490578944086SLukas Czerner return ret; 490678944086SLukas Czerner } 490778944086SLukas Czerner bitmap = e4b.bd_bitmap; 490828739eeaSLukas Czerner 490928739eeaSLukas Czerner ext4_lock_group(sb, group); 49103d56b8d2STao Ma if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) && 49113d56b8d2STao Ma minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) 49123d56b8d2STao Ma goto out; 49133d56b8d2STao Ma 491478944086SLukas Czerner start = (e4b.bd_info->bb_first_free > start) ? 491578944086SLukas Czerner e4b.bd_info->bb_first_free : start; 49167360d173SLukas Czerner 4917913eed83SLukas Czerner while (start <= max) { 4918913eed83SLukas Czerner start = mb_find_next_zero_bit(bitmap, max + 1, start); 4919913eed83SLukas Czerner if (start > max) 49207360d173SLukas Czerner break; 4921913eed83SLukas Czerner next = mb_find_next_bit(bitmap, max + 1, start); 49227360d173SLukas Czerner 49237360d173SLukas Czerner if ((next - start) >= minblocks) { 4924d9f34504STheodore Ts'o ext4_trim_extent(sb, start, 492578944086SLukas Czerner next - start, group, &e4b); 49267360d173SLukas Czerner count += next - start; 49277360d173SLukas Czerner } 4928169ddc3eSTao Ma free_count += next - start; 49297360d173SLukas Czerner start = next + 1; 49307360d173SLukas Czerner 49317360d173SLukas Czerner if (fatal_signal_pending(current)) { 49327360d173SLukas Czerner count = -ERESTARTSYS; 49337360d173SLukas Czerner break; 49347360d173SLukas Czerner } 49357360d173SLukas Czerner 49367360d173SLukas Czerner if (need_resched()) { 49377360d173SLukas Czerner ext4_unlock_group(sb, group); 49387360d173SLukas Czerner cond_resched(); 49397360d173SLukas Czerner ext4_lock_group(sb, group); 49407360d173SLukas Czerner } 49417360d173SLukas Czerner 4942169ddc3eSTao Ma if ((e4b.bd_info->bb_free - free_count) < minblocks) 49437360d173SLukas Czerner break; 49447360d173SLukas Czerner } 49453d56b8d2STao Ma 49463d56b8d2STao Ma if (!ret) 49473d56b8d2STao Ma EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info); 49483d56b8d2STao Ma out: 49497360d173SLukas Czerner ext4_unlock_group(sb, group); 495078944086SLukas Czerner ext4_mb_unload_buddy(&e4b); 49517360d173SLukas Czerner 49527360d173SLukas Czerner ext4_debug("trimmed %d blocks in the group %d\n", 49537360d173SLukas Czerner count, group); 49547360d173SLukas Czerner 49557360d173SLukas Czerner return count; 49567360d173SLukas Czerner } 49577360d173SLukas Czerner 49587360d173SLukas Czerner /** 49597360d173SLukas Czerner * ext4_trim_fs() -- trim ioctl handle function 49607360d173SLukas Czerner * @sb: superblock for filesystem 49617360d173SLukas Czerner * @range: fstrim_range structure 49627360d173SLukas Czerner * 49637360d173SLukas Czerner * start: First Byte to trim 49647360d173SLukas Czerner * len: number of Bytes to trim from start 49657360d173SLukas Czerner * minlen: minimum extent length in Bytes 49667360d173SLukas Czerner * ext4_trim_fs goes through all allocation groups containing Bytes from 49677360d173SLukas Czerner * start to start+len. For each such a group ext4_trim_all_free function 49687360d173SLukas Czerner * is invoked to trim all free space. 49697360d173SLukas Czerner */ 49707360d173SLukas Czerner int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) 49717360d173SLukas Czerner { 497278944086SLukas Czerner struct ext4_group_info *grp; 4973913eed83SLukas Czerner ext4_group_t group, first_group, last_group; 49747137d7a4STheodore Ts'o ext4_grpblk_t cnt = 0, first_cluster, last_cluster; 4975913eed83SLukas Czerner uint64_t start, end, minlen, trimmed = 0; 49760f0a25bfSJan Kara ext4_fsblk_t first_data_blk = 49770f0a25bfSJan Kara le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 4978913eed83SLukas Czerner ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es); 49797360d173SLukas Czerner int ret = 0; 49807360d173SLukas Czerner 49817360d173SLukas Czerner start = range->start >> sb->s_blocksize_bits; 4982913eed83SLukas Czerner end = start + (range->len >> sb->s_blocksize_bits) - 1; 49837360d173SLukas Czerner minlen = range->minlen >> sb->s_blocksize_bits; 49847360d173SLukas Czerner 4985913eed83SLukas Czerner if (unlikely(minlen > EXT4_CLUSTERS_PER_GROUP(sb)) || 4986913eed83SLukas Czerner unlikely(start >= max_blks)) 49877360d173SLukas Czerner return -EINVAL; 4988913eed83SLukas Czerner if (end >= max_blks) 4989913eed83SLukas Czerner end = max_blks - 1; 4990913eed83SLukas Czerner if (end <= first_data_blk) 499122f10457STao Ma goto out; 4992913eed83SLukas Czerner if (start < first_data_blk) 49930f0a25bfSJan Kara start = first_data_blk; 49947360d173SLukas Czerner 4995913eed83SLukas Czerner /* Determine first and last group to examine based on start and end */ 49967360d173SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, 49977137d7a4STheodore Ts'o &first_group, &first_cluster); 4998913eed83SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end, 49997137d7a4STheodore Ts'o &last_group, &last_cluster); 50007360d173SLukas Czerner 5001913eed83SLukas Czerner /* end now represents the last cluster to discard in this group */ 5002913eed83SLukas Czerner end = EXT4_CLUSTERS_PER_GROUP(sb) - 1; 50037360d173SLukas Czerner 50047360d173SLukas Czerner for (group = first_group; group <= last_group; group++) { 500578944086SLukas Czerner grp = ext4_get_group_info(sb, group); 500678944086SLukas Czerner /* We only do this if the grp has never been initialized */ 500778944086SLukas Czerner if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 500878944086SLukas Czerner ret = ext4_mb_init_group(sb, group); 500978944086SLukas Czerner if (ret) 50107360d173SLukas Czerner break; 50117360d173SLukas Czerner } 50127360d173SLukas Czerner 50130ba08517STao Ma /* 5014913eed83SLukas Czerner * For all the groups except the last one, last cluster will 5015913eed83SLukas Czerner * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to 5016913eed83SLukas Czerner * change it for the last group, note that last_cluster is 5017913eed83SLukas Czerner * already computed earlier by ext4_get_group_no_and_offset() 50180ba08517STao Ma */ 5019913eed83SLukas Czerner if (group == last_group) 5020913eed83SLukas Czerner end = last_cluster; 50217360d173SLukas Czerner 502278944086SLukas Czerner if (grp->bb_free >= minlen) { 50237137d7a4STheodore Ts'o cnt = ext4_trim_all_free(sb, group, first_cluster, 5024913eed83SLukas Czerner end, minlen); 50257360d173SLukas Czerner if (cnt < 0) { 50267360d173SLukas Czerner ret = cnt; 50277360d173SLukas Czerner break; 50287360d173SLukas Czerner } 50297360d173SLukas Czerner trimmed += cnt; 503021e7fd22SLukas Czerner } 5031913eed83SLukas Czerner 5032913eed83SLukas Czerner /* 5033913eed83SLukas Czerner * For every group except the first one, we are sure 5034913eed83SLukas Czerner * that the first cluster to discard will be cluster #0. 5035913eed83SLukas Czerner */ 50367137d7a4STheodore Ts'o first_cluster = 0; 50377360d173SLukas Czerner } 50387360d173SLukas Czerner range->len = trimmed * sb->s_blocksize; 50397360d173SLukas Czerner 50403d56b8d2STao Ma if (!ret) 50413d56b8d2STao Ma atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen); 50423d56b8d2STao Ma 504322f10457STao Ma out: 50447360d173SLukas Czerner return ret; 50457360d173SLukas Czerner } 5046