1c9de560dSAlex Tomas /* 2c9de560dSAlex Tomas * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com 3c9de560dSAlex Tomas * Written by Alex Tomas <alex@clusterfs.com> 4c9de560dSAlex Tomas * 5c9de560dSAlex Tomas * This program is free software; you can redistribute it and/or modify 6c9de560dSAlex Tomas * it under the terms of the GNU General Public License version 2 as 7c9de560dSAlex Tomas * published by the Free Software Foundation. 8c9de560dSAlex Tomas * 9c9de560dSAlex Tomas * This program is distributed in the hope that it will be useful, 10c9de560dSAlex Tomas * but WITHOUT ANY WARRANTY; without even the implied warranty of 11c9de560dSAlex Tomas * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12c9de560dSAlex Tomas * GNU General Public License for more details. 13c9de560dSAlex Tomas * 14c9de560dSAlex Tomas * You should have received a copy of the GNU General Public Licens 15c9de560dSAlex Tomas * along with this program; if not, write to the Free Software 16c9de560dSAlex Tomas * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- 17c9de560dSAlex Tomas */ 18c9de560dSAlex Tomas 19c9de560dSAlex Tomas 20c9de560dSAlex Tomas /* 21c9de560dSAlex Tomas * mballoc.c contains the multiblocks allocation routines 22c9de560dSAlex Tomas */ 23c9de560dSAlex Tomas 248f6e39a7SMingming Cao #include "mballoc.h" 25c9de560dSAlex Tomas /* 26c9de560dSAlex Tomas * MUSTDO: 27c9de560dSAlex Tomas * - test ext4_ext_search_left() and ext4_ext_search_right() 28c9de560dSAlex Tomas * - search for metadata in few groups 29c9de560dSAlex Tomas * 30c9de560dSAlex Tomas * TODO v4: 31c9de560dSAlex Tomas * - normalization should take into account whether file is still open 32c9de560dSAlex Tomas * - discard preallocations if no free space left (policy?) 33c9de560dSAlex Tomas * - don't normalize tails 34c9de560dSAlex Tomas * - quota 35c9de560dSAlex Tomas * - reservation for superuser 36c9de560dSAlex Tomas * 37c9de560dSAlex Tomas * TODO v3: 38c9de560dSAlex Tomas * - bitmap read-ahead (proposed by Oleg Drokin aka green) 39c9de560dSAlex Tomas * - track min/max extents in each group for better group selection 40c9de560dSAlex Tomas * - mb_mark_used() may allocate chunk right after splitting buddy 41c9de560dSAlex Tomas * - tree of groups sorted by number of free blocks 42c9de560dSAlex Tomas * - error handling 43c9de560dSAlex Tomas */ 44c9de560dSAlex Tomas 45c9de560dSAlex Tomas /* 46c9de560dSAlex Tomas * The allocation request involve request for multiple number of blocks 47c9de560dSAlex Tomas * near to the goal(block) value specified. 48c9de560dSAlex Tomas * 49c9de560dSAlex Tomas * During initialization phase of the allocator we decide to use the group 50c9de560dSAlex Tomas * preallocation or inode preallocation depending on the size file. The 51c9de560dSAlex Tomas * size of the file could be the resulting file size we would have after 52c9de560dSAlex Tomas * allocation or the current file size which ever is larger. If the size is 53c9de560dSAlex Tomas * less that sbi->s_mb_stream_request we select the group 54c9de560dSAlex Tomas * preallocation. The default value of s_mb_stream_request is 16 55c9de560dSAlex Tomas * blocks. This can also be tuned via 56c9de560dSAlex Tomas * /proc/fs/ext4/<partition>/stream_req. The value is represented in terms 57c9de560dSAlex Tomas * of number of blocks. 58c9de560dSAlex Tomas * 59c9de560dSAlex Tomas * The main motivation for having small file use group preallocation is to 60c9de560dSAlex Tomas * ensure that we have small file closer in the disk. 61c9de560dSAlex Tomas * 62c9de560dSAlex Tomas * First stage the allocator looks at the inode prealloc list 63c9de560dSAlex Tomas * ext4_inode_info->i_prealloc_list contain list of prealloc spaces for 64c9de560dSAlex Tomas * this particular inode. The inode prealloc space is represented as: 65c9de560dSAlex Tomas * 66c9de560dSAlex Tomas * pa_lstart -> the logical start block for this prealloc space 67c9de560dSAlex Tomas * pa_pstart -> the physical start block for this prealloc space 68c9de560dSAlex Tomas * pa_len -> lenght for this prealloc space 69c9de560dSAlex Tomas * pa_free -> free space available in this prealloc space 70c9de560dSAlex Tomas * 71c9de560dSAlex Tomas * The inode preallocation space is used looking at the _logical_ start 72c9de560dSAlex Tomas * block. If only the logical file block falls within the range of prealloc 73c9de560dSAlex Tomas * space we will consume the particular prealloc space. This make sure that 74c9de560dSAlex Tomas * that the we have contiguous physical blocks representing the file blocks 75c9de560dSAlex Tomas * 76c9de560dSAlex Tomas * The important thing to be noted in case of inode prealloc space is that 77c9de560dSAlex Tomas * we don't modify the values associated to inode prealloc space except 78c9de560dSAlex Tomas * pa_free. 79c9de560dSAlex Tomas * 80c9de560dSAlex Tomas * If we are not able to find blocks in the inode prealloc space and if we 81c9de560dSAlex Tomas * have the group allocation flag set then we look at the locality group 82c9de560dSAlex Tomas * prealloc space. These are per CPU prealloc list repreasented as 83c9de560dSAlex Tomas * 84c9de560dSAlex Tomas * ext4_sb_info.s_locality_groups[smp_processor_id()] 85c9de560dSAlex Tomas * 86c9de560dSAlex Tomas * The reason for having a per cpu locality group is to reduce the contention 87c9de560dSAlex Tomas * between CPUs. It is possible to get scheduled at this point. 88c9de560dSAlex Tomas * 89c9de560dSAlex Tomas * The locality group prealloc space is used looking at whether we have 90c9de560dSAlex Tomas * enough free space (pa_free) withing the prealloc space. 91c9de560dSAlex Tomas * 92c9de560dSAlex Tomas * If we can't allocate blocks via inode prealloc or/and locality group 93c9de560dSAlex Tomas * prealloc then we look at the buddy cache. The buddy cache is represented 94c9de560dSAlex Tomas * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 95c9de560dSAlex Tomas * mapped to the buddy and bitmap information regarding different 96c9de560dSAlex Tomas * groups. The buddy information is attached to buddy cache inode so that 97c9de560dSAlex Tomas * we can access them through the page cache. The information regarding 98c9de560dSAlex Tomas * each group is loaded via ext4_mb_load_buddy. The information involve 99c9de560dSAlex Tomas * block bitmap and buddy information. The information are stored in the 100c9de560dSAlex Tomas * inode as: 101c9de560dSAlex Tomas * 102c9de560dSAlex Tomas * { page } 103c9de560dSAlex Tomas * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]... 104c9de560dSAlex Tomas * 105c9de560dSAlex Tomas * 106c9de560dSAlex Tomas * one block each for bitmap and buddy information. So for each group we 107c9de560dSAlex Tomas * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / 108c9de560dSAlex Tomas * blocksize) blocks. So it can have information regarding groups_per_page 109c9de560dSAlex Tomas * which is blocks_per_page/2 110c9de560dSAlex Tomas * 111c9de560dSAlex Tomas * The buddy cache inode is not stored on disk. The inode is thrown 112c9de560dSAlex Tomas * away when the filesystem is unmounted. 113c9de560dSAlex Tomas * 114c9de560dSAlex Tomas * We look for count number of blocks in the buddy cache. If we were able 115c9de560dSAlex Tomas * to locate that many free blocks we return with additional information 116c9de560dSAlex Tomas * regarding rest of the contiguous physical block available 117c9de560dSAlex Tomas * 118c9de560dSAlex Tomas * Before allocating blocks via buddy cache we normalize the request 119c9de560dSAlex Tomas * blocks. This ensure we ask for more blocks that we needed. The extra 120c9de560dSAlex Tomas * blocks that we get after allocation is added to the respective prealloc 121c9de560dSAlex Tomas * list. In case of inode preallocation we follow a list of heuristics 122c9de560dSAlex Tomas * based on file size. This can be found in ext4_mb_normalize_request. If 123c9de560dSAlex Tomas * we are doing a group prealloc we try to normalize the request to 124c9de560dSAlex Tomas * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is set to 125c9de560dSAlex Tomas * 512 blocks. This can be tuned via 126c9de560dSAlex Tomas * /proc/fs/ext4/<partition/group_prealloc. The value is represented in 127c9de560dSAlex Tomas * terms of number of blocks. If we have mounted the file system with -O 128c9de560dSAlex Tomas * stripe=<value> option the group prealloc request is normalized to the 129c9de560dSAlex Tomas * stripe value (sbi->s_stripe) 130c9de560dSAlex Tomas * 131c9de560dSAlex Tomas * The regular allocator(using the buddy cache) support few tunables. 132c9de560dSAlex Tomas * 133c9de560dSAlex Tomas * /proc/fs/ext4/<partition>/min_to_scan 134c9de560dSAlex Tomas * /proc/fs/ext4/<partition>/max_to_scan 135c9de560dSAlex Tomas * /proc/fs/ext4/<partition>/order2_req 136c9de560dSAlex Tomas * 137c9de560dSAlex Tomas * The regular allocator use buddy scan only if the request len is power of 138c9de560dSAlex Tomas * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 139c9de560dSAlex Tomas * value of s_mb_order2_reqs can be tuned via 140c9de560dSAlex Tomas * /proc/fs/ext4/<partition>/order2_req. If the request len is equal to 141c9de560dSAlex Tomas * stripe size (sbi->s_stripe), we try to search for contigous block in 142c9de560dSAlex Tomas * stripe size. This should result in better allocation on RAID setup. If 143c9de560dSAlex Tomas * not we search in the specific group using bitmap for best extents. The 144c9de560dSAlex Tomas * tunable min_to_scan and max_to_scan controll the behaviour here. 145c9de560dSAlex Tomas * min_to_scan indicate how long the mballoc __must__ look for a best 146c9de560dSAlex Tomas * extent and max_to_scanindicate how long the mballoc __can__ look for a 147c9de560dSAlex Tomas * best extent in the found extents. Searching for the blocks starts with 148c9de560dSAlex Tomas * the group specified as the goal value in allocation context via 149c9de560dSAlex Tomas * ac_g_ex. Each group is first checked based on the criteria whether it 150c9de560dSAlex Tomas * can used for allocation. ext4_mb_good_group explains how the groups are 151c9de560dSAlex Tomas * checked. 152c9de560dSAlex Tomas * 153c9de560dSAlex Tomas * Both the prealloc space are getting populated as above. So for the first 154c9de560dSAlex Tomas * request we will hit the buddy cache which will result in this prealloc 155c9de560dSAlex Tomas * space getting filled. The prealloc space is then later used for the 156c9de560dSAlex Tomas * subsequent request. 157c9de560dSAlex Tomas */ 158c9de560dSAlex Tomas 159c9de560dSAlex Tomas /* 160c9de560dSAlex Tomas * mballoc operates on the following data: 161c9de560dSAlex Tomas * - on-disk bitmap 162c9de560dSAlex Tomas * - in-core buddy (actually includes buddy and bitmap) 163c9de560dSAlex Tomas * - preallocation descriptors (PAs) 164c9de560dSAlex Tomas * 165c9de560dSAlex Tomas * there are two types of preallocations: 166c9de560dSAlex Tomas * - inode 167c9de560dSAlex Tomas * assiged to specific inode and can be used for this inode only. 168c9de560dSAlex Tomas * it describes part of inode's space preallocated to specific 169c9de560dSAlex Tomas * physical blocks. any block from that preallocated can be used 170c9de560dSAlex Tomas * independent. the descriptor just tracks number of blocks left 171c9de560dSAlex Tomas * unused. so, before taking some block from descriptor, one must 172c9de560dSAlex Tomas * make sure corresponded logical block isn't allocated yet. this 173c9de560dSAlex Tomas * also means that freeing any block within descriptor's range 174c9de560dSAlex Tomas * must discard all preallocated blocks. 175c9de560dSAlex Tomas * - locality group 176c9de560dSAlex Tomas * assigned to specific locality group which does not translate to 177c9de560dSAlex Tomas * permanent set of inodes: inode can join and leave group. space 178c9de560dSAlex Tomas * from this type of preallocation can be used for any inode. thus 179c9de560dSAlex Tomas * it's consumed from the beginning to the end. 180c9de560dSAlex Tomas * 181c9de560dSAlex Tomas * relation between them can be expressed as: 182c9de560dSAlex Tomas * in-core buddy = on-disk bitmap + preallocation descriptors 183c9de560dSAlex Tomas * 184c9de560dSAlex Tomas * this mean blocks mballoc considers used are: 185c9de560dSAlex Tomas * - allocated blocks (persistent) 186c9de560dSAlex Tomas * - preallocated blocks (non-persistent) 187c9de560dSAlex Tomas * 188c9de560dSAlex Tomas * consistency in mballoc world means that at any time a block is either 189c9de560dSAlex Tomas * free or used in ALL structures. notice: "any time" should not be read 190c9de560dSAlex Tomas * literally -- time is discrete and delimited by locks. 191c9de560dSAlex Tomas * 192c9de560dSAlex Tomas * to keep it simple, we don't use block numbers, instead we count number of 193c9de560dSAlex Tomas * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 194c9de560dSAlex Tomas * 195c9de560dSAlex Tomas * all operations can be expressed as: 196c9de560dSAlex Tomas * - init buddy: buddy = on-disk + PAs 197c9de560dSAlex Tomas * - new PA: buddy += N; PA = N 198c9de560dSAlex Tomas * - use inode PA: on-disk += N; PA -= N 199c9de560dSAlex Tomas * - discard inode PA buddy -= on-disk - PA; PA = 0 200c9de560dSAlex Tomas * - use locality group PA on-disk += N; PA -= N 201c9de560dSAlex Tomas * - discard locality group PA buddy -= PA; PA = 0 202c9de560dSAlex Tomas * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 203c9de560dSAlex Tomas * is used in real operation because we can't know actual used 204c9de560dSAlex Tomas * bits from PA, only from on-disk bitmap 205c9de560dSAlex Tomas * 206c9de560dSAlex Tomas * if we follow this strict logic, then all operations above should be atomic. 207c9de560dSAlex Tomas * given some of them can block, we'd have to use something like semaphores 208c9de560dSAlex Tomas * killing performance on high-end SMP hardware. let's try to relax it using 209c9de560dSAlex Tomas * the following knowledge: 210c9de560dSAlex Tomas * 1) if buddy is referenced, it's already initialized 211c9de560dSAlex Tomas * 2) while block is used in buddy and the buddy is referenced, 212c9de560dSAlex Tomas * nobody can re-allocate that block 213c9de560dSAlex Tomas * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 214c9de560dSAlex Tomas * bit set and PA claims same block, it's OK. IOW, one can set bit in 215c9de560dSAlex Tomas * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 216c9de560dSAlex Tomas * block 217c9de560dSAlex Tomas * 218c9de560dSAlex Tomas * so, now we're building a concurrency table: 219c9de560dSAlex Tomas * - init buddy vs. 220c9de560dSAlex Tomas * - new PA 221c9de560dSAlex Tomas * blocks for PA are allocated in the buddy, buddy must be referenced 222c9de560dSAlex Tomas * until PA is linked to allocation group to avoid concurrent buddy init 223c9de560dSAlex Tomas * - use inode PA 224c9de560dSAlex Tomas * we need to make sure that either on-disk bitmap or PA has uptodate data 225c9de560dSAlex Tomas * given (3) we care that PA-=N operation doesn't interfere with init 226c9de560dSAlex Tomas * - discard inode PA 227c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 228c9de560dSAlex Tomas * - use locality group PA 229c9de560dSAlex Tomas * again PA-=N must be serialized with init 230c9de560dSAlex Tomas * - discard locality group PA 231c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 232c9de560dSAlex Tomas * - new PA vs. 233c9de560dSAlex Tomas * - use inode PA 234c9de560dSAlex Tomas * i_data_sem serializes them 235c9de560dSAlex Tomas * - discard inode PA 236c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 237c9de560dSAlex Tomas * - use locality group PA 238c9de560dSAlex Tomas * some mutex should serialize them 239c9de560dSAlex Tomas * - discard locality group PA 240c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 241c9de560dSAlex Tomas * - use inode PA 242c9de560dSAlex Tomas * - use inode PA 243c9de560dSAlex Tomas * i_data_sem or another mutex should serializes them 244c9de560dSAlex Tomas * - discard inode PA 245c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 246c9de560dSAlex Tomas * - use locality group PA 247c9de560dSAlex Tomas * nothing wrong here -- they're different PAs covering different blocks 248c9de560dSAlex Tomas * - discard locality group PA 249c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 250c9de560dSAlex Tomas * 251c9de560dSAlex Tomas * now we're ready to make few consequences: 252c9de560dSAlex Tomas * - PA is referenced and while it is no discard is possible 253c9de560dSAlex Tomas * - PA is referenced until block isn't marked in on-disk bitmap 254c9de560dSAlex Tomas * - PA changes only after on-disk bitmap 255c9de560dSAlex Tomas * - discard must not compete with init. either init is done before 256c9de560dSAlex Tomas * any discard or they're serialized somehow 257c9de560dSAlex Tomas * - buddy init as sum of on-disk bitmap and PAs is done atomically 258c9de560dSAlex Tomas * 259c9de560dSAlex Tomas * a special case when we've used PA to emptiness. no need to modify buddy 260c9de560dSAlex Tomas * in this case, but we should care about concurrent init 261c9de560dSAlex Tomas * 262c9de560dSAlex Tomas */ 263c9de560dSAlex Tomas 264c9de560dSAlex Tomas /* 265c9de560dSAlex Tomas * Logic in few words: 266c9de560dSAlex Tomas * 267c9de560dSAlex Tomas * - allocation: 268c9de560dSAlex Tomas * load group 269c9de560dSAlex Tomas * find blocks 270c9de560dSAlex Tomas * mark bits in on-disk bitmap 271c9de560dSAlex Tomas * release group 272c9de560dSAlex Tomas * 273c9de560dSAlex Tomas * - use preallocation: 274c9de560dSAlex Tomas * find proper PA (per-inode or group) 275c9de560dSAlex Tomas * load group 276c9de560dSAlex Tomas * mark bits in on-disk bitmap 277c9de560dSAlex Tomas * release group 278c9de560dSAlex Tomas * release PA 279c9de560dSAlex Tomas * 280c9de560dSAlex Tomas * - free: 281c9de560dSAlex Tomas * load group 282c9de560dSAlex Tomas * mark bits in on-disk bitmap 283c9de560dSAlex Tomas * release group 284c9de560dSAlex Tomas * 285c9de560dSAlex Tomas * - discard preallocations in group: 286c9de560dSAlex Tomas * mark PAs deleted 287c9de560dSAlex Tomas * move them onto local list 288c9de560dSAlex Tomas * load on-disk bitmap 289c9de560dSAlex Tomas * load group 290c9de560dSAlex Tomas * remove PA from object (inode or locality group) 291c9de560dSAlex Tomas * mark free blocks in-core 292c9de560dSAlex Tomas * 293c9de560dSAlex Tomas * - discard inode's preallocations: 294c9de560dSAlex Tomas */ 295c9de560dSAlex Tomas 296c9de560dSAlex Tomas /* 297c9de560dSAlex Tomas * Locking rules 298c9de560dSAlex Tomas * 299c9de560dSAlex Tomas * Locks: 300c9de560dSAlex Tomas * - bitlock on a group (group) 301c9de560dSAlex Tomas * - object (inode/locality) (object) 302c9de560dSAlex Tomas * - per-pa lock (pa) 303c9de560dSAlex Tomas * 304c9de560dSAlex Tomas * Paths: 305c9de560dSAlex Tomas * - new pa 306c9de560dSAlex Tomas * object 307c9de560dSAlex Tomas * group 308c9de560dSAlex Tomas * 309c9de560dSAlex Tomas * - find and use pa: 310c9de560dSAlex Tomas * pa 311c9de560dSAlex Tomas * 312c9de560dSAlex Tomas * - release consumed pa: 313c9de560dSAlex Tomas * pa 314c9de560dSAlex Tomas * group 315c9de560dSAlex Tomas * object 316c9de560dSAlex Tomas * 317c9de560dSAlex Tomas * - generate in-core bitmap: 318c9de560dSAlex Tomas * group 319c9de560dSAlex Tomas * pa 320c9de560dSAlex Tomas * 321c9de560dSAlex Tomas * - discard all for given object (inode, locality group): 322c9de560dSAlex Tomas * object 323c9de560dSAlex Tomas * pa 324c9de560dSAlex Tomas * group 325c9de560dSAlex Tomas * 326c9de560dSAlex Tomas * - discard all for given group: 327c9de560dSAlex Tomas * group 328c9de560dSAlex Tomas * pa 329c9de560dSAlex Tomas * group 330c9de560dSAlex Tomas * object 331c9de560dSAlex Tomas * 332c9de560dSAlex Tomas */ 333c9de560dSAlex Tomas 334ffad0a44SAneesh Kumar K.V static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 335ffad0a44SAneesh Kumar K.V { 336c9de560dSAlex Tomas #if BITS_PER_LONG == 64 337ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 7UL) << 3; 338ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~7UL); 339c9de560dSAlex Tomas #elif BITS_PER_LONG == 32 340ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 3UL) << 3; 341ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~3UL); 342c9de560dSAlex Tomas #else 343c9de560dSAlex Tomas #error "how many bits you are?!" 344c9de560dSAlex Tomas #endif 345ffad0a44SAneesh Kumar K.V return addr; 346ffad0a44SAneesh Kumar K.V } 347c9de560dSAlex Tomas 348c9de560dSAlex Tomas static inline int mb_test_bit(int bit, void *addr) 349c9de560dSAlex Tomas { 350c9de560dSAlex Tomas /* 351c9de560dSAlex Tomas * ext4_test_bit on architecture like powerpc 352c9de560dSAlex Tomas * needs unsigned long aligned address 353c9de560dSAlex Tomas */ 354ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 355c9de560dSAlex Tomas return ext4_test_bit(bit, addr); 356c9de560dSAlex Tomas } 357c9de560dSAlex Tomas 358c9de560dSAlex Tomas static inline void mb_set_bit(int bit, void *addr) 359c9de560dSAlex Tomas { 360ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 361c9de560dSAlex Tomas ext4_set_bit(bit, addr); 362c9de560dSAlex Tomas } 363c9de560dSAlex Tomas 364c9de560dSAlex Tomas static inline void mb_set_bit_atomic(spinlock_t *lock, int bit, void *addr) 365c9de560dSAlex Tomas { 366ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 367c9de560dSAlex Tomas ext4_set_bit_atomic(lock, bit, addr); 368c9de560dSAlex Tomas } 369c9de560dSAlex Tomas 370c9de560dSAlex Tomas static inline void mb_clear_bit(int bit, void *addr) 371c9de560dSAlex Tomas { 372ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 373c9de560dSAlex Tomas ext4_clear_bit(bit, addr); 374c9de560dSAlex Tomas } 375c9de560dSAlex Tomas 376c9de560dSAlex Tomas static inline void mb_clear_bit_atomic(spinlock_t *lock, int bit, void *addr) 377c9de560dSAlex Tomas { 378ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 379c9de560dSAlex Tomas ext4_clear_bit_atomic(lock, bit, addr); 380c9de560dSAlex Tomas } 381c9de560dSAlex Tomas 382ffad0a44SAneesh Kumar K.V static inline int mb_find_next_zero_bit(void *addr, int max, int start) 383ffad0a44SAneesh Kumar K.V { 384e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 385ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 386e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 387ffad0a44SAneesh Kumar K.V start += fix; 388ffad0a44SAneesh Kumar K.V 389e7dfb246SAneesh Kumar K.V ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 390e7dfb246SAneesh Kumar K.V if (ret > max) 391e7dfb246SAneesh Kumar K.V return max; 392e7dfb246SAneesh Kumar K.V return ret; 393ffad0a44SAneesh Kumar K.V } 394ffad0a44SAneesh Kumar K.V 395ffad0a44SAneesh Kumar K.V static inline int mb_find_next_bit(void *addr, int max, int start) 396ffad0a44SAneesh Kumar K.V { 397e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 398ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 399e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 400ffad0a44SAneesh Kumar K.V start += fix; 401ffad0a44SAneesh Kumar K.V 402e7dfb246SAneesh Kumar K.V ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 403e7dfb246SAneesh Kumar K.V if (ret > max) 404e7dfb246SAneesh Kumar K.V return max; 405e7dfb246SAneesh Kumar K.V return ret; 406ffad0a44SAneesh Kumar K.V } 407ffad0a44SAneesh Kumar K.V 408c9de560dSAlex Tomas static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 409c9de560dSAlex Tomas { 410c9de560dSAlex Tomas char *bb; 411c9de560dSAlex Tomas 412c9de560dSAlex Tomas BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); 413c9de560dSAlex Tomas BUG_ON(max == NULL); 414c9de560dSAlex Tomas 415c9de560dSAlex Tomas if (order > e4b->bd_blkbits + 1) { 416c9de560dSAlex Tomas *max = 0; 417c9de560dSAlex Tomas return NULL; 418c9de560dSAlex Tomas } 419c9de560dSAlex Tomas 420c9de560dSAlex Tomas /* at order 0 we see each particular block */ 421c9de560dSAlex Tomas *max = 1 << (e4b->bd_blkbits + 3); 422c9de560dSAlex Tomas if (order == 0) 423c9de560dSAlex Tomas return EXT4_MB_BITMAP(e4b); 424c9de560dSAlex Tomas 425c9de560dSAlex Tomas bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 426c9de560dSAlex Tomas *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 427c9de560dSAlex Tomas 428c9de560dSAlex Tomas return bb; 429c9de560dSAlex Tomas } 430c9de560dSAlex Tomas 431c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 432c9de560dSAlex Tomas static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 433c9de560dSAlex Tomas int first, int count) 434c9de560dSAlex Tomas { 435c9de560dSAlex Tomas int i; 436c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 437c9de560dSAlex Tomas 438c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 439c9de560dSAlex Tomas return; 440c9de560dSAlex Tomas BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group)); 441c9de560dSAlex Tomas for (i = 0; i < count; i++) { 442c9de560dSAlex Tomas if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 443c9de560dSAlex Tomas ext4_fsblk_t blocknr; 444c9de560dSAlex Tomas blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); 445c9de560dSAlex Tomas blocknr += first + i; 446c9de560dSAlex Tomas blocknr += 447c9de560dSAlex Tomas le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 448c9de560dSAlex Tomas 44946e665e9SHarvey Harrison ext4_error(sb, __func__, "double-free of inode" 450a9df9a49STheodore Ts'o " %lu's block %llu(bit %u in group %u)", 451c9de560dSAlex Tomas inode ? inode->i_ino : 0, blocknr, 452c9de560dSAlex Tomas first + i, e4b->bd_group); 453c9de560dSAlex Tomas } 454c9de560dSAlex Tomas mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 455c9de560dSAlex Tomas } 456c9de560dSAlex Tomas } 457c9de560dSAlex Tomas 458c9de560dSAlex Tomas static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 459c9de560dSAlex Tomas { 460c9de560dSAlex Tomas int i; 461c9de560dSAlex Tomas 462c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 463c9de560dSAlex Tomas return; 464c9de560dSAlex Tomas BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group)); 465c9de560dSAlex Tomas for (i = 0; i < count; i++) { 466c9de560dSAlex Tomas BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 467c9de560dSAlex Tomas mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 468c9de560dSAlex Tomas } 469c9de560dSAlex Tomas } 470c9de560dSAlex Tomas 471c9de560dSAlex Tomas static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 472c9de560dSAlex Tomas { 473c9de560dSAlex Tomas if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 474c9de560dSAlex Tomas unsigned char *b1, *b2; 475c9de560dSAlex Tomas int i; 476c9de560dSAlex Tomas b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 477c9de560dSAlex Tomas b2 = (unsigned char *) bitmap; 478c9de560dSAlex Tomas for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 479c9de560dSAlex Tomas if (b1[i] != b2[i]) { 480a9df9a49STheodore Ts'o printk(KERN_ERR "corruption in group %u " 4814776004fSTheodore Ts'o "at byte %u(%u): %x in copy != %x " 4824776004fSTheodore Ts'o "on disk/prealloc\n", 483c9de560dSAlex Tomas e4b->bd_group, i, i * 8, b1[i], b2[i]); 484c9de560dSAlex Tomas BUG(); 485c9de560dSAlex Tomas } 486c9de560dSAlex Tomas } 487c9de560dSAlex Tomas } 488c9de560dSAlex Tomas } 489c9de560dSAlex Tomas 490c9de560dSAlex Tomas #else 491c9de560dSAlex Tomas static inline void mb_free_blocks_double(struct inode *inode, 492c9de560dSAlex Tomas struct ext4_buddy *e4b, int first, int count) 493c9de560dSAlex Tomas { 494c9de560dSAlex Tomas return; 495c9de560dSAlex Tomas } 496c9de560dSAlex Tomas static inline void mb_mark_used_double(struct ext4_buddy *e4b, 497c9de560dSAlex Tomas int first, int count) 498c9de560dSAlex Tomas { 499c9de560dSAlex Tomas return; 500c9de560dSAlex Tomas } 501c9de560dSAlex Tomas static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 502c9de560dSAlex Tomas { 503c9de560dSAlex Tomas return; 504c9de560dSAlex Tomas } 505c9de560dSAlex Tomas #endif 506c9de560dSAlex Tomas 507c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 508c9de560dSAlex Tomas 509c9de560dSAlex Tomas #define MB_CHECK_ASSERT(assert) \ 510c9de560dSAlex Tomas do { \ 511c9de560dSAlex Tomas if (!(assert)) { \ 512c9de560dSAlex Tomas printk(KERN_EMERG \ 513c9de560dSAlex Tomas "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 514c9de560dSAlex Tomas function, file, line, # assert); \ 515c9de560dSAlex Tomas BUG(); \ 516c9de560dSAlex Tomas } \ 517c9de560dSAlex Tomas } while (0) 518c9de560dSAlex Tomas 519c9de560dSAlex Tomas static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 520c9de560dSAlex Tomas const char *function, int line) 521c9de560dSAlex Tomas { 522c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 523c9de560dSAlex Tomas int order = e4b->bd_blkbits + 1; 524c9de560dSAlex Tomas int max; 525c9de560dSAlex Tomas int max2; 526c9de560dSAlex Tomas int i; 527c9de560dSAlex Tomas int j; 528c9de560dSAlex Tomas int k; 529c9de560dSAlex Tomas int count; 530c9de560dSAlex Tomas struct ext4_group_info *grp; 531c9de560dSAlex Tomas int fragments = 0; 532c9de560dSAlex Tomas int fstart; 533c9de560dSAlex Tomas struct list_head *cur; 534c9de560dSAlex Tomas void *buddy; 535c9de560dSAlex Tomas void *buddy2; 536c9de560dSAlex Tomas 537c9de560dSAlex Tomas { 538c9de560dSAlex Tomas static int mb_check_counter; 539c9de560dSAlex Tomas if (mb_check_counter++ % 100 != 0) 540c9de560dSAlex Tomas return 0; 541c9de560dSAlex Tomas } 542c9de560dSAlex Tomas 543c9de560dSAlex Tomas while (order > 1) { 544c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 545c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy); 546c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order - 1, &max2); 547c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy2); 548c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy != buddy2); 549c9de560dSAlex Tomas MB_CHECK_ASSERT(max * 2 == max2); 550c9de560dSAlex Tomas 551c9de560dSAlex Tomas count = 0; 552c9de560dSAlex Tomas for (i = 0; i < max; i++) { 553c9de560dSAlex Tomas 554c9de560dSAlex Tomas if (mb_test_bit(i, buddy)) { 555c9de560dSAlex Tomas /* only single bit in buddy2 may be 1 */ 556c9de560dSAlex Tomas if (!mb_test_bit(i << 1, buddy2)) { 557c9de560dSAlex Tomas MB_CHECK_ASSERT( 558c9de560dSAlex Tomas mb_test_bit((i<<1)+1, buddy2)); 559c9de560dSAlex Tomas } else if (!mb_test_bit((i << 1) + 1, buddy2)) { 560c9de560dSAlex Tomas MB_CHECK_ASSERT( 561c9de560dSAlex Tomas mb_test_bit(i << 1, buddy2)); 562c9de560dSAlex Tomas } 563c9de560dSAlex Tomas continue; 564c9de560dSAlex Tomas } 565c9de560dSAlex Tomas 566c9de560dSAlex Tomas /* both bits in buddy2 must be 0 */ 567c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 568c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 569c9de560dSAlex Tomas 570c9de560dSAlex Tomas for (j = 0; j < (1 << order); j++) { 571c9de560dSAlex Tomas k = (i * (1 << order)) + j; 572c9de560dSAlex Tomas MB_CHECK_ASSERT( 573c9de560dSAlex Tomas !mb_test_bit(k, EXT4_MB_BITMAP(e4b))); 574c9de560dSAlex Tomas } 575c9de560dSAlex Tomas count++; 576c9de560dSAlex Tomas } 577c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 578c9de560dSAlex Tomas order--; 579c9de560dSAlex Tomas } 580c9de560dSAlex Tomas 581c9de560dSAlex Tomas fstart = -1; 582c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 583c9de560dSAlex Tomas for (i = 0; i < max; i++) { 584c9de560dSAlex Tomas if (!mb_test_bit(i, buddy)) { 585c9de560dSAlex Tomas MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 586c9de560dSAlex Tomas if (fstart == -1) { 587c9de560dSAlex Tomas fragments++; 588c9de560dSAlex Tomas fstart = i; 589c9de560dSAlex Tomas } 590c9de560dSAlex Tomas continue; 591c9de560dSAlex Tomas } 592c9de560dSAlex Tomas fstart = -1; 593c9de560dSAlex Tomas /* check used bits only */ 594c9de560dSAlex Tomas for (j = 0; j < e4b->bd_blkbits + 1; j++) { 595c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, j, &max2); 596c9de560dSAlex Tomas k = i >> j; 597c9de560dSAlex Tomas MB_CHECK_ASSERT(k < max2); 598c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 599c9de560dSAlex Tomas } 600c9de560dSAlex Tomas } 601c9de560dSAlex Tomas MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 602c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 603c9de560dSAlex Tomas 604c9de560dSAlex Tomas grp = ext4_get_group_info(sb, e4b->bd_group); 605c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 606c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 607c9de560dSAlex Tomas ext4_group_t groupnr; 608c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 60960bd63d1SSolofo Ramangalahy pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 61060bd63d1SSolofo Ramangalahy ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 611c9de560dSAlex Tomas MB_CHECK_ASSERT(groupnr == e4b->bd_group); 61260bd63d1SSolofo Ramangalahy for (i = 0; i < pa->pa_len; i++) 613c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 614c9de560dSAlex Tomas } 615c9de560dSAlex Tomas return 0; 616c9de560dSAlex Tomas } 617c9de560dSAlex Tomas #undef MB_CHECK_ASSERT 618c9de560dSAlex Tomas #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 61946e665e9SHarvey Harrison __FILE__, __func__, __LINE__) 620c9de560dSAlex Tomas #else 621c9de560dSAlex Tomas #define mb_check_buddy(e4b) 622c9de560dSAlex Tomas #endif 623c9de560dSAlex Tomas 624c9de560dSAlex Tomas /* FIXME!! need more doc */ 625c9de560dSAlex Tomas static void ext4_mb_mark_free_simple(struct super_block *sb, 626c9de560dSAlex Tomas void *buddy, unsigned first, int len, 627c9de560dSAlex Tomas struct ext4_group_info *grp) 628c9de560dSAlex Tomas { 629c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 630c9de560dSAlex Tomas unsigned short min; 631c9de560dSAlex Tomas unsigned short max; 632c9de560dSAlex Tomas unsigned short chunk; 633c9de560dSAlex Tomas unsigned short border; 634c9de560dSAlex Tomas 635b73fce69SValerie Clement BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb)); 636c9de560dSAlex Tomas 637c9de560dSAlex Tomas border = 2 << sb->s_blocksize_bits; 638c9de560dSAlex Tomas 639c9de560dSAlex Tomas while (len > 0) { 640c9de560dSAlex Tomas /* find how many blocks can be covered since this position */ 641c9de560dSAlex Tomas max = ffs(first | border) - 1; 642c9de560dSAlex Tomas 643c9de560dSAlex Tomas /* find how many blocks of power 2 we need to mark */ 644c9de560dSAlex Tomas min = fls(len) - 1; 645c9de560dSAlex Tomas 646c9de560dSAlex Tomas if (max < min) 647c9de560dSAlex Tomas min = max; 648c9de560dSAlex Tomas chunk = 1 << min; 649c9de560dSAlex Tomas 650c9de560dSAlex Tomas /* mark multiblock chunks only */ 651c9de560dSAlex Tomas grp->bb_counters[min]++; 652c9de560dSAlex Tomas if (min > 0) 653c9de560dSAlex Tomas mb_clear_bit(first >> min, 654c9de560dSAlex Tomas buddy + sbi->s_mb_offsets[min]); 655c9de560dSAlex Tomas 656c9de560dSAlex Tomas len -= chunk; 657c9de560dSAlex Tomas first += chunk; 658c9de560dSAlex Tomas } 659c9de560dSAlex Tomas } 660c9de560dSAlex Tomas 661c9de560dSAlex Tomas static void ext4_mb_generate_buddy(struct super_block *sb, 662c9de560dSAlex Tomas void *buddy, void *bitmap, ext4_group_t group) 663c9de560dSAlex Tomas { 664c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 665c9de560dSAlex Tomas unsigned short max = EXT4_BLOCKS_PER_GROUP(sb); 666c9de560dSAlex Tomas unsigned short i = 0; 667c9de560dSAlex Tomas unsigned short first; 668c9de560dSAlex Tomas unsigned short len; 669c9de560dSAlex Tomas unsigned free = 0; 670c9de560dSAlex Tomas unsigned fragments = 0; 671c9de560dSAlex Tomas unsigned long long period = get_cycles(); 672c9de560dSAlex Tomas 673c9de560dSAlex Tomas /* initialize buddy from bitmap which is aggregation 674c9de560dSAlex Tomas * of on-disk bitmap and preallocations */ 675ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, 0); 676c9de560dSAlex Tomas grp->bb_first_free = i; 677c9de560dSAlex Tomas while (i < max) { 678c9de560dSAlex Tomas fragments++; 679c9de560dSAlex Tomas first = i; 680ffad0a44SAneesh Kumar K.V i = mb_find_next_bit(bitmap, max, i); 681c9de560dSAlex Tomas len = i - first; 682c9de560dSAlex Tomas free += len; 683c9de560dSAlex Tomas if (len > 1) 684c9de560dSAlex Tomas ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 685c9de560dSAlex Tomas else 686c9de560dSAlex Tomas grp->bb_counters[0]++; 687c9de560dSAlex Tomas if (i < max) 688ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, i); 689c9de560dSAlex Tomas } 690c9de560dSAlex Tomas grp->bb_fragments = fragments; 691c9de560dSAlex Tomas 692c9de560dSAlex Tomas if (free != grp->bb_free) { 69346e665e9SHarvey Harrison ext4_error(sb, __func__, 694a9df9a49STheodore Ts'o "EXT4-fs: group %u: %u blocks in bitmap, %u in gd", 695c9de560dSAlex Tomas group, free, grp->bb_free); 696e56eb659SAneesh Kumar K.V /* 697e56eb659SAneesh Kumar K.V * If we intent to continue, we consider group descritor 698e56eb659SAneesh Kumar K.V * corrupt and update bb_free using bitmap value 699e56eb659SAneesh Kumar K.V */ 700c9de560dSAlex Tomas grp->bb_free = free; 701c9de560dSAlex Tomas } 702c9de560dSAlex Tomas 703c9de560dSAlex Tomas clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 704c9de560dSAlex Tomas 705c9de560dSAlex Tomas period = get_cycles() - period; 706c9de560dSAlex Tomas spin_lock(&EXT4_SB(sb)->s_bal_lock); 707c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_buddies_generated++; 708c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_generation_time += period; 709c9de560dSAlex Tomas spin_unlock(&EXT4_SB(sb)->s_bal_lock); 710c9de560dSAlex Tomas } 711c9de560dSAlex Tomas 712c9de560dSAlex Tomas /* The buddy information is attached the buddy cache inode 713c9de560dSAlex Tomas * for convenience. The information regarding each group 714c9de560dSAlex Tomas * is loaded via ext4_mb_load_buddy. The information involve 715c9de560dSAlex Tomas * block bitmap and buddy information. The information are 716c9de560dSAlex Tomas * stored in the inode as 717c9de560dSAlex Tomas * 718c9de560dSAlex Tomas * { page } 719c9de560dSAlex Tomas * [ group 0 buddy][ group 0 bitmap] [group 1][ group 1]... 720c9de560dSAlex Tomas * 721c9de560dSAlex Tomas * 722c9de560dSAlex Tomas * one block each for bitmap and buddy information. 723c9de560dSAlex Tomas * So for each group we take up 2 blocks. A page can 724c9de560dSAlex Tomas * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. 725c9de560dSAlex Tomas * So it can have information regarding groups_per_page which 726c9de560dSAlex Tomas * is blocks_per_page/2 727c9de560dSAlex Tomas */ 728c9de560dSAlex Tomas 729c9de560dSAlex Tomas static int ext4_mb_init_cache(struct page *page, char *incore) 730c9de560dSAlex Tomas { 731c9de560dSAlex Tomas int blocksize; 732c9de560dSAlex Tomas int blocks_per_page; 733c9de560dSAlex Tomas int groups_per_page; 734c9de560dSAlex Tomas int err = 0; 735c9de560dSAlex Tomas int i; 736c9de560dSAlex Tomas ext4_group_t first_group; 737c9de560dSAlex Tomas int first_block; 738c9de560dSAlex Tomas struct super_block *sb; 739c9de560dSAlex Tomas struct buffer_head *bhs; 740c9de560dSAlex Tomas struct buffer_head **bh; 741c9de560dSAlex Tomas struct inode *inode; 742c9de560dSAlex Tomas char *data; 743c9de560dSAlex Tomas char *bitmap; 744c9de560dSAlex Tomas 745c9de560dSAlex Tomas mb_debug("init page %lu\n", page->index); 746c9de560dSAlex Tomas 747c9de560dSAlex Tomas inode = page->mapping->host; 748c9de560dSAlex Tomas sb = inode->i_sb; 749c9de560dSAlex Tomas blocksize = 1 << inode->i_blkbits; 750c9de560dSAlex Tomas blocks_per_page = PAGE_CACHE_SIZE / blocksize; 751c9de560dSAlex Tomas 752c9de560dSAlex Tomas groups_per_page = blocks_per_page >> 1; 753c9de560dSAlex Tomas if (groups_per_page == 0) 754c9de560dSAlex Tomas groups_per_page = 1; 755c9de560dSAlex Tomas 756c9de560dSAlex Tomas /* allocate buffer_heads to read bitmaps */ 757c9de560dSAlex Tomas if (groups_per_page > 1) { 758c9de560dSAlex Tomas err = -ENOMEM; 759c9de560dSAlex Tomas i = sizeof(struct buffer_head *) * groups_per_page; 760c9de560dSAlex Tomas bh = kzalloc(i, GFP_NOFS); 761c9de560dSAlex Tomas if (bh == NULL) 762c9de560dSAlex Tomas goto out; 763c9de560dSAlex Tomas } else 764c9de560dSAlex Tomas bh = &bhs; 765c9de560dSAlex Tomas 766c9de560dSAlex Tomas first_group = page->index * blocks_per_page / 2; 767c9de560dSAlex Tomas 768c9de560dSAlex Tomas /* read all groups the page covers into the cache */ 769c9de560dSAlex Tomas for (i = 0; i < groups_per_page; i++) { 770c9de560dSAlex Tomas struct ext4_group_desc *desc; 771c9de560dSAlex Tomas 772c9de560dSAlex Tomas if (first_group + i >= EXT4_SB(sb)->s_groups_count) 773c9de560dSAlex Tomas break; 774c9de560dSAlex Tomas 775c9de560dSAlex Tomas err = -EIO; 776c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, first_group + i, NULL); 777c9de560dSAlex Tomas if (desc == NULL) 778c9de560dSAlex Tomas goto out; 779c9de560dSAlex Tomas 780c9de560dSAlex Tomas err = -ENOMEM; 781c9de560dSAlex Tomas bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); 782c9de560dSAlex Tomas if (bh[i] == NULL) 783c9de560dSAlex Tomas goto out; 784c9de560dSAlex Tomas 785c806e68fSFrederic Bohe if (buffer_uptodate(bh[i]) && 786c806e68fSFrederic Bohe !(desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) 787c9de560dSAlex Tomas continue; 788c9de560dSAlex Tomas 789c806e68fSFrederic Bohe lock_buffer(bh[i]); 790b5f10eedSEric Sandeen spin_lock(sb_bgl_lock(EXT4_SB(sb), first_group + i)); 791c9de560dSAlex Tomas if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 792c9de560dSAlex Tomas ext4_init_block_bitmap(sb, bh[i], 793c9de560dSAlex Tomas first_group + i, desc); 794c9de560dSAlex Tomas set_buffer_uptodate(bh[i]); 795c9de560dSAlex Tomas unlock_buffer(bh[i]); 796b5f10eedSEric Sandeen spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i)); 797c9de560dSAlex Tomas continue; 798c9de560dSAlex Tomas } 799b5f10eedSEric Sandeen spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i)); 800c9de560dSAlex Tomas get_bh(bh[i]); 801c9de560dSAlex Tomas bh[i]->b_end_io = end_buffer_read_sync; 802c9de560dSAlex Tomas submit_bh(READ, bh[i]); 803a9df9a49STheodore Ts'o mb_debug("read bitmap for group %u\n", first_group + i); 804c9de560dSAlex Tomas } 805c9de560dSAlex Tomas 806c9de560dSAlex Tomas /* wait for I/O completion */ 807c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 808c9de560dSAlex Tomas wait_on_buffer(bh[i]); 809c9de560dSAlex Tomas 810c9de560dSAlex Tomas err = -EIO; 811c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 812c9de560dSAlex Tomas if (!buffer_uptodate(bh[i])) 813c9de560dSAlex Tomas goto out; 814c9de560dSAlex Tomas 81531b481dcSMingming Cao err = 0; 816c9de560dSAlex Tomas first_block = page->index * blocks_per_page; 817c9de560dSAlex Tomas for (i = 0; i < blocks_per_page; i++) { 818c9de560dSAlex Tomas int group; 819c9de560dSAlex Tomas struct ext4_group_info *grinfo; 820c9de560dSAlex Tomas 821c9de560dSAlex Tomas group = (first_block + i) >> 1; 822c9de560dSAlex Tomas if (group >= EXT4_SB(sb)->s_groups_count) 823c9de560dSAlex Tomas break; 824c9de560dSAlex Tomas 825c9de560dSAlex Tomas /* 826c9de560dSAlex Tomas * data carry information regarding this 827c9de560dSAlex Tomas * particular group in the format specified 828c9de560dSAlex Tomas * above 829c9de560dSAlex Tomas * 830c9de560dSAlex Tomas */ 831c9de560dSAlex Tomas data = page_address(page) + (i * blocksize); 832c9de560dSAlex Tomas bitmap = bh[group - first_group]->b_data; 833c9de560dSAlex Tomas 834c9de560dSAlex Tomas /* 835c9de560dSAlex Tomas * We place the buddy block and bitmap block 836c9de560dSAlex Tomas * close together 837c9de560dSAlex Tomas */ 838c9de560dSAlex Tomas if ((first_block + i) & 1) { 839c9de560dSAlex Tomas /* this is block of buddy */ 840c9de560dSAlex Tomas BUG_ON(incore == NULL); 841c9de560dSAlex Tomas mb_debug("put buddy for group %u in page %lu/%x\n", 842c9de560dSAlex Tomas group, page->index, i * blocksize); 843c9de560dSAlex Tomas memset(data, 0xff, blocksize); 844c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, group); 845c9de560dSAlex Tomas grinfo->bb_fragments = 0; 846c9de560dSAlex Tomas memset(grinfo->bb_counters, 0, 847c9de560dSAlex Tomas sizeof(unsigned short)*(sb->s_blocksize_bits+2)); 848c9de560dSAlex Tomas /* 849c9de560dSAlex Tomas * incore got set to the group block bitmap below 850c9de560dSAlex Tomas */ 851c9de560dSAlex Tomas ext4_mb_generate_buddy(sb, data, incore, group); 852c9de560dSAlex Tomas incore = NULL; 853c9de560dSAlex Tomas } else { 854c9de560dSAlex Tomas /* this is block of bitmap */ 855c9de560dSAlex Tomas BUG_ON(incore != NULL); 856c9de560dSAlex Tomas mb_debug("put bitmap for group %u in page %lu/%x\n", 857c9de560dSAlex Tomas group, page->index, i * blocksize); 858c9de560dSAlex Tomas 859c9de560dSAlex Tomas /* see comments in ext4_mb_put_pa() */ 860c9de560dSAlex Tomas ext4_lock_group(sb, group); 861c9de560dSAlex Tomas memcpy(data, bitmap, blocksize); 862c9de560dSAlex Tomas 863c9de560dSAlex Tomas /* mark all preallocated blks used in in-core bitmap */ 864c9de560dSAlex Tomas ext4_mb_generate_from_pa(sb, data, group); 865c9de560dSAlex Tomas ext4_unlock_group(sb, group); 866c9de560dSAlex Tomas 867c9de560dSAlex Tomas /* set incore so that the buddy information can be 868c9de560dSAlex Tomas * generated using this 869c9de560dSAlex Tomas */ 870c9de560dSAlex Tomas incore = data; 871c9de560dSAlex Tomas } 872c9de560dSAlex Tomas } 873c9de560dSAlex Tomas SetPageUptodate(page); 874c9de560dSAlex Tomas 875c9de560dSAlex Tomas out: 876c9de560dSAlex Tomas if (bh) { 877c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 878c9de560dSAlex Tomas brelse(bh[i]); 879c9de560dSAlex Tomas if (bh != &bhs) 880c9de560dSAlex Tomas kfree(bh); 881c9de560dSAlex Tomas } 882c9de560dSAlex Tomas return err; 883c9de560dSAlex Tomas } 884c9de560dSAlex Tomas 8854ddfef7bSEric Sandeen static noinline_for_stack int 8864ddfef7bSEric Sandeen ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 887c9de560dSAlex Tomas struct ext4_buddy *e4b) 888c9de560dSAlex Tomas { 889c9de560dSAlex Tomas int blocks_per_page; 890c9de560dSAlex Tomas int block; 891c9de560dSAlex Tomas int pnum; 892c9de560dSAlex Tomas int poff; 893c9de560dSAlex Tomas struct page *page; 894fdf6c7a7SShen Feng int ret; 895920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 896920313a7SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 897920313a7SAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 898c9de560dSAlex Tomas 899a9df9a49STheodore Ts'o mb_debug("load group %u\n", group); 900c9de560dSAlex Tomas 901c9de560dSAlex Tomas blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 902920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 903c9de560dSAlex Tomas 904c9de560dSAlex Tomas e4b->bd_blkbits = sb->s_blocksize_bits; 905c9de560dSAlex Tomas e4b->bd_info = ext4_get_group_info(sb, group); 906c9de560dSAlex Tomas e4b->bd_sb = sb; 907c9de560dSAlex Tomas e4b->bd_group = group; 908c9de560dSAlex Tomas e4b->bd_buddy_page = NULL; 909c9de560dSAlex Tomas e4b->bd_bitmap_page = NULL; 910920313a7SAneesh Kumar K.V e4b->alloc_semp = &grp->alloc_sem; 911920313a7SAneesh Kumar K.V 912920313a7SAneesh Kumar K.V /* Take the read lock on the group alloc 913920313a7SAneesh Kumar K.V * sem. This would make sure a parallel 914920313a7SAneesh Kumar K.V * ext4_mb_init_group happening on other 915920313a7SAneesh Kumar K.V * groups mapped by the page is blocked 916920313a7SAneesh Kumar K.V * till we are done with allocation 917920313a7SAneesh Kumar K.V */ 918920313a7SAneesh Kumar K.V down_read(e4b->alloc_semp); 919c9de560dSAlex Tomas 920c9de560dSAlex Tomas /* 921c9de560dSAlex Tomas * the buddy cache inode stores the block bitmap 922c9de560dSAlex Tomas * and buddy information in consecutive blocks. 923c9de560dSAlex Tomas * So for each group we need two blocks. 924c9de560dSAlex Tomas */ 925c9de560dSAlex Tomas block = group * 2; 926c9de560dSAlex Tomas pnum = block / blocks_per_page; 927c9de560dSAlex Tomas poff = block % blocks_per_page; 928c9de560dSAlex Tomas 929c9de560dSAlex Tomas /* we could use find_or_create_page(), but it locks page 930c9de560dSAlex Tomas * what we'd like to avoid in fast path ... */ 931c9de560dSAlex Tomas page = find_get_page(inode->i_mapping, pnum); 932c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 933c9de560dSAlex Tomas if (page) 934920313a7SAneesh Kumar K.V /* 935920313a7SAneesh Kumar K.V * drop the page reference and try 936920313a7SAneesh Kumar K.V * to get the page with lock. If we 937920313a7SAneesh Kumar K.V * are not uptodate that implies 938920313a7SAneesh Kumar K.V * somebody just created the page but 939920313a7SAneesh Kumar K.V * is yet to initialize the same. So 940920313a7SAneesh Kumar K.V * wait for it to initialize. 941920313a7SAneesh Kumar K.V */ 942c9de560dSAlex Tomas page_cache_release(page); 943c9de560dSAlex Tomas page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 944c9de560dSAlex Tomas if (page) { 945c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 946c9de560dSAlex Tomas if (!PageUptodate(page)) { 947fdf6c7a7SShen Feng ret = ext4_mb_init_cache(page, NULL); 948fdf6c7a7SShen Feng if (ret) { 949fdf6c7a7SShen Feng unlock_page(page); 950fdf6c7a7SShen Feng goto err; 951fdf6c7a7SShen Feng } 952c9de560dSAlex Tomas mb_cmp_bitmaps(e4b, page_address(page) + 953c9de560dSAlex Tomas (poff * sb->s_blocksize)); 954c9de560dSAlex Tomas } 955c9de560dSAlex Tomas unlock_page(page); 956c9de560dSAlex Tomas } 957c9de560dSAlex Tomas } 958fdf6c7a7SShen Feng if (page == NULL || !PageUptodate(page)) { 959fdf6c7a7SShen Feng ret = -EIO; 960c9de560dSAlex Tomas goto err; 961fdf6c7a7SShen Feng } 962c9de560dSAlex Tomas e4b->bd_bitmap_page = page; 963c9de560dSAlex Tomas e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 964c9de560dSAlex Tomas mark_page_accessed(page); 965c9de560dSAlex Tomas 966c9de560dSAlex Tomas block++; 967c9de560dSAlex Tomas pnum = block / blocks_per_page; 968c9de560dSAlex Tomas poff = block % blocks_per_page; 969c9de560dSAlex Tomas 970c9de560dSAlex Tomas page = find_get_page(inode->i_mapping, pnum); 971c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 972c9de560dSAlex Tomas if (page) 973c9de560dSAlex Tomas page_cache_release(page); 974c9de560dSAlex Tomas page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 975c9de560dSAlex Tomas if (page) { 976c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 977fdf6c7a7SShen Feng if (!PageUptodate(page)) { 978fdf6c7a7SShen Feng ret = ext4_mb_init_cache(page, e4b->bd_bitmap); 979fdf6c7a7SShen Feng if (ret) { 980fdf6c7a7SShen Feng unlock_page(page); 981fdf6c7a7SShen Feng goto err; 982fdf6c7a7SShen Feng } 983fdf6c7a7SShen Feng } 984c9de560dSAlex Tomas unlock_page(page); 985c9de560dSAlex Tomas } 986c9de560dSAlex Tomas } 987fdf6c7a7SShen Feng if (page == NULL || !PageUptodate(page)) { 988fdf6c7a7SShen Feng ret = -EIO; 989c9de560dSAlex Tomas goto err; 990fdf6c7a7SShen Feng } 991c9de560dSAlex Tomas e4b->bd_buddy_page = page; 992c9de560dSAlex Tomas e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 993c9de560dSAlex Tomas mark_page_accessed(page); 994c9de560dSAlex Tomas 995c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 996c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 997c9de560dSAlex Tomas 998c9de560dSAlex Tomas return 0; 999c9de560dSAlex Tomas 1000c9de560dSAlex Tomas err: 1001c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 1002c9de560dSAlex Tomas page_cache_release(e4b->bd_bitmap_page); 1003c9de560dSAlex Tomas if (e4b->bd_buddy_page) 1004c9de560dSAlex Tomas page_cache_release(e4b->bd_buddy_page); 1005c9de560dSAlex Tomas e4b->bd_buddy = NULL; 1006c9de560dSAlex Tomas e4b->bd_bitmap = NULL; 1007920313a7SAneesh Kumar K.V 1008920313a7SAneesh Kumar K.V /* Done with the buddy cache */ 1009920313a7SAneesh Kumar K.V up_read(e4b->alloc_semp); 1010fdf6c7a7SShen Feng return ret; 1011c9de560dSAlex Tomas } 1012c9de560dSAlex Tomas 1013c9de560dSAlex Tomas static void ext4_mb_release_desc(struct ext4_buddy *e4b) 1014c9de560dSAlex Tomas { 1015c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 1016c9de560dSAlex Tomas page_cache_release(e4b->bd_bitmap_page); 1017c9de560dSAlex Tomas if (e4b->bd_buddy_page) 1018c9de560dSAlex Tomas page_cache_release(e4b->bd_buddy_page); 1019920313a7SAneesh Kumar K.V /* Done with the buddy cache */ 1020920313a7SAneesh Kumar K.V up_read(e4b->alloc_semp); 1021c9de560dSAlex Tomas } 1022c9de560dSAlex Tomas 1023c9de560dSAlex Tomas 1024c9de560dSAlex Tomas static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1025c9de560dSAlex Tomas { 1026c9de560dSAlex Tomas int order = 1; 1027c9de560dSAlex Tomas void *bb; 1028c9de560dSAlex Tomas 1029c9de560dSAlex Tomas BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); 1030c9de560dSAlex Tomas BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1031c9de560dSAlex Tomas 1032c9de560dSAlex Tomas bb = EXT4_MB_BUDDY(e4b); 1033c9de560dSAlex Tomas while (order <= e4b->bd_blkbits + 1) { 1034c9de560dSAlex Tomas block = block >> 1; 1035c9de560dSAlex Tomas if (!mb_test_bit(block, bb)) { 1036c9de560dSAlex Tomas /* this block is part of buddy of order 'order' */ 1037c9de560dSAlex Tomas return order; 1038c9de560dSAlex Tomas } 1039c9de560dSAlex Tomas bb += 1 << (e4b->bd_blkbits - order); 1040c9de560dSAlex Tomas order++; 1041c9de560dSAlex Tomas } 1042c9de560dSAlex Tomas return 0; 1043c9de560dSAlex Tomas } 1044c9de560dSAlex Tomas 1045c9de560dSAlex Tomas static void mb_clear_bits(spinlock_t *lock, void *bm, int cur, int len) 1046c9de560dSAlex Tomas { 1047c9de560dSAlex Tomas __u32 *addr; 1048c9de560dSAlex Tomas 1049c9de560dSAlex Tomas len = cur + len; 1050c9de560dSAlex Tomas while (cur < len) { 1051c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1052c9de560dSAlex Tomas /* fast path: clear whole word at once */ 1053c9de560dSAlex Tomas addr = bm + (cur >> 3); 1054c9de560dSAlex Tomas *addr = 0; 1055c9de560dSAlex Tomas cur += 32; 1056c9de560dSAlex Tomas continue; 1057c9de560dSAlex Tomas } 1058c9de560dSAlex Tomas mb_clear_bit_atomic(lock, cur, bm); 1059c9de560dSAlex Tomas cur++; 1060c9de560dSAlex Tomas } 1061c9de560dSAlex Tomas } 1062c9de560dSAlex Tomas 1063c9de560dSAlex Tomas static void mb_set_bits(spinlock_t *lock, void *bm, int cur, int len) 1064c9de560dSAlex Tomas { 1065c9de560dSAlex Tomas __u32 *addr; 1066c9de560dSAlex Tomas 1067c9de560dSAlex Tomas len = cur + len; 1068c9de560dSAlex Tomas while (cur < len) { 1069c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1070c9de560dSAlex Tomas /* fast path: set whole word at once */ 1071c9de560dSAlex Tomas addr = bm + (cur >> 3); 1072c9de560dSAlex Tomas *addr = 0xffffffff; 1073c9de560dSAlex Tomas cur += 32; 1074c9de560dSAlex Tomas continue; 1075c9de560dSAlex Tomas } 1076c9de560dSAlex Tomas mb_set_bit_atomic(lock, cur, bm); 1077c9de560dSAlex Tomas cur++; 1078c9de560dSAlex Tomas } 1079c9de560dSAlex Tomas } 1080c9de560dSAlex Tomas 10817e5a8cddSShen Feng static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1082c9de560dSAlex Tomas int first, int count) 10833a06d778SAneesh Kumar K.V __releases(bitlock) 10843a06d778SAneesh Kumar K.V __acquires(bitlock) 1085c9de560dSAlex Tomas { 1086c9de560dSAlex Tomas int block = 0; 1087c9de560dSAlex Tomas int max = 0; 1088c9de560dSAlex Tomas int order; 1089c9de560dSAlex Tomas void *buddy; 1090c9de560dSAlex Tomas void *buddy2; 1091c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 1092c9de560dSAlex Tomas 1093c9de560dSAlex Tomas BUG_ON(first + count > (sb->s_blocksize << 3)); 1094c9de560dSAlex Tomas BUG_ON(!ext4_is_group_locked(sb, e4b->bd_group)); 1095c9de560dSAlex Tomas mb_check_buddy(e4b); 1096c9de560dSAlex Tomas mb_free_blocks_double(inode, e4b, first, count); 1097c9de560dSAlex Tomas 1098c9de560dSAlex Tomas e4b->bd_info->bb_free += count; 1099c9de560dSAlex Tomas if (first < e4b->bd_info->bb_first_free) 1100c9de560dSAlex Tomas e4b->bd_info->bb_first_free = first; 1101c9de560dSAlex Tomas 1102c9de560dSAlex Tomas /* let's maintain fragments counter */ 1103c9de560dSAlex Tomas if (first != 0) 1104c9de560dSAlex Tomas block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); 1105c9de560dSAlex Tomas if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) 1106c9de560dSAlex Tomas max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); 1107c9de560dSAlex Tomas if (block && max) 1108c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1109c9de560dSAlex Tomas else if (!block && !max) 1110c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1111c9de560dSAlex Tomas 1112c9de560dSAlex Tomas /* let's maintain buddy itself */ 1113c9de560dSAlex Tomas while (count-- > 0) { 1114c9de560dSAlex Tomas block = first++; 1115c9de560dSAlex Tomas order = 0; 1116c9de560dSAlex Tomas 1117c9de560dSAlex Tomas if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { 1118c9de560dSAlex Tomas ext4_fsblk_t blocknr; 1119c9de560dSAlex Tomas blocknr = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb); 1120c9de560dSAlex Tomas blocknr += block; 1121c9de560dSAlex Tomas blocknr += 1122c9de560dSAlex Tomas le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 11237e5a8cddSShen Feng ext4_unlock_group(sb, e4b->bd_group); 112446e665e9SHarvey Harrison ext4_error(sb, __func__, "double-free of inode" 1125a9df9a49STheodore Ts'o " %lu's block %llu(bit %u in group %u)", 1126c9de560dSAlex Tomas inode ? inode->i_ino : 0, blocknr, block, 1127c9de560dSAlex Tomas e4b->bd_group); 11287e5a8cddSShen Feng ext4_lock_group(sb, e4b->bd_group); 1129c9de560dSAlex Tomas } 1130c9de560dSAlex Tomas mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); 1131c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]++; 1132c9de560dSAlex Tomas 1133c9de560dSAlex Tomas /* start of the buddy */ 1134c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 1135c9de560dSAlex Tomas 1136c9de560dSAlex Tomas do { 1137c9de560dSAlex Tomas block &= ~1UL; 1138c9de560dSAlex Tomas if (mb_test_bit(block, buddy) || 1139c9de560dSAlex Tomas mb_test_bit(block + 1, buddy)) 1140c9de560dSAlex Tomas break; 1141c9de560dSAlex Tomas 1142c9de560dSAlex Tomas /* both the buddies are free, try to coalesce them */ 1143c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order + 1, &max); 1144c9de560dSAlex Tomas 1145c9de560dSAlex Tomas if (!buddy2) 1146c9de560dSAlex Tomas break; 1147c9de560dSAlex Tomas 1148c9de560dSAlex Tomas if (order > 0) { 1149c9de560dSAlex Tomas /* for special purposes, we don't set 1150c9de560dSAlex Tomas * free bits in bitmap */ 1151c9de560dSAlex Tomas mb_set_bit(block, buddy); 1152c9de560dSAlex Tomas mb_set_bit(block + 1, buddy); 1153c9de560dSAlex Tomas } 1154c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]--; 1155c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]--; 1156c9de560dSAlex Tomas 1157c9de560dSAlex Tomas block = block >> 1; 1158c9de560dSAlex Tomas order++; 1159c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]++; 1160c9de560dSAlex Tomas 1161c9de560dSAlex Tomas mb_clear_bit(block, buddy2); 1162c9de560dSAlex Tomas buddy = buddy2; 1163c9de560dSAlex Tomas } while (1); 1164c9de560dSAlex Tomas } 1165c9de560dSAlex Tomas mb_check_buddy(e4b); 1166c9de560dSAlex Tomas } 1167c9de560dSAlex Tomas 1168c9de560dSAlex Tomas static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, 1169c9de560dSAlex Tomas int needed, struct ext4_free_extent *ex) 1170c9de560dSAlex Tomas { 1171c9de560dSAlex Tomas int next = block; 1172c9de560dSAlex Tomas int max; 1173c9de560dSAlex Tomas int ord; 1174c9de560dSAlex Tomas void *buddy; 1175c9de560dSAlex Tomas 1176c9de560dSAlex Tomas BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group)); 1177c9de560dSAlex Tomas BUG_ON(ex == NULL); 1178c9de560dSAlex Tomas 1179c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 1180c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1181c9de560dSAlex Tomas BUG_ON(block >= max); 1182c9de560dSAlex Tomas if (mb_test_bit(block, buddy)) { 1183c9de560dSAlex Tomas ex->fe_len = 0; 1184c9de560dSAlex Tomas ex->fe_start = 0; 1185c9de560dSAlex Tomas ex->fe_group = 0; 1186c9de560dSAlex Tomas return 0; 1187c9de560dSAlex Tomas } 1188c9de560dSAlex Tomas 1189c9de560dSAlex Tomas /* FIXME dorp order completely ? */ 1190c9de560dSAlex Tomas if (likely(order == 0)) { 1191c9de560dSAlex Tomas /* find actual order */ 1192c9de560dSAlex Tomas order = mb_find_order_for_block(e4b, block); 1193c9de560dSAlex Tomas block = block >> order; 1194c9de560dSAlex Tomas } 1195c9de560dSAlex Tomas 1196c9de560dSAlex Tomas ex->fe_len = 1 << order; 1197c9de560dSAlex Tomas ex->fe_start = block << order; 1198c9de560dSAlex Tomas ex->fe_group = e4b->bd_group; 1199c9de560dSAlex Tomas 1200c9de560dSAlex Tomas /* calc difference from given start */ 1201c9de560dSAlex Tomas next = next - ex->fe_start; 1202c9de560dSAlex Tomas ex->fe_len -= next; 1203c9de560dSAlex Tomas ex->fe_start += next; 1204c9de560dSAlex Tomas 1205c9de560dSAlex Tomas while (needed > ex->fe_len && 1206c9de560dSAlex Tomas (buddy = mb_find_buddy(e4b, order, &max))) { 1207c9de560dSAlex Tomas 1208c9de560dSAlex Tomas if (block + 1 >= max) 1209c9de560dSAlex Tomas break; 1210c9de560dSAlex Tomas 1211c9de560dSAlex Tomas next = (block + 1) * (1 << order); 1212c9de560dSAlex Tomas if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) 1213c9de560dSAlex Tomas break; 1214c9de560dSAlex Tomas 1215c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, next); 1216c9de560dSAlex Tomas 1217c9de560dSAlex Tomas order = ord; 1218c9de560dSAlex Tomas block = next >> order; 1219c9de560dSAlex Tomas ex->fe_len += 1 << order; 1220c9de560dSAlex Tomas } 1221c9de560dSAlex Tomas 1222c9de560dSAlex Tomas BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); 1223c9de560dSAlex Tomas return ex->fe_len; 1224c9de560dSAlex Tomas } 1225c9de560dSAlex Tomas 1226c9de560dSAlex Tomas static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 1227c9de560dSAlex Tomas { 1228c9de560dSAlex Tomas int ord; 1229c9de560dSAlex Tomas int mlen = 0; 1230c9de560dSAlex Tomas int max = 0; 1231c9de560dSAlex Tomas int cur; 1232c9de560dSAlex Tomas int start = ex->fe_start; 1233c9de560dSAlex Tomas int len = ex->fe_len; 1234c9de560dSAlex Tomas unsigned ret = 0; 1235c9de560dSAlex Tomas int len0 = len; 1236c9de560dSAlex Tomas void *buddy; 1237c9de560dSAlex Tomas 1238c9de560dSAlex Tomas BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 1239c9de560dSAlex Tomas BUG_ON(e4b->bd_group != ex->fe_group); 1240c9de560dSAlex Tomas BUG_ON(!ext4_is_group_locked(e4b->bd_sb, e4b->bd_group)); 1241c9de560dSAlex Tomas mb_check_buddy(e4b); 1242c9de560dSAlex Tomas mb_mark_used_double(e4b, start, len); 1243c9de560dSAlex Tomas 1244c9de560dSAlex Tomas e4b->bd_info->bb_free -= len; 1245c9de560dSAlex Tomas if (e4b->bd_info->bb_first_free == start) 1246c9de560dSAlex Tomas e4b->bd_info->bb_first_free += len; 1247c9de560dSAlex Tomas 1248c9de560dSAlex Tomas /* let's maintain fragments counter */ 1249c9de560dSAlex Tomas if (start != 0) 1250c9de560dSAlex Tomas mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); 1251c9de560dSAlex Tomas if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 1252c9de560dSAlex Tomas max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); 1253c9de560dSAlex Tomas if (mlen && max) 1254c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1255c9de560dSAlex Tomas else if (!mlen && !max) 1256c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1257c9de560dSAlex Tomas 1258c9de560dSAlex Tomas /* let's maintain buddy itself */ 1259c9de560dSAlex Tomas while (len) { 1260c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, start); 1261c9de560dSAlex Tomas 1262c9de560dSAlex Tomas if (((start >> ord) << ord) == start && len >= (1 << ord)) { 1263c9de560dSAlex Tomas /* the whole chunk may be allocated at once! */ 1264c9de560dSAlex Tomas mlen = 1 << ord; 1265c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1266c9de560dSAlex Tomas BUG_ON((start >> ord) >= max); 1267c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1268c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1269c9de560dSAlex Tomas start += mlen; 1270c9de560dSAlex Tomas len -= mlen; 1271c9de560dSAlex Tomas BUG_ON(len < 0); 1272c9de560dSAlex Tomas continue; 1273c9de560dSAlex Tomas } 1274c9de560dSAlex Tomas 1275c9de560dSAlex Tomas /* store for history */ 1276c9de560dSAlex Tomas if (ret == 0) 1277c9de560dSAlex Tomas ret = len | (ord << 16); 1278c9de560dSAlex Tomas 1279c9de560dSAlex Tomas /* we have to split large buddy */ 1280c9de560dSAlex Tomas BUG_ON(ord <= 0); 1281c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1282c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1283c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1284c9de560dSAlex Tomas 1285c9de560dSAlex Tomas ord--; 1286c9de560dSAlex Tomas cur = (start >> ord) & ~1U; 1287c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1288c9de560dSAlex Tomas mb_clear_bit(cur, buddy); 1289c9de560dSAlex Tomas mb_clear_bit(cur + 1, buddy); 1290c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1291c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1292c9de560dSAlex Tomas } 1293c9de560dSAlex Tomas 1294c9de560dSAlex Tomas mb_set_bits(sb_bgl_lock(EXT4_SB(e4b->bd_sb), ex->fe_group), 1295c9de560dSAlex Tomas EXT4_MB_BITMAP(e4b), ex->fe_start, len0); 1296c9de560dSAlex Tomas mb_check_buddy(e4b); 1297c9de560dSAlex Tomas 1298c9de560dSAlex Tomas return ret; 1299c9de560dSAlex Tomas } 1300c9de560dSAlex Tomas 1301c9de560dSAlex Tomas /* 1302c9de560dSAlex Tomas * Must be called under group lock! 1303c9de560dSAlex Tomas */ 1304c9de560dSAlex Tomas static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 1305c9de560dSAlex Tomas struct ext4_buddy *e4b) 1306c9de560dSAlex Tomas { 1307c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1308c9de560dSAlex Tomas int ret; 1309c9de560dSAlex Tomas 1310c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 1311c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1312c9de560dSAlex Tomas 1313c9de560dSAlex Tomas ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 1314c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 1315c9de560dSAlex Tomas ret = mb_mark_used(e4b, &ac->ac_b_ex); 1316c9de560dSAlex Tomas 1317c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 1318c9de560dSAlex Tomas * allocated blocks for history */ 1319c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 1320c9de560dSAlex Tomas 1321c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 1322c9de560dSAlex Tomas ac->ac_tail = ret & 0xffff; 1323c9de560dSAlex Tomas ac->ac_buddy = ret >> 16; 1324c9de560dSAlex Tomas 1325c9de560dSAlex Tomas /* XXXXXXX: SUCH A HORRIBLE **CK */ 1326c9de560dSAlex Tomas /*FIXME!! Why ? */ 1327c9de560dSAlex Tomas ac->ac_bitmap_page = e4b->bd_bitmap_page; 1328c9de560dSAlex Tomas get_page(ac->ac_bitmap_page); 1329c9de560dSAlex Tomas ac->ac_buddy_page = e4b->bd_buddy_page; 1330c9de560dSAlex Tomas get_page(ac->ac_buddy_page); 1331c9de560dSAlex Tomas 1332c9de560dSAlex Tomas /* store last allocated for subsequent stream allocation */ 1333c9de560dSAlex Tomas if ((ac->ac_flags & EXT4_MB_HINT_DATA)) { 1334c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 1335c9de560dSAlex Tomas sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 1336c9de560dSAlex Tomas sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 1337c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 1338c9de560dSAlex Tomas } 1339c9de560dSAlex Tomas } 1340c9de560dSAlex Tomas 1341c9de560dSAlex Tomas /* 1342c9de560dSAlex Tomas * regular allocator, for general purposes allocation 1343c9de560dSAlex Tomas */ 1344c9de560dSAlex Tomas 1345c9de560dSAlex Tomas static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 1346c9de560dSAlex Tomas struct ext4_buddy *e4b, 1347c9de560dSAlex Tomas int finish_group) 1348c9de560dSAlex Tomas { 1349c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1350c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1351c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1352c9de560dSAlex Tomas struct ext4_free_extent ex; 1353c9de560dSAlex Tomas int max; 1354c9de560dSAlex Tomas 1355032115fcSAneesh Kumar K.V if (ac->ac_status == AC_STATUS_FOUND) 1356032115fcSAneesh Kumar K.V return; 1357c9de560dSAlex Tomas /* 1358c9de560dSAlex Tomas * We don't want to scan for a whole year 1359c9de560dSAlex Tomas */ 1360c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan && 1361c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1362c9de560dSAlex Tomas ac->ac_status = AC_STATUS_BREAK; 1363c9de560dSAlex Tomas return; 1364c9de560dSAlex Tomas } 1365c9de560dSAlex Tomas 1366c9de560dSAlex Tomas /* 1367c9de560dSAlex Tomas * Haven't found good chunk so far, let's continue 1368c9de560dSAlex Tomas */ 1369c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) 1370c9de560dSAlex Tomas return; 1371c9de560dSAlex Tomas 1372c9de560dSAlex Tomas if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 1373c9de560dSAlex Tomas && bex->fe_group == e4b->bd_group) { 1374c9de560dSAlex Tomas /* recheck chunk's availability - we don't know 1375c9de560dSAlex Tomas * when it was found (within this lock-unlock 1376c9de560dSAlex Tomas * period or not) */ 1377c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); 1378c9de560dSAlex Tomas if (max >= gex->fe_len) { 1379c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1380c9de560dSAlex Tomas return; 1381c9de560dSAlex Tomas } 1382c9de560dSAlex Tomas } 1383c9de560dSAlex Tomas } 1384c9de560dSAlex Tomas 1385c9de560dSAlex Tomas /* 1386c9de560dSAlex Tomas * The routine checks whether found extent is good enough. If it is, 1387c9de560dSAlex Tomas * then the extent gets marked used and flag is set to the context 1388c9de560dSAlex Tomas * to stop scanning. Otherwise, the extent is compared with the 1389c9de560dSAlex Tomas * previous found extent and if new one is better, then it's stored 1390c9de560dSAlex Tomas * in the context. Later, the best found extent will be used, if 1391c9de560dSAlex Tomas * mballoc can't find good enough extent. 1392c9de560dSAlex Tomas * 1393c9de560dSAlex Tomas * FIXME: real allocation policy is to be designed yet! 1394c9de560dSAlex Tomas */ 1395c9de560dSAlex Tomas static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 1396c9de560dSAlex Tomas struct ext4_free_extent *ex, 1397c9de560dSAlex Tomas struct ext4_buddy *e4b) 1398c9de560dSAlex Tomas { 1399c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1400c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1401c9de560dSAlex Tomas 1402c9de560dSAlex Tomas BUG_ON(ex->fe_len <= 0); 1403c9de560dSAlex Tomas BUG_ON(ex->fe_len >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 1404c9de560dSAlex Tomas BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 1405c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 1406c9de560dSAlex Tomas 1407c9de560dSAlex Tomas ac->ac_found++; 1408c9de560dSAlex Tomas 1409c9de560dSAlex Tomas /* 1410c9de560dSAlex Tomas * The special case - take what you catch first 1411c9de560dSAlex Tomas */ 1412c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1413c9de560dSAlex Tomas *bex = *ex; 1414c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1415c9de560dSAlex Tomas return; 1416c9de560dSAlex Tomas } 1417c9de560dSAlex Tomas 1418c9de560dSAlex Tomas /* 1419c9de560dSAlex Tomas * Let's check whether the chuck is good enough 1420c9de560dSAlex Tomas */ 1421c9de560dSAlex Tomas if (ex->fe_len == gex->fe_len) { 1422c9de560dSAlex Tomas *bex = *ex; 1423c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1424c9de560dSAlex Tomas return; 1425c9de560dSAlex Tomas } 1426c9de560dSAlex Tomas 1427c9de560dSAlex Tomas /* 1428c9de560dSAlex Tomas * If this is first found extent, just store it in the context 1429c9de560dSAlex Tomas */ 1430c9de560dSAlex Tomas if (bex->fe_len == 0) { 1431c9de560dSAlex Tomas *bex = *ex; 1432c9de560dSAlex Tomas return; 1433c9de560dSAlex Tomas } 1434c9de560dSAlex Tomas 1435c9de560dSAlex Tomas /* 1436c9de560dSAlex Tomas * If new found extent is better, store it in the context 1437c9de560dSAlex Tomas */ 1438c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) { 1439c9de560dSAlex Tomas /* if the request isn't satisfied, any found extent 1440c9de560dSAlex Tomas * larger than previous best one is better */ 1441c9de560dSAlex Tomas if (ex->fe_len > bex->fe_len) 1442c9de560dSAlex Tomas *bex = *ex; 1443c9de560dSAlex Tomas } else if (ex->fe_len > gex->fe_len) { 1444c9de560dSAlex Tomas /* if the request is satisfied, then we try to find 1445c9de560dSAlex Tomas * an extent that still satisfy the request, but is 1446c9de560dSAlex Tomas * smaller than previous one */ 1447c9de560dSAlex Tomas if (ex->fe_len < bex->fe_len) 1448c9de560dSAlex Tomas *bex = *ex; 1449c9de560dSAlex Tomas } 1450c9de560dSAlex Tomas 1451c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 0); 1452c9de560dSAlex Tomas } 1453c9de560dSAlex Tomas 1454c9de560dSAlex Tomas static int ext4_mb_try_best_found(struct ext4_allocation_context *ac, 1455c9de560dSAlex Tomas struct ext4_buddy *e4b) 1456c9de560dSAlex Tomas { 1457c9de560dSAlex Tomas struct ext4_free_extent ex = ac->ac_b_ex; 1458c9de560dSAlex Tomas ext4_group_t group = ex.fe_group; 1459c9de560dSAlex Tomas int max; 1460c9de560dSAlex Tomas int err; 1461c9de560dSAlex Tomas 1462c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1463c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1464c9de560dSAlex Tomas if (err) 1465c9de560dSAlex Tomas return err; 1466c9de560dSAlex Tomas 1467c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 1468c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); 1469c9de560dSAlex Tomas 1470c9de560dSAlex Tomas if (max > 0) { 1471c9de560dSAlex Tomas ac->ac_b_ex = ex; 1472c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1473c9de560dSAlex Tomas } 1474c9de560dSAlex Tomas 1475c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1476c9de560dSAlex Tomas ext4_mb_release_desc(e4b); 1477c9de560dSAlex Tomas 1478c9de560dSAlex Tomas return 0; 1479c9de560dSAlex Tomas } 1480c9de560dSAlex Tomas 1481c9de560dSAlex Tomas static int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 1482c9de560dSAlex Tomas struct ext4_buddy *e4b) 1483c9de560dSAlex Tomas { 1484c9de560dSAlex Tomas ext4_group_t group = ac->ac_g_ex.fe_group; 1485c9de560dSAlex Tomas int max; 1486c9de560dSAlex Tomas int err; 1487c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1488c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 1489c9de560dSAlex Tomas struct ext4_free_extent ex; 1490c9de560dSAlex Tomas 1491c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) 1492c9de560dSAlex Tomas return 0; 1493c9de560dSAlex Tomas 1494c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1495c9de560dSAlex Tomas if (err) 1496c9de560dSAlex Tomas return err; 1497c9de560dSAlex Tomas 1498c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 1499c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, 1500c9de560dSAlex Tomas ac->ac_g_ex.fe_len, &ex); 1501c9de560dSAlex Tomas 1502c9de560dSAlex Tomas if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { 1503c9de560dSAlex Tomas ext4_fsblk_t start; 1504c9de560dSAlex Tomas 1505c9de560dSAlex Tomas start = (e4b->bd_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb)) + 1506c9de560dSAlex Tomas ex.fe_start + le32_to_cpu(es->s_first_data_block); 1507c9de560dSAlex Tomas /* use do_div to get remainder (would be 64-bit modulo) */ 1508c9de560dSAlex Tomas if (do_div(start, sbi->s_stripe) == 0) { 1509c9de560dSAlex Tomas ac->ac_found++; 1510c9de560dSAlex Tomas ac->ac_b_ex = ex; 1511c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1512c9de560dSAlex Tomas } 1513c9de560dSAlex Tomas } else if (max >= ac->ac_g_ex.fe_len) { 1514c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1515c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1516c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1517c9de560dSAlex Tomas ac->ac_found++; 1518c9de560dSAlex Tomas ac->ac_b_ex = ex; 1519c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1520c9de560dSAlex Tomas } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 1521c9de560dSAlex Tomas /* Sometimes, caller may want to merge even small 1522c9de560dSAlex Tomas * number of blocks to an existing extent */ 1523c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1524c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1525c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1526c9de560dSAlex Tomas ac->ac_found++; 1527c9de560dSAlex Tomas ac->ac_b_ex = ex; 1528c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1529c9de560dSAlex Tomas } 1530c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1531c9de560dSAlex Tomas ext4_mb_release_desc(e4b); 1532c9de560dSAlex Tomas 1533c9de560dSAlex Tomas return 0; 1534c9de560dSAlex Tomas } 1535c9de560dSAlex Tomas 1536c9de560dSAlex Tomas /* 1537c9de560dSAlex Tomas * The routine scans buddy structures (not bitmap!) from given order 1538c9de560dSAlex Tomas * to max order and tries to find big enough chunk to satisfy the req 1539c9de560dSAlex Tomas */ 1540c9de560dSAlex Tomas static void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 1541c9de560dSAlex Tomas struct ext4_buddy *e4b) 1542c9de560dSAlex Tomas { 1543c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1544c9de560dSAlex Tomas struct ext4_group_info *grp = e4b->bd_info; 1545c9de560dSAlex Tomas void *buddy; 1546c9de560dSAlex Tomas int i; 1547c9de560dSAlex Tomas int k; 1548c9de560dSAlex Tomas int max; 1549c9de560dSAlex Tomas 1550c9de560dSAlex Tomas BUG_ON(ac->ac_2order <= 0); 1551c9de560dSAlex Tomas for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { 1552c9de560dSAlex Tomas if (grp->bb_counters[i] == 0) 1553c9de560dSAlex Tomas continue; 1554c9de560dSAlex Tomas 1555c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, i, &max); 1556c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1557c9de560dSAlex Tomas 1558ffad0a44SAneesh Kumar K.V k = mb_find_next_zero_bit(buddy, max, 0); 1559c9de560dSAlex Tomas BUG_ON(k >= max); 1560c9de560dSAlex Tomas 1561c9de560dSAlex Tomas ac->ac_found++; 1562c9de560dSAlex Tomas 1563c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 1 << i; 1564c9de560dSAlex Tomas ac->ac_b_ex.fe_start = k << i; 1565c9de560dSAlex Tomas ac->ac_b_ex.fe_group = e4b->bd_group; 1566c9de560dSAlex Tomas 1567c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1568c9de560dSAlex Tomas 1569c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); 1570c9de560dSAlex Tomas 1571c9de560dSAlex Tomas if (EXT4_SB(sb)->s_mb_stats) 1572c9de560dSAlex Tomas atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 1573c9de560dSAlex Tomas 1574c9de560dSAlex Tomas break; 1575c9de560dSAlex Tomas } 1576c9de560dSAlex Tomas } 1577c9de560dSAlex Tomas 1578c9de560dSAlex Tomas /* 1579c9de560dSAlex Tomas * The routine scans the group and measures all found extents. 1580c9de560dSAlex Tomas * In order to optimize scanning, caller must pass number of 1581c9de560dSAlex Tomas * free blocks in the group, so the routine can know upper limit. 1582c9de560dSAlex Tomas */ 1583c9de560dSAlex Tomas static void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 1584c9de560dSAlex Tomas struct ext4_buddy *e4b) 1585c9de560dSAlex Tomas { 1586c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1587c9de560dSAlex Tomas void *bitmap = EXT4_MB_BITMAP(e4b); 1588c9de560dSAlex Tomas struct ext4_free_extent ex; 1589c9de560dSAlex Tomas int i; 1590c9de560dSAlex Tomas int free; 1591c9de560dSAlex Tomas 1592c9de560dSAlex Tomas free = e4b->bd_info->bb_free; 1593c9de560dSAlex Tomas BUG_ON(free <= 0); 1594c9de560dSAlex Tomas 1595c9de560dSAlex Tomas i = e4b->bd_info->bb_first_free; 1596c9de560dSAlex Tomas 1597c9de560dSAlex Tomas while (free && ac->ac_status == AC_STATUS_CONTINUE) { 1598ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, 1599c9de560dSAlex Tomas EXT4_BLOCKS_PER_GROUP(sb), i); 1600c9de560dSAlex Tomas if (i >= EXT4_BLOCKS_PER_GROUP(sb)) { 160126346ff6SAneesh Kumar K.V /* 1602e56eb659SAneesh Kumar K.V * IF we have corrupt bitmap, we won't find any 160326346ff6SAneesh Kumar K.V * free blocks even though group info says we 160426346ff6SAneesh Kumar K.V * we have free blocks 160526346ff6SAneesh Kumar K.V */ 160646e665e9SHarvey Harrison ext4_error(sb, __func__, "%d free blocks as per " 1607fde4d95aSTheodore Ts'o "group info. But bitmap says 0", 160826346ff6SAneesh Kumar K.V free); 1609c9de560dSAlex Tomas break; 1610c9de560dSAlex Tomas } 1611c9de560dSAlex Tomas 1612c9de560dSAlex Tomas mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); 1613c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 161426346ff6SAneesh Kumar K.V if (free < ex.fe_len) { 161546e665e9SHarvey Harrison ext4_error(sb, __func__, "%d free blocks as per " 1616fde4d95aSTheodore Ts'o "group info. But got %d blocks", 161726346ff6SAneesh Kumar K.V free, ex.fe_len); 1618e56eb659SAneesh Kumar K.V /* 1619e56eb659SAneesh Kumar K.V * The number of free blocks differs. This mostly 1620e56eb659SAneesh Kumar K.V * indicate that the bitmap is corrupt. So exit 1621e56eb659SAneesh Kumar K.V * without claiming the space. 1622e56eb659SAneesh Kumar K.V */ 1623e56eb659SAneesh Kumar K.V break; 162426346ff6SAneesh Kumar K.V } 1625c9de560dSAlex Tomas 1626c9de560dSAlex Tomas ext4_mb_measure_extent(ac, &ex, e4b); 1627c9de560dSAlex Tomas 1628c9de560dSAlex Tomas i += ex.fe_len; 1629c9de560dSAlex Tomas free -= ex.fe_len; 1630c9de560dSAlex Tomas } 1631c9de560dSAlex Tomas 1632c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 1); 1633c9de560dSAlex Tomas } 1634c9de560dSAlex Tomas 1635c9de560dSAlex Tomas /* 1636c9de560dSAlex Tomas * This is a special case for storages like raid5 1637c9de560dSAlex Tomas * we try to find stripe-aligned chunks for stripe-size requests 1638c9de560dSAlex Tomas * XXX should do so at least for multiples of stripe size as well 1639c9de560dSAlex Tomas */ 1640c9de560dSAlex Tomas static void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 1641c9de560dSAlex Tomas struct ext4_buddy *e4b) 1642c9de560dSAlex Tomas { 1643c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1644c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 1645c9de560dSAlex Tomas void *bitmap = EXT4_MB_BITMAP(e4b); 1646c9de560dSAlex Tomas struct ext4_free_extent ex; 1647c9de560dSAlex Tomas ext4_fsblk_t first_group_block; 1648c9de560dSAlex Tomas ext4_fsblk_t a; 1649c9de560dSAlex Tomas ext4_grpblk_t i; 1650c9de560dSAlex Tomas int max; 1651c9de560dSAlex Tomas 1652c9de560dSAlex Tomas BUG_ON(sbi->s_stripe == 0); 1653c9de560dSAlex Tomas 1654c9de560dSAlex Tomas /* find first stripe-aligned block in group */ 1655c9de560dSAlex Tomas first_group_block = e4b->bd_group * EXT4_BLOCKS_PER_GROUP(sb) 1656c9de560dSAlex Tomas + le32_to_cpu(sbi->s_es->s_first_data_block); 1657c9de560dSAlex Tomas a = first_group_block + sbi->s_stripe - 1; 1658c9de560dSAlex Tomas do_div(a, sbi->s_stripe); 1659c9de560dSAlex Tomas i = (a * sbi->s_stripe) - first_group_block; 1660c9de560dSAlex Tomas 1661c9de560dSAlex Tomas while (i < EXT4_BLOCKS_PER_GROUP(sb)) { 1662c9de560dSAlex Tomas if (!mb_test_bit(i, bitmap)) { 1663c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); 1664c9de560dSAlex Tomas if (max >= sbi->s_stripe) { 1665c9de560dSAlex Tomas ac->ac_found++; 1666c9de560dSAlex Tomas ac->ac_b_ex = ex; 1667c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1668c9de560dSAlex Tomas break; 1669c9de560dSAlex Tomas } 1670c9de560dSAlex Tomas } 1671c9de560dSAlex Tomas i += sbi->s_stripe; 1672c9de560dSAlex Tomas } 1673c9de560dSAlex Tomas } 1674c9de560dSAlex Tomas 1675c9de560dSAlex Tomas static int ext4_mb_good_group(struct ext4_allocation_context *ac, 1676c9de560dSAlex Tomas ext4_group_t group, int cr) 1677c9de560dSAlex Tomas { 1678c9de560dSAlex Tomas unsigned free, fragments; 1679c9de560dSAlex Tomas unsigned i, bits; 1680c9de560dSAlex Tomas struct ext4_group_desc *desc; 1681c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 1682c9de560dSAlex Tomas 1683c9de560dSAlex Tomas BUG_ON(cr < 0 || cr >= 4); 1684c9de560dSAlex Tomas BUG_ON(EXT4_MB_GRP_NEED_INIT(grp)); 1685c9de560dSAlex Tomas 1686c9de560dSAlex Tomas free = grp->bb_free; 1687c9de560dSAlex Tomas fragments = grp->bb_fragments; 1688c9de560dSAlex Tomas if (free == 0) 1689c9de560dSAlex Tomas return 0; 1690c9de560dSAlex Tomas if (fragments == 0) 1691c9de560dSAlex Tomas return 0; 1692c9de560dSAlex Tomas 1693c9de560dSAlex Tomas switch (cr) { 1694c9de560dSAlex Tomas case 0: 1695c9de560dSAlex Tomas BUG_ON(ac->ac_2order == 0); 1696c9de560dSAlex Tomas /* If this group is uninitialized, skip it initially */ 1697c9de560dSAlex Tomas desc = ext4_get_group_desc(ac->ac_sb, group, NULL); 1698c9de560dSAlex Tomas if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) 1699c9de560dSAlex Tomas return 0; 1700c9de560dSAlex Tomas 1701c9de560dSAlex Tomas bits = ac->ac_sb->s_blocksize_bits + 1; 1702c9de560dSAlex Tomas for (i = ac->ac_2order; i <= bits; i++) 1703c9de560dSAlex Tomas if (grp->bb_counters[i] > 0) 1704c9de560dSAlex Tomas return 1; 1705c9de560dSAlex Tomas break; 1706c9de560dSAlex Tomas case 1: 1707c9de560dSAlex Tomas if ((free / fragments) >= ac->ac_g_ex.fe_len) 1708c9de560dSAlex Tomas return 1; 1709c9de560dSAlex Tomas break; 1710c9de560dSAlex Tomas case 2: 1711c9de560dSAlex Tomas if (free >= ac->ac_g_ex.fe_len) 1712c9de560dSAlex Tomas return 1; 1713c9de560dSAlex Tomas break; 1714c9de560dSAlex Tomas case 3: 1715c9de560dSAlex Tomas return 1; 1716c9de560dSAlex Tomas default: 1717c9de560dSAlex Tomas BUG(); 1718c9de560dSAlex Tomas } 1719c9de560dSAlex Tomas 1720c9de560dSAlex Tomas return 0; 1721c9de560dSAlex Tomas } 1722c9de560dSAlex Tomas 1723920313a7SAneesh Kumar K.V /* 1724920313a7SAneesh Kumar K.V * lock the group_info alloc_sem of all the groups 1725920313a7SAneesh Kumar K.V * belonging to the same buddy cache page. This 1726920313a7SAneesh Kumar K.V * make sure other parallel operation on the buddy 1727920313a7SAneesh Kumar K.V * cache doesn't happen whild holding the buddy cache 1728920313a7SAneesh Kumar K.V * lock 1729920313a7SAneesh Kumar K.V */ 1730920313a7SAneesh Kumar K.V int ext4_mb_get_buddy_cache_lock(struct super_block *sb, ext4_group_t group) 1731920313a7SAneesh Kumar K.V { 1732920313a7SAneesh Kumar K.V int i; 1733920313a7SAneesh Kumar K.V int block, pnum; 1734920313a7SAneesh Kumar K.V int blocks_per_page; 1735920313a7SAneesh Kumar K.V int groups_per_page; 1736920313a7SAneesh Kumar K.V ext4_group_t first_group; 1737920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1738920313a7SAneesh Kumar K.V 1739920313a7SAneesh Kumar K.V blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1740920313a7SAneesh Kumar K.V /* 1741920313a7SAneesh Kumar K.V * the buddy cache inode stores the block bitmap 1742920313a7SAneesh Kumar K.V * and buddy information in consecutive blocks. 1743920313a7SAneesh Kumar K.V * So for each group we need two blocks. 1744920313a7SAneesh Kumar K.V */ 1745920313a7SAneesh Kumar K.V block = group * 2; 1746920313a7SAneesh Kumar K.V pnum = block / blocks_per_page; 1747920313a7SAneesh Kumar K.V first_group = pnum * blocks_per_page / 2; 1748920313a7SAneesh Kumar K.V 1749920313a7SAneesh Kumar K.V groups_per_page = blocks_per_page >> 1; 1750920313a7SAneesh Kumar K.V if (groups_per_page == 0) 1751920313a7SAneesh Kumar K.V groups_per_page = 1; 1752920313a7SAneesh Kumar K.V /* read all groups the page covers into the cache */ 1753920313a7SAneesh Kumar K.V for (i = 0; i < groups_per_page; i++) { 1754920313a7SAneesh Kumar K.V 1755920313a7SAneesh Kumar K.V if ((first_group + i) >= EXT4_SB(sb)->s_groups_count) 1756920313a7SAneesh Kumar K.V break; 1757920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, first_group + i); 1758920313a7SAneesh Kumar K.V /* take all groups write allocation 1759920313a7SAneesh Kumar K.V * semaphore. This make sure there is 1760920313a7SAneesh Kumar K.V * no block allocation going on in any 1761920313a7SAneesh Kumar K.V * of that groups 1762920313a7SAneesh Kumar K.V */ 1763920313a7SAneesh Kumar K.V down_write(&grp->alloc_sem); 1764920313a7SAneesh Kumar K.V } 1765920313a7SAneesh Kumar K.V return i; 1766920313a7SAneesh Kumar K.V } 1767920313a7SAneesh Kumar K.V 1768920313a7SAneesh Kumar K.V void ext4_mb_put_buddy_cache_lock(struct super_block *sb, 1769920313a7SAneesh Kumar K.V ext4_group_t group, int locked_group) 1770920313a7SAneesh Kumar K.V { 1771920313a7SAneesh Kumar K.V int i; 1772920313a7SAneesh Kumar K.V int block, pnum; 1773920313a7SAneesh Kumar K.V int blocks_per_page; 1774920313a7SAneesh Kumar K.V ext4_group_t first_group; 1775920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1776920313a7SAneesh Kumar K.V 1777920313a7SAneesh Kumar K.V blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1778920313a7SAneesh Kumar K.V /* 1779920313a7SAneesh Kumar K.V * the buddy cache inode stores the block bitmap 1780920313a7SAneesh Kumar K.V * and buddy information in consecutive blocks. 1781920313a7SAneesh Kumar K.V * So for each group we need two blocks. 1782920313a7SAneesh Kumar K.V */ 1783920313a7SAneesh Kumar K.V block = group * 2; 1784920313a7SAneesh Kumar K.V pnum = block / blocks_per_page; 1785920313a7SAneesh Kumar K.V first_group = pnum * blocks_per_page / 2; 1786920313a7SAneesh Kumar K.V /* release locks on all the groups */ 1787920313a7SAneesh Kumar K.V for (i = 0; i < locked_group; i++) { 1788920313a7SAneesh Kumar K.V 1789920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, first_group + i); 1790920313a7SAneesh Kumar K.V /* take all groups write allocation 1791920313a7SAneesh Kumar K.V * semaphore. This make sure there is 1792920313a7SAneesh Kumar K.V * no block allocation going on in any 1793920313a7SAneesh Kumar K.V * of that groups 1794920313a7SAneesh Kumar K.V */ 1795920313a7SAneesh Kumar K.V up_write(&grp->alloc_sem); 1796920313a7SAneesh Kumar K.V } 1797920313a7SAneesh Kumar K.V 1798920313a7SAneesh Kumar K.V } 1799920313a7SAneesh Kumar K.V 1800920313a7SAneesh Kumar K.V static int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) 1801920313a7SAneesh Kumar K.V { 1802920313a7SAneesh Kumar K.V 1803920313a7SAneesh Kumar K.V int ret; 1804920313a7SAneesh Kumar K.V void *bitmap; 1805920313a7SAneesh Kumar K.V int blocks_per_page; 1806920313a7SAneesh Kumar K.V int block, pnum, poff; 1807920313a7SAneesh Kumar K.V int num_grp_locked = 0; 1808920313a7SAneesh Kumar K.V struct ext4_group_info *this_grp; 1809920313a7SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 1810920313a7SAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 1811920313a7SAneesh Kumar K.V struct page *page = NULL, *bitmap_page = NULL; 1812920313a7SAneesh Kumar K.V 1813920313a7SAneesh Kumar K.V mb_debug("init group %lu\n", group); 1814920313a7SAneesh Kumar K.V blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1815920313a7SAneesh Kumar K.V this_grp = ext4_get_group_info(sb, group); 1816920313a7SAneesh Kumar K.V /* 1817920313a7SAneesh Kumar K.V * This ensures we don't add group 1818920313a7SAneesh Kumar K.V * to this buddy cache via resize 1819920313a7SAneesh Kumar K.V */ 1820920313a7SAneesh Kumar K.V num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group); 1821920313a7SAneesh Kumar K.V if (!EXT4_MB_GRP_NEED_INIT(this_grp)) { 1822920313a7SAneesh Kumar K.V /* 1823920313a7SAneesh Kumar K.V * somebody initialized the group 1824920313a7SAneesh Kumar K.V * return without doing anything 1825920313a7SAneesh Kumar K.V */ 1826920313a7SAneesh Kumar K.V ret = 0; 1827920313a7SAneesh Kumar K.V goto err; 1828920313a7SAneesh Kumar K.V } 1829920313a7SAneesh Kumar K.V /* 1830920313a7SAneesh Kumar K.V * the buddy cache inode stores the block bitmap 1831920313a7SAneesh Kumar K.V * and buddy information in consecutive blocks. 1832920313a7SAneesh Kumar K.V * So for each group we need two blocks. 1833920313a7SAneesh Kumar K.V */ 1834920313a7SAneesh Kumar K.V block = group * 2; 1835920313a7SAneesh Kumar K.V pnum = block / blocks_per_page; 1836920313a7SAneesh Kumar K.V poff = block % blocks_per_page; 1837920313a7SAneesh Kumar K.V page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1838920313a7SAneesh Kumar K.V if (page) { 1839920313a7SAneesh Kumar K.V BUG_ON(page->mapping != inode->i_mapping); 1840920313a7SAneesh Kumar K.V ret = ext4_mb_init_cache(page, NULL); 1841920313a7SAneesh Kumar K.V if (ret) { 1842920313a7SAneesh Kumar K.V unlock_page(page); 1843920313a7SAneesh Kumar K.V goto err; 1844920313a7SAneesh Kumar K.V } 1845920313a7SAneesh Kumar K.V unlock_page(page); 1846920313a7SAneesh Kumar K.V } 1847920313a7SAneesh Kumar K.V if (page == NULL || !PageUptodate(page)) { 1848920313a7SAneesh Kumar K.V ret = -EIO; 1849920313a7SAneesh Kumar K.V goto err; 1850920313a7SAneesh Kumar K.V } 1851920313a7SAneesh Kumar K.V mark_page_accessed(page); 1852920313a7SAneesh Kumar K.V bitmap_page = page; 1853920313a7SAneesh Kumar K.V bitmap = page_address(page) + (poff * sb->s_blocksize); 1854920313a7SAneesh Kumar K.V 1855920313a7SAneesh Kumar K.V /* init buddy cache */ 1856920313a7SAneesh Kumar K.V block++; 1857920313a7SAneesh Kumar K.V pnum = block / blocks_per_page; 1858920313a7SAneesh Kumar K.V poff = block % blocks_per_page; 1859920313a7SAneesh Kumar K.V page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1860920313a7SAneesh Kumar K.V if (page == bitmap_page) { 1861920313a7SAneesh Kumar K.V /* 1862920313a7SAneesh Kumar K.V * If both the bitmap and buddy are in 1863920313a7SAneesh Kumar K.V * the same page we don't need to force 1864920313a7SAneesh Kumar K.V * init the buddy 1865920313a7SAneesh Kumar K.V */ 1866920313a7SAneesh Kumar K.V unlock_page(page); 1867920313a7SAneesh Kumar K.V } else if (page) { 1868920313a7SAneesh Kumar K.V BUG_ON(page->mapping != inode->i_mapping); 1869920313a7SAneesh Kumar K.V ret = ext4_mb_init_cache(page, bitmap); 1870920313a7SAneesh Kumar K.V if (ret) { 1871920313a7SAneesh Kumar K.V unlock_page(page); 1872920313a7SAneesh Kumar K.V goto err; 1873920313a7SAneesh Kumar K.V } 1874920313a7SAneesh Kumar K.V unlock_page(page); 1875920313a7SAneesh Kumar K.V } 1876920313a7SAneesh Kumar K.V if (page == NULL || !PageUptodate(page)) { 1877920313a7SAneesh Kumar K.V ret = -EIO; 1878920313a7SAneesh Kumar K.V goto err; 1879920313a7SAneesh Kumar K.V } 1880920313a7SAneesh Kumar K.V mark_page_accessed(page); 1881920313a7SAneesh Kumar K.V err: 1882920313a7SAneesh Kumar K.V ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked); 1883920313a7SAneesh Kumar K.V if (bitmap_page) 1884920313a7SAneesh Kumar K.V page_cache_release(bitmap_page); 1885920313a7SAneesh Kumar K.V if (page) 1886920313a7SAneesh Kumar K.V page_cache_release(page); 1887920313a7SAneesh Kumar K.V return ret; 1888920313a7SAneesh Kumar K.V } 1889920313a7SAneesh Kumar K.V 18904ddfef7bSEric Sandeen static noinline_for_stack int 18914ddfef7bSEric Sandeen ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 1892c9de560dSAlex Tomas { 1893c9de560dSAlex Tomas ext4_group_t group; 1894c9de560dSAlex Tomas ext4_group_t i; 1895c9de560dSAlex Tomas int cr; 1896c9de560dSAlex Tomas int err = 0; 1897c9de560dSAlex Tomas int bsbits; 1898c9de560dSAlex Tomas struct ext4_sb_info *sbi; 1899c9de560dSAlex Tomas struct super_block *sb; 1900c9de560dSAlex Tomas struct ext4_buddy e4b; 1901c9de560dSAlex Tomas loff_t size, isize; 1902c9de560dSAlex Tomas 1903c9de560dSAlex Tomas sb = ac->ac_sb; 1904c9de560dSAlex Tomas sbi = EXT4_SB(sb); 1905c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1906c9de560dSAlex Tomas 1907c9de560dSAlex Tomas /* first, try the goal */ 1908c9de560dSAlex Tomas err = ext4_mb_find_by_goal(ac, &e4b); 1909c9de560dSAlex Tomas if (err || ac->ac_status == AC_STATUS_FOUND) 1910c9de560dSAlex Tomas goto out; 1911c9de560dSAlex Tomas 1912c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 1913c9de560dSAlex Tomas goto out; 1914c9de560dSAlex Tomas 1915c9de560dSAlex Tomas /* 1916c9de560dSAlex Tomas * ac->ac2_order is set only if the fe_len is a power of 2 1917c9de560dSAlex Tomas * if ac2_order is set we also set criteria to 0 so that we 1918c9de560dSAlex Tomas * try exact allocation using buddy. 1919c9de560dSAlex Tomas */ 1920c9de560dSAlex Tomas i = fls(ac->ac_g_ex.fe_len); 1921c9de560dSAlex Tomas ac->ac_2order = 0; 1922c9de560dSAlex Tomas /* 1923c9de560dSAlex Tomas * We search using buddy data only if the order of the request 1924c9de560dSAlex Tomas * is greater than equal to the sbi_s_mb_order2_reqs 1925c9de560dSAlex Tomas * You can tune it via /proc/fs/ext4/<partition>/order2_req 1926c9de560dSAlex Tomas */ 1927c9de560dSAlex Tomas if (i >= sbi->s_mb_order2_reqs) { 1928c9de560dSAlex Tomas /* 1929c9de560dSAlex Tomas * This should tell if fe_len is exactly power of 2 1930c9de560dSAlex Tomas */ 1931c9de560dSAlex Tomas if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) 1932c9de560dSAlex Tomas ac->ac_2order = i - 1; 1933c9de560dSAlex Tomas } 1934c9de560dSAlex Tomas 1935c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 1936c9de560dSAlex Tomas /* if stream allocation is enabled, use global goal */ 1937c9de560dSAlex Tomas size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; 1938c9de560dSAlex Tomas isize = i_size_read(ac->ac_inode) >> bsbits; 1939c9de560dSAlex Tomas if (size < isize) 1940c9de560dSAlex Tomas size = isize; 1941c9de560dSAlex Tomas 1942c9de560dSAlex Tomas if (size < sbi->s_mb_stream_request && 1943c9de560dSAlex Tomas (ac->ac_flags & EXT4_MB_HINT_DATA)) { 1944c9de560dSAlex Tomas /* TBD: may be hot point */ 1945c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 1946c9de560dSAlex Tomas ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 1947c9de560dSAlex Tomas ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 1948c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 1949c9de560dSAlex Tomas } 1950c9de560dSAlex Tomas /* Let's just scan groups to find more-less suitable blocks */ 1951c9de560dSAlex Tomas cr = ac->ac_2order ? 0 : 1; 1952c9de560dSAlex Tomas /* 1953c9de560dSAlex Tomas * cr == 0 try to get exact allocation, 1954c9de560dSAlex Tomas * cr == 3 try to get anything 1955c9de560dSAlex Tomas */ 1956c9de560dSAlex Tomas repeat: 1957c9de560dSAlex Tomas for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 1958c9de560dSAlex Tomas ac->ac_criteria = cr; 1959ed8f9c75SAneesh Kumar K.V /* 1960ed8f9c75SAneesh Kumar K.V * searching for the right group start 1961ed8f9c75SAneesh Kumar K.V * from the goal value specified 1962ed8f9c75SAneesh Kumar K.V */ 1963ed8f9c75SAneesh Kumar K.V group = ac->ac_g_ex.fe_group; 1964ed8f9c75SAneesh Kumar K.V 1965c9de560dSAlex Tomas for (i = 0; i < EXT4_SB(sb)->s_groups_count; group++, i++) { 1966c9de560dSAlex Tomas struct ext4_group_info *grp; 1967c9de560dSAlex Tomas struct ext4_group_desc *desc; 1968c9de560dSAlex Tomas 1969c9de560dSAlex Tomas if (group == EXT4_SB(sb)->s_groups_count) 1970c9de560dSAlex Tomas group = 0; 1971c9de560dSAlex Tomas 1972c9de560dSAlex Tomas /* quick check to skip empty groups */ 1973920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 1974c9de560dSAlex Tomas if (grp->bb_free == 0) 1975c9de560dSAlex Tomas continue; 1976c9de560dSAlex Tomas 1977c9de560dSAlex Tomas /* 1978c9de560dSAlex Tomas * if the group is already init we check whether it is 1979c9de560dSAlex Tomas * a good group and if not we don't load the buddy 1980c9de560dSAlex Tomas */ 1981c9de560dSAlex Tomas if (EXT4_MB_GRP_NEED_INIT(grp)) { 1982c9de560dSAlex Tomas /* 1983c9de560dSAlex Tomas * we need full data about the group 1984c9de560dSAlex Tomas * to make a good selection 1985c9de560dSAlex Tomas */ 1986920313a7SAneesh Kumar K.V err = ext4_mb_init_group(sb, group); 1987c9de560dSAlex Tomas if (err) 1988c9de560dSAlex Tomas goto out; 1989c9de560dSAlex Tomas } 1990c9de560dSAlex Tomas 1991c9de560dSAlex Tomas /* 1992c9de560dSAlex Tomas * If the particular group doesn't satisfy our 1993c9de560dSAlex Tomas * criteria we continue with the next group 1994c9de560dSAlex Tomas */ 1995c9de560dSAlex Tomas if (!ext4_mb_good_group(ac, group, cr)) 1996c9de560dSAlex Tomas continue; 1997c9de560dSAlex Tomas 1998c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 1999c9de560dSAlex Tomas if (err) 2000c9de560dSAlex Tomas goto out; 2001c9de560dSAlex Tomas 2002c9de560dSAlex Tomas ext4_lock_group(sb, group); 2003c9de560dSAlex Tomas if (!ext4_mb_good_group(ac, group, cr)) { 2004c9de560dSAlex Tomas /* someone did allocation from this group */ 2005c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2006c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 2007c9de560dSAlex Tomas continue; 2008c9de560dSAlex Tomas } 2009c9de560dSAlex Tomas 2010c9de560dSAlex Tomas ac->ac_groups_scanned++; 2011c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, group, NULL); 2012c9de560dSAlex Tomas if (cr == 0 || (desc->bg_flags & 2013c9de560dSAlex Tomas cpu_to_le16(EXT4_BG_BLOCK_UNINIT) && 2014c9de560dSAlex Tomas ac->ac_2order != 0)) 2015c9de560dSAlex Tomas ext4_mb_simple_scan_group(ac, &e4b); 2016c9de560dSAlex Tomas else if (cr == 1 && 2017c9de560dSAlex Tomas ac->ac_g_ex.fe_len == sbi->s_stripe) 2018c9de560dSAlex Tomas ext4_mb_scan_aligned(ac, &e4b); 2019c9de560dSAlex Tomas else 2020c9de560dSAlex Tomas ext4_mb_complex_scan_group(ac, &e4b); 2021c9de560dSAlex Tomas 2022c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2023c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 2024c9de560dSAlex Tomas 2025c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_CONTINUE) 2026c9de560dSAlex Tomas break; 2027c9de560dSAlex Tomas } 2028c9de560dSAlex Tomas } 2029c9de560dSAlex Tomas 2030c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2031c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2032c9de560dSAlex Tomas /* 2033c9de560dSAlex Tomas * We've been searching too long. Let's try to allocate 2034c9de560dSAlex Tomas * the best chunk we've found so far 2035c9de560dSAlex Tomas */ 2036c9de560dSAlex Tomas 2037c9de560dSAlex Tomas ext4_mb_try_best_found(ac, &e4b); 2038c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_FOUND) { 2039c9de560dSAlex Tomas /* 2040c9de560dSAlex Tomas * Someone more lucky has already allocated it. 2041c9de560dSAlex Tomas * The only thing we can do is just take first 2042c9de560dSAlex Tomas * found block(s) 2043c9de560dSAlex Tomas printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); 2044c9de560dSAlex Tomas */ 2045c9de560dSAlex Tomas ac->ac_b_ex.fe_group = 0; 2046c9de560dSAlex Tomas ac->ac_b_ex.fe_start = 0; 2047c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 0; 2048c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 2049c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_FIRST; 2050c9de560dSAlex Tomas cr = 3; 2051c9de560dSAlex Tomas atomic_inc(&sbi->s_mb_lost_chunks); 2052c9de560dSAlex Tomas goto repeat; 2053c9de560dSAlex Tomas } 2054c9de560dSAlex Tomas } 2055c9de560dSAlex Tomas out: 2056c9de560dSAlex Tomas return err; 2057c9de560dSAlex Tomas } 2058c9de560dSAlex Tomas 2059c9de560dSAlex Tomas #ifdef EXT4_MB_HISTORY 2060c9de560dSAlex Tomas struct ext4_mb_proc_session { 2061c9de560dSAlex Tomas struct ext4_mb_history *history; 2062c9de560dSAlex Tomas struct super_block *sb; 2063c9de560dSAlex Tomas int start; 2064c9de560dSAlex Tomas int max; 2065c9de560dSAlex Tomas }; 2066c9de560dSAlex Tomas 2067c9de560dSAlex Tomas static void *ext4_mb_history_skip_empty(struct ext4_mb_proc_session *s, 2068c9de560dSAlex Tomas struct ext4_mb_history *hs, 2069c9de560dSAlex Tomas int first) 2070c9de560dSAlex Tomas { 2071c9de560dSAlex Tomas if (hs == s->history + s->max) 2072c9de560dSAlex Tomas hs = s->history; 2073c9de560dSAlex Tomas if (!first && hs == s->history + s->start) 2074c9de560dSAlex Tomas return NULL; 2075c9de560dSAlex Tomas while (hs->orig.fe_len == 0) { 2076c9de560dSAlex Tomas hs++; 2077c9de560dSAlex Tomas if (hs == s->history + s->max) 2078c9de560dSAlex Tomas hs = s->history; 2079c9de560dSAlex Tomas if (hs == s->history + s->start) 2080c9de560dSAlex Tomas return NULL; 2081c9de560dSAlex Tomas } 2082c9de560dSAlex Tomas return hs; 2083c9de560dSAlex Tomas } 2084c9de560dSAlex Tomas 2085c9de560dSAlex Tomas static void *ext4_mb_seq_history_start(struct seq_file *seq, loff_t *pos) 2086c9de560dSAlex Tomas { 2087c9de560dSAlex Tomas struct ext4_mb_proc_session *s = seq->private; 2088c9de560dSAlex Tomas struct ext4_mb_history *hs; 2089c9de560dSAlex Tomas int l = *pos; 2090c9de560dSAlex Tomas 2091c9de560dSAlex Tomas if (l == 0) 2092c9de560dSAlex Tomas return SEQ_START_TOKEN; 2093c9de560dSAlex Tomas hs = ext4_mb_history_skip_empty(s, s->history + s->start, 1); 2094c9de560dSAlex Tomas if (!hs) 2095c9de560dSAlex Tomas return NULL; 2096c9de560dSAlex Tomas while (--l && (hs = ext4_mb_history_skip_empty(s, ++hs, 0)) != NULL); 2097c9de560dSAlex Tomas return hs; 2098c9de560dSAlex Tomas } 2099c9de560dSAlex Tomas 2100c9de560dSAlex Tomas static void *ext4_mb_seq_history_next(struct seq_file *seq, void *v, 2101c9de560dSAlex Tomas loff_t *pos) 2102c9de560dSAlex Tomas { 2103c9de560dSAlex Tomas struct ext4_mb_proc_session *s = seq->private; 2104c9de560dSAlex Tomas struct ext4_mb_history *hs = v; 2105c9de560dSAlex Tomas 2106c9de560dSAlex Tomas ++*pos; 2107c9de560dSAlex Tomas if (v == SEQ_START_TOKEN) 2108c9de560dSAlex Tomas return ext4_mb_history_skip_empty(s, s->history + s->start, 1); 2109c9de560dSAlex Tomas else 2110c9de560dSAlex Tomas return ext4_mb_history_skip_empty(s, ++hs, 0); 2111c9de560dSAlex Tomas } 2112c9de560dSAlex Tomas 2113c9de560dSAlex Tomas static int ext4_mb_seq_history_show(struct seq_file *seq, void *v) 2114c9de560dSAlex Tomas { 2115c9de560dSAlex Tomas char buf[25], buf2[25], buf3[25], *fmt; 2116c9de560dSAlex Tomas struct ext4_mb_history *hs = v; 2117c9de560dSAlex Tomas 2118c9de560dSAlex Tomas if (v == SEQ_START_TOKEN) { 2119c9de560dSAlex Tomas seq_printf(seq, "%-5s %-8s %-23s %-23s %-23s %-5s " 2120c9de560dSAlex Tomas "%-5s %-2s %-5s %-5s %-5s %-6s\n", 2121c9de560dSAlex Tomas "pid", "inode", "original", "goal", "result", "found", 2122c9de560dSAlex Tomas "grps", "cr", "flags", "merge", "tail", "broken"); 2123c9de560dSAlex Tomas return 0; 2124c9de560dSAlex Tomas } 2125c9de560dSAlex Tomas 2126c9de560dSAlex Tomas if (hs->op == EXT4_MB_HISTORY_ALLOC) { 2127c9de560dSAlex Tomas fmt = "%-5u %-8u %-23s %-23s %-23s %-5u %-5u %-2u " 2128c9de560dSAlex Tomas "%-5u %-5s %-5u %-6u\n"; 2129a9df9a49STheodore Ts'o sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group, 2130c9de560dSAlex Tomas hs->result.fe_start, hs->result.fe_len, 2131c9de560dSAlex Tomas hs->result.fe_logical); 2132a9df9a49STheodore Ts'o sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group, 2133c9de560dSAlex Tomas hs->orig.fe_start, hs->orig.fe_len, 2134c9de560dSAlex Tomas hs->orig.fe_logical); 2135a9df9a49STheodore Ts'o sprintf(buf3, "%u/%d/%u@%u", hs->goal.fe_group, 2136c9de560dSAlex Tomas hs->goal.fe_start, hs->goal.fe_len, 2137c9de560dSAlex Tomas hs->goal.fe_logical); 2138c9de560dSAlex Tomas seq_printf(seq, fmt, hs->pid, hs->ino, buf, buf3, buf2, 2139c9de560dSAlex Tomas hs->found, hs->groups, hs->cr, hs->flags, 2140c9de560dSAlex Tomas hs->merged ? "M" : "", hs->tail, 2141c9de560dSAlex Tomas hs->buddy ? 1 << hs->buddy : 0); 2142c9de560dSAlex Tomas } else if (hs->op == EXT4_MB_HISTORY_PREALLOC) { 2143c9de560dSAlex Tomas fmt = "%-5u %-8u %-23s %-23s %-23s\n"; 2144a9df9a49STheodore Ts'o sprintf(buf2, "%u/%d/%u@%u", hs->result.fe_group, 2145c9de560dSAlex Tomas hs->result.fe_start, hs->result.fe_len, 2146c9de560dSAlex Tomas hs->result.fe_logical); 2147a9df9a49STheodore Ts'o sprintf(buf, "%u/%d/%u@%u", hs->orig.fe_group, 2148c9de560dSAlex Tomas hs->orig.fe_start, hs->orig.fe_len, 2149c9de560dSAlex Tomas hs->orig.fe_logical); 2150c9de560dSAlex Tomas seq_printf(seq, fmt, hs->pid, hs->ino, buf, "", buf2); 2151c9de560dSAlex Tomas } else if (hs->op == EXT4_MB_HISTORY_DISCARD) { 2152a9df9a49STheodore Ts'o sprintf(buf2, "%u/%d/%u", hs->result.fe_group, 2153c9de560dSAlex Tomas hs->result.fe_start, hs->result.fe_len); 2154c9de560dSAlex Tomas seq_printf(seq, "%-5u %-8u %-23s discard\n", 2155c9de560dSAlex Tomas hs->pid, hs->ino, buf2); 2156c9de560dSAlex Tomas } else if (hs->op == EXT4_MB_HISTORY_FREE) { 2157a9df9a49STheodore Ts'o sprintf(buf2, "%u/%d/%u", hs->result.fe_group, 2158c9de560dSAlex Tomas hs->result.fe_start, hs->result.fe_len); 2159c9de560dSAlex Tomas seq_printf(seq, "%-5u %-8u %-23s free\n", 2160c9de560dSAlex Tomas hs->pid, hs->ino, buf2); 2161c9de560dSAlex Tomas } 2162c9de560dSAlex Tomas return 0; 2163c9de560dSAlex Tomas } 2164c9de560dSAlex Tomas 2165c9de560dSAlex Tomas static void ext4_mb_seq_history_stop(struct seq_file *seq, void *v) 2166c9de560dSAlex Tomas { 2167c9de560dSAlex Tomas } 2168c9de560dSAlex Tomas 2169c9de560dSAlex Tomas static struct seq_operations ext4_mb_seq_history_ops = { 2170c9de560dSAlex Tomas .start = ext4_mb_seq_history_start, 2171c9de560dSAlex Tomas .next = ext4_mb_seq_history_next, 2172c9de560dSAlex Tomas .stop = ext4_mb_seq_history_stop, 2173c9de560dSAlex Tomas .show = ext4_mb_seq_history_show, 2174c9de560dSAlex Tomas }; 2175c9de560dSAlex Tomas 2176c9de560dSAlex Tomas static int ext4_mb_seq_history_open(struct inode *inode, struct file *file) 2177c9de560dSAlex Tomas { 2178c9de560dSAlex Tomas struct super_block *sb = PDE(inode)->data; 2179c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2180c9de560dSAlex Tomas struct ext4_mb_proc_session *s; 2181c9de560dSAlex Tomas int rc; 2182c9de560dSAlex Tomas int size; 2183c9de560dSAlex Tomas 218474767c5aSShen Feng if (unlikely(sbi->s_mb_history == NULL)) 218574767c5aSShen Feng return -ENOMEM; 2186c9de560dSAlex Tomas s = kmalloc(sizeof(*s), GFP_KERNEL); 2187c9de560dSAlex Tomas if (s == NULL) 2188c9de560dSAlex Tomas return -ENOMEM; 2189c9de560dSAlex Tomas s->sb = sb; 2190c9de560dSAlex Tomas size = sizeof(struct ext4_mb_history) * sbi->s_mb_history_max; 2191c9de560dSAlex Tomas s->history = kmalloc(size, GFP_KERNEL); 2192c9de560dSAlex Tomas if (s->history == NULL) { 2193c9de560dSAlex Tomas kfree(s); 2194c9de560dSAlex Tomas return -ENOMEM; 2195c9de560dSAlex Tomas } 2196c9de560dSAlex Tomas 2197c9de560dSAlex Tomas spin_lock(&sbi->s_mb_history_lock); 2198c9de560dSAlex Tomas memcpy(s->history, sbi->s_mb_history, size); 2199c9de560dSAlex Tomas s->max = sbi->s_mb_history_max; 2200c9de560dSAlex Tomas s->start = sbi->s_mb_history_cur % s->max; 2201c9de560dSAlex Tomas spin_unlock(&sbi->s_mb_history_lock); 2202c9de560dSAlex Tomas 2203c9de560dSAlex Tomas rc = seq_open(file, &ext4_mb_seq_history_ops); 2204c9de560dSAlex Tomas if (rc == 0) { 2205c9de560dSAlex Tomas struct seq_file *m = (struct seq_file *)file->private_data; 2206c9de560dSAlex Tomas m->private = s; 2207c9de560dSAlex Tomas } else { 2208c9de560dSAlex Tomas kfree(s->history); 2209c9de560dSAlex Tomas kfree(s); 2210c9de560dSAlex Tomas } 2211c9de560dSAlex Tomas return rc; 2212c9de560dSAlex Tomas 2213c9de560dSAlex Tomas } 2214c9de560dSAlex Tomas 2215c9de560dSAlex Tomas static int ext4_mb_seq_history_release(struct inode *inode, struct file *file) 2216c9de560dSAlex Tomas { 2217c9de560dSAlex Tomas struct seq_file *seq = (struct seq_file *)file->private_data; 2218c9de560dSAlex Tomas struct ext4_mb_proc_session *s = seq->private; 2219c9de560dSAlex Tomas kfree(s->history); 2220c9de560dSAlex Tomas kfree(s); 2221c9de560dSAlex Tomas return seq_release(inode, file); 2222c9de560dSAlex Tomas } 2223c9de560dSAlex Tomas 2224c9de560dSAlex Tomas static ssize_t ext4_mb_seq_history_write(struct file *file, 2225c9de560dSAlex Tomas const char __user *buffer, 2226c9de560dSAlex Tomas size_t count, loff_t *ppos) 2227c9de560dSAlex Tomas { 2228c9de560dSAlex Tomas struct seq_file *seq = (struct seq_file *)file->private_data; 2229c9de560dSAlex Tomas struct ext4_mb_proc_session *s = seq->private; 2230c9de560dSAlex Tomas struct super_block *sb = s->sb; 2231c9de560dSAlex Tomas char str[32]; 2232c9de560dSAlex Tomas int value; 2233c9de560dSAlex Tomas 2234c9de560dSAlex Tomas if (count >= sizeof(str)) { 2235c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: %s string too long, max %u bytes\n", 2236c9de560dSAlex Tomas "mb_history", (int)sizeof(str)); 2237c9de560dSAlex Tomas return -EOVERFLOW; 2238c9de560dSAlex Tomas } 2239c9de560dSAlex Tomas 2240c9de560dSAlex Tomas if (copy_from_user(str, buffer, count)) 2241c9de560dSAlex Tomas return -EFAULT; 2242c9de560dSAlex Tomas 2243c9de560dSAlex Tomas value = simple_strtol(str, NULL, 0); 2244c9de560dSAlex Tomas if (value < 0) 2245c9de560dSAlex Tomas return -ERANGE; 2246c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_history_filter = value; 2247c9de560dSAlex Tomas 2248c9de560dSAlex Tomas return count; 2249c9de560dSAlex Tomas } 2250c9de560dSAlex Tomas 2251c9de560dSAlex Tomas static struct file_operations ext4_mb_seq_history_fops = { 2252c9de560dSAlex Tomas .owner = THIS_MODULE, 2253c9de560dSAlex Tomas .open = ext4_mb_seq_history_open, 2254c9de560dSAlex Tomas .read = seq_read, 2255c9de560dSAlex Tomas .write = ext4_mb_seq_history_write, 2256c9de560dSAlex Tomas .llseek = seq_lseek, 2257c9de560dSAlex Tomas .release = ext4_mb_seq_history_release, 2258c9de560dSAlex Tomas }; 2259c9de560dSAlex Tomas 2260c9de560dSAlex Tomas static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2261c9de560dSAlex Tomas { 2262c9de560dSAlex Tomas struct super_block *sb = seq->private; 2263c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2264c9de560dSAlex Tomas ext4_group_t group; 2265c9de560dSAlex Tomas 2266c9de560dSAlex Tomas if (*pos < 0 || *pos >= sbi->s_groups_count) 2267c9de560dSAlex Tomas return NULL; 2268c9de560dSAlex Tomas 2269c9de560dSAlex Tomas group = *pos + 1; 2270a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2271c9de560dSAlex Tomas } 2272c9de560dSAlex Tomas 2273c9de560dSAlex Tomas static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2274c9de560dSAlex Tomas { 2275c9de560dSAlex Tomas struct super_block *sb = seq->private; 2276c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2277c9de560dSAlex Tomas ext4_group_t group; 2278c9de560dSAlex Tomas 2279c9de560dSAlex Tomas ++*pos; 2280c9de560dSAlex Tomas if (*pos < 0 || *pos >= sbi->s_groups_count) 2281c9de560dSAlex Tomas return NULL; 2282c9de560dSAlex Tomas group = *pos + 1; 2283a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2284c9de560dSAlex Tomas } 2285c9de560dSAlex Tomas 2286c9de560dSAlex Tomas static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2287c9de560dSAlex Tomas { 2288c9de560dSAlex Tomas struct super_block *sb = seq->private; 2289a9df9a49STheodore Ts'o ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2290c9de560dSAlex Tomas int i; 2291c9de560dSAlex Tomas int err; 2292c9de560dSAlex Tomas struct ext4_buddy e4b; 2293c9de560dSAlex Tomas struct sg { 2294c9de560dSAlex Tomas struct ext4_group_info info; 2295c9de560dSAlex Tomas unsigned short counters[16]; 2296c9de560dSAlex Tomas } sg; 2297c9de560dSAlex Tomas 2298c9de560dSAlex Tomas group--; 2299c9de560dSAlex Tomas if (group == 0) 2300c9de560dSAlex Tomas seq_printf(seq, "#%-5s: %-5s %-5s %-5s " 2301c9de560dSAlex Tomas "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " 2302c9de560dSAlex Tomas "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", 2303c9de560dSAlex Tomas "group", "free", "frags", "first", 2304c9de560dSAlex Tomas "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", 2305c9de560dSAlex Tomas "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); 2306c9de560dSAlex Tomas 2307c9de560dSAlex Tomas i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 2308c9de560dSAlex Tomas sizeof(struct ext4_group_info); 2309c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2310c9de560dSAlex Tomas if (err) { 2311a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: I/O error\n", group); 2312c9de560dSAlex Tomas return 0; 2313c9de560dSAlex Tomas } 2314c9de560dSAlex Tomas ext4_lock_group(sb, group); 2315c9de560dSAlex Tomas memcpy(&sg, ext4_get_group_info(sb, group), i); 2316c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2317c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 2318c9de560dSAlex Tomas 2319a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 2320c9de560dSAlex Tomas sg.info.bb_fragments, sg.info.bb_first_free); 2321c9de560dSAlex Tomas for (i = 0; i <= 13; i++) 2322c9de560dSAlex Tomas seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? 2323c9de560dSAlex Tomas sg.info.bb_counters[i] : 0); 2324c9de560dSAlex Tomas seq_printf(seq, " ]\n"); 2325c9de560dSAlex Tomas 2326c9de560dSAlex Tomas return 0; 2327c9de560dSAlex Tomas } 2328c9de560dSAlex Tomas 2329c9de560dSAlex Tomas static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 2330c9de560dSAlex Tomas { 2331c9de560dSAlex Tomas } 2332c9de560dSAlex Tomas 2333c9de560dSAlex Tomas static struct seq_operations ext4_mb_seq_groups_ops = { 2334c9de560dSAlex Tomas .start = ext4_mb_seq_groups_start, 2335c9de560dSAlex Tomas .next = ext4_mb_seq_groups_next, 2336c9de560dSAlex Tomas .stop = ext4_mb_seq_groups_stop, 2337c9de560dSAlex Tomas .show = ext4_mb_seq_groups_show, 2338c9de560dSAlex Tomas }; 2339c9de560dSAlex Tomas 2340c9de560dSAlex Tomas static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) 2341c9de560dSAlex Tomas { 2342c9de560dSAlex Tomas struct super_block *sb = PDE(inode)->data; 2343c9de560dSAlex Tomas int rc; 2344c9de560dSAlex Tomas 2345c9de560dSAlex Tomas rc = seq_open(file, &ext4_mb_seq_groups_ops); 2346c9de560dSAlex Tomas if (rc == 0) { 2347c9de560dSAlex Tomas struct seq_file *m = (struct seq_file *)file->private_data; 2348c9de560dSAlex Tomas m->private = sb; 2349c9de560dSAlex Tomas } 2350c9de560dSAlex Tomas return rc; 2351c9de560dSAlex Tomas 2352c9de560dSAlex Tomas } 2353c9de560dSAlex Tomas 2354c9de560dSAlex Tomas static struct file_operations ext4_mb_seq_groups_fops = { 2355c9de560dSAlex Tomas .owner = THIS_MODULE, 2356c9de560dSAlex Tomas .open = ext4_mb_seq_groups_open, 2357c9de560dSAlex Tomas .read = seq_read, 2358c9de560dSAlex Tomas .llseek = seq_lseek, 2359c9de560dSAlex Tomas .release = seq_release, 2360c9de560dSAlex Tomas }; 2361c9de560dSAlex Tomas 2362c9de560dSAlex Tomas static void ext4_mb_history_release(struct super_block *sb) 2363c9de560dSAlex Tomas { 2364c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2365c9de560dSAlex Tomas 23669f6200bbSTheodore Ts'o if (sbi->s_proc != NULL) { 23679f6200bbSTheodore Ts'o remove_proc_entry("mb_groups", sbi->s_proc); 23689f6200bbSTheodore Ts'o remove_proc_entry("mb_history", sbi->s_proc); 23699f6200bbSTheodore Ts'o } 2370c9de560dSAlex Tomas kfree(sbi->s_mb_history); 2371c9de560dSAlex Tomas } 2372c9de560dSAlex Tomas 2373c9de560dSAlex Tomas static void ext4_mb_history_init(struct super_block *sb) 2374c9de560dSAlex Tomas { 2375c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2376c9de560dSAlex Tomas int i; 2377c9de560dSAlex Tomas 23789f6200bbSTheodore Ts'o if (sbi->s_proc != NULL) { 23799f6200bbSTheodore Ts'o proc_create_data("mb_history", S_IRUGO, sbi->s_proc, 238046fe74f2SDenis V. Lunev &ext4_mb_seq_history_fops, sb); 23819f6200bbSTheodore Ts'o proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, 238246fe74f2SDenis V. Lunev &ext4_mb_seq_groups_fops, sb); 2383c9de560dSAlex Tomas } 2384c9de560dSAlex Tomas 2385c9de560dSAlex Tomas sbi->s_mb_history_max = 1000; 2386c9de560dSAlex Tomas sbi->s_mb_history_cur = 0; 2387c9de560dSAlex Tomas spin_lock_init(&sbi->s_mb_history_lock); 2388c9de560dSAlex Tomas i = sbi->s_mb_history_max * sizeof(struct ext4_mb_history); 238974767c5aSShen Feng sbi->s_mb_history = kzalloc(i, GFP_KERNEL); 2390c9de560dSAlex Tomas /* if we can't allocate history, then we simple won't use it */ 2391c9de560dSAlex Tomas } 2392c9de560dSAlex Tomas 23934ddfef7bSEric Sandeen static noinline_for_stack void 23944ddfef7bSEric Sandeen ext4_mb_store_history(struct ext4_allocation_context *ac) 2395c9de560dSAlex Tomas { 2396c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 2397c9de560dSAlex Tomas struct ext4_mb_history h; 2398c9de560dSAlex Tomas 2399c9de560dSAlex Tomas if (unlikely(sbi->s_mb_history == NULL)) 2400c9de560dSAlex Tomas return; 2401c9de560dSAlex Tomas 2402c9de560dSAlex Tomas if (!(ac->ac_op & sbi->s_mb_history_filter)) 2403c9de560dSAlex Tomas return; 2404c9de560dSAlex Tomas 2405c9de560dSAlex Tomas h.op = ac->ac_op; 2406c9de560dSAlex Tomas h.pid = current->pid; 2407c9de560dSAlex Tomas h.ino = ac->ac_inode ? ac->ac_inode->i_ino : 0; 2408c9de560dSAlex Tomas h.orig = ac->ac_o_ex; 2409c9de560dSAlex Tomas h.result = ac->ac_b_ex; 2410c9de560dSAlex Tomas h.flags = ac->ac_flags; 2411c9de560dSAlex Tomas h.found = ac->ac_found; 2412c9de560dSAlex Tomas h.groups = ac->ac_groups_scanned; 2413c9de560dSAlex Tomas h.cr = ac->ac_criteria; 2414c9de560dSAlex Tomas h.tail = ac->ac_tail; 2415c9de560dSAlex Tomas h.buddy = ac->ac_buddy; 2416c9de560dSAlex Tomas h.merged = 0; 2417c9de560dSAlex Tomas if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) { 2418c9de560dSAlex Tomas if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 2419c9de560dSAlex Tomas ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 2420c9de560dSAlex Tomas h.merged = 1; 2421c9de560dSAlex Tomas h.goal = ac->ac_g_ex; 2422c9de560dSAlex Tomas h.result = ac->ac_f_ex; 2423c9de560dSAlex Tomas } 2424c9de560dSAlex Tomas 2425c9de560dSAlex Tomas spin_lock(&sbi->s_mb_history_lock); 2426c9de560dSAlex Tomas memcpy(sbi->s_mb_history + sbi->s_mb_history_cur, &h, sizeof(h)); 2427c9de560dSAlex Tomas if (++sbi->s_mb_history_cur >= sbi->s_mb_history_max) 2428c9de560dSAlex Tomas sbi->s_mb_history_cur = 0; 2429c9de560dSAlex Tomas spin_unlock(&sbi->s_mb_history_lock); 2430c9de560dSAlex Tomas } 2431c9de560dSAlex Tomas 2432c9de560dSAlex Tomas #else 2433c9de560dSAlex Tomas #define ext4_mb_history_release(sb) 2434c9de560dSAlex Tomas #define ext4_mb_history_init(sb) 2435c9de560dSAlex Tomas #endif 2436c9de560dSAlex Tomas 24375f21b0e6SFrederic Bohe 24385f21b0e6SFrederic Bohe /* Create and initialize ext4_group_info data for the given group. */ 2439920313a7SAneesh Kumar K.V int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 24405f21b0e6SFrederic Bohe struct ext4_group_desc *desc) 24415f21b0e6SFrederic Bohe { 24425f21b0e6SFrederic Bohe int i, len; 24435f21b0e6SFrederic Bohe int metalen = 0; 24445f21b0e6SFrederic Bohe struct ext4_sb_info *sbi = EXT4_SB(sb); 24455f21b0e6SFrederic Bohe struct ext4_group_info **meta_group_info; 24465f21b0e6SFrederic Bohe 24475f21b0e6SFrederic Bohe /* 24485f21b0e6SFrederic Bohe * First check if this group is the first of a reserved block. 24495f21b0e6SFrederic Bohe * If it's true, we have to allocate a new table of pointers 24505f21b0e6SFrederic Bohe * to ext4_group_info structures 24515f21b0e6SFrederic Bohe */ 24525f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 24535f21b0e6SFrederic Bohe metalen = sizeof(*meta_group_info) << 24545f21b0e6SFrederic Bohe EXT4_DESC_PER_BLOCK_BITS(sb); 24555f21b0e6SFrederic Bohe meta_group_info = kmalloc(metalen, GFP_KERNEL); 24565f21b0e6SFrederic Bohe if (meta_group_info == NULL) { 24575f21b0e6SFrederic Bohe printk(KERN_ERR "EXT4-fs: can't allocate mem for a " 24585f21b0e6SFrederic Bohe "buddy group\n"); 24595f21b0e6SFrederic Bohe goto exit_meta_group_info; 24605f21b0e6SFrederic Bohe } 24615f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = 24625f21b0e6SFrederic Bohe meta_group_info; 24635f21b0e6SFrederic Bohe } 24645f21b0e6SFrederic Bohe 24655f21b0e6SFrederic Bohe /* 24665f21b0e6SFrederic Bohe * calculate needed size. if change bb_counters size, 24675f21b0e6SFrederic Bohe * don't forget about ext4_mb_generate_buddy() 24685f21b0e6SFrederic Bohe */ 24695f21b0e6SFrederic Bohe len = offsetof(typeof(**meta_group_info), 24705f21b0e6SFrederic Bohe bb_counters[sb->s_blocksize_bits + 2]); 24715f21b0e6SFrederic Bohe 24725f21b0e6SFrederic Bohe meta_group_info = 24735f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; 24745f21b0e6SFrederic Bohe i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 24755f21b0e6SFrederic Bohe 24765f21b0e6SFrederic Bohe meta_group_info[i] = kzalloc(len, GFP_KERNEL); 24775f21b0e6SFrederic Bohe if (meta_group_info[i] == NULL) { 24785f21b0e6SFrederic Bohe printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n"); 24795f21b0e6SFrederic Bohe goto exit_group_info; 24805f21b0e6SFrederic Bohe } 24815f21b0e6SFrederic Bohe set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 24825f21b0e6SFrederic Bohe &(meta_group_info[i]->bb_state)); 24835f21b0e6SFrederic Bohe 24845f21b0e6SFrederic Bohe /* 24855f21b0e6SFrederic Bohe * initialize bb_free to be able to skip 24865f21b0e6SFrederic Bohe * empty groups without initialization 24875f21b0e6SFrederic Bohe */ 24885f21b0e6SFrederic Bohe if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 24895f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 24905f21b0e6SFrederic Bohe ext4_free_blocks_after_init(sb, group, desc); 24915f21b0e6SFrederic Bohe } else { 24925f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 24935f21b0e6SFrederic Bohe le16_to_cpu(desc->bg_free_blocks_count); 24945f21b0e6SFrederic Bohe } 24955f21b0e6SFrederic Bohe 24965f21b0e6SFrederic Bohe INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 2497920313a7SAneesh Kumar K.V init_rwsem(&meta_group_info[i]->alloc_sem); 2498c894058dSAneesh Kumar K.V meta_group_info[i]->bb_free_root.rb_node = NULL;; 24995f21b0e6SFrederic Bohe 25005f21b0e6SFrederic Bohe #ifdef DOUBLE_CHECK 25015f21b0e6SFrederic Bohe { 25025f21b0e6SFrederic Bohe struct buffer_head *bh; 25035f21b0e6SFrederic Bohe meta_group_info[i]->bb_bitmap = 25045f21b0e6SFrederic Bohe kmalloc(sb->s_blocksize, GFP_KERNEL); 25055f21b0e6SFrederic Bohe BUG_ON(meta_group_info[i]->bb_bitmap == NULL); 25065f21b0e6SFrederic Bohe bh = ext4_read_block_bitmap(sb, group); 25075f21b0e6SFrederic Bohe BUG_ON(bh == NULL); 25085f21b0e6SFrederic Bohe memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, 25095f21b0e6SFrederic Bohe sb->s_blocksize); 25105f21b0e6SFrederic Bohe put_bh(bh); 25115f21b0e6SFrederic Bohe } 25125f21b0e6SFrederic Bohe #endif 25135f21b0e6SFrederic Bohe 25145f21b0e6SFrederic Bohe return 0; 25155f21b0e6SFrederic Bohe 25165f21b0e6SFrederic Bohe exit_group_info: 25175f21b0e6SFrederic Bohe /* If a meta_group_info table has been allocated, release it now */ 25185f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) 25195f21b0e6SFrederic Bohe kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); 25205f21b0e6SFrederic Bohe exit_meta_group_info: 25215f21b0e6SFrederic Bohe return -ENOMEM; 25225f21b0e6SFrederic Bohe } /* ext4_mb_add_groupinfo */ 25235f21b0e6SFrederic Bohe 25245f21b0e6SFrederic Bohe /* 25255f21b0e6SFrederic Bohe * Update an existing group. 25265f21b0e6SFrederic Bohe * This function is used for online resize 25275f21b0e6SFrederic Bohe */ 25285f21b0e6SFrederic Bohe void ext4_mb_update_group_info(struct ext4_group_info *grp, ext4_grpblk_t add) 25295f21b0e6SFrederic Bohe { 25305f21b0e6SFrederic Bohe grp->bb_free += add; 25315f21b0e6SFrederic Bohe } 25325f21b0e6SFrederic Bohe 2533c9de560dSAlex Tomas static int ext4_mb_init_backend(struct super_block *sb) 2534c9de560dSAlex Tomas { 2535c9de560dSAlex Tomas ext4_group_t i; 25365f21b0e6SFrederic Bohe int metalen; 2537c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 25385f21b0e6SFrederic Bohe struct ext4_super_block *es = sbi->s_es; 25395f21b0e6SFrederic Bohe int num_meta_group_infos; 25405f21b0e6SFrederic Bohe int num_meta_group_infos_max; 25415f21b0e6SFrederic Bohe int array_size; 2542c9de560dSAlex Tomas struct ext4_group_info **meta_group_info; 25435f21b0e6SFrederic Bohe struct ext4_group_desc *desc; 2544c9de560dSAlex Tomas 25455f21b0e6SFrederic Bohe /* This is the number of blocks used by GDT */ 25465f21b0e6SFrederic Bohe num_meta_group_infos = (sbi->s_groups_count + EXT4_DESC_PER_BLOCK(sb) - 25475f21b0e6SFrederic Bohe 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); 25485f21b0e6SFrederic Bohe 25495f21b0e6SFrederic Bohe /* 25505f21b0e6SFrederic Bohe * This is the total number of blocks used by GDT including 25515f21b0e6SFrederic Bohe * the number of reserved blocks for GDT. 25525f21b0e6SFrederic Bohe * The s_group_info array is allocated with this value 25535f21b0e6SFrederic Bohe * to allow a clean online resize without a complex 25545f21b0e6SFrederic Bohe * manipulation of pointer. 25555f21b0e6SFrederic Bohe * The drawback is the unused memory when no resize 25565f21b0e6SFrederic Bohe * occurs but it's very low in terms of pages 25575f21b0e6SFrederic Bohe * (see comments below) 25585f21b0e6SFrederic Bohe * Need to handle this properly when META_BG resizing is allowed 25595f21b0e6SFrederic Bohe */ 25605f21b0e6SFrederic Bohe num_meta_group_infos_max = num_meta_group_infos + 25615f21b0e6SFrederic Bohe le16_to_cpu(es->s_reserved_gdt_blocks); 25625f21b0e6SFrederic Bohe 25635f21b0e6SFrederic Bohe /* 25645f21b0e6SFrederic Bohe * array_size is the size of s_group_info array. We round it 25655f21b0e6SFrederic Bohe * to the next power of two because this approximation is done 25665f21b0e6SFrederic Bohe * internally by kmalloc so we can have some more memory 25675f21b0e6SFrederic Bohe * for free here (e.g. may be used for META_BG resize). 25685f21b0e6SFrederic Bohe */ 25695f21b0e6SFrederic Bohe array_size = 1; 25705f21b0e6SFrederic Bohe while (array_size < sizeof(*sbi->s_group_info) * 25715f21b0e6SFrederic Bohe num_meta_group_infos_max) 25725f21b0e6SFrederic Bohe array_size = array_size << 1; 2573c9de560dSAlex Tomas /* An 8TB filesystem with 64-bit pointers requires a 4096 byte 2574c9de560dSAlex Tomas * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. 2575c9de560dSAlex Tomas * So a two level scheme suffices for now. */ 25765f21b0e6SFrederic Bohe sbi->s_group_info = kmalloc(array_size, GFP_KERNEL); 2577c9de560dSAlex Tomas if (sbi->s_group_info == NULL) { 2578c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n"); 2579c9de560dSAlex Tomas return -ENOMEM; 2580c9de560dSAlex Tomas } 2581c9de560dSAlex Tomas sbi->s_buddy_cache = new_inode(sb); 2582c9de560dSAlex Tomas if (sbi->s_buddy_cache == NULL) { 2583c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: can't get new inode\n"); 2584c9de560dSAlex Tomas goto err_freesgi; 2585c9de560dSAlex Tomas } 2586c9de560dSAlex Tomas EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 2587c9de560dSAlex Tomas 2588c9de560dSAlex Tomas metalen = sizeof(*meta_group_info) << EXT4_DESC_PER_BLOCK_BITS(sb); 2589c9de560dSAlex Tomas for (i = 0; i < num_meta_group_infos; i++) { 2590c9de560dSAlex Tomas if ((i + 1) == num_meta_group_infos) 2591c9de560dSAlex Tomas metalen = sizeof(*meta_group_info) * 2592c9de560dSAlex Tomas (sbi->s_groups_count - 2593c9de560dSAlex Tomas (i << EXT4_DESC_PER_BLOCK_BITS(sb))); 2594c9de560dSAlex Tomas meta_group_info = kmalloc(metalen, GFP_KERNEL); 2595c9de560dSAlex Tomas if (meta_group_info == NULL) { 2596c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: can't allocate mem for a " 2597c9de560dSAlex Tomas "buddy group\n"); 2598c9de560dSAlex Tomas goto err_freemeta; 2599c9de560dSAlex Tomas } 2600c9de560dSAlex Tomas sbi->s_group_info[i] = meta_group_info; 2601c9de560dSAlex Tomas } 2602c9de560dSAlex Tomas 2603c9de560dSAlex Tomas for (i = 0; i < sbi->s_groups_count; i++) { 2604c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, i, NULL); 2605c9de560dSAlex Tomas if (desc == NULL) { 2606c9de560dSAlex Tomas printk(KERN_ERR 2607a9df9a49STheodore Ts'o "EXT4-fs: can't read descriptor %u\n", i); 2608c9de560dSAlex Tomas goto err_freebuddy; 2609c9de560dSAlex Tomas } 26105f21b0e6SFrederic Bohe if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 26115f21b0e6SFrederic Bohe goto err_freebuddy; 2612c9de560dSAlex Tomas } 2613c9de560dSAlex Tomas 2614c9de560dSAlex Tomas return 0; 2615c9de560dSAlex Tomas 2616c9de560dSAlex Tomas err_freebuddy: 2617f1fa3342SRoel Kluin while (i-- > 0) 2618c9de560dSAlex Tomas kfree(ext4_get_group_info(sb, i)); 2619c9de560dSAlex Tomas i = num_meta_group_infos; 2620c9de560dSAlex Tomas err_freemeta: 2621f1fa3342SRoel Kluin while (i-- > 0) 2622c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2623c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2624c9de560dSAlex Tomas err_freesgi: 2625c9de560dSAlex Tomas kfree(sbi->s_group_info); 2626c9de560dSAlex Tomas return -ENOMEM; 2627c9de560dSAlex Tomas } 2628c9de560dSAlex Tomas 2629c9de560dSAlex Tomas int ext4_mb_init(struct super_block *sb, int needs_recovery) 2630c9de560dSAlex Tomas { 2631c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 26326be2ded1SAneesh Kumar K.V unsigned i, j; 2633c9de560dSAlex Tomas unsigned offset; 2634c9de560dSAlex Tomas unsigned max; 263574767c5aSShen Feng int ret; 2636c9de560dSAlex Tomas 2637c9de560dSAlex Tomas i = (sb->s_blocksize_bits + 2) * sizeof(unsigned short); 2638c9de560dSAlex Tomas 2639c9de560dSAlex Tomas sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 2640c9de560dSAlex Tomas if (sbi->s_mb_offsets == NULL) { 2641c9de560dSAlex Tomas return -ENOMEM; 2642c9de560dSAlex Tomas } 2643ff7ef329SYasunori Goto 2644ff7ef329SYasunori Goto i = (sb->s_blocksize_bits + 2) * sizeof(unsigned int); 2645c9de560dSAlex Tomas sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 2646c9de560dSAlex Tomas if (sbi->s_mb_maxs == NULL) { 2647c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 2648c9de560dSAlex Tomas return -ENOMEM; 2649c9de560dSAlex Tomas } 2650c9de560dSAlex Tomas 2651c9de560dSAlex Tomas /* order 0 is regular bitmap */ 2652c9de560dSAlex Tomas sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 2653c9de560dSAlex Tomas sbi->s_mb_offsets[0] = 0; 2654c9de560dSAlex Tomas 2655c9de560dSAlex Tomas i = 1; 2656c9de560dSAlex Tomas offset = 0; 2657c9de560dSAlex Tomas max = sb->s_blocksize << 2; 2658c9de560dSAlex Tomas do { 2659c9de560dSAlex Tomas sbi->s_mb_offsets[i] = offset; 2660c9de560dSAlex Tomas sbi->s_mb_maxs[i] = max; 2661c9de560dSAlex Tomas offset += 1 << (sb->s_blocksize_bits - i); 2662c9de560dSAlex Tomas max = max >> 1; 2663c9de560dSAlex Tomas i++; 2664c9de560dSAlex Tomas } while (i <= sb->s_blocksize_bits + 1); 2665c9de560dSAlex Tomas 2666c9de560dSAlex Tomas /* init file for buddy data */ 266774767c5aSShen Feng ret = ext4_mb_init_backend(sb); 266874767c5aSShen Feng if (ret != 0) { 2669c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2670c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 267174767c5aSShen Feng return ret; 2672c9de560dSAlex Tomas } 2673c9de560dSAlex Tomas 2674c9de560dSAlex Tomas spin_lock_init(&sbi->s_md_lock); 2675c9de560dSAlex Tomas spin_lock_init(&sbi->s_bal_lock); 2676c9de560dSAlex Tomas 2677c9de560dSAlex Tomas sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 2678c9de560dSAlex Tomas sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 2679c9de560dSAlex Tomas sbi->s_mb_stats = MB_DEFAULT_STATS; 2680c9de560dSAlex Tomas sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 2681c9de560dSAlex Tomas sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 2682c9de560dSAlex Tomas sbi->s_mb_history_filter = EXT4_MB_HISTORY_DEFAULT; 2683c9de560dSAlex Tomas sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; 2684c9de560dSAlex Tomas 2685730c213cSEric Sandeen sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 2686c9de560dSAlex Tomas if (sbi->s_locality_groups == NULL) { 2687c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2688c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 2689c9de560dSAlex Tomas return -ENOMEM; 2690c9de560dSAlex Tomas } 2691730c213cSEric Sandeen for_each_possible_cpu(i) { 2692c9de560dSAlex Tomas struct ext4_locality_group *lg; 2693730c213cSEric Sandeen lg = per_cpu_ptr(sbi->s_locality_groups, i); 2694c9de560dSAlex Tomas mutex_init(&lg->lg_mutex); 26956be2ded1SAneesh Kumar K.V for (j = 0; j < PREALLOC_TB_SIZE; j++) 26966be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 2697c9de560dSAlex Tomas spin_lock_init(&lg->lg_prealloc_lock); 2698c9de560dSAlex Tomas } 2699c9de560dSAlex Tomas 2700c9de560dSAlex Tomas ext4_mb_init_per_dev_proc(sb); 2701c9de560dSAlex Tomas ext4_mb_history_init(sb); 2702c9de560dSAlex Tomas 27030390131bSFrank Mayhar if (sbi->s_journal) 27043e624fc7STheodore Ts'o sbi->s_journal->j_commit_callback = release_blocks_on_commit; 27053e624fc7STheodore Ts'o 27064776004fSTheodore Ts'o printk(KERN_INFO "EXT4-fs: mballoc enabled\n"); 2707c9de560dSAlex Tomas return 0; 2708c9de560dSAlex Tomas } 2709c9de560dSAlex Tomas 2710c9de560dSAlex Tomas /* need to called with ext4 group lock (ext4_lock_group) */ 2711c9de560dSAlex Tomas static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) 2712c9de560dSAlex Tomas { 2713c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 2714c9de560dSAlex Tomas struct list_head *cur, *tmp; 2715c9de560dSAlex Tomas int count = 0; 2716c9de560dSAlex Tomas 2717c9de560dSAlex Tomas list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 2718c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 2719c9de560dSAlex Tomas list_del(&pa->pa_group_list); 2720c9de560dSAlex Tomas count++; 2721688f05a0SAneesh Kumar K.V kmem_cache_free(ext4_pspace_cachep, pa); 2722c9de560dSAlex Tomas } 2723c9de560dSAlex Tomas if (count) 2724c9de560dSAlex Tomas mb_debug("mballoc: %u PAs left\n", count); 2725c9de560dSAlex Tomas 2726c9de560dSAlex Tomas } 2727c9de560dSAlex Tomas 2728c9de560dSAlex Tomas int ext4_mb_release(struct super_block *sb) 2729c9de560dSAlex Tomas { 2730c9de560dSAlex Tomas ext4_group_t i; 2731c9de560dSAlex Tomas int num_meta_group_infos; 2732c9de560dSAlex Tomas struct ext4_group_info *grinfo; 2733c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2734c9de560dSAlex Tomas 2735c9de560dSAlex Tomas if (sbi->s_group_info) { 2736c9de560dSAlex Tomas for (i = 0; i < sbi->s_groups_count; i++) { 2737c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, i); 2738c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 2739c9de560dSAlex Tomas kfree(grinfo->bb_bitmap); 2740c9de560dSAlex Tomas #endif 2741c9de560dSAlex Tomas ext4_lock_group(sb, i); 2742c9de560dSAlex Tomas ext4_mb_cleanup_pa(grinfo); 2743c9de560dSAlex Tomas ext4_unlock_group(sb, i); 2744c9de560dSAlex Tomas kfree(grinfo); 2745c9de560dSAlex Tomas } 2746c9de560dSAlex Tomas num_meta_group_infos = (sbi->s_groups_count + 2747c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK(sb) - 1) >> 2748c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK_BITS(sb); 2749c9de560dSAlex Tomas for (i = 0; i < num_meta_group_infos; i++) 2750c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2751c9de560dSAlex Tomas kfree(sbi->s_group_info); 2752c9de560dSAlex Tomas } 2753c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2754c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 2755c9de560dSAlex Tomas if (sbi->s_buddy_cache) 2756c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2757c9de560dSAlex Tomas if (sbi->s_mb_stats) { 2758c9de560dSAlex Tomas printk(KERN_INFO 2759c9de560dSAlex Tomas "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n", 2760c9de560dSAlex Tomas atomic_read(&sbi->s_bal_allocated), 2761c9de560dSAlex Tomas atomic_read(&sbi->s_bal_reqs), 2762c9de560dSAlex Tomas atomic_read(&sbi->s_bal_success)); 2763c9de560dSAlex Tomas printk(KERN_INFO 2764c9de560dSAlex Tomas "EXT4-fs: mballoc: %u extents scanned, %u goal hits, " 2765c9de560dSAlex Tomas "%u 2^N hits, %u breaks, %u lost\n", 2766c9de560dSAlex Tomas atomic_read(&sbi->s_bal_ex_scanned), 2767c9de560dSAlex Tomas atomic_read(&sbi->s_bal_goals), 2768c9de560dSAlex Tomas atomic_read(&sbi->s_bal_2orders), 2769c9de560dSAlex Tomas atomic_read(&sbi->s_bal_breaks), 2770c9de560dSAlex Tomas atomic_read(&sbi->s_mb_lost_chunks)); 2771c9de560dSAlex Tomas printk(KERN_INFO 2772c9de560dSAlex Tomas "EXT4-fs: mballoc: %lu generated and it took %Lu\n", 2773c9de560dSAlex Tomas sbi->s_mb_buddies_generated++, 2774c9de560dSAlex Tomas sbi->s_mb_generation_time); 2775c9de560dSAlex Tomas printk(KERN_INFO 2776c9de560dSAlex Tomas "EXT4-fs: mballoc: %u preallocated, %u discarded\n", 2777c9de560dSAlex Tomas atomic_read(&sbi->s_mb_preallocated), 2778c9de560dSAlex Tomas atomic_read(&sbi->s_mb_discarded)); 2779c9de560dSAlex Tomas } 2780c9de560dSAlex Tomas 2781730c213cSEric Sandeen free_percpu(sbi->s_locality_groups); 2782c9de560dSAlex Tomas ext4_mb_history_release(sb); 2783c9de560dSAlex Tomas ext4_mb_destroy_per_dev_proc(sb); 2784c9de560dSAlex Tomas 2785c9de560dSAlex Tomas return 0; 2786c9de560dSAlex Tomas } 2787c9de560dSAlex Tomas 27883e624fc7STheodore Ts'o /* 27893e624fc7STheodore Ts'o * This function is called by the jbd2 layer once the commit has finished, 27903e624fc7STheodore Ts'o * so we know we can free the blocks that were released with that commit. 27913e624fc7STheodore Ts'o */ 27923e624fc7STheodore Ts'o static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) 2793c9de560dSAlex Tomas { 27943e624fc7STheodore Ts'o struct super_block *sb = journal->j_private; 2795c9de560dSAlex Tomas struct ext4_buddy e4b; 2796c894058dSAneesh Kumar K.V struct ext4_group_info *db; 2797c894058dSAneesh Kumar K.V int err, count = 0, count2 = 0; 2798c894058dSAneesh Kumar K.V struct ext4_free_data *entry; 27998a0aba73STheodore Ts'o ext4_fsblk_t discard_block; 28003e624fc7STheodore Ts'o struct list_head *l, *ltmp; 2801c9de560dSAlex Tomas 28023e624fc7STheodore Ts'o list_for_each_safe(l, ltmp, &txn->t_private_list) { 28033e624fc7STheodore Ts'o entry = list_entry(l, struct ext4_free_data, list); 2804c9de560dSAlex Tomas 2805a9df9a49STheodore Ts'o mb_debug("gonna free %u blocks in group %u (0x%p):", 2806c894058dSAneesh Kumar K.V entry->count, entry->group, entry); 2807c9de560dSAlex Tomas 2808c894058dSAneesh Kumar K.V err = ext4_mb_load_buddy(sb, entry->group, &e4b); 2809c9de560dSAlex Tomas /* we expect to find existing buddy because it's pinned */ 2810c9de560dSAlex Tomas BUG_ON(err != 0); 2811c9de560dSAlex Tomas 2812c894058dSAneesh Kumar K.V db = e4b.bd_info; 2813c9de560dSAlex Tomas /* there are blocks to put in buddy to make them really free */ 2814c894058dSAneesh Kumar K.V count += entry->count; 2815c9de560dSAlex Tomas count2++; 2816c894058dSAneesh Kumar K.V ext4_lock_group(sb, entry->group); 2817c894058dSAneesh Kumar K.V /* Take it out of per group rb tree */ 2818c894058dSAneesh Kumar K.V rb_erase(&entry->node, &(db->bb_free_root)); 2819c894058dSAneesh Kumar K.V mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); 2820c9de560dSAlex Tomas 2821c894058dSAneesh Kumar K.V if (!db->bb_free_root.rb_node) { 2822c894058dSAneesh Kumar K.V /* No more items in the per group rb tree 2823c894058dSAneesh Kumar K.V * balance refcounts from ext4_mb_free_metadata() 2824c894058dSAneesh Kumar K.V */ 2825c9de560dSAlex Tomas page_cache_release(e4b.bd_buddy_page); 2826c9de560dSAlex Tomas page_cache_release(e4b.bd_bitmap_page); 2827c894058dSAneesh Kumar K.V } 2828c894058dSAneesh Kumar K.V ext4_unlock_group(sb, entry->group); 28298a0aba73STheodore Ts'o discard_block = (ext4_fsblk_t) entry->group * EXT4_BLOCKS_PER_GROUP(sb) 28308a0aba73STheodore Ts'o + entry->start_blk 28318a0aba73STheodore Ts'o + le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 28328a0aba73STheodore Ts'o trace_mark(ext4_discard_blocks, "dev %s blk %llu count %u", sb->s_id, 28338a0aba73STheodore Ts'o (unsigned long long) discard_block, entry->count); 28348a0aba73STheodore Ts'o sb_issue_discard(sb, discard_block, entry->count); 2835c9de560dSAlex Tomas 2836c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 2837c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 28383e624fc7STheodore Ts'o } 2839c9de560dSAlex Tomas 2840c9de560dSAlex Tomas mb_debug("freed %u blocks in %u structures\n", count, count2); 2841c9de560dSAlex Tomas } 2842c9de560dSAlex Tomas 2843c9de560dSAlex Tomas #define EXT4_MB_STATS_NAME "stats" 2844c9de560dSAlex Tomas #define EXT4_MB_MAX_TO_SCAN_NAME "max_to_scan" 2845c9de560dSAlex Tomas #define EXT4_MB_MIN_TO_SCAN_NAME "min_to_scan" 2846c9de560dSAlex Tomas #define EXT4_MB_ORDER2_REQ "order2_req" 2847c9de560dSAlex Tomas #define EXT4_MB_STREAM_REQ "stream_req" 2848c9de560dSAlex Tomas #define EXT4_MB_GROUP_PREALLOC "group_prealloc" 2849c9de560dSAlex Tomas 2850c9de560dSAlex Tomas static int ext4_mb_init_per_dev_proc(struct super_block *sb) 2851c9de560dSAlex Tomas { 28520b09923eSManish Katiyar #ifdef CONFIG_PROC_FS 2853c9de560dSAlex Tomas mode_t mode = S_IFREG | S_IRUGO | S_IWUSR; 2854c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2855c9de560dSAlex Tomas struct proc_dir_entry *proc; 2856c9de560dSAlex Tomas 28579f6200bbSTheodore Ts'o if (sbi->s_proc == NULL) 2858cfbe7e4fSShen Feng return -EINVAL; 2859c9de560dSAlex Tomas 28605e8814f2STheodore Ts'o EXT4_PROC_HANDLER(EXT4_MB_STATS_NAME, mb_stats); 28615e8814f2STheodore Ts'o EXT4_PROC_HANDLER(EXT4_MB_MAX_TO_SCAN_NAME, mb_max_to_scan); 28625e8814f2STheodore Ts'o EXT4_PROC_HANDLER(EXT4_MB_MIN_TO_SCAN_NAME, mb_min_to_scan); 28635e8814f2STheodore Ts'o EXT4_PROC_HANDLER(EXT4_MB_ORDER2_REQ, mb_order2_reqs); 28645e8814f2STheodore Ts'o EXT4_PROC_HANDLER(EXT4_MB_STREAM_REQ, mb_stream_request); 28655e8814f2STheodore Ts'o EXT4_PROC_HANDLER(EXT4_MB_GROUP_PREALLOC, mb_group_prealloc); 2866c9de560dSAlex Tomas return 0; 2867c9de560dSAlex Tomas 2868c9de560dSAlex Tomas err_out: 28699f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_proc); 28709f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_proc); 28719f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_proc); 28729f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_proc); 28739f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_proc); 28749f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_proc); 2875c9de560dSAlex Tomas return -ENOMEM; 28760b09923eSManish Katiyar #else 28770b09923eSManish Katiyar return 0; 28780b09923eSManish Katiyar #endif 2879c9de560dSAlex Tomas } 2880c9de560dSAlex Tomas 2881c9de560dSAlex Tomas static int ext4_mb_destroy_per_dev_proc(struct super_block *sb) 2882c9de560dSAlex Tomas { 28830b09923eSManish Katiyar #ifdef CONFIG_PROC_FS 2884c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2885c9de560dSAlex Tomas 28869f6200bbSTheodore Ts'o if (sbi->s_proc == NULL) 2887c9de560dSAlex Tomas return -EINVAL; 2888c9de560dSAlex Tomas 28899f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_GROUP_PREALLOC, sbi->s_proc); 28909f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_STREAM_REQ, sbi->s_proc); 28919f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_ORDER2_REQ, sbi->s_proc); 28929f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_MIN_TO_SCAN_NAME, sbi->s_proc); 28939f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_MAX_TO_SCAN_NAME, sbi->s_proc); 28949f6200bbSTheodore Ts'o remove_proc_entry(EXT4_MB_STATS_NAME, sbi->s_proc); 28950b09923eSManish Katiyar #endif 2896c9de560dSAlex Tomas return 0; 2897c9de560dSAlex Tomas } 2898c9de560dSAlex Tomas 2899c9de560dSAlex Tomas int __init init_ext4_mballoc(void) 2900c9de560dSAlex Tomas { 2901c9de560dSAlex Tomas ext4_pspace_cachep = 2902c9de560dSAlex Tomas kmem_cache_create("ext4_prealloc_space", 2903c9de560dSAlex Tomas sizeof(struct ext4_prealloc_space), 2904c9de560dSAlex Tomas 0, SLAB_RECLAIM_ACCOUNT, NULL); 2905c9de560dSAlex Tomas if (ext4_pspace_cachep == NULL) 2906c9de560dSAlex Tomas return -ENOMEM; 2907c9de560dSAlex Tomas 2908256bdb49SEric Sandeen ext4_ac_cachep = 2909256bdb49SEric Sandeen kmem_cache_create("ext4_alloc_context", 2910256bdb49SEric Sandeen sizeof(struct ext4_allocation_context), 2911256bdb49SEric Sandeen 0, SLAB_RECLAIM_ACCOUNT, NULL); 2912256bdb49SEric Sandeen if (ext4_ac_cachep == NULL) { 2913256bdb49SEric Sandeen kmem_cache_destroy(ext4_pspace_cachep); 2914256bdb49SEric Sandeen return -ENOMEM; 2915256bdb49SEric Sandeen } 2916c894058dSAneesh Kumar K.V 2917c894058dSAneesh Kumar K.V ext4_free_ext_cachep = 2918c894058dSAneesh Kumar K.V kmem_cache_create("ext4_free_block_extents", 2919c894058dSAneesh Kumar K.V sizeof(struct ext4_free_data), 2920c894058dSAneesh Kumar K.V 0, SLAB_RECLAIM_ACCOUNT, NULL); 2921c894058dSAneesh Kumar K.V if (ext4_free_ext_cachep == NULL) { 2922c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_pspace_cachep); 2923c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_ac_cachep); 2924c894058dSAneesh Kumar K.V return -ENOMEM; 2925c894058dSAneesh Kumar K.V } 2926c9de560dSAlex Tomas return 0; 2927c9de560dSAlex Tomas } 2928c9de560dSAlex Tomas 2929c9de560dSAlex Tomas void exit_ext4_mballoc(void) 2930c9de560dSAlex Tomas { 2931c9de560dSAlex Tomas /* XXX: synchronize_rcu(); */ 2932c9de560dSAlex Tomas kmem_cache_destroy(ext4_pspace_cachep); 2933256bdb49SEric Sandeen kmem_cache_destroy(ext4_ac_cachep); 2934c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_free_ext_cachep); 2935c9de560dSAlex Tomas } 2936c9de560dSAlex Tomas 2937c9de560dSAlex Tomas 2938c9de560dSAlex Tomas /* 2939c9de560dSAlex Tomas * Check quota and mark choosed space (ac->ac_b_ex) non-free in bitmaps 2940c9de560dSAlex Tomas * Returns 0 if success or error code 2941c9de560dSAlex Tomas */ 29424ddfef7bSEric Sandeen static noinline_for_stack int 29434ddfef7bSEric Sandeen ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 2944498e5f24STheodore Ts'o handle_t *handle, unsigned int reserv_blks) 2945c9de560dSAlex Tomas { 2946c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 2947c9de560dSAlex Tomas struct ext4_super_block *es; 2948c9de560dSAlex Tomas struct ext4_group_desc *gdp; 2949c9de560dSAlex Tomas struct buffer_head *gdp_bh; 2950c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2951c9de560dSAlex Tomas struct super_block *sb; 2952c9de560dSAlex Tomas ext4_fsblk_t block; 2953519deca0SAneesh Kumar K.V int err, len; 2954c9de560dSAlex Tomas 2955c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 2956c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len <= 0); 2957c9de560dSAlex Tomas 2958c9de560dSAlex Tomas sb = ac->ac_sb; 2959c9de560dSAlex Tomas sbi = EXT4_SB(sb); 2960c9de560dSAlex Tomas es = sbi->s_es; 2961c9de560dSAlex Tomas 2962c9de560dSAlex Tomas 2963c9de560dSAlex Tomas err = -EIO; 2964574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 2965c9de560dSAlex Tomas if (!bitmap_bh) 2966c9de560dSAlex Tomas goto out_err; 2967c9de560dSAlex Tomas 2968c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 2969c9de560dSAlex Tomas if (err) 2970c9de560dSAlex Tomas goto out_err; 2971c9de560dSAlex Tomas 2972c9de560dSAlex Tomas err = -EIO; 2973c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 2974c9de560dSAlex Tomas if (!gdp) 2975c9de560dSAlex Tomas goto out_err; 2976c9de560dSAlex Tomas 2977a9df9a49STheodore Ts'o ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 297803cddb80SAneesh Kumar K.V gdp->bg_free_blocks_count); 297903cddb80SAneesh Kumar K.V 2980c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gdp_bh); 2981c9de560dSAlex Tomas if (err) 2982c9de560dSAlex Tomas goto out_err; 2983c9de560dSAlex Tomas 2984c9de560dSAlex Tomas block = ac->ac_b_ex.fe_group * EXT4_BLOCKS_PER_GROUP(sb) 2985c9de560dSAlex Tomas + ac->ac_b_ex.fe_start 2986c9de560dSAlex Tomas + le32_to_cpu(es->s_first_data_block); 2987c9de560dSAlex Tomas 2988519deca0SAneesh Kumar K.V len = ac->ac_b_ex.fe_len; 2989519deca0SAneesh Kumar K.V if (in_range(ext4_block_bitmap(sb, gdp), block, len) || 2990519deca0SAneesh Kumar K.V in_range(ext4_inode_bitmap(sb, gdp), block, len) || 2991c9de560dSAlex Tomas in_range(block, ext4_inode_table(sb, gdp), 2992519deca0SAneesh Kumar K.V EXT4_SB(sb)->s_itb_per_group) || 2993519deca0SAneesh Kumar K.V in_range(block + len - 1, ext4_inode_table(sb, gdp), 2994c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group)) { 299546e665e9SHarvey Harrison ext4_error(sb, __func__, 2996c9de560dSAlex Tomas "Allocating block in system zone - block = %llu", 2997c9de560dSAlex Tomas block); 2998519deca0SAneesh Kumar K.V /* File system mounted not to panic on error 2999519deca0SAneesh Kumar K.V * Fix the bitmap and repeat the block allocation 3000519deca0SAneesh Kumar K.V * We leak some of the blocks here. 3001519deca0SAneesh Kumar K.V */ 3002519deca0SAneesh Kumar K.V mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), 3003519deca0SAneesh Kumar K.V bitmap_bh->b_data, ac->ac_b_ex.fe_start, 3004519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len); 30050390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 3006519deca0SAneesh Kumar K.V if (!err) 3007519deca0SAneesh Kumar K.V err = -EAGAIN; 3008519deca0SAneesh Kumar K.V goto out_err; 3009c9de560dSAlex Tomas } 3010c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 3011c9de560dSAlex Tomas { 3012c9de560dSAlex Tomas int i; 3013c9de560dSAlex Tomas for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 3014c9de560dSAlex Tomas BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 3015c9de560dSAlex Tomas bitmap_bh->b_data)); 3016c9de560dSAlex Tomas } 3017c9de560dSAlex Tomas } 3018c9de560dSAlex Tomas #endif 3019c9de560dSAlex Tomas mb_set_bits(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group), bitmap_bh->b_data, 3020c9de560dSAlex Tomas ac->ac_b_ex.fe_start, ac->ac_b_ex.fe_len); 3021c9de560dSAlex Tomas 3022c9de560dSAlex Tomas spin_lock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group)); 3023c9de560dSAlex Tomas if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 3024c9de560dSAlex Tomas gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 3025c9de560dSAlex Tomas gdp->bg_free_blocks_count = 3026c9de560dSAlex Tomas cpu_to_le16(ext4_free_blocks_after_init(sb, 3027c9de560dSAlex Tomas ac->ac_b_ex.fe_group, 3028c9de560dSAlex Tomas gdp)); 3029c9de560dSAlex Tomas } 3030e8546d06SMarcin Slusarz le16_add_cpu(&gdp->bg_free_blocks_count, -ac->ac_b_ex.fe_len); 3031c9de560dSAlex Tomas gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); 3032c9de560dSAlex Tomas spin_unlock(sb_bgl_lock(sbi, ac->ac_b_ex.fe_group)); 30336bc6e63fSAneesh Kumar K.V percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); 3034d2a17637SMingming Cao /* 30356bc6e63fSAneesh Kumar K.V * Now reduce the dirty block count also. Should not go negative 3036d2a17637SMingming Cao */ 30376bc6e63fSAneesh Kumar K.V if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 30386bc6e63fSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 30396bc6e63fSAneesh Kumar K.V percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks); 30406bc6e63fSAneesh Kumar K.V else 30416bc6e63fSAneesh Kumar K.V percpu_counter_sub(&sbi->s_dirtyblocks_counter, 30426bc6e63fSAneesh Kumar K.V ac->ac_b_ex.fe_len); 3043c9de560dSAlex Tomas 3044772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 3045772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, 3046772cb7c8SJose R. Santos ac->ac_b_ex.fe_group); 3047772cb7c8SJose R. Santos spin_lock(sb_bgl_lock(sbi, flex_group)); 3048772cb7c8SJose R. Santos sbi->s_flex_groups[flex_group].free_blocks -= ac->ac_b_ex.fe_len; 3049772cb7c8SJose R. Santos spin_unlock(sb_bgl_lock(sbi, flex_group)); 3050772cb7c8SJose R. Santos } 3051772cb7c8SJose R. Santos 30520390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 3053c9de560dSAlex Tomas if (err) 3054c9de560dSAlex Tomas goto out_err; 30550390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 3056c9de560dSAlex Tomas 3057c9de560dSAlex Tomas out_err: 3058c9de560dSAlex Tomas sb->s_dirt = 1; 305942a10addSAneesh Kumar K.V brelse(bitmap_bh); 3060c9de560dSAlex Tomas return err; 3061c9de560dSAlex Tomas } 3062c9de560dSAlex Tomas 3063c9de560dSAlex Tomas /* 3064c9de560dSAlex Tomas * here we normalize request for locality group 3065c9de560dSAlex Tomas * Group request are normalized to s_strip size if we set the same via mount 3066c9de560dSAlex Tomas * option. If not we set it to s_mb_group_prealloc which can be configured via 3067c9de560dSAlex Tomas * /proc/fs/ext4/<partition>/group_prealloc 3068c9de560dSAlex Tomas * 3069c9de560dSAlex Tomas * XXX: should we try to preallocate more than the group has now? 3070c9de560dSAlex Tomas */ 3071c9de560dSAlex Tomas static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 3072c9de560dSAlex Tomas { 3073c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3074c9de560dSAlex Tomas struct ext4_locality_group *lg = ac->ac_lg; 3075c9de560dSAlex Tomas 3076c9de560dSAlex Tomas BUG_ON(lg == NULL); 3077c9de560dSAlex Tomas if (EXT4_SB(sb)->s_stripe) 3078c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe; 3079c9de560dSAlex Tomas else 3080c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 308160bd63d1SSolofo Ramangalahy mb_debug("#%u: goal %u blocks for locality group\n", 3082c9de560dSAlex Tomas current->pid, ac->ac_g_ex.fe_len); 3083c9de560dSAlex Tomas } 3084c9de560dSAlex Tomas 3085c9de560dSAlex Tomas /* 3086c9de560dSAlex Tomas * Normalization means making request better in terms of 3087c9de560dSAlex Tomas * size and alignment 3088c9de560dSAlex Tomas */ 30894ddfef7bSEric Sandeen static noinline_for_stack void 30904ddfef7bSEric Sandeen ext4_mb_normalize_request(struct ext4_allocation_context *ac, 3091c9de560dSAlex Tomas struct ext4_allocation_request *ar) 3092c9de560dSAlex Tomas { 3093c9de560dSAlex Tomas int bsbits, max; 3094c9de560dSAlex Tomas ext4_lblk_t end; 3095c9de560dSAlex Tomas loff_t size, orig_size, start_off; 3096c9de560dSAlex Tomas ext4_lblk_t start, orig_start; 3097c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 30989a0762c5SAneesh Kumar K.V struct ext4_prealloc_space *pa; 3099c9de560dSAlex Tomas 3100c9de560dSAlex Tomas /* do normalize only data requests, metadata requests 3101c9de560dSAlex Tomas do not need preallocation */ 3102c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3103c9de560dSAlex Tomas return; 3104c9de560dSAlex Tomas 3105c9de560dSAlex Tomas /* sometime caller may want exact blocks */ 3106c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 3107c9de560dSAlex Tomas return; 3108c9de560dSAlex Tomas 3109c9de560dSAlex Tomas /* caller may indicate that preallocation isn't 3110c9de560dSAlex Tomas * required (it's a tail, for example) */ 3111c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 3112c9de560dSAlex Tomas return; 3113c9de560dSAlex Tomas 3114c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 3115c9de560dSAlex Tomas ext4_mb_normalize_group_request(ac); 3116c9de560dSAlex Tomas return ; 3117c9de560dSAlex Tomas } 3118c9de560dSAlex Tomas 3119c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 3120c9de560dSAlex Tomas 3121c9de560dSAlex Tomas /* first, let's learn actual file size 3122c9de560dSAlex Tomas * given current request is allocated */ 3123c9de560dSAlex Tomas size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; 3124c9de560dSAlex Tomas size = size << bsbits; 3125c9de560dSAlex Tomas if (size < i_size_read(ac->ac_inode)) 3126c9de560dSAlex Tomas size = i_size_read(ac->ac_inode); 3127c9de560dSAlex Tomas 31281930479cSValerie Clement /* max size of free chunks */ 31291930479cSValerie Clement max = 2 << bsbits; 3130c9de560dSAlex Tomas 31311930479cSValerie Clement #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 31321930479cSValerie Clement (req <= (size) || max <= (chunk_size)) 3133c9de560dSAlex Tomas 3134c9de560dSAlex Tomas /* first, try to predict filesize */ 3135c9de560dSAlex Tomas /* XXX: should this table be tunable? */ 3136c9de560dSAlex Tomas start_off = 0; 3137c9de560dSAlex Tomas if (size <= 16 * 1024) { 3138c9de560dSAlex Tomas size = 16 * 1024; 3139c9de560dSAlex Tomas } else if (size <= 32 * 1024) { 3140c9de560dSAlex Tomas size = 32 * 1024; 3141c9de560dSAlex Tomas } else if (size <= 64 * 1024) { 3142c9de560dSAlex Tomas size = 64 * 1024; 3143c9de560dSAlex Tomas } else if (size <= 128 * 1024) { 3144c9de560dSAlex Tomas size = 128 * 1024; 3145c9de560dSAlex Tomas } else if (size <= 256 * 1024) { 3146c9de560dSAlex Tomas size = 256 * 1024; 3147c9de560dSAlex Tomas } else if (size <= 512 * 1024) { 3148c9de560dSAlex Tomas size = 512 * 1024; 3149c9de560dSAlex Tomas } else if (size <= 1024 * 1024) { 3150c9de560dSAlex Tomas size = 1024 * 1024; 31511930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 3152c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 31531930479cSValerie Clement (21 - bsbits)) << 21; 31541930479cSValerie Clement size = 2 * 1024 * 1024; 31551930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 3156c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3157c9de560dSAlex Tomas (22 - bsbits)) << 22; 3158c9de560dSAlex Tomas size = 4 * 1024 * 1024; 3159c9de560dSAlex Tomas } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, 31601930479cSValerie Clement (8<<20)>>bsbits, max, 8 * 1024)) { 3161c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 3162c9de560dSAlex Tomas (23 - bsbits)) << 23; 3163c9de560dSAlex Tomas size = 8 * 1024 * 1024; 3164c9de560dSAlex Tomas } else { 3165c9de560dSAlex Tomas start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; 3166c9de560dSAlex Tomas size = ac->ac_o_ex.fe_len << bsbits; 3167c9de560dSAlex Tomas } 3168c9de560dSAlex Tomas orig_size = size = size >> bsbits; 3169c9de560dSAlex Tomas orig_start = start = start_off >> bsbits; 3170c9de560dSAlex Tomas 3171c9de560dSAlex Tomas /* don't cover already allocated blocks in selected range */ 3172c9de560dSAlex Tomas if (ar->pleft && start <= ar->lleft) { 3173c9de560dSAlex Tomas size -= ar->lleft + 1 - start; 3174c9de560dSAlex Tomas start = ar->lleft + 1; 3175c9de560dSAlex Tomas } 3176c9de560dSAlex Tomas if (ar->pright && start + size - 1 >= ar->lright) 3177c9de560dSAlex Tomas size -= start + size - ar->lright; 3178c9de560dSAlex Tomas 3179c9de560dSAlex Tomas end = start + size; 3180c9de560dSAlex Tomas 3181c9de560dSAlex Tomas /* check we don't cross already preallocated blocks */ 3182c9de560dSAlex Tomas rcu_read_lock(); 31839a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3184498e5f24STheodore Ts'o ext4_lblk_t pa_end; 3185c9de560dSAlex Tomas 3186c9de560dSAlex Tomas if (pa->pa_deleted) 3187c9de560dSAlex Tomas continue; 3188c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3189c9de560dSAlex Tomas if (pa->pa_deleted) { 3190c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3191c9de560dSAlex Tomas continue; 3192c9de560dSAlex Tomas } 3193c9de560dSAlex Tomas 3194c9de560dSAlex Tomas pa_end = pa->pa_lstart + pa->pa_len; 3195c9de560dSAlex Tomas 3196c9de560dSAlex Tomas /* PA must not overlap original request */ 3197c9de560dSAlex Tomas BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || 3198c9de560dSAlex Tomas ac->ac_o_ex.fe_logical < pa->pa_lstart)); 3199c9de560dSAlex Tomas 3200c9de560dSAlex Tomas /* skip PA normalized request doesn't overlap with */ 3201c9de560dSAlex Tomas if (pa->pa_lstart >= end) { 3202c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3203c9de560dSAlex Tomas continue; 3204c9de560dSAlex Tomas } 3205c9de560dSAlex Tomas if (pa_end <= start) { 3206c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3207c9de560dSAlex Tomas continue; 3208c9de560dSAlex Tomas } 3209c9de560dSAlex Tomas BUG_ON(pa->pa_lstart <= start && pa_end >= end); 3210c9de560dSAlex Tomas 3211c9de560dSAlex Tomas if (pa_end <= ac->ac_o_ex.fe_logical) { 3212c9de560dSAlex Tomas BUG_ON(pa_end < start); 3213c9de560dSAlex Tomas start = pa_end; 3214c9de560dSAlex Tomas } 3215c9de560dSAlex Tomas 3216c9de560dSAlex Tomas if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { 3217c9de560dSAlex Tomas BUG_ON(pa->pa_lstart > end); 3218c9de560dSAlex Tomas end = pa->pa_lstart; 3219c9de560dSAlex Tomas } 3220c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3221c9de560dSAlex Tomas } 3222c9de560dSAlex Tomas rcu_read_unlock(); 3223c9de560dSAlex Tomas size = end - start; 3224c9de560dSAlex Tomas 3225c9de560dSAlex Tomas /* XXX: extra loop to check we really don't overlap preallocations */ 3226c9de560dSAlex Tomas rcu_read_lock(); 32279a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3228498e5f24STheodore Ts'o ext4_lblk_t pa_end; 3229c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3230c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 3231c9de560dSAlex Tomas pa_end = pa->pa_lstart + pa->pa_len; 3232c9de560dSAlex Tomas BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); 3233c9de560dSAlex Tomas } 3234c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3235c9de560dSAlex Tomas } 3236c9de560dSAlex Tomas rcu_read_unlock(); 3237c9de560dSAlex Tomas 3238c9de560dSAlex Tomas if (start + size <= ac->ac_o_ex.fe_logical && 3239c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical) { 3240c9de560dSAlex Tomas printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n", 3241c9de560dSAlex Tomas (unsigned long) start, (unsigned long) size, 3242c9de560dSAlex Tomas (unsigned long) ac->ac_o_ex.fe_logical); 3243c9de560dSAlex Tomas } 3244c9de560dSAlex Tomas BUG_ON(start + size <= ac->ac_o_ex.fe_logical && 3245c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical); 3246c9de560dSAlex Tomas BUG_ON(size <= 0 || size >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 3247c9de560dSAlex Tomas 3248c9de560dSAlex Tomas /* now prepare goal request */ 3249c9de560dSAlex Tomas 3250c9de560dSAlex Tomas /* XXX: is it better to align blocks WRT to logical 3251c9de560dSAlex Tomas * placement or satisfy big request as is */ 3252c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = start; 3253c9de560dSAlex Tomas ac->ac_g_ex.fe_len = size; 3254c9de560dSAlex Tomas 3255c9de560dSAlex Tomas /* define goal start in order to merge */ 3256c9de560dSAlex Tomas if (ar->pright && (ar->lright == (start + size))) { 3257c9de560dSAlex Tomas /* merge to the right */ 3258c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 3259c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 3260c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 3261c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3262c9de560dSAlex Tomas } 3263c9de560dSAlex Tomas if (ar->pleft && (ar->lleft + 1 == start)) { 3264c9de560dSAlex Tomas /* merge to the left */ 3265c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 3266c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 3267c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 3268c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3269c9de560dSAlex Tomas } 3270c9de560dSAlex Tomas 3271c9de560dSAlex Tomas mb_debug("goal: %u(was %u) blocks at %u\n", (unsigned) size, 3272c9de560dSAlex Tomas (unsigned) orig_size, (unsigned) start); 3273c9de560dSAlex Tomas } 3274c9de560dSAlex Tomas 3275c9de560dSAlex Tomas static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 3276c9de560dSAlex Tomas { 3277c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3278c9de560dSAlex Tomas 3279c9de560dSAlex Tomas if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { 3280c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_reqs); 3281c9de560dSAlex Tomas atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 3282c9de560dSAlex Tomas if (ac->ac_o_ex.fe_len >= ac->ac_g_ex.fe_len) 3283c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_success); 3284c9de560dSAlex Tomas atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 3285c9de560dSAlex Tomas if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 3286c9de560dSAlex Tomas ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 3287c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_goals); 3288c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan) 3289c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_breaks); 3290c9de560dSAlex Tomas } 3291c9de560dSAlex Tomas 3292c9de560dSAlex Tomas ext4_mb_store_history(ac); 3293c9de560dSAlex Tomas } 3294c9de560dSAlex Tomas 3295c9de560dSAlex Tomas /* 3296c9de560dSAlex Tomas * use blocks preallocated to inode 3297c9de560dSAlex Tomas */ 3298c9de560dSAlex Tomas static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 3299c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3300c9de560dSAlex Tomas { 3301c9de560dSAlex Tomas ext4_fsblk_t start; 3302c9de560dSAlex Tomas ext4_fsblk_t end; 3303c9de560dSAlex Tomas int len; 3304c9de560dSAlex Tomas 3305c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3306c9de560dSAlex Tomas start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 3307c9de560dSAlex Tomas end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len); 3308c9de560dSAlex Tomas len = end - start; 3309c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 3310c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3311c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3312c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3313c9de560dSAlex Tomas ac->ac_pa = pa; 3314c9de560dSAlex Tomas 3315c9de560dSAlex Tomas BUG_ON(start < pa->pa_pstart); 3316c9de560dSAlex Tomas BUG_ON(start + len > pa->pa_pstart + pa->pa_len); 3317c9de560dSAlex Tomas BUG_ON(pa->pa_free < len); 3318c9de560dSAlex Tomas pa->pa_free -= len; 3319c9de560dSAlex Tomas 332060bd63d1SSolofo Ramangalahy mb_debug("use %llu/%u from inode pa %p\n", start, len, pa); 3321c9de560dSAlex Tomas } 3322c9de560dSAlex Tomas 3323c9de560dSAlex Tomas /* 3324c9de560dSAlex Tomas * use blocks preallocated to locality group 3325c9de560dSAlex Tomas */ 3326c9de560dSAlex Tomas static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 3327c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3328c9de560dSAlex Tomas { 332903cddb80SAneesh Kumar K.V unsigned int len = ac->ac_o_ex.fe_len; 33306be2ded1SAneesh Kumar K.V 3331c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 3332c9de560dSAlex Tomas &ac->ac_b_ex.fe_group, 3333c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3334c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3335c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3336c9de560dSAlex Tomas ac->ac_pa = pa; 3337c9de560dSAlex Tomas 3338c9de560dSAlex Tomas /* we don't correct pa_pstart or pa_plen here to avoid 333926346ff6SAneesh Kumar K.V * possible race when the group is being loaded concurrently 3340c9de560dSAlex Tomas * instead we correct pa later, after blocks are marked 334126346ff6SAneesh Kumar K.V * in on-disk bitmap -- see ext4_mb_release_context() 334226346ff6SAneesh Kumar K.V * Other CPUs are prevented from allocating from this pa by lg_mutex 3343c9de560dSAlex Tomas */ 3344c9de560dSAlex Tomas mb_debug("use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); 3345c9de560dSAlex Tomas } 3346c9de560dSAlex Tomas 3347c9de560dSAlex Tomas /* 33485e745b04SAneesh Kumar K.V * Return the prealloc space that have minimal distance 33495e745b04SAneesh Kumar K.V * from the goal block. @cpa is the prealloc 33505e745b04SAneesh Kumar K.V * space that is having currently known minimal distance 33515e745b04SAneesh Kumar K.V * from the goal block. 33525e745b04SAneesh Kumar K.V */ 33535e745b04SAneesh Kumar K.V static struct ext4_prealloc_space * 33545e745b04SAneesh Kumar K.V ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 33555e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, 33565e745b04SAneesh Kumar K.V struct ext4_prealloc_space *cpa) 33575e745b04SAneesh Kumar K.V { 33585e745b04SAneesh Kumar K.V ext4_fsblk_t cur_distance, new_distance; 33595e745b04SAneesh Kumar K.V 33605e745b04SAneesh Kumar K.V if (cpa == NULL) { 33615e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 33625e745b04SAneesh Kumar K.V return pa; 33635e745b04SAneesh Kumar K.V } 33645e745b04SAneesh Kumar K.V cur_distance = abs(goal_block - cpa->pa_pstart); 33655e745b04SAneesh Kumar K.V new_distance = abs(goal_block - pa->pa_pstart); 33665e745b04SAneesh Kumar K.V 33675e745b04SAneesh Kumar K.V if (cur_distance < new_distance) 33685e745b04SAneesh Kumar K.V return cpa; 33695e745b04SAneesh Kumar K.V 33705e745b04SAneesh Kumar K.V /* drop the previous reference */ 33715e745b04SAneesh Kumar K.V atomic_dec(&cpa->pa_count); 33725e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 33735e745b04SAneesh Kumar K.V return pa; 33745e745b04SAneesh Kumar K.V } 33755e745b04SAneesh Kumar K.V 33765e745b04SAneesh Kumar K.V /* 3377c9de560dSAlex Tomas * search goal blocks in preallocated space 3378c9de560dSAlex Tomas */ 33794ddfef7bSEric Sandeen static noinline_for_stack int 33804ddfef7bSEric Sandeen ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 3381c9de560dSAlex Tomas { 33826be2ded1SAneesh Kumar K.V int order, i; 3383c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3384c9de560dSAlex Tomas struct ext4_locality_group *lg; 33855e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, *cpa = NULL; 33865e745b04SAneesh Kumar K.V ext4_fsblk_t goal_block; 3387c9de560dSAlex Tomas 3388c9de560dSAlex Tomas /* only data can be preallocated */ 3389c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3390c9de560dSAlex Tomas return 0; 3391c9de560dSAlex Tomas 3392c9de560dSAlex Tomas /* first, try per-file preallocation */ 3393c9de560dSAlex Tomas rcu_read_lock(); 33949a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3395c9de560dSAlex Tomas 3396c9de560dSAlex Tomas /* all fields in this condition don't change, 3397c9de560dSAlex Tomas * so we can skip locking for them */ 3398c9de560dSAlex Tomas if (ac->ac_o_ex.fe_logical < pa->pa_lstart || 3399c9de560dSAlex Tomas ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len) 3400c9de560dSAlex Tomas continue; 3401c9de560dSAlex Tomas 3402c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3403c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3404c9de560dSAlex Tomas if (pa->pa_deleted == 0 && pa->pa_free) { 3405c9de560dSAlex Tomas atomic_inc(&pa->pa_count); 3406c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 3407c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3408c9de560dSAlex Tomas ac->ac_criteria = 10; 3409c9de560dSAlex Tomas rcu_read_unlock(); 3410c9de560dSAlex Tomas return 1; 3411c9de560dSAlex Tomas } 3412c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3413c9de560dSAlex Tomas } 3414c9de560dSAlex Tomas rcu_read_unlock(); 3415c9de560dSAlex Tomas 3416c9de560dSAlex Tomas /* can we use group allocation? */ 3417c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 3418c9de560dSAlex Tomas return 0; 3419c9de560dSAlex Tomas 3420c9de560dSAlex Tomas /* inode may have no locality group for some reason */ 3421c9de560dSAlex Tomas lg = ac->ac_lg; 3422c9de560dSAlex Tomas if (lg == NULL) 3423c9de560dSAlex Tomas return 0; 34246be2ded1SAneesh Kumar K.V order = fls(ac->ac_o_ex.fe_len) - 1; 34256be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 34266be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 34276be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 3428c9de560dSAlex Tomas 34295e745b04SAneesh Kumar K.V goal_block = ac->ac_g_ex.fe_group * EXT4_BLOCKS_PER_GROUP(ac->ac_sb) + 34305e745b04SAneesh Kumar K.V ac->ac_g_ex.fe_start + 34315e745b04SAneesh Kumar K.V le32_to_cpu(EXT4_SB(ac->ac_sb)->s_es->s_first_data_block); 34325e745b04SAneesh Kumar K.V /* 34335e745b04SAneesh Kumar K.V * search for the prealloc space that is having 34345e745b04SAneesh Kumar K.V * minimal distance from the goal block. 34355e745b04SAneesh Kumar K.V */ 34366be2ded1SAneesh Kumar K.V for (i = order; i < PREALLOC_TB_SIZE; i++) { 3437c9de560dSAlex Tomas rcu_read_lock(); 34386be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], 34396be2ded1SAneesh Kumar K.V pa_inode_list) { 3440c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 34416be2ded1SAneesh Kumar K.V if (pa->pa_deleted == 0 && 34426be2ded1SAneesh Kumar K.V pa->pa_free >= ac->ac_o_ex.fe_len) { 34435e745b04SAneesh Kumar K.V 34445e745b04SAneesh Kumar K.V cpa = ext4_mb_check_group_pa(goal_block, 34455e745b04SAneesh Kumar K.V pa, cpa); 34465e745b04SAneesh Kumar K.V } 3447c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 34485e745b04SAneesh Kumar K.V } 34495e745b04SAneesh Kumar K.V rcu_read_unlock(); 34505e745b04SAneesh Kumar K.V } 34515e745b04SAneesh Kumar K.V if (cpa) { 34525e745b04SAneesh Kumar K.V ext4_mb_use_group_pa(ac, cpa); 3453c9de560dSAlex Tomas ac->ac_criteria = 20; 3454c9de560dSAlex Tomas return 1; 3455c9de560dSAlex Tomas } 3456c9de560dSAlex Tomas return 0; 3457c9de560dSAlex Tomas } 3458c9de560dSAlex Tomas 3459c9de560dSAlex Tomas /* 3460c9de560dSAlex Tomas * the function goes through all preallocation in this group and marks them 3461c9de560dSAlex Tomas * used in in-core bitmap. buddy must be generated from this bitmap 3462c9de560dSAlex Tomas * Need to be called with ext4 group lock (ext4_lock_group) 3463c9de560dSAlex Tomas */ 3464c9de560dSAlex Tomas static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 3465c9de560dSAlex Tomas ext4_group_t group) 3466c9de560dSAlex Tomas { 3467c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3468c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3469c9de560dSAlex Tomas struct list_head *cur; 3470c9de560dSAlex Tomas ext4_group_t groupnr; 3471c9de560dSAlex Tomas ext4_grpblk_t start; 3472c9de560dSAlex Tomas int preallocated = 0; 3473c9de560dSAlex Tomas int count = 0; 3474c9de560dSAlex Tomas int len; 3475c9de560dSAlex Tomas 3476c9de560dSAlex Tomas /* all form of preallocation discards first load group, 3477c9de560dSAlex Tomas * so the only competing code is preallocation use. 3478c9de560dSAlex Tomas * we don't need any locking here 3479c9de560dSAlex Tomas * notice we do NOT ignore preallocations with pa_deleted 3480c9de560dSAlex Tomas * otherwise we could leave used blocks available for 3481c9de560dSAlex Tomas * allocation in buddy when concurrent ext4_mb_put_pa() 3482c9de560dSAlex Tomas * is dropping preallocation 3483c9de560dSAlex Tomas */ 3484c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 3485c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3486c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3487c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3488c9de560dSAlex Tomas &groupnr, &start); 3489c9de560dSAlex Tomas len = pa->pa_len; 3490c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3491c9de560dSAlex Tomas if (unlikely(len == 0)) 3492c9de560dSAlex Tomas continue; 3493c9de560dSAlex Tomas BUG_ON(groupnr != group); 3494c9de560dSAlex Tomas mb_set_bits(sb_bgl_lock(EXT4_SB(sb), group), 3495c9de560dSAlex Tomas bitmap, start, len); 3496c9de560dSAlex Tomas preallocated += len; 3497c9de560dSAlex Tomas count++; 3498c9de560dSAlex Tomas } 3499a9df9a49STheodore Ts'o mb_debug("prellocated %u for group %u\n", preallocated, group); 3500c9de560dSAlex Tomas } 3501c9de560dSAlex Tomas 3502c9de560dSAlex Tomas static void ext4_mb_pa_callback(struct rcu_head *head) 3503c9de560dSAlex Tomas { 3504c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3505c9de560dSAlex Tomas pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 3506c9de560dSAlex Tomas kmem_cache_free(ext4_pspace_cachep, pa); 3507c9de560dSAlex Tomas } 3508c9de560dSAlex Tomas 3509c9de560dSAlex Tomas /* 3510c9de560dSAlex Tomas * drops a reference to preallocated space descriptor 3511c9de560dSAlex Tomas * if this was the last reference and the space is consumed 3512c9de560dSAlex Tomas */ 3513c9de560dSAlex Tomas static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 3514c9de560dSAlex Tomas struct super_block *sb, struct ext4_prealloc_space *pa) 3515c9de560dSAlex Tomas { 3516a9df9a49STheodore Ts'o ext4_group_t grp; 3517c9de560dSAlex Tomas 3518c9de560dSAlex Tomas if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) 3519c9de560dSAlex Tomas return; 3520c9de560dSAlex Tomas 3521c9de560dSAlex Tomas /* in this short window concurrent discard can set pa_deleted */ 3522c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3523c9de560dSAlex Tomas if (pa->pa_deleted == 1) { 3524c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3525c9de560dSAlex Tomas return; 3526c9de560dSAlex Tomas } 3527c9de560dSAlex Tomas 3528c9de560dSAlex Tomas pa->pa_deleted = 1; 3529c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3530c9de560dSAlex Tomas 3531c9de560dSAlex Tomas /* -1 is to protect from crossing allocation group */ 3532c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart - 1, &grp, NULL); 3533c9de560dSAlex Tomas 3534c9de560dSAlex Tomas /* 3535c9de560dSAlex Tomas * possible race: 3536c9de560dSAlex Tomas * 3537c9de560dSAlex Tomas * P1 (buddy init) P2 (regular allocation) 3538c9de560dSAlex Tomas * find block B in PA 3539c9de560dSAlex Tomas * copy on-disk bitmap to buddy 3540c9de560dSAlex Tomas * mark B in on-disk bitmap 3541c9de560dSAlex Tomas * drop PA from group 3542c9de560dSAlex Tomas * mark all PAs in buddy 3543c9de560dSAlex Tomas * 3544c9de560dSAlex Tomas * thus, P1 initializes buddy with B available. to prevent this 3545c9de560dSAlex Tomas * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 3546c9de560dSAlex Tomas * against that pair 3547c9de560dSAlex Tomas */ 3548c9de560dSAlex Tomas ext4_lock_group(sb, grp); 3549c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3550c9de560dSAlex Tomas ext4_unlock_group(sb, grp); 3551c9de560dSAlex Tomas 3552c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3553c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3554c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3555c9de560dSAlex Tomas 3556c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3557c9de560dSAlex Tomas } 3558c9de560dSAlex Tomas 3559c9de560dSAlex Tomas /* 3560c9de560dSAlex Tomas * creates new preallocated space for given inode 3561c9de560dSAlex Tomas */ 35624ddfef7bSEric Sandeen static noinline_for_stack int 35634ddfef7bSEric Sandeen ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 3564c9de560dSAlex Tomas { 3565c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3566c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3567c9de560dSAlex Tomas struct ext4_group_info *grp; 3568c9de560dSAlex Tomas struct ext4_inode_info *ei; 3569c9de560dSAlex Tomas 3570c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3571c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3572c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3573c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3574c9de560dSAlex Tomas 3575c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3576c9de560dSAlex Tomas if (pa == NULL) 3577c9de560dSAlex Tomas return -ENOMEM; 3578c9de560dSAlex Tomas 3579c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { 3580c9de560dSAlex Tomas int winl; 3581c9de560dSAlex Tomas int wins; 3582c9de560dSAlex Tomas int win; 3583c9de560dSAlex Tomas int offs; 3584c9de560dSAlex Tomas 3585c9de560dSAlex Tomas /* we can't allocate as much as normalizer wants. 3586c9de560dSAlex Tomas * so, found space must get proper lstart 3587c9de560dSAlex Tomas * to cover original request */ 3588c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 3589c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 3590c9de560dSAlex Tomas 3591c9de560dSAlex Tomas /* we're limited by original request in that 3592c9de560dSAlex Tomas * logical block must be covered any way 3593c9de560dSAlex Tomas * winl is window we can move our chunk within */ 3594c9de560dSAlex Tomas winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; 3595c9de560dSAlex Tomas 3596c9de560dSAlex Tomas /* also, we should cover whole original request */ 3597c9de560dSAlex Tomas wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len; 3598c9de560dSAlex Tomas 3599c9de560dSAlex Tomas /* the smallest one defines real window */ 3600c9de560dSAlex Tomas win = min(winl, wins); 3601c9de560dSAlex Tomas 3602c9de560dSAlex Tomas offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len; 3603c9de560dSAlex Tomas if (offs && offs < win) 3604c9de560dSAlex Tomas win = offs; 3605c9de560dSAlex Tomas 3606c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win; 3607c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 3608c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 3609c9de560dSAlex Tomas } 3610c9de560dSAlex Tomas 3611c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3612c9de560dSAlex Tomas * allocated blocks for history */ 3613c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3614c9de560dSAlex Tomas 3615c9de560dSAlex Tomas pa->pa_lstart = ac->ac_b_ex.fe_logical; 3616c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3617c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3618c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3619c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3620c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 3621c9de560dSAlex Tomas pa->pa_deleted = 0; 3622c9de560dSAlex Tomas pa->pa_linear = 0; 3623c9de560dSAlex Tomas 3624c9de560dSAlex Tomas mb_debug("new inode pa %p: %llu/%u for %u\n", pa, 3625c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 3626c9de560dSAlex Tomas 3627c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 3628c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3629c9de560dSAlex Tomas 3630c9de560dSAlex Tomas ei = EXT4_I(ac->ac_inode); 3631c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3632c9de560dSAlex Tomas 3633c9de560dSAlex Tomas pa->pa_obj_lock = &ei->i_prealloc_lock; 3634c9de560dSAlex Tomas pa->pa_inode = ac->ac_inode; 3635c9de560dSAlex Tomas 3636c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3637c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3638c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3639c9de560dSAlex Tomas 3640c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3641c9de560dSAlex Tomas list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); 3642c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3643c9de560dSAlex Tomas 3644c9de560dSAlex Tomas return 0; 3645c9de560dSAlex Tomas } 3646c9de560dSAlex Tomas 3647c9de560dSAlex Tomas /* 3648c9de560dSAlex Tomas * creates new preallocated space for locality group inodes belongs to 3649c9de560dSAlex Tomas */ 36504ddfef7bSEric Sandeen static noinline_for_stack int 36514ddfef7bSEric Sandeen ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 3652c9de560dSAlex Tomas { 3653c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3654c9de560dSAlex Tomas struct ext4_locality_group *lg; 3655c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3656c9de560dSAlex Tomas struct ext4_group_info *grp; 3657c9de560dSAlex Tomas 3658c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3659c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3660c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3661c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3662c9de560dSAlex Tomas 3663c9de560dSAlex Tomas BUG_ON(ext4_pspace_cachep == NULL); 3664c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3665c9de560dSAlex Tomas if (pa == NULL) 3666c9de560dSAlex Tomas return -ENOMEM; 3667c9de560dSAlex Tomas 3668c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3669c9de560dSAlex Tomas * allocated blocks for history */ 3670c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3671c9de560dSAlex Tomas 3672c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3673c9de560dSAlex Tomas pa->pa_lstart = pa->pa_pstart; 3674c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3675c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3676c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3677c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 36786be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3679c9de560dSAlex Tomas pa->pa_deleted = 0; 3680c9de560dSAlex Tomas pa->pa_linear = 1; 3681c9de560dSAlex Tomas 3682c9de560dSAlex Tomas mb_debug("new group pa %p: %llu/%u for %u\n", pa, 3683c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 3684c9de560dSAlex Tomas 3685c9de560dSAlex Tomas ext4_mb_use_group_pa(ac, pa); 3686c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3687c9de560dSAlex Tomas 3688c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3689c9de560dSAlex Tomas lg = ac->ac_lg; 3690c9de560dSAlex Tomas BUG_ON(lg == NULL); 3691c9de560dSAlex Tomas 3692c9de560dSAlex Tomas pa->pa_obj_lock = &lg->lg_prealloc_lock; 3693c9de560dSAlex Tomas pa->pa_inode = NULL; 3694c9de560dSAlex Tomas 3695c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3696c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3697c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3698c9de560dSAlex Tomas 36996be2ded1SAneesh Kumar K.V /* 37006be2ded1SAneesh Kumar K.V * We will later add the new pa to the right bucket 37016be2ded1SAneesh Kumar K.V * after updating the pa_free in ext4_mb_release_context 37026be2ded1SAneesh Kumar K.V */ 3703c9de560dSAlex Tomas return 0; 3704c9de560dSAlex Tomas } 3705c9de560dSAlex Tomas 3706c9de560dSAlex Tomas static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 3707c9de560dSAlex Tomas { 3708c9de560dSAlex Tomas int err; 3709c9de560dSAlex Tomas 3710c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 3711c9de560dSAlex Tomas err = ext4_mb_new_group_pa(ac); 3712c9de560dSAlex Tomas else 3713c9de560dSAlex Tomas err = ext4_mb_new_inode_pa(ac); 3714c9de560dSAlex Tomas return err; 3715c9de560dSAlex Tomas } 3716c9de560dSAlex Tomas 3717c9de560dSAlex Tomas /* 3718c9de560dSAlex Tomas * finds all unused blocks in on-disk bitmap, frees them in 3719c9de560dSAlex Tomas * in-core bitmap and buddy. 3720c9de560dSAlex Tomas * @pa must be unlinked from inode and group lists, so that 3721c9de560dSAlex Tomas * nobody else can find/use it. 3722c9de560dSAlex Tomas * the caller MUST hold group/inode locks. 3723c9de560dSAlex Tomas * TODO: optimize the case when there are no in-core structures yet 3724c9de560dSAlex Tomas */ 37254ddfef7bSEric Sandeen static noinline_for_stack int 37264ddfef7bSEric Sandeen ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 3727c83617dbSAneesh Kumar K.V struct ext4_prealloc_space *pa, 3728c83617dbSAneesh Kumar K.V struct ext4_allocation_context *ac) 3729c9de560dSAlex Tomas { 3730c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3731c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3732498e5f24STheodore Ts'o unsigned int end; 3733498e5f24STheodore Ts'o unsigned int next; 3734c9de560dSAlex Tomas ext4_group_t group; 3735c9de560dSAlex Tomas ext4_grpblk_t bit; 3736c9de560dSAlex Tomas sector_t start; 3737c9de560dSAlex Tomas int err = 0; 3738c9de560dSAlex Tomas int free = 0; 3739c9de560dSAlex Tomas 3740c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3741c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3742c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3743c9de560dSAlex Tomas end = bit + pa->pa_len; 3744c9de560dSAlex Tomas 3745256bdb49SEric Sandeen if (ac) { 3746256bdb49SEric Sandeen ac->ac_sb = sb; 3747256bdb49SEric Sandeen ac->ac_inode = pa->pa_inode; 3748256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_DISCARD; 3749256bdb49SEric Sandeen } 3750c9de560dSAlex Tomas 3751c9de560dSAlex Tomas while (bit < end) { 3752ffad0a44SAneesh Kumar K.V bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 3753c9de560dSAlex Tomas if (bit >= end) 3754c9de560dSAlex Tomas break; 3755ffad0a44SAneesh Kumar K.V next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 3756c9de560dSAlex Tomas start = group * EXT4_BLOCKS_PER_GROUP(sb) + bit + 3757c9de560dSAlex Tomas le32_to_cpu(sbi->s_es->s_first_data_block); 3758c9de560dSAlex Tomas mb_debug(" free preallocated %u/%u in group %u\n", 3759c9de560dSAlex Tomas (unsigned) start, (unsigned) next - bit, 3760c9de560dSAlex Tomas (unsigned) group); 3761c9de560dSAlex Tomas free += next - bit; 3762c9de560dSAlex Tomas 3763256bdb49SEric Sandeen if (ac) { 3764256bdb49SEric Sandeen ac->ac_b_ex.fe_group = group; 3765256bdb49SEric Sandeen ac->ac_b_ex.fe_start = bit; 3766256bdb49SEric Sandeen ac->ac_b_ex.fe_len = next - bit; 3767256bdb49SEric Sandeen ac->ac_b_ex.fe_logical = 0; 3768256bdb49SEric Sandeen ext4_mb_store_history(ac); 3769256bdb49SEric Sandeen } 3770c9de560dSAlex Tomas 3771c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 3772c9de560dSAlex Tomas bit = next + 1; 3773c9de560dSAlex Tomas } 3774c9de560dSAlex Tomas if (free != pa->pa_free) { 377526346ff6SAneesh Kumar K.V printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n", 3776c9de560dSAlex Tomas pa, (unsigned long) pa->pa_lstart, 3777c9de560dSAlex Tomas (unsigned long) pa->pa_pstart, 3778c9de560dSAlex Tomas (unsigned long) pa->pa_len); 3779fde4d95aSTheodore Ts'o ext4_error(sb, __func__, "free %u, pa_free %u", 378026346ff6SAneesh Kumar K.V free, pa->pa_free); 3781e56eb659SAneesh Kumar K.V /* 3782e56eb659SAneesh Kumar K.V * pa is already deleted so we use the value obtained 3783e56eb659SAneesh Kumar K.V * from the bitmap and continue. 3784e56eb659SAneesh Kumar K.V */ 3785c9de560dSAlex Tomas } 3786c9de560dSAlex Tomas atomic_add(free, &sbi->s_mb_discarded); 3787c9de560dSAlex Tomas 3788c9de560dSAlex Tomas return err; 3789c9de560dSAlex Tomas } 3790c9de560dSAlex Tomas 37914ddfef7bSEric Sandeen static noinline_for_stack int 37924ddfef7bSEric Sandeen ext4_mb_release_group_pa(struct ext4_buddy *e4b, 3793c83617dbSAneesh Kumar K.V struct ext4_prealloc_space *pa, 3794c83617dbSAneesh Kumar K.V struct ext4_allocation_context *ac) 3795c9de560dSAlex Tomas { 3796c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3797c9de560dSAlex Tomas ext4_group_t group; 3798c9de560dSAlex Tomas ext4_grpblk_t bit; 3799c9de560dSAlex Tomas 3800256bdb49SEric Sandeen if (ac) 3801256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_DISCARD; 3802c9de560dSAlex Tomas 3803c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3804c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3805c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3806c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 3807c9de560dSAlex Tomas atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 3808c9de560dSAlex Tomas 3809256bdb49SEric Sandeen if (ac) { 3810256bdb49SEric Sandeen ac->ac_sb = sb; 3811256bdb49SEric Sandeen ac->ac_inode = NULL; 3812256bdb49SEric Sandeen ac->ac_b_ex.fe_group = group; 3813256bdb49SEric Sandeen ac->ac_b_ex.fe_start = bit; 3814256bdb49SEric Sandeen ac->ac_b_ex.fe_len = pa->pa_len; 3815256bdb49SEric Sandeen ac->ac_b_ex.fe_logical = 0; 3816256bdb49SEric Sandeen ext4_mb_store_history(ac); 3817256bdb49SEric Sandeen } 3818c9de560dSAlex Tomas 3819c9de560dSAlex Tomas return 0; 3820c9de560dSAlex Tomas } 3821c9de560dSAlex Tomas 3822c9de560dSAlex Tomas /* 3823c9de560dSAlex Tomas * releases all preallocations in given group 3824c9de560dSAlex Tomas * 3825c9de560dSAlex Tomas * first, we need to decide discard policy: 3826c9de560dSAlex Tomas * - when do we discard 3827c9de560dSAlex Tomas * 1) ENOSPC 3828c9de560dSAlex Tomas * - how many do we discard 3829c9de560dSAlex Tomas * 1) how many requested 3830c9de560dSAlex Tomas */ 38314ddfef7bSEric Sandeen static noinline_for_stack int 38324ddfef7bSEric Sandeen ext4_mb_discard_group_preallocations(struct super_block *sb, 3833c9de560dSAlex Tomas ext4_group_t group, int needed) 3834c9de560dSAlex Tomas { 3835c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3836c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3837c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3838c83617dbSAneesh Kumar K.V struct ext4_allocation_context *ac; 3839c9de560dSAlex Tomas struct list_head list; 3840c9de560dSAlex Tomas struct ext4_buddy e4b; 3841c9de560dSAlex Tomas int err; 3842c9de560dSAlex Tomas int busy = 0; 3843c9de560dSAlex Tomas int free = 0; 3844c9de560dSAlex Tomas 3845a9df9a49STheodore Ts'o mb_debug("discard preallocation for group %u\n", group); 3846c9de560dSAlex Tomas 3847c9de560dSAlex Tomas if (list_empty(&grp->bb_prealloc_list)) 3848c9de560dSAlex Tomas return 0; 3849c9de560dSAlex Tomas 3850574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 3851c9de560dSAlex Tomas if (bitmap_bh == NULL) { 3852ce89f46cSAneesh Kumar K.V ext4_error(sb, __func__, "Error in reading block " 3853a9df9a49STheodore Ts'o "bitmap for %u", group); 3854ce89f46cSAneesh Kumar K.V return 0; 3855c9de560dSAlex Tomas } 3856c9de560dSAlex Tomas 3857c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3858ce89f46cSAneesh Kumar K.V if (err) { 3859ce89f46cSAneesh Kumar K.V ext4_error(sb, __func__, "Error in loading buddy " 3860a9df9a49STheodore Ts'o "information for %u", group); 3861ce89f46cSAneesh Kumar K.V put_bh(bitmap_bh); 3862ce89f46cSAneesh Kumar K.V return 0; 3863ce89f46cSAneesh Kumar K.V } 3864c9de560dSAlex Tomas 3865c9de560dSAlex Tomas if (needed == 0) 3866c9de560dSAlex Tomas needed = EXT4_BLOCKS_PER_GROUP(sb) + 1; 3867c9de560dSAlex Tomas 3868c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3869c83617dbSAneesh Kumar K.V ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 3870c9de560dSAlex Tomas repeat: 3871c9de560dSAlex Tomas ext4_lock_group(sb, group); 3872c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, 3873c9de560dSAlex Tomas &grp->bb_prealloc_list, pa_group_list) { 3874c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3875c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3876c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3877c9de560dSAlex Tomas busy = 1; 3878c9de560dSAlex Tomas continue; 3879c9de560dSAlex Tomas } 3880c9de560dSAlex Tomas if (pa->pa_deleted) { 3881c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3882c9de560dSAlex Tomas continue; 3883c9de560dSAlex Tomas } 3884c9de560dSAlex Tomas 3885c9de560dSAlex Tomas /* seems this one can be freed ... */ 3886c9de560dSAlex Tomas pa->pa_deleted = 1; 3887c9de560dSAlex Tomas 3888c9de560dSAlex Tomas /* we can trust pa_free ... */ 3889c9de560dSAlex Tomas free += pa->pa_free; 3890c9de560dSAlex Tomas 3891c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3892c9de560dSAlex Tomas 3893c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3894c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3895c9de560dSAlex Tomas } 3896c9de560dSAlex Tomas 3897c9de560dSAlex Tomas /* if we still need more blocks and some PAs were used, try again */ 3898c9de560dSAlex Tomas if (free < needed && busy) { 3899c9de560dSAlex Tomas busy = 0; 3900c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3901c9de560dSAlex Tomas /* 3902c9de560dSAlex Tomas * Yield the CPU here so that we don't get soft lockup 3903c9de560dSAlex Tomas * in non preempt case. 3904c9de560dSAlex Tomas */ 3905c9de560dSAlex Tomas yield(); 3906c9de560dSAlex Tomas goto repeat; 3907c9de560dSAlex Tomas } 3908c9de560dSAlex Tomas 3909c9de560dSAlex Tomas /* found anything to free? */ 3910c9de560dSAlex Tomas if (list_empty(&list)) { 3911c9de560dSAlex Tomas BUG_ON(free != 0); 3912c9de560dSAlex Tomas goto out; 3913c9de560dSAlex Tomas } 3914c9de560dSAlex Tomas 3915c9de560dSAlex Tomas /* now free all selected PAs */ 3916c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3917c9de560dSAlex Tomas 3918c9de560dSAlex Tomas /* remove from object (inode or locality group) */ 3919c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3920c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3921c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3922c9de560dSAlex Tomas 3923c9de560dSAlex Tomas if (pa->pa_linear) 3924c83617dbSAneesh Kumar K.V ext4_mb_release_group_pa(&e4b, pa, ac); 3925c9de560dSAlex Tomas else 3926c83617dbSAneesh Kumar K.V ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); 3927c9de560dSAlex Tomas 3928c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 3929c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3930c9de560dSAlex Tomas } 3931c9de560dSAlex Tomas 3932c9de560dSAlex Tomas out: 3933c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3934c83617dbSAneesh Kumar K.V if (ac) 3935c83617dbSAneesh Kumar K.V kmem_cache_free(ext4_ac_cachep, ac); 3936c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 3937c9de560dSAlex Tomas put_bh(bitmap_bh); 3938c9de560dSAlex Tomas return free; 3939c9de560dSAlex Tomas } 3940c9de560dSAlex Tomas 3941c9de560dSAlex Tomas /* 3942c9de560dSAlex Tomas * releases all non-used preallocated blocks for given inode 3943c9de560dSAlex Tomas * 3944c9de560dSAlex Tomas * It's important to discard preallocations under i_data_sem 3945c9de560dSAlex Tomas * We don't want another block to be served from the prealloc 3946c9de560dSAlex Tomas * space when we are discarding the inode prealloc space. 3947c9de560dSAlex Tomas * 3948c9de560dSAlex Tomas * FIXME!! Make sure it is valid at all the call sites 3949c9de560dSAlex Tomas */ 3950c2ea3fdeSTheodore Ts'o void ext4_discard_preallocations(struct inode *inode) 3951c9de560dSAlex Tomas { 3952c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(inode); 3953c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 3954c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3955c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3956c83617dbSAneesh Kumar K.V struct ext4_allocation_context *ac; 3957c9de560dSAlex Tomas ext4_group_t group = 0; 3958c9de560dSAlex Tomas struct list_head list; 3959c9de560dSAlex Tomas struct ext4_buddy e4b; 3960c9de560dSAlex Tomas int err; 3961c9de560dSAlex Tomas 3962c2ea3fdeSTheodore Ts'o if (!S_ISREG(inode->i_mode)) { 3963c9de560dSAlex Tomas /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ 3964c9de560dSAlex Tomas return; 3965c9de560dSAlex Tomas } 3966c9de560dSAlex Tomas 3967c9de560dSAlex Tomas mb_debug("discard preallocation for inode %lu\n", inode->i_ino); 3968c9de560dSAlex Tomas 3969c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3970c9de560dSAlex Tomas 3971c83617dbSAneesh Kumar K.V ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 3972c9de560dSAlex Tomas repeat: 3973c9de560dSAlex Tomas /* first, collect all pa's in the inode */ 3974c9de560dSAlex Tomas spin_lock(&ei->i_prealloc_lock); 3975c9de560dSAlex Tomas while (!list_empty(&ei->i_prealloc_list)) { 3976c9de560dSAlex Tomas pa = list_entry(ei->i_prealloc_list.next, 3977c9de560dSAlex Tomas struct ext4_prealloc_space, pa_inode_list); 3978c9de560dSAlex Tomas BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); 3979c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3980c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3981c9de560dSAlex Tomas /* this shouldn't happen often - nobody should 3982c9de560dSAlex Tomas * use preallocation while we're discarding it */ 3983c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3984c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3985c9de560dSAlex Tomas printk(KERN_ERR "uh-oh! used pa while discarding\n"); 3986c9de560dSAlex Tomas WARN_ON(1); 3987c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 3988c9de560dSAlex Tomas goto repeat; 3989c9de560dSAlex Tomas 3990c9de560dSAlex Tomas } 3991c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 3992c9de560dSAlex Tomas pa->pa_deleted = 1; 3993c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3994c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3995c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3996c9de560dSAlex Tomas continue; 3997c9de560dSAlex Tomas } 3998c9de560dSAlex Tomas 3999c9de560dSAlex Tomas /* someone is deleting pa right now */ 4000c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 4001c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 4002c9de560dSAlex Tomas 4003c9de560dSAlex Tomas /* we have to wait here because pa_deleted 4004c9de560dSAlex Tomas * doesn't mean pa is already unlinked from 4005c9de560dSAlex Tomas * the list. as we might be called from 4006c9de560dSAlex Tomas * ->clear_inode() the inode will get freed 4007c9de560dSAlex Tomas * and concurrent thread which is unlinking 4008c9de560dSAlex Tomas * pa from inode's list may access already 4009c9de560dSAlex Tomas * freed memory, bad-bad-bad */ 4010c9de560dSAlex Tomas 4011c9de560dSAlex Tomas /* XXX: if this happens too often, we can 4012c9de560dSAlex Tomas * add a flag to force wait only in case 4013c9de560dSAlex Tomas * of ->clear_inode(), but not in case of 4014c9de560dSAlex Tomas * regular truncate */ 4015c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 4016c9de560dSAlex Tomas goto repeat; 4017c9de560dSAlex Tomas } 4018c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 4019c9de560dSAlex Tomas 4020c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 4021c9de560dSAlex Tomas BUG_ON(pa->pa_linear != 0); 4022c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); 4023c9de560dSAlex Tomas 4024c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 4025ce89f46cSAneesh Kumar K.V if (err) { 4026ce89f46cSAneesh Kumar K.V ext4_error(sb, __func__, "Error in loading buddy " 4027a9df9a49STheodore Ts'o "information for %u", group); 4028ce89f46cSAneesh Kumar K.V continue; 4029ce89f46cSAneesh Kumar K.V } 4030c9de560dSAlex Tomas 4031574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 4032c9de560dSAlex Tomas if (bitmap_bh == NULL) { 4033ce89f46cSAneesh Kumar K.V ext4_error(sb, __func__, "Error in reading block " 4034a9df9a49STheodore Ts'o "bitmap for %u", group); 4035c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 4036ce89f46cSAneesh Kumar K.V continue; 4037c9de560dSAlex Tomas } 4038c9de560dSAlex Tomas 4039c9de560dSAlex Tomas ext4_lock_group(sb, group); 4040c9de560dSAlex Tomas list_del(&pa->pa_group_list); 4041c83617dbSAneesh Kumar K.V ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa, ac); 4042c9de560dSAlex Tomas ext4_unlock_group(sb, group); 4043c9de560dSAlex Tomas 4044c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 4045c9de560dSAlex Tomas put_bh(bitmap_bh); 4046c9de560dSAlex Tomas 4047c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 4048c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 4049c9de560dSAlex Tomas } 4050c83617dbSAneesh Kumar K.V if (ac) 4051c83617dbSAneesh Kumar K.V kmem_cache_free(ext4_ac_cachep, ac); 4052c9de560dSAlex Tomas } 4053c9de560dSAlex Tomas 4054c9de560dSAlex Tomas /* 4055c9de560dSAlex Tomas * finds all preallocated spaces and return blocks being freed to them 4056c9de560dSAlex Tomas * if preallocated space becomes full (no block is used from the space) 4057c9de560dSAlex Tomas * then the function frees space in buddy 4058c9de560dSAlex Tomas * XXX: at the moment, truncate (which is the only way to free blocks) 4059c9de560dSAlex Tomas * discards all preallocations 4060c9de560dSAlex Tomas */ 4061c9de560dSAlex Tomas static void ext4_mb_return_to_preallocation(struct inode *inode, 4062c9de560dSAlex Tomas struct ext4_buddy *e4b, 4063c9de560dSAlex Tomas sector_t block, int count) 4064c9de560dSAlex Tomas { 4065c9de560dSAlex Tomas BUG_ON(!list_empty(&EXT4_I(inode)->i_prealloc_list)); 4066c9de560dSAlex Tomas } 4067c9de560dSAlex Tomas #ifdef MB_DEBUG 4068c9de560dSAlex Tomas static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 4069c9de560dSAlex Tomas { 4070c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 4071c9de560dSAlex Tomas ext4_group_t i; 4072c9de560dSAlex Tomas 4073c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: Can't allocate:" 4074c9de560dSAlex Tomas " Allocation context details:\n"); 4075c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: status %d flags %d\n", 4076c9de560dSAlex Tomas ac->ac_status, ac->ac_flags); 4077c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, " 4078c9de560dSAlex Tomas "best %lu/%lu/%lu@%lu cr %d\n", 4079c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_group, 4080c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_start, 4081c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_len, 4082c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_logical, 4083c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_group, 4084c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_start, 4085c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_len, 4086c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_logical, 4087c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_group, 4088c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_start, 4089c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_len, 4090c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_logical, 4091c9de560dSAlex Tomas (int)ac->ac_criteria); 4092c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned, 4093c9de560dSAlex Tomas ac->ac_found); 4094c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: groups: \n"); 4095c9de560dSAlex Tomas for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) { 4096c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, i); 4097c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 4098c9de560dSAlex Tomas ext4_grpblk_t start; 4099c9de560dSAlex Tomas struct list_head *cur; 4100c9de560dSAlex Tomas ext4_lock_group(sb, i); 4101c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 4102c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, 4103c9de560dSAlex Tomas pa_group_list); 4104c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 4105c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 4106c9de560dSAlex Tomas NULL, &start); 4107c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 4108c9de560dSAlex Tomas printk(KERN_ERR "PA:%lu:%d:%u \n", i, 4109c9de560dSAlex Tomas start, pa->pa_len); 4110c9de560dSAlex Tomas } 411160bd63d1SSolofo Ramangalahy ext4_unlock_group(sb, i); 4112c9de560dSAlex Tomas 4113c9de560dSAlex Tomas if (grp->bb_free == 0) 4114c9de560dSAlex Tomas continue; 4115c9de560dSAlex Tomas printk(KERN_ERR "%lu: %d/%d \n", 4116c9de560dSAlex Tomas i, grp->bb_free, grp->bb_fragments); 4117c9de560dSAlex Tomas } 4118c9de560dSAlex Tomas printk(KERN_ERR "\n"); 4119c9de560dSAlex Tomas } 4120c9de560dSAlex Tomas #else 4121c9de560dSAlex Tomas static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 4122c9de560dSAlex Tomas { 4123c9de560dSAlex Tomas return; 4124c9de560dSAlex Tomas } 4125c9de560dSAlex Tomas #endif 4126c9de560dSAlex Tomas 4127c9de560dSAlex Tomas /* 4128c9de560dSAlex Tomas * We use locality group preallocation for small size file. The size of the 4129c9de560dSAlex Tomas * file is determined by the current size or the resulting size after 4130c9de560dSAlex Tomas * allocation which ever is larger 4131c9de560dSAlex Tomas * 4132c9de560dSAlex Tomas * One can tune this size via /proc/fs/ext4/<partition>/stream_req 4133c9de560dSAlex Tomas */ 4134c9de560dSAlex Tomas static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 4135c9de560dSAlex Tomas { 4136c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 4137c9de560dSAlex Tomas int bsbits = ac->ac_sb->s_blocksize_bits; 4138c9de560dSAlex Tomas loff_t size, isize; 4139c9de560dSAlex Tomas 4140c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 4141c9de560dSAlex Tomas return; 4142c9de560dSAlex Tomas 4143c9de560dSAlex Tomas size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; 4144c9de560dSAlex Tomas isize = i_size_read(ac->ac_inode) >> bsbits; 4145c9de560dSAlex Tomas size = max(size, isize); 4146c9de560dSAlex Tomas 4147c9de560dSAlex Tomas /* don't use group allocation for large files */ 4148c9de560dSAlex Tomas if (size >= sbi->s_mb_stream_request) 4149c9de560dSAlex Tomas return; 4150c9de560dSAlex Tomas 4151c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 4152c9de560dSAlex Tomas return; 4153c9de560dSAlex Tomas 4154c9de560dSAlex Tomas BUG_ON(ac->ac_lg != NULL); 4155c9de560dSAlex Tomas /* 4156c9de560dSAlex Tomas * locality group prealloc space are per cpu. The reason for having 4157c9de560dSAlex Tomas * per cpu locality group is to reduce the contention between block 4158c9de560dSAlex Tomas * request from multiple CPUs. 4159c9de560dSAlex Tomas */ 4160730c213cSEric Sandeen ac->ac_lg = per_cpu_ptr(sbi->s_locality_groups, raw_smp_processor_id()); 4161c9de560dSAlex Tomas 4162c9de560dSAlex Tomas /* we're going to use group allocation */ 4163c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 4164c9de560dSAlex Tomas 4165c9de560dSAlex Tomas /* serialize all allocations in the group */ 4166c9de560dSAlex Tomas mutex_lock(&ac->ac_lg->lg_mutex); 4167c9de560dSAlex Tomas } 4168c9de560dSAlex Tomas 41694ddfef7bSEric Sandeen static noinline_for_stack int 41704ddfef7bSEric Sandeen ext4_mb_initialize_context(struct ext4_allocation_context *ac, 4171c9de560dSAlex Tomas struct ext4_allocation_request *ar) 4172c9de560dSAlex Tomas { 4173c9de560dSAlex Tomas struct super_block *sb = ar->inode->i_sb; 4174c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 4175c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 4176c9de560dSAlex Tomas ext4_group_t group; 4177498e5f24STheodore Ts'o unsigned int len; 4178498e5f24STheodore Ts'o ext4_fsblk_t goal; 4179c9de560dSAlex Tomas ext4_grpblk_t block; 4180c9de560dSAlex Tomas 4181c9de560dSAlex Tomas /* we can't allocate > group size */ 4182c9de560dSAlex Tomas len = ar->len; 4183c9de560dSAlex Tomas 4184c9de560dSAlex Tomas /* just a dirty hack to filter too big requests */ 4185c9de560dSAlex Tomas if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10) 4186c9de560dSAlex Tomas len = EXT4_BLOCKS_PER_GROUP(sb) - 10; 4187c9de560dSAlex Tomas 4188c9de560dSAlex Tomas /* start searching from the goal */ 4189c9de560dSAlex Tomas goal = ar->goal; 4190c9de560dSAlex Tomas if (goal < le32_to_cpu(es->s_first_data_block) || 4191c9de560dSAlex Tomas goal >= ext4_blocks_count(es)) 4192c9de560dSAlex Tomas goal = le32_to_cpu(es->s_first_data_block); 4193c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, goal, &group, &block); 4194c9de560dSAlex Tomas 4195c9de560dSAlex Tomas /* set up allocation goals */ 4196c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ar->logical; 4197c9de560dSAlex Tomas ac->ac_b_ex.fe_group = 0; 4198c9de560dSAlex Tomas ac->ac_b_ex.fe_start = 0; 4199c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 0; 4200c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 4201c9de560dSAlex Tomas ac->ac_groups_scanned = 0; 4202c9de560dSAlex Tomas ac->ac_ex_scanned = 0; 4203c9de560dSAlex Tomas ac->ac_found = 0; 4204c9de560dSAlex Tomas ac->ac_sb = sb; 4205c9de560dSAlex Tomas ac->ac_inode = ar->inode; 4206c9de560dSAlex Tomas ac->ac_o_ex.fe_logical = ar->logical; 4207c9de560dSAlex Tomas ac->ac_o_ex.fe_group = group; 4208c9de560dSAlex Tomas ac->ac_o_ex.fe_start = block; 4209c9de560dSAlex Tomas ac->ac_o_ex.fe_len = len; 4210c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = ar->logical; 4211c9de560dSAlex Tomas ac->ac_g_ex.fe_group = group; 4212c9de560dSAlex Tomas ac->ac_g_ex.fe_start = block; 4213c9de560dSAlex Tomas ac->ac_g_ex.fe_len = len; 4214c9de560dSAlex Tomas ac->ac_f_ex.fe_len = 0; 4215c9de560dSAlex Tomas ac->ac_flags = ar->flags; 4216c9de560dSAlex Tomas ac->ac_2order = 0; 4217c9de560dSAlex Tomas ac->ac_criteria = 0; 4218c9de560dSAlex Tomas ac->ac_pa = NULL; 4219c9de560dSAlex Tomas ac->ac_bitmap_page = NULL; 4220c9de560dSAlex Tomas ac->ac_buddy_page = NULL; 4221c9de560dSAlex Tomas ac->ac_lg = NULL; 4222c9de560dSAlex Tomas 4223c9de560dSAlex Tomas /* we have to define context: we'll we work with a file or 4224c9de560dSAlex Tomas * locality group. this is a policy, actually */ 4225c9de560dSAlex Tomas ext4_mb_group_or_file(ac); 4226c9de560dSAlex Tomas 4227c9de560dSAlex Tomas mb_debug("init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " 4228c9de560dSAlex Tomas "left: %u/%u, right %u/%u to %swritable\n", 4229c9de560dSAlex Tomas (unsigned) ar->len, (unsigned) ar->logical, 4230c9de560dSAlex Tomas (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 4231c9de560dSAlex Tomas (unsigned) ar->lleft, (unsigned) ar->pleft, 4232c9de560dSAlex Tomas (unsigned) ar->lright, (unsigned) ar->pright, 4233c9de560dSAlex Tomas atomic_read(&ar->inode->i_writecount) ? "" : "non-"); 4234c9de560dSAlex Tomas return 0; 4235c9de560dSAlex Tomas 4236c9de560dSAlex Tomas } 4237c9de560dSAlex Tomas 42386be2ded1SAneesh Kumar K.V static noinline_for_stack void 42396be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(struct super_block *sb, 42406be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg, 42416be2ded1SAneesh Kumar K.V int order, int total_entries) 42426be2ded1SAneesh Kumar K.V { 42436be2ded1SAneesh Kumar K.V ext4_group_t group = 0; 42446be2ded1SAneesh Kumar K.V struct ext4_buddy e4b; 42456be2ded1SAneesh Kumar K.V struct list_head discard_list; 42466be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa, *tmp; 42476be2ded1SAneesh Kumar K.V struct ext4_allocation_context *ac; 42486be2ded1SAneesh Kumar K.V 42496be2ded1SAneesh Kumar K.V mb_debug("discard locality group preallocation\n"); 42506be2ded1SAneesh Kumar K.V 42516be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&discard_list); 42526be2ded1SAneesh Kumar K.V ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 42536be2ded1SAneesh Kumar K.V 42546be2ded1SAneesh Kumar K.V spin_lock(&lg->lg_prealloc_lock); 42556be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 42566be2ded1SAneesh Kumar K.V pa_inode_list) { 42576be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 42586be2ded1SAneesh Kumar K.V if (atomic_read(&pa->pa_count)) { 42596be2ded1SAneesh Kumar K.V /* 42606be2ded1SAneesh Kumar K.V * This is the pa that we just used 42616be2ded1SAneesh Kumar K.V * for block allocation. So don't 42626be2ded1SAneesh Kumar K.V * free that 42636be2ded1SAneesh Kumar K.V */ 42646be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 42656be2ded1SAneesh Kumar K.V continue; 42666be2ded1SAneesh Kumar K.V } 42676be2ded1SAneesh Kumar K.V if (pa->pa_deleted) { 42686be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 42696be2ded1SAneesh Kumar K.V continue; 42706be2ded1SAneesh Kumar K.V } 42716be2ded1SAneesh Kumar K.V /* only lg prealloc space */ 42726be2ded1SAneesh Kumar K.V BUG_ON(!pa->pa_linear); 42736be2ded1SAneesh Kumar K.V 42746be2ded1SAneesh Kumar K.V /* seems this one can be freed ... */ 42756be2ded1SAneesh Kumar K.V pa->pa_deleted = 1; 42766be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 42776be2ded1SAneesh Kumar K.V 42786be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 42796be2ded1SAneesh Kumar K.V list_add(&pa->u.pa_tmp_list, &discard_list); 42806be2ded1SAneesh Kumar K.V 42816be2ded1SAneesh Kumar K.V total_entries--; 42826be2ded1SAneesh Kumar K.V if (total_entries <= 5) { 42836be2ded1SAneesh Kumar K.V /* 42846be2ded1SAneesh Kumar K.V * we want to keep only 5 entries 42856be2ded1SAneesh Kumar K.V * allowing it to grow to 8. This 42866be2ded1SAneesh Kumar K.V * mak sure we don't call discard 42876be2ded1SAneesh Kumar K.V * soon for this list. 42886be2ded1SAneesh Kumar K.V */ 42896be2ded1SAneesh Kumar K.V break; 42906be2ded1SAneesh Kumar K.V } 42916be2ded1SAneesh Kumar K.V } 42926be2ded1SAneesh Kumar K.V spin_unlock(&lg->lg_prealloc_lock); 42936be2ded1SAneesh Kumar K.V 42946be2ded1SAneesh Kumar K.V list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 42956be2ded1SAneesh Kumar K.V 42966be2ded1SAneesh Kumar K.V ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); 42976be2ded1SAneesh Kumar K.V if (ext4_mb_load_buddy(sb, group, &e4b)) { 42986be2ded1SAneesh Kumar K.V ext4_error(sb, __func__, "Error in loading buddy " 4299a9df9a49STheodore Ts'o "information for %u", group); 43006be2ded1SAneesh Kumar K.V continue; 43016be2ded1SAneesh Kumar K.V } 43026be2ded1SAneesh Kumar K.V ext4_lock_group(sb, group); 43036be2ded1SAneesh Kumar K.V list_del(&pa->pa_group_list); 43046be2ded1SAneesh Kumar K.V ext4_mb_release_group_pa(&e4b, pa, ac); 43056be2ded1SAneesh Kumar K.V ext4_unlock_group(sb, group); 43066be2ded1SAneesh Kumar K.V 43076be2ded1SAneesh Kumar K.V ext4_mb_release_desc(&e4b); 43086be2ded1SAneesh Kumar K.V list_del(&pa->u.pa_tmp_list); 43096be2ded1SAneesh Kumar K.V call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 43106be2ded1SAneesh Kumar K.V } 43116be2ded1SAneesh Kumar K.V if (ac) 43126be2ded1SAneesh Kumar K.V kmem_cache_free(ext4_ac_cachep, ac); 43136be2ded1SAneesh Kumar K.V } 43146be2ded1SAneesh Kumar K.V 43156be2ded1SAneesh Kumar K.V /* 43166be2ded1SAneesh Kumar K.V * We have incremented pa_count. So it cannot be freed at this 43176be2ded1SAneesh Kumar K.V * point. Also we hold lg_mutex. So no parallel allocation is 43186be2ded1SAneesh Kumar K.V * possible from this lg. That means pa_free cannot be updated. 43196be2ded1SAneesh Kumar K.V * 43206be2ded1SAneesh Kumar K.V * A parallel ext4_mb_discard_group_preallocations is possible. 43216be2ded1SAneesh Kumar K.V * which can cause the lg_prealloc_list to be updated. 43226be2ded1SAneesh Kumar K.V */ 43236be2ded1SAneesh Kumar K.V 43246be2ded1SAneesh Kumar K.V static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 43256be2ded1SAneesh Kumar K.V { 43266be2ded1SAneesh Kumar K.V int order, added = 0, lg_prealloc_count = 1; 43276be2ded1SAneesh Kumar K.V struct super_block *sb = ac->ac_sb; 43286be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg = ac->ac_lg; 43296be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 43306be2ded1SAneesh Kumar K.V 43316be2ded1SAneesh Kumar K.V order = fls(pa->pa_free) - 1; 43326be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 43336be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 43346be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 43356be2ded1SAneesh Kumar K.V /* Add the prealloc space to lg */ 43366be2ded1SAneesh Kumar K.V rcu_read_lock(); 43376be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 43386be2ded1SAneesh Kumar K.V pa_inode_list) { 43396be2ded1SAneesh Kumar K.V spin_lock(&tmp_pa->pa_lock); 43406be2ded1SAneesh Kumar K.V if (tmp_pa->pa_deleted) { 43416be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 43426be2ded1SAneesh Kumar K.V continue; 43436be2ded1SAneesh Kumar K.V } 43446be2ded1SAneesh Kumar K.V if (!added && pa->pa_free < tmp_pa->pa_free) { 43456be2ded1SAneesh Kumar K.V /* Add to the tail of the previous entry */ 43466be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 43476be2ded1SAneesh Kumar K.V &tmp_pa->pa_inode_list); 43486be2ded1SAneesh Kumar K.V added = 1; 43496be2ded1SAneesh Kumar K.V /* 43506be2ded1SAneesh Kumar K.V * we want to count the total 43516be2ded1SAneesh Kumar K.V * number of entries in the list 43526be2ded1SAneesh Kumar K.V */ 43536be2ded1SAneesh Kumar K.V } 43546be2ded1SAneesh Kumar K.V spin_unlock(&tmp_pa->pa_lock); 43556be2ded1SAneesh Kumar K.V lg_prealloc_count++; 43566be2ded1SAneesh Kumar K.V } 43576be2ded1SAneesh Kumar K.V if (!added) 43586be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 43596be2ded1SAneesh Kumar K.V &lg->lg_prealloc_list[order]); 43606be2ded1SAneesh Kumar K.V rcu_read_unlock(); 43616be2ded1SAneesh Kumar K.V 43626be2ded1SAneesh Kumar K.V /* Now trim the list to be not more than 8 elements */ 43636be2ded1SAneesh Kumar K.V if (lg_prealloc_count > 8) { 43646be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(sb, lg, 43656be2ded1SAneesh Kumar K.V order, lg_prealloc_count); 43666be2ded1SAneesh Kumar K.V return; 43676be2ded1SAneesh Kumar K.V } 43686be2ded1SAneesh Kumar K.V return ; 43696be2ded1SAneesh Kumar K.V } 43706be2ded1SAneesh Kumar K.V 4371c9de560dSAlex Tomas /* 4372c9de560dSAlex Tomas * release all resource we used in allocation 4373c9de560dSAlex Tomas */ 4374c9de560dSAlex Tomas static int ext4_mb_release_context(struct ext4_allocation_context *ac) 4375c9de560dSAlex Tomas { 43766be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa = ac->ac_pa; 43776be2ded1SAneesh Kumar K.V if (pa) { 43786be2ded1SAneesh Kumar K.V if (pa->pa_linear) { 4379c9de560dSAlex Tomas /* see comment in ext4_mb_use_group_pa() */ 43806be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 43816be2ded1SAneesh Kumar K.V pa->pa_pstart += ac->ac_b_ex.fe_len; 43826be2ded1SAneesh Kumar K.V pa->pa_lstart += ac->ac_b_ex.fe_len; 43836be2ded1SAneesh Kumar K.V pa->pa_free -= ac->ac_b_ex.fe_len; 43846be2ded1SAneesh Kumar K.V pa->pa_len -= ac->ac_b_ex.fe_len; 43856be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 43866be2ded1SAneesh Kumar K.V /* 43876be2ded1SAneesh Kumar K.V * We want to add the pa to the right bucket. 43886be2ded1SAneesh Kumar K.V * Remove it from the list and while adding 43896be2ded1SAneesh Kumar K.V * make sure the list to which we are adding 43906be2ded1SAneesh Kumar K.V * doesn't grow big. 43916be2ded1SAneesh Kumar K.V */ 43926be2ded1SAneesh Kumar K.V if (likely(pa->pa_free)) { 43936be2ded1SAneesh Kumar K.V spin_lock(pa->pa_obj_lock); 43946be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 43956be2ded1SAneesh Kumar K.V spin_unlock(pa->pa_obj_lock); 43966be2ded1SAneesh Kumar K.V ext4_mb_add_n_trim(ac); 4397c9de560dSAlex Tomas } 43986be2ded1SAneesh Kumar K.V } 43996be2ded1SAneesh Kumar K.V ext4_mb_put_pa(ac, ac->ac_sb, pa); 4400c9de560dSAlex Tomas } 4401c9de560dSAlex Tomas if (ac->ac_bitmap_page) 4402c9de560dSAlex Tomas page_cache_release(ac->ac_bitmap_page); 4403c9de560dSAlex Tomas if (ac->ac_buddy_page) 4404c9de560dSAlex Tomas page_cache_release(ac->ac_buddy_page); 4405c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 4406c9de560dSAlex Tomas mutex_unlock(&ac->ac_lg->lg_mutex); 4407c9de560dSAlex Tomas ext4_mb_collect_stats(ac); 4408c9de560dSAlex Tomas return 0; 4409c9de560dSAlex Tomas } 4410c9de560dSAlex Tomas 4411c9de560dSAlex Tomas static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 4412c9de560dSAlex Tomas { 4413c9de560dSAlex Tomas ext4_group_t i; 4414c9de560dSAlex Tomas int ret; 4415c9de560dSAlex Tomas int freed = 0; 4416c9de560dSAlex Tomas 4417c9de560dSAlex Tomas for (i = 0; i < EXT4_SB(sb)->s_groups_count && needed > 0; i++) { 4418c9de560dSAlex Tomas ret = ext4_mb_discard_group_preallocations(sb, i, needed); 4419c9de560dSAlex Tomas freed += ret; 4420c9de560dSAlex Tomas needed -= ret; 4421c9de560dSAlex Tomas } 4422c9de560dSAlex Tomas 4423c9de560dSAlex Tomas return freed; 4424c9de560dSAlex Tomas } 4425c9de560dSAlex Tomas 4426c9de560dSAlex Tomas /* 4427c9de560dSAlex Tomas * Main entry point into mballoc to allocate blocks 4428c9de560dSAlex Tomas * it tries to use preallocation first, then falls back 4429c9de560dSAlex Tomas * to usual allocation 4430c9de560dSAlex Tomas */ 4431c9de560dSAlex Tomas ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 4432c9de560dSAlex Tomas struct ext4_allocation_request *ar, int *errp) 4433c9de560dSAlex Tomas { 44346bc6e63fSAneesh Kumar K.V int freed; 4435256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 4436c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4437c9de560dSAlex Tomas struct super_block *sb; 4438c9de560dSAlex Tomas ext4_fsblk_t block = 0; 4439498e5f24STheodore Ts'o unsigned int inquota; 4440498e5f24STheodore Ts'o unsigned int reserv_blks = 0; 4441c9de560dSAlex Tomas 4442c9de560dSAlex Tomas sb = ar->inode->i_sb; 4443c9de560dSAlex Tomas sbi = EXT4_SB(sb); 4444c9de560dSAlex Tomas 4445d2a17637SMingming Cao if (!EXT4_I(ar->inode)->i_delalloc_reserved_flag) { 4446d2a17637SMingming Cao /* 4447d2a17637SMingming Cao * With delalloc we already reserved the blocks 4448d2a17637SMingming Cao */ 4449030ba6bcSAneesh Kumar K.V while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) { 4450030ba6bcSAneesh Kumar K.V /* let others to free the space */ 4451030ba6bcSAneesh Kumar K.V yield(); 4452030ba6bcSAneesh Kumar K.V ar->len = ar->len >> 1; 4453030ba6bcSAneesh Kumar K.V } 4454030ba6bcSAneesh Kumar K.V if (!ar->len) { 445507031431SMingming Cao *errp = -ENOSPC; 445607031431SMingming Cao return 0; 445707031431SMingming Cao } 44586bc6e63fSAneesh Kumar K.V reserv_blks = ar->len; 4459a30d542aSAneesh Kumar K.V } 4460c9de560dSAlex Tomas while (ar->len && DQUOT_ALLOC_BLOCK(ar->inode, ar->len)) { 4461c9de560dSAlex Tomas ar->flags |= EXT4_MB_HINT_NOPREALLOC; 4462c9de560dSAlex Tomas ar->len--; 4463c9de560dSAlex Tomas } 4464c9de560dSAlex Tomas if (ar->len == 0) { 4465c9de560dSAlex Tomas *errp = -EDQUOT; 4466c9de560dSAlex Tomas return 0; 4467c9de560dSAlex Tomas } 4468c9de560dSAlex Tomas inquota = ar->len; 4469c9de560dSAlex Tomas 4470d2a17637SMingming Cao if (EXT4_I(ar->inode)->i_delalloc_reserved_flag) 4471d2a17637SMingming Cao ar->flags |= EXT4_MB_DELALLOC_RESERVED; 4472d2a17637SMingming Cao 4473256bdb49SEric Sandeen ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 4474256bdb49SEric Sandeen if (!ac) { 4475363d4251SShen Feng ar->len = 0; 4476256bdb49SEric Sandeen *errp = -ENOMEM; 4477363d4251SShen Feng goto out1; 4478256bdb49SEric Sandeen } 4479256bdb49SEric Sandeen 4480256bdb49SEric Sandeen *errp = ext4_mb_initialize_context(ac, ar); 4481c9de560dSAlex Tomas if (*errp) { 4482c9de560dSAlex Tomas ar->len = 0; 4483363d4251SShen Feng goto out2; 4484c9de560dSAlex Tomas } 4485c9de560dSAlex Tomas 4486256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 4487256bdb49SEric Sandeen if (!ext4_mb_use_preallocated(ac)) { 4488256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_ALLOC; 4489256bdb49SEric Sandeen ext4_mb_normalize_request(ac, ar); 4490c9de560dSAlex Tomas repeat: 4491c9de560dSAlex Tomas /* allocate space in core */ 4492256bdb49SEric Sandeen ext4_mb_regular_allocator(ac); 4493c9de560dSAlex Tomas 4494c9de560dSAlex Tomas /* as we've just preallocated more space than 4495c9de560dSAlex Tomas * user requested orinally, we store allocated 4496c9de560dSAlex Tomas * space in a special descriptor */ 4497256bdb49SEric Sandeen if (ac->ac_status == AC_STATUS_FOUND && 4498256bdb49SEric Sandeen ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 4499256bdb49SEric Sandeen ext4_mb_new_preallocation(ac); 4500c9de560dSAlex Tomas } 4501c9de560dSAlex Tomas 4502256bdb49SEric Sandeen if (likely(ac->ac_status == AC_STATUS_FOUND)) { 45036bc6e63fSAneesh Kumar K.V *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks); 4504519deca0SAneesh Kumar K.V if (*errp == -EAGAIN) { 4505519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_group = 0; 4506519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_start = 0; 4507519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len = 0; 4508519deca0SAneesh Kumar K.V ac->ac_status = AC_STATUS_CONTINUE; 4509519deca0SAneesh Kumar K.V goto repeat; 4510519deca0SAneesh Kumar K.V } else if (*errp) { 4511519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len = 0; 4512519deca0SAneesh Kumar K.V ar->len = 0; 4513519deca0SAneesh Kumar K.V ext4_mb_show_ac(ac); 4514519deca0SAneesh Kumar K.V } else { 4515256bdb49SEric Sandeen block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 4516256bdb49SEric Sandeen ar->len = ac->ac_b_ex.fe_len; 4517519deca0SAneesh Kumar K.V } 4518c9de560dSAlex Tomas } else { 4519256bdb49SEric Sandeen freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 4520c9de560dSAlex Tomas if (freed) 4521c9de560dSAlex Tomas goto repeat; 4522c9de560dSAlex Tomas *errp = -ENOSPC; 4523256bdb49SEric Sandeen ac->ac_b_ex.fe_len = 0; 4524c9de560dSAlex Tomas ar->len = 0; 4525256bdb49SEric Sandeen ext4_mb_show_ac(ac); 4526c9de560dSAlex Tomas } 4527c9de560dSAlex Tomas 4528256bdb49SEric Sandeen ext4_mb_release_context(ac); 4529c9de560dSAlex Tomas 4530363d4251SShen Feng out2: 4531363d4251SShen Feng kmem_cache_free(ext4_ac_cachep, ac); 4532363d4251SShen Feng out1: 4533c9de560dSAlex Tomas if (ar->len < inquota) 4534c9de560dSAlex Tomas DQUOT_FREE_BLOCK(ar->inode, inquota - ar->len); 4535c9de560dSAlex Tomas 4536c9de560dSAlex Tomas return block; 4537c9de560dSAlex Tomas } 4538c9de560dSAlex Tomas 4539c894058dSAneesh Kumar K.V /* 4540c894058dSAneesh Kumar K.V * We can merge two free data extents only if the physical blocks 4541c894058dSAneesh Kumar K.V * are contiguous, AND the extents were freed by the same transaction, 4542c894058dSAneesh Kumar K.V * AND the blocks are associated with the same group. 4543c894058dSAneesh Kumar K.V */ 4544c894058dSAneesh Kumar K.V static int can_merge(struct ext4_free_data *entry1, 4545c894058dSAneesh Kumar K.V struct ext4_free_data *entry2) 4546c894058dSAneesh Kumar K.V { 4547c894058dSAneesh Kumar K.V if ((entry1->t_tid == entry2->t_tid) && 4548c894058dSAneesh Kumar K.V (entry1->group == entry2->group) && 4549c894058dSAneesh Kumar K.V ((entry1->start_blk + entry1->count) == entry2->start_blk)) 4550c894058dSAneesh Kumar K.V return 1; 4551c894058dSAneesh Kumar K.V return 0; 4552c894058dSAneesh Kumar K.V } 4553c894058dSAneesh Kumar K.V 45544ddfef7bSEric Sandeen static noinline_for_stack int 45554ddfef7bSEric Sandeen ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 4556c9de560dSAlex Tomas ext4_group_t group, ext4_grpblk_t block, int count) 4557c9de560dSAlex Tomas { 4558c9de560dSAlex Tomas struct ext4_group_info *db = e4b->bd_info; 4559c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 4560c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 4561c894058dSAneesh Kumar K.V struct ext4_free_data *entry, *new_entry; 4562c894058dSAneesh Kumar K.V struct rb_node **n = &db->bb_free_root.rb_node, *node; 4563c894058dSAneesh Kumar K.V struct rb_node *parent = NULL, *new_node; 4564c894058dSAneesh Kumar K.V 45650390131bSFrank Mayhar BUG_ON(!ext4_handle_valid(handle)); 4566c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 4567c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 4568c9de560dSAlex Tomas 4569c894058dSAneesh Kumar K.V new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); 4570c894058dSAneesh Kumar K.V new_entry->start_blk = block; 4571c894058dSAneesh Kumar K.V new_entry->group = group; 4572c894058dSAneesh Kumar K.V new_entry->count = count; 4573c894058dSAneesh Kumar K.V new_entry->t_tid = handle->h_transaction->t_tid; 4574c894058dSAneesh Kumar K.V new_node = &new_entry->node; 4575c9de560dSAlex Tomas 4576c9de560dSAlex Tomas ext4_lock_group(sb, group); 4577c894058dSAneesh Kumar K.V if (!*n) { 4578c894058dSAneesh Kumar K.V /* first free block exent. We need to 4579c894058dSAneesh Kumar K.V protect buddy cache from being freed, 4580c9de560dSAlex Tomas * otherwise we'll refresh it from 4581c9de560dSAlex Tomas * on-disk bitmap and lose not-yet-available 4582c9de560dSAlex Tomas * blocks */ 4583c9de560dSAlex Tomas page_cache_get(e4b->bd_buddy_page); 4584c9de560dSAlex Tomas page_cache_get(e4b->bd_bitmap_page); 4585c894058dSAneesh Kumar K.V } 4586c894058dSAneesh Kumar K.V while (*n) { 4587c894058dSAneesh Kumar K.V parent = *n; 4588c894058dSAneesh Kumar K.V entry = rb_entry(parent, struct ext4_free_data, node); 4589c894058dSAneesh Kumar K.V if (block < entry->start_blk) 4590c894058dSAneesh Kumar K.V n = &(*n)->rb_left; 4591c894058dSAneesh Kumar K.V else if (block >= (entry->start_blk + entry->count)) 4592c894058dSAneesh Kumar K.V n = &(*n)->rb_right; 4593c894058dSAneesh Kumar K.V else { 4594ae2d9fb1SAneesh Kumar K.V ext4_unlock_group(sb, group); 4595c894058dSAneesh Kumar K.V ext4_error(sb, __func__, 4596fde4d95aSTheodore Ts'o "Double free of blocks %d (%d %d)", 4597c894058dSAneesh Kumar K.V block, entry->start_blk, entry->count); 4598c894058dSAneesh Kumar K.V return 0; 4599c9de560dSAlex Tomas } 4600c9de560dSAlex Tomas } 4601c9de560dSAlex Tomas 4602c894058dSAneesh Kumar K.V rb_link_node(new_node, parent, n); 4603c894058dSAneesh Kumar K.V rb_insert_color(new_node, &db->bb_free_root); 4604c894058dSAneesh Kumar K.V 4605c894058dSAneesh Kumar K.V /* Now try to see the extent can be merged to left and right */ 4606c894058dSAneesh Kumar K.V node = rb_prev(new_node); 4607c894058dSAneesh Kumar K.V if (node) { 4608c894058dSAneesh Kumar K.V entry = rb_entry(node, struct ext4_free_data, node); 4609c894058dSAneesh Kumar K.V if (can_merge(entry, new_entry)) { 4610c894058dSAneesh Kumar K.V new_entry->start_blk = entry->start_blk; 4611c894058dSAneesh Kumar K.V new_entry->count += entry->count; 4612c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 4613c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 4614c894058dSAneesh Kumar K.V list_del(&entry->list); 4615c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4616c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 4617c9de560dSAlex Tomas } 4618c9de560dSAlex Tomas } 4619c894058dSAneesh Kumar K.V 4620c894058dSAneesh Kumar K.V node = rb_next(new_node); 4621c894058dSAneesh Kumar K.V if (node) { 4622c894058dSAneesh Kumar K.V entry = rb_entry(node, struct ext4_free_data, node); 4623c894058dSAneesh Kumar K.V if (can_merge(new_entry, entry)) { 4624c894058dSAneesh Kumar K.V new_entry->count += entry->count; 4625c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 4626c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 4627c894058dSAneesh Kumar K.V list_del(&entry->list); 4628c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4629c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 4630c894058dSAneesh Kumar K.V } 4631c894058dSAneesh Kumar K.V } 46323e624fc7STheodore Ts'o /* Add the extent to transaction's private list */ 4633c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 46343e624fc7STheodore Ts'o list_add(&new_entry->list, &handle->h_transaction->t_private_list); 4635c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4636c9de560dSAlex Tomas ext4_unlock_group(sb, group); 4637c9de560dSAlex Tomas return 0; 4638c9de560dSAlex Tomas } 4639c9de560dSAlex Tomas 4640c9de560dSAlex Tomas /* 4641c9de560dSAlex Tomas * Main entry point into mballoc to free blocks 4642c9de560dSAlex Tomas */ 4643c9de560dSAlex Tomas void ext4_mb_free_blocks(handle_t *handle, struct inode *inode, 4644c9de560dSAlex Tomas unsigned long block, unsigned long count, 4645c9de560dSAlex Tomas int metadata, unsigned long *freed) 4646c9de560dSAlex Tomas { 464726346ff6SAneesh Kumar K.V struct buffer_head *bitmap_bh = NULL; 4648c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 4649256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 4650c9de560dSAlex Tomas struct ext4_group_desc *gdp; 4651c9de560dSAlex Tomas struct ext4_super_block *es; 4652498e5f24STheodore Ts'o unsigned int overflow; 4653c9de560dSAlex Tomas ext4_grpblk_t bit; 4654c9de560dSAlex Tomas struct buffer_head *gd_bh; 4655c9de560dSAlex Tomas ext4_group_t block_group; 4656c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4657c9de560dSAlex Tomas struct ext4_buddy e4b; 4658c9de560dSAlex Tomas int err = 0; 4659c9de560dSAlex Tomas int ret; 4660c9de560dSAlex Tomas 4661c9de560dSAlex Tomas *freed = 0; 4662c9de560dSAlex Tomas 4663c9de560dSAlex Tomas sbi = EXT4_SB(sb); 4664c9de560dSAlex Tomas es = EXT4_SB(sb)->s_es; 4665c9de560dSAlex Tomas if (block < le32_to_cpu(es->s_first_data_block) || 4666c9de560dSAlex Tomas block + count < block || 4667c9de560dSAlex Tomas block + count > ext4_blocks_count(es)) { 466846e665e9SHarvey Harrison ext4_error(sb, __func__, 4669c9de560dSAlex Tomas "Freeing blocks not in datazone - " 4670c9de560dSAlex Tomas "block = %lu, count = %lu", block, count); 4671c9de560dSAlex Tomas goto error_return; 4672c9de560dSAlex Tomas } 4673c9de560dSAlex Tomas 4674c9de560dSAlex Tomas ext4_debug("freeing block %lu\n", block); 4675c9de560dSAlex Tomas 4676256bdb49SEric Sandeen ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 4677256bdb49SEric Sandeen if (ac) { 4678256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_FREE; 4679256bdb49SEric Sandeen ac->ac_inode = inode; 4680256bdb49SEric Sandeen ac->ac_sb = sb; 4681256bdb49SEric Sandeen } 4682c9de560dSAlex Tomas 4683c9de560dSAlex Tomas do_more: 4684c9de560dSAlex Tomas overflow = 0; 4685c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 4686c9de560dSAlex Tomas 4687c9de560dSAlex Tomas /* 4688c9de560dSAlex Tomas * Check to see if we are freeing blocks across a group 4689c9de560dSAlex Tomas * boundary. 4690c9de560dSAlex Tomas */ 4691c9de560dSAlex Tomas if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { 4692c9de560dSAlex Tomas overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb); 4693c9de560dSAlex Tomas count -= overflow; 4694c9de560dSAlex Tomas } 4695574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, block_group); 4696ce89f46cSAneesh Kumar K.V if (!bitmap_bh) { 4697ce89f46cSAneesh Kumar K.V err = -EIO; 4698c9de560dSAlex Tomas goto error_return; 4699ce89f46cSAneesh Kumar K.V } 4700c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 4701ce89f46cSAneesh Kumar K.V if (!gdp) { 4702ce89f46cSAneesh Kumar K.V err = -EIO; 4703c9de560dSAlex Tomas goto error_return; 4704ce89f46cSAneesh Kumar K.V } 4705c9de560dSAlex Tomas 4706c9de560dSAlex Tomas if (in_range(ext4_block_bitmap(sb, gdp), block, count) || 4707c9de560dSAlex Tomas in_range(ext4_inode_bitmap(sb, gdp), block, count) || 4708c9de560dSAlex Tomas in_range(block, ext4_inode_table(sb, gdp), 4709c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group) || 4710c9de560dSAlex Tomas in_range(block + count - 1, ext4_inode_table(sb, gdp), 4711c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group)) { 4712c9de560dSAlex Tomas 471346e665e9SHarvey Harrison ext4_error(sb, __func__, 4714c9de560dSAlex Tomas "Freeing blocks in system zone - " 4715c9de560dSAlex Tomas "Block = %lu, count = %lu", block, count); 4716519deca0SAneesh Kumar K.V /* err = 0. ext4_std_error should be a no op */ 4717519deca0SAneesh Kumar K.V goto error_return; 4718c9de560dSAlex Tomas } 4719c9de560dSAlex Tomas 4720c9de560dSAlex Tomas BUFFER_TRACE(bitmap_bh, "getting write access"); 4721c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 4722c9de560dSAlex Tomas if (err) 4723c9de560dSAlex Tomas goto error_return; 4724c9de560dSAlex Tomas 4725c9de560dSAlex Tomas /* 4726c9de560dSAlex Tomas * We are about to modify some metadata. Call the journal APIs 4727c9de560dSAlex Tomas * to unshare ->b_data if a currently-committing transaction is 4728c9de560dSAlex Tomas * using it 4729c9de560dSAlex Tomas */ 4730c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "get_write_access"); 4731c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gd_bh); 4732c9de560dSAlex Tomas if (err) 4733c9de560dSAlex Tomas goto error_return; 4734c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 4735c9de560dSAlex Tomas { 4736c9de560dSAlex Tomas int i; 4737c9de560dSAlex Tomas for (i = 0; i < count; i++) 4738c9de560dSAlex Tomas BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 4739c9de560dSAlex Tomas } 4740c9de560dSAlex Tomas #endif 4741c9de560dSAlex Tomas mb_clear_bits(sb_bgl_lock(sbi, block_group), bitmap_bh->b_data, 4742c9de560dSAlex Tomas bit, count); 4743c9de560dSAlex Tomas 4744c9de560dSAlex Tomas /* We dirtied the bitmap block */ 4745c9de560dSAlex Tomas BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 47460390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 4747920313a7SAneesh Kumar K.V if (err) 4748920313a7SAneesh Kumar K.V goto error_return; 4749c9de560dSAlex Tomas 4750256bdb49SEric Sandeen if (ac) { 4751256bdb49SEric Sandeen ac->ac_b_ex.fe_group = block_group; 4752256bdb49SEric Sandeen ac->ac_b_ex.fe_start = bit; 4753256bdb49SEric Sandeen ac->ac_b_ex.fe_len = count; 4754256bdb49SEric Sandeen ext4_mb_store_history(ac); 4755256bdb49SEric Sandeen } 4756c9de560dSAlex Tomas 4757920313a7SAneesh Kumar K.V err = ext4_mb_load_buddy(sb, block_group, &e4b); 4758920313a7SAneesh Kumar K.V if (err) 4759920313a7SAneesh Kumar K.V goto error_return; 47600390131bSFrank Mayhar if (metadata && ext4_handle_valid(handle)) { 4761c9de560dSAlex Tomas /* blocks being freed are metadata. these blocks shouldn't 4762c9de560dSAlex Tomas * be used until this transaction is committed */ 4763c9de560dSAlex Tomas ext4_mb_free_metadata(handle, &e4b, block_group, bit, count); 4764c9de560dSAlex Tomas } else { 4765c9de560dSAlex Tomas ext4_lock_group(sb, block_group); 47667e5a8cddSShen Feng mb_free_blocks(inode, &e4b, bit, count); 4767c9de560dSAlex Tomas ext4_mb_return_to_preallocation(inode, &e4b, block, count); 4768c9de560dSAlex Tomas ext4_unlock_group(sb, block_group); 4769c9de560dSAlex Tomas } 4770c9de560dSAlex Tomas 4771c9de560dSAlex Tomas spin_lock(sb_bgl_lock(sbi, block_group)); 4772e8546d06SMarcin Slusarz le16_add_cpu(&gdp->bg_free_blocks_count, count); 4773c9de560dSAlex Tomas gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); 4774c9de560dSAlex Tomas spin_unlock(sb_bgl_lock(sbi, block_group)); 4775c9de560dSAlex Tomas percpu_counter_add(&sbi->s_freeblocks_counter, count); 4776c9de560dSAlex Tomas 4777772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 4778772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 4779772cb7c8SJose R. Santos spin_lock(sb_bgl_lock(sbi, flex_group)); 4780772cb7c8SJose R. Santos sbi->s_flex_groups[flex_group].free_blocks += count; 4781772cb7c8SJose R. Santos spin_unlock(sb_bgl_lock(sbi, flex_group)); 4782772cb7c8SJose R. Santos } 4783772cb7c8SJose R. Santos 4784c9de560dSAlex Tomas ext4_mb_release_desc(&e4b); 4785c9de560dSAlex Tomas 4786c9de560dSAlex Tomas *freed += count; 4787c9de560dSAlex Tomas 4788c9de560dSAlex Tomas /* And the group descriptor block */ 4789c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 47900390131bSFrank Mayhar ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 4791c9de560dSAlex Tomas if (!err) 4792c9de560dSAlex Tomas err = ret; 4793c9de560dSAlex Tomas 4794c9de560dSAlex Tomas if (overflow && !err) { 4795c9de560dSAlex Tomas block += count; 4796c9de560dSAlex Tomas count = overflow; 4797c9de560dSAlex Tomas put_bh(bitmap_bh); 4798c9de560dSAlex Tomas goto do_more; 4799c9de560dSAlex Tomas } 4800c9de560dSAlex Tomas sb->s_dirt = 1; 4801c9de560dSAlex Tomas error_return: 4802c9de560dSAlex Tomas brelse(bitmap_bh); 4803c9de560dSAlex Tomas ext4_std_error(sb, err); 4804256bdb49SEric Sandeen if (ac) 4805256bdb49SEric Sandeen kmem_cache_free(ext4_ac_cachep, ac); 4806c9de560dSAlex Tomas return; 4807c9de560dSAlex Tomas } 4808