1c9de560dSAlex Tomas /* 2c9de560dSAlex Tomas * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com 3c9de560dSAlex Tomas * Written by Alex Tomas <alex@clusterfs.com> 4c9de560dSAlex Tomas * 5c9de560dSAlex Tomas * This program is free software; you can redistribute it and/or modify 6c9de560dSAlex Tomas * it under the terms of the GNU General Public License version 2 as 7c9de560dSAlex Tomas * published by the Free Software Foundation. 8c9de560dSAlex Tomas * 9c9de560dSAlex Tomas * This program is distributed in the hope that it will be useful, 10c9de560dSAlex Tomas * but WITHOUT ANY WARRANTY; without even the implied warranty of 11c9de560dSAlex Tomas * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12c9de560dSAlex Tomas * GNU General Public License for more details. 13c9de560dSAlex Tomas * 14c9de560dSAlex Tomas * You should have received a copy of the GNU General Public Licens 15c9de560dSAlex Tomas * along with this program; if not, write to the Free Software 16c9de560dSAlex Tomas * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111- 17c9de560dSAlex Tomas */ 18c9de560dSAlex Tomas 19c9de560dSAlex Tomas 20c9de560dSAlex Tomas /* 21c9de560dSAlex Tomas * mballoc.c contains the multiblocks allocation routines 22c9de560dSAlex Tomas */ 23c9de560dSAlex Tomas 248f6e39a7SMingming Cao #include "mballoc.h" 256ba495e9STheodore Ts'o #include <linux/debugfs.h> 265a0e3ad6STejun Heo #include <linux/slab.h> 279bffad1eSTheodore Ts'o #include <trace/events/ext4.h> 289bffad1eSTheodore Ts'o 29c9de560dSAlex Tomas /* 30c9de560dSAlex Tomas * MUSTDO: 31c9de560dSAlex Tomas * - test ext4_ext_search_left() and ext4_ext_search_right() 32c9de560dSAlex Tomas * - search for metadata in few groups 33c9de560dSAlex Tomas * 34c9de560dSAlex Tomas * TODO v4: 35c9de560dSAlex Tomas * - normalization should take into account whether file is still open 36c9de560dSAlex Tomas * - discard preallocations if no free space left (policy?) 37c9de560dSAlex Tomas * - don't normalize tails 38c9de560dSAlex Tomas * - quota 39c9de560dSAlex Tomas * - reservation for superuser 40c9de560dSAlex Tomas * 41c9de560dSAlex Tomas * TODO v3: 42c9de560dSAlex Tomas * - bitmap read-ahead (proposed by Oleg Drokin aka green) 43c9de560dSAlex Tomas * - track min/max extents in each group for better group selection 44c9de560dSAlex Tomas * - mb_mark_used() may allocate chunk right after splitting buddy 45c9de560dSAlex Tomas * - tree of groups sorted by number of free blocks 46c9de560dSAlex Tomas * - error handling 47c9de560dSAlex Tomas */ 48c9de560dSAlex Tomas 49c9de560dSAlex Tomas /* 50c9de560dSAlex Tomas * The allocation request involve request for multiple number of blocks 51c9de560dSAlex Tomas * near to the goal(block) value specified. 52c9de560dSAlex Tomas * 53b713a5ecSTheodore Ts'o * During initialization phase of the allocator we decide to use the 54b713a5ecSTheodore Ts'o * group preallocation or inode preallocation depending on the size of 55b713a5ecSTheodore Ts'o * the file. The size of the file could be the resulting file size we 56b713a5ecSTheodore Ts'o * would have after allocation, or the current file size, which ever 57b713a5ecSTheodore Ts'o * is larger. If the size is less than sbi->s_mb_stream_request we 58b713a5ecSTheodore Ts'o * select to use the group preallocation. The default value of 59b713a5ecSTheodore Ts'o * s_mb_stream_request is 16 blocks. This can also be tuned via 60b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in 61b713a5ecSTheodore Ts'o * terms of number of blocks. 62c9de560dSAlex Tomas * 63c9de560dSAlex Tomas * The main motivation for having small file use group preallocation is to 64b713a5ecSTheodore Ts'o * ensure that we have small files closer together on the disk. 65c9de560dSAlex Tomas * 66b713a5ecSTheodore Ts'o * First stage the allocator looks at the inode prealloc list, 67b713a5ecSTheodore Ts'o * ext4_inode_info->i_prealloc_list, which contains list of prealloc 68b713a5ecSTheodore Ts'o * spaces for this particular inode. The inode prealloc space is 69b713a5ecSTheodore Ts'o * represented as: 70c9de560dSAlex Tomas * 71c9de560dSAlex Tomas * pa_lstart -> the logical start block for this prealloc space 72c9de560dSAlex Tomas * pa_pstart -> the physical start block for this prealloc space 731537a363SDaniel Mack * pa_len -> length for this prealloc space 74c9de560dSAlex Tomas * pa_free -> free space available in this prealloc space 75c9de560dSAlex Tomas * 76c9de560dSAlex Tomas * The inode preallocation space is used looking at the _logical_ start 77c9de560dSAlex Tomas * block. If only the logical file block falls within the range of prealloc 78c9de560dSAlex Tomas * space we will consume the particular prealloc space. This make sure that 79c9de560dSAlex Tomas * that the we have contiguous physical blocks representing the file blocks 80c9de560dSAlex Tomas * 81c9de560dSAlex Tomas * The important thing to be noted in case of inode prealloc space is that 82c9de560dSAlex Tomas * we don't modify the values associated to inode prealloc space except 83c9de560dSAlex Tomas * pa_free. 84c9de560dSAlex Tomas * 85c9de560dSAlex Tomas * If we are not able to find blocks in the inode prealloc space and if we 86c9de560dSAlex Tomas * have the group allocation flag set then we look at the locality group 87c9de560dSAlex Tomas * prealloc space. These are per CPU prealloc list repreasented as 88c9de560dSAlex Tomas * 89c9de560dSAlex Tomas * ext4_sb_info.s_locality_groups[smp_processor_id()] 90c9de560dSAlex Tomas * 91c9de560dSAlex Tomas * The reason for having a per cpu locality group is to reduce the contention 92c9de560dSAlex Tomas * between CPUs. It is possible to get scheduled at this point. 93c9de560dSAlex Tomas * 94c9de560dSAlex Tomas * The locality group prealloc space is used looking at whether we have 9525985edcSLucas De Marchi * enough free space (pa_free) within the prealloc space. 96c9de560dSAlex Tomas * 97c9de560dSAlex Tomas * If we can't allocate blocks via inode prealloc or/and locality group 98c9de560dSAlex Tomas * prealloc then we look at the buddy cache. The buddy cache is represented 99c9de560dSAlex Tomas * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets 100c9de560dSAlex Tomas * mapped to the buddy and bitmap information regarding different 101c9de560dSAlex Tomas * groups. The buddy information is attached to buddy cache inode so that 102c9de560dSAlex Tomas * we can access them through the page cache. The information regarding 103c9de560dSAlex Tomas * each group is loaded via ext4_mb_load_buddy. The information involve 104c9de560dSAlex Tomas * block bitmap and buddy information. The information are stored in the 105c9de560dSAlex Tomas * inode as: 106c9de560dSAlex Tomas * 107c9de560dSAlex Tomas * { page } 108c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 109c9de560dSAlex Tomas * 110c9de560dSAlex Tomas * 111c9de560dSAlex Tomas * one block each for bitmap and buddy information. So for each group we 112c9de560dSAlex Tomas * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE / 113c9de560dSAlex Tomas * blocksize) blocks. So it can have information regarding groups_per_page 114c9de560dSAlex Tomas * which is blocks_per_page/2 115c9de560dSAlex Tomas * 116c9de560dSAlex Tomas * The buddy cache inode is not stored on disk. The inode is thrown 117c9de560dSAlex Tomas * away when the filesystem is unmounted. 118c9de560dSAlex Tomas * 119c9de560dSAlex Tomas * We look for count number of blocks in the buddy cache. If we were able 120c9de560dSAlex Tomas * to locate that many free blocks we return with additional information 121c9de560dSAlex Tomas * regarding rest of the contiguous physical block available 122c9de560dSAlex Tomas * 123c9de560dSAlex Tomas * Before allocating blocks via buddy cache we normalize the request 124c9de560dSAlex Tomas * blocks. This ensure we ask for more blocks that we needed. The extra 125c9de560dSAlex Tomas * blocks that we get after allocation is added to the respective prealloc 126c9de560dSAlex Tomas * list. In case of inode preallocation we follow a list of heuristics 127c9de560dSAlex Tomas * based on file size. This can be found in ext4_mb_normalize_request. If 128c9de560dSAlex Tomas * we are doing a group prealloc we try to normalize the request to 129b713a5ecSTheodore Ts'o * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is 130c9de560dSAlex Tomas * 512 blocks. This can be tuned via 131b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in 132c9de560dSAlex Tomas * terms of number of blocks. If we have mounted the file system with -O 133c9de560dSAlex Tomas * stripe=<value> option the group prealloc request is normalized to the 134c9de560dSAlex Tomas * stripe value (sbi->s_stripe) 135c9de560dSAlex Tomas * 136b713a5ecSTheodore Ts'o * The regular allocator(using the buddy cache) supports few tunables. 137c9de560dSAlex Tomas * 138b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_min_to_scan 139b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_max_to_scan 140b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req 141c9de560dSAlex Tomas * 142b713a5ecSTheodore Ts'o * The regular allocator uses buddy scan only if the request len is power of 143c9de560dSAlex Tomas * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The 144c9de560dSAlex Tomas * value of s_mb_order2_reqs can be tuned via 145b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_order2_req. If the request len is equal to 146af901ca1SAndré Goddard Rosa * stripe size (sbi->s_stripe), we try to search for contiguous block in 147b713a5ecSTheodore Ts'o * stripe size. This should result in better allocation on RAID setups. If 148b713a5ecSTheodore Ts'o * not, we search in the specific group using bitmap for best extents. The 149b713a5ecSTheodore Ts'o * tunable min_to_scan and max_to_scan control the behaviour here. 150c9de560dSAlex Tomas * min_to_scan indicate how long the mballoc __must__ look for a best 151b713a5ecSTheodore Ts'o * extent and max_to_scan indicates how long the mballoc __can__ look for a 152c9de560dSAlex Tomas * best extent in the found extents. Searching for the blocks starts with 153c9de560dSAlex Tomas * the group specified as the goal value in allocation context via 154c9de560dSAlex Tomas * ac_g_ex. Each group is first checked based on the criteria whether it 155c9de560dSAlex Tomas * can used for allocation. ext4_mb_good_group explains how the groups are 156c9de560dSAlex Tomas * checked. 157c9de560dSAlex Tomas * 158c9de560dSAlex Tomas * Both the prealloc space are getting populated as above. So for the first 159c9de560dSAlex Tomas * request we will hit the buddy cache which will result in this prealloc 160c9de560dSAlex Tomas * space getting filled. The prealloc space is then later used for the 161c9de560dSAlex Tomas * subsequent request. 162c9de560dSAlex Tomas */ 163c9de560dSAlex Tomas 164c9de560dSAlex Tomas /* 165c9de560dSAlex Tomas * mballoc operates on the following data: 166c9de560dSAlex Tomas * - on-disk bitmap 167c9de560dSAlex Tomas * - in-core buddy (actually includes buddy and bitmap) 168c9de560dSAlex Tomas * - preallocation descriptors (PAs) 169c9de560dSAlex Tomas * 170c9de560dSAlex Tomas * there are two types of preallocations: 171c9de560dSAlex Tomas * - inode 172c9de560dSAlex Tomas * assiged to specific inode and can be used for this inode only. 173c9de560dSAlex Tomas * it describes part of inode's space preallocated to specific 174c9de560dSAlex Tomas * physical blocks. any block from that preallocated can be used 175c9de560dSAlex Tomas * independent. the descriptor just tracks number of blocks left 176c9de560dSAlex Tomas * unused. so, before taking some block from descriptor, one must 177c9de560dSAlex Tomas * make sure corresponded logical block isn't allocated yet. this 178c9de560dSAlex Tomas * also means that freeing any block within descriptor's range 179c9de560dSAlex Tomas * must discard all preallocated blocks. 180c9de560dSAlex Tomas * - locality group 181c9de560dSAlex Tomas * assigned to specific locality group which does not translate to 182c9de560dSAlex Tomas * permanent set of inodes: inode can join and leave group. space 183c9de560dSAlex Tomas * from this type of preallocation can be used for any inode. thus 184c9de560dSAlex Tomas * it's consumed from the beginning to the end. 185c9de560dSAlex Tomas * 186c9de560dSAlex Tomas * relation between them can be expressed as: 187c9de560dSAlex Tomas * in-core buddy = on-disk bitmap + preallocation descriptors 188c9de560dSAlex Tomas * 189c9de560dSAlex Tomas * this mean blocks mballoc considers used are: 190c9de560dSAlex Tomas * - allocated blocks (persistent) 191c9de560dSAlex Tomas * - preallocated blocks (non-persistent) 192c9de560dSAlex Tomas * 193c9de560dSAlex Tomas * consistency in mballoc world means that at any time a block is either 194c9de560dSAlex Tomas * free or used in ALL structures. notice: "any time" should not be read 195c9de560dSAlex Tomas * literally -- time is discrete and delimited by locks. 196c9de560dSAlex Tomas * 197c9de560dSAlex Tomas * to keep it simple, we don't use block numbers, instead we count number of 198c9de560dSAlex Tomas * blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA. 199c9de560dSAlex Tomas * 200c9de560dSAlex Tomas * all operations can be expressed as: 201c9de560dSAlex Tomas * - init buddy: buddy = on-disk + PAs 202c9de560dSAlex Tomas * - new PA: buddy += N; PA = N 203c9de560dSAlex Tomas * - use inode PA: on-disk += N; PA -= N 204c9de560dSAlex Tomas * - discard inode PA buddy -= on-disk - PA; PA = 0 205c9de560dSAlex Tomas * - use locality group PA on-disk += N; PA -= N 206c9de560dSAlex Tomas * - discard locality group PA buddy -= PA; PA = 0 207c9de560dSAlex Tomas * note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap 208c9de560dSAlex Tomas * is used in real operation because we can't know actual used 209c9de560dSAlex Tomas * bits from PA, only from on-disk bitmap 210c9de560dSAlex Tomas * 211c9de560dSAlex Tomas * if we follow this strict logic, then all operations above should be atomic. 212c9de560dSAlex Tomas * given some of them can block, we'd have to use something like semaphores 213c9de560dSAlex Tomas * killing performance on high-end SMP hardware. let's try to relax it using 214c9de560dSAlex Tomas * the following knowledge: 215c9de560dSAlex Tomas * 1) if buddy is referenced, it's already initialized 216c9de560dSAlex Tomas * 2) while block is used in buddy and the buddy is referenced, 217c9de560dSAlex Tomas * nobody can re-allocate that block 218c9de560dSAlex Tomas * 3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has 219c9de560dSAlex Tomas * bit set and PA claims same block, it's OK. IOW, one can set bit in 220c9de560dSAlex Tomas * on-disk bitmap if buddy has same bit set or/and PA covers corresponded 221c9de560dSAlex Tomas * block 222c9de560dSAlex Tomas * 223c9de560dSAlex Tomas * so, now we're building a concurrency table: 224c9de560dSAlex Tomas * - init buddy vs. 225c9de560dSAlex Tomas * - new PA 226c9de560dSAlex Tomas * blocks for PA are allocated in the buddy, buddy must be referenced 227c9de560dSAlex Tomas * until PA is linked to allocation group to avoid concurrent buddy init 228c9de560dSAlex Tomas * - use inode PA 229c9de560dSAlex Tomas * we need to make sure that either on-disk bitmap or PA has uptodate data 230c9de560dSAlex Tomas * given (3) we care that PA-=N operation doesn't interfere with init 231c9de560dSAlex Tomas * - discard inode PA 232c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 233c9de560dSAlex Tomas * - use locality group PA 234c9de560dSAlex Tomas * again PA-=N must be serialized with init 235c9de560dSAlex Tomas * - discard locality group PA 236c9de560dSAlex Tomas * the simplest way would be to have buddy initialized by the discard 237c9de560dSAlex Tomas * - new PA vs. 238c9de560dSAlex Tomas * - use inode PA 239c9de560dSAlex Tomas * i_data_sem serializes them 240c9de560dSAlex Tomas * - discard inode PA 241c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 242c9de560dSAlex Tomas * - use locality group PA 243c9de560dSAlex Tomas * some mutex should serialize them 244c9de560dSAlex Tomas * - discard locality group PA 245c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 246c9de560dSAlex Tomas * - use inode PA 247c9de560dSAlex Tomas * - use inode PA 248c9de560dSAlex Tomas * i_data_sem or another mutex should serializes them 249c9de560dSAlex Tomas * - discard inode PA 250c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 251c9de560dSAlex Tomas * - use locality group PA 252c9de560dSAlex Tomas * nothing wrong here -- they're different PAs covering different blocks 253c9de560dSAlex Tomas * - discard locality group PA 254c9de560dSAlex Tomas * discard process must wait until PA isn't used by another process 255c9de560dSAlex Tomas * 256c9de560dSAlex Tomas * now we're ready to make few consequences: 257c9de560dSAlex Tomas * - PA is referenced and while it is no discard is possible 258c9de560dSAlex Tomas * - PA is referenced until block isn't marked in on-disk bitmap 259c9de560dSAlex Tomas * - PA changes only after on-disk bitmap 260c9de560dSAlex Tomas * - discard must not compete with init. either init is done before 261c9de560dSAlex Tomas * any discard or they're serialized somehow 262c9de560dSAlex Tomas * - buddy init as sum of on-disk bitmap and PAs is done atomically 263c9de560dSAlex Tomas * 264c9de560dSAlex Tomas * a special case when we've used PA to emptiness. no need to modify buddy 265c9de560dSAlex Tomas * in this case, but we should care about concurrent init 266c9de560dSAlex Tomas * 267c9de560dSAlex Tomas */ 268c9de560dSAlex Tomas 269c9de560dSAlex Tomas /* 270c9de560dSAlex Tomas * Logic in few words: 271c9de560dSAlex Tomas * 272c9de560dSAlex Tomas * - allocation: 273c9de560dSAlex Tomas * load group 274c9de560dSAlex Tomas * find blocks 275c9de560dSAlex Tomas * mark bits in on-disk bitmap 276c9de560dSAlex Tomas * release group 277c9de560dSAlex Tomas * 278c9de560dSAlex Tomas * - use preallocation: 279c9de560dSAlex Tomas * find proper PA (per-inode or group) 280c9de560dSAlex Tomas * load group 281c9de560dSAlex Tomas * mark bits in on-disk bitmap 282c9de560dSAlex Tomas * release group 283c9de560dSAlex Tomas * release PA 284c9de560dSAlex Tomas * 285c9de560dSAlex Tomas * - free: 286c9de560dSAlex Tomas * load group 287c9de560dSAlex Tomas * mark bits in on-disk bitmap 288c9de560dSAlex Tomas * release group 289c9de560dSAlex Tomas * 290c9de560dSAlex Tomas * - discard preallocations in group: 291c9de560dSAlex Tomas * mark PAs deleted 292c9de560dSAlex Tomas * move them onto local list 293c9de560dSAlex Tomas * load on-disk bitmap 294c9de560dSAlex Tomas * load group 295c9de560dSAlex Tomas * remove PA from object (inode or locality group) 296c9de560dSAlex Tomas * mark free blocks in-core 297c9de560dSAlex Tomas * 298c9de560dSAlex Tomas * - discard inode's preallocations: 299c9de560dSAlex Tomas */ 300c9de560dSAlex Tomas 301c9de560dSAlex Tomas /* 302c9de560dSAlex Tomas * Locking rules 303c9de560dSAlex Tomas * 304c9de560dSAlex Tomas * Locks: 305c9de560dSAlex Tomas * - bitlock on a group (group) 306c9de560dSAlex Tomas * - object (inode/locality) (object) 307c9de560dSAlex Tomas * - per-pa lock (pa) 308c9de560dSAlex Tomas * 309c9de560dSAlex Tomas * Paths: 310c9de560dSAlex Tomas * - new pa 311c9de560dSAlex Tomas * object 312c9de560dSAlex Tomas * group 313c9de560dSAlex Tomas * 314c9de560dSAlex Tomas * - find and use pa: 315c9de560dSAlex Tomas * pa 316c9de560dSAlex Tomas * 317c9de560dSAlex Tomas * - release consumed pa: 318c9de560dSAlex Tomas * pa 319c9de560dSAlex Tomas * group 320c9de560dSAlex Tomas * object 321c9de560dSAlex Tomas * 322c9de560dSAlex Tomas * - generate in-core bitmap: 323c9de560dSAlex Tomas * group 324c9de560dSAlex Tomas * pa 325c9de560dSAlex Tomas * 326c9de560dSAlex Tomas * - discard all for given object (inode, locality group): 327c9de560dSAlex Tomas * object 328c9de560dSAlex Tomas * pa 329c9de560dSAlex Tomas * group 330c9de560dSAlex Tomas * 331c9de560dSAlex Tomas * - discard all for given group: 332c9de560dSAlex Tomas * group 333c9de560dSAlex Tomas * pa 334c9de560dSAlex Tomas * group 335c9de560dSAlex Tomas * object 336c9de560dSAlex Tomas * 337c9de560dSAlex Tomas */ 338c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_pspace_cachep; 339c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_ac_cachep; 340c3a326a6SAneesh Kumar K.V static struct kmem_cache *ext4_free_ext_cachep; 341fb1813f4SCurt Wohlgemuth 342fb1813f4SCurt Wohlgemuth /* We create slab caches for groupinfo data structures based on the 343fb1813f4SCurt Wohlgemuth * superblock block size. There will be one per mounted filesystem for 344fb1813f4SCurt Wohlgemuth * each unique s_blocksize_bits */ 3452892c15dSEric Sandeen #define NR_GRPINFO_CACHES 8 346fb1813f4SCurt Wohlgemuth static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES]; 347fb1813f4SCurt Wohlgemuth 3482892c15dSEric Sandeen static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = { 3492892c15dSEric Sandeen "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k", 3502892c15dSEric Sandeen "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k", 3512892c15dSEric Sandeen "ext4_groupinfo_64k", "ext4_groupinfo_128k" 3522892c15dSEric Sandeen }; 3532892c15dSEric Sandeen 354c3a326a6SAneesh Kumar K.V static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 355c3a326a6SAneesh Kumar K.V ext4_group_t group); 3567a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 3577a2fcbf7SAneesh Kumar K.V ext4_group_t group); 358c3a326a6SAneesh Kumar K.V static void release_blocks_on_commit(journal_t *journal, transaction_t *txn); 359c3a326a6SAneesh Kumar K.V 360ffad0a44SAneesh Kumar K.V static inline void *mb_correct_addr_and_bit(int *bit, void *addr) 361ffad0a44SAneesh Kumar K.V { 362c9de560dSAlex Tomas #if BITS_PER_LONG == 64 363ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 7UL) << 3; 364ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~7UL); 365c9de560dSAlex Tomas #elif BITS_PER_LONG == 32 366ffad0a44SAneesh Kumar K.V *bit += ((unsigned long) addr & 3UL) << 3; 367ffad0a44SAneesh Kumar K.V addr = (void *) ((unsigned long) addr & ~3UL); 368c9de560dSAlex Tomas #else 369c9de560dSAlex Tomas #error "how many bits you are?!" 370c9de560dSAlex Tomas #endif 371ffad0a44SAneesh Kumar K.V return addr; 372ffad0a44SAneesh Kumar K.V } 373c9de560dSAlex Tomas 374c9de560dSAlex Tomas static inline int mb_test_bit(int bit, void *addr) 375c9de560dSAlex Tomas { 376c9de560dSAlex Tomas /* 377c9de560dSAlex Tomas * ext4_test_bit on architecture like powerpc 378c9de560dSAlex Tomas * needs unsigned long aligned address 379c9de560dSAlex Tomas */ 380ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 381c9de560dSAlex Tomas return ext4_test_bit(bit, addr); 382c9de560dSAlex Tomas } 383c9de560dSAlex Tomas 384c9de560dSAlex Tomas static inline void mb_set_bit(int bit, void *addr) 385c9de560dSAlex Tomas { 386ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 387c9de560dSAlex Tomas ext4_set_bit(bit, addr); 388c9de560dSAlex Tomas } 389c9de560dSAlex Tomas 390c9de560dSAlex Tomas static inline void mb_clear_bit(int bit, void *addr) 391c9de560dSAlex Tomas { 392ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&bit, addr); 393c9de560dSAlex Tomas ext4_clear_bit(bit, addr); 394c9de560dSAlex Tomas } 395c9de560dSAlex Tomas 396ffad0a44SAneesh Kumar K.V static inline int mb_find_next_zero_bit(void *addr, int max, int start) 397ffad0a44SAneesh Kumar K.V { 398e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 399ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 400e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 401ffad0a44SAneesh Kumar K.V start += fix; 402ffad0a44SAneesh Kumar K.V 403e7dfb246SAneesh Kumar K.V ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix; 404e7dfb246SAneesh Kumar K.V if (ret > max) 405e7dfb246SAneesh Kumar K.V return max; 406e7dfb246SAneesh Kumar K.V return ret; 407ffad0a44SAneesh Kumar K.V } 408ffad0a44SAneesh Kumar K.V 409ffad0a44SAneesh Kumar K.V static inline int mb_find_next_bit(void *addr, int max, int start) 410ffad0a44SAneesh Kumar K.V { 411e7dfb246SAneesh Kumar K.V int fix = 0, ret, tmpmax; 412ffad0a44SAneesh Kumar K.V addr = mb_correct_addr_and_bit(&fix, addr); 413e7dfb246SAneesh Kumar K.V tmpmax = max + fix; 414ffad0a44SAneesh Kumar K.V start += fix; 415ffad0a44SAneesh Kumar K.V 416e7dfb246SAneesh Kumar K.V ret = ext4_find_next_bit(addr, tmpmax, start) - fix; 417e7dfb246SAneesh Kumar K.V if (ret > max) 418e7dfb246SAneesh Kumar K.V return max; 419e7dfb246SAneesh Kumar K.V return ret; 420ffad0a44SAneesh Kumar K.V } 421ffad0a44SAneesh Kumar K.V 422c9de560dSAlex Tomas static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max) 423c9de560dSAlex Tomas { 424c9de560dSAlex Tomas char *bb; 425c9de560dSAlex Tomas 426c9de560dSAlex Tomas BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); 427c9de560dSAlex Tomas BUG_ON(max == NULL); 428c9de560dSAlex Tomas 429c9de560dSAlex Tomas if (order > e4b->bd_blkbits + 1) { 430c9de560dSAlex Tomas *max = 0; 431c9de560dSAlex Tomas return NULL; 432c9de560dSAlex Tomas } 433c9de560dSAlex Tomas 434c9de560dSAlex Tomas /* at order 0 we see each particular block */ 43584b775a3SColy Li if (order == 0) { 436c9de560dSAlex Tomas *max = 1 << (e4b->bd_blkbits + 3); 437c9de560dSAlex Tomas return EXT4_MB_BITMAP(e4b); 43884b775a3SColy Li } 439c9de560dSAlex Tomas 440c9de560dSAlex Tomas bb = EXT4_MB_BUDDY(e4b) + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order]; 441c9de560dSAlex Tomas *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order]; 442c9de560dSAlex Tomas 443c9de560dSAlex Tomas return bb; 444c9de560dSAlex Tomas } 445c9de560dSAlex Tomas 446c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 447c9de560dSAlex Tomas static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b, 448c9de560dSAlex Tomas int first, int count) 449c9de560dSAlex Tomas { 450c9de560dSAlex Tomas int i; 451c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 452c9de560dSAlex Tomas 453c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 454c9de560dSAlex Tomas return; 455bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 456c9de560dSAlex Tomas for (i = 0; i < count; i++) { 457c9de560dSAlex Tomas if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) { 458c9de560dSAlex Tomas ext4_fsblk_t blocknr; 4595661bd68SAkinobu Mita 4605661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 461c9de560dSAlex Tomas blocknr += first + i; 4625d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 463e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 464e29136f8STheodore Ts'o blocknr, 465e29136f8STheodore Ts'o "freeing block already freed " 466e29136f8STheodore Ts'o "(bit %u)", 467e29136f8STheodore Ts'o first + i); 468c9de560dSAlex Tomas } 469c9de560dSAlex Tomas mb_clear_bit(first + i, e4b->bd_info->bb_bitmap); 470c9de560dSAlex Tomas } 471c9de560dSAlex Tomas } 472c9de560dSAlex Tomas 473c9de560dSAlex Tomas static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count) 474c9de560dSAlex Tomas { 475c9de560dSAlex Tomas int i; 476c9de560dSAlex Tomas 477c9de560dSAlex Tomas if (unlikely(e4b->bd_info->bb_bitmap == NULL)) 478c9de560dSAlex Tomas return; 479bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 480c9de560dSAlex Tomas for (i = 0; i < count; i++) { 481c9de560dSAlex Tomas BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap)); 482c9de560dSAlex Tomas mb_set_bit(first + i, e4b->bd_info->bb_bitmap); 483c9de560dSAlex Tomas } 484c9de560dSAlex Tomas } 485c9de560dSAlex Tomas 486c9de560dSAlex Tomas static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 487c9de560dSAlex Tomas { 488c9de560dSAlex Tomas if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) { 489c9de560dSAlex Tomas unsigned char *b1, *b2; 490c9de560dSAlex Tomas int i; 491c9de560dSAlex Tomas b1 = (unsigned char *) e4b->bd_info->bb_bitmap; 492c9de560dSAlex Tomas b2 = (unsigned char *) bitmap; 493c9de560dSAlex Tomas for (i = 0; i < e4b->bd_sb->s_blocksize; i++) { 494c9de560dSAlex Tomas if (b1[i] != b2[i]) { 495a9df9a49STheodore Ts'o printk(KERN_ERR "corruption in group %u " 4964776004fSTheodore Ts'o "at byte %u(%u): %x in copy != %x " 4974776004fSTheodore Ts'o "on disk/prealloc\n", 498c9de560dSAlex Tomas e4b->bd_group, i, i * 8, b1[i], b2[i]); 499c9de560dSAlex Tomas BUG(); 500c9de560dSAlex Tomas } 501c9de560dSAlex Tomas } 502c9de560dSAlex Tomas } 503c9de560dSAlex Tomas } 504c9de560dSAlex Tomas 505c9de560dSAlex Tomas #else 506c9de560dSAlex Tomas static inline void mb_free_blocks_double(struct inode *inode, 507c9de560dSAlex Tomas struct ext4_buddy *e4b, int first, int count) 508c9de560dSAlex Tomas { 509c9de560dSAlex Tomas return; 510c9de560dSAlex Tomas } 511c9de560dSAlex Tomas static inline void mb_mark_used_double(struct ext4_buddy *e4b, 512c9de560dSAlex Tomas int first, int count) 513c9de560dSAlex Tomas { 514c9de560dSAlex Tomas return; 515c9de560dSAlex Tomas } 516c9de560dSAlex Tomas static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap) 517c9de560dSAlex Tomas { 518c9de560dSAlex Tomas return; 519c9de560dSAlex Tomas } 520c9de560dSAlex Tomas #endif 521c9de560dSAlex Tomas 522c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 523c9de560dSAlex Tomas 524c9de560dSAlex Tomas #define MB_CHECK_ASSERT(assert) \ 525c9de560dSAlex Tomas do { \ 526c9de560dSAlex Tomas if (!(assert)) { \ 527c9de560dSAlex Tomas printk(KERN_EMERG \ 528c9de560dSAlex Tomas "Assertion failure in %s() at %s:%d: \"%s\"\n", \ 529c9de560dSAlex Tomas function, file, line, # assert); \ 530c9de560dSAlex Tomas BUG(); \ 531c9de560dSAlex Tomas } \ 532c9de560dSAlex Tomas } while (0) 533c9de560dSAlex Tomas 534c9de560dSAlex Tomas static int __mb_check_buddy(struct ext4_buddy *e4b, char *file, 535c9de560dSAlex Tomas const char *function, int line) 536c9de560dSAlex Tomas { 537c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 538c9de560dSAlex Tomas int order = e4b->bd_blkbits + 1; 539c9de560dSAlex Tomas int max; 540c9de560dSAlex Tomas int max2; 541c9de560dSAlex Tomas int i; 542c9de560dSAlex Tomas int j; 543c9de560dSAlex Tomas int k; 544c9de560dSAlex Tomas int count; 545c9de560dSAlex Tomas struct ext4_group_info *grp; 546c9de560dSAlex Tomas int fragments = 0; 547c9de560dSAlex Tomas int fstart; 548c9de560dSAlex Tomas struct list_head *cur; 549c9de560dSAlex Tomas void *buddy; 550c9de560dSAlex Tomas void *buddy2; 551c9de560dSAlex Tomas 552c9de560dSAlex Tomas { 553c9de560dSAlex Tomas static int mb_check_counter; 554c9de560dSAlex Tomas if (mb_check_counter++ % 100 != 0) 555c9de560dSAlex Tomas return 0; 556c9de560dSAlex Tomas } 557c9de560dSAlex Tomas 558c9de560dSAlex Tomas while (order > 1) { 559c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 560c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy); 561c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order - 1, &max2); 562c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy2); 563c9de560dSAlex Tomas MB_CHECK_ASSERT(buddy != buddy2); 564c9de560dSAlex Tomas MB_CHECK_ASSERT(max * 2 == max2); 565c9de560dSAlex Tomas 566c9de560dSAlex Tomas count = 0; 567c9de560dSAlex Tomas for (i = 0; i < max; i++) { 568c9de560dSAlex Tomas 569c9de560dSAlex Tomas if (mb_test_bit(i, buddy)) { 570c9de560dSAlex Tomas /* only single bit in buddy2 may be 1 */ 571c9de560dSAlex Tomas if (!mb_test_bit(i << 1, buddy2)) { 572c9de560dSAlex Tomas MB_CHECK_ASSERT( 573c9de560dSAlex Tomas mb_test_bit((i<<1)+1, buddy2)); 574c9de560dSAlex Tomas } else if (!mb_test_bit((i << 1) + 1, buddy2)) { 575c9de560dSAlex Tomas MB_CHECK_ASSERT( 576c9de560dSAlex Tomas mb_test_bit(i << 1, buddy2)); 577c9de560dSAlex Tomas } 578c9de560dSAlex Tomas continue; 579c9de560dSAlex Tomas } 580c9de560dSAlex Tomas 581c9de560dSAlex Tomas /* both bits in buddy2 must be 0 */ 582c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2)); 583c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2)); 584c9de560dSAlex Tomas 585c9de560dSAlex Tomas for (j = 0; j < (1 << order); j++) { 586c9de560dSAlex Tomas k = (i * (1 << order)) + j; 587c9de560dSAlex Tomas MB_CHECK_ASSERT( 588c9de560dSAlex Tomas !mb_test_bit(k, EXT4_MB_BITMAP(e4b))); 589c9de560dSAlex Tomas } 590c9de560dSAlex Tomas count++; 591c9de560dSAlex Tomas } 592c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count); 593c9de560dSAlex Tomas order--; 594c9de560dSAlex Tomas } 595c9de560dSAlex Tomas 596c9de560dSAlex Tomas fstart = -1; 597c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, 0, &max); 598c9de560dSAlex Tomas for (i = 0; i < max; i++) { 599c9de560dSAlex Tomas if (!mb_test_bit(i, buddy)) { 600c9de560dSAlex Tomas MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free); 601c9de560dSAlex Tomas if (fstart == -1) { 602c9de560dSAlex Tomas fragments++; 603c9de560dSAlex Tomas fstart = i; 604c9de560dSAlex Tomas } 605c9de560dSAlex Tomas continue; 606c9de560dSAlex Tomas } 607c9de560dSAlex Tomas fstart = -1; 608c9de560dSAlex Tomas /* check used bits only */ 609c9de560dSAlex Tomas for (j = 0; j < e4b->bd_blkbits + 1; j++) { 610c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, j, &max2); 611c9de560dSAlex Tomas k = i >> j; 612c9de560dSAlex Tomas MB_CHECK_ASSERT(k < max2); 613c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k, buddy2)); 614c9de560dSAlex Tomas } 615c9de560dSAlex Tomas } 616c9de560dSAlex Tomas MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info)); 617c9de560dSAlex Tomas MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments); 618c9de560dSAlex Tomas 619c9de560dSAlex Tomas grp = ext4_get_group_info(sb, e4b->bd_group); 620c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 621c9de560dSAlex Tomas ext4_group_t groupnr; 622c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 62360bd63d1SSolofo Ramangalahy pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 62460bd63d1SSolofo Ramangalahy ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k); 625c9de560dSAlex Tomas MB_CHECK_ASSERT(groupnr == e4b->bd_group); 62660bd63d1SSolofo Ramangalahy for (i = 0; i < pa->pa_len; i++) 627c9de560dSAlex Tomas MB_CHECK_ASSERT(mb_test_bit(k + i, buddy)); 628c9de560dSAlex Tomas } 629c9de560dSAlex Tomas return 0; 630c9de560dSAlex Tomas } 631c9de560dSAlex Tomas #undef MB_CHECK_ASSERT 632c9de560dSAlex Tomas #define mb_check_buddy(e4b) __mb_check_buddy(e4b, \ 63346e665e9SHarvey Harrison __FILE__, __func__, __LINE__) 634c9de560dSAlex Tomas #else 635c9de560dSAlex Tomas #define mb_check_buddy(e4b) 636c9de560dSAlex Tomas #endif 637c9de560dSAlex Tomas 6387c786059SColy Li /* 6397c786059SColy Li * Divide blocks started from @first with length @len into 6407c786059SColy Li * smaller chunks with power of 2 blocks. 6417c786059SColy Li * Clear the bits in bitmap which the blocks of the chunk(s) covered, 6427c786059SColy Li * then increase bb_counters[] for corresponded chunk size. 6437c786059SColy Li */ 644c9de560dSAlex Tomas static void ext4_mb_mark_free_simple(struct super_block *sb, 645a36b4498SEric Sandeen void *buddy, ext4_grpblk_t first, ext4_grpblk_t len, 646c9de560dSAlex Tomas struct ext4_group_info *grp) 647c9de560dSAlex Tomas { 648c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 649a36b4498SEric Sandeen ext4_grpblk_t min; 650a36b4498SEric Sandeen ext4_grpblk_t max; 651a36b4498SEric Sandeen ext4_grpblk_t chunk; 652c9de560dSAlex Tomas unsigned short border; 653c9de560dSAlex Tomas 654b73fce69SValerie Clement BUG_ON(len > EXT4_BLOCKS_PER_GROUP(sb)); 655c9de560dSAlex Tomas 656c9de560dSAlex Tomas border = 2 << sb->s_blocksize_bits; 657c9de560dSAlex Tomas 658c9de560dSAlex Tomas while (len > 0) { 659c9de560dSAlex Tomas /* find how many blocks can be covered since this position */ 660c9de560dSAlex Tomas max = ffs(first | border) - 1; 661c9de560dSAlex Tomas 662c9de560dSAlex Tomas /* find how many blocks of power 2 we need to mark */ 663c9de560dSAlex Tomas min = fls(len) - 1; 664c9de560dSAlex Tomas 665c9de560dSAlex Tomas if (max < min) 666c9de560dSAlex Tomas min = max; 667c9de560dSAlex Tomas chunk = 1 << min; 668c9de560dSAlex Tomas 669c9de560dSAlex Tomas /* mark multiblock chunks only */ 670c9de560dSAlex Tomas grp->bb_counters[min]++; 671c9de560dSAlex Tomas if (min > 0) 672c9de560dSAlex Tomas mb_clear_bit(first >> min, 673c9de560dSAlex Tomas buddy + sbi->s_mb_offsets[min]); 674c9de560dSAlex Tomas 675c9de560dSAlex Tomas len -= chunk; 676c9de560dSAlex Tomas first += chunk; 677c9de560dSAlex Tomas } 678c9de560dSAlex Tomas } 679c9de560dSAlex Tomas 6808a57d9d6SCurt Wohlgemuth /* 6818a57d9d6SCurt Wohlgemuth * Cache the order of the largest free extent we have available in this block 6828a57d9d6SCurt Wohlgemuth * group. 6838a57d9d6SCurt Wohlgemuth */ 6848a57d9d6SCurt Wohlgemuth static void 6858a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) 6868a57d9d6SCurt Wohlgemuth { 6878a57d9d6SCurt Wohlgemuth int i; 6888a57d9d6SCurt Wohlgemuth int bits; 6898a57d9d6SCurt Wohlgemuth 6908a57d9d6SCurt Wohlgemuth grp->bb_largest_free_order = -1; /* uninit */ 6918a57d9d6SCurt Wohlgemuth 6928a57d9d6SCurt Wohlgemuth bits = sb->s_blocksize_bits + 1; 6938a57d9d6SCurt Wohlgemuth for (i = bits; i >= 0; i--) { 6948a57d9d6SCurt Wohlgemuth if (grp->bb_counters[i] > 0) { 6958a57d9d6SCurt Wohlgemuth grp->bb_largest_free_order = i; 6968a57d9d6SCurt Wohlgemuth break; 6978a57d9d6SCurt Wohlgemuth } 6988a57d9d6SCurt Wohlgemuth } 6998a57d9d6SCurt Wohlgemuth } 7008a57d9d6SCurt Wohlgemuth 701089ceeccSEric Sandeen static noinline_for_stack 702089ceeccSEric Sandeen void ext4_mb_generate_buddy(struct super_block *sb, 703c9de560dSAlex Tomas void *buddy, void *bitmap, ext4_group_t group) 704c9de560dSAlex Tomas { 705c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 706a36b4498SEric Sandeen ext4_grpblk_t max = EXT4_BLOCKS_PER_GROUP(sb); 707a36b4498SEric Sandeen ext4_grpblk_t i = 0; 708a36b4498SEric Sandeen ext4_grpblk_t first; 709a36b4498SEric Sandeen ext4_grpblk_t len; 710c9de560dSAlex Tomas unsigned free = 0; 711c9de560dSAlex Tomas unsigned fragments = 0; 712c9de560dSAlex Tomas unsigned long long period = get_cycles(); 713c9de560dSAlex Tomas 714c9de560dSAlex Tomas /* initialize buddy from bitmap which is aggregation 715c9de560dSAlex Tomas * of on-disk bitmap and preallocations */ 716ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, 0); 717c9de560dSAlex Tomas grp->bb_first_free = i; 718c9de560dSAlex Tomas while (i < max) { 719c9de560dSAlex Tomas fragments++; 720c9de560dSAlex Tomas first = i; 721ffad0a44SAneesh Kumar K.V i = mb_find_next_bit(bitmap, max, i); 722c9de560dSAlex Tomas len = i - first; 723c9de560dSAlex Tomas free += len; 724c9de560dSAlex Tomas if (len > 1) 725c9de560dSAlex Tomas ext4_mb_mark_free_simple(sb, buddy, first, len, grp); 726c9de560dSAlex Tomas else 727c9de560dSAlex Tomas grp->bb_counters[0]++; 728c9de560dSAlex Tomas if (i < max) 729ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, max, i); 730c9de560dSAlex Tomas } 731c9de560dSAlex Tomas grp->bb_fragments = fragments; 732c9de560dSAlex Tomas 733c9de560dSAlex Tomas if (free != grp->bb_free) { 734e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, 735e29136f8STheodore Ts'o "%u blocks in bitmap, %u in gd", 736e29136f8STheodore Ts'o free, grp->bb_free); 737e56eb659SAneesh Kumar K.V /* 738e56eb659SAneesh Kumar K.V * If we intent to continue, we consider group descritor 739e56eb659SAneesh Kumar K.V * corrupt and update bb_free using bitmap value 740e56eb659SAneesh Kumar K.V */ 741c9de560dSAlex Tomas grp->bb_free = free; 742c9de560dSAlex Tomas } 7438a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, grp); 744c9de560dSAlex Tomas 745c9de560dSAlex Tomas clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state)); 746c9de560dSAlex Tomas 747c9de560dSAlex Tomas period = get_cycles() - period; 748c9de560dSAlex Tomas spin_lock(&EXT4_SB(sb)->s_bal_lock); 749c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_buddies_generated++; 750c9de560dSAlex Tomas EXT4_SB(sb)->s_mb_generation_time += period; 751c9de560dSAlex Tomas spin_unlock(&EXT4_SB(sb)->s_bal_lock); 752c9de560dSAlex Tomas } 753c9de560dSAlex Tomas 754c9de560dSAlex Tomas /* The buddy information is attached the buddy cache inode 755c9de560dSAlex Tomas * for convenience. The information regarding each group 756c9de560dSAlex Tomas * is loaded via ext4_mb_load_buddy. The information involve 757c9de560dSAlex Tomas * block bitmap and buddy information. The information are 758c9de560dSAlex Tomas * stored in the inode as 759c9de560dSAlex Tomas * 760c9de560dSAlex Tomas * { page } 761c3a326a6SAneesh Kumar K.V * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]... 762c9de560dSAlex Tomas * 763c9de560dSAlex Tomas * 764c9de560dSAlex Tomas * one block each for bitmap and buddy information. 765c9de560dSAlex Tomas * So for each group we take up 2 blocks. A page can 766c9de560dSAlex Tomas * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize) blocks. 767c9de560dSAlex Tomas * So it can have information regarding groups_per_page which 768c9de560dSAlex Tomas * is blocks_per_page/2 7698a57d9d6SCurt Wohlgemuth * 7708a57d9d6SCurt Wohlgemuth * Locking note: This routine takes the block group lock of all groups 7718a57d9d6SCurt Wohlgemuth * for this page; do not hold this lock when calling this routine! 772c9de560dSAlex Tomas */ 773c9de560dSAlex Tomas 774c9de560dSAlex Tomas static int ext4_mb_init_cache(struct page *page, char *incore) 775c9de560dSAlex Tomas { 7768df9675fSTheodore Ts'o ext4_group_t ngroups; 777c9de560dSAlex Tomas int blocksize; 778c9de560dSAlex Tomas int blocks_per_page; 779c9de560dSAlex Tomas int groups_per_page; 780c9de560dSAlex Tomas int err = 0; 781c9de560dSAlex Tomas int i; 782c9de560dSAlex Tomas ext4_group_t first_group; 783c9de560dSAlex Tomas int first_block; 784c9de560dSAlex Tomas struct super_block *sb; 785c9de560dSAlex Tomas struct buffer_head *bhs; 786c9de560dSAlex Tomas struct buffer_head **bh; 787c9de560dSAlex Tomas struct inode *inode; 788c9de560dSAlex Tomas char *data; 789c9de560dSAlex Tomas char *bitmap; 790c9de560dSAlex Tomas 7916ba495e9STheodore Ts'o mb_debug(1, "init page %lu\n", page->index); 792c9de560dSAlex Tomas 793c9de560dSAlex Tomas inode = page->mapping->host; 794c9de560dSAlex Tomas sb = inode->i_sb; 7958df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 796c9de560dSAlex Tomas blocksize = 1 << inode->i_blkbits; 797c9de560dSAlex Tomas blocks_per_page = PAGE_CACHE_SIZE / blocksize; 798c9de560dSAlex Tomas 799c9de560dSAlex Tomas groups_per_page = blocks_per_page >> 1; 800c9de560dSAlex Tomas if (groups_per_page == 0) 801c9de560dSAlex Tomas groups_per_page = 1; 802c9de560dSAlex Tomas 803c9de560dSAlex Tomas /* allocate buffer_heads to read bitmaps */ 804c9de560dSAlex Tomas if (groups_per_page > 1) { 805c9de560dSAlex Tomas err = -ENOMEM; 806c9de560dSAlex Tomas i = sizeof(struct buffer_head *) * groups_per_page; 807c9de560dSAlex Tomas bh = kzalloc(i, GFP_NOFS); 808c9de560dSAlex Tomas if (bh == NULL) 809c9de560dSAlex Tomas goto out; 810c9de560dSAlex Tomas } else 811c9de560dSAlex Tomas bh = &bhs; 812c9de560dSAlex Tomas 813c9de560dSAlex Tomas first_group = page->index * blocks_per_page / 2; 814c9de560dSAlex Tomas 815c9de560dSAlex Tomas /* read all groups the page covers into the cache */ 816c9de560dSAlex Tomas for (i = 0; i < groups_per_page; i++) { 817c9de560dSAlex Tomas struct ext4_group_desc *desc; 818c9de560dSAlex Tomas 8198df9675fSTheodore Ts'o if (first_group + i >= ngroups) 820c9de560dSAlex Tomas break; 821c9de560dSAlex Tomas 822c9de560dSAlex Tomas err = -EIO; 823c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, first_group + i, NULL); 824c9de560dSAlex Tomas if (desc == NULL) 825c9de560dSAlex Tomas goto out; 826c9de560dSAlex Tomas 827c9de560dSAlex Tomas err = -ENOMEM; 828c9de560dSAlex Tomas bh[i] = sb_getblk(sb, ext4_block_bitmap(sb, desc)); 829c9de560dSAlex Tomas if (bh[i] == NULL) 830c9de560dSAlex Tomas goto out; 831c9de560dSAlex Tomas 8322ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh[i])) 833c9de560dSAlex Tomas continue; 834c9de560dSAlex Tomas 835c806e68fSFrederic Bohe lock_buffer(bh[i]); 8362ccb5fb9SAneesh Kumar K.V if (bitmap_uptodate(bh[i])) { 8372ccb5fb9SAneesh Kumar K.V unlock_buffer(bh[i]); 8382ccb5fb9SAneesh Kumar K.V continue; 8392ccb5fb9SAneesh Kumar K.V } 840955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, first_group + i); 841c9de560dSAlex Tomas if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 842c9de560dSAlex Tomas ext4_init_block_bitmap(sb, bh[i], 843c9de560dSAlex Tomas first_group + i, desc); 8442ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh[i]); 845c9de560dSAlex Tomas set_buffer_uptodate(bh[i]); 846955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, first_group + i); 8473300bedaSAneesh Kumar K.V unlock_buffer(bh[i]); 848c9de560dSAlex Tomas continue; 849c9de560dSAlex Tomas } 850955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, first_group + i); 8512ccb5fb9SAneesh Kumar K.V if (buffer_uptodate(bh[i])) { 8522ccb5fb9SAneesh Kumar K.V /* 8532ccb5fb9SAneesh Kumar K.V * if not uninit if bh is uptodate, 8542ccb5fb9SAneesh Kumar K.V * bitmap is also uptodate 8552ccb5fb9SAneesh Kumar K.V */ 8562ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh[i]); 8572ccb5fb9SAneesh Kumar K.V unlock_buffer(bh[i]); 8582ccb5fb9SAneesh Kumar K.V continue; 8592ccb5fb9SAneesh Kumar K.V } 860c9de560dSAlex Tomas get_bh(bh[i]); 8612ccb5fb9SAneesh Kumar K.V /* 8622ccb5fb9SAneesh Kumar K.V * submit the buffer_head for read. We can 8632ccb5fb9SAneesh Kumar K.V * safely mark the bitmap as uptodate now. 8642ccb5fb9SAneesh Kumar K.V * We do it here so the bitmap uptodate bit 8652ccb5fb9SAneesh Kumar K.V * get set with buffer lock held. 8662ccb5fb9SAneesh Kumar K.V */ 8672ccb5fb9SAneesh Kumar K.V set_bitmap_uptodate(bh[i]); 868c9de560dSAlex Tomas bh[i]->b_end_io = end_buffer_read_sync; 869c9de560dSAlex Tomas submit_bh(READ, bh[i]); 8706ba495e9STheodore Ts'o mb_debug(1, "read bitmap for group %u\n", first_group + i); 871c9de560dSAlex Tomas } 872c9de560dSAlex Tomas 873c9de560dSAlex Tomas /* wait for I/O completion */ 874c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 875c9de560dSAlex Tomas wait_on_buffer(bh[i]); 876c9de560dSAlex Tomas 877c9de560dSAlex Tomas err = -EIO; 878c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 879c9de560dSAlex Tomas if (!buffer_uptodate(bh[i])) 880c9de560dSAlex Tomas goto out; 881c9de560dSAlex Tomas 88231b481dcSMingming Cao err = 0; 883c9de560dSAlex Tomas first_block = page->index * blocks_per_page; 88429eaf024SAneesh Kumar K.V /* init the page */ 88529eaf024SAneesh Kumar K.V memset(page_address(page), 0xff, PAGE_CACHE_SIZE); 886c9de560dSAlex Tomas for (i = 0; i < blocks_per_page; i++) { 887c9de560dSAlex Tomas int group; 888c9de560dSAlex Tomas struct ext4_group_info *grinfo; 889c9de560dSAlex Tomas 890c9de560dSAlex Tomas group = (first_block + i) >> 1; 8918df9675fSTheodore Ts'o if (group >= ngroups) 892c9de560dSAlex Tomas break; 893c9de560dSAlex Tomas 894c9de560dSAlex Tomas /* 895c9de560dSAlex Tomas * data carry information regarding this 896c9de560dSAlex Tomas * particular group in the format specified 897c9de560dSAlex Tomas * above 898c9de560dSAlex Tomas * 899c9de560dSAlex Tomas */ 900c9de560dSAlex Tomas data = page_address(page) + (i * blocksize); 901c9de560dSAlex Tomas bitmap = bh[group - first_group]->b_data; 902c9de560dSAlex Tomas 903c9de560dSAlex Tomas /* 904c9de560dSAlex Tomas * We place the buddy block and bitmap block 905c9de560dSAlex Tomas * close together 906c9de560dSAlex Tomas */ 907c9de560dSAlex Tomas if ((first_block + i) & 1) { 908c9de560dSAlex Tomas /* this is block of buddy */ 909c9de560dSAlex Tomas BUG_ON(incore == NULL); 9106ba495e9STheodore Ts'o mb_debug(1, "put buddy for group %u in page %lu/%x\n", 911c9de560dSAlex Tomas group, page->index, i * blocksize); 912f307333eSTheodore Ts'o trace_ext4_mb_buddy_bitmap_load(sb, group); 913c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, group); 914c9de560dSAlex Tomas grinfo->bb_fragments = 0; 915c9de560dSAlex Tomas memset(grinfo->bb_counters, 0, 9161927805eSEric Sandeen sizeof(*grinfo->bb_counters) * 9171927805eSEric Sandeen (sb->s_blocksize_bits+2)); 918c9de560dSAlex Tomas /* 919c9de560dSAlex Tomas * incore got set to the group block bitmap below 920c9de560dSAlex Tomas */ 9217a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, group); 922c9de560dSAlex Tomas ext4_mb_generate_buddy(sb, data, incore, group); 9237a2fcbf7SAneesh Kumar K.V ext4_unlock_group(sb, group); 924c9de560dSAlex Tomas incore = NULL; 925c9de560dSAlex Tomas } else { 926c9de560dSAlex Tomas /* this is block of bitmap */ 927c9de560dSAlex Tomas BUG_ON(incore != NULL); 9286ba495e9STheodore Ts'o mb_debug(1, "put bitmap for group %u in page %lu/%x\n", 929c9de560dSAlex Tomas group, page->index, i * blocksize); 930f307333eSTheodore Ts'o trace_ext4_mb_bitmap_load(sb, group); 931c9de560dSAlex Tomas 932c9de560dSAlex Tomas /* see comments in ext4_mb_put_pa() */ 933c9de560dSAlex Tomas ext4_lock_group(sb, group); 934c9de560dSAlex Tomas memcpy(data, bitmap, blocksize); 935c9de560dSAlex Tomas 936c9de560dSAlex Tomas /* mark all preallocated blks used in in-core bitmap */ 937c9de560dSAlex Tomas ext4_mb_generate_from_pa(sb, data, group); 9387a2fcbf7SAneesh Kumar K.V ext4_mb_generate_from_freelist(sb, data, group); 939c9de560dSAlex Tomas ext4_unlock_group(sb, group); 940c9de560dSAlex Tomas 941c9de560dSAlex Tomas /* set incore so that the buddy information can be 942c9de560dSAlex Tomas * generated using this 943c9de560dSAlex Tomas */ 944c9de560dSAlex Tomas incore = data; 945c9de560dSAlex Tomas } 946c9de560dSAlex Tomas } 947c9de560dSAlex Tomas SetPageUptodate(page); 948c9de560dSAlex Tomas 949c9de560dSAlex Tomas out: 950c9de560dSAlex Tomas if (bh) { 951c9de560dSAlex Tomas for (i = 0; i < groups_per_page && bh[i]; i++) 952c9de560dSAlex Tomas brelse(bh[i]); 953c9de560dSAlex Tomas if (bh != &bhs) 954c9de560dSAlex Tomas kfree(bh); 955c9de560dSAlex Tomas } 956c9de560dSAlex Tomas return err; 957c9de560dSAlex Tomas } 958c9de560dSAlex Tomas 9598a57d9d6SCurt Wohlgemuth /* 960eee4adc7SEric Sandeen * lock the group_info alloc_sem of all the groups 961eee4adc7SEric Sandeen * belonging to the same buddy cache page. This 962eee4adc7SEric Sandeen * make sure other parallel operation on the buddy 963eee4adc7SEric Sandeen * cache doesn't happen whild holding the buddy cache 964eee4adc7SEric Sandeen * lock 965eee4adc7SEric Sandeen */ 966eee4adc7SEric Sandeen static int ext4_mb_get_buddy_cache_lock(struct super_block *sb, 967eee4adc7SEric Sandeen ext4_group_t group) 968eee4adc7SEric Sandeen { 969eee4adc7SEric Sandeen int i; 970eee4adc7SEric Sandeen int block, pnum; 971eee4adc7SEric Sandeen int blocks_per_page; 972eee4adc7SEric Sandeen int groups_per_page; 973eee4adc7SEric Sandeen ext4_group_t ngroups = ext4_get_groups_count(sb); 974eee4adc7SEric Sandeen ext4_group_t first_group; 975eee4adc7SEric Sandeen struct ext4_group_info *grp; 976eee4adc7SEric Sandeen 977eee4adc7SEric Sandeen blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 978eee4adc7SEric Sandeen /* 979eee4adc7SEric Sandeen * the buddy cache inode stores the block bitmap 980eee4adc7SEric Sandeen * and buddy information in consecutive blocks. 981eee4adc7SEric Sandeen * So for each group we need two blocks. 982eee4adc7SEric Sandeen */ 983eee4adc7SEric Sandeen block = group * 2; 984eee4adc7SEric Sandeen pnum = block / blocks_per_page; 985eee4adc7SEric Sandeen first_group = pnum * blocks_per_page / 2; 986eee4adc7SEric Sandeen 987eee4adc7SEric Sandeen groups_per_page = blocks_per_page >> 1; 988eee4adc7SEric Sandeen if (groups_per_page == 0) 989eee4adc7SEric Sandeen groups_per_page = 1; 990eee4adc7SEric Sandeen /* read all groups the page covers into the cache */ 991eee4adc7SEric Sandeen for (i = 0; i < groups_per_page; i++) { 992eee4adc7SEric Sandeen 993eee4adc7SEric Sandeen if ((first_group + i) >= ngroups) 994eee4adc7SEric Sandeen break; 995eee4adc7SEric Sandeen grp = ext4_get_group_info(sb, first_group + i); 996eee4adc7SEric Sandeen /* take all groups write allocation 997eee4adc7SEric Sandeen * semaphore. This make sure there is 998eee4adc7SEric Sandeen * no block allocation going on in any 999eee4adc7SEric Sandeen * of that groups 1000eee4adc7SEric Sandeen */ 1001eee4adc7SEric Sandeen down_write_nested(&grp->alloc_sem, i); 1002eee4adc7SEric Sandeen } 1003eee4adc7SEric Sandeen return i; 1004eee4adc7SEric Sandeen } 1005eee4adc7SEric Sandeen 1006eee4adc7SEric Sandeen static void ext4_mb_put_buddy_cache_lock(struct super_block *sb, 1007eee4adc7SEric Sandeen ext4_group_t group, int locked_group) 1008eee4adc7SEric Sandeen { 1009eee4adc7SEric Sandeen int i; 1010eee4adc7SEric Sandeen int block, pnum; 1011eee4adc7SEric Sandeen int blocks_per_page; 1012eee4adc7SEric Sandeen ext4_group_t first_group; 1013eee4adc7SEric Sandeen struct ext4_group_info *grp; 1014eee4adc7SEric Sandeen 1015eee4adc7SEric Sandeen blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1016eee4adc7SEric Sandeen /* 1017eee4adc7SEric Sandeen * the buddy cache inode stores the block bitmap 1018eee4adc7SEric Sandeen * and buddy information in consecutive blocks. 1019eee4adc7SEric Sandeen * So for each group we need two blocks. 1020eee4adc7SEric Sandeen */ 1021eee4adc7SEric Sandeen block = group * 2; 1022eee4adc7SEric Sandeen pnum = block / blocks_per_page; 1023eee4adc7SEric Sandeen first_group = pnum * blocks_per_page / 2; 1024eee4adc7SEric Sandeen /* release locks on all the groups */ 1025eee4adc7SEric Sandeen for (i = 0; i < locked_group; i++) { 1026eee4adc7SEric Sandeen 1027eee4adc7SEric Sandeen grp = ext4_get_group_info(sb, first_group + i); 1028eee4adc7SEric Sandeen /* take all groups write allocation 1029eee4adc7SEric Sandeen * semaphore. This make sure there is 1030eee4adc7SEric Sandeen * no block allocation going on in any 1031eee4adc7SEric Sandeen * of that groups 1032eee4adc7SEric Sandeen */ 1033eee4adc7SEric Sandeen up_write(&grp->alloc_sem); 1034eee4adc7SEric Sandeen } 1035eee4adc7SEric Sandeen 1036eee4adc7SEric Sandeen } 1037eee4adc7SEric Sandeen 1038eee4adc7SEric Sandeen /* 10398a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 10408a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 10418a57d9d6SCurt Wohlgemuth * calling this routine! 10428a57d9d6SCurt Wohlgemuth */ 1043b6a758ecSAneesh Kumar K.V static noinline_for_stack 1044b6a758ecSAneesh Kumar K.V int ext4_mb_init_group(struct super_block *sb, ext4_group_t group) 1045b6a758ecSAneesh Kumar K.V { 1046b6a758ecSAneesh Kumar K.V 1047b6a758ecSAneesh Kumar K.V int ret = 0; 1048b6a758ecSAneesh Kumar K.V void *bitmap; 1049b6a758ecSAneesh Kumar K.V int blocks_per_page; 1050b6a758ecSAneesh Kumar K.V int block, pnum, poff; 1051b6a758ecSAneesh Kumar K.V int num_grp_locked = 0; 1052b6a758ecSAneesh Kumar K.V struct ext4_group_info *this_grp; 1053b6a758ecSAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 1054b6a758ecSAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 1055b6a758ecSAneesh Kumar K.V struct page *page = NULL, *bitmap_page = NULL; 1056b6a758ecSAneesh Kumar K.V 1057b6a758ecSAneesh Kumar K.V mb_debug(1, "init group %u\n", group); 1058b6a758ecSAneesh Kumar K.V blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1059b6a758ecSAneesh Kumar K.V this_grp = ext4_get_group_info(sb, group); 1060b6a758ecSAneesh Kumar K.V /* 106108c3a813SAneesh Kumar K.V * This ensures that we don't reinit the buddy cache 106208c3a813SAneesh Kumar K.V * page which map to the group from which we are already 106308c3a813SAneesh Kumar K.V * allocating. If we are looking at the buddy cache we would 106408c3a813SAneesh Kumar K.V * have taken a reference using ext4_mb_load_buddy and that 106508c3a813SAneesh Kumar K.V * would have taken the alloc_sem lock. 1066b6a758ecSAneesh Kumar K.V */ 1067b6a758ecSAneesh Kumar K.V num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, group); 1068b6a758ecSAneesh Kumar K.V if (!EXT4_MB_GRP_NEED_INIT(this_grp)) { 1069b6a758ecSAneesh Kumar K.V /* 1070b6a758ecSAneesh Kumar K.V * somebody initialized the group 1071b6a758ecSAneesh Kumar K.V * return without doing anything 1072b6a758ecSAneesh Kumar K.V */ 1073b6a758ecSAneesh Kumar K.V ret = 0; 1074b6a758ecSAneesh Kumar K.V goto err; 1075b6a758ecSAneesh Kumar K.V } 1076b6a758ecSAneesh Kumar K.V /* 1077b6a758ecSAneesh Kumar K.V * the buddy cache inode stores the block bitmap 1078b6a758ecSAneesh Kumar K.V * and buddy information in consecutive blocks. 1079b6a758ecSAneesh Kumar K.V * So for each group we need two blocks. 1080b6a758ecSAneesh Kumar K.V */ 1081b6a758ecSAneesh Kumar K.V block = group * 2; 1082b6a758ecSAneesh Kumar K.V pnum = block / blocks_per_page; 1083b6a758ecSAneesh Kumar K.V poff = block % blocks_per_page; 1084b6a758ecSAneesh Kumar K.V page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1085b6a758ecSAneesh Kumar K.V if (page) { 1086b6a758ecSAneesh Kumar K.V BUG_ON(page->mapping != inode->i_mapping); 1087b6a758ecSAneesh Kumar K.V ret = ext4_mb_init_cache(page, NULL); 1088b6a758ecSAneesh Kumar K.V if (ret) { 1089b6a758ecSAneesh Kumar K.V unlock_page(page); 1090b6a758ecSAneesh Kumar K.V goto err; 1091b6a758ecSAneesh Kumar K.V } 1092b6a758ecSAneesh Kumar K.V unlock_page(page); 1093b6a758ecSAneesh Kumar K.V } 1094b6a758ecSAneesh Kumar K.V if (page == NULL || !PageUptodate(page)) { 1095b6a758ecSAneesh Kumar K.V ret = -EIO; 1096b6a758ecSAneesh Kumar K.V goto err; 1097b6a758ecSAneesh Kumar K.V } 1098b6a758ecSAneesh Kumar K.V mark_page_accessed(page); 1099b6a758ecSAneesh Kumar K.V bitmap_page = page; 1100b6a758ecSAneesh Kumar K.V bitmap = page_address(page) + (poff * sb->s_blocksize); 1101b6a758ecSAneesh Kumar K.V 1102b6a758ecSAneesh Kumar K.V /* init buddy cache */ 1103b6a758ecSAneesh Kumar K.V block++; 1104b6a758ecSAneesh Kumar K.V pnum = block / blocks_per_page; 1105b6a758ecSAneesh Kumar K.V poff = block % blocks_per_page; 1106b6a758ecSAneesh Kumar K.V page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1107b6a758ecSAneesh Kumar K.V if (page == bitmap_page) { 1108b6a758ecSAneesh Kumar K.V /* 1109b6a758ecSAneesh Kumar K.V * If both the bitmap and buddy are in 1110b6a758ecSAneesh Kumar K.V * the same page we don't need to force 1111b6a758ecSAneesh Kumar K.V * init the buddy 1112b6a758ecSAneesh Kumar K.V */ 1113b6a758ecSAneesh Kumar K.V unlock_page(page); 1114b6a758ecSAneesh Kumar K.V } else if (page) { 1115b6a758ecSAneesh Kumar K.V BUG_ON(page->mapping != inode->i_mapping); 1116b6a758ecSAneesh Kumar K.V ret = ext4_mb_init_cache(page, bitmap); 1117b6a758ecSAneesh Kumar K.V if (ret) { 1118b6a758ecSAneesh Kumar K.V unlock_page(page); 1119b6a758ecSAneesh Kumar K.V goto err; 1120b6a758ecSAneesh Kumar K.V } 1121b6a758ecSAneesh Kumar K.V unlock_page(page); 1122b6a758ecSAneesh Kumar K.V } 1123b6a758ecSAneesh Kumar K.V if (page == NULL || !PageUptodate(page)) { 1124b6a758ecSAneesh Kumar K.V ret = -EIO; 1125b6a758ecSAneesh Kumar K.V goto err; 1126b6a758ecSAneesh Kumar K.V } 1127b6a758ecSAneesh Kumar K.V mark_page_accessed(page); 1128b6a758ecSAneesh Kumar K.V err: 1129b6a758ecSAneesh Kumar K.V ext4_mb_put_buddy_cache_lock(sb, group, num_grp_locked); 1130b6a758ecSAneesh Kumar K.V if (bitmap_page) 1131b6a758ecSAneesh Kumar K.V page_cache_release(bitmap_page); 1132b6a758ecSAneesh Kumar K.V if (page) 1133b6a758ecSAneesh Kumar K.V page_cache_release(page); 1134b6a758ecSAneesh Kumar K.V return ret; 1135b6a758ecSAneesh Kumar K.V } 1136b6a758ecSAneesh Kumar K.V 11378a57d9d6SCurt Wohlgemuth /* 11388a57d9d6SCurt Wohlgemuth * Locking note: This routine calls ext4_mb_init_cache(), which takes the 11398a57d9d6SCurt Wohlgemuth * block group lock of all groups for this page; do not hold the BG lock when 11408a57d9d6SCurt Wohlgemuth * calling this routine! 11418a57d9d6SCurt Wohlgemuth */ 11424ddfef7bSEric Sandeen static noinline_for_stack int 11434ddfef7bSEric Sandeen ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group, 1144c9de560dSAlex Tomas struct ext4_buddy *e4b) 1145c9de560dSAlex Tomas { 1146c9de560dSAlex Tomas int blocks_per_page; 1147c9de560dSAlex Tomas int block; 1148c9de560dSAlex Tomas int pnum; 1149c9de560dSAlex Tomas int poff; 1150c9de560dSAlex Tomas struct page *page; 1151fdf6c7a7SShen Feng int ret; 1152920313a7SAneesh Kumar K.V struct ext4_group_info *grp; 1153920313a7SAneesh Kumar K.V struct ext4_sb_info *sbi = EXT4_SB(sb); 1154920313a7SAneesh Kumar K.V struct inode *inode = sbi->s_buddy_cache; 1155c9de560dSAlex Tomas 11566ba495e9STheodore Ts'o mb_debug(1, "load group %u\n", group); 1157c9de560dSAlex Tomas 1158c9de560dSAlex Tomas blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize; 1159920313a7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 1160c9de560dSAlex Tomas 1161c9de560dSAlex Tomas e4b->bd_blkbits = sb->s_blocksize_bits; 1162c9de560dSAlex Tomas e4b->bd_info = ext4_get_group_info(sb, group); 1163c9de560dSAlex Tomas e4b->bd_sb = sb; 1164c9de560dSAlex Tomas e4b->bd_group = group; 1165c9de560dSAlex Tomas e4b->bd_buddy_page = NULL; 1166c9de560dSAlex Tomas e4b->bd_bitmap_page = NULL; 1167920313a7SAneesh Kumar K.V e4b->alloc_semp = &grp->alloc_sem; 1168920313a7SAneesh Kumar K.V 1169920313a7SAneesh Kumar K.V /* Take the read lock on the group alloc 1170920313a7SAneesh Kumar K.V * sem. This would make sure a parallel 1171920313a7SAneesh Kumar K.V * ext4_mb_init_group happening on other 1172920313a7SAneesh Kumar K.V * groups mapped by the page is blocked 1173920313a7SAneesh Kumar K.V * till we are done with allocation 1174920313a7SAneesh Kumar K.V */ 1175f41c0750SAneesh Kumar K.V repeat_load_buddy: 1176920313a7SAneesh Kumar K.V down_read(e4b->alloc_semp); 1177c9de560dSAlex Tomas 1178f41c0750SAneesh Kumar K.V if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 1179f41c0750SAneesh Kumar K.V /* we need to check for group need init flag 1180f41c0750SAneesh Kumar K.V * with alloc_semp held so that we can be sure 1181f41c0750SAneesh Kumar K.V * that new blocks didn't get added to the group 1182f41c0750SAneesh Kumar K.V * when we are loading the buddy cache 1183f41c0750SAneesh Kumar K.V */ 1184f41c0750SAneesh Kumar K.V up_read(e4b->alloc_semp); 1185f41c0750SAneesh Kumar K.V /* 1186f41c0750SAneesh Kumar K.V * we need full data about the group 1187f41c0750SAneesh Kumar K.V * to make a good selection 1188f41c0750SAneesh Kumar K.V */ 1189f41c0750SAneesh Kumar K.V ret = ext4_mb_init_group(sb, group); 1190f41c0750SAneesh Kumar K.V if (ret) 1191f41c0750SAneesh Kumar K.V return ret; 1192f41c0750SAneesh Kumar K.V goto repeat_load_buddy; 1193f41c0750SAneesh Kumar K.V } 1194f41c0750SAneesh Kumar K.V 1195c9de560dSAlex Tomas /* 1196c9de560dSAlex Tomas * the buddy cache inode stores the block bitmap 1197c9de560dSAlex Tomas * and buddy information in consecutive blocks. 1198c9de560dSAlex Tomas * So for each group we need two blocks. 1199c9de560dSAlex Tomas */ 1200c9de560dSAlex Tomas block = group * 2; 1201c9de560dSAlex Tomas pnum = block / blocks_per_page; 1202c9de560dSAlex Tomas poff = block % blocks_per_page; 1203c9de560dSAlex Tomas 1204c9de560dSAlex Tomas /* we could use find_or_create_page(), but it locks page 1205c9de560dSAlex Tomas * what we'd like to avoid in fast path ... */ 1206c9de560dSAlex Tomas page = find_get_page(inode->i_mapping, pnum); 1207c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1208c9de560dSAlex Tomas if (page) 1209920313a7SAneesh Kumar K.V /* 1210920313a7SAneesh Kumar K.V * drop the page reference and try 1211920313a7SAneesh Kumar K.V * to get the page with lock. If we 1212920313a7SAneesh Kumar K.V * are not uptodate that implies 1213920313a7SAneesh Kumar K.V * somebody just created the page but 1214920313a7SAneesh Kumar K.V * is yet to initialize the same. So 1215920313a7SAneesh Kumar K.V * wait for it to initialize. 1216920313a7SAneesh Kumar K.V */ 1217c9de560dSAlex Tomas page_cache_release(page); 1218c9de560dSAlex Tomas page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1219c9de560dSAlex Tomas if (page) { 1220c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 1221c9de560dSAlex Tomas if (!PageUptodate(page)) { 1222fdf6c7a7SShen Feng ret = ext4_mb_init_cache(page, NULL); 1223fdf6c7a7SShen Feng if (ret) { 1224fdf6c7a7SShen Feng unlock_page(page); 1225fdf6c7a7SShen Feng goto err; 1226fdf6c7a7SShen Feng } 1227c9de560dSAlex Tomas mb_cmp_bitmaps(e4b, page_address(page) + 1228c9de560dSAlex Tomas (poff * sb->s_blocksize)); 1229c9de560dSAlex Tomas } 1230c9de560dSAlex Tomas unlock_page(page); 1231c9de560dSAlex Tomas } 1232c9de560dSAlex Tomas } 1233fdf6c7a7SShen Feng if (page == NULL || !PageUptodate(page)) { 1234fdf6c7a7SShen Feng ret = -EIO; 1235c9de560dSAlex Tomas goto err; 1236fdf6c7a7SShen Feng } 1237c9de560dSAlex Tomas e4b->bd_bitmap_page = page; 1238c9de560dSAlex Tomas e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize); 1239c9de560dSAlex Tomas mark_page_accessed(page); 1240c9de560dSAlex Tomas 1241c9de560dSAlex Tomas block++; 1242c9de560dSAlex Tomas pnum = block / blocks_per_page; 1243c9de560dSAlex Tomas poff = block % blocks_per_page; 1244c9de560dSAlex Tomas 1245c9de560dSAlex Tomas page = find_get_page(inode->i_mapping, pnum); 1246c9de560dSAlex Tomas if (page == NULL || !PageUptodate(page)) { 1247c9de560dSAlex Tomas if (page) 1248c9de560dSAlex Tomas page_cache_release(page); 1249c9de560dSAlex Tomas page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS); 1250c9de560dSAlex Tomas if (page) { 1251c9de560dSAlex Tomas BUG_ON(page->mapping != inode->i_mapping); 1252fdf6c7a7SShen Feng if (!PageUptodate(page)) { 1253fdf6c7a7SShen Feng ret = ext4_mb_init_cache(page, e4b->bd_bitmap); 1254fdf6c7a7SShen Feng if (ret) { 1255fdf6c7a7SShen Feng unlock_page(page); 1256fdf6c7a7SShen Feng goto err; 1257fdf6c7a7SShen Feng } 1258fdf6c7a7SShen Feng } 1259c9de560dSAlex Tomas unlock_page(page); 1260c9de560dSAlex Tomas } 1261c9de560dSAlex Tomas } 1262fdf6c7a7SShen Feng if (page == NULL || !PageUptodate(page)) { 1263fdf6c7a7SShen Feng ret = -EIO; 1264c9de560dSAlex Tomas goto err; 1265fdf6c7a7SShen Feng } 1266c9de560dSAlex Tomas e4b->bd_buddy_page = page; 1267c9de560dSAlex Tomas e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize); 1268c9de560dSAlex Tomas mark_page_accessed(page); 1269c9de560dSAlex Tomas 1270c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 1271c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 1272c9de560dSAlex Tomas 1273c9de560dSAlex Tomas return 0; 1274c9de560dSAlex Tomas 1275c9de560dSAlex Tomas err: 1276*26626f11SYang Ruirui if (page) 1277*26626f11SYang Ruirui page_cache_release(page); 1278c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 1279c9de560dSAlex Tomas page_cache_release(e4b->bd_bitmap_page); 1280c9de560dSAlex Tomas if (e4b->bd_buddy_page) 1281c9de560dSAlex Tomas page_cache_release(e4b->bd_buddy_page); 1282c9de560dSAlex Tomas e4b->bd_buddy = NULL; 1283c9de560dSAlex Tomas e4b->bd_bitmap = NULL; 1284920313a7SAneesh Kumar K.V 1285920313a7SAneesh Kumar K.V /* Done with the buddy cache */ 1286920313a7SAneesh Kumar K.V up_read(e4b->alloc_semp); 1287fdf6c7a7SShen Feng return ret; 1288c9de560dSAlex Tomas } 1289c9de560dSAlex Tomas 1290e39e07fdSJing Zhang static void ext4_mb_unload_buddy(struct ext4_buddy *e4b) 1291c9de560dSAlex Tomas { 1292c9de560dSAlex Tomas if (e4b->bd_bitmap_page) 1293c9de560dSAlex Tomas page_cache_release(e4b->bd_bitmap_page); 1294c9de560dSAlex Tomas if (e4b->bd_buddy_page) 1295c9de560dSAlex Tomas page_cache_release(e4b->bd_buddy_page); 1296920313a7SAneesh Kumar K.V /* Done with the buddy cache */ 12978556e8f3SAneesh Kumar K.V if (e4b->alloc_semp) 1298920313a7SAneesh Kumar K.V up_read(e4b->alloc_semp); 1299c9de560dSAlex Tomas } 1300c9de560dSAlex Tomas 1301c9de560dSAlex Tomas 1302c9de560dSAlex Tomas static int mb_find_order_for_block(struct ext4_buddy *e4b, int block) 1303c9de560dSAlex Tomas { 1304c9de560dSAlex Tomas int order = 1; 1305c9de560dSAlex Tomas void *bb; 1306c9de560dSAlex Tomas 1307c9de560dSAlex Tomas BUG_ON(EXT4_MB_BITMAP(e4b) == EXT4_MB_BUDDY(e4b)); 1308c9de560dSAlex Tomas BUG_ON(block >= (1 << (e4b->bd_blkbits + 3))); 1309c9de560dSAlex Tomas 1310c9de560dSAlex Tomas bb = EXT4_MB_BUDDY(e4b); 1311c9de560dSAlex Tomas while (order <= e4b->bd_blkbits + 1) { 1312c9de560dSAlex Tomas block = block >> 1; 1313c9de560dSAlex Tomas if (!mb_test_bit(block, bb)) { 1314c9de560dSAlex Tomas /* this block is part of buddy of order 'order' */ 1315c9de560dSAlex Tomas return order; 1316c9de560dSAlex Tomas } 1317c9de560dSAlex Tomas bb += 1 << (e4b->bd_blkbits - order); 1318c9de560dSAlex Tomas order++; 1319c9de560dSAlex Tomas } 1320c9de560dSAlex Tomas return 0; 1321c9de560dSAlex Tomas } 1322c9de560dSAlex Tomas 1323955ce5f5SAneesh Kumar K.V static void mb_clear_bits(void *bm, int cur, int len) 1324c9de560dSAlex Tomas { 1325c9de560dSAlex Tomas __u32 *addr; 1326c9de560dSAlex Tomas 1327c9de560dSAlex Tomas len = cur + len; 1328c9de560dSAlex Tomas while (cur < len) { 1329c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1330c9de560dSAlex Tomas /* fast path: clear whole word at once */ 1331c9de560dSAlex Tomas addr = bm + (cur >> 3); 1332c9de560dSAlex Tomas *addr = 0; 1333c9de560dSAlex Tomas cur += 32; 1334c9de560dSAlex Tomas continue; 1335c9de560dSAlex Tomas } 1336e8134b27SAneesh Kumar K.V mb_clear_bit(cur, bm); 1337c9de560dSAlex Tomas cur++; 1338c9de560dSAlex Tomas } 1339c9de560dSAlex Tomas } 1340c9de560dSAlex Tomas 1341955ce5f5SAneesh Kumar K.V static void mb_set_bits(void *bm, int cur, int len) 1342c9de560dSAlex Tomas { 1343c9de560dSAlex Tomas __u32 *addr; 1344c9de560dSAlex Tomas 1345c9de560dSAlex Tomas len = cur + len; 1346c9de560dSAlex Tomas while (cur < len) { 1347c9de560dSAlex Tomas if ((cur & 31) == 0 && (len - cur) >= 32) { 1348c9de560dSAlex Tomas /* fast path: set whole word at once */ 1349c9de560dSAlex Tomas addr = bm + (cur >> 3); 1350c9de560dSAlex Tomas *addr = 0xffffffff; 1351c9de560dSAlex Tomas cur += 32; 1352c9de560dSAlex Tomas continue; 1353c9de560dSAlex Tomas } 1354e8134b27SAneesh Kumar K.V mb_set_bit(cur, bm); 1355c9de560dSAlex Tomas cur++; 1356c9de560dSAlex Tomas } 1357c9de560dSAlex Tomas } 1358c9de560dSAlex Tomas 13597e5a8cddSShen Feng static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b, 1360c9de560dSAlex Tomas int first, int count) 1361c9de560dSAlex Tomas { 1362c9de560dSAlex Tomas int block = 0; 1363c9de560dSAlex Tomas int max = 0; 1364c9de560dSAlex Tomas int order; 1365c9de560dSAlex Tomas void *buddy; 1366c9de560dSAlex Tomas void *buddy2; 1367c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 1368c9de560dSAlex Tomas 1369c9de560dSAlex Tomas BUG_ON(first + count > (sb->s_blocksize << 3)); 1370bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group)); 1371c9de560dSAlex Tomas mb_check_buddy(e4b); 1372c9de560dSAlex Tomas mb_free_blocks_double(inode, e4b, first, count); 1373c9de560dSAlex Tomas 1374c9de560dSAlex Tomas e4b->bd_info->bb_free += count; 1375c9de560dSAlex Tomas if (first < e4b->bd_info->bb_first_free) 1376c9de560dSAlex Tomas e4b->bd_info->bb_first_free = first; 1377c9de560dSAlex Tomas 1378c9de560dSAlex Tomas /* let's maintain fragments counter */ 1379c9de560dSAlex Tomas if (first != 0) 1380c9de560dSAlex Tomas block = !mb_test_bit(first - 1, EXT4_MB_BITMAP(e4b)); 1381c9de560dSAlex Tomas if (first + count < EXT4_SB(sb)->s_mb_maxs[0]) 1382c9de560dSAlex Tomas max = !mb_test_bit(first + count, EXT4_MB_BITMAP(e4b)); 1383c9de560dSAlex Tomas if (block && max) 1384c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1385c9de560dSAlex Tomas else if (!block && !max) 1386c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1387c9de560dSAlex Tomas 1388c9de560dSAlex Tomas /* let's maintain buddy itself */ 1389c9de560dSAlex Tomas while (count-- > 0) { 1390c9de560dSAlex Tomas block = first++; 1391c9de560dSAlex Tomas order = 0; 1392c9de560dSAlex Tomas 1393c9de560dSAlex Tomas if (!mb_test_bit(block, EXT4_MB_BITMAP(e4b))) { 1394c9de560dSAlex Tomas ext4_fsblk_t blocknr; 13955661bd68SAkinobu Mita 13965661bd68SAkinobu Mita blocknr = ext4_group_first_block_no(sb, e4b->bd_group); 1397c9de560dSAlex Tomas blocknr += block; 13985d1b1b3fSAneesh Kumar K.V ext4_grp_locked_error(sb, e4b->bd_group, 1399e29136f8STheodore Ts'o inode ? inode->i_ino : 0, 1400e29136f8STheodore Ts'o blocknr, 1401e29136f8STheodore Ts'o "freeing already freed block " 1402e29136f8STheodore Ts'o "(bit %u)", block); 1403c9de560dSAlex Tomas } 1404c9de560dSAlex Tomas mb_clear_bit(block, EXT4_MB_BITMAP(e4b)); 1405c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]++; 1406c9de560dSAlex Tomas 1407c9de560dSAlex Tomas /* start of the buddy */ 1408c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 1409c9de560dSAlex Tomas 1410c9de560dSAlex Tomas do { 1411c9de560dSAlex Tomas block &= ~1UL; 1412c9de560dSAlex Tomas if (mb_test_bit(block, buddy) || 1413c9de560dSAlex Tomas mb_test_bit(block + 1, buddy)) 1414c9de560dSAlex Tomas break; 1415c9de560dSAlex Tomas 1416c9de560dSAlex Tomas /* both the buddies are free, try to coalesce them */ 1417c9de560dSAlex Tomas buddy2 = mb_find_buddy(e4b, order + 1, &max); 1418c9de560dSAlex Tomas 1419c9de560dSAlex Tomas if (!buddy2) 1420c9de560dSAlex Tomas break; 1421c9de560dSAlex Tomas 1422c9de560dSAlex Tomas if (order > 0) { 1423c9de560dSAlex Tomas /* for special purposes, we don't set 1424c9de560dSAlex Tomas * free bits in bitmap */ 1425c9de560dSAlex Tomas mb_set_bit(block, buddy); 1426c9de560dSAlex Tomas mb_set_bit(block + 1, buddy); 1427c9de560dSAlex Tomas } 1428c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]--; 1429c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]--; 1430c9de560dSAlex Tomas 1431c9de560dSAlex Tomas block = block >> 1; 1432c9de560dSAlex Tomas order++; 1433c9de560dSAlex Tomas e4b->bd_info->bb_counters[order]++; 1434c9de560dSAlex Tomas 1435c9de560dSAlex Tomas mb_clear_bit(block, buddy2); 1436c9de560dSAlex Tomas buddy = buddy2; 1437c9de560dSAlex Tomas } while (1); 1438c9de560dSAlex Tomas } 14398a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(sb, e4b->bd_info); 1440c9de560dSAlex Tomas mb_check_buddy(e4b); 1441c9de560dSAlex Tomas } 1442c9de560dSAlex Tomas 1443c9de560dSAlex Tomas static int mb_find_extent(struct ext4_buddy *e4b, int order, int block, 1444c9de560dSAlex Tomas int needed, struct ext4_free_extent *ex) 1445c9de560dSAlex Tomas { 1446c9de560dSAlex Tomas int next = block; 1447c9de560dSAlex Tomas int max; 1448c9de560dSAlex Tomas int ord; 1449c9de560dSAlex Tomas void *buddy; 1450c9de560dSAlex Tomas 1451bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1452c9de560dSAlex Tomas BUG_ON(ex == NULL); 1453c9de560dSAlex Tomas 1454c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, order, &max); 1455c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1456c9de560dSAlex Tomas BUG_ON(block >= max); 1457c9de560dSAlex Tomas if (mb_test_bit(block, buddy)) { 1458c9de560dSAlex Tomas ex->fe_len = 0; 1459c9de560dSAlex Tomas ex->fe_start = 0; 1460c9de560dSAlex Tomas ex->fe_group = 0; 1461c9de560dSAlex Tomas return 0; 1462c9de560dSAlex Tomas } 1463c9de560dSAlex Tomas 1464c9de560dSAlex Tomas /* FIXME dorp order completely ? */ 1465c9de560dSAlex Tomas if (likely(order == 0)) { 1466c9de560dSAlex Tomas /* find actual order */ 1467c9de560dSAlex Tomas order = mb_find_order_for_block(e4b, block); 1468c9de560dSAlex Tomas block = block >> order; 1469c9de560dSAlex Tomas } 1470c9de560dSAlex Tomas 1471c9de560dSAlex Tomas ex->fe_len = 1 << order; 1472c9de560dSAlex Tomas ex->fe_start = block << order; 1473c9de560dSAlex Tomas ex->fe_group = e4b->bd_group; 1474c9de560dSAlex Tomas 1475c9de560dSAlex Tomas /* calc difference from given start */ 1476c9de560dSAlex Tomas next = next - ex->fe_start; 1477c9de560dSAlex Tomas ex->fe_len -= next; 1478c9de560dSAlex Tomas ex->fe_start += next; 1479c9de560dSAlex Tomas 1480c9de560dSAlex Tomas while (needed > ex->fe_len && 1481c9de560dSAlex Tomas (buddy = mb_find_buddy(e4b, order, &max))) { 1482c9de560dSAlex Tomas 1483c9de560dSAlex Tomas if (block + 1 >= max) 1484c9de560dSAlex Tomas break; 1485c9de560dSAlex Tomas 1486c9de560dSAlex Tomas next = (block + 1) * (1 << order); 1487c9de560dSAlex Tomas if (mb_test_bit(next, EXT4_MB_BITMAP(e4b))) 1488c9de560dSAlex Tomas break; 1489c9de560dSAlex Tomas 1490c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, next); 1491c9de560dSAlex Tomas 1492c9de560dSAlex Tomas order = ord; 1493c9de560dSAlex Tomas block = next >> order; 1494c9de560dSAlex Tomas ex->fe_len += 1 << order; 1495c9de560dSAlex Tomas } 1496c9de560dSAlex Tomas 1497c9de560dSAlex Tomas BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3))); 1498c9de560dSAlex Tomas return ex->fe_len; 1499c9de560dSAlex Tomas } 1500c9de560dSAlex Tomas 1501c9de560dSAlex Tomas static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex) 1502c9de560dSAlex Tomas { 1503c9de560dSAlex Tomas int ord; 1504c9de560dSAlex Tomas int mlen = 0; 1505c9de560dSAlex Tomas int max = 0; 1506c9de560dSAlex Tomas int cur; 1507c9de560dSAlex Tomas int start = ex->fe_start; 1508c9de560dSAlex Tomas int len = ex->fe_len; 1509c9de560dSAlex Tomas unsigned ret = 0; 1510c9de560dSAlex Tomas int len0 = len; 1511c9de560dSAlex Tomas void *buddy; 1512c9de560dSAlex Tomas 1513c9de560dSAlex Tomas BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3)); 1514c9de560dSAlex Tomas BUG_ON(e4b->bd_group != ex->fe_group); 1515bc8e6740SVincent Minet assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group)); 1516c9de560dSAlex Tomas mb_check_buddy(e4b); 1517c9de560dSAlex Tomas mb_mark_used_double(e4b, start, len); 1518c9de560dSAlex Tomas 1519c9de560dSAlex Tomas e4b->bd_info->bb_free -= len; 1520c9de560dSAlex Tomas if (e4b->bd_info->bb_first_free == start) 1521c9de560dSAlex Tomas e4b->bd_info->bb_first_free += len; 1522c9de560dSAlex Tomas 1523c9de560dSAlex Tomas /* let's maintain fragments counter */ 1524c9de560dSAlex Tomas if (start != 0) 1525c9de560dSAlex Tomas mlen = !mb_test_bit(start - 1, EXT4_MB_BITMAP(e4b)); 1526c9de560dSAlex Tomas if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0]) 1527c9de560dSAlex Tomas max = !mb_test_bit(start + len, EXT4_MB_BITMAP(e4b)); 1528c9de560dSAlex Tomas if (mlen && max) 1529c9de560dSAlex Tomas e4b->bd_info->bb_fragments++; 1530c9de560dSAlex Tomas else if (!mlen && !max) 1531c9de560dSAlex Tomas e4b->bd_info->bb_fragments--; 1532c9de560dSAlex Tomas 1533c9de560dSAlex Tomas /* let's maintain buddy itself */ 1534c9de560dSAlex Tomas while (len) { 1535c9de560dSAlex Tomas ord = mb_find_order_for_block(e4b, start); 1536c9de560dSAlex Tomas 1537c9de560dSAlex Tomas if (((start >> ord) << ord) == start && len >= (1 << ord)) { 1538c9de560dSAlex Tomas /* the whole chunk may be allocated at once! */ 1539c9de560dSAlex Tomas mlen = 1 << ord; 1540c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1541c9de560dSAlex Tomas BUG_ON((start >> ord) >= max); 1542c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1543c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1544c9de560dSAlex Tomas start += mlen; 1545c9de560dSAlex Tomas len -= mlen; 1546c9de560dSAlex Tomas BUG_ON(len < 0); 1547c9de560dSAlex Tomas continue; 1548c9de560dSAlex Tomas } 1549c9de560dSAlex Tomas 1550c9de560dSAlex Tomas /* store for history */ 1551c9de560dSAlex Tomas if (ret == 0) 1552c9de560dSAlex Tomas ret = len | (ord << 16); 1553c9de560dSAlex Tomas 1554c9de560dSAlex Tomas /* we have to split large buddy */ 1555c9de560dSAlex Tomas BUG_ON(ord <= 0); 1556c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1557c9de560dSAlex Tomas mb_set_bit(start >> ord, buddy); 1558c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]--; 1559c9de560dSAlex Tomas 1560c9de560dSAlex Tomas ord--; 1561c9de560dSAlex Tomas cur = (start >> ord) & ~1U; 1562c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, ord, &max); 1563c9de560dSAlex Tomas mb_clear_bit(cur, buddy); 1564c9de560dSAlex Tomas mb_clear_bit(cur + 1, buddy); 1565c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1566c9de560dSAlex Tomas e4b->bd_info->bb_counters[ord]++; 1567c9de560dSAlex Tomas } 15688a57d9d6SCurt Wohlgemuth mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info); 1569c9de560dSAlex Tomas 1570955ce5f5SAneesh Kumar K.V mb_set_bits(EXT4_MB_BITMAP(e4b), ex->fe_start, len0); 1571c9de560dSAlex Tomas mb_check_buddy(e4b); 1572c9de560dSAlex Tomas 1573c9de560dSAlex Tomas return ret; 1574c9de560dSAlex Tomas } 1575c9de560dSAlex Tomas 1576c9de560dSAlex Tomas /* 1577c9de560dSAlex Tomas * Must be called under group lock! 1578c9de560dSAlex Tomas */ 1579c9de560dSAlex Tomas static void ext4_mb_use_best_found(struct ext4_allocation_context *ac, 1580c9de560dSAlex Tomas struct ext4_buddy *e4b) 1581c9de560dSAlex Tomas { 1582c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1583c9de560dSAlex Tomas int ret; 1584c9de560dSAlex Tomas 1585c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group); 1586c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 1587c9de560dSAlex Tomas 1588c9de560dSAlex Tomas ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len); 1589c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical; 1590c9de560dSAlex Tomas ret = mb_mark_used(e4b, &ac->ac_b_ex); 1591c9de560dSAlex Tomas 1592c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 1593c9de560dSAlex Tomas * allocated blocks for history */ 1594c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 1595c9de560dSAlex Tomas 1596c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 1597c9de560dSAlex Tomas ac->ac_tail = ret & 0xffff; 1598c9de560dSAlex Tomas ac->ac_buddy = ret >> 16; 1599c9de560dSAlex Tomas 1600c3a326a6SAneesh Kumar K.V /* 1601c3a326a6SAneesh Kumar K.V * take the page reference. We want the page to be pinned 1602c3a326a6SAneesh Kumar K.V * so that we don't get a ext4_mb_init_cache_call for this 1603c3a326a6SAneesh Kumar K.V * group until we update the bitmap. That would mean we 1604c3a326a6SAneesh Kumar K.V * double allocate blocks. The reference is dropped 1605c3a326a6SAneesh Kumar K.V * in ext4_mb_release_context 1606c3a326a6SAneesh Kumar K.V */ 1607c9de560dSAlex Tomas ac->ac_bitmap_page = e4b->bd_bitmap_page; 1608c9de560dSAlex Tomas get_page(ac->ac_bitmap_page); 1609c9de560dSAlex Tomas ac->ac_buddy_page = e4b->bd_buddy_page; 1610c9de560dSAlex Tomas get_page(ac->ac_buddy_page); 16118556e8f3SAneesh Kumar K.V /* on allocation we use ac to track the held semaphore */ 16128556e8f3SAneesh Kumar K.V ac->alloc_semp = e4b->alloc_semp; 16138556e8f3SAneesh Kumar K.V e4b->alloc_semp = NULL; 1614c9de560dSAlex Tomas /* store last allocated for subsequent stream allocation */ 16154ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 1616c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 1617c9de560dSAlex Tomas sbi->s_mb_last_group = ac->ac_f_ex.fe_group; 1618c9de560dSAlex Tomas sbi->s_mb_last_start = ac->ac_f_ex.fe_start; 1619c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 1620c9de560dSAlex Tomas } 1621c9de560dSAlex Tomas } 1622c9de560dSAlex Tomas 1623c9de560dSAlex Tomas /* 1624c9de560dSAlex Tomas * regular allocator, for general purposes allocation 1625c9de560dSAlex Tomas */ 1626c9de560dSAlex Tomas 1627c9de560dSAlex Tomas static void ext4_mb_check_limits(struct ext4_allocation_context *ac, 1628c9de560dSAlex Tomas struct ext4_buddy *e4b, 1629c9de560dSAlex Tomas int finish_group) 1630c9de560dSAlex Tomas { 1631c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1632c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1633c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1634c9de560dSAlex Tomas struct ext4_free_extent ex; 1635c9de560dSAlex Tomas int max; 1636c9de560dSAlex Tomas 1637032115fcSAneesh Kumar K.V if (ac->ac_status == AC_STATUS_FOUND) 1638032115fcSAneesh Kumar K.V return; 1639c9de560dSAlex Tomas /* 1640c9de560dSAlex Tomas * We don't want to scan for a whole year 1641c9de560dSAlex Tomas */ 1642c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan && 1643c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1644c9de560dSAlex Tomas ac->ac_status = AC_STATUS_BREAK; 1645c9de560dSAlex Tomas return; 1646c9de560dSAlex Tomas } 1647c9de560dSAlex Tomas 1648c9de560dSAlex Tomas /* 1649c9de560dSAlex Tomas * Haven't found good chunk so far, let's continue 1650c9de560dSAlex Tomas */ 1651c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) 1652c9de560dSAlex Tomas return; 1653c9de560dSAlex Tomas 1654c9de560dSAlex Tomas if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan) 1655c9de560dSAlex Tomas && bex->fe_group == e4b->bd_group) { 1656c9de560dSAlex Tomas /* recheck chunk's availability - we don't know 1657c9de560dSAlex Tomas * when it was found (within this lock-unlock 1658c9de560dSAlex Tomas * period or not) */ 1659c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex); 1660c9de560dSAlex Tomas if (max >= gex->fe_len) { 1661c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1662c9de560dSAlex Tomas return; 1663c9de560dSAlex Tomas } 1664c9de560dSAlex Tomas } 1665c9de560dSAlex Tomas } 1666c9de560dSAlex Tomas 1667c9de560dSAlex Tomas /* 1668c9de560dSAlex Tomas * The routine checks whether found extent is good enough. If it is, 1669c9de560dSAlex Tomas * then the extent gets marked used and flag is set to the context 1670c9de560dSAlex Tomas * to stop scanning. Otherwise, the extent is compared with the 1671c9de560dSAlex Tomas * previous found extent and if new one is better, then it's stored 1672c9de560dSAlex Tomas * in the context. Later, the best found extent will be used, if 1673c9de560dSAlex Tomas * mballoc can't find good enough extent. 1674c9de560dSAlex Tomas * 1675c9de560dSAlex Tomas * FIXME: real allocation policy is to be designed yet! 1676c9de560dSAlex Tomas */ 1677c9de560dSAlex Tomas static void ext4_mb_measure_extent(struct ext4_allocation_context *ac, 1678c9de560dSAlex Tomas struct ext4_free_extent *ex, 1679c9de560dSAlex Tomas struct ext4_buddy *e4b) 1680c9de560dSAlex Tomas { 1681c9de560dSAlex Tomas struct ext4_free_extent *bex = &ac->ac_b_ex; 1682c9de560dSAlex Tomas struct ext4_free_extent *gex = &ac->ac_g_ex; 1683c9de560dSAlex Tomas 1684c9de560dSAlex Tomas BUG_ON(ex->fe_len <= 0); 16858d03c7a0SEric Sandeen BUG_ON(ex->fe_len > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 1686c9de560dSAlex Tomas BUG_ON(ex->fe_start >= EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 1687c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_CONTINUE); 1688c9de560dSAlex Tomas 1689c9de560dSAlex Tomas ac->ac_found++; 1690c9de560dSAlex Tomas 1691c9de560dSAlex Tomas /* 1692c9de560dSAlex Tomas * The special case - take what you catch first 1693c9de560dSAlex Tomas */ 1694c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 1695c9de560dSAlex Tomas *bex = *ex; 1696c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1697c9de560dSAlex Tomas return; 1698c9de560dSAlex Tomas } 1699c9de560dSAlex Tomas 1700c9de560dSAlex Tomas /* 1701c9de560dSAlex Tomas * Let's check whether the chuck is good enough 1702c9de560dSAlex Tomas */ 1703c9de560dSAlex Tomas if (ex->fe_len == gex->fe_len) { 1704c9de560dSAlex Tomas *bex = *ex; 1705c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1706c9de560dSAlex Tomas return; 1707c9de560dSAlex Tomas } 1708c9de560dSAlex Tomas 1709c9de560dSAlex Tomas /* 1710c9de560dSAlex Tomas * If this is first found extent, just store it in the context 1711c9de560dSAlex Tomas */ 1712c9de560dSAlex Tomas if (bex->fe_len == 0) { 1713c9de560dSAlex Tomas *bex = *ex; 1714c9de560dSAlex Tomas return; 1715c9de560dSAlex Tomas } 1716c9de560dSAlex Tomas 1717c9de560dSAlex Tomas /* 1718c9de560dSAlex Tomas * If new found extent is better, store it in the context 1719c9de560dSAlex Tomas */ 1720c9de560dSAlex Tomas if (bex->fe_len < gex->fe_len) { 1721c9de560dSAlex Tomas /* if the request isn't satisfied, any found extent 1722c9de560dSAlex Tomas * larger than previous best one is better */ 1723c9de560dSAlex Tomas if (ex->fe_len > bex->fe_len) 1724c9de560dSAlex Tomas *bex = *ex; 1725c9de560dSAlex Tomas } else if (ex->fe_len > gex->fe_len) { 1726c9de560dSAlex Tomas /* if the request is satisfied, then we try to find 1727c9de560dSAlex Tomas * an extent that still satisfy the request, but is 1728c9de560dSAlex Tomas * smaller than previous one */ 1729c9de560dSAlex Tomas if (ex->fe_len < bex->fe_len) 1730c9de560dSAlex Tomas *bex = *ex; 1731c9de560dSAlex Tomas } 1732c9de560dSAlex Tomas 1733c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 0); 1734c9de560dSAlex Tomas } 1735c9de560dSAlex Tomas 1736089ceeccSEric Sandeen static noinline_for_stack 1737089ceeccSEric Sandeen int ext4_mb_try_best_found(struct ext4_allocation_context *ac, 1738c9de560dSAlex Tomas struct ext4_buddy *e4b) 1739c9de560dSAlex Tomas { 1740c9de560dSAlex Tomas struct ext4_free_extent ex = ac->ac_b_ex; 1741c9de560dSAlex Tomas ext4_group_t group = ex.fe_group; 1742c9de560dSAlex Tomas int max; 1743c9de560dSAlex Tomas int err; 1744c9de560dSAlex Tomas 1745c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1746c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1747c9de560dSAlex Tomas if (err) 1748c9de560dSAlex Tomas return err; 1749c9de560dSAlex Tomas 1750c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 1751c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex); 1752c9de560dSAlex Tomas 1753c9de560dSAlex Tomas if (max > 0) { 1754c9de560dSAlex Tomas ac->ac_b_ex = ex; 1755c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1756c9de560dSAlex Tomas } 1757c9de560dSAlex Tomas 1758c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1759e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 1760c9de560dSAlex Tomas 1761c9de560dSAlex Tomas return 0; 1762c9de560dSAlex Tomas } 1763c9de560dSAlex Tomas 1764089ceeccSEric Sandeen static noinline_for_stack 1765089ceeccSEric Sandeen int ext4_mb_find_by_goal(struct ext4_allocation_context *ac, 1766c9de560dSAlex Tomas struct ext4_buddy *e4b) 1767c9de560dSAlex Tomas { 1768c9de560dSAlex Tomas ext4_group_t group = ac->ac_g_ex.fe_group; 1769c9de560dSAlex Tomas int max; 1770c9de560dSAlex Tomas int err; 1771c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 1772c9de560dSAlex Tomas struct ext4_free_extent ex; 1773c9de560dSAlex Tomas 1774c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL)) 1775c9de560dSAlex Tomas return 0; 1776c9de560dSAlex Tomas 1777c9de560dSAlex Tomas err = ext4_mb_load_buddy(ac->ac_sb, group, e4b); 1778c9de560dSAlex Tomas if (err) 1779c9de560dSAlex Tomas return err; 1780c9de560dSAlex Tomas 1781c9de560dSAlex Tomas ext4_lock_group(ac->ac_sb, group); 1782c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start, 1783c9de560dSAlex Tomas ac->ac_g_ex.fe_len, &ex); 1784c9de560dSAlex Tomas 1785c9de560dSAlex Tomas if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) { 1786c9de560dSAlex Tomas ext4_fsblk_t start; 1787c9de560dSAlex Tomas 17885661bd68SAkinobu Mita start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) + 17895661bd68SAkinobu Mita ex.fe_start; 1790c9de560dSAlex Tomas /* use do_div to get remainder (would be 64-bit modulo) */ 1791c9de560dSAlex Tomas if (do_div(start, sbi->s_stripe) == 0) { 1792c9de560dSAlex Tomas ac->ac_found++; 1793c9de560dSAlex Tomas ac->ac_b_ex = ex; 1794c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1795c9de560dSAlex Tomas } 1796c9de560dSAlex Tomas } else if (max >= ac->ac_g_ex.fe_len) { 1797c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1798c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1799c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1800c9de560dSAlex Tomas ac->ac_found++; 1801c9de560dSAlex Tomas ac->ac_b_ex = ex; 1802c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1803c9de560dSAlex Tomas } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) { 1804c9de560dSAlex Tomas /* Sometimes, caller may want to merge even small 1805c9de560dSAlex Tomas * number of blocks to an existing extent */ 1806c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 1807c9de560dSAlex Tomas BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group); 1808c9de560dSAlex Tomas BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start); 1809c9de560dSAlex Tomas ac->ac_found++; 1810c9de560dSAlex Tomas ac->ac_b_ex = ex; 1811c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1812c9de560dSAlex Tomas } 1813c9de560dSAlex Tomas ext4_unlock_group(ac->ac_sb, group); 1814e39e07fdSJing Zhang ext4_mb_unload_buddy(e4b); 1815c9de560dSAlex Tomas 1816c9de560dSAlex Tomas return 0; 1817c9de560dSAlex Tomas } 1818c9de560dSAlex Tomas 1819c9de560dSAlex Tomas /* 1820c9de560dSAlex Tomas * The routine scans buddy structures (not bitmap!) from given order 1821c9de560dSAlex Tomas * to max order and tries to find big enough chunk to satisfy the req 1822c9de560dSAlex Tomas */ 1823089ceeccSEric Sandeen static noinline_for_stack 1824089ceeccSEric Sandeen void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac, 1825c9de560dSAlex Tomas struct ext4_buddy *e4b) 1826c9de560dSAlex Tomas { 1827c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1828c9de560dSAlex Tomas struct ext4_group_info *grp = e4b->bd_info; 1829c9de560dSAlex Tomas void *buddy; 1830c9de560dSAlex Tomas int i; 1831c9de560dSAlex Tomas int k; 1832c9de560dSAlex Tomas int max; 1833c9de560dSAlex Tomas 1834c9de560dSAlex Tomas BUG_ON(ac->ac_2order <= 0); 1835c9de560dSAlex Tomas for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) { 1836c9de560dSAlex Tomas if (grp->bb_counters[i] == 0) 1837c9de560dSAlex Tomas continue; 1838c9de560dSAlex Tomas 1839c9de560dSAlex Tomas buddy = mb_find_buddy(e4b, i, &max); 1840c9de560dSAlex Tomas BUG_ON(buddy == NULL); 1841c9de560dSAlex Tomas 1842ffad0a44SAneesh Kumar K.V k = mb_find_next_zero_bit(buddy, max, 0); 1843c9de560dSAlex Tomas BUG_ON(k >= max); 1844c9de560dSAlex Tomas 1845c9de560dSAlex Tomas ac->ac_found++; 1846c9de560dSAlex Tomas 1847c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 1 << i; 1848c9de560dSAlex Tomas ac->ac_b_ex.fe_start = k << i; 1849c9de560dSAlex Tomas ac->ac_b_ex.fe_group = e4b->bd_group; 1850c9de560dSAlex Tomas 1851c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1852c9de560dSAlex Tomas 1853c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len); 1854c9de560dSAlex Tomas 1855c9de560dSAlex Tomas if (EXT4_SB(sb)->s_mb_stats) 1856c9de560dSAlex Tomas atomic_inc(&EXT4_SB(sb)->s_bal_2orders); 1857c9de560dSAlex Tomas 1858c9de560dSAlex Tomas break; 1859c9de560dSAlex Tomas } 1860c9de560dSAlex Tomas } 1861c9de560dSAlex Tomas 1862c9de560dSAlex Tomas /* 1863c9de560dSAlex Tomas * The routine scans the group and measures all found extents. 1864c9de560dSAlex Tomas * In order to optimize scanning, caller must pass number of 1865c9de560dSAlex Tomas * free blocks in the group, so the routine can know upper limit. 1866c9de560dSAlex Tomas */ 1867089ceeccSEric Sandeen static noinline_for_stack 1868089ceeccSEric Sandeen void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac, 1869c9de560dSAlex Tomas struct ext4_buddy *e4b) 1870c9de560dSAlex Tomas { 1871c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1872c9de560dSAlex Tomas void *bitmap = EXT4_MB_BITMAP(e4b); 1873c9de560dSAlex Tomas struct ext4_free_extent ex; 1874c9de560dSAlex Tomas int i; 1875c9de560dSAlex Tomas int free; 1876c9de560dSAlex Tomas 1877c9de560dSAlex Tomas free = e4b->bd_info->bb_free; 1878c9de560dSAlex Tomas BUG_ON(free <= 0); 1879c9de560dSAlex Tomas 1880c9de560dSAlex Tomas i = e4b->bd_info->bb_first_free; 1881c9de560dSAlex Tomas 1882c9de560dSAlex Tomas while (free && ac->ac_status == AC_STATUS_CONTINUE) { 1883ffad0a44SAneesh Kumar K.V i = mb_find_next_zero_bit(bitmap, 1884c9de560dSAlex Tomas EXT4_BLOCKS_PER_GROUP(sb), i); 1885c9de560dSAlex Tomas if (i >= EXT4_BLOCKS_PER_GROUP(sb)) { 188626346ff6SAneesh Kumar K.V /* 1887e56eb659SAneesh Kumar K.V * IF we have corrupt bitmap, we won't find any 188826346ff6SAneesh Kumar K.V * free blocks even though group info says we 188926346ff6SAneesh Kumar K.V * we have free blocks 189026346ff6SAneesh Kumar K.V */ 1891e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 1892e29136f8STheodore Ts'o "%d free blocks as per " 1893fde4d95aSTheodore Ts'o "group info. But bitmap says 0", 189426346ff6SAneesh Kumar K.V free); 1895c9de560dSAlex Tomas break; 1896c9de560dSAlex Tomas } 1897c9de560dSAlex Tomas 1898c9de560dSAlex Tomas mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex); 1899c9de560dSAlex Tomas BUG_ON(ex.fe_len <= 0); 190026346ff6SAneesh Kumar K.V if (free < ex.fe_len) { 1901e29136f8STheodore Ts'o ext4_grp_locked_error(sb, e4b->bd_group, 0, 0, 1902e29136f8STheodore Ts'o "%d free blocks as per " 1903fde4d95aSTheodore Ts'o "group info. But got %d blocks", 190426346ff6SAneesh Kumar K.V free, ex.fe_len); 1905e56eb659SAneesh Kumar K.V /* 1906e56eb659SAneesh Kumar K.V * The number of free blocks differs. This mostly 1907e56eb659SAneesh Kumar K.V * indicate that the bitmap is corrupt. So exit 1908e56eb659SAneesh Kumar K.V * without claiming the space. 1909e56eb659SAneesh Kumar K.V */ 1910e56eb659SAneesh Kumar K.V break; 191126346ff6SAneesh Kumar K.V } 1912c9de560dSAlex Tomas 1913c9de560dSAlex Tomas ext4_mb_measure_extent(ac, &ex, e4b); 1914c9de560dSAlex Tomas 1915c9de560dSAlex Tomas i += ex.fe_len; 1916c9de560dSAlex Tomas free -= ex.fe_len; 1917c9de560dSAlex Tomas } 1918c9de560dSAlex Tomas 1919c9de560dSAlex Tomas ext4_mb_check_limits(ac, e4b, 1); 1920c9de560dSAlex Tomas } 1921c9de560dSAlex Tomas 1922c9de560dSAlex Tomas /* 1923c9de560dSAlex Tomas * This is a special case for storages like raid5 1924506bf2d8SEric Sandeen * we try to find stripe-aligned chunks for stripe-size-multiple requests 1925c9de560dSAlex Tomas */ 1926089ceeccSEric Sandeen static noinline_for_stack 1927089ceeccSEric Sandeen void ext4_mb_scan_aligned(struct ext4_allocation_context *ac, 1928c9de560dSAlex Tomas struct ext4_buddy *e4b) 1929c9de560dSAlex Tomas { 1930c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 1931c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 1932c9de560dSAlex Tomas void *bitmap = EXT4_MB_BITMAP(e4b); 1933c9de560dSAlex Tomas struct ext4_free_extent ex; 1934c9de560dSAlex Tomas ext4_fsblk_t first_group_block; 1935c9de560dSAlex Tomas ext4_fsblk_t a; 1936c9de560dSAlex Tomas ext4_grpblk_t i; 1937c9de560dSAlex Tomas int max; 1938c9de560dSAlex Tomas 1939c9de560dSAlex Tomas BUG_ON(sbi->s_stripe == 0); 1940c9de560dSAlex Tomas 1941c9de560dSAlex Tomas /* find first stripe-aligned block in group */ 19425661bd68SAkinobu Mita first_group_block = ext4_group_first_block_no(sb, e4b->bd_group); 19435661bd68SAkinobu Mita 1944c9de560dSAlex Tomas a = first_group_block + sbi->s_stripe - 1; 1945c9de560dSAlex Tomas do_div(a, sbi->s_stripe); 1946c9de560dSAlex Tomas i = (a * sbi->s_stripe) - first_group_block; 1947c9de560dSAlex Tomas 1948c9de560dSAlex Tomas while (i < EXT4_BLOCKS_PER_GROUP(sb)) { 1949c9de560dSAlex Tomas if (!mb_test_bit(i, bitmap)) { 1950c9de560dSAlex Tomas max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex); 1951c9de560dSAlex Tomas if (max >= sbi->s_stripe) { 1952c9de560dSAlex Tomas ac->ac_found++; 1953c9de560dSAlex Tomas ac->ac_b_ex = ex; 1954c9de560dSAlex Tomas ext4_mb_use_best_found(ac, e4b); 1955c9de560dSAlex Tomas break; 1956c9de560dSAlex Tomas } 1957c9de560dSAlex Tomas } 1958c9de560dSAlex Tomas i += sbi->s_stripe; 1959c9de560dSAlex Tomas } 1960c9de560dSAlex Tomas } 1961c9de560dSAlex Tomas 19628a57d9d6SCurt Wohlgemuth /* This is now called BEFORE we load the buddy bitmap. */ 1963c9de560dSAlex Tomas static int ext4_mb_good_group(struct ext4_allocation_context *ac, 1964c9de560dSAlex Tomas ext4_group_t group, int cr) 1965c9de560dSAlex Tomas { 1966c9de560dSAlex Tomas unsigned free, fragments; 1967a4912123STheodore Ts'o int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb)); 1968c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group); 1969c9de560dSAlex Tomas 1970c9de560dSAlex Tomas BUG_ON(cr < 0 || cr >= 4); 19718a57d9d6SCurt Wohlgemuth 19728a57d9d6SCurt Wohlgemuth /* We only do this if the grp has never been initialized */ 19738a57d9d6SCurt Wohlgemuth if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) { 19748a57d9d6SCurt Wohlgemuth int ret = ext4_mb_init_group(ac->ac_sb, group); 19758a57d9d6SCurt Wohlgemuth if (ret) 19768a57d9d6SCurt Wohlgemuth return 0; 19778a57d9d6SCurt Wohlgemuth } 1978c9de560dSAlex Tomas 1979c9de560dSAlex Tomas free = grp->bb_free; 1980c9de560dSAlex Tomas fragments = grp->bb_fragments; 1981c9de560dSAlex Tomas if (free == 0) 1982c9de560dSAlex Tomas return 0; 1983c9de560dSAlex Tomas if (fragments == 0) 1984c9de560dSAlex Tomas return 0; 1985c9de560dSAlex Tomas 1986c9de560dSAlex Tomas switch (cr) { 1987c9de560dSAlex Tomas case 0: 1988c9de560dSAlex Tomas BUG_ON(ac->ac_2order == 0); 1989c9de560dSAlex Tomas 19908a57d9d6SCurt Wohlgemuth if (grp->bb_largest_free_order < ac->ac_2order) 19918a57d9d6SCurt Wohlgemuth return 0; 19928a57d9d6SCurt Wohlgemuth 1993a4912123STheodore Ts'o /* Avoid using the first bg of a flexgroup for data files */ 1994a4912123STheodore Ts'o if ((ac->ac_flags & EXT4_MB_HINT_DATA) && 1995a4912123STheodore Ts'o (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) && 1996a4912123STheodore Ts'o ((group % flex_size) == 0)) 1997a4912123STheodore Ts'o return 0; 1998a4912123STheodore Ts'o 1999c9de560dSAlex Tomas return 1; 2000c9de560dSAlex Tomas case 1: 2001c9de560dSAlex Tomas if ((free / fragments) >= ac->ac_g_ex.fe_len) 2002c9de560dSAlex Tomas return 1; 2003c9de560dSAlex Tomas break; 2004c9de560dSAlex Tomas case 2: 2005c9de560dSAlex Tomas if (free >= ac->ac_g_ex.fe_len) 2006c9de560dSAlex Tomas return 1; 2007c9de560dSAlex Tomas break; 2008c9de560dSAlex Tomas case 3: 2009c9de560dSAlex Tomas return 1; 2010c9de560dSAlex Tomas default: 2011c9de560dSAlex Tomas BUG(); 2012c9de560dSAlex Tomas } 2013c9de560dSAlex Tomas 2014c9de560dSAlex Tomas return 0; 2015c9de560dSAlex Tomas } 2016c9de560dSAlex Tomas 20174ddfef7bSEric Sandeen static noinline_for_stack int 20184ddfef7bSEric Sandeen ext4_mb_regular_allocator(struct ext4_allocation_context *ac) 2019c9de560dSAlex Tomas { 20208df9675fSTheodore Ts'o ext4_group_t ngroups, group, i; 2021c9de560dSAlex Tomas int cr; 2022c9de560dSAlex Tomas int err = 0; 2023c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2024c9de560dSAlex Tomas struct super_block *sb; 2025c9de560dSAlex Tomas struct ext4_buddy e4b; 2026c9de560dSAlex Tomas 2027c9de560dSAlex Tomas sb = ac->ac_sb; 2028c9de560dSAlex Tomas sbi = EXT4_SB(sb); 20298df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 2030fb0a387dSEric Sandeen /* non-extent files are limited to low blocks/groups */ 203112e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))) 2032fb0a387dSEric Sandeen ngroups = sbi->s_blockfile_groups; 2033fb0a387dSEric Sandeen 2034c9de560dSAlex Tomas BUG_ON(ac->ac_status == AC_STATUS_FOUND); 2035c9de560dSAlex Tomas 2036c9de560dSAlex Tomas /* first, try the goal */ 2037c9de560dSAlex Tomas err = ext4_mb_find_by_goal(ac, &e4b); 2038c9de560dSAlex Tomas if (err || ac->ac_status == AC_STATUS_FOUND) 2039c9de560dSAlex Tomas goto out; 2040c9de560dSAlex Tomas 2041c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2042c9de560dSAlex Tomas goto out; 2043c9de560dSAlex Tomas 2044c9de560dSAlex Tomas /* 2045c9de560dSAlex Tomas * ac->ac2_order is set only if the fe_len is a power of 2 2046c9de560dSAlex Tomas * if ac2_order is set we also set criteria to 0 so that we 2047c9de560dSAlex Tomas * try exact allocation using buddy. 2048c9de560dSAlex Tomas */ 2049c9de560dSAlex Tomas i = fls(ac->ac_g_ex.fe_len); 2050c9de560dSAlex Tomas ac->ac_2order = 0; 2051c9de560dSAlex Tomas /* 2052c9de560dSAlex Tomas * We search using buddy data only if the order of the request 2053c9de560dSAlex Tomas * is greater than equal to the sbi_s_mb_order2_reqs 2054b713a5ecSTheodore Ts'o * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req 2055c9de560dSAlex Tomas */ 2056c9de560dSAlex Tomas if (i >= sbi->s_mb_order2_reqs) { 2057c9de560dSAlex Tomas /* 2058c9de560dSAlex Tomas * This should tell if fe_len is exactly power of 2 2059c9de560dSAlex Tomas */ 2060c9de560dSAlex Tomas if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0) 2061c9de560dSAlex Tomas ac->ac_2order = i - 1; 2062c9de560dSAlex Tomas } 2063c9de560dSAlex Tomas 20644ba74d00STheodore Ts'o /* if stream allocation is enabled, use global goal */ 20654ba74d00STheodore Ts'o if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) { 2066c9de560dSAlex Tomas /* TBD: may be hot point */ 2067c9de560dSAlex Tomas spin_lock(&sbi->s_md_lock); 2068c9de560dSAlex Tomas ac->ac_g_ex.fe_group = sbi->s_mb_last_group; 2069c9de560dSAlex Tomas ac->ac_g_ex.fe_start = sbi->s_mb_last_start; 2070c9de560dSAlex Tomas spin_unlock(&sbi->s_md_lock); 2071c9de560dSAlex Tomas } 20724ba74d00STheodore Ts'o 2073c9de560dSAlex Tomas /* Let's just scan groups to find more-less suitable blocks */ 2074c9de560dSAlex Tomas cr = ac->ac_2order ? 0 : 1; 2075c9de560dSAlex Tomas /* 2076c9de560dSAlex Tomas * cr == 0 try to get exact allocation, 2077c9de560dSAlex Tomas * cr == 3 try to get anything 2078c9de560dSAlex Tomas */ 2079c9de560dSAlex Tomas repeat: 2080c9de560dSAlex Tomas for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) { 2081c9de560dSAlex Tomas ac->ac_criteria = cr; 2082ed8f9c75SAneesh Kumar K.V /* 2083ed8f9c75SAneesh Kumar K.V * searching for the right group start 2084ed8f9c75SAneesh Kumar K.V * from the goal value specified 2085ed8f9c75SAneesh Kumar K.V */ 2086ed8f9c75SAneesh Kumar K.V group = ac->ac_g_ex.fe_group; 2087ed8f9c75SAneesh Kumar K.V 20888df9675fSTheodore Ts'o for (i = 0; i < ngroups; group++, i++) { 20898df9675fSTheodore Ts'o if (group == ngroups) 2090c9de560dSAlex Tomas group = 0; 2091c9de560dSAlex Tomas 20928a57d9d6SCurt Wohlgemuth /* This now checks without needing the buddy page */ 20938a57d9d6SCurt Wohlgemuth if (!ext4_mb_good_group(ac, group, cr)) 2094c9de560dSAlex Tomas continue; 2095c9de560dSAlex Tomas 2096c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2097c9de560dSAlex Tomas if (err) 2098c9de560dSAlex Tomas goto out; 2099c9de560dSAlex Tomas 2100c9de560dSAlex Tomas ext4_lock_group(sb, group); 21018a57d9d6SCurt Wohlgemuth 21028a57d9d6SCurt Wohlgemuth /* 21038a57d9d6SCurt Wohlgemuth * We need to check again after locking the 21048a57d9d6SCurt Wohlgemuth * block group 21058a57d9d6SCurt Wohlgemuth */ 2106c9de560dSAlex Tomas if (!ext4_mb_good_group(ac, group, cr)) { 2107c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2108e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2109c9de560dSAlex Tomas continue; 2110c9de560dSAlex Tomas } 2111c9de560dSAlex Tomas 2112c9de560dSAlex Tomas ac->ac_groups_scanned++; 211375507efbSTheodore Ts'o if (cr == 0) 2114c9de560dSAlex Tomas ext4_mb_simple_scan_group(ac, &e4b); 2115506bf2d8SEric Sandeen else if (cr == 1 && sbi->s_stripe && 2116506bf2d8SEric Sandeen !(ac->ac_g_ex.fe_len % sbi->s_stripe)) 2117c9de560dSAlex Tomas ext4_mb_scan_aligned(ac, &e4b); 2118c9de560dSAlex Tomas else 2119c9de560dSAlex Tomas ext4_mb_complex_scan_group(ac, &e4b); 2120c9de560dSAlex Tomas 2121c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2122e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2123c9de560dSAlex Tomas 2124c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_CONTINUE) 2125c9de560dSAlex Tomas break; 2126c9de560dSAlex Tomas } 2127c9de560dSAlex Tomas } 2128c9de560dSAlex Tomas 2129c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND && 2130c9de560dSAlex Tomas !(ac->ac_flags & EXT4_MB_HINT_FIRST)) { 2131c9de560dSAlex Tomas /* 2132c9de560dSAlex Tomas * We've been searching too long. Let's try to allocate 2133c9de560dSAlex Tomas * the best chunk we've found so far 2134c9de560dSAlex Tomas */ 2135c9de560dSAlex Tomas 2136c9de560dSAlex Tomas ext4_mb_try_best_found(ac, &e4b); 2137c9de560dSAlex Tomas if (ac->ac_status != AC_STATUS_FOUND) { 2138c9de560dSAlex Tomas /* 2139c9de560dSAlex Tomas * Someone more lucky has already allocated it. 2140c9de560dSAlex Tomas * The only thing we can do is just take first 2141c9de560dSAlex Tomas * found block(s) 2142c9de560dSAlex Tomas printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n"); 2143c9de560dSAlex Tomas */ 2144c9de560dSAlex Tomas ac->ac_b_ex.fe_group = 0; 2145c9de560dSAlex Tomas ac->ac_b_ex.fe_start = 0; 2146c9de560dSAlex Tomas ac->ac_b_ex.fe_len = 0; 2147c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 2148c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_FIRST; 2149c9de560dSAlex Tomas cr = 3; 2150c9de560dSAlex Tomas atomic_inc(&sbi->s_mb_lost_chunks); 2151c9de560dSAlex Tomas goto repeat; 2152c9de560dSAlex Tomas } 2153c9de560dSAlex Tomas } 2154c9de560dSAlex Tomas out: 2155c9de560dSAlex Tomas return err; 2156c9de560dSAlex Tomas } 2157c9de560dSAlex Tomas 2158c9de560dSAlex Tomas static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos) 2159c9de560dSAlex Tomas { 2160c9de560dSAlex Tomas struct super_block *sb = seq->private; 2161c9de560dSAlex Tomas ext4_group_t group; 2162c9de560dSAlex Tomas 21638df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2164c9de560dSAlex Tomas return NULL; 2165c9de560dSAlex Tomas group = *pos + 1; 2166a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2167c9de560dSAlex Tomas } 2168c9de560dSAlex Tomas 2169c9de560dSAlex Tomas static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos) 2170c9de560dSAlex Tomas { 2171c9de560dSAlex Tomas struct super_block *sb = seq->private; 2172c9de560dSAlex Tomas ext4_group_t group; 2173c9de560dSAlex Tomas 2174c9de560dSAlex Tomas ++*pos; 21758df9675fSTheodore Ts'o if (*pos < 0 || *pos >= ext4_get_groups_count(sb)) 2176c9de560dSAlex Tomas return NULL; 2177c9de560dSAlex Tomas group = *pos + 1; 2178a9df9a49STheodore Ts'o return (void *) ((unsigned long) group); 2179c9de560dSAlex Tomas } 2180c9de560dSAlex Tomas 2181c9de560dSAlex Tomas static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v) 2182c9de560dSAlex Tomas { 2183c9de560dSAlex Tomas struct super_block *sb = seq->private; 2184a9df9a49STheodore Ts'o ext4_group_t group = (ext4_group_t) ((unsigned long) v); 2185c9de560dSAlex Tomas int i; 2186c9de560dSAlex Tomas int err; 2187c9de560dSAlex Tomas struct ext4_buddy e4b; 2188c9de560dSAlex Tomas struct sg { 2189c9de560dSAlex Tomas struct ext4_group_info info; 2190a36b4498SEric Sandeen ext4_grpblk_t counters[16]; 2191c9de560dSAlex Tomas } sg; 2192c9de560dSAlex Tomas 2193c9de560dSAlex Tomas group--; 2194c9de560dSAlex Tomas if (group == 0) 2195c9de560dSAlex Tomas seq_printf(seq, "#%-5s: %-5s %-5s %-5s " 2196c9de560dSAlex Tomas "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s " 2197c9de560dSAlex Tomas "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n", 2198c9de560dSAlex Tomas "group", "free", "frags", "first", 2199c9de560dSAlex Tomas "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6", 2200c9de560dSAlex Tomas "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13"); 2201c9de560dSAlex Tomas 2202c9de560dSAlex Tomas i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) + 2203c9de560dSAlex Tomas sizeof(struct ext4_group_info); 2204c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 2205c9de560dSAlex Tomas if (err) { 2206a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: I/O error\n", group); 2207c9de560dSAlex Tomas return 0; 2208c9de560dSAlex Tomas } 2209c9de560dSAlex Tomas ext4_lock_group(sb, group); 2210c9de560dSAlex Tomas memcpy(&sg, ext4_get_group_info(sb, group), i); 2211c9de560dSAlex Tomas ext4_unlock_group(sb, group); 2212e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 2213c9de560dSAlex Tomas 2214a9df9a49STheodore Ts'o seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free, 2215c9de560dSAlex Tomas sg.info.bb_fragments, sg.info.bb_first_free); 2216c9de560dSAlex Tomas for (i = 0; i <= 13; i++) 2217c9de560dSAlex Tomas seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ? 2218c9de560dSAlex Tomas sg.info.bb_counters[i] : 0); 2219c9de560dSAlex Tomas seq_printf(seq, " ]\n"); 2220c9de560dSAlex Tomas 2221c9de560dSAlex Tomas return 0; 2222c9de560dSAlex Tomas } 2223c9de560dSAlex Tomas 2224c9de560dSAlex Tomas static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v) 2225c9de560dSAlex Tomas { 2226c9de560dSAlex Tomas } 2227c9de560dSAlex Tomas 22287f1346a9STobias Klauser static const struct seq_operations ext4_mb_seq_groups_ops = { 2229c9de560dSAlex Tomas .start = ext4_mb_seq_groups_start, 2230c9de560dSAlex Tomas .next = ext4_mb_seq_groups_next, 2231c9de560dSAlex Tomas .stop = ext4_mb_seq_groups_stop, 2232c9de560dSAlex Tomas .show = ext4_mb_seq_groups_show, 2233c9de560dSAlex Tomas }; 2234c9de560dSAlex Tomas 2235c9de560dSAlex Tomas static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file) 2236c9de560dSAlex Tomas { 2237c9de560dSAlex Tomas struct super_block *sb = PDE(inode)->data; 2238c9de560dSAlex Tomas int rc; 2239c9de560dSAlex Tomas 2240c9de560dSAlex Tomas rc = seq_open(file, &ext4_mb_seq_groups_ops); 2241c9de560dSAlex Tomas if (rc == 0) { 2242a271fe85SJoe Perches struct seq_file *m = file->private_data; 2243c9de560dSAlex Tomas m->private = sb; 2244c9de560dSAlex Tomas } 2245c9de560dSAlex Tomas return rc; 2246c9de560dSAlex Tomas 2247c9de560dSAlex Tomas } 2248c9de560dSAlex Tomas 22497f1346a9STobias Klauser static const struct file_operations ext4_mb_seq_groups_fops = { 2250c9de560dSAlex Tomas .owner = THIS_MODULE, 2251c9de560dSAlex Tomas .open = ext4_mb_seq_groups_open, 2252c9de560dSAlex Tomas .read = seq_read, 2253c9de560dSAlex Tomas .llseek = seq_lseek, 2254c9de560dSAlex Tomas .release = seq_release, 2255c9de560dSAlex Tomas }; 2256c9de560dSAlex Tomas 2257fb1813f4SCurt Wohlgemuth static struct kmem_cache *get_groupinfo_cache(int blocksize_bits) 2258fb1813f4SCurt Wohlgemuth { 2259fb1813f4SCurt Wohlgemuth int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 2260fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index]; 2261fb1813f4SCurt Wohlgemuth 2262fb1813f4SCurt Wohlgemuth BUG_ON(!cachep); 2263fb1813f4SCurt Wohlgemuth return cachep; 2264fb1813f4SCurt Wohlgemuth } 22655f21b0e6SFrederic Bohe 22665f21b0e6SFrederic Bohe /* Create and initialize ext4_group_info data for the given group. */ 2267920313a7SAneesh Kumar K.V int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, 22685f21b0e6SFrederic Bohe struct ext4_group_desc *desc) 22695f21b0e6SFrederic Bohe { 2270fb1813f4SCurt Wohlgemuth int i; 22715f21b0e6SFrederic Bohe int metalen = 0; 22725f21b0e6SFrederic Bohe struct ext4_sb_info *sbi = EXT4_SB(sb); 22735f21b0e6SFrederic Bohe struct ext4_group_info **meta_group_info; 2274fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 22755f21b0e6SFrederic Bohe 22765f21b0e6SFrederic Bohe /* 22775f21b0e6SFrederic Bohe * First check if this group is the first of a reserved block. 22785f21b0e6SFrederic Bohe * If it's true, we have to allocate a new table of pointers 22795f21b0e6SFrederic Bohe * to ext4_group_info structures 22805f21b0e6SFrederic Bohe */ 22815f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) { 22825f21b0e6SFrederic Bohe metalen = sizeof(*meta_group_info) << 22835f21b0e6SFrederic Bohe EXT4_DESC_PER_BLOCK_BITS(sb); 22845f21b0e6SFrederic Bohe meta_group_info = kmalloc(metalen, GFP_KERNEL); 22855f21b0e6SFrederic Bohe if (meta_group_info == NULL) { 22865f21b0e6SFrederic Bohe printk(KERN_ERR "EXT4-fs: can't allocate mem for a " 22875f21b0e6SFrederic Bohe "buddy group\n"); 22885f21b0e6SFrederic Bohe goto exit_meta_group_info; 22895f21b0e6SFrederic Bohe } 22905f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = 22915f21b0e6SFrederic Bohe meta_group_info; 22925f21b0e6SFrederic Bohe } 22935f21b0e6SFrederic Bohe 22945f21b0e6SFrederic Bohe meta_group_info = 22955f21b0e6SFrederic Bohe sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]; 22965f21b0e6SFrederic Bohe i = group & (EXT4_DESC_PER_BLOCK(sb) - 1); 22975f21b0e6SFrederic Bohe 2298fb1813f4SCurt Wohlgemuth meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL); 22995f21b0e6SFrederic Bohe if (meta_group_info[i] == NULL) { 23005f21b0e6SFrederic Bohe printk(KERN_ERR "EXT4-fs: can't allocate buddy mem\n"); 23015f21b0e6SFrederic Bohe goto exit_group_info; 23025f21b0e6SFrederic Bohe } 2303fb1813f4SCurt Wohlgemuth memset(meta_group_info[i], 0, kmem_cache_size(cachep)); 23045f21b0e6SFrederic Bohe set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, 23055f21b0e6SFrederic Bohe &(meta_group_info[i]->bb_state)); 23065f21b0e6SFrederic Bohe 23075f21b0e6SFrederic Bohe /* 23085f21b0e6SFrederic Bohe * initialize bb_free to be able to skip 23095f21b0e6SFrederic Bohe * empty groups without initialization 23105f21b0e6SFrederic Bohe */ 23115f21b0e6SFrederic Bohe if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 23125f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 23135f21b0e6SFrederic Bohe ext4_free_blocks_after_init(sb, group, desc); 23145f21b0e6SFrederic Bohe } else { 23155f21b0e6SFrederic Bohe meta_group_info[i]->bb_free = 2316560671a0SAneesh Kumar K.V ext4_free_blks_count(sb, desc); 23175f21b0e6SFrederic Bohe } 23185f21b0e6SFrederic Bohe 23195f21b0e6SFrederic Bohe INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); 2320920313a7SAneesh Kumar K.V init_rwsem(&meta_group_info[i]->alloc_sem); 232164e290ecSVenkatesh Pallipadi meta_group_info[i]->bb_free_root = RB_ROOT; 23228a57d9d6SCurt Wohlgemuth meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ 23235f21b0e6SFrederic Bohe 23245f21b0e6SFrederic Bohe #ifdef DOUBLE_CHECK 23255f21b0e6SFrederic Bohe { 23265f21b0e6SFrederic Bohe struct buffer_head *bh; 23275f21b0e6SFrederic Bohe meta_group_info[i]->bb_bitmap = 23285f21b0e6SFrederic Bohe kmalloc(sb->s_blocksize, GFP_KERNEL); 23295f21b0e6SFrederic Bohe BUG_ON(meta_group_info[i]->bb_bitmap == NULL); 23305f21b0e6SFrederic Bohe bh = ext4_read_block_bitmap(sb, group); 23315f21b0e6SFrederic Bohe BUG_ON(bh == NULL); 23325f21b0e6SFrederic Bohe memcpy(meta_group_info[i]->bb_bitmap, bh->b_data, 23335f21b0e6SFrederic Bohe sb->s_blocksize); 23345f21b0e6SFrederic Bohe put_bh(bh); 23355f21b0e6SFrederic Bohe } 23365f21b0e6SFrederic Bohe #endif 23375f21b0e6SFrederic Bohe 23385f21b0e6SFrederic Bohe return 0; 23395f21b0e6SFrederic Bohe 23405f21b0e6SFrederic Bohe exit_group_info: 23415f21b0e6SFrederic Bohe /* If a meta_group_info table has been allocated, release it now */ 23425f21b0e6SFrederic Bohe if (group % EXT4_DESC_PER_BLOCK(sb) == 0) 23435f21b0e6SFrederic Bohe kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]); 23445f21b0e6SFrederic Bohe exit_meta_group_info: 23455f21b0e6SFrederic Bohe return -ENOMEM; 23465f21b0e6SFrederic Bohe } /* ext4_mb_add_groupinfo */ 23475f21b0e6SFrederic Bohe 2348c9de560dSAlex Tomas static int ext4_mb_init_backend(struct super_block *sb) 2349c9de560dSAlex Tomas { 23508df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 2351c9de560dSAlex Tomas ext4_group_t i; 2352c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 23535f21b0e6SFrederic Bohe struct ext4_super_block *es = sbi->s_es; 23545f21b0e6SFrederic Bohe int num_meta_group_infos; 23555f21b0e6SFrederic Bohe int num_meta_group_infos_max; 23565f21b0e6SFrederic Bohe int array_size; 23575f21b0e6SFrederic Bohe struct ext4_group_desc *desc; 2358fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep; 2359c9de560dSAlex Tomas 23605f21b0e6SFrederic Bohe /* This is the number of blocks used by GDT */ 23618df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 23625f21b0e6SFrederic Bohe 1) >> EXT4_DESC_PER_BLOCK_BITS(sb); 23635f21b0e6SFrederic Bohe 23645f21b0e6SFrederic Bohe /* 23655f21b0e6SFrederic Bohe * This is the total number of blocks used by GDT including 23665f21b0e6SFrederic Bohe * the number of reserved blocks for GDT. 23675f21b0e6SFrederic Bohe * The s_group_info array is allocated with this value 23685f21b0e6SFrederic Bohe * to allow a clean online resize without a complex 23695f21b0e6SFrederic Bohe * manipulation of pointer. 23705f21b0e6SFrederic Bohe * The drawback is the unused memory when no resize 23715f21b0e6SFrederic Bohe * occurs but it's very low in terms of pages 23725f21b0e6SFrederic Bohe * (see comments below) 23735f21b0e6SFrederic Bohe * Need to handle this properly when META_BG resizing is allowed 23745f21b0e6SFrederic Bohe */ 23755f21b0e6SFrederic Bohe num_meta_group_infos_max = num_meta_group_infos + 23765f21b0e6SFrederic Bohe le16_to_cpu(es->s_reserved_gdt_blocks); 23775f21b0e6SFrederic Bohe 23785f21b0e6SFrederic Bohe /* 23795f21b0e6SFrederic Bohe * array_size is the size of s_group_info array. We round it 23805f21b0e6SFrederic Bohe * to the next power of two because this approximation is done 23815f21b0e6SFrederic Bohe * internally by kmalloc so we can have some more memory 23825f21b0e6SFrederic Bohe * for free here (e.g. may be used for META_BG resize). 23835f21b0e6SFrederic Bohe */ 23845f21b0e6SFrederic Bohe array_size = 1; 23855f21b0e6SFrederic Bohe while (array_size < sizeof(*sbi->s_group_info) * 23865f21b0e6SFrederic Bohe num_meta_group_infos_max) 23875f21b0e6SFrederic Bohe array_size = array_size << 1; 2388c9de560dSAlex Tomas /* An 8TB filesystem with 64-bit pointers requires a 4096 byte 2389c9de560dSAlex Tomas * kmalloc. A 128kb malloc should suffice for a 256TB filesystem. 2390c9de560dSAlex Tomas * So a two level scheme suffices for now. */ 23914596fe07SEric Sandeen sbi->s_group_info = kzalloc(array_size, GFP_KERNEL); 2392c9de560dSAlex Tomas if (sbi->s_group_info == NULL) { 2393c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: can't allocate buddy meta group\n"); 2394c9de560dSAlex Tomas return -ENOMEM; 2395c9de560dSAlex Tomas } 2396c9de560dSAlex Tomas sbi->s_buddy_cache = new_inode(sb); 2397c9de560dSAlex Tomas if (sbi->s_buddy_cache == NULL) { 2398c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: can't get new inode\n"); 2399c9de560dSAlex Tomas goto err_freesgi; 2400c9de560dSAlex Tomas } 240185fe4025SChristoph Hellwig sbi->s_buddy_cache->i_ino = get_next_ino(); 2402c9de560dSAlex Tomas EXT4_I(sbi->s_buddy_cache)->i_disksize = 0; 24038df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 2404c9de560dSAlex Tomas desc = ext4_get_group_desc(sb, i, NULL); 2405c9de560dSAlex Tomas if (desc == NULL) { 2406c9de560dSAlex Tomas printk(KERN_ERR 2407a9df9a49STheodore Ts'o "EXT4-fs: can't read descriptor %u\n", i); 2408c9de560dSAlex Tomas goto err_freebuddy; 2409c9de560dSAlex Tomas } 24105f21b0e6SFrederic Bohe if (ext4_mb_add_groupinfo(sb, i, desc) != 0) 24115f21b0e6SFrederic Bohe goto err_freebuddy; 2412c9de560dSAlex Tomas } 2413c9de560dSAlex Tomas 2414c9de560dSAlex Tomas return 0; 2415c9de560dSAlex Tomas 2416c9de560dSAlex Tomas err_freebuddy: 2417fb1813f4SCurt Wohlgemuth cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2418f1fa3342SRoel Kluin while (i-- > 0) 2419fb1813f4SCurt Wohlgemuth kmem_cache_free(cachep, ext4_get_group_info(sb, i)); 2420c9de560dSAlex Tomas i = num_meta_group_infos; 2421f1fa3342SRoel Kluin while (i-- > 0) 2422c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2423c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2424c9de560dSAlex Tomas err_freesgi: 2425c9de560dSAlex Tomas kfree(sbi->s_group_info); 2426c9de560dSAlex Tomas return -ENOMEM; 2427c9de560dSAlex Tomas } 2428c9de560dSAlex Tomas 24292892c15dSEric Sandeen static void ext4_groupinfo_destroy_slabs(void) 24302892c15dSEric Sandeen { 24312892c15dSEric Sandeen int i; 24322892c15dSEric Sandeen 24332892c15dSEric Sandeen for (i = 0; i < NR_GRPINFO_CACHES; i++) { 24342892c15dSEric Sandeen if (ext4_groupinfo_caches[i]) 24352892c15dSEric Sandeen kmem_cache_destroy(ext4_groupinfo_caches[i]); 24362892c15dSEric Sandeen ext4_groupinfo_caches[i] = NULL; 24372892c15dSEric Sandeen } 24382892c15dSEric Sandeen } 24392892c15dSEric Sandeen 24402892c15dSEric Sandeen static int ext4_groupinfo_create_slab(size_t size) 24412892c15dSEric Sandeen { 24422892c15dSEric Sandeen static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex); 24432892c15dSEric Sandeen int slab_size; 24442892c15dSEric Sandeen int blocksize_bits = order_base_2(size); 24452892c15dSEric Sandeen int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE; 24462892c15dSEric Sandeen struct kmem_cache *cachep; 24472892c15dSEric Sandeen 24482892c15dSEric Sandeen if (cache_index >= NR_GRPINFO_CACHES) 24492892c15dSEric Sandeen return -EINVAL; 24502892c15dSEric Sandeen 24512892c15dSEric Sandeen if (unlikely(cache_index < 0)) 24522892c15dSEric Sandeen cache_index = 0; 24532892c15dSEric Sandeen 24542892c15dSEric Sandeen mutex_lock(&ext4_grpinfo_slab_create_mutex); 24552892c15dSEric Sandeen if (ext4_groupinfo_caches[cache_index]) { 24562892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 24572892c15dSEric Sandeen return 0; /* Already created */ 24582892c15dSEric Sandeen } 24592892c15dSEric Sandeen 24602892c15dSEric Sandeen slab_size = offsetof(struct ext4_group_info, 24612892c15dSEric Sandeen bb_counters[blocksize_bits + 2]); 24622892c15dSEric Sandeen 24632892c15dSEric Sandeen cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index], 24642892c15dSEric Sandeen slab_size, 0, SLAB_RECLAIM_ACCOUNT, 24652892c15dSEric Sandeen NULL); 24662892c15dSEric Sandeen 24672892c15dSEric Sandeen mutex_unlock(&ext4_grpinfo_slab_create_mutex); 24682892c15dSEric Sandeen if (!cachep) { 24692892c15dSEric Sandeen printk(KERN_EMERG "EXT4: no memory for groupinfo slab cache\n"); 24702892c15dSEric Sandeen return -ENOMEM; 24712892c15dSEric Sandeen } 24722892c15dSEric Sandeen 24732892c15dSEric Sandeen ext4_groupinfo_caches[cache_index] = cachep; 24742892c15dSEric Sandeen 24752892c15dSEric Sandeen return 0; 24762892c15dSEric Sandeen } 24772892c15dSEric Sandeen 2478c9de560dSAlex Tomas int ext4_mb_init(struct super_block *sb, int needs_recovery) 2479c9de560dSAlex Tomas { 2480c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 24816be2ded1SAneesh Kumar K.V unsigned i, j; 2482c9de560dSAlex Tomas unsigned offset; 2483c9de560dSAlex Tomas unsigned max; 248474767c5aSShen Feng int ret; 2485c9de560dSAlex Tomas 24861927805eSEric Sandeen i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets); 2487c9de560dSAlex Tomas 2488c9de560dSAlex Tomas sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL); 2489c9de560dSAlex Tomas if (sbi->s_mb_offsets == NULL) { 2490fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 2491fb1813f4SCurt Wohlgemuth goto out; 2492c9de560dSAlex Tomas } 2493ff7ef329SYasunori Goto 24941927805eSEric Sandeen i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs); 2495c9de560dSAlex Tomas sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL); 2496c9de560dSAlex Tomas if (sbi->s_mb_maxs == NULL) { 2497fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 2498fb1813f4SCurt Wohlgemuth goto out; 2499fb1813f4SCurt Wohlgemuth } 2500fb1813f4SCurt Wohlgemuth 25012892c15dSEric Sandeen ret = ext4_groupinfo_create_slab(sb->s_blocksize); 25022892c15dSEric Sandeen if (ret < 0) 2503fb1813f4SCurt Wohlgemuth goto out; 2504c9de560dSAlex Tomas 2505c9de560dSAlex Tomas /* order 0 is regular bitmap */ 2506c9de560dSAlex Tomas sbi->s_mb_maxs[0] = sb->s_blocksize << 3; 2507c9de560dSAlex Tomas sbi->s_mb_offsets[0] = 0; 2508c9de560dSAlex Tomas 2509c9de560dSAlex Tomas i = 1; 2510c9de560dSAlex Tomas offset = 0; 2511c9de560dSAlex Tomas max = sb->s_blocksize << 2; 2512c9de560dSAlex Tomas do { 2513c9de560dSAlex Tomas sbi->s_mb_offsets[i] = offset; 2514c9de560dSAlex Tomas sbi->s_mb_maxs[i] = max; 2515c9de560dSAlex Tomas offset += 1 << (sb->s_blocksize_bits - i); 2516c9de560dSAlex Tomas max = max >> 1; 2517c9de560dSAlex Tomas i++; 2518c9de560dSAlex Tomas } while (i <= sb->s_blocksize_bits + 1); 2519c9de560dSAlex Tomas 2520c9de560dSAlex Tomas /* init file for buddy data */ 252174767c5aSShen Feng ret = ext4_mb_init_backend(sb); 252274767c5aSShen Feng if (ret != 0) { 2523fb1813f4SCurt Wohlgemuth goto out; 2524c9de560dSAlex Tomas } 2525c9de560dSAlex Tomas 2526c9de560dSAlex Tomas spin_lock_init(&sbi->s_md_lock); 2527c9de560dSAlex Tomas spin_lock_init(&sbi->s_bal_lock); 2528c9de560dSAlex Tomas 2529c9de560dSAlex Tomas sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN; 2530c9de560dSAlex Tomas sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN; 2531c9de560dSAlex Tomas sbi->s_mb_stats = MB_DEFAULT_STATS; 2532c9de560dSAlex Tomas sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD; 2533c9de560dSAlex Tomas sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS; 2534c9de560dSAlex Tomas sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC; 2535c9de560dSAlex Tomas 2536730c213cSEric Sandeen sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group); 2537c9de560dSAlex Tomas if (sbi->s_locality_groups == NULL) { 2538fb1813f4SCurt Wohlgemuth ret = -ENOMEM; 2539fb1813f4SCurt Wohlgemuth goto out; 2540c9de560dSAlex Tomas } 2541730c213cSEric Sandeen for_each_possible_cpu(i) { 2542c9de560dSAlex Tomas struct ext4_locality_group *lg; 2543730c213cSEric Sandeen lg = per_cpu_ptr(sbi->s_locality_groups, i); 2544c9de560dSAlex Tomas mutex_init(&lg->lg_mutex); 25456be2ded1SAneesh Kumar K.V for (j = 0; j < PREALLOC_TB_SIZE; j++) 25466be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&lg->lg_prealloc_list[j]); 2547c9de560dSAlex Tomas spin_lock_init(&lg->lg_prealloc_lock); 2548c9de560dSAlex Tomas } 2549c9de560dSAlex Tomas 2550296c355cSTheodore Ts'o if (sbi->s_proc) 2551296c355cSTheodore Ts'o proc_create_data("mb_groups", S_IRUGO, sbi->s_proc, 2552296c355cSTheodore Ts'o &ext4_mb_seq_groups_fops, sb); 2553c9de560dSAlex Tomas 25540390131bSFrank Mayhar if (sbi->s_journal) 25553e624fc7STheodore Ts'o sbi->s_journal->j_commit_callback = release_blocks_on_commit; 2556fb1813f4SCurt Wohlgemuth out: 2557fb1813f4SCurt Wohlgemuth if (ret) { 2558fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_offsets); 2559fb1813f4SCurt Wohlgemuth kfree(sbi->s_mb_maxs); 2560fb1813f4SCurt Wohlgemuth } 2561fb1813f4SCurt Wohlgemuth return ret; 2562c9de560dSAlex Tomas } 2563c9de560dSAlex Tomas 2564955ce5f5SAneesh Kumar K.V /* need to called with the ext4 group lock held */ 2565c9de560dSAlex Tomas static void ext4_mb_cleanup_pa(struct ext4_group_info *grp) 2566c9de560dSAlex Tomas { 2567c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 2568c9de560dSAlex Tomas struct list_head *cur, *tmp; 2569c9de560dSAlex Tomas int count = 0; 2570c9de560dSAlex Tomas 2571c9de560dSAlex Tomas list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) { 2572c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 2573c9de560dSAlex Tomas list_del(&pa->pa_group_list); 2574c9de560dSAlex Tomas count++; 2575688f05a0SAneesh Kumar K.V kmem_cache_free(ext4_pspace_cachep, pa); 2576c9de560dSAlex Tomas } 2577c9de560dSAlex Tomas if (count) 25786ba495e9STheodore Ts'o mb_debug(1, "mballoc: %u PAs left\n", count); 2579c9de560dSAlex Tomas 2580c9de560dSAlex Tomas } 2581c9de560dSAlex Tomas 2582c9de560dSAlex Tomas int ext4_mb_release(struct super_block *sb) 2583c9de560dSAlex Tomas { 25848df9675fSTheodore Ts'o ext4_group_t ngroups = ext4_get_groups_count(sb); 2585c9de560dSAlex Tomas ext4_group_t i; 2586c9de560dSAlex Tomas int num_meta_group_infos; 2587c9de560dSAlex Tomas struct ext4_group_info *grinfo; 2588c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 2589fb1813f4SCurt Wohlgemuth struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits); 2590c9de560dSAlex Tomas 2591c9de560dSAlex Tomas if (sbi->s_group_info) { 25928df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 2593c9de560dSAlex Tomas grinfo = ext4_get_group_info(sb, i); 2594c9de560dSAlex Tomas #ifdef DOUBLE_CHECK 2595c9de560dSAlex Tomas kfree(grinfo->bb_bitmap); 2596c9de560dSAlex Tomas #endif 2597c9de560dSAlex Tomas ext4_lock_group(sb, i); 2598c9de560dSAlex Tomas ext4_mb_cleanup_pa(grinfo); 2599c9de560dSAlex Tomas ext4_unlock_group(sb, i); 2600fb1813f4SCurt Wohlgemuth kmem_cache_free(cachep, grinfo); 2601c9de560dSAlex Tomas } 26028df9675fSTheodore Ts'o num_meta_group_infos = (ngroups + 2603c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK(sb) - 1) >> 2604c9de560dSAlex Tomas EXT4_DESC_PER_BLOCK_BITS(sb); 2605c9de560dSAlex Tomas for (i = 0; i < num_meta_group_infos; i++) 2606c9de560dSAlex Tomas kfree(sbi->s_group_info[i]); 2607c9de560dSAlex Tomas kfree(sbi->s_group_info); 2608c9de560dSAlex Tomas } 2609c9de560dSAlex Tomas kfree(sbi->s_mb_offsets); 2610c9de560dSAlex Tomas kfree(sbi->s_mb_maxs); 2611c9de560dSAlex Tomas if (sbi->s_buddy_cache) 2612c9de560dSAlex Tomas iput(sbi->s_buddy_cache); 2613c9de560dSAlex Tomas if (sbi->s_mb_stats) { 2614c9de560dSAlex Tomas printk(KERN_INFO 2615c9de560dSAlex Tomas "EXT4-fs: mballoc: %u blocks %u reqs (%u success)\n", 2616c9de560dSAlex Tomas atomic_read(&sbi->s_bal_allocated), 2617c9de560dSAlex Tomas atomic_read(&sbi->s_bal_reqs), 2618c9de560dSAlex Tomas atomic_read(&sbi->s_bal_success)); 2619c9de560dSAlex Tomas printk(KERN_INFO 2620c9de560dSAlex Tomas "EXT4-fs: mballoc: %u extents scanned, %u goal hits, " 2621c9de560dSAlex Tomas "%u 2^N hits, %u breaks, %u lost\n", 2622c9de560dSAlex Tomas atomic_read(&sbi->s_bal_ex_scanned), 2623c9de560dSAlex Tomas atomic_read(&sbi->s_bal_goals), 2624c9de560dSAlex Tomas atomic_read(&sbi->s_bal_2orders), 2625c9de560dSAlex Tomas atomic_read(&sbi->s_bal_breaks), 2626c9de560dSAlex Tomas atomic_read(&sbi->s_mb_lost_chunks)); 2627c9de560dSAlex Tomas printk(KERN_INFO 2628c9de560dSAlex Tomas "EXT4-fs: mballoc: %lu generated and it took %Lu\n", 2629c9de560dSAlex Tomas sbi->s_mb_buddies_generated++, 2630c9de560dSAlex Tomas sbi->s_mb_generation_time); 2631c9de560dSAlex Tomas printk(KERN_INFO 2632c9de560dSAlex Tomas "EXT4-fs: mballoc: %u preallocated, %u discarded\n", 2633c9de560dSAlex Tomas atomic_read(&sbi->s_mb_preallocated), 2634c9de560dSAlex Tomas atomic_read(&sbi->s_mb_discarded)); 2635c9de560dSAlex Tomas } 2636c9de560dSAlex Tomas 2637730c213cSEric Sandeen free_percpu(sbi->s_locality_groups); 2638296c355cSTheodore Ts'o if (sbi->s_proc) 2639296c355cSTheodore Ts'o remove_proc_entry("mb_groups", sbi->s_proc); 2640c9de560dSAlex Tomas 2641c9de560dSAlex Tomas return 0; 2642c9de560dSAlex Tomas } 2643c9de560dSAlex Tomas 264477ca6cdfSLukas Czerner static inline int ext4_issue_discard(struct super_block *sb, 26455c521830SJiaying Zhang ext4_group_t block_group, ext4_grpblk_t block, int count) 26465c521830SJiaying Zhang { 26475c521830SJiaying Zhang ext4_fsblk_t discard_block; 26485c521830SJiaying Zhang 26495c521830SJiaying Zhang discard_block = block + ext4_group_first_block_no(sb, block_group); 26505c521830SJiaying Zhang trace_ext4_discard_blocks(sb, 26515c521830SJiaying Zhang (unsigned long long) discard_block, count); 265293259636SLukas Czerner return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0); 26535c521830SJiaying Zhang } 26545c521830SJiaying Zhang 26553e624fc7STheodore Ts'o /* 26563e624fc7STheodore Ts'o * This function is called by the jbd2 layer once the commit has finished, 26573e624fc7STheodore Ts'o * so we know we can free the blocks that were released with that commit. 26583e624fc7STheodore Ts'o */ 26593e624fc7STheodore Ts'o static void release_blocks_on_commit(journal_t *journal, transaction_t *txn) 2660c9de560dSAlex Tomas { 26613e624fc7STheodore Ts'o struct super_block *sb = journal->j_private; 2662c9de560dSAlex Tomas struct ext4_buddy e4b; 2663c894058dSAneesh Kumar K.V struct ext4_group_info *db; 266493259636SLukas Czerner int err, ret, count = 0, count2 = 0; 2665c894058dSAneesh Kumar K.V struct ext4_free_data *entry; 26663e624fc7STheodore Ts'o struct list_head *l, *ltmp; 2667c9de560dSAlex Tomas 26683e624fc7STheodore Ts'o list_for_each_safe(l, ltmp, &txn->t_private_list) { 26693e624fc7STheodore Ts'o entry = list_entry(l, struct ext4_free_data, list); 2670c9de560dSAlex Tomas 26716ba495e9STheodore Ts'o mb_debug(1, "gonna free %u blocks in group %u (0x%p):", 2672c894058dSAneesh Kumar K.V entry->count, entry->group, entry); 2673c9de560dSAlex Tomas 267493259636SLukas Czerner if (test_opt(sb, DISCARD)) { 267593259636SLukas Czerner ret = ext4_issue_discard(sb, entry->group, 26765c521830SJiaying Zhang entry->start_blk, entry->count); 267793259636SLukas Czerner if (unlikely(ret == -EOPNOTSUPP)) { 267893259636SLukas Czerner ext4_warning(sb, "discard not supported, " 267993259636SLukas Czerner "disabling"); 268093259636SLukas Czerner clear_opt(sb, DISCARD); 268193259636SLukas Czerner } 268293259636SLukas Czerner } 2683b90f6870STheodore Ts'o 2684c894058dSAneesh Kumar K.V err = ext4_mb_load_buddy(sb, entry->group, &e4b); 2685c9de560dSAlex Tomas /* we expect to find existing buddy because it's pinned */ 2686c9de560dSAlex Tomas BUG_ON(err != 0); 2687c9de560dSAlex Tomas 2688c894058dSAneesh Kumar K.V db = e4b.bd_info; 2689c9de560dSAlex Tomas /* there are blocks to put in buddy to make them really free */ 2690c894058dSAneesh Kumar K.V count += entry->count; 2691c9de560dSAlex Tomas count2++; 2692c894058dSAneesh Kumar K.V ext4_lock_group(sb, entry->group); 2693c894058dSAneesh Kumar K.V /* Take it out of per group rb tree */ 2694c894058dSAneesh Kumar K.V rb_erase(&entry->node, &(db->bb_free_root)); 2695c894058dSAneesh Kumar K.V mb_free_blocks(NULL, &e4b, entry->start_blk, entry->count); 2696c9de560dSAlex Tomas 2697c894058dSAneesh Kumar K.V if (!db->bb_free_root.rb_node) { 2698c894058dSAneesh Kumar K.V /* No more items in the per group rb tree 2699c894058dSAneesh Kumar K.V * balance refcounts from ext4_mb_free_metadata() 2700c894058dSAneesh Kumar K.V */ 2701c9de560dSAlex Tomas page_cache_release(e4b.bd_buddy_page); 2702c9de560dSAlex Tomas page_cache_release(e4b.bd_bitmap_page); 2703c894058dSAneesh Kumar K.V } 2704c894058dSAneesh Kumar K.V ext4_unlock_group(sb, entry->group); 2705c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 2706e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 27073e624fc7STheodore Ts'o } 2708c9de560dSAlex Tomas 27096ba495e9STheodore Ts'o mb_debug(1, "freed %u blocks in %u structures\n", count, count2); 2710c9de560dSAlex Tomas } 2711c9de560dSAlex Tomas 27126ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 27136ba495e9STheodore Ts'o u8 mb_enable_debug __read_mostly; 27146ba495e9STheodore Ts'o 27156ba495e9STheodore Ts'o static struct dentry *debugfs_dir; 27166ba495e9STheodore Ts'o static struct dentry *debugfs_debug; 27176ba495e9STheodore Ts'o 27186ba495e9STheodore Ts'o static void __init ext4_create_debugfs_entry(void) 27196ba495e9STheodore Ts'o { 27206ba495e9STheodore Ts'o debugfs_dir = debugfs_create_dir("ext4", NULL); 27216ba495e9STheodore Ts'o if (debugfs_dir) 27226ba495e9STheodore Ts'o debugfs_debug = debugfs_create_u8("mballoc-debug", 27236ba495e9STheodore Ts'o S_IRUGO | S_IWUSR, 27246ba495e9STheodore Ts'o debugfs_dir, 27256ba495e9STheodore Ts'o &mb_enable_debug); 27266ba495e9STheodore Ts'o } 27276ba495e9STheodore Ts'o 27286ba495e9STheodore Ts'o static void ext4_remove_debugfs_entry(void) 27296ba495e9STheodore Ts'o { 27306ba495e9STheodore Ts'o debugfs_remove(debugfs_debug); 27316ba495e9STheodore Ts'o debugfs_remove(debugfs_dir); 27326ba495e9STheodore Ts'o } 27336ba495e9STheodore Ts'o 27346ba495e9STheodore Ts'o #else 27356ba495e9STheodore Ts'o 27366ba495e9STheodore Ts'o static void __init ext4_create_debugfs_entry(void) 27376ba495e9STheodore Ts'o { 27386ba495e9STheodore Ts'o } 27396ba495e9STheodore Ts'o 27406ba495e9STheodore Ts'o static void ext4_remove_debugfs_entry(void) 27416ba495e9STheodore Ts'o { 27426ba495e9STheodore Ts'o } 27436ba495e9STheodore Ts'o 27446ba495e9STheodore Ts'o #endif 27456ba495e9STheodore Ts'o 27465dabfc78STheodore Ts'o int __init ext4_init_mballoc(void) 2747c9de560dSAlex Tomas { 274816828088STheodore Ts'o ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space, 274916828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 2750c9de560dSAlex Tomas if (ext4_pspace_cachep == NULL) 2751c9de560dSAlex Tomas return -ENOMEM; 2752c9de560dSAlex Tomas 275316828088STheodore Ts'o ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context, 275416828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 2755256bdb49SEric Sandeen if (ext4_ac_cachep == NULL) { 2756256bdb49SEric Sandeen kmem_cache_destroy(ext4_pspace_cachep); 2757256bdb49SEric Sandeen return -ENOMEM; 2758256bdb49SEric Sandeen } 2759c894058dSAneesh Kumar K.V 276016828088STheodore Ts'o ext4_free_ext_cachep = KMEM_CACHE(ext4_free_data, 276116828088STheodore Ts'o SLAB_RECLAIM_ACCOUNT); 2762c894058dSAneesh Kumar K.V if (ext4_free_ext_cachep == NULL) { 2763c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_pspace_cachep); 2764c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_ac_cachep); 2765c894058dSAneesh Kumar K.V return -ENOMEM; 2766c894058dSAneesh Kumar K.V } 27676ba495e9STheodore Ts'o ext4_create_debugfs_entry(); 2768c9de560dSAlex Tomas return 0; 2769c9de560dSAlex Tomas } 2770c9de560dSAlex Tomas 27715dabfc78STheodore Ts'o void ext4_exit_mballoc(void) 2772c9de560dSAlex Tomas { 27733e03f9caSJesper Dangaard Brouer /* 27743e03f9caSJesper Dangaard Brouer * Wait for completion of call_rcu()'s on ext4_pspace_cachep 27753e03f9caSJesper Dangaard Brouer * before destroying the slab cache. 27763e03f9caSJesper Dangaard Brouer */ 27773e03f9caSJesper Dangaard Brouer rcu_barrier(); 2778c9de560dSAlex Tomas kmem_cache_destroy(ext4_pspace_cachep); 2779256bdb49SEric Sandeen kmem_cache_destroy(ext4_ac_cachep); 2780c894058dSAneesh Kumar K.V kmem_cache_destroy(ext4_free_ext_cachep); 27812892c15dSEric Sandeen ext4_groupinfo_destroy_slabs(); 27826ba495e9STheodore Ts'o ext4_remove_debugfs_entry(); 2783c9de560dSAlex Tomas } 2784c9de560dSAlex Tomas 2785c9de560dSAlex Tomas 2786c9de560dSAlex Tomas /* 278773b2c716SUwe Kleine-König * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps 2788c9de560dSAlex Tomas * Returns 0 if success or error code 2789c9de560dSAlex Tomas */ 27904ddfef7bSEric Sandeen static noinline_for_stack int 27914ddfef7bSEric Sandeen ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac, 2792498e5f24STheodore Ts'o handle_t *handle, unsigned int reserv_blks) 2793c9de560dSAlex Tomas { 2794c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 2795c9de560dSAlex Tomas struct ext4_group_desc *gdp; 2796c9de560dSAlex Tomas struct buffer_head *gdp_bh; 2797c9de560dSAlex Tomas struct ext4_sb_info *sbi; 2798c9de560dSAlex Tomas struct super_block *sb; 2799c9de560dSAlex Tomas ext4_fsblk_t block; 2800519deca0SAneesh Kumar K.V int err, len; 2801c9de560dSAlex Tomas 2802c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 2803c9de560dSAlex Tomas BUG_ON(ac->ac_b_ex.fe_len <= 0); 2804c9de560dSAlex Tomas 2805c9de560dSAlex Tomas sb = ac->ac_sb; 2806c9de560dSAlex Tomas sbi = EXT4_SB(sb); 2807c9de560dSAlex Tomas 2808c9de560dSAlex Tomas err = -EIO; 2809574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group); 2810c9de560dSAlex Tomas if (!bitmap_bh) 2811c9de560dSAlex Tomas goto out_err; 2812c9de560dSAlex Tomas 2813c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 2814c9de560dSAlex Tomas if (err) 2815c9de560dSAlex Tomas goto out_err; 2816c9de560dSAlex Tomas 2817c9de560dSAlex Tomas err = -EIO; 2818c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh); 2819c9de560dSAlex Tomas if (!gdp) 2820c9de560dSAlex Tomas goto out_err; 2821c9de560dSAlex Tomas 2822a9df9a49STheodore Ts'o ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group, 28239fd9784cSThadeu Lima de Souza Cascardo ext4_free_blks_count(sb, gdp)); 282403cddb80SAneesh Kumar K.V 2825c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gdp_bh); 2826c9de560dSAlex Tomas if (err) 2827c9de560dSAlex Tomas goto out_err; 2828c9de560dSAlex Tomas 2829bda00de7SAkinobu Mita block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 2830c9de560dSAlex Tomas 2831519deca0SAneesh Kumar K.V len = ac->ac_b_ex.fe_len; 28326fd058f7STheodore Ts'o if (!ext4_data_block_valid(sbi, block, len)) { 283312062dddSEric Sandeen ext4_error(sb, "Allocating blocks %llu-%llu which overlap " 28346fd058f7STheodore Ts'o "fs metadata\n", block, block+len); 2835519deca0SAneesh Kumar K.V /* File system mounted not to panic on error 2836519deca0SAneesh Kumar K.V * Fix the bitmap and repeat the block allocation 2837519deca0SAneesh Kumar K.V * We leak some of the blocks here. 2838519deca0SAneesh Kumar K.V */ 2839955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2840955ce5f5SAneesh Kumar K.V mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start, 2841519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len); 2842955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 28430390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 2844519deca0SAneesh Kumar K.V if (!err) 2845519deca0SAneesh Kumar K.V err = -EAGAIN; 2846519deca0SAneesh Kumar K.V goto out_err; 2847c9de560dSAlex Tomas } 2848955ce5f5SAneesh Kumar K.V 2849955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, ac->ac_b_ex.fe_group); 2850c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 2851c9de560dSAlex Tomas { 2852c9de560dSAlex Tomas int i; 2853c9de560dSAlex Tomas for (i = 0; i < ac->ac_b_ex.fe_len; i++) { 2854c9de560dSAlex Tomas BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i, 2855c9de560dSAlex Tomas bitmap_bh->b_data)); 2856c9de560dSAlex Tomas } 2857c9de560dSAlex Tomas } 2858c9de560dSAlex Tomas #endif 2859955ce5f5SAneesh Kumar K.V mb_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,ac->ac_b_ex.fe_len); 2860c9de560dSAlex Tomas if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) { 2861c9de560dSAlex Tomas gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT); 2862560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, 2863560671a0SAneesh Kumar K.V ext4_free_blocks_after_init(sb, 2864560671a0SAneesh Kumar K.V ac->ac_b_ex.fe_group, gdp)); 2865c9de560dSAlex Tomas } 2866560671a0SAneesh Kumar K.V len = ext4_free_blks_count(sb, gdp) - ac->ac_b_ex.fe_len; 2867560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, len); 2868c9de560dSAlex Tomas gdp->bg_checksum = ext4_group_desc_csum(sbi, ac->ac_b_ex.fe_group, gdp); 2869955ce5f5SAneesh Kumar K.V 2870955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 28716bc6e63fSAneesh Kumar K.V percpu_counter_sub(&sbi->s_freeblocks_counter, ac->ac_b_ex.fe_len); 2872d2a17637SMingming Cao /* 28736bc6e63fSAneesh Kumar K.V * Now reduce the dirty block count also. Should not go negative 2874d2a17637SMingming Cao */ 28756bc6e63fSAneesh Kumar K.V if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED)) 28766bc6e63fSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 28776bc6e63fSAneesh Kumar K.V percpu_counter_sub(&sbi->s_dirtyblocks_counter, reserv_blks); 2878c9de560dSAlex Tomas 2879772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 2880772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, 2881772cb7c8SJose R. Santos ac->ac_b_ex.fe_group); 28829f24e420STheodore Ts'o atomic_sub(ac->ac_b_ex.fe_len, 28839f24e420STheodore Ts'o &sbi->s_flex_groups[flex_group].free_blocks); 2884772cb7c8SJose R. Santos } 2885772cb7c8SJose R. Santos 28860390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 2887c9de560dSAlex Tomas if (err) 2888c9de560dSAlex Tomas goto out_err; 28890390131bSFrank Mayhar err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh); 2890c9de560dSAlex Tomas 2891c9de560dSAlex Tomas out_err: 2892a0375156STheodore Ts'o ext4_mark_super_dirty(sb); 289342a10addSAneesh Kumar K.V brelse(bitmap_bh); 2894c9de560dSAlex Tomas return err; 2895c9de560dSAlex Tomas } 2896c9de560dSAlex Tomas 2897c9de560dSAlex Tomas /* 2898c9de560dSAlex Tomas * here we normalize request for locality group 2899c9de560dSAlex Tomas * Group request are normalized to s_strip size if we set the same via mount 2900c9de560dSAlex Tomas * option. If not we set it to s_mb_group_prealloc which can be configured via 2901b713a5ecSTheodore Ts'o * /sys/fs/ext4/<partition>/mb_group_prealloc 2902c9de560dSAlex Tomas * 2903c9de560dSAlex Tomas * XXX: should we try to preallocate more than the group has now? 2904c9de560dSAlex Tomas */ 2905c9de560dSAlex Tomas static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac) 2906c9de560dSAlex Tomas { 2907c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 2908c9de560dSAlex Tomas struct ext4_locality_group *lg = ac->ac_lg; 2909c9de560dSAlex Tomas 2910c9de560dSAlex Tomas BUG_ON(lg == NULL); 2911c9de560dSAlex Tomas if (EXT4_SB(sb)->s_stripe) 2912c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe; 2913c9de560dSAlex Tomas else 2914c9de560dSAlex Tomas ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc; 29156ba495e9STheodore Ts'o mb_debug(1, "#%u: goal %u blocks for locality group\n", 2916c9de560dSAlex Tomas current->pid, ac->ac_g_ex.fe_len); 2917c9de560dSAlex Tomas } 2918c9de560dSAlex Tomas 2919c9de560dSAlex Tomas /* 2920c9de560dSAlex Tomas * Normalization means making request better in terms of 2921c9de560dSAlex Tomas * size and alignment 2922c9de560dSAlex Tomas */ 29234ddfef7bSEric Sandeen static noinline_for_stack void 29244ddfef7bSEric Sandeen ext4_mb_normalize_request(struct ext4_allocation_context *ac, 2925c9de560dSAlex Tomas struct ext4_allocation_request *ar) 2926c9de560dSAlex Tomas { 2927c9de560dSAlex Tomas int bsbits, max; 2928c9de560dSAlex Tomas ext4_lblk_t end; 2929c9de560dSAlex Tomas loff_t size, orig_size, start_off; 29305a0790c2SAndi Kleen ext4_lblk_t start; 2931c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 29329a0762c5SAneesh Kumar K.V struct ext4_prealloc_space *pa; 2933c9de560dSAlex Tomas 2934c9de560dSAlex Tomas /* do normalize only data requests, metadata requests 2935c9de560dSAlex Tomas do not need preallocation */ 2936c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 2937c9de560dSAlex Tomas return; 2938c9de560dSAlex Tomas 2939c9de560dSAlex Tomas /* sometime caller may want exact blocks */ 2940c9de560dSAlex Tomas if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 2941c9de560dSAlex Tomas return; 2942c9de560dSAlex Tomas 2943c9de560dSAlex Tomas /* caller may indicate that preallocation isn't 2944c9de560dSAlex Tomas * required (it's a tail, for example) */ 2945c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC) 2946c9de560dSAlex Tomas return; 2947c9de560dSAlex Tomas 2948c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) { 2949c9de560dSAlex Tomas ext4_mb_normalize_group_request(ac); 2950c9de560dSAlex Tomas return ; 2951c9de560dSAlex Tomas } 2952c9de560dSAlex Tomas 2953c9de560dSAlex Tomas bsbits = ac->ac_sb->s_blocksize_bits; 2954c9de560dSAlex Tomas 2955c9de560dSAlex Tomas /* first, let's learn actual file size 2956c9de560dSAlex Tomas * given current request is allocated */ 2957c9de560dSAlex Tomas size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; 2958c9de560dSAlex Tomas size = size << bsbits; 2959c9de560dSAlex Tomas if (size < i_size_read(ac->ac_inode)) 2960c9de560dSAlex Tomas size = i_size_read(ac->ac_inode); 29615a0790c2SAndi Kleen orig_size = size; 2962c9de560dSAlex Tomas 29631930479cSValerie Clement /* max size of free chunks */ 29641930479cSValerie Clement max = 2 << bsbits; 2965c9de560dSAlex Tomas 29661930479cSValerie Clement #define NRL_CHECK_SIZE(req, size, max, chunk_size) \ 29671930479cSValerie Clement (req <= (size) || max <= (chunk_size)) 2968c9de560dSAlex Tomas 2969c9de560dSAlex Tomas /* first, try to predict filesize */ 2970c9de560dSAlex Tomas /* XXX: should this table be tunable? */ 2971c9de560dSAlex Tomas start_off = 0; 2972c9de560dSAlex Tomas if (size <= 16 * 1024) { 2973c9de560dSAlex Tomas size = 16 * 1024; 2974c9de560dSAlex Tomas } else if (size <= 32 * 1024) { 2975c9de560dSAlex Tomas size = 32 * 1024; 2976c9de560dSAlex Tomas } else if (size <= 64 * 1024) { 2977c9de560dSAlex Tomas size = 64 * 1024; 2978c9de560dSAlex Tomas } else if (size <= 128 * 1024) { 2979c9de560dSAlex Tomas size = 128 * 1024; 2980c9de560dSAlex Tomas } else if (size <= 256 * 1024) { 2981c9de560dSAlex Tomas size = 256 * 1024; 2982c9de560dSAlex Tomas } else if (size <= 512 * 1024) { 2983c9de560dSAlex Tomas size = 512 * 1024; 2984c9de560dSAlex Tomas } else if (size <= 1024 * 1024) { 2985c9de560dSAlex Tomas size = 1024 * 1024; 29861930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) { 2987c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 29881930479cSValerie Clement (21 - bsbits)) << 21; 29891930479cSValerie Clement size = 2 * 1024 * 1024; 29901930479cSValerie Clement } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) { 2991c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 2992c9de560dSAlex Tomas (22 - bsbits)) << 22; 2993c9de560dSAlex Tomas size = 4 * 1024 * 1024; 2994c9de560dSAlex Tomas } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len, 29951930479cSValerie Clement (8<<20)>>bsbits, max, 8 * 1024)) { 2996c9de560dSAlex Tomas start_off = ((loff_t)ac->ac_o_ex.fe_logical >> 2997c9de560dSAlex Tomas (23 - bsbits)) << 23; 2998c9de560dSAlex Tomas size = 8 * 1024 * 1024; 2999c9de560dSAlex Tomas } else { 3000c9de560dSAlex Tomas start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits; 3001c9de560dSAlex Tomas size = ac->ac_o_ex.fe_len << bsbits; 3002c9de560dSAlex Tomas } 30035a0790c2SAndi Kleen size = size >> bsbits; 30045a0790c2SAndi Kleen start = start_off >> bsbits; 3005c9de560dSAlex Tomas 3006c9de560dSAlex Tomas /* don't cover already allocated blocks in selected range */ 3007c9de560dSAlex Tomas if (ar->pleft && start <= ar->lleft) { 3008c9de560dSAlex Tomas size -= ar->lleft + 1 - start; 3009c9de560dSAlex Tomas start = ar->lleft + 1; 3010c9de560dSAlex Tomas } 3011c9de560dSAlex Tomas if (ar->pright && start + size - 1 >= ar->lright) 3012c9de560dSAlex Tomas size -= start + size - ar->lright; 3013c9de560dSAlex Tomas 3014c9de560dSAlex Tomas end = start + size; 3015c9de560dSAlex Tomas 3016c9de560dSAlex Tomas /* check we don't cross already preallocated blocks */ 3017c9de560dSAlex Tomas rcu_read_lock(); 30189a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3019498e5f24STheodore Ts'o ext4_lblk_t pa_end; 3020c9de560dSAlex Tomas 3021c9de560dSAlex Tomas if (pa->pa_deleted) 3022c9de560dSAlex Tomas continue; 3023c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3024c9de560dSAlex Tomas if (pa->pa_deleted) { 3025c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3026c9de560dSAlex Tomas continue; 3027c9de560dSAlex Tomas } 3028c9de560dSAlex Tomas 3029c9de560dSAlex Tomas pa_end = pa->pa_lstart + pa->pa_len; 3030c9de560dSAlex Tomas 3031c9de560dSAlex Tomas /* PA must not overlap original request */ 3032c9de560dSAlex Tomas BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end || 3033c9de560dSAlex Tomas ac->ac_o_ex.fe_logical < pa->pa_lstart)); 3034c9de560dSAlex Tomas 303538877f4eSEric Sandeen /* skip PAs this normalized request doesn't overlap with */ 303638877f4eSEric Sandeen if (pa->pa_lstart >= end || pa_end <= start) { 3037c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3038c9de560dSAlex Tomas continue; 3039c9de560dSAlex Tomas } 3040c9de560dSAlex Tomas BUG_ON(pa->pa_lstart <= start && pa_end >= end); 3041c9de560dSAlex Tomas 304238877f4eSEric Sandeen /* adjust start or end to be adjacent to this pa */ 3043c9de560dSAlex Tomas if (pa_end <= ac->ac_o_ex.fe_logical) { 3044c9de560dSAlex Tomas BUG_ON(pa_end < start); 3045c9de560dSAlex Tomas start = pa_end; 304638877f4eSEric Sandeen } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) { 3047c9de560dSAlex Tomas BUG_ON(pa->pa_lstart > end); 3048c9de560dSAlex Tomas end = pa->pa_lstart; 3049c9de560dSAlex Tomas } 3050c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3051c9de560dSAlex Tomas } 3052c9de560dSAlex Tomas rcu_read_unlock(); 3053c9de560dSAlex Tomas size = end - start; 3054c9de560dSAlex Tomas 3055c9de560dSAlex Tomas /* XXX: extra loop to check we really don't overlap preallocations */ 3056c9de560dSAlex Tomas rcu_read_lock(); 30579a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3058498e5f24STheodore Ts'o ext4_lblk_t pa_end; 3059c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3060c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 3061c9de560dSAlex Tomas pa_end = pa->pa_lstart + pa->pa_len; 3062c9de560dSAlex Tomas BUG_ON(!(start >= pa_end || end <= pa->pa_lstart)); 3063c9de560dSAlex Tomas } 3064c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3065c9de560dSAlex Tomas } 3066c9de560dSAlex Tomas rcu_read_unlock(); 3067c9de560dSAlex Tomas 3068c9de560dSAlex Tomas if (start + size <= ac->ac_o_ex.fe_logical && 3069c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical) { 3070c9de560dSAlex Tomas printk(KERN_ERR "start %lu, size %lu, fe_logical %lu\n", 3071c9de560dSAlex Tomas (unsigned long) start, (unsigned long) size, 3072c9de560dSAlex Tomas (unsigned long) ac->ac_o_ex.fe_logical); 3073c9de560dSAlex Tomas } 3074c9de560dSAlex Tomas BUG_ON(start + size <= ac->ac_o_ex.fe_logical && 3075c9de560dSAlex Tomas start > ac->ac_o_ex.fe_logical); 30768d03c7a0SEric Sandeen BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb)); 3077c9de560dSAlex Tomas 3078c9de560dSAlex Tomas /* now prepare goal request */ 3079c9de560dSAlex Tomas 3080c9de560dSAlex Tomas /* XXX: is it better to align blocks WRT to logical 3081c9de560dSAlex Tomas * placement or satisfy big request as is */ 3082c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = start; 3083c9de560dSAlex Tomas ac->ac_g_ex.fe_len = size; 3084c9de560dSAlex Tomas 3085c9de560dSAlex Tomas /* define goal start in order to merge */ 3086c9de560dSAlex Tomas if (ar->pright && (ar->lright == (start + size))) { 3087c9de560dSAlex Tomas /* merge to the right */ 3088c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, 3089c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 3090c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 3091c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3092c9de560dSAlex Tomas } 3093c9de560dSAlex Tomas if (ar->pleft && (ar->lleft + 1 == start)) { 3094c9de560dSAlex Tomas /* merge to the left */ 3095c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, 3096c9de560dSAlex Tomas &ac->ac_f_ex.fe_group, 3097c9de560dSAlex Tomas &ac->ac_f_ex.fe_start); 3098c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; 3099c9de560dSAlex Tomas } 3100c9de560dSAlex Tomas 31016ba495e9STheodore Ts'o mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size, 3102c9de560dSAlex Tomas (unsigned) orig_size, (unsigned) start); 3103c9de560dSAlex Tomas } 3104c9de560dSAlex Tomas 3105c9de560dSAlex Tomas static void ext4_mb_collect_stats(struct ext4_allocation_context *ac) 3106c9de560dSAlex Tomas { 3107c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3108c9de560dSAlex Tomas 3109c9de560dSAlex Tomas if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) { 3110c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_reqs); 3111c9de560dSAlex Tomas atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated); 3112291dae47SCurt Wohlgemuth if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len) 3113c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_success); 3114c9de560dSAlex Tomas atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned); 3115c9de560dSAlex Tomas if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start && 3116c9de560dSAlex Tomas ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group) 3117c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_goals); 3118c9de560dSAlex Tomas if (ac->ac_found > sbi->s_mb_max_to_scan) 3119c9de560dSAlex Tomas atomic_inc(&sbi->s_bal_breaks); 3120c9de560dSAlex Tomas } 3121c9de560dSAlex Tomas 3122296c355cSTheodore Ts'o if (ac->ac_op == EXT4_MB_HISTORY_ALLOC) 3123296c355cSTheodore Ts'o trace_ext4_mballoc_alloc(ac); 3124296c355cSTheodore Ts'o else 3125296c355cSTheodore Ts'o trace_ext4_mballoc_prealloc(ac); 3126c9de560dSAlex Tomas } 3127c9de560dSAlex Tomas 3128c9de560dSAlex Tomas /* 3129b844167eSCurt Wohlgemuth * Called on failure; free up any blocks from the inode PA for this 3130b844167eSCurt Wohlgemuth * context. We don't need this for MB_GROUP_PA because we only change 3131b844167eSCurt Wohlgemuth * pa_free in ext4_mb_release_context(), but on failure, we've already 3132b844167eSCurt Wohlgemuth * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed. 3133b844167eSCurt Wohlgemuth */ 3134b844167eSCurt Wohlgemuth static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac) 3135b844167eSCurt Wohlgemuth { 3136b844167eSCurt Wohlgemuth struct ext4_prealloc_space *pa = ac->ac_pa; 3137b844167eSCurt Wohlgemuth int len; 3138b844167eSCurt Wohlgemuth 3139b844167eSCurt Wohlgemuth if (pa && pa->pa_type == MB_INODE_PA) { 3140b844167eSCurt Wohlgemuth len = ac->ac_b_ex.fe_len; 3141b844167eSCurt Wohlgemuth pa->pa_free += len; 3142b844167eSCurt Wohlgemuth } 3143b844167eSCurt Wohlgemuth 3144b844167eSCurt Wohlgemuth } 3145b844167eSCurt Wohlgemuth 3146b844167eSCurt Wohlgemuth /* 3147c9de560dSAlex Tomas * use blocks preallocated to inode 3148c9de560dSAlex Tomas */ 3149c9de560dSAlex Tomas static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac, 3150c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3151c9de560dSAlex Tomas { 3152c9de560dSAlex Tomas ext4_fsblk_t start; 3153c9de560dSAlex Tomas ext4_fsblk_t end; 3154c9de560dSAlex Tomas int len; 3155c9de560dSAlex Tomas 3156c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3157c9de560dSAlex Tomas start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart); 3158c9de560dSAlex Tomas end = min(pa->pa_pstart + pa->pa_len, start + ac->ac_o_ex.fe_len); 3159c9de560dSAlex Tomas len = end - start; 3160c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group, 3161c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3162c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3163c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3164c9de560dSAlex Tomas ac->ac_pa = pa; 3165c9de560dSAlex Tomas 3166c9de560dSAlex Tomas BUG_ON(start < pa->pa_pstart); 3167c9de560dSAlex Tomas BUG_ON(start + len > pa->pa_pstart + pa->pa_len); 3168c9de560dSAlex Tomas BUG_ON(pa->pa_free < len); 3169c9de560dSAlex Tomas pa->pa_free -= len; 3170c9de560dSAlex Tomas 31716ba495e9STheodore Ts'o mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa); 3172c9de560dSAlex Tomas } 3173c9de560dSAlex Tomas 3174c9de560dSAlex Tomas /* 3175c9de560dSAlex Tomas * use blocks preallocated to locality group 3176c9de560dSAlex Tomas */ 3177c9de560dSAlex Tomas static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac, 3178c9de560dSAlex Tomas struct ext4_prealloc_space *pa) 3179c9de560dSAlex Tomas { 318003cddb80SAneesh Kumar K.V unsigned int len = ac->ac_o_ex.fe_len; 31816be2ded1SAneesh Kumar K.V 3182c9de560dSAlex Tomas ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart, 3183c9de560dSAlex Tomas &ac->ac_b_ex.fe_group, 3184c9de560dSAlex Tomas &ac->ac_b_ex.fe_start); 3185c9de560dSAlex Tomas ac->ac_b_ex.fe_len = len; 3186c9de560dSAlex Tomas ac->ac_status = AC_STATUS_FOUND; 3187c9de560dSAlex Tomas ac->ac_pa = pa; 3188c9de560dSAlex Tomas 3189c9de560dSAlex Tomas /* we don't correct pa_pstart or pa_plen here to avoid 319026346ff6SAneesh Kumar K.V * possible race when the group is being loaded concurrently 3191c9de560dSAlex Tomas * instead we correct pa later, after blocks are marked 319226346ff6SAneesh Kumar K.V * in on-disk bitmap -- see ext4_mb_release_context() 319326346ff6SAneesh Kumar K.V * Other CPUs are prevented from allocating from this pa by lg_mutex 3194c9de560dSAlex Tomas */ 31956ba495e9STheodore Ts'o mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa); 3196c9de560dSAlex Tomas } 3197c9de560dSAlex Tomas 3198c9de560dSAlex Tomas /* 31995e745b04SAneesh Kumar K.V * Return the prealloc space that have minimal distance 32005e745b04SAneesh Kumar K.V * from the goal block. @cpa is the prealloc 32015e745b04SAneesh Kumar K.V * space that is having currently known minimal distance 32025e745b04SAneesh Kumar K.V * from the goal block. 32035e745b04SAneesh Kumar K.V */ 32045e745b04SAneesh Kumar K.V static struct ext4_prealloc_space * 32055e745b04SAneesh Kumar K.V ext4_mb_check_group_pa(ext4_fsblk_t goal_block, 32065e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, 32075e745b04SAneesh Kumar K.V struct ext4_prealloc_space *cpa) 32085e745b04SAneesh Kumar K.V { 32095e745b04SAneesh Kumar K.V ext4_fsblk_t cur_distance, new_distance; 32105e745b04SAneesh Kumar K.V 32115e745b04SAneesh Kumar K.V if (cpa == NULL) { 32125e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 32135e745b04SAneesh Kumar K.V return pa; 32145e745b04SAneesh Kumar K.V } 32155e745b04SAneesh Kumar K.V cur_distance = abs(goal_block - cpa->pa_pstart); 32165e745b04SAneesh Kumar K.V new_distance = abs(goal_block - pa->pa_pstart); 32175e745b04SAneesh Kumar K.V 32185a54b2f1SColy Li if (cur_distance <= new_distance) 32195e745b04SAneesh Kumar K.V return cpa; 32205e745b04SAneesh Kumar K.V 32215e745b04SAneesh Kumar K.V /* drop the previous reference */ 32225e745b04SAneesh Kumar K.V atomic_dec(&cpa->pa_count); 32235e745b04SAneesh Kumar K.V atomic_inc(&pa->pa_count); 32245e745b04SAneesh Kumar K.V return pa; 32255e745b04SAneesh Kumar K.V } 32265e745b04SAneesh Kumar K.V 32275e745b04SAneesh Kumar K.V /* 3228c9de560dSAlex Tomas * search goal blocks in preallocated space 3229c9de560dSAlex Tomas */ 32304ddfef7bSEric Sandeen static noinline_for_stack int 32314ddfef7bSEric Sandeen ext4_mb_use_preallocated(struct ext4_allocation_context *ac) 3232c9de560dSAlex Tomas { 32336be2ded1SAneesh Kumar K.V int order, i; 3234c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(ac->ac_inode); 3235c9de560dSAlex Tomas struct ext4_locality_group *lg; 32365e745b04SAneesh Kumar K.V struct ext4_prealloc_space *pa, *cpa = NULL; 32375e745b04SAneesh Kumar K.V ext4_fsblk_t goal_block; 3238c9de560dSAlex Tomas 3239c9de560dSAlex Tomas /* only data can be preallocated */ 3240c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3241c9de560dSAlex Tomas return 0; 3242c9de560dSAlex Tomas 3243c9de560dSAlex Tomas /* first, try per-file preallocation */ 3244c9de560dSAlex Tomas rcu_read_lock(); 32459a0762c5SAneesh Kumar K.V list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) { 3246c9de560dSAlex Tomas 3247c9de560dSAlex Tomas /* all fields in this condition don't change, 3248c9de560dSAlex Tomas * so we can skip locking for them */ 3249c9de560dSAlex Tomas if (ac->ac_o_ex.fe_logical < pa->pa_lstart || 3250c9de560dSAlex Tomas ac->ac_o_ex.fe_logical >= pa->pa_lstart + pa->pa_len) 3251c9de560dSAlex Tomas continue; 3252c9de560dSAlex Tomas 3253fb0a387dSEric Sandeen /* non-extent files can't have physical blocks past 2^32 */ 325412e9b892SDmitry Monakhov if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) && 3255fb0a387dSEric Sandeen pa->pa_pstart + pa->pa_len > EXT4_MAX_BLOCK_FILE_PHYS) 3256fb0a387dSEric Sandeen continue; 3257fb0a387dSEric Sandeen 3258c9de560dSAlex Tomas /* found preallocated blocks, use them */ 3259c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3260c9de560dSAlex Tomas if (pa->pa_deleted == 0 && pa->pa_free) { 3261c9de560dSAlex Tomas atomic_inc(&pa->pa_count); 3262c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 3263c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3264c9de560dSAlex Tomas ac->ac_criteria = 10; 3265c9de560dSAlex Tomas rcu_read_unlock(); 3266c9de560dSAlex Tomas return 1; 3267c9de560dSAlex Tomas } 3268c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3269c9de560dSAlex Tomas } 3270c9de560dSAlex Tomas rcu_read_unlock(); 3271c9de560dSAlex Tomas 3272c9de560dSAlex Tomas /* can we use group allocation? */ 3273c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)) 3274c9de560dSAlex Tomas return 0; 3275c9de560dSAlex Tomas 3276c9de560dSAlex Tomas /* inode may have no locality group for some reason */ 3277c9de560dSAlex Tomas lg = ac->ac_lg; 3278c9de560dSAlex Tomas if (lg == NULL) 3279c9de560dSAlex Tomas return 0; 32806be2ded1SAneesh Kumar K.V order = fls(ac->ac_o_ex.fe_len) - 1; 32816be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 32826be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 32836be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 3284c9de560dSAlex Tomas 3285bda00de7SAkinobu Mita goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex); 32865e745b04SAneesh Kumar K.V /* 32875e745b04SAneesh Kumar K.V * search for the prealloc space that is having 32885e745b04SAneesh Kumar K.V * minimal distance from the goal block. 32895e745b04SAneesh Kumar K.V */ 32906be2ded1SAneesh Kumar K.V for (i = order; i < PREALLOC_TB_SIZE; i++) { 3291c9de560dSAlex Tomas rcu_read_lock(); 32926be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i], 32936be2ded1SAneesh Kumar K.V pa_inode_list) { 3294c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 32956be2ded1SAneesh Kumar K.V if (pa->pa_deleted == 0 && 32966be2ded1SAneesh Kumar K.V pa->pa_free >= ac->ac_o_ex.fe_len) { 32975e745b04SAneesh Kumar K.V 32985e745b04SAneesh Kumar K.V cpa = ext4_mb_check_group_pa(goal_block, 32995e745b04SAneesh Kumar K.V pa, cpa); 33005e745b04SAneesh Kumar K.V } 3301c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 33025e745b04SAneesh Kumar K.V } 33035e745b04SAneesh Kumar K.V rcu_read_unlock(); 33045e745b04SAneesh Kumar K.V } 33055e745b04SAneesh Kumar K.V if (cpa) { 33065e745b04SAneesh Kumar K.V ext4_mb_use_group_pa(ac, cpa); 3307c9de560dSAlex Tomas ac->ac_criteria = 20; 3308c9de560dSAlex Tomas return 1; 3309c9de560dSAlex Tomas } 3310c9de560dSAlex Tomas return 0; 3311c9de560dSAlex Tomas } 3312c9de560dSAlex Tomas 3313c9de560dSAlex Tomas /* 33147a2fcbf7SAneesh Kumar K.V * the function goes through all block freed in the group 33157a2fcbf7SAneesh Kumar K.V * but not yet committed and marks them used in in-core bitmap. 33167a2fcbf7SAneesh Kumar K.V * buddy must be generated from this bitmap 3317955ce5f5SAneesh Kumar K.V * Need to be called with the ext4 group lock held 33187a2fcbf7SAneesh Kumar K.V */ 33197a2fcbf7SAneesh Kumar K.V static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap, 33207a2fcbf7SAneesh Kumar K.V ext4_group_t group) 33217a2fcbf7SAneesh Kumar K.V { 33227a2fcbf7SAneesh Kumar K.V struct rb_node *n; 33237a2fcbf7SAneesh Kumar K.V struct ext4_group_info *grp; 33247a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 33257a2fcbf7SAneesh Kumar K.V 33267a2fcbf7SAneesh Kumar K.V grp = ext4_get_group_info(sb, group); 33277a2fcbf7SAneesh Kumar K.V n = rb_first(&(grp->bb_free_root)); 33287a2fcbf7SAneesh Kumar K.V 33297a2fcbf7SAneesh Kumar K.V while (n) { 33307a2fcbf7SAneesh Kumar K.V entry = rb_entry(n, struct ext4_free_data, node); 3331955ce5f5SAneesh Kumar K.V mb_set_bits(bitmap, entry->start_blk, entry->count); 33327a2fcbf7SAneesh Kumar K.V n = rb_next(n); 33337a2fcbf7SAneesh Kumar K.V } 33347a2fcbf7SAneesh Kumar K.V return; 33357a2fcbf7SAneesh Kumar K.V } 33367a2fcbf7SAneesh Kumar K.V 33377a2fcbf7SAneesh Kumar K.V /* 3338c9de560dSAlex Tomas * the function goes through all preallocation in this group and marks them 3339c9de560dSAlex Tomas * used in in-core bitmap. buddy must be generated from this bitmap 3340955ce5f5SAneesh Kumar K.V * Need to be called with ext4 group lock held 3341c9de560dSAlex Tomas */ 3342089ceeccSEric Sandeen static noinline_for_stack 3343089ceeccSEric Sandeen void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap, 3344c9de560dSAlex Tomas ext4_group_t group) 3345c9de560dSAlex Tomas { 3346c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3347c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3348c9de560dSAlex Tomas struct list_head *cur; 3349c9de560dSAlex Tomas ext4_group_t groupnr; 3350c9de560dSAlex Tomas ext4_grpblk_t start; 3351c9de560dSAlex Tomas int preallocated = 0; 3352c9de560dSAlex Tomas int count = 0; 3353c9de560dSAlex Tomas int len; 3354c9de560dSAlex Tomas 3355c9de560dSAlex Tomas /* all form of preallocation discards first load group, 3356c9de560dSAlex Tomas * so the only competing code is preallocation use. 3357c9de560dSAlex Tomas * we don't need any locking here 3358c9de560dSAlex Tomas * notice we do NOT ignore preallocations with pa_deleted 3359c9de560dSAlex Tomas * otherwise we could leave used blocks available for 3360c9de560dSAlex Tomas * allocation in buddy when concurrent ext4_mb_put_pa() 3361c9de560dSAlex Tomas * is dropping preallocation 3362c9de560dSAlex Tomas */ 3363c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 3364c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list); 3365c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3366c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3367c9de560dSAlex Tomas &groupnr, &start); 3368c9de560dSAlex Tomas len = pa->pa_len; 3369c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3370c9de560dSAlex Tomas if (unlikely(len == 0)) 3371c9de560dSAlex Tomas continue; 3372c9de560dSAlex Tomas BUG_ON(groupnr != group); 3373955ce5f5SAneesh Kumar K.V mb_set_bits(bitmap, start, len); 3374c9de560dSAlex Tomas preallocated += len; 3375c9de560dSAlex Tomas count++; 3376c9de560dSAlex Tomas } 33776ba495e9STheodore Ts'o mb_debug(1, "prellocated %u for group %u\n", preallocated, group); 3378c9de560dSAlex Tomas } 3379c9de560dSAlex Tomas 3380c9de560dSAlex Tomas static void ext4_mb_pa_callback(struct rcu_head *head) 3381c9de560dSAlex Tomas { 3382c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3383c9de560dSAlex Tomas pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu); 3384c9de560dSAlex Tomas kmem_cache_free(ext4_pspace_cachep, pa); 3385c9de560dSAlex Tomas } 3386c9de560dSAlex Tomas 3387c9de560dSAlex Tomas /* 3388c9de560dSAlex Tomas * drops a reference to preallocated space descriptor 3389c9de560dSAlex Tomas * if this was the last reference and the space is consumed 3390c9de560dSAlex Tomas */ 3391c9de560dSAlex Tomas static void ext4_mb_put_pa(struct ext4_allocation_context *ac, 3392c9de560dSAlex Tomas struct super_block *sb, struct ext4_prealloc_space *pa) 3393c9de560dSAlex Tomas { 3394a9df9a49STheodore Ts'o ext4_group_t grp; 3395d33a1976SEric Sandeen ext4_fsblk_t grp_blk; 3396c9de560dSAlex Tomas 3397c9de560dSAlex Tomas if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) 3398c9de560dSAlex Tomas return; 3399c9de560dSAlex Tomas 3400c9de560dSAlex Tomas /* in this short window concurrent discard can set pa_deleted */ 3401c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3402c9de560dSAlex Tomas if (pa->pa_deleted == 1) { 3403c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3404c9de560dSAlex Tomas return; 3405c9de560dSAlex Tomas } 3406c9de560dSAlex Tomas 3407c9de560dSAlex Tomas pa->pa_deleted = 1; 3408c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3409c9de560dSAlex Tomas 3410d33a1976SEric Sandeen grp_blk = pa->pa_pstart; 3411cc0fb9adSAneesh Kumar K.V /* 3412cc0fb9adSAneesh Kumar K.V * If doing group-based preallocation, pa_pstart may be in the 3413cc0fb9adSAneesh Kumar K.V * next group when pa is used up 3414cc0fb9adSAneesh Kumar K.V */ 3415cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 3416d33a1976SEric Sandeen grp_blk--; 3417d33a1976SEric Sandeen 3418d33a1976SEric Sandeen ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL); 3419c9de560dSAlex Tomas 3420c9de560dSAlex Tomas /* 3421c9de560dSAlex Tomas * possible race: 3422c9de560dSAlex Tomas * 3423c9de560dSAlex Tomas * P1 (buddy init) P2 (regular allocation) 3424c9de560dSAlex Tomas * find block B in PA 3425c9de560dSAlex Tomas * copy on-disk bitmap to buddy 3426c9de560dSAlex Tomas * mark B in on-disk bitmap 3427c9de560dSAlex Tomas * drop PA from group 3428c9de560dSAlex Tomas * mark all PAs in buddy 3429c9de560dSAlex Tomas * 3430c9de560dSAlex Tomas * thus, P1 initializes buddy with B available. to prevent this 3431c9de560dSAlex Tomas * we make "copy" and "mark all PAs" atomic and serialize "drop PA" 3432c9de560dSAlex Tomas * against that pair 3433c9de560dSAlex Tomas */ 3434c9de560dSAlex Tomas ext4_lock_group(sb, grp); 3435c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3436c9de560dSAlex Tomas ext4_unlock_group(sb, grp); 3437c9de560dSAlex Tomas 3438c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3439c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3440c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3441c9de560dSAlex Tomas 3442c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3443c9de560dSAlex Tomas } 3444c9de560dSAlex Tomas 3445c9de560dSAlex Tomas /* 3446c9de560dSAlex Tomas * creates new preallocated space for given inode 3447c9de560dSAlex Tomas */ 34484ddfef7bSEric Sandeen static noinline_for_stack int 34494ddfef7bSEric Sandeen ext4_mb_new_inode_pa(struct ext4_allocation_context *ac) 3450c9de560dSAlex Tomas { 3451c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3452c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3453c9de560dSAlex Tomas struct ext4_group_info *grp; 3454c9de560dSAlex Tomas struct ext4_inode_info *ei; 3455c9de560dSAlex Tomas 3456c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3457c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3458c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3459c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3460c9de560dSAlex Tomas 3461c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3462c9de560dSAlex Tomas if (pa == NULL) 3463c9de560dSAlex Tomas return -ENOMEM; 3464c9de560dSAlex Tomas 3465c9de560dSAlex Tomas if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) { 3466c9de560dSAlex Tomas int winl; 3467c9de560dSAlex Tomas int wins; 3468c9de560dSAlex Tomas int win; 3469c9de560dSAlex Tomas int offs; 3470c9de560dSAlex Tomas 3471c9de560dSAlex Tomas /* we can't allocate as much as normalizer wants. 3472c9de560dSAlex Tomas * so, found space must get proper lstart 3473c9de560dSAlex Tomas * to cover original request */ 3474c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical); 3475c9de560dSAlex Tomas BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len); 3476c9de560dSAlex Tomas 3477c9de560dSAlex Tomas /* we're limited by original request in that 3478c9de560dSAlex Tomas * logical block must be covered any way 3479c9de560dSAlex Tomas * winl is window we can move our chunk within */ 3480c9de560dSAlex Tomas winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical; 3481c9de560dSAlex Tomas 3482c9de560dSAlex Tomas /* also, we should cover whole original request */ 3483c9de560dSAlex Tomas wins = ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len; 3484c9de560dSAlex Tomas 3485c9de560dSAlex Tomas /* the smallest one defines real window */ 3486c9de560dSAlex Tomas win = min(winl, wins); 3487c9de560dSAlex Tomas 3488c9de560dSAlex Tomas offs = ac->ac_o_ex.fe_logical % ac->ac_b_ex.fe_len; 3489c9de560dSAlex Tomas if (offs && offs < win) 3490c9de560dSAlex Tomas win = offs; 3491c9de560dSAlex Tomas 3492c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical - win; 3493c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical); 3494c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len); 3495c9de560dSAlex Tomas } 3496c9de560dSAlex Tomas 3497c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3498c9de560dSAlex Tomas * allocated blocks for history */ 3499c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3500c9de560dSAlex Tomas 3501c9de560dSAlex Tomas pa->pa_lstart = ac->ac_b_ex.fe_logical; 3502c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3503c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3504c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3505c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3506c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 3507d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3508d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 3509c9de560dSAlex Tomas pa->pa_deleted = 0; 3510cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_INODE_PA; 3511c9de560dSAlex Tomas 35126ba495e9STheodore Ts'o mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa, 3513c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 35149bffad1eSTheodore Ts'o trace_ext4_mb_new_inode_pa(ac, pa); 3515c9de560dSAlex Tomas 3516c9de560dSAlex Tomas ext4_mb_use_inode_pa(ac, pa); 3517c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3518c9de560dSAlex Tomas 3519c9de560dSAlex Tomas ei = EXT4_I(ac->ac_inode); 3520c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3521c9de560dSAlex Tomas 3522c9de560dSAlex Tomas pa->pa_obj_lock = &ei->i_prealloc_lock; 3523c9de560dSAlex Tomas pa->pa_inode = ac->ac_inode; 3524c9de560dSAlex Tomas 3525c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3526c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3527c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3528c9de560dSAlex Tomas 3529c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3530c9de560dSAlex Tomas list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list); 3531c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3532c9de560dSAlex Tomas 3533c9de560dSAlex Tomas return 0; 3534c9de560dSAlex Tomas } 3535c9de560dSAlex Tomas 3536c9de560dSAlex Tomas /* 3537c9de560dSAlex Tomas * creates new preallocated space for locality group inodes belongs to 3538c9de560dSAlex Tomas */ 35394ddfef7bSEric Sandeen static noinline_for_stack int 35404ddfef7bSEric Sandeen ext4_mb_new_group_pa(struct ext4_allocation_context *ac) 3541c9de560dSAlex Tomas { 3542c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 3543c9de560dSAlex Tomas struct ext4_locality_group *lg; 3544c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3545c9de560dSAlex Tomas struct ext4_group_info *grp; 3546c9de560dSAlex Tomas 3547c9de560dSAlex Tomas /* preallocate only when found space is larger then requested */ 3548c9de560dSAlex Tomas BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len); 3549c9de560dSAlex Tomas BUG_ON(ac->ac_status != AC_STATUS_FOUND); 3550c9de560dSAlex Tomas BUG_ON(!S_ISREG(ac->ac_inode->i_mode)); 3551c9de560dSAlex Tomas 3552c9de560dSAlex Tomas BUG_ON(ext4_pspace_cachep == NULL); 3553c9de560dSAlex Tomas pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS); 3554c9de560dSAlex Tomas if (pa == NULL) 3555c9de560dSAlex Tomas return -ENOMEM; 3556c9de560dSAlex Tomas 3557c9de560dSAlex Tomas /* preallocation can change ac_b_ex, thus we store actually 3558c9de560dSAlex Tomas * allocated blocks for history */ 3559c9de560dSAlex Tomas ac->ac_f_ex = ac->ac_b_ex; 3560c9de560dSAlex Tomas 3561c9de560dSAlex Tomas pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 3562c9de560dSAlex Tomas pa->pa_lstart = pa->pa_pstart; 3563c9de560dSAlex Tomas pa->pa_len = ac->ac_b_ex.fe_len; 3564c9de560dSAlex Tomas pa->pa_free = pa->pa_len; 3565c9de560dSAlex Tomas atomic_set(&pa->pa_count, 1); 3566c9de560dSAlex Tomas spin_lock_init(&pa->pa_lock); 35676be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_inode_list); 3568d794bf8eSAneesh Kumar K.V INIT_LIST_HEAD(&pa->pa_group_list); 3569c9de560dSAlex Tomas pa->pa_deleted = 0; 3570cc0fb9adSAneesh Kumar K.V pa->pa_type = MB_GROUP_PA; 3571c9de560dSAlex Tomas 35726ba495e9STheodore Ts'o mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa, 3573c9de560dSAlex Tomas pa->pa_pstart, pa->pa_len, pa->pa_lstart); 35749bffad1eSTheodore Ts'o trace_ext4_mb_new_group_pa(ac, pa); 3575c9de560dSAlex Tomas 3576c9de560dSAlex Tomas ext4_mb_use_group_pa(ac, pa); 3577c9de560dSAlex Tomas atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated); 3578c9de560dSAlex Tomas 3579c9de560dSAlex Tomas grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group); 3580c9de560dSAlex Tomas lg = ac->ac_lg; 3581c9de560dSAlex Tomas BUG_ON(lg == NULL); 3582c9de560dSAlex Tomas 3583c9de560dSAlex Tomas pa->pa_obj_lock = &lg->lg_prealloc_lock; 3584c9de560dSAlex Tomas pa->pa_inode = NULL; 3585c9de560dSAlex Tomas 3586c9de560dSAlex Tomas ext4_lock_group(sb, ac->ac_b_ex.fe_group); 3587c9de560dSAlex Tomas list_add(&pa->pa_group_list, &grp->bb_prealloc_list); 3588c9de560dSAlex Tomas ext4_unlock_group(sb, ac->ac_b_ex.fe_group); 3589c9de560dSAlex Tomas 35906be2ded1SAneesh Kumar K.V /* 35916be2ded1SAneesh Kumar K.V * We will later add the new pa to the right bucket 35926be2ded1SAneesh Kumar K.V * after updating the pa_free in ext4_mb_release_context 35936be2ded1SAneesh Kumar K.V */ 3594c9de560dSAlex Tomas return 0; 3595c9de560dSAlex Tomas } 3596c9de560dSAlex Tomas 3597c9de560dSAlex Tomas static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac) 3598c9de560dSAlex Tomas { 3599c9de560dSAlex Tomas int err; 3600c9de560dSAlex Tomas 3601c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 3602c9de560dSAlex Tomas err = ext4_mb_new_group_pa(ac); 3603c9de560dSAlex Tomas else 3604c9de560dSAlex Tomas err = ext4_mb_new_inode_pa(ac); 3605c9de560dSAlex Tomas return err; 3606c9de560dSAlex Tomas } 3607c9de560dSAlex Tomas 3608c9de560dSAlex Tomas /* 3609c9de560dSAlex Tomas * finds all unused blocks in on-disk bitmap, frees them in 3610c9de560dSAlex Tomas * in-core bitmap and buddy. 3611c9de560dSAlex Tomas * @pa must be unlinked from inode and group lists, so that 3612c9de560dSAlex Tomas * nobody else can find/use it. 3613c9de560dSAlex Tomas * the caller MUST hold group/inode locks. 3614c9de560dSAlex Tomas * TODO: optimize the case when there are no in-core structures yet 3615c9de560dSAlex Tomas */ 36164ddfef7bSEric Sandeen static noinline_for_stack int 36174ddfef7bSEric Sandeen ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh, 36183e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 3619c9de560dSAlex Tomas { 3620c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3621c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 3622498e5f24STheodore Ts'o unsigned int end; 3623498e5f24STheodore Ts'o unsigned int next; 3624c9de560dSAlex Tomas ext4_group_t group; 3625c9de560dSAlex Tomas ext4_grpblk_t bit; 3626ba80b101STheodore Ts'o unsigned long long grp_blk_start; 3627c9de560dSAlex Tomas int err = 0; 3628c9de560dSAlex Tomas int free = 0; 3629c9de560dSAlex Tomas 3630c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3631c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3632ba80b101STheodore Ts'o grp_blk_start = pa->pa_pstart - bit; 3633c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3634c9de560dSAlex Tomas end = bit + pa->pa_len; 3635c9de560dSAlex Tomas 3636c9de560dSAlex Tomas while (bit < end) { 3637ffad0a44SAneesh Kumar K.V bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit); 3638c9de560dSAlex Tomas if (bit >= end) 3639c9de560dSAlex Tomas break; 3640ffad0a44SAneesh Kumar K.V next = mb_find_next_bit(bitmap_bh->b_data, end, bit); 36416ba495e9STheodore Ts'o mb_debug(1, " free preallocated %u/%u in group %u\n", 36425a0790c2SAndi Kleen (unsigned) ext4_group_first_block_no(sb, group) + bit, 36435a0790c2SAndi Kleen (unsigned) next - bit, (unsigned) group); 3644c9de560dSAlex Tomas free += next - bit; 3645c9de560dSAlex Tomas 36463e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit); 36473e1e5f50SEric Sandeen trace_ext4_mb_release_inode_pa(sb, pa->pa_inode, pa, 36483e1e5f50SEric Sandeen grp_blk_start + bit, next - bit); 3649c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, next - bit); 3650c9de560dSAlex Tomas bit = next + 1; 3651c9de560dSAlex Tomas } 3652c9de560dSAlex Tomas if (free != pa->pa_free) { 365326346ff6SAneesh Kumar K.V printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n", 3654c9de560dSAlex Tomas pa, (unsigned long) pa->pa_lstart, 3655c9de560dSAlex Tomas (unsigned long) pa->pa_pstart, 3656c9de560dSAlex Tomas (unsigned long) pa->pa_len); 3657e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u", 365826346ff6SAneesh Kumar K.V free, pa->pa_free); 3659e56eb659SAneesh Kumar K.V /* 3660e56eb659SAneesh Kumar K.V * pa is already deleted so we use the value obtained 3661e56eb659SAneesh Kumar K.V * from the bitmap and continue. 3662e56eb659SAneesh Kumar K.V */ 3663c9de560dSAlex Tomas } 3664c9de560dSAlex Tomas atomic_add(free, &sbi->s_mb_discarded); 3665c9de560dSAlex Tomas 3666c9de560dSAlex Tomas return err; 3667c9de560dSAlex Tomas } 3668c9de560dSAlex Tomas 36694ddfef7bSEric Sandeen static noinline_for_stack int 36704ddfef7bSEric Sandeen ext4_mb_release_group_pa(struct ext4_buddy *e4b, 36713e1e5f50SEric Sandeen struct ext4_prealloc_space *pa) 3672c9de560dSAlex Tomas { 3673c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 3674c9de560dSAlex Tomas ext4_group_t group; 3675c9de560dSAlex Tomas ext4_grpblk_t bit; 3676c9de560dSAlex Tomas 36773e1e5f50SEric Sandeen trace_ext4_mb_release_group_pa(sb, pa); 3678c9de560dSAlex Tomas BUG_ON(pa->pa_deleted == 0); 3679c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit); 3680c9de560dSAlex Tomas BUG_ON(group != e4b->bd_group && pa->pa_len != 0); 3681c9de560dSAlex Tomas mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len); 3682c9de560dSAlex Tomas atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded); 36833e1e5f50SEric Sandeen trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len); 3684c9de560dSAlex Tomas 3685c9de560dSAlex Tomas return 0; 3686c9de560dSAlex Tomas } 3687c9de560dSAlex Tomas 3688c9de560dSAlex Tomas /* 3689c9de560dSAlex Tomas * releases all preallocations in given group 3690c9de560dSAlex Tomas * 3691c9de560dSAlex Tomas * first, we need to decide discard policy: 3692c9de560dSAlex Tomas * - when do we discard 3693c9de560dSAlex Tomas * 1) ENOSPC 3694c9de560dSAlex Tomas * - how many do we discard 3695c9de560dSAlex Tomas * 1) how many requested 3696c9de560dSAlex Tomas */ 36974ddfef7bSEric Sandeen static noinline_for_stack int 36984ddfef7bSEric Sandeen ext4_mb_discard_group_preallocations(struct super_block *sb, 3699c9de560dSAlex Tomas ext4_group_t group, int needed) 3700c9de560dSAlex Tomas { 3701c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, group); 3702c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3703c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3704c9de560dSAlex Tomas struct list_head list; 3705c9de560dSAlex Tomas struct ext4_buddy e4b; 3706c9de560dSAlex Tomas int err; 3707c9de560dSAlex Tomas int busy = 0; 3708c9de560dSAlex Tomas int free = 0; 3709c9de560dSAlex Tomas 37106ba495e9STheodore Ts'o mb_debug(1, "discard preallocation for group %u\n", group); 3711c9de560dSAlex Tomas 3712c9de560dSAlex Tomas if (list_empty(&grp->bb_prealloc_list)) 3713c9de560dSAlex Tomas return 0; 3714c9de560dSAlex Tomas 3715574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 3716c9de560dSAlex Tomas if (bitmap_bh == NULL) { 371712062dddSEric Sandeen ext4_error(sb, "Error reading block bitmap for %u", group); 3718ce89f46cSAneesh Kumar K.V return 0; 3719c9de560dSAlex Tomas } 3720c9de560dSAlex Tomas 3721c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3722ce89f46cSAneesh Kumar K.V if (err) { 372312062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", group); 3724ce89f46cSAneesh Kumar K.V put_bh(bitmap_bh); 3725ce89f46cSAneesh Kumar K.V return 0; 3726ce89f46cSAneesh Kumar K.V } 3727c9de560dSAlex Tomas 3728c9de560dSAlex Tomas if (needed == 0) 3729c9de560dSAlex Tomas needed = EXT4_BLOCKS_PER_GROUP(sb) + 1; 3730c9de560dSAlex Tomas 3731c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3732c9de560dSAlex Tomas repeat: 3733c9de560dSAlex Tomas ext4_lock_group(sb, group); 3734c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, 3735c9de560dSAlex Tomas &grp->bb_prealloc_list, pa_group_list) { 3736c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3737c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3738c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3739c9de560dSAlex Tomas busy = 1; 3740c9de560dSAlex Tomas continue; 3741c9de560dSAlex Tomas } 3742c9de560dSAlex Tomas if (pa->pa_deleted) { 3743c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3744c9de560dSAlex Tomas continue; 3745c9de560dSAlex Tomas } 3746c9de560dSAlex Tomas 3747c9de560dSAlex Tomas /* seems this one can be freed ... */ 3748c9de560dSAlex Tomas pa->pa_deleted = 1; 3749c9de560dSAlex Tomas 3750c9de560dSAlex Tomas /* we can trust pa_free ... */ 3751c9de560dSAlex Tomas free += pa->pa_free; 3752c9de560dSAlex Tomas 3753c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3754c9de560dSAlex Tomas 3755c9de560dSAlex Tomas list_del(&pa->pa_group_list); 3756c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3757c9de560dSAlex Tomas } 3758c9de560dSAlex Tomas 3759c9de560dSAlex Tomas /* if we still need more blocks and some PAs were used, try again */ 3760c9de560dSAlex Tomas if (free < needed && busy) { 3761c9de560dSAlex Tomas busy = 0; 3762c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3763c9de560dSAlex Tomas /* 3764c9de560dSAlex Tomas * Yield the CPU here so that we don't get soft lockup 3765c9de560dSAlex Tomas * in non preempt case. 3766c9de560dSAlex Tomas */ 3767c9de560dSAlex Tomas yield(); 3768c9de560dSAlex Tomas goto repeat; 3769c9de560dSAlex Tomas } 3770c9de560dSAlex Tomas 3771c9de560dSAlex Tomas /* found anything to free? */ 3772c9de560dSAlex Tomas if (list_empty(&list)) { 3773c9de560dSAlex Tomas BUG_ON(free != 0); 3774c9de560dSAlex Tomas goto out; 3775c9de560dSAlex Tomas } 3776c9de560dSAlex Tomas 3777c9de560dSAlex Tomas /* now free all selected PAs */ 3778c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3779c9de560dSAlex Tomas 3780c9de560dSAlex Tomas /* remove from object (inode or locality group) */ 3781c9de560dSAlex Tomas spin_lock(pa->pa_obj_lock); 3782c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3783c9de560dSAlex Tomas spin_unlock(pa->pa_obj_lock); 3784c9de560dSAlex Tomas 3785cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) 37863e1e5f50SEric Sandeen ext4_mb_release_group_pa(&e4b, pa); 3787c9de560dSAlex Tomas else 37883e1e5f50SEric Sandeen ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 3789c9de560dSAlex Tomas 3790c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 3791c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3792c9de560dSAlex Tomas } 3793c9de560dSAlex Tomas 3794c9de560dSAlex Tomas out: 3795c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3796e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3797c9de560dSAlex Tomas put_bh(bitmap_bh); 3798c9de560dSAlex Tomas return free; 3799c9de560dSAlex Tomas } 3800c9de560dSAlex Tomas 3801c9de560dSAlex Tomas /* 3802c9de560dSAlex Tomas * releases all non-used preallocated blocks for given inode 3803c9de560dSAlex Tomas * 3804c9de560dSAlex Tomas * It's important to discard preallocations under i_data_sem 3805c9de560dSAlex Tomas * We don't want another block to be served from the prealloc 3806c9de560dSAlex Tomas * space when we are discarding the inode prealloc space. 3807c9de560dSAlex Tomas * 3808c9de560dSAlex Tomas * FIXME!! Make sure it is valid at all the call sites 3809c9de560dSAlex Tomas */ 3810c2ea3fdeSTheodore Ts'o void ext4_discard_preallocations(struct inode *inode) 3811c9de560dSAlex Tomas { 3812c9de560dSAlex Tomas struct ext4_inode_info *ei = EXT4_I(inode); 3813c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 3814c9de560dSAlex Tomas struct buffer_head *bitmap_bh = NULL; 3815c9de560dSAlex Tomas struct ext4_prealloc_space *pa, *tmp; 3816c9de560dSAlex Tomas ext4_group_t group = 0; 3817c9de560dSAlex Tomas struct list_head list; 3818c9de560dSAlex Tomas struct ext4_buddy e4b; 3819c9de560dSAlex Tomas int err; 3820c9de560dSAlex Tomas 3821c2ea3fdeSTheodore Ts'o if (!S_ISREG(inode->i_mode)) { 3822c9de560dSAlex Tomas /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/ 3823c9de560dSAlex Tomas return; 3824c9de560dSAlex Tomas } 3825c9de560dSAlex Tomas 38266ba495e9STheodore Ts'o mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino); 38279bffad1eSTheodore Ts'o trace_ext4_discard_preallocations(inode); 3828c9de560dSAlex Tomas 3829c9de560dSAlex Tomas INIT_LIST_HEAD(&list); 3830c9de560dSAlex Tomas 3831c9de560dSAlex Tomas repeat: 3832c9de560dSAlex Tomas /* first, collect all pa's in the inode */ 3833c9de560dSAlex Tomas spin_lock(&ei->i_prealloc_lock); 3834c9de560dSAlex Tomas while (!list_empty(&ei->i_prealloc_list)) { 3835c9de560dSAlex Tomas pa = list_entry(ei->i_prealloc_list.next, 3836c9de560dSAlex Tomas struct ext4_prealloc_space, pa_inode_list); 3837c9de560dSAlex Tomas BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock); 3838c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3839c9de560dSAlex Tomas if (atomic_read(&pa->pa_count)) { 3840c9de560dSAlex Tomas /* this shouldn't happen often - nobody should 3841c9de560dSAlex Tomas * use preallocation while we're discarding it */ 3842c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3843c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3844c9de560dSAlex Tomas printk(KERN_ERR "uh-oh! used pa while discarding\n"); 3845c9de560dSAlex Tomas WARN_ON(1); 3846c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 3847c9de560dSAlex Tomas goto repeat; 3848c9de560dSAlex Tomas 3849c9de560dSAlex Tomas } 3850c9de560dSAlex Tomas if (pa->pa_deleted == 0) { 3851c9de560dSAlex Tomas pa->pa_deleted = 1; 3852c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3853c9de560dSAlex Tomas list_del_rcu(&pa->pa_inode_list); 3854c9de560dSAlex Tomas list_add(&pa->u.pa_tmp_list, &list); 3855c9de560dSAlex Tomas continue; 3856c9de560dSAlex Tomas } 3857c9de560dSAlex Tomas 3858c9de560dSAlex Tomas /* someone is deleting pa right now */ 3859c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 3860c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3861c9de560dSAlex Tomas 3862c9de560dSAlex Tomas /* we have to wait here because pa_deleted 3863c9de560dSAlex Tomas * doesn't mean pa is already unlinked from 3864c9de560dSAlex Tomas * the list. as we might be called from 3865c9de560dSAlex Tomas * ->clear_inode() the inode will get freed 3866c9de560dSAlex Tomas * and concurrent thread which is unlinking 3867c9de560dSAlex Tomas * pa from inode's list may access already 3868c9de560dSAlex Tomas * freed memory, bad-bad-bad */ 3869c9de560dSAlex Tomas 3870c9de560dSAlex Tomas /* XXX: if this happens too often, we can 3871c9de560dSAlex Tomas * add a flag to force wait only in case 3872c9de560dSAlex Tomas * of ->clear_inode(), but not in case of 3873c9de560dSAlex Tomas * regular truncate */ 3874c9de560dSAlex Tomas schedule_timeout_uninterruptible(HZ); 3875c9de560dSAlex Tomas goto repeat; 3876c9de560dSAlex Tomas } 3877c9de560dSAlex Tomas spin_unlock(&ei->i_prealloc_lock); 3878c9de560dSAlex Tomas 3879c9de560dSAlex Tomas list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) { 3880cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_INODE_PA); 3881c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); 3882c9de560dSAlex Tomas 3883c9de560dSAlex Tomas err = ext4_mb_load_buddy(sb, group, &e4b); 3884ce89f46cSAneesh Kumar K.V if (err) { 388512062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", 388612062dddSEric Sandeen group); 3887ce89f46cSAneesh Kumar K.V continue; 3888ce89f46cSAneesh Kumar K.V } 3889c9de560dSAlex Tomas 3890574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, group); 3891c9de560dSAlex Tomas if (bitmap_bh == NULL) { 389212062dddSEric Sandeen ext4_error(sb, "Error reading block bitmap for %u", 389312062dddSEric Sandeen group); 3894e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3895ce89f46cSAneesh Kumar K.V continue; 3896c9de560dSAlex Tomas } 3897c9de560dSAlex Tomas 3898c9de560dSAlex Tomas ext4_lock_group(sb, group); 3899c9de560dSAlex Tomas list_del(&pa->pa_group_list); 39003e1e5f50SEric Sandeen ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa); 3901c9de560dSAlex Tomas ext4_unlock_group(sb, group); 3902c9de560dSAlex Tomas 3903e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 3904c9de560dSAlex Tomas put_bh(bitmap_bh); 3905c9de560dSAlex Tomas 3906c9de560dSAlex Tomas list_del(&pa->u.pa_tmp_list); 3907c9de560dSAlex Tomas call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 3908c9de560dSAlex Tomas } 3909c9de560dSAlex Tomas } 3910c9de560dSAlex Tomas 39116ba495e9STheodore Ts'o #ifdef CONFIG_EXT4_DEBUG 3912c9de560dSAlex Tomas static void ext4_mb_show_ac(struct ext4_allocation_context *ac) 3913c9de560dSAlex Tomas { 3914c9de560dSAlex Tomas struct super_block *sb = ac->ac_sb; 39158df9675fSTheodore Ts'o ext4_group_t ngroups, i; 3916c9de560dSAlex Tomas 39174dd89fc6STheodore Ts'o if (!mb_enable_debug || 39184dd89fc6STheodore Ts'o (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) 3919e3570639SEric Sandeen return; 3920e3570639SEric Sandeen 3921c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: Can't allocate:" 3922c9de560dSAlex Tomas " Allocation context details:\n"); 3923c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: status %d flags %d\n", 3924c9de560dSAlex Tomas ac->ac_status, ac->ac_flags); 3925c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: orig %lu/%lu/%lu@%lu, goal %lu/%lu/%lu@%lu, " 3926c9de560dSAlex Tomas "best %lu/%lu/%lu@%lu cr %d\n", 3927c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_group, 3928c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_start, 3929c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_len, 3930c9de560dSAlex Tomas (unsigned long)ac->ac_o_ex.fe_logical, 3931c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_group, 3932c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_start, 3933c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_len, 3934c9de560dSAlex Tomas (unsigned long)ac->ac_g_ex.fe_logical, 3935c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_group, 3936c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_start, 3937c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_len, 3938c9de560dSAlex Tomas (unsigned long)ac->ac_b_ex.fe_logical, 3939c9de560dSAlex Tomas (int)ac->ac_criteria); 3940c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: %lu scanned, %d found\n", ac->ac_ex_scanned, 3941c9de560dSAlex Tomas ac->ac_found); 3942c9de560dSAlex Tomas printk(KERN_ERR "EXT4-fs: groups: \n"); 39438df9675fSTheodore Ts'o ngroups = ext4_get_groups_count(sb); 39448df9675fSTheodore Ts'o for (i = 0; i < ngroups; i++) { 3945c9de560dSAlex Tomas struct ext4_group_info *grp = ext4_get_group_info(sb, i); 3946c9de560dSAlex Tomas struct ext4_prealloc_space *pa; 3947c9de560dSAlex Tomas ext4_grpblk_t start; 3948c9de560dSAlex Tomas struct list_head *cur; 3949c9de560dSAlex Tomas ext4_lock_group(sb, i); 3950c9de560dSAlex Tomas list_for_each(cur, &grp->bb_prealloc_list) { 3951c9de560dSAlex Tomas pa = list_entry(cur, struct ext4_prealloc_space, 3952c9de560dSAlex Tomas pa_group_list); 3953c9de560dSAlex Tomas spin_lock(&pa->pa_lock); 3954c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, pa->pa_pstart, 3955c9de560dSAlex Tomas NULL, &start); 3956c9de560dSAlex Tomas spin_unlock(&pa->pa_lock); 39571c718505SAkira Fujita printk(KERN_ERR "PA:%u:%d:%u \n", i, 3958c9de560dSAlex Tomas start, pa->pa_len); 3959c9de560dSAlex Tomas } 396060bd63d1SSolofo Ramangalahy ext4_unlock_group(sb, i); 3961c9de560dSAlex Tomas 3962c9de560dSAlex Tomas if (grp->bb_free == 0) 3963c9de560dSAlex Tomas continue; 39641c718505SAkira Fujita printk(KERN_ERR "%u: %d/%d \n", 3965c9de560dSAlex Tomas i, grp->bb_free, grp->bb_fragments); 3966c9de560dSAlex Tomas } 3967c9de560dSAlex Tomas printk(KERN_ERR "\n"); 3968c9de560dSAlex Tomas } 3969c9de560dSAlex Tomas #else 3970c9de560dSAlex Tomas static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac) 3971c9de560dSAlex Tomas { 3972c9de560dSAlex Tomas return; 3973c9de560dSAlex Tomas } 3974c9de560dSAlex Tomas #endif 3975c9de560dSAlex Tomas 3976c9de560dSAlex Tomas /* 3977c9de560dSAlex Tomas * We use locality group preallocation for small size file. The size of the 3978c9de560dSAlex Tomas * file is determined by the current size or the resulting size after 3979c9de560dSAlex Tomas * allocation which ever is larger 3980c9de560dSAlex Tomas * 3981b713a5ecSTheodore Ts'o * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req 3982c9de560dSAlex Tomas */ 3983c9de560dSAlex Tomas static void ext4_mb_group_or_file(struct ext4_allocation_context *ac) 3984c9de560dSAlex Tomas { 3985c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); 3986c9de560dSAlex Tomas int bsbits = ac->ac_sb->s_blocksize_bits; 3987c9de560dSAlex Tomas loff_t size, isize; 3988c9de560dSAlex Tomas 3989c9de560dSAlex Tomas if (!(ac->ac_flags & EXT4_MB_HINT_DATA)) 3990c9de560dSAlex Tomas return; 3991c9de560dSAlex Tomas 39924ba74d00STheodore Ts'o if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY)) 39934ba74d00STheodore Ts'o return; 39944ba74d00STheodore Ts'o 3995c9de560dSAlex Tomas size = ac->ac_o_ex.fe_logical + ac->ac_o_ex.fe_len; 399650797481STheodore Ts'o isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1) 399750797481STheodore Ts'o >> bsbits; 3998c9de560dSAlex Tomas 399950797481STheodore Ts'o if ((size == isize) && 400050797481STheodore Ts'o !ext4_fs_is_busy(sbi) && 400150797481STheodore Ts'o (atomic_read(&ac->ac_inode->i_writecount) == 0)) { 400250797481STheodore Ts'o ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC; 400350797481STheodore Ts'o return; 400450797481STheodore Ts'o } 400550797481STheodore Ts'o 4006c9de560dSAlex Tomas /* don't use group allocation for large files */ 400771780577STheodore Ts'o size = max(size, isize); 4008cc483f10STao Ma if (size > sbi->s_mb_stream_request) { 40094ba74d00STheodore Ts'o ac->ac_flags |= EXT4_MB_STREAM_ALLOC; 4010c9de560dSAlex Tomas return; 40114ba74d00STheodore Ts'o } 4012c9de560dSAlex Tomas 4013c9de560dSAlex Tomas BUG_ON(ac->ac_lg != NULL); 4014c9de560dSAlex Tomas /* 4015c9de560dSAlex Tomas * locality group prealloc space are per cpu. The reason for having 4016c9de560dSAlex Tomas * per cpu locality group is to reduce the contention between block 4017c9de560dSAlex Tomas * request from multiple CPUs. 4018c9de560dSAlex Tomas */ 4019ca0c9584SChristoph Lameter ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups); 4020c9de560dSAlex Tomas 4021c9de560dSAlex Tomas /* we're going to use group allocation */ 4022c9de560dSAlex Tomas ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC; 4023c9de560dSAlex Tomas 4024c9de560dSAlex Tomas /* serialize all allocations in the group */ 4025c9de560dSAlex Tomas mutex_lock(&ac->ac_lg->lg_mutex); 4026c9de560dSAlex Tomas } 4027c9de560dSAlex Tomas 40284ddfef7bSEric Sandeen static noinline_for_stack int 40294ddfef7bSEric Sandeen ext4_mb_initialize_context(struct ext4_allocation_context *ac, 4030c9de560dSAlex Tomas struct ext4_allocation_request *ar) 4031c9de560dSAlex Tomas { 4032c9de560dSAlex Tomas struct super_block *sb = ar->inode->i_sb; 4033c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 4034c9de560dSAlex Tomas struct ext4_super_block *es = sbi->s_es; 4035c9de560dSAlex Tomas ext4_group_t group; 4036498e5f24STheodore Ts'o unsigned int len; 4037498e5f24STheodore Ts'o ext4_fsblk_t goal; 4038c9de560dSAlex Tomas ext4_grpblk_t block; 4039c9de560dSAlex Tomas 4040c9de560dSAlex Tomas /* we can't allocate > group size */ 4041c9de560dSAlex Tomas len = ar->len; 4042c9de560dSAlex Tomas 4043c9de560dSAlex Tomas /* just a dirty hack to filter too big requests */ 4044c9de560dSAlex Tomas if (len >= EXT4_BLOCKS_PER_GROUP(sb) - 10) 4045c9de560dSAlex Tomas len = EXT4_BLOCKS_PER_GROUP(sb) - 10; 4046c9de560dSAlex Tomas 4047c9de560dSAlex Tomas /* start searching from the goal */ 4048c9de560dSAlex Tomas goal = ar->goal; 4049c9de560dSAlex Tomas if (goal < le32_to_cpu(es->s_first_data_block) || 4050c9de560dSAlex Tomas goal >= ext4_blocks_count(es)) 4051c9de560dSAlex Tomas goal = le32_to_cpu(es->s_first_data_block); 4052c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, goal, &group, &block); 4053c9de560dSAlex Tomas 4054c9de560dSAlex Tomas /* set up allocation goals */ 4055833576b3STheodore Ts'o memset(ac, 0, sizeof(struct ext4_allocation_context)); 4056c9de560dSAlex Tomas ac->ac_b_ex.fe_logical = ar->logical; 4057c9de560dSAlex Tomas ac->ac_status = AC_STATUS_CONTINUE; 4058c9de560dSAlex Tomas ac->ac_sb = sb; 4059c9de560dSAlex Tomas ac->ac_inode = ar->inode; 4060c9de560dSAlex Tomas ac->ac_o_ex.fe_logical = ar->logical; 4061c9de560dSAlex Tomas ac->ac_o_ex.fe_group = group; 4062c9de560dSAlex Tomas ac->ac_o_ex.fe_start = block; 4063c9de560dSAlex Tomas ac->ac_o_ex.fe_len = len; 4064c9de560dSAlex Tomas ac->ac_g_ex.fe_logical = ar->logical; 4065c9de560dSAlex Tomas ac->ac_g_ex.fe_group = group; 4066c9de560dSAlex Tomas ac->ac_g_ex.fe_start = block; 4067c9de560dSAlex Tomas ac->ac_g_ex.fe_len = len; 4068c9de560dSAlex Tomas ac->ac_flags = ar->flags; 4069c9de560dSAlex Tomas 4070c9de560dSAlex Tomas /* we have to define context: we'll we work with a file or 4071c9de560dSAlex Tomas * locality group. this is a policy, actually */ 4072c9de560dSAlex Tomas ext4_mb_group_or_file(ac); 4073c9de560dSAlex Tomas 40746ba495e9STheodore Ts'o mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, " 4075c9de560dSAlex Tomas "left: %u/%u, right %u/%u to %swritable\n", 4076c9de560dSAlex Tomas (unsigned) ar->len, (unsigned) ar->logical, 4077c9de560dSAlex Tomas (unsigned) ar->goal, ac->ac_flags, ac->ac_2order, 4078c9de560dSAlex Tomas (unsigned) ar->lleft, (unsigned) ar->pleft, 4079c9de560dSAlex Tomas (unsigned) ar->lright, (unsigned) ar->pright, 4080c9de560dSAlex Tomas atomic_read(&ar->inode->i_writecount) ? "" : "non-"); 4081c9de560dSAlex Tomas return 0; 4082c9de560dSAlex Tomas 4083c9de560dSAlex Tomas } 4084c9de560dSAlex Tomas 40856be2ded1SAneesh Kumar K.V static noinline_for_stack void 40866be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(struct super_block *sb, 40876be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg, 40886be2ded1SAneesh Kumar K.V int order, int total_entries) 40896be2ded1SAneesh Kumar K.V { 40906be2ded1SAneesh Kumar K.V ext4_group_t group = 0; 40916be2ded1SAneesh Kumar K.V struct ext4_buddy e4b; 40926be2ded1SAneesh Kumar K.V struct list_head discard_list; 40936be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa, *tmp; 40946be2ded1SAneesh Kumar K.V 40956ba495e9STheodore Ts'o mb_debug(1, "discard locality group preallocation\n"); 40966be2ded1SAneesh Kumar K.V 40976be2ded1SAneesh Kumar K.V INIT_LIST_HEAD(&discard_list); 40986be2ded1SAneesh Kumar K.V 40996be2ded1SAneesh Kumar K.V spin_lock(&lg->lg_prealloc_lock); 41006be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order], 41016be2ded1SAneesh Kumar K.V pa_inode_list) { 41026be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 41036be2ded1SAneesh Kumar K.V if (atomic_read(&pa->pa_count)) { 41046be2ded1SAneesh Kumar K.V /* 41056be2ded1SAneesh Kumar K.V * This is the pa that we just used 41066be2ded1SAneesh Kumar K.V * for block allocation. So don't 41076be2ded1SAneesh Kumar K.V * free that 41086be2ded1SAneesh Kumar K.V */ 41096be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 41106be2ded1SAneesh Kumar K.V continue; 41116be2ded1SAneesh Kumar K.V } 41126be2ded1SAneesh Kumar K.V if (pa->pa_deleted) { 41136be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 41146be2ded1SAneesh Kumar K.V continue; 41156be2ded1SAneesh Kumar K.V } 41166be2ded1SAneesh Kumar K.V /* only lg prealloc space */ 4117cc0fb9adSAneesh Kumar K.V BUG_ON(pa->pa_type != MB_GROUP_PA); 41186be2ded1SAneesh Kumar K.V 41196be2ded1SAneesh Kumar K.V /* seems this one can be freed ... */ 41206be2ded1SAneesh Kumar K.V pa->pa_deleted = 1; 41216be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 41226be2ded1SAneesh Kumar K.V 41236be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 41246be2ded1SAneesh Kumar K.V list_add(&pa->u.pa_tmp_list, &discard_list); 41256be2ded1SAneesh Kumar K.V 41266be2ded1SAneesh Kumar K.V total_entries--; 41276be2ded1SAneesh Kumar K.V if (total_entries <= 5) { 41286be2ded1SAneesh Kumar K.V /* 41296be2ded1SAneesh Kumar K.V * we want to keep only 5 entries 41306be2ded1SAneesh Kumar K.V * allowing it to grow to 8. This 41316be2ded1SAneesh Kumar K.V * mak sure we don't call discard 41326be2ded1SAneesh Kumar K.V * soon for this list. 41336be2ded1SAneesh Kumar K.V */ 41346be2ded1SAneesh Kumar K.V break; 41356be2ded1SAneesh Kumar K.V } 41366be2ded1SAneesh Kumar K.V } 41376be2ded1SAneesh Kumar K.V spin_unlock(&lg->lg_prealloc_lock); 41386be2ded1SAneesh Kumar K.V 41396be2ded1SAneesh Kumar K.V list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) { 41406be2ded1SAneesh Kumar K.V 41416be2ded1SAneesh Kumar K.V ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL); 41426be2ded1SAneesh Kumar K.V if (ext4_mb_load_buddy(sb, group, &e4b)) { 414312062dddSEric Sandeen ext4_error(sb, "Error loading buddy information for %u", 414412062dddSEric Sandeen group); 41456be2ded1SAneesh Kumar K.V continue; 41466be2ded1SAneesh Kumar K.V } 41476be2ded1SAneesh Kumar K.V ext4_lock_group(sb, group); 41486be2ded1SAneesh Kumar K.V list_del(&pa->pa_group_list); 41493e1e5f50SEric Sandeen ext4_mb_release_group_pa(&e4b, pa); 41506be2ded1SAneesh Kumar K.V ext4_unlock_group(sb, group); 41516be2ded1SAneesh Kumar K.V 4152e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 41536be2ded1SAneesh Kumar K.V list_del(&pa->u.pa_tmp_list); 41546be2ded1SAneesh Kumar K.V call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback); 41556be2ded1SAneesh Kumar K.V } 41566be2ded1SAneesh Kumar K.V } 41576be2ded1SAneesh Kumar K.V 41586be2ded1SAneesh Kumar K.V /* 41596be2ded1SAneesh Kumar K.V * We have incremented pa_count. So it cannot be freed at this 41606be2ded1SAneesh Kumar K.V * point. Also we hold lg_mutex. So no parallel allocation is 41616be2ded1SAneesh Kumar K.V * possible from this lg. That means pa_free cannot be updated. 41626be2ded1SAneesh Kumar K.V * 41636be2ded1SAneesh Kumar K.V * A parallel ext4_mb_discard_group_preallocations is possible. 41646be2ded1SAneesh Kumar K.V * which can cause the lg_prealloc_list to be updated. 41656be2ded1SAneesh Kumar K.V */ 41666be2ded1SAneesh Kumar K.V 41676be2ded1SAneesh Kumar K.V static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac) 41686be2ded1SAneesh Kumar K.V { 41696be2ded1SAneesh Kumar K.V int order, added = 0, lg_prealloc_count = 1; 41706be2ded1SAneesh Kumar K.V struct super_block *sb = ac->ac_sb; 41716be2ded1SAneesh Kumar K.V struct ext4_locality_group *lg = ac->ac_lg; 41726be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa; 41736be2ded1SAneesh Kumar K.V 41746be2ded1SAneesh Kumar K.V order = fls(pa->pa_free) - 1; 41756be2ded1SAneesh Kumar K.V if (order > PREALLOC_TB_SIZE - 1) 41766be2ded1SAneesh Kumar K.V /* The max size of hash table is PREALLOC_TB_SIZE */ 41776be2ded1SAneesh Kumar K.V order = PREALLOC_TB_SIZE - 1; 41786be2ded1SAneesh Kumar K.V /* Add the prealloc space to lg */ 41796be2ded1SAneesh Kumar K.V rcu_read_lock(); 41806be2ded1SAneesh Kumar K.V list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order], 41816be2ded1SAneesh Kumar K.V pa_inode_list) { 41826be2ded1SAneesh Kumar K.V spin_lock(&tmp_pa->pa_lock); 41836be2ded1SAneesh Kumar K.V if (tmp_pa->pa_deleted) { 4184e7c9e3e9STheodore Ts'o spin_unlock(&tmp_pa->pa_lock); 41856be2ded1SAneesh Kumar K.V continue; 41866be2ded1SAneesh Kumar K.V } 41876be2ded1SAneesh Kumar K.V if (!added && pa->pa_free < tmp_pa->pa_free) { 41886be2ded1SAneesh Kumar K.V /* Add to the tail of the previous entry */ 41896be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 41906be2ded1SAneesh Kumar K.V &tmp_pa->pa_inode_list); 41916be2ded1SAneesh Kumar K.V added = 1; 41926be2ded1SAneesh Kumar K.V /* 41936be2ded1SAneesh Kumar K.V * we want to count the total 41946be2ded1SAneesh Kumar K.V * number of entries in the list 41956be2ded1SAneesh Kumar K.V */ 41966be2ded1SAneesh Kumar K.V } 41976be2ded1SAneesh Kumar K.V spin_unlock(&tmp_pa->pa_lock); 41986be2ded1SAneesh Kumar K.V lg_prealloc_count++; 41996be2ded1SAneesh Kumar K.V } 42006be2ded1SAneesh Kumar K.V if (!added) 42016be2ded1SAneesh Kumar K.V list_add_tail_rcu(&pa->pa_inode_list, 42026be2ded1SAneesh Kumar K.V &lg->lg_prealloc_list[order]); 42036be2ded1SAneesh Kumar K.V rcu_read_unlock(); 42046be2ded1SAneesh Kumar K.V 42056be2ded1SAneesh Kumar K.V /* Now trim the list to be not more than 8 elements */ 42066be2ded1SAneesh Kumar K.V if (lg_prealloc_count > 8) { 42076be2ded1SAneesh Kumar K.V ext4_mb_discard_lg_preallocations(sb, lg, 42086be2ded1SAneesh Kumar K.V order, lg_prealloc_count); 42096be2ded1SAneesh Kumar K.V return; 42106be2ded1SAneesh Kumar K.V } 42116be2ded1SAneesh Kumar K.V return ; 42126be2ded1SAneesh Kumar K.V } 42136be2ded1SAneesh Kumar K.V 4214c9de560dSAlex Tomas /* 4215c9de560dSAlex Tomas * release all resource we used in allocation 4216c9de560dSAlex Tomas */ 4217c9de560dSAlex Tomas static int ext4_mb_release_context(struct ext4_allocation_context *ac) 4218c9de560dSAlex Tomas { 42196be2ded1SAneesh Kumar K.V struct ext4_prealloc_space *pa = ac->ac_pa; 42206be2ded1SAneesh Kumar K.V if (pa) { 4221cc0fb9adSAneesh Kumar K.V if (pa->pa_type == MB_GROUP_PA) { 4222c9de560dSAlex Tomas /* see comment in ext4_mb_use_group_pa() */ 42236be2ded1SAneesh Kumar K.V spin_lock(&pa->pa_lock); 42246be2ded1SAneesh Kumar K.V pa->pa_pstart += ac->ac_b_ex.fe_len; 42256be2ded1SAneesh Kumar K.V pa->pa_lstart += ac->ac_b_ex.fe_len; 42266be2ded1SAneesh Kumar K.V pa->pa_free -= ac->ac_b_ex.fe_len; 42276be2ded1SAneesh Kumar K.V pa->pa_len -= ac->ac_b_ex.fe_len; 42286be2ded1SAneesh Kumar K.V spin_unlock(&pa->pa_lock); 4229ba443916SAneesh Kumar K.V } 4230ba443916SAneesh Kumar K.V } 4231ba443916SAneesh Kumar K.V if (ac->alloc_semp) 4232ba443916SAneesh Kumar K.V up_read(ac->alloc_semp); 4233ba443916SAneesh Kumar K.V if (pa) { 42346be2ded1SAneesh Kumar K.V /* 42356be2ded1SAneesh Kumar K.V * We want to add the pa to the right bucket. 42366be2ded1SAneesh Kumar K.V * Remove it from the list and while adding 42376be2ded1SAneesh Kumar K.V * make sure the list to which we are adding 4238ba443916SAneesh Kumar K.V * doesn't grow big. We need to release 4239ba443916SAneesh Kumar K.V * alloc_semp before calling ext4_mb_add_n_trim() 42406be2ded1SAneesh Kumar K.V */ 4241cc0fb9adSAneesh Kumar K.V if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) { 42426be2ded1SAneesh Kumar K.V spin_lock(pa->pa_obj_lock); 42436be2ded1SAneesh Kumar K.V list_del_rcu(&pa->pa_inode_list); 42446be2ded1SAneesh Kumar K.V spin_unlock(pa->pa_obj_lock); 42456be2ded1SAneesh Kumar K.V ext4_mb_add_n_trim(ac); 4246c9de560dSAlex Tomas } 42476be2ded1SAneesh Kumar K.V ext4_mb_put_pa(ac, ac->ac_sb, pa); 4248c9de560dSAlex Tomas } 4249c9de560dSAlex Tomas if (ac->ac_bitmap_page) 4250c9de560dSAlex Tomas page_cache_release(ac->ac_bitmap_page); 4251c9de560dSAlex Tomas if (ac->ac_buddy_page) 4252c9de560dSAlex Tomas page_cache_release(ac->ac_buddy_page); 4253c9de560dSAlex Tomas if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) 4254c9de560dSAlex Tomas mutex_unlock(&ac->ac_lg->lg_mutex); 4255c9de560dSAlex Tomas ext4_mb_collect_stats(ac); 4256c9de560dSAlex Tomas return 0; 4257c9de560dSAlex Tomas } 4258c9de560dSAlex Tomas 4259c9de560dSAlex Tomas static int ext4_mb_discard_preallocations(struct super_block *sb, int needed) 4260c9de560dSAlex Tomas { 42618df9675fSTheodore Ts'o ext4_group_t i, ngroups = ext4_get_groups_count(sb); 4262c9de560dSAlex Tomas int ret; 4263c9de560dSAlex Tomas int freed = 0; 4264c9de560dSAlex Tomas 42659bffad1eSTheodore Ts'o trace_ext4_mb_discard_preallocations(sb, needed); 42668df9675fSTheodore Ts'o for (i = 0; i < ngroups && needed > 0; i++) { 4267c9de560dSAlex Tomas ret = ext4_mb_discard_group_preallocations(sb, i, needed); 4268c9de560dSAlex Tomas freed += ret; 4269c9de560dSAlex Tomas needed -= ret; 4270c9de560dSAlex Tomas } 4271c9de560dSAlex Tomas 4272c9de560dSAlex Tomas return freed; 4273c9de560dSAlex Tomas } 4274c9de560dSAlex Tomas 4275c9de560dSAlex Tomas /* 4276c9de560dSAlex Tomas * Main entry point into mballoc to allocate blocks 4277c9de560dSAlex Tomas * it tries to use preallocation first, then falls back 4278c9de560dSAlex Tomas * to usual allocation 4279c9de560dSAlex Tomas */ 4280c9de560dSAlex Tomas ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle, 4281c9de560dSAlex Tomas struct ext4_allocation_request *ar, int *errp) 4282c9de560dSAlex Tomas { 42836bc6e63fSAneesh Kumar K.V int freed; 4284256bdb49SEric Sandeen struct ext4_allocation_context *ac = NULL; 4285c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4286c9de560dSAlex Tomas struct super_block *sb; 4287c9de560dSAlex Tomas ext4_fsblk_t block = 0; 428860e58e0fSMingming Cao unsigned int inquota = 0; 4289498e5f24STheodore Ts'o unsigned int reserv_blks = 0; 4290c9de560dSAlex Tomas 4291c9de560dSAlex Tomas sb = ar->inode->i_sb; 4292c9de560dSAlex Tomas sbi = EXT4_SB(sb); 4293c9de560dSAlex Tomas 42949bffad1eSTheodore Ts'o trace_ext4_request_blocks(ar); 4295ba80b101STheodore Ts'o 4296d2a17637SMingming Cao /* 429760e58e0fSMingming Cao * For delayed allocation, we could skip the ENOSPC and 429860e58e0fSMingming Cao * EDQUOT check, as blocks and quotas have been already 429960e58e0fSMingming Cao * reserved when data being copied into pagecache. 430060e58e0fSMingming Cao */ 4301f2321097STheodore Ts'o if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED)) 430260e58e0fSMingming Cao ar->flags |= EXT4_MB_DELALLOC_RESERVED; 430360e58e0fSMingming Cao else { 430460e58e0fSMingming Cao /* Without delayed allocation we need to verify 430560e58e0fSMingming Cao * there is enough free blocks to do block allocation 430660e58e0fSMingming Cao * and verify allocation doesn't exceed the quota limits. 4307d2a17637SMingming Cao */ 4308030ba6bcSAneesh Kumar K.V while (ar->len && ext4_claim_free_blocks(sbi, ar->len)) { 4309030ba6bcSAneesh Kumar K.V /* let others to free the space */ 4310030ba6bcSAneesh Kumar K.V yield(); 4311030ba6bcSAneesh Kumar K.V ar->len = ar->len >> 1; 4312030ba6bcSAneesh Kumar K.V } 4313030ba6bcSAneesh Kumar K.V if (!ar->len) { 431407031431SMingming Cao *errp = -ENOSPC; 431507031431SMingming Cao return 0; 431607031431SMingming Cao } 43176bc6e63fSAneesh Kumar K.V reserv_blks = ar->len; 43185dd4056dSChristoph Hellwig while (ar->len && dquot_alloc_block(ar->inode, ar->len)) { 4319c9de560dSAlex Tomas ar->flags |= EXT4_MB_HINT_NOPREALLOC; 4320c9de560dSAlex Tomas ar->len--; 4321c9de560dSAlex Tomas } 432260e58e0fSMingming Cao inquota = ar->len; 4323c9de560dSAlex Tomas if (ar->len == 0) { 4324c9de560dSAlex Tomas *errp = -EDQUOT; 43256c7a120aSAditya Kali goto out; 4326c9de560dSAlex Tomas } 432760e58e0fSMingming Cao } 4328d2a17637SMingming Cao 4329256bdb49SEric Sandeen ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS); 4330833576b3STheodore Ts'o if (!ac) { 4331363d4251SShen Feng ar->len = 0; 4332256bdb49SEric Sandeen *errp = -ENOMEM; 43336c7a120aSAditya Kali goto out; 4334256bdb49SEric Sandeen } 4335256bdb49SEric Sandeen 4336256bdb49SEric Sandeen *errp = ext4_mb_initialize_context(ac, ar); 4337c9de560dSAlex Tomas if (*errp) { 4338c9de560dSAlex Tomas ar->len = 0; 43396c7a120aSAditya Kali goto out; 4340c9de560dSAlex Tomas } 4341c9de560dSAlex Tomas 4342256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_PREALLOC; 4343256bdb49SEric Sandeen if (!ext4_mb_use_preallocated(ac)) { 4344256bdb49SEric Sandeen ac->ac_op = EXT4_MB_HISTORY_ALLOC; 4345256bdb49SEric Sandeen ext4_mb_normalize_request(ac, ar); 4346c9de560dSAlex Tomas repeat: 4347c9de560dSAlex Tomas /* allocate space in core */ 43486c7a120aSAditya Kali *errp = ext4_mb_regular_allocator(ac); 43496c7a120aSAditya Kali if (*errp) 43506c7a120aSAditya Kali goto errout; 4351c9de560dSAlex Tomas 4352c9de560dSAlex Tomas /* as we've just preallocated more space than 4353c9de560dSAlex Tomas * user requested orinally, we store allocated 4354c9de560dSAlex Tomas * space in a special descriptor */ 4355256bdb49SEric Sandeen if (ac->ac_status == AC_STATUS_FOUND && 4356256bdb49SEric Sandeen ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len) 4357256bdb49SEric Sandeen ext4_mb_new_preallocation(ac); 4358c9de560dSAlex Tomas } 4359256bdb49SEric Sandeen if (likely(ac->ac_status == AC_STATUS_FOUND)) { 43606bc6e63fSAneesh Kumar K.V *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_blks); 4361519deca0SAneesh Kumar K.V if (*errp == -EAGAIN) { 43628556e8f3SAneesh Kumar K.V /* 43638556e8f3SAneesh Kumar K.V * drop the reference that we took 43648556e8f3SAneesh Kumar K.V * in ext4_mb_use_best_found 43658556e8f3SAneesh Kumar K.V */ 43668556e8f3SAneesh Kumar K.V ext4_mb_release_context(ac); 4367519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_group = 0; 4368519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_start = 0; 4369519deca0SAneesh Kumar K.V ac->ac_b_ex.fe_len = 0; 4370519deca0SAneesh Kumar K.V ac->ac_status = AC_STATUS_CONTINUE; 4371519deca0SAneesh Kumar K.V goto repeat; 43726c7a120aSAditya Kali } else if (*errp) 43736c7a120aSAditya Kali errout: 4374b844167eSCurt Wohlgemuth ext4_discard_allocated_blocks(ac); 43756c7a120aSAditya Kali else { 4376256bdb49SEric Sandeen block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex); 4377256bdb49SEric Sandeen ar->len = ac->ac_b_ex.fe_len; 4378519deca0SAneesh Kumar K.V } 4379c9de560dSAlex Tomas } else { 4380256bdb49SEric Sandeen freed = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len); 4381c9de560dSAlex Tomas if (freed) 4382c9de560dSAlex Tomas goto repeat; 4383c9de560dSAlex Tomas *errp = -ENOSPC; 43846c7a120aSAditya Kali } 43856c7a120aSAditya Kali 43866c7a120aSAditya Kali if (*errp) { 4387256bdb49SEric Sandeen ac->ac_b_ex.fe_len = 0; 4388c9de560dSAlex Tomas ar->len = 0; 4389256bdb49SEric Sandeen ext4_mb_show_ac(ac); 4390c9de560dSAlex Tomas } 4391256bdb49SEric Sandeen ext4_mb_release_context(ac); 43926c7a120aSAditya Kali out: 43936c7a120aSAditya Kali if (ac) 4394363d4251SShen Feng kmem_cache_free(ext4_ac_cachep, ac); 439560e58e0fSMingming Cao if (inquota && ar->len < inquota) 43965dd4056dSChristoph Hellwig dquot_free_block(ar->inode, inquota - ar->len); 43970087d9fbSAneesh Kumar K.V if (!ar->len) { 4398f2321097STheodore Ts'o if (!ext4_test_inode_state(ar->inode, 4399f2321097STheodore Ts'o EXT4_STATE_DELALLOC_RESERVED)) 44000087d9fbSAneesh Kumar K.V /* release all the reserved blocks if non delalloc */ 44010087d9fbSAneesh Kumar K.V percpu_counter_sub(&sbi->s_dirtyblocks_counter, 44020087d9fbSAneesh Kumar K.V reserv_blks); 44030087d9fbSAneesh Kumar K.V } 4404c9de560dSAlex Tomas 44059bffad1eSTheodore Ts'o trace_ext4_allocate_blocks(ar, (unsigned long long)block); 4406ba80b101STheodore Ts'o 4407c9de560dSAlex Tomas return block; 4408c9de560dSAlex Tomas } 4409c9de560dSAlex Tomas 4410c894058dSAneesh Kumar K.V /* 4411c894058dSAneesh Kumar K.V * We can merge two free data extents only if the physical blocks 4412c894058dSAneesh Kumar K.V * are contiguous, AND the extents were freed by the same transaction, 4413c894058dSAneesh Kumar K.V * AND the blocks are associated with the same group. 4414c894058dSAneesh Kumar K.V */ 4415c894058dSAneesh Kumar K.V static int can_merge(struct ext4_free_data *entry1, 4416c894058dSAneesh Kumar K.V struct ext4_free_data *entry2) 4417c894058dSAneesh Kumar K.V { 4418c894058dSAneesh Kumar K.V if ((entry1->t_tid == entry2->t_tid) && 4419c894058dSAneesh Kumar K.V (entry1->group == entry2->group) && 4420c894058dSAneesh Kumar K.V ((entry1->start_blk + entry1->count) == entry2->start_blk)) 4421c894058dSAneesh Kumar K.V return 1; 4422c894058dSAneesh Kumar K.V return 0; 4423c894058dSAneesh Kumar K.V } 4424c894058dSAneesh Kumar K.V 44254ddfef7bSEric Sandeen static noinline_for_stack int 44264ddfef7bSEric Sandeen ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b, 44277a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry) 4428c9de560dSAlex Tomas { 4429e29136f8STheodore Ts'o ext4_group_t group = e4b->bd_group; 44307a2fcbf7SAneesh Kumar K.V ext4_grpblk_t block; 44317a2fcbf7SAneesh Kumar K.V struct ext4_free_data *entry; 4432c9de560dSAlex Tomas struct ext4_group_info *db = e4b->bd_info; 4433c9de560dSAlex Tomas struct super_block *sb = e4b->bd_sb; 4434c9de560dSAlex Tomas struct ext4_sb_info *sbi = EXT4_SB(sb); 4435c894058dSAneesh Kumar K.V struct rb_node **n = &db->bb_free_root.rb_node, *node; 4436c894058dSAneesh Kumar K.V struct rb_node *parent = NULL, *new_node; 4437c894058dSAneesh Kumar K.V 44380390131bSFrank Mayhar BUG_ON(!ext4_handle_valid(handle)); 4439c9de560dSAlex Tomas BUG_ON(e4b->bd_bitmap_page == NULL); 4440c9de560dSAlex Tomas BUG_ON(e4b->bd_buddy_page == NULL); 4441c9de560dSAlex Tomas 4442c894058dSAneesh Kumar K.V new_node = &new_entry->node; 44437a2fcbf7SAneesh Kumar K.V block = new_entry->start_blk; 4444c9de560dSAlex Tomas 4445c894058dSAneesh Kumar K.V if (!*n) { 4446c894058dSAneesh Kumar K.V /* first free block exent. We need to 4447c894058dSAneesh Kumar K.V protect buddy cache from being freed, 4448c9de560dSAlex Tomas * otherwise we'll refresh it from 4449c9de560dSAlex Tomas * on-disk bitmap and lose not-yet-available 4450c9de560dSAlex Tomas * blocks */ 4451c9de560dSAlex Tomas page_cache_get(e4b->bd_buddy_page); 4452c9de560dSAlex Tomas page_cache_get(e4b->bd_bitmap_page); 4453c894058dSAneesh Kumar K.V } 4454c894058dSAneesh Kumar K.V while (*n) { 4455c894058dSAneesh Kumar K.V parent = *n; 4456c894058dSAneesh Kumar K.V entry = rb_entry(parent, struct ext4_free_data, node); 4457c894058dSAneesh Kumar K.V if (block < entry->start_blk) 4458c894058dSAneesh Kumar K.V n = &(*n)->rb_left; 4459c894058dSAneesh Kumar K.V else if (block >= (entry->start_blk + entry->count)) 4460c894058dSAneesh Kumar K.V n = &(*n)->rb_right; 4461c894058dSAneesh Kumar K.V else { 4462e29136f8STheodore Ts'o ext4_grp_locked_error(sb, group, 0, 4463e29136f8STheodore Ts'o ext4_group_first_block_no(sb, group) + block, 4464e29136f8STheodore Ts'o "Block already on to-be-freed list"); 4465c894058dSAneesh Kumar K.V return 0; 4466c9de560dSAlex Tomas } 4467c9de560dSAlex Tomas } 4468c9de560dSAlex Tomas 4469c894058dSAneesh Kumar K.V rb_link_node(new_node, parent, n); 4470c894058dSAneesh Kumar K.V rb_insert_color(new_node, &db->bb_free_root); 4471c894058dSAneesh Kumar K.V 4472c894058dSAneesh Kumar K.V /* Now try to see the extent can be merged to left and right */ 4473c894058dSAneesh Kumar K.V node = rb_prev(new_node); 4474c894058dSAneesh Kumar K.V if (node) { 4475c894058dSAneesh Kumar K.V entry = rb_entry(node, struct ext4_free_data, node); 4476c894058dSAneesh Kumar K.V if (can_merge(entry, new_entry)) { 4477c894058dSAneesh Kumar K.V new_entry->start_blk = entry->start_blk; 4478c894058dSAneesh Kumar K.V new_entry->count += entry->count; 4479c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 4480c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 4481c894058dSAneesh Kumar K.V list_del(&entry->list); 4482c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4483c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 4484c9de560dSAlex Tomas } 4485c9de560dSAlex Tomas } 4486c894058dSAneesh Kumar K.V 4487c894058dSAneesh Kumar K.V node = rb_next(new_node); 4488c894058dSAneesh Kumar K.V if (node) { 4489c894058dSAneesh Kumar K.V entry = rb_entry(node, struct ext4_free_data, node); 4490c894058dSAneesh Kumar K.V if (can_merge(new_entry, entry)) { 4491c894058dSAneesh Kumar K.V new_entry->count += entry->count; 4492c894058dSAneesh Kumar K.V rb_erase(node, &(db->bb_free_root)); 4493c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 4494c894058dSAneesh Kumar K.V list_del(&entry->list); 4495c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4496c894058dSAneesh Kumar K.V kmem_cache_free(ext4_free_ext_cachep, entry); 4497c894058dSAneesh Kumar K.V } 4498c894058dSAneesh Kumar K.V } 44993e624fc7STheodore Ts'o /* Add the extent to transaction's private list */ 4500c894058dSAneesh Kumar K.V spin_lock(&sbi->s_md_lock); 45013e624fc7STheodore Ts'o list_add(&new_entry->list, &handle->h_transaction->t_private_list); 4502c894058dSAneesh Kumar K.V spin_unlock(&sbi->s_md_lock); 4503c9de560dSAlex Tomas return 0; 4504c9de560dSAlex Tomas } 4505c9de560dSAlex Tomas 450644338711STheodore Ts'o /** 450744338711STheodore Ts'o * ext4_free_blocks() -- Free given blocks and update quota 450844338711STheodore Ts'o * @handle: handle for this transaction 450944338711STheodore Ts'o * @inode: inode 451044338711STheodore Ts'o * @block: start physical block to free 451144338711STheodore Ts'o * @count: number of blocks to count 451244338711STheodore Ts'o * @metadata: Are these metadata blocks 4513c9de560dSAlex Tomas */ 451444338711STheodore Ts'o void ext4_free_blocks(handle_t *handle, struct inode *inode, 4515e6362609STheodore Ts'o struct buffer_head *bh, ext4_fsblk_t block, 4516e6362609STheodore Ts'o unsigned long count, int flags) 4517c9de560dSAlex Tomas { 451826346ff6SAneesh Kumar K.V struct buffer_head *bitmap_bh = NULL; 4519c9de560dSAlex Tomas struct super_block *sb = inode->i_sb; 4520c9de560dSAlex Tomas struct ext4_group_desc *gdp; 452144338711STheodore Ts'o unsigned long freed = 0; 4522498e5f24STheodore Ts'o unsigned int overflow; 4523c9de560dSAlex Tomas ext4_grpblk_t bit; 4524c9de560dSAlex Tomas struct buffer_head *gd_bh; 4525c9de560dSAlex Tomas ext4_group_t block_group; 4526c9de560dSAlex Tomas struct ext4_sb_info *sbi; 4527c9de560dSAlex Tomas struct ext4_buddy e4b; 4528c9de560dSAlex Tomas int err = 0; 4529c9de560dSAlex Tomas int ret; 4530c9de560dSAlex Tomas 4531e6362609STheodore Ts'o if (bh) { 4532e6362609STheodore Ts'o if (block) 4533e6362609STheodore Ts'o BUG_ON(block != bh->b_blocknr); 4534e6362609STheodore Ts'o else 4535e6362609STheodore Ts'o block = bh->b_blocknr; 4536e6362609STheodore Ts'o } 4537c9de560dSAlex Tomas 4538c9de560dSAlex Tomas sbi = EXT4_SB(sb); 45391f2acb60STheodore Ts'o if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) && 45401f2acb60STheodore Ts'o !ext4_data_block_valid(sbi, block, count)) { 454112062dddSEric Sandeen ext4_error(sb, "Freeing blocks not in datazone - " 45420610b6e9STheodore Ts'o "block = %llu, count = %lu", block, count); 4543c9de560dSAlex Tomas goto error_return; 4544c9de560dSAlex Tomas } 4545c9de560dSAlex Tomas 45460610b6e9STheodore Ts'o ext4_debug("freeing block %llu\n", block); 4547e6362609STheodore Ts'o trace_ext4_free_blocks(inode, block, count, flags); 4548e6362609STheodore Ts'o 4549e6362609STheodore Ts'o if (flags & EXT4_FREE_BLOCKS_FORGET) { 4550e6362609STheodore Ts'o struct buffer_head *tbh = bh; 4551e6362609STheodore Ts'o int i; 4552e6362609STheodore Ts'o 4553e6362609STheodore Ts'o BUG_ON(bh && (count > 1)); 4554e6362609STheodore Ts'o 4555e6362609STheodore Ts'o for (i = 0; i < count; i++) { 4556e6362609STheodore Ts'o if (!bh) 4557e6362609STheodore Ts'o tbh = sb_find_get_block(inode->i_sb, 4558e6362609STheodore Ts'o block + i); 455987783690SNamhyung Kim if (unlikely(!tbh)) 456087783690SNamhyung Kim continue; 4561e6362609STheodore Ts'o ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA, 4562e6362609STheodore Ts'o inode, tbh, block + i); 4563e6362609STheodore Ts'o } 4564e6362609STheodore Ts'o } 4565e6362609STheodore Ts'o 4566e6362609STheodore Ts'o /* 4567e6362609STheodore Ts'o * We need to make sure we don't reuse the freed block until 4568e6362609STheodore Ts'o * after the transaction is committed, which we can do by 4569e6362609STheodore Ts'o * treating the block as metadata, below. We make an 4570e6362609STheodore Ts'o * exception if the inode is to be written in writeback mode 4571e6362609STheodore Ts'o * since writeback mode has weak data consistency guarantees. 4572e6362609STheodore Ts'o */ 4573e6362609STheodore Ts'o if (!ext4_should_writeback_data(inode)) 4574e6362609STheodore Ts'o flags |= EXT4_FREE_BLOCKS_METADATA; 4575c9de560dSAlex Tomas 4576c9de560dSAlex Tomas do_more: 4577c9de560dSAlex Tomas overflow = 0; 4578c9de560dSAlex Tomas ext4_get_group_no_and_offset(sb, block, &block_group, &bit); 4579c9de560dSAlex Tomas 4580c9de560dSAlex Tomas /* 4581c9de560dSAlex Tomas * Check to see if we are freeing blocks across a group 4582c9de560dSAlex Tomas * boundary. 4583c9de560dSAlex Tomas */ 4584c9de560dSAlex Tomas if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) { 4585c9de560dSAlex Tomas overflow = bit + count - EXT4_BLOCKS_PER_GROUP(sb); 4586c9de560dSAlex Tomas count -= overflow; 4587c9de560dSAlex Tomas } 4588574ca174STheodore Ts'o bitmap_bh = ext4_read_block_bitmap(sb, block_group); 4589ce89f46cSAneesh Kumar K.V if (!bitmap_bh) { 4590ce89f46cSAneesh Kumar K.V err = -EIO; 4591c9de560dSAlex Tomas goto error_return; 4592ce89f46cSAneesh Kumar K.V } 4593c9de560dSAlex Tomas gdp = ext4_get_group_desc(sb, block_group, &gd_bh); 4594ce89f46cSAneesh Kumar K.V if (!gdp) { 4595ce89f46cSAneesh Kumar K.V err = -EIO; 4596c9de560dSAlex Tomas goto error_return; 4597ce89f46cSAneesh Kumar K.V } 4598c9de560dSAlex Tomas 4599c9de560dSAlex Tomas if (in_range(ext4_block_bitmap(sb, gdp), block, count) || 4600c9de560dSAlex Tomas in_range(ext4_inode_bitmap(sb, gdp), block, count) || 4601c9de560dSAlex Tomas in_range(block, ext4_inode_table(sb, gdp), 4602c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group) || 4603c9de560dSAlex Tomas in_range(block + count - 1, ext4_inode_table(sb, gdp), 4604c9de560dSAlex Tomas EXT4_SB(sb)->s_itb_per_group)) { 4605c9de560dSAlex Tomas 460612062dddSEric Sandeen ext4_error(sb, "Freeing blocks in system zone - " 46070610b6e9STheodore Ts'o "Block = %llu, count = %lu", block, count); 4608519deca0SAneesh Kumar K.V /* err = 0. ext4_std_error should be a no op */ 4609519deca0SAneesh Kumar K.V goto error_return; 4610c9de560dSAlex Tomas } 4611c9de560dSAlex Tomas 4612c9de560dSAlex Tomas BUFFER_TRACE(bitmap_bh, "getting write access"); 4613c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, bitmap_bh); 4614c9de560dSAlex Tomas if (err) 4615c9de560dSAlex Tomas goto error_return; 4616c9de560dSAlex Tomas 4617c9de560dSAlex Tomas /* 4618c9de560dSAlex Tomas * We are about to modify some metadata. Call the journal APIs 4619c9de560dSAlex Tomas * to unshare ->b_data if a currently-committing transaction is 4620c9de560dSAlex Tomas * using it 4621c9de560dSAlex Tomas */ 4622c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "get_write_access"); 4623c9de560dSAlex Tomas err = ext4_journal_get_write_access(handle, gd_bh); 4624c9de560dSAlex Tomas if (err) 4625c9de560dSAlex Tomas goto error_return; 4626c9de560dSAlex Tomas #ifdef AGGRESSIVE_CHECK 4627c9de560dSAlex Tomas { 4628c9de560dSAlex Tomas int i; 4629c9de560dSAlex Tomas for (i = 0; i < count; i++) 4630c9de560dSAlex Tomas BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data)); 4631c9de560dSAlex Tomas } 4632c9de560dSAlex Tomas #endif 46333e1e5f50SEric Sandeen trace_ext4_mballoc_free(sb, inode, block_group, bit, count); 4634c9de560dSAlex Tomas 4635920313a7SAneesh Kumar K.V err = ext4_mb_load_buddy(sb, block_group, &e4b); 4636920313a7SAneesh Kumar K.V if (err) 4637920313a7SAneesh Kumar K.V goto error_return; 4638e6362609STheodore Ts'o 4639e6362609STheodore Ts'o if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) { 46407a2fcbf7SAneesh Kumar K.V struct ext4_free_data *new_entry; 46417a2fcbf7SAneesh Kumar K.V /* 46427a2fcbf7SAneesh Kumar K.V * blocks being freed are metadata. these blocks shouldn't 46437a2fcbf7SAneesh Kumar K.V * be used until this transaction is committed 46447a2fcbf7SAneesh Kumar K.V */ 46457a2fcbf7SAneesh Kumar K.V new_entry = kmem_cache_alloc(ext4_free_ext_cachep, GFP_NOFS); 4646b72143abSTheodore Ts'o if (!new_entry) { 4647b72143abSTheodore Ts'o err = -ENOMEM; 4648b72143abSTheodore Ts'o goto error_return; 4649b72143abSTheodore Ts'o } 46507a2fcbf7SAneesh Kumar K.V new_entry->start_blk = bit; 46517a2fcbf7SAneesh Kumar K.V new_entry->group = block_group; 46527a2fcbf7SAneesh Kumar K.V new_entry->count = count; 46537a2fcbf7SAneesh Kumar K.V new_entry->t_tid = handle->h_transaction->t_tid; 4654955ce5f5SAneesh Kumar K.V 46557a2fcbf7SAneesh Kumar K.V ext4_lock_group(sb, block_group); 4656955ce5f5SAneesh Kumar K.V mb_clear_bits(bitmap_bh->b_data, bit, count); 46577a2fcbf7SAneesh Kumar K.V ext4_mb_free_metadata(handle, &e4b, new_entry); 4658c9de560dSAlex Tomas } else { 46597a2fcbf7SAneesh Kumar K.V /* need to update group_info->bb_free and bitmap 46607a2fcbf7SAneesh Kumar K.V * with group lock held. generate_buddy look at 46617a2fcbf7SAneesh Kumar K.V * them with group lock_held 46627a2fcbf7SAneesh Kumar K.V */ 4663955ce5f5SAneesh Kumar K.V ext4_lock_group(sb, block_group); 4664955ce5f5SAneesh Kumar K.V mb_clear_bits(bitmap_bh->b_data, bit, count); 46657e5a8cddSShen Feng mb_free_blocks(inode, &e4b, bit, count); 4666c9de560dSAlex Tomas } 4667c9de560dSAlex Tomas 4668560671a0SAneesh Kumar K.V ret = ext4_free_blks_count(sb, gdp) + count; 4669560671a0SAneesh Kumar K.V ext4_free_blks_set(sb, gdp, ret); 4670c9de560dSAlex Tomas gdp->bg_checksum = ext4_group_desc_csum(sbi, block_group, gdp); 4671955ce5f5SAneesh Kumar K.V ext4_unlock_group(sb, block_group); 4672c9de560dSAlex Tomas percpu_counter_add(&sbi->s_freeblocks_counter, count); 4673c9de560dSAlex Tomas 4674772cb7c8SJose R. Santos if (sbi->s_log_groups_per_flex) { 4675772cb7c8SJose R. Santos ext4_group_t flex_group = ext4_flex_group(sbi, block_group); 46769f24e420STheodore Ts'o atomic_add(count, &sbi->s_flex_groups[flex_group].free_blocks); 4677772cb7c8SJose R. Santos } 4678772cb7c8SJose R. Santos 4679e39e07fdSJing Zhang ext4_mb_unload_buddy(&e4b); 4680c9de560dSAlex Tomas 468144338711STheodore Ts'o freed += count; 4682c9de560dSAlex Tomas 46837a2fcbf7SAneesh Kumar K.V /* We dirtied the bitmap block */ 46847a2fcbf7SAneesh Kumar K.V BUFFER_TRACE(bitmap_bh, "dirtied bitmap block"); 46857a2fcbf7SAneesh Kumar K.V err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh); 46867a2fcbf7SAneesh Kumar K.V 4687c9de560dSAlex Tomas /* And the group descriptor block */ 4688c9de560dSAlex Tomas BUFFER_TRACE(gd_bh, "dirtied group descriptor block"); 46890390131bSFrank Mayhar ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh); 4690c9de560dSAlex Tomas if (!err) 4691c9de560dSAlex Tomas err = ret; 4692c9de560dSAlex Tomas 4693c9de560dSAlex Tomas if (overflow && !err) { 4694c9de560dSAlex Tomas block += count; 4695c9de560dSAlex Tomas count = overflow; 4696c9de560dSAlex Tomas put_bh(bitmap_bh); 4697c9de560dSAlex Tomas goto do_more; 4698c9de560dSAlex Tomas } 4699a0375156STheodore Ts'o ext4_mark_super_dirty(sb); 4700c9de560dSAlex Tomas error_return: 470144338711STheodore Ts'o if (freed) 47025dd4056dSChristoph Hellwig dquot_free_block(inode, freed); 4703c9de560dSAlex Tomas brelse(bitmap_bh); 4704c9de560dSAlex Tomas ext4_std_error(sb, err); 4705c9de560dSAlex Tomas return; 4706c9de560dSAlex Tomas } 47077360d173SLukas Czerner 47087360d173SLukas Czerner /** 47097360d173SLukas Czerner * ext4_trim_extent -- function to TRIM one single free extent in the group 47107360d173SLukas Czerner * @sb: super block for the file system 47117360d173SLukas Czerner * @start: starting block of the free extent in the alloc. group 47127360d173SLukas Czerner * @count: number of blocks to TRIM 47137360d173SLukas Czerner * @group: alloc. group we are working with 47147360d173SLukas Czerner * @e4b: ext4 buddy for the group 47157360d173SLukas Czerner * 47167360d173SLukas Czerner * Trim "count" blocks starting at "start" in the "group". To assure that no 47177360d173SLukas Czerner * one will allocate those blocks, mark it as used in buddy bitmap. This must 47187360d173SLukas Czerner * be called with under the group lock. 47197360d173SLukas Czerner */ 47207360d173SLukas Czerner static int ext4_trim_extent(struct super_block *sb, int start, int count, 47217360d173SLukas Czerner ext4_group_t group, struct ext4_buddy *e4b) 47227360d173SLukas Czerner { 47237360d173SLukas Czerner struct ext4_free_extent ex; 47247360d173SLukas Czerner int ret = 0; 47257360d173SLukas Czerner 47267360d173SLukas Czerner assert_spin_locked(ext4_group_lock_ptr(sb, group)); 47277360d173SLukas Czerner 47287360d173SLukas Czerner ex.fe_start = start; 47297360d173SLukas Czerner ex.fe_group = group; 47307360d173SLukas Czerner ex.fe_len = count; 47317360d173SLukas Czerner 47327360d173SLukas Czerner /* 47337360d173SLukas Czerner * Mark blocks used, so no one can reuse them while 47347360d173SLukas Czerner * being trimmed. 47357360d173SLukas Czerner */ 47367360d173SLukas Czerner mb_mark_used(e4b, &ex); 47377360d173SLukas Czerner ext4_unlock_group(sb, group); 47387360d173SLukas Czerner 47397360d173SLukas Czerner ret = ext4_issue_discard(sb, group, start, count); 47407360d173SLukas Czerner 47417360d173SLukas Czerner ext4_lock_group(sb, group); 47427360d173SLukas Czerner mb_free_blocks(NULL, e4b, start, ex.fe_len); 47437360d173SLukas Czerner return ret; 47447360d173SLukas Czerner } 47457360d173SLukas Czerner 47467360d173SLukas Czerner /** 47477360d173SLukas Czerner * ext4_trim_all_free -- function to trim all free space in alloc. group 47487360d173SLukas Czerner * @sb: super block for file system 47497360d173SLukas Czerner * @e4b: ext4 buddy 47507360d173SLukas Czerner * @start: first group block to examine 47517360d173SLukas Czerner * @max: last group block to examine 47527360d173SLukas Czerner * @minblocks: minimum extent block count 47537360d173SLukas Czerner * 47547360d173SLukas Czerner * ext4_trim_all_free walks through group's buddy bitmap searching for free 47557360d173SLukas Czerner * extents. When the free block is found, ext4_trim_extent is called to TRIM 47567360d173SLukas Czerner * the extent. 47577360d173SLukas Czerner * 47587360d173SLukas Czerner * 47597360d173SLukas Czerner * ext4_trim_all_free walks through group's block bitmap searching for free 47607360d173SLukas Czerner * extents. When the free extent is found, mark it as used in group buddy 47617360d173SLukas Czerner * bitmap. Then issue a TRIM command on this extent and free the extent in 47627360d173SLukas Czerner * the group buddy bitmap. This is done until whole group is scanned. 47637360d173SLukas Czerner */ 47640b75a840SLukas Czerner static ext4_grpblk_t 47650b75a840SLukas Czerner ext4_trim_all_free(struct super_block *sb, struct ext4_buddy *e4b, 47667360d173SLukas Czerner ext4_grpblk_t start, ext4_grpblk_t max, ext4_grpblk_t minblocks) 47677360d173SLukas Czerner { 47687360d173SLukas Czerner void *bitmap; 47697360d173SLukas Czerner ext4_grpblk_t next, count = 0; 47707360d173SLukas Czerner ext4_group_t group; 47717360d173SLukas Czerner int ret = 0; 47727360d173SLukas Czerner 47737360d173SLukas Czerner BUG_ON(e4b == NULL); 47747360d173SLukas Czerner 47757360d173SLukas Czerner bitmap = e4b->bd_bitmap; 47767360d173SLukas Czerner group = e4b->bd_group; 47777360d173SLukas Czerner start = (e4b->bd_info->bb_first_free > start) ? 47787360d173SLukas Czerner e4b->bd_info->bb_first_free : start; 47797360d173SLukas Czerner ext4_lock_group(sb, group); 47807360d173SLukas Czerner 47817360d173SLukas Czerner while (start < max) { 47827360d173SLukas Czerner start = mb_find_next_zero_bit(bitmap, max, start); 47837360d173SLukas Czerner if (start >= max) 47847360d173SLukas Czerner break; 47857360d173SLukas Czerner next = mb_find_next_bit(bitmap, max, start); 47867360d173SLukas Czerner 47877360d173SLukas Czerner if ((next - start) >= minblocks) { 47887360d173SLukas Czerner ret = ext4_trim_extent(sb, start, 47897360d173SLukas Czerner next - start, group, e4b); 47907360d173SLukas Czerner if (ret < 0) 47917360d173SLukas Czerner break; 47927360d173SLukas Czerner count += next - start; 47937360d173SLukas Czerner } 47947360d173SLukas Czerner start = next + 1; 47957360d173SLukas Czerner 47967360d173SLukas Czerner if (fatal_signal_pending(current)) { 47977360d173SLukas Czerner count = -ERESTARTSYS; 47987360d173SLukas Czerner break; 47997360d173SLukas Czerner } 48007360d173SLukas Czerner 48017360d173SLukas Czerner if (need_resched()) { 48027360d173SLukas Czerner ext4_unlock_group(sb, group); 48037360d173SLukas Czerner cond_resched(); 48047360d173SLukas Czerner ext4_lock_group(sb, group); 48057360d173SLukas Czerner } 48067360d173SLukas Czerner 48077360d173SLukas Czerner if ((e4b->bd_info->bb_free - count) < minblocks) 48087360d173SLukas Czerner break; 48097360d173SLukas Czerner } 48107360d173SLukas Czerner ext4_unlock_group(sb, group); 48117360d173SLukas Czerner 48127360d173SLukas Czerner ext4_debug("trimmed %d blocks in the group %d\n", 48137360d173SLukas Czerner count, group); 48147360d173SLukas Czerner 48157360d173SLukas Czerner if (ret < 0) 48167360d173SLukas Czerner count = ret; 48177360d173SLukas Czerner 48187360d173SLukas Czerner return count; 48197360d173SLukas Czerner } 48207360d173SLukas Czerner 48217360d173SLukas Czerner /** 48227360d173SLukas Czerner * ext4_trim_fs() -- trim ioctl handle function 48237360d173SLukas Czerner * @sb: superblock for filesystem 48247360d173SLukas Czerner * @range: fstrim_range structure 48257360d173SLukas Czerner * 48267360d173SLukas Czerner * start: First Byte to trim 48277360d173SLukas Czerner * len: number of Bytes to trim from start 48287360d173SLukas Czerner * minlen: minimum extent length in Bytes 48297360d173SLukas Czerner * ext4_trim_fs goes through all allocation groups containing Bytes from 48307360d173SLukas Czerner * start to start+len. For each such a group ext4_trim_all_free function 48317360d173SLukas Czerner * is invoked to trim all free space. 48327360d173SLukas Czerner */ 48337360d173SLukas Czerner int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range) 48347360d173SLukas Czerner { 48357360d173SLukas Czerner struct ext4_buddy e4b; 48367360d173SLukas Czerner ext4_group_t first_group, last_group; 48377360d173SLukas Czerner ext4_group_t group, ngroups = ext4_get_groups_count(sb); 48387360d173SLukas Czerner ext4_grpblk_t cnt = 0, first_block, last_block; 48397360d173SLukas Czerner uint64_t start, len, minlen, trimmed; 48400f0a25bfSJan Kara ext4_fsblk_t first_data_blk = 48410f0a25bfSJan Kara le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block); 48427360d173SLukas Czerner int ret = 0; 48437360d173SLukas Czerner 48447360d173SLukas Czerner start = range->start >> sb->s_blocksize_bits; 48457360d173SLukas Czerner len = range->len >> sb->s_blocksize_bits; 48467360d173SLukas Czerner minlen = range->minlen >> sb->s_blocksize_bits; 48477360d173SLukas Czerner trimmed = 0; 48487360d173SLukas Czerner 48497360d173SLukas Czerner if (unlikely(minlen > EXT4_BLOCKS_PER_GROUP(sb))) 48507360d173SLukas Czerner return -EINVAL; 48510f0a25bfSJan Kara if (start < first_data_blk) { 48520f0a25bfSJan Kara len -= first_data_blk - start; 48530f0a25bfSJan Kara start = first_data_blk; 48540f0a25bfSJan Kara } 48557360d173SLukas Czerner 48567360d173SLukas Czerner /* Determine first and last group to examine based on start and len */ 48577360d173SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start, 48587360d173SLukas Czerner &first_group, &first_block); 48597360d173SLukas Czerner ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) (start + len), 48607360d173SLukas Czerner &last_group, &last_block); 48617360d173SLukas Czerner last_group = (last_group > ngroups - 1) ? ngroups - 1 : last_group; 48627360d173SLukas Czerner last_block = EXT4_BLOCKS_PER_GROUP(sb); 48637360d173SLukas Czerner 48647360d173SLukas Czerner if (first_group > last_group) 48657360d173SLukas Czerner return -EINVAL; 48667360d173SLukas Czerner 48677360d173SLukas Czerner for (group = first_group; group <= last_group; group++) { 48687360d173SLukas Czerner ret = ext4_mb_load_buddy(sb, group, &e4b); 48697360d173SLukas Czerner if (ret) { 48707360d173SLukas Czerner ext4_error(sb, "Error in loading buddy " 48717360d173SLukas Czerner "information for %u", group); 48727360d173SLukas Czerner break; 48737360d173SLukas Czerner } 48747360d173SLukas Czerner 48750ba08517STao Ma /* 48760ba08517STao Ma * For all the groups except the last one, last block will 48770ba08517STao Ma * always be EXT4_BLOCKS_PER_GROUP(sb), so we only need to 48780ba08517STao Ma * change it for the last group in which case start + 48790ba08517STao Ma * len < EXT4_BLOCKS_PER_GROUP(sb). 48800ba08517STao Ma */ 48810ba08517STao Ma if (first_block + len < EXT4_BLOCKS_PER_GROUP(sb)) 4882ca6e909fSJan Kara last_block = first_block + len; 48830ba08517STao Ma len -= last_block - first_block; 48847360d173SLukas Czerner 48857360d173SLukas Czerner if (e4b.bd_info->bb_free >= minlen) { 48867360d173SLukas Czerner cnt = ext4_trim_all_free(sb, &e4b, first_block, 48877360d173SLukas Czerner last_block, minlen); 48887360d173SLukas Czerner if (cnt < 0) { 48897360d173SLukas Czerner ret = cnt; 48907360d173SLukas Czerner ext4_mb_unload_buddy(&e4b); 48917360d173SLukas Czerner break; 48927360d173SLukas Czerner } 48937360d173SLukas Czerner } 48947360d173SLukas Czerner ext4_mb_unload_buddy(&e4b); 48957360d173SLukas Czerner trimmed += cnt; 48967360d173SLukas Czerner first_block = 0; 48977360d173SLukas Czerner } 48987360d173SLukas Czerner range->len = trimmed * sb->s_blocksize; 48997360d173SLukas Czerner 49007360d173SLukas Czerner return ret; 49017360d173SLukas Czerner } 4902