linux/fs/ext4/mballoc.c
<<
>>
Prefs
   1/*
   2 * Copyright (c) 2003-2006, Cluster File Systems, Inc, info@clusterfs.com
   3 * Written by Alex Tomas <alex@clusterfs.com>
   4 *
   5 * This program is free software; you can redistribute it and/or modify
   6 * it under the terms of the GNU General Public License version 2 as
   7 * published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope that it will be useful,
  10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 * GNU General Public License for more details.
  13 *
  14 * You should have received a copy of the GNU General Public Licens
  15 * along with this program; if not, write to the Free Software
  16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-
  17 */
  18
  19
  20/*
  21 * mballoc.c contains the multiblocks allocation routines
  22 */
  23
  24#include "ext4_jbd2.h"
  25#include "mballoc.h"
  26#include <linux/debugfs.h>
  27#include <linux/slab.h>
  28#include <trace/events/ext4.h>
  29
  30/*
  31 * MUSTDO:
  32 *   - test ext4_ext_search_left() and ext4_ext_search_right()
  33 *   - search for metadata in few groups
  34 *
  35 * TODO v4:
  36 *   - normalization should take into account whether file is still open
  37 *   - discard preallocations if no free space left (policy?)
  38 *   - don't normalize tails
  39 *   - quota
  40 *   - reservation for superuser
  41 *
  42 * TODO v3:
  43 *   - bitmap read-ahead (proposed by Oleg Drokin aka green)
  44 *   - track min/max extents in each group for better group selection
  45 *   - mb_mark_used() may allocate chunk right after splitting buddy
  46 *   - tree of groups sorted by number of free blocks
  47 *   - error handling
  48 */
  49
  50/*
  51 * The allocation request involve request for multiple number of blocks
  52 * near to the goal(block) value specified.
  53 *
  54 * During initialization phase of the allocator we decide to use the
  55 * group preallocation or inode preallocation depending on the size of
  56 * the file. The size of the file could be the resulting file size we
  57 * would have after allocation, or the current file size, which ever
  58 * is larger. If the size is less than sbi->s_mb_stream_request we
  59 * select to use the group preallocation. The default value of
  60 * s_mb_stream_request is 16 blocks. This can also be tuned via
  61 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
  62 * terms of number of blocks.
  63 *
  64 * The main motivation for having small file use group preallocation is to
  65 * ensure that we have small files closer together on the disk.
  66 *
  67 * First stage the allocator looks at the inode prealloc list,
  68 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
  69 * spaces for this particular inode. The inode prealloc space is
  70 * represented as:
  71 *
  72 * pa_lstart -> the logical start block for this prealloc space
  73 * pa_pstart -> the physical start block for this prealloc space
  74 * pa_len    -> length for this prealloc space (in clusters)
  75 * pa_free   ->  free space available in this prealloc space (in clusters)
  76 *
  77 * The inode preallocation space is used looking at the _logical_ start
  78 * block. If only the logical file block falls within the range of prealloc
  79 * space we will consume the particular prealloc space. This makes sure that
  80 * we have contiguous physical blocks representing the file blocks
  81 *
  82 * The important thing to be noted in case of inode prealloc space is that
  83 * we don't modify the values associated to inode prealloc space except
  84 * pa_free.
  85 *
  86 * If we are not able to find blocks in the inode prealloc space and if we
  87 * have the group allocation flag set then we look at the locality group
  88 * prealloc space. These are per CPU prealloc list represented as
  89 *
  90 * ext4_sb_info.s_locality_groups[smp_processor_id()]
  91 *
  92 * The reason for having a per cpu locality group is to reduce the contention
  93 * between CPUs. It is possible to get scheduled at this point.
  94 *
  95 * The locality group prealloc space is used looking at whether we have
  96 * enough free space (pa_free) within the prealloc space.
  97 *
  98 * If we can't allocate blocks via inode prealloc or/and locality group
  99 * prealloc then we look at the buddy cache. The buddy cache is represented
 100 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
 101 * mapped to the buddy and bitmap information regarding different
 102 * groups. The buddy information is attached to buddy cache inode so that
 103 * we can access them through the page cache. The information regarding
 104 * each group is loaded via ext4_mb_load_buddy.  The information involve
 105 * block bitmap and buddy information. The information are stored in the
 106 * inode as:
 107 *
 108 *  {                        page                        }
 109 *  [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
 110 *
 111 *
 112 * one block each for bitmap and buddy information.  So for each group we
 113 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
 114 * blocksize) blocks.  So it can have information regarding groups_per_page
 115 * which is blocks_per_page/2
 116 *
 117 * The buddy cache inode is not stored on disk. The inode is thrown
 118 * away when the filesystem is unmounted.
 119 *
 120 * We look for count number of blocks in the buddy cache. If we were able
 121 * to locate that many free blocks we return with additional information
 122 * regarding rest of the contiguous physical block available
 123 *
 124 * Before allocating blocks via buddy cache we normalize the request
 125 * blocks. This ensure we ask for more blocks that we needed. The extra
 126 * blocks that we get after allocation is added to the respective prealloc
 127 * list. In case of inode preallocation we follow a list of heuristics
 128 * based on file size. This can be found in ext4_mb_normalize_request. If
 129 * we are doing a group prealloc we try to normalize the request to
 130 * sbi->s_mb_group_prealloc.  The default value of s_mb_group_prealloc is
 131 * dependent on the cluster size; for non-bigalloc file systems, it is
 132 * 512 blocks. This can be tuned via
 133 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
 134 * terms of number of blocks. If we have mounted the file system with -O
 135 * stripe=<value> option the group prealloc request is normalized to the
 136 * the smallest multiple of the stripe value (sbi->s_stripe) which is
 137 * greater than the default mb_group_prealloc.
 138 *
 139 * The regular allocator (using the buddy cache) supports a few tunables.
 140 *
 141 * /sys/fs/ext4/<partition>/mb_min_to_scan
 142 * /sys/fs/ext4/<partition>/mb_max_to_scan
 143 * /sys/fs/ext4/<partition>/mb_order2_req
 144 *
 145 * The regular allocator uses buddy scan only if the request len is power of
 146 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
 147 * value of s_mb_order2_reqs can be tuned via
 148 * /sys/fs/ext4/<partition>/mb_order2_req.  If the request len is equal to
 149 * stripe size (sbi->s_stripe), we try to search for contiguous block in
 150 * stripe size. This should result in better allocation on RAID setups. If
 151 * not, we search in the specific group using bitmap for best extents. The
 152 * tunable min_to_scan and max_to_scan control the behaviour here.
 153 * min_to_scan indicate how long the mballoc __must__ look for a best
 154 * extent and max_to_scan indicates how long the mballoc __can__ look for a
 155 * best extent in the found extents. Searching for the blocks starts with
 156 * the group specified as the goal value in allocation context via
 157 * ac_g_ex. Each group is first checked based on the criteria whether it
 158 * can be used for allocation. ext4_mb_good_group explains how the groups are
 159 * checked.
 160 *
 161 * Both the prealloc space are getting populated as above. So for the first
 162 * request we will hit the buddy cache which will result in this prealloc
 163 * space getting filled. The prealloc space is then later used for the
 164 * subsequent request.
 165 */
 166
 167/*
 168 * mballoc operates on the following data:
 169 *  - on-disk bitmap
 170 *  - in-core buddy (actually includes buddy and bitmap)
 171 *  - preallocation descriptors (PAs)
 172 *
 173 * there are two types of preallocations:
 174 *  - inode
 175 *    assiged to specific inode and can be used for this inode only.
 176 *    it describes part of inode's space preallocated to specific
 177 *    physical blocks. any block from that preallocated can be used
 178 *    independent. the descriptor just tracks number of blocks left
 179 *    unused. so, before taking some block from descriptor, one must
 180 *    make sure corresponded logical block isn't allocated yet. this
 181 *    also means that freeing any block within descriptor's range
 182 *    must discard all preallocated blocks.
 183 *  - locality group
 184 *    assigned to specific locality group which does not translate to
 185 *    permanent set of inodes: inode can join and leave group. space
 186 *    from this type of preallocation can be used for any inode. thus
 187 *    it's consumed from the beginning to the end.
 188 *
 189 * relation between them can be expressed as:
 190 *    in-core buddy = on-disk bitmap + preallocation descriptors
 191 *
 192 * this mean blocks mballoc considers used are:
 193 *  - allocated blocks (persistent)
 194 *  - preallocated blocks (non-persistent)
 195 *
 196 * consistency in mballoc world means that at any time a block is either
 197 * free or used in ALL structures. notice: "any time" should not be read
 198 * literally -- time is discrete and delimited by locks.
 199 *
 200 *  to keep it simple, we don't use block numbers, instead we count number of
 201 *  blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
 202 *
 203 * all operations can be expressed as:
 204 *  - init buddy:                       buddy = on-disk + PAs
 205 *  - new PA:                           buddy += N; PA = N
 206 *  - use inode PA:                     on-disk += N; PA -= N
 207 *  - discard inode PA                  buddy -= on-disk - PA; PA = 0
 208 *  - use locality group PA             on-disk += N; PA -= N
 209 *  - discard locality group PA         buddy -= PA; PA = 0
 210 *  note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
 211 *        is used in real operation because we can't know actual used
 212 *        bits from PA, only from on-disk bitmap
 213 *
 214 * if we follow this strict logic, then all operations above should be atomic.
 215 * given some of them can block, we'd have to use something like semaphores
 216 * killing performance on high-end SMP hardware. let's try to relax it using
 217 * the following knowledge:
 218 *  1) if buddy is referenced, it's already initialized
 219 *  2) while block is used in buddy and the buddy is referenced,
 220 *     nobody can re-allocate that block
 221 *  3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
 222 *     bit set and PA claims same block, it's OK. IOW, one can set bit in
 223 *     on-disk bitmap if buddy has same bit set or/and PA covers corresponded
 224 *     block
 225 *
 226 * so, now we're building a concurrency table:
 227 *  - init buddy vs.
 228 *    - new PA
 229 *      blocks for PA are allocated in the buddy, buddy must be referenced
 230 *      until PA is linked to allocation group to avoid concurrent buddy init
 231 *    - use inode PA
 232 *      we need to make sure that either on-disk bitmap or PA has uptodate data
 233 *      given (3) we care that PA-=N operation doesn't interfere with init
 234 *    - discard inode PA
 235 *      the simplest way would be to have buddy initialized by the discard
 236 *    - use locality group PA
 237 *      again PA-=N must be serialized with init
 238 *    - discard locality group PA
 239 *      the simplest way would be to have buddy initialized by the discard
 240 *  - new PA vs.
 241 *    - use inode PA
 242 *      i_data_sem serializes them
 243 *    - discard inode PA
 244 *      discard process must wait until PA isn't used by another process
 245 *    - use locality group PA
 246 *      some mutex should serialize them
 247 *    - discard locality group PA
 248 *      discard process must wait until PA isn't used by another process
 249 *  - use inode PA
 250 *    - use inode PA
 251 *      i_data_sem or another mutex should serializes them
 252 *    - discard inode PA
 253 *      discard process must wait until PA isn't used by another process
 254 *    - use locality group PA
 255 *      nothing wrong here -- they're different PAs covering different blocks
 256 *    - discard locality group PA
 257 *      discard process must wait until PA isn't used by another process
 258 *
 259 * now we're ready to make few consequences:
 260 *  - PA is referenced and while it is no discard is possible
 261 *  - PA is referenced until block isn't marked in on-disk bitmap
 262 *  - PA changes only after on-disk bitmap
 263 *  - discard must not compete with init. either init is done before
 264 *    any discard or they're serialized somehow
 265 *  - buddy init as sum of on-disk bitmap and PAs is done atomically
 266 *
 267 * a special case when we've used PA to emptiness. no need to modify buddy
 268 * in this case, but we should care about concurrent init
 269 *
 270 */
 271
 272 /*
 273 * Logic in few words:
 274 *
 275 *  - allocation:
 276 *    load group
 277 *    find blocks
 278 *    mark bits in on-disk bitmap
 279 *    release group
 280 *
 281 *  - use preallocation:
 282 *    find proper PA (per-inode or group)
 283 *    load group
 284 *    mark bits in on-disk bitmap
 285 *    release group
 286 *    release PA
 287 *
 288 *  - free:
 289 *    load group
 290 *    mark bits in on-disk bitmap
 291 *    release group
 292 *
 293 *  - discard preallocations in group:
 294 *    mark PAs deleted
 295 *    move them onto local list
 296 *    load on-disk bitmap
 297 *    load group
 298 *    remove PA from object (inode or locality group)
 299 *    mark free blocks in-core
 300 *
 301 *  - discard inode's preallocations:
 302 */
 303
 304/*
 305 * Locking rules
 306 *
 307 * Locks:
 308 *  - bitlock on a group        (group)
 309 *  - object (inode/locality)   (object)
 310 *  - per-pa lock               (pa)
 311 *
 312 * Paths:
 313 *  - new pa
 314 *    object
 315 *    group
 316 *
 317 *  - find and use pa:
 318 *    pa
 319 *
 320 *  - release consumed pa:
 321 *    pa
 322 *    group
 323 *    object
 324 *
 325 *  - generate in-core bitmap:
 326 *    group
 327 *        pa
 328 *
 329 *  - discard all for given object (inode, locality group):
 330 *    object
 331 *        pa
 332 *    group
 333 *
 334 *  - discard all for given group:
 335 *    group
 336 *        pa
 337 *    group
 338 *        object
 339 *
 340 */
 341static struct kmem_cache *ext4_pspace_cachep;
 342static struct kmem_cache *ext4_ac_cachep;
 343static struct kmem_cache *ext4_free_data_cachep;
 344
 345/* We create slab caches for groupinfo data structures based on the
 346 * superblock block size.  There will be one per mounted filesystem for
 347 * each unique s_blocksize_bits */
 348#define NR_GRPINFO_CACHES 8
 349static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
 350
 351static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
 352        "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
 353        "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
 354        "ext4_groupinfo_64k", "ext4_groupinfo_128k"
 355};
 356
 357static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 358                                        ext4_group_t group);
 359static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
 360                                                ext4_group_t group);
 361static void ext4_free_data_callback(struct super_block *sb,
 362                                struct ext4_journal_cb_entry *jce, int rc);
 363
 364static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
 365{
 366#if BITS_PER_LONG == 64
 367        *bit += ((unsigned long) addr & 7UL) << 3;
 368        addr = (void *) ((unsigned long) addr & ~7UL);
 369#elif BITS_PER_LONG == 32
 370        *bit += ((unsigned long) addr & 3UL) << 3;
 371        addr = (void *) ((unsigned long) addr & ~3UL);
 372#else
 373#error "how many bits you are?!"
 374#endif
 375        return addr;
 376}
 377
 378static inline int mb_test_bit(int bit, void *addr)
 379{
 380        /*
 381         * ext4_test_bit on architecture like powerpc
 382         * needs unsigned long aligned address
 383         */
 384        addr = mb_correct_addr_and_bit(&bit, addr);
 385        return ext4_test_bit(bit, addr);
 386}
 387
 388static inline void mb_set_bit(int bit, void *addr)
 389{
 390        addr = mb_correct_addr_and_bit(&bit, addr);
 391        ext4_set_bit(bit, addr);
 392}
 393
 394static inline void mb_clear_bit(int bit, void *addr)
 395{
 396        addr = mb_correct_addr_and_bit(&bit, addr);
 397        ext4_clear_bit(bit, addr);
 398}
 399
 400static inline int mb_find_next_zero_bit(void *addr, int max, int start)
 401{
 402        int fix = 0, ret, tmpmax;
 403        addr = mb_correct_addr_and_bit(&fix, addr);
 404        tmpmax = max + fix;
 405        start += fix;
 406
 407        ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
 408        if (ret > max)
 409                return max;
 410        return ret;
 411}
 412
 413static inline int mb_find_next_bit(void *addr, int max, int start)
 414{
 415        int fix = 0, ret, tmpmax;
 416        addr = mb_correct_addr_and_bit(&fix, addr);
 417        tmpmax = max + fix;
 418        start += fix;
 419
 420        ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
 421        if (ret > max)
 422                return max;
 423        return ret;
 424}
 425
 426static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
 427{
 428        char *bb;
 429
 430        BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
 431        BUG_ON(max == NULL);
 432
 433        if (order > e4b->bd_blkbits + 1) {
 434                *max = 0;
 435                return NULL;
 436        }
 437
 438        /* at order 0 we see each particular block */
 439        if (order == 0) {
 440                *max = 1 << (e4b->bd_blkbits + 3);
 441                return e4b->bd_bitmap;
 442        }
 443
 444        bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
 445        *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
 446
 447        return bb;
 448}
 449
 450#ifdef DOUBLE_CHECK
 451static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
 452                           int first, int count)
 453{
 454        int i;
 455        struct super_block *sb = e4b->bd_sb;
 456
 457        if (unlikely(e4b->bd_info->bb_bitmap == NULL))
 458                return;
 459        assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
 460        for (i = 0; i < count; i++) {
 461                if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
 462                        ext4_fsblk_t blocknr;
 463
 464                        blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
 465                        blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
 466                        ext4_grp_locked_error(sb, e4b->bd_group,
 467                                              inode ? inode->i_ino : 0,
 468                                              blocknr,
 469                                              "freeing block already freed "
 470                                              "(bit %u)",
 471                                              first + i);
 472                }
 473                mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
 474        }
 475}
 476
 477static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
 478{
 479        int i;
 480
 481        if (unlikely(e4b->bd_info->bb_bitmap == NULL))
 482                return;
 483        assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
 484        for (i = 0; i < count; i++) {
 485                BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
 486                mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
 487        }
 488}
 489
 490static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
 491{
 492        if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
 493                unsigned char *b1, *b2;
 494                int i;
 495                b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
 496                b2 = (unsigned char *) bitmap;
 497                for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
 498                        if (b1[i] != b2[i]) {
 499                                ext4_msg(e4b->bd_sb, KERN_ERR,
 500                                         "corruption in group %u "
 501                                         "at byte %u(%u): %x in copy != %x "
 502                                         "on disk/prealloc",
 503                                         e4b->bd_group, i, i * 8, b1[i], b2[i]);
 504                                BUG();
 505                        }
 506                }
 507        }
 508}
 509
 510#else
 511static inline void mb_free_blocks_double(struct inode *inode,
 512                                struct ext4_buddy *e4b, int first, int count)
 513{
 514        return;
 515}
 516static inline void mb_mark_used_double(struct ext4_buddy *e4b,
 517                                                int first, int count)
 518{
 519        return;
 520}
 521static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
 522{
 523        return;
 524}
 525#endif
 526
 527#ifdef AGGRESSIVE_CHECK
 528
 529#define MB_CHECK_ASSERT(assert)                                         \
 530do {                                                                    \
 531        if (!(assert)) {                                                \
 532                printk(KERN_EMERG                                       \
 533                        "Assertion failure in %s() at %s:%d: \"%s\"\n", \
 534                        function, file, line, # assert);                \
 535                BUG();                                                  \
 536        }                                                               \
 537} while (0)
 538
 539static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
 540                                const char *function, int line)
 541{
 542        struct super_block *sb = e4b->bd_sb;
 543        int order = e4b->bd_blkbits + 1;
 544        int max;
 545        int max2;
 546        int i;
 547        int j;
 548        int k;
 549        int count;
 550        struct ext4_group_info *grp;
 551        int fragments = 0;
 552        int fstart;
 553        struct list_head *cur;
 554        void *buddy;
 555        void *buddy2;
 556
 557        {
 558                static int mb_check_counter;
 559                if (mb_check_counter++ % 100 != 0)
 560                        return 0;
 561        }
 562
 563        while (order > 1) {
 564                buddy = mb_find_buddy(e4b, order, &max);
 565                MB_CHECK_ASSERT(buddy);
 566                buddy2 = mb_find_buddy(e4b, order - 1, &max2);
 567                MB_CHECK_ASSERT(buddy2);
 568                MB_CHECK_ASSERT(buddy != buddy2);
 569                MB_CHECK_ASSERT(max * 2 == max2);
 570
 571                count = 0;
 572                for (i = 0; i < max; i++) {
 573
 574                        if (mb_test_bit(i, buddy)) {
 575                                /* only single bit in buddy2 may be 1 */
 576                                if (!mb_test_bit(i << 1, buddy2)) {
 577                                        MB_CHECK_ASSERT(
 578                                                mb_test_bit((i<<1)+1, buddy2));
 579                                } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
 580                                        MB_CHECK_ASSERT(
 581                                                mb_test_bit(i << 1, buddy2));
 582                                }
 583                                continue;
 584                        }
 585
 586                        /* both bits in buddy2 must be 1 */
 587                        MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
 588                        MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
 589
 590                        for (j = 0; j < (1 << order); j++) {
 591                                k = (i * (1 << order)) + j;
 592                                MB_CHECK_ASSERT(
 593                                        !mb_test_bit(k, e4b->bd_bitmap));
 594                        }
 595                        count++;
 596                }
 597                MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
 598                order--;
 599        }
 600
 601        fstart = -1;
 602        buddy = mb_find_buddy(e4b, 0, &max);
 603        for (i = 0; i < max; i++) {
 604                if (!mb_test_bit(i, buddy)) {
 605                        MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
 606                        if (fstart == -1) {
 607                                fragments++;
 608                                fstart = i;
 609                        }
 610                        continue;
 611                }
 612                fstart = -1;
 613                /* check used bits only */
 614                for (j = 0; j < e4b->bd_blkbits + 1; j++) {
 615                        buddy2 = mb_find_buddy(e4b, j, &max2);
 616                        k = i >> j;
 617                        MB_CHECK_ASSERT(k < max2);
 618                        MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
 619                }
 620        }
 621        MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
 622        MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
 623
 624        grp = ext4_get_group_info(sb, e4b->bd_group);
 625        list_for_each(cur, &grp->bb_prealloc_list) {
 626                ext4_group_t groupnr;
 627                struct ext4_prealloc_space *pa;
 628                pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
 629                ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
 630                MB_CHECK_ASSERT(groupnr == e4b->bd_group);
 631                for (i = 0; i < pa->pa_len; i++)
 632                        MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
 633        }
 634        return 0;
 635}
 636#undef MB_CHECK_ASSERT
 637#define mb_check_buddy(e4b) __mb_check_buddy(e4b,       \
 638                                        __FILE__, __func__, __LINE__)
 639#else
 640#define mb_check_buddy(e4b)
 641#endif
 642
 643/*
 644 * Divide blocks started from @first with length @len into
 645 * smaller chunks with power of 2 blocks.
 646 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
 647 * then increase bb_counters[] for corresponded chunk size.
 648 */
 649static void ext4_mb_mark_free_simple(struct super_block *sb,
 650                                void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
 651                                        struct ext4_group_info *grp)
 652{
 653        struct ext4_sb_info *sbi = EXT4_SB(sb);
 654        ext4_grpblk_t min;
 655        ext4_grpblk_t max;
 656        ext4_grpblk_t chunk;
 657        unsigned short border;
 658
 659        BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
 660
 661        border = 2 << sb->s_blocksize_bits;
 662
 663        while (len > 0) {
 664                /* find how many blocks can be covered since this position */
 665                max = ffs(first | border) - 1;
 666
 667                /* find how many blocks of power 2 we need to mark */
 668                min = fls(len) - 1;
 669
 670                if (max < min)
 671                        min = max;
 672                chunk = 1 << min;
 673
 674                /* mark multiblock chunks only */
 675                grp->bb_counters[min]++;
 676                if (min > 0)
 677                        mb_clear_bit(first >> min,
 678                                     buddy + sbi->s_mb_offsets[min]);
 679
 680                len -= chunk;
 681                first += chunk;
 682        }
 683}
 684
 685/*
 686 * Cache the order of the largest free extent we have available in this block
 687 * group.
 688 */
 689static void
 690mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
 691{
 692        int i;
 693        int bits;
 694
 695        grp->bb_largest_free_order = -1; /* uninit */
 696
 697        bits = sb->s_blocksize_bits + 1;
 698        for (i = bits; i >= 0; i--) {
 699                if (grp->bb_counters[i] > 0) {
 700                        grp->bb_largest_free_order = i;
 701                        break;
 702                }
 703        }
 704}
 705
 706static noinline_for_stack
 707void ext4_mb_generate_buddy(struct super_block *sb,
 708                                void *buddy, void *bitmap, ext4_group_t group)
 709{
 710        struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 711        ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
 712        ext4_grpblk_t i = 0;
 713        ext4_grpblk_t first;
 714        ext4_grpblk_t len;
 715        unsigned free = 0;
 716        unsigned fragments = 0;
 717        unsigned long long period = get_cycles();
 718
 719        /* initialize buddy from bitmap which is aggregation
 720         * of on-disk bitmap and preallocations */
 721        i = mb_find_next_zero_bit(bitmap, max, 0);
 722        grp->bb_first_free = i;
 723        while (i < max) {
 724                fragments++;
 725                first = i;
 726                i = mb_find_next_bit(bitmap, max, i);
 727                len = i - first;
 728                free += len;
 729                if (len > 1)
 730                        ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
 731                else
 732                        grp->bb_counters[0]++;
 733                if (i < max)
 734                        i = mb_find_next_zero_bit(bitmap, max, i);
 735        }
 736        grp->bb_fragments = fragments;
 737
 738        if (free != grp->bb_free) {
 739                ext4_grp_locked_error(sb, group, 0, 0,
 740                                      "%u clusters in bitmap, %u in gd",
 741                                      free, grp->bb_free);
 742                /*
 743                 * If we intent to continue, we consider group descritor
 744                 * corrupt and update bb_free using bitmap value
 745                 */
 746                grp->bb_free = free;
 747        }
 748        mb_set_largest_free_order(sb, grp);
 749
 750        clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
 751
 752        period = get_cycles() - period;
 753        spin_lock(&EXT4_SB(sb)->s_bal_lock);
 754        EXT4_SB(sb)->s_mb_buddies_generated++;
 755        EXT4_SB(sb)->s_mb_generation_time += period;
 756        spin_unlock(&EXT4_SB(sb)->s_bal_lock);
 757}
 758
 759/* The buddy information is attached the buddy cache inode
 760 * for convenience. The information regarding each group
 761 * is loaded via ext4_mb_load_buddy. The information involve
 762 * block bitmap and buddy information. The information are
 763 * stored in the inode as
 764 *
 765 * {                        page                        }
 766 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
 767 *
 768 *
 769 * one block each for bitmap and buddy information.
 770 * So for each group we take up 2 blocks. A page can
 771 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize)  blocks.
 772 * So it can have information regarding groups_per_page which
 773 * is blocks_per_page/2
 774 *
 775 * Locking note:  This routine takes the block group lock of all groups
 776 * for this page; do not hold this lock when calling this routine!
 777 */
 778
 779static int ext4_mb_init_cache(struct page *page, char *incore)
 780{
 781        ext4_group_t ngroups;
 782        int blocksize;
 783        int blocks_per_page;
 784        int groups_per_page;
 785        int err = 0;
 786        int i;
 787        ext4_group_t first_group, group;
 788        int first_block;
 789        struct super_block *sb;
 790        struct buffer_head *bhs;
 791        struct buffer_head **bh = NULL;
 792        struct inode *inode;
 793        char *data;
 794        char *bitmap;
 795        struct ext4_group_info *grinfo;
 796
 797        mb_debug(1, "init page %lu\n", page->index);
 798
 799        inode = page->mapping->host;
 800        sb = inode->i_sb;
 801        ngroups = ext4_get_groups_count(sb);
 802        blocksize = 1 << inode->i_blkbits;
 803        blocks_per_page = PAGE_CACHE_SIZE / blocksize;
 804
 805        groups_per_page = blocks_per_page >> 1;
 806        if (groups_per_page == 0)
 807                groups_per_page = 1;
 808
 809        /* allocate buffer_heads to read bitmaps */
 810        if (groups_per_page > 1) {
 811                i = sizeof(struct buffer_head *) * groups_per_page;
 812                bh = kzalloc(i, GFP_NOFS);
 813                if (bh == NULL) {
 814                        err = -ENOMEM;
 815                        goto out;
 816                }
 817        } else
 818                bh = &bhs;
 819
 820        first_group = page->index * blocks_per_page / 2;
 821
 822        /* read all groups the page covers into the cache */
 823        for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
 824                if (group >= ngroups)
 825                        break;
 826
 827                grinfo = ext4_get_group_info(sb, group);
 828                /*
 829                 * If page is uptodate then we came here after online resize
 830                 * which added some new uninitialized group info structs, so
 831                 * we must skip all initialized uptodate buddies on the page,
 832                 * which may be currently in use by an allocating task.
 833                 */
 834                if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
 835                        bh[i] = NULL;
 836                        continue;
 837                }
 838                if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
 839                        err = -ENOMEM;
 840                        goto out;
 841                }
 842                mb_debug(1, "read bitmap for group %u\n", group);
 843        }
 844
 845        /* wait for I/O completion */
 846        for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
 847                if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
 848                        err = -EIO;
 849                        goto out;
 850                }
 851        }
 852
 853        first_block = page->index * blocks_per_page;
 854        for (i = 0; i < blocks_per_page; i++) {
 855                int group;
 856
 857                group = (first_block + i) >> 1;
 858                if (group >= ngroups)
 859                        break;
 860
 861                if (!bh[group - first_group])
 862                        /* skip initialized uptodate buddy */
 863                        continue;
 864
 865                /*
 866                 * data carry information regarding this
 867                 * particular group in the format specified
 868                 * above
 869                 *
 870                 */
 871                data = page_address(page) + (i * blocksize);
 872                bitmap = bh[group - first_group]->b_data;
 873
 874                /*
 875                 * We place the buddy block and bitmap block
 876                 * close together
 877                 */
 878                if ((first_block + i) & 1) {
 879                        /* this is block of buddy */
 880                        BUG_ON(incore == NULL);
 881                        mb_debug(1, "put buddy for group %u in page %lu/%x\n",
 882                                group, page->index, i * blocksize);
 883                        trace_ext4_mb_buddy_bitmap_load(sb, group);
 884                        grinfo = ext4_get_group_info(sb, group);
 885                        grinfo->bb_fragments = 0;
 886                        memset(grinfo->bb_counters, 0,
 887                               sizeof(*grinfo->bb_counters) *
 888                                (sb->s_blocksize_bits+2));
 889                        /*
 890                         * incore got set to the group block bitmap below
 891                         */
 892                        ext4_lock_group(sb, group);
 893                        /* init the buddy */
 894                        memset(data, 0xff, blocksize);
 895                        ext4_mb_generate_buddy(sb, data, incore, group);
 896                        ext4_unlock_group(sb, group);
 897                        incore = NULL;
 898                } else {
 899                        /* this is block of bitmap */
 900                        BUG_ON(incore != NULL);
 901                        mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
 902                                group, page->index, i * blocksize);
 903                        trace_ext4_mb_bitmap_load(sb, group);
 904
 905                        /* see comments in ext4_mb_put_pa() */
 906                        ext4_lock_group(sb, group);
 907                        memcpy(data, bitmap, blocksize);
 908
 909                        /* mark all preallocated blks used in in-core bitmap */
 910                        ext4_mb_generate_from_pa(sb, data, group);
 911                        ext4_mb_generate_from_freelist(sb, data, group);
 912                        ext4_unlock_group(sb, group);
 913
 914                        /* set incore so that the buddy information can be
 915                         * generated using this
 916                         */
 917                        incore = data;
 918                }
 919        }
 920        SetPageUptodate(page);
 921
 922out:
 923        if (bh) {
 924                for (i = 0; i < groups_per_page; i++)
 925                        brelse(bh[i]);
 926                if (bh != &bhs)
 927                        kfree(bh);
 928        }
 929        return err;
 930}
 931
 932/*
 933 * Lock the buddy and bitmap pages. This make sure other parallel init_group
 934 * on the same buddy page doesn't happen whild holding the buddy page lock.
 935 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
 936 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
 937 */
 938static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
 939                ext4_group_t group, struct ext4_buddy *e4b)
 940{
 941        struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
 942        int block, pnum, poff;
 943        int blocks_per_page;
 944        struct page *page;
 945
 946        e4b->bd_buddy_page = NULL;
 947        e4b->bd_bitmap_page = NULL;
 948
 949        blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
 950        /*
 951         * the buddy cache inode stores the block bitmap
 952         * and buddy information in consecutive blocks.
 953         * So for each group we need two blocks.
 954         */
 955        block = group * 2;
 956        pnum = block / blocks_per_page;
 957        poff = block % blocks_per_page;
 958        page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
 959        if (!page)
 960                return -EIO;
 961        BUG_ON(page->mapping != inode->i_mapping);
 962        e4b->bd_bitmap_page = page;
 963        e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
 964
 965        if (blocks_per_page >= 2) {
 966                /* buddy and bitmap are on the same page */
 967                return 0;
 968        }
 969
 970        block++;
 971        pnum = block / blocks_per_page;
 972        page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
 973        if (!page)
 974                return -EIO;
 975        BUG_ON(page->mapping != inode->i_mapping);
 976        e4b->bd_buddy_page = page;
 977        return 0;
 978}
 979
 980static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
 981{
 982        if (e4b->bd_bitmap_page) {
 983                unlock_page(e4b->bd_bitmap_page);
 984                page_cache_release(e4b->bd_bitmap_page);
 985        }
 986        if (e4b->bd_buddy_page) {
 987                unlock_page(e4b->bd_buddy_page);
 988                page_cache_release(e4b->bd_buddy_page);
 989        }
 990}
 991
 992/*
 993 * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
 994 * block group lock of all groups for this page; do not hold the BG lock when
 995 * calling this routine!
 996 */
 997static noinline_for_stack
 998int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
 999{
1000
1001        struct ext4_group_info *this_grp;
1002        struct ext4_buddy e4b;
1003        struct page *page;
1004        int ret = 0;
1005
1006        mb_debug(1, "init group %u\n", group);
1007        this_grp = ext4_get_group_info(sb, group);
1008        /*
1009         * This ensures that we don't reinit the buddy cache
1010         * page which map to the group from which we are already
1011         * allocating. If we are looking at the buddy cache we would
1012         * have taken a reference using ext4_mb_load_buddy and that
1013         * would have pinned buddy page to page cache.
1014         */
1015        ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1016        if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1017                /*
1018                 * somebody initialized the group
1019                 * return without doing anything
1020                 */
1021                goto err;
1022        }
1023
1024        page = e4b.bd_bitmap_page;
1025        ret = ext4_mb_init_cache(page, NULL);
1026        if (ret)
1027                goto err;
1028        if (!PageUptodate(page)) {
1029                ret = -EIO;
1030                goto err;
1031        }
1032        mark_page_accessed(page);
1033
1034        if (e4b.bd_buddy_page == NULL) {
1035                /*
1036                 * If both the bitmap and buddy are in
1037                 * the same page we don't need to force
1038                 * init the buddy
1039                 */
1040                ret = 0;
1041                goto err;
1042        }
1043        /* init buddy cache */
1044        page = e4b.bd_buddy_page;
1045        ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1046        if (ret)
1047                goto err;
1048        if (!PageUptodate(page)) {
1049                ret = -EIO;
1050                goto err;
1051        }
1052        mark_page_accessed(page);
1053err:
1054        ext4_mb_put_buddy_page_lock(&e4b);
1055        return ret;
1056}
1057
1058/*
1059 * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1060 * block group lock of all groups for this page; do not hold the BG lock when
1061 * calling this routine!
1062 */
1063static noinline_for_stack int
1064ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1065                                        struct ext4_buddy *e4b)
1066{
1067        int blocks_per_page;
1068        int block;
1069        int pnum;
1070        int poff;
1071        struct page *page;
1072        int ret;
1073        struct ext4_group_info *grp;
1074        struct ext4_sb_info *sbi = EXT4_SB(sb);
1075        struct inode *inode = sbi->s_buddy_cache;
1076
1077        mb_debug(1, "load group %u\n", group);
1078
1079        blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1080        grp = ext4_get_group_info(sb, group);
1081
1082        e4b->bd_blkbits = sb->s_blocksize_bits;
1083        e4b->bd_info = grp;
1084        e4b->bd_sb = sb;
1085        e4b->bd_group = group;
1086        e4b->bd_buddy_page = NULL;
1087        e4b->bd_bitmap_page = NULL;
1088
1089        if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1090                /*
1091                 * we need full data about the group
1092                 * to make a good selection
1093                 */
1094                ret = ext4_mb_init_group(sb, group);
1095                if (ret)
1096                        return ret;
1097        }
1098
1099        /*
1100         * the buddy cache inode stores the block bitmap
1101         * and buddy information in consecutive blocks.
1102         * So for each group we need two blocks.
1103         */
1104        block = group * 2;
1105        pnum = block / blocks_per_page;
1106        poff = block % blocks_per_page;
1107
1108        /* we could use find_or_create_page(), but it locks page
1109         * what we'd like to avoid in fast path ... */
1110        page = find_get_page(inode->i_mapping, pnum);
1111        if (page == NULL || !PageUptodate(page)) {
1112                if (page)
1113                        /*
1114                         * drop the page reference and try
1115                         * to get the page with lock. If we
1116                         * are not uptodate that implies
1117                         * somebody just created the page but
1118                         * is yet to initialize the same. So
1119                         * wait for it to initialize.
1120                         */
1121                        page_cache_release(page);
1122                page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1123                if (page) {
1124                        BUG_ON(page->mapping != inode->i_mapping);
1125                        if (!PageUptodate(page)) {
1126                                ret = ext4_mb_init_cache(page, NULL);
1127                                if (ret) {
1128                                        unlock_page(page);
1129                                        goto err;
1130                                }
1131                                mb_cmp_bitmaps(e4b, page_address(page) +
1132                                               (poff * sb->s_blocksize));
1133                        }
1134                        unlock_page(page);
1135                }
1136        }
1137        if (page == NULL || !PageUptodate(page)) {
1138                ret = -EIO;
1139                goto err;
1140        }
1141        e4b->bd_bitmap_page = page;
1142        e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1143        mark_page_accessed(page);
1144
1145        block++;
1146        pnum = block / blocks_per_page;
1147        poff = block % blocks_per_page;
1148
1149        page = find_get_page(inode->i_mapping, pnum);
1150        if (page == NULL || !PageUptodate(page)) {
1151                if (page)
1152                        page_cache_release(page);
1153                page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1154                if (page) {
1155                        BUG_ON(page->mapping != inode->i_mapping);
1156                        if (!PageUptodate(page)) {
1157                                ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1158                                if (ret) {
1159                                        unlock_page(page);
1160                                        goto err;
1161                                }
1162                        }
1163                        unlock_page(page);
1164                }
1165        }
1166        if (page == NULL || !PageUptodate(page)) {
1167                ret = -EIO;
1168                goto err;
1169        }
1170        e4b->bd_buddy_page = page;
1171        e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1172        mark_page_accessed(page);
1173
1174        BUG_ON(e4b->bd_bitmap_page == NULL);
1175        BUG_ON(e4b->bd_buddy_page == NULL);
1176
1177        return 0;
1178
1179err:
1180        if (page)
1181                page_cache_release(page);
1182        if (e4b->bd_bitmap_page)
1183                page_cache_release(e4b->bd_bitmap_page);
1184        if (e4b->bd_buddy_page)
1185                page_cache_release(e4b->bd_buddy_page);
1186        e4b->bd_buddy = NULL;
1187        e4b->bd_bitmap = NULL;
1188        return ret;
1189}
1190
1191static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1192{
1193        if (e4b->bd_bitmap_page)
1194                page_cache_release(e4b->bd_bitmap_page);
1195        if (e4b->bd_buddy_page)
1196                page_cache_release(e4b->bd_buddy_page);
1197}
1198
1199
1200static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1201{
1202        int order = 1;
1203        void *bb;
1204
1205        BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1206        BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1207
1208        bb = e4b->bd_buddy;
1209        while (order <= e4b->bd_blkbits + 1) {
1210                block = block >> 1;
1211                if (!mb_test_bit(block, bb)) {
1212                        /* this block is part of buddy of order 'order' */
1213                        return order;
1214                }
1215                bb += 1 << (e4b->bd_blkbits - order);
1216                order++;
1217        }
1218        return 0;
1219}
1220
1221static void mb_clear_bits(void *bm, int cur, int len)
1222{
1223        __u32 *addr;
1224
1225        len = cur + len;
1226        while (cur < len) {
1227                if ((cur & 31) == 0 && (len - cur) >= 32) {
1228                        /* fast path: clear whole word at once */
1229                        addr = bm + (cur >> 3);
1230                        *addr = 0;
1231                        cur += 32;
1232                        continue;
1233                }
1234                mb_clear_bit(cur, bm);
1235                cur++;
1236        }
1237}
1238
1239void ext4_set_bits(void *bm, int cur, int len)
1240{
1241        __u32 *addr;
1242
1243        len = cur + len;
1244        while (cur < len) {
1245                if ((cur & 31) == 0 && (len - cur) >= 32) {
1246                        /* fast path: set whole word at once */
1247                        addr = bm + (cur >> 3);
1248                        *addr = 0xffffffff;
1249                        cur += 32;
1250                        continue;
1251                }
1252                mb_set_bit(cur, bm);
1253                cur++;
1254        }
1255}
1256
1257static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1258                          int first, int count)
1259{
1260        int block = 0;
1261        int max = 0;
1262        int order;
1263        void *buddy;
1264        void *buddy2;
1265        struct super_block *sb = e4b->bd_sb;
1266
1267        BUG_ON(first + count > (sb->s_blocksize << 3));
1268        assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1269        mb_check_buddy(e4b);
1270        mb_free_blocks_double(inode, e4b, first, count);
1271
1272        e4b->bd_info->bb_free += count;
1273        if (first < e4b->bd_info->bb_first_free)
1274                e4b->bd_info->bb_first_free = first;
1275
1276        /* let's maintain fragments counter */
1277        if (first != 0)
1278                block = !mb_test_bit(first - 1, e4b->bd_bitmap);
1279        if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1280                max = !mb_test_bit(first + count, e4b->bd_bitmap);
1281        if (block && max)
1282                e4b->bd_info->bb_fragments--;
1283        else if (!block && !max)
1284                e4b->bd_info->bb_fragments++;
1285
1286        /* let's maintain buddy itself */
1287        while (count-- > 0) {
1288                block = first++;
1289                order = 0;
1290
1291                if (!mb_test_bit(block, e4b->bd_bitmap)) {
1292                        ext4_fsblk_t blocknr;
1293
1294                        blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1295                        blocknr += EXT4_C2B(EXT4_SB(sb), block);
1296                        ext4_grp_locked_error(sb, e4b->bd_group,
1297                                              inode ? inode->i_ino : 0,
1298                                              blocknr,
1299                                              "freeing already freed block "
1300                                              "(bit %u)", block);
1301                }
1302                mb_clear_bit(block, e4b->bd_bitmap);
1303                e4b->bd_info->bb_counters[order]++;
1304
1305                /* start of the buddy */
1306                buddy = mb_find_buddy(e4b, order, &max);
1307
1308                do {
1309                        block &= ~1UL;
1310                        if (mb_test_bit(block, buddy) ||
1311                                        mb_test_bit(block + 1, buddy))
1312                                break;
1313
1314                        /* both the buddies are free, try to coalesce them */
1315                        buddy2 = mb_find_buddy(e4b, order + 1, &max);
1316
1317                        if (!buddy2)
1318                                break;
1319
1320                        if (order > 0) {
1321                                /* for special purposes, we don't set
1322                                 * free bits in bitmap */
1323                                mb_set_bit(block, buddy);
1324                                mb_set_bit(block + 1, buddy);
1325                        }
1326                        e4b->bd_info->bb_counters[order]--;
1327                        e4b->bd_info->bb_counters[order]--;
1328
1329                        block = block >> 1;
1330                        order++;
1331                        e4b->bd_info->bb_counters[order]++;
1332
1333                        mb_clear_bit(block, buddy2);
1334                        buddy = buddy2;
1335                } while (1);
1336        }
1337        mb_set_largest_free_order(sb, e4b->bd_info);
1338        mb_check_buddy(e4b);
1339}
1340
1341static int mb_find_extent(struct ext4_buddy *e4b, int order, int block,
1342                                int needed, struct ext4_free_extent *ex)
1343{
1344        int next = block;
1345        int max;
1346        void *buddy;
1347
1348        assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1349        BUG_ON(ex == NULL);
1350
1351        buddy = mb_find_buddy(e4b, order, &max);
1352        BUG_ON(buddy == NULL);
1353        BUG_ON(block >= max);
1354        if (mb_test_bit(block, buddy)) {
1355                ex->fe_len = 0;
1356                ex->fe_start = 0;
1357                ex->fe_group = 0;
1358                return 0;
1359        }
1360
1361        /* FIXME dorp order completely ? */
1362        if (likely(order == 0)) {
1363                /* find actual order */
1364                order = mb_find_order_for_block(e4b, block);
1365                block = block >> order;
1366        }
1367
1368        ex->fe_len = 1 << order;
1369        ex->fe_start = block << order;
1370        ex->fe_group = e4b->bd_group;
1371
1372        /* calc difference from given start */
1373        next = next - ex->fe_start;
1374        ex->fe_len -= next;
1375        ex->fe_start += next;
1376
1377        while (needed > ex->fe_len &&
1378               (buddy = mb_find_buddy(e4b, order, &max))) {
1379
1380                if (block + 1 >= max)
1381                        break;
1382
1383                next = (block + 1) * (1 << order);
1384                if (mb_test_bit(next, e4b->bd_bitmap))
1385                        break;
1386
1387                order = mb_find_order_for_block(e4b, next);
1388
1389                block = next >> order;
1390                ex->fe_len += 1 << order;
1391        }
1392
1393        BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1394        return ex->fe_len;
1395}
1396
1397static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1398{
1399        int ord;
1400        int mlen = 0;
1401        int max = 0;
1402        int cur;
1403        int start = ex->fe_start;
1404        int len = ex->fe_len;
1405        unsigned ret = 0;
1406        int len0 = len;
1407        void *buddy;
1408
1409        BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1410        BUG_ON(e4b->bd_group != ex->fe_group);
1411        assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1412        mb_check_buddy(e4b);
1413        mb_mark_used_double(e4b, start, len);
1414
1415        e4b->bd_info->bb_free -= len;
1416        if (e4b->bd_info->bb_first_free == start)
1417                e4b->bd_info->bb_first_free += len;
1418
1419        /* let's maintain fragments counter */
1420        if (start != 0)
1421                mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1422        if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1423                max = !mb_test_bit(start + len, e4b->bd_bitmap);
1424        if (mlen && max)
1425                e4b->bd_info->bb_fragments++;
1426        else if (!mlen && !max)
1427                e4b->bd_info->bb_fragments--;
1428
1429        /* let's maintain buddy itself */
1430        while (len) {
1431                ord = mb_find_order_for_block(e4b, start);
1432
1433                if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1434                        /* the whole chunk may be allocated at once! */
1435                        mlen = 1 << ord;
1436                        buddy = mb_find_buddy(e4b, ord, &max);
1437                        BUG_ON((start >> ord) >= max);
1438                        mb_set_bit(start >> ord, buddy);
1439                        e4b->bd_info->bb_counters[ord]--;
1440                        start += mlen;
1441                        len -= mlen;
1442                        BUG_ON(len < 0);
1443                        continue;
1444                }
1445
1446                /* store for history */
1447                if (ret == 0)
1448                        ret = len | (ord << 16);
1449
1450                /* we have to split large buddy */
1451                BUG_ON(ord <= 0);
1452                buddy = mb_find_buddy(e4b, ord, &max);
1453                mb_set_bit(start >> ord, buddy);
1454                e4b->bd_info->bb_counters[ord]--;
1455
1456                ord--;
1457                cur = (start >> ord) & ~1U;
1458                buddy = mb_find_buddy(e4b, ord, &max);
1459                mb_clear_bit(cur, buddy);
1460                mb_clear_bit(cur + 1, buddy);
1461                e4b->bd_info->bb_counters[ord]++;
1462                e4b->bd_info->bb_counters[ord]++;
1463        }
1464        mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1465
1466        ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1467        mb_check_buddy(e4b);
1468
1469        return ret;
1470}
1471
1472/*
1473 * Must be called under group lock!
1474 */
1475static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1476                                        struct ext4_buddy *e4b)
1477{
1478        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1479        int ret;
1480
1481        BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1482        BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1483
1484        ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1485        ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1486        ret = mb_mark_used(e4b, &ac->ac_b_ex);
1487
1488        /* preallocation can change ac_b_ex, thus we store actually
1489         * allocated blocks for history */
1490        ac->ac_f_ex = ac->ac_b_ex;
1491
1492        ac->ac_status = AC_STATUS_FOUND;
1493        ac->ac_tail = ret & 0xffff;
1494        ac->ac_buddy = ret >> 16;
1495
1496        /*
1497         * take the page reference. We want the page to be pinned
1498         * so that we don't get a ext4_mb_init_cache_call for this
1499         * group until we update the bitmap. That would mean we
1500         * double allocate blocks. The reference is dropped
1501         * in ext4_mb_release_context
1502         */
1503        ac->ac_bitmap_page = e4b->bd_bitmap_page;
1504        get_page(ac->ac_bitmap_page);
1505        ac->ac_buddy_page = e4b->bd_buddy_page;
1506        get_page(ac->ac_buddy_page);
1507        /* store last allocated for subsequent stream allocation */
1508        if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1509                spin_lock(&sbi->s_md_lock);
1510                sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1511                sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1512                spin_unlock(&sbi->s_md_lock);
1513        }
1514}
1515
1516/*
1517 * regular allocator, for general purposes allocation
1518 */
1519
1520static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1521                                        struct ext4_buddy *e4b,
1522                                        int finish_group)
1523{
1524        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1525        struct ext4_free_extent *bex = &ac->ac_b_ex;
1526        struct ext4_free_extent *gex = &ac->ac_g_ex;
1527        struct ext4_free_extent ex;
1528        int max;
1529
1530        if (ac->ac_status == AC_STATUS_FOUND)
1531                return;
1532        /*
1533         * We don't want to scan for a whole year
1534         */
1535        if (ac->ac_found > sbi->s_mb_max_to_scan &&
1536                        !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1537                ac->ac_status = AC_STATUS_BREAK;
1538                return;
1539        }
1540
1541        /*
1542         * Haven't found good chunk so far, let's continue
1543         */
1544        if (bex->fe_len < gex->fe_len)
1545                return;
1546
1547        if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1548                        && bex->fe_group == e4b->bd_group) {
1549                /* recheck chunk's availability - we don't know
1550                 * when it was found (within this lock-unlock
1551                 * period or not) */
1552                max = mb_find_extent(e4b, 0, bex->fe_start, gex->fe_len, &ex);
1553                if (max >= gex->fe_len) {
1554                        ext4_mb_use_best_found(ac, e4b);
1555                        return;
1556                }
1557        }
1558}
1559
1560/*
1561 * The routine checks whether found extent is good enough. If it is,
1562 * then the extent gets marked used and flag is set to the context
1563 * to stop scanning. Otherwise, the extent is compared with the
1564 * previous found extent and if new one is better, then it's stored
1565 * in the context. Later, the best found extent will be used, if
1566 * mballoc can't find good enough extent.
1567 *
1568 * FIXME: real allocation policy is to be designed yet!
1569 */
1570static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1571                                        struct ext4_free_extent *ex,
1572                                        struct ext4_buddy *e4b)
1573{
1574        struct ext4_free_extent *bex = &ac->ac_b_ex;
1575        struct ext4_free_extent *gex = &ac->ac_g_ex;
1576
1577        BUG_ON(ex->fe_len <= 0);
1578        BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1579        BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1580        BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1581
1582        ac->ac_found++;
1583
1584        /*
1585         * The special case - take what you catch first
1586         */
1587        if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1588                *bex = *ex;
1589                ext4_mb_use_best_found(ac, e4b);
1590                return;
1591        }
1592
1593        /*
1594         * Let's check whether the chuck is good enough
1595         */
1596        if (ex->fe_len == gex->fe_len) {
1597                *bex = *ex;
1598                ext4_mb_use_best_found(ac, e4b);
1599                return;
1600        }
1601
1602        /*
1603         * If this is first found extent, just store it in the context
1604         */
1605        if (bex->fe_len == 0) {
1606                *bex = *ex;
1607                return;
1608        }
1609
1610        /*
1611         * If new found extent is better, store it in the context
1612         */
1613        if (bex->fe_len < gex->fe_len) {
1614                /* if the request isn't satisfied, any found extent
1615                 * larger than previous best one is better */
1616                if (ex->fe_len > bex->fe_len)
1617                        *bex = *ex;
1618        } else if (ex->fe_len > gex->fe_len) {
1619                /* if the request is satisfied, then we try to find
1620                 * an extent that still satisfy the request, but is
1621                 * smaller than previous one */
1622                if (ex->fe_len < bex->fe_len)
1623                        *bex = *ex;
1624        }
1625
1626        ext4_mb_check_limits(ac, e4b, 0);
1627}
1628
1629static noinline_for_stack
1630int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1631                                        struct ext4_buddy *e4b)
1632{
1633        struct ext4_free_extent ex = ac->ac_b_ex;
1634        ext4_group_t group = ex.fe_group;
1635        int max;
1636        int err;
1637
1638        BUG_ON(ex.fe_len <= 0);
1639        err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1640        if (err)
1641                return err;
1642
1643        ext4_lock_group(ac->ac_sb, group);
1644        max = mb_find_extent(e4b, 0, ex.fe_start, ex.fe_len, &ex);
1645
1646        if (max > 0) {
1647                ac->ac_b_ex = ex;
1648                ext4_mb_use_best_found(ac, e4b);
1649        }
1650
1651        ext4_unlock_group(ac->ac_sb, group);
1652        ext4_mb_unload_buddy(e4b);
1653
1654        return 0;
1655}
1656
1657static noinline_for_stack
1658int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1659                                struct ext4_buddy *e4b)
1660{
1661        ext4_group_t group = ac->ac_g_ex.fe_group;
1662        int max;
1663        int err;
1664        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1665        struct ext4_free_extent ex;
1666
1667        if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1668                return 0;
1669
1670        err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1671        if (err)
1672                return err;
1673
1674        ext4_lock_group(ac->ac_sb, group);
1675        max = mb_find_extent(e4b, 0, ac->ac_g_ex.fe_start,
1676                             ac->ac_g_ex.fe_len, &ex);
1677
1678        if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1679                ext4_fsblk_t start;
1680
1681                start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1682                        ex.fe_start;
1683                /* use do_div to get remainder (would be 64-bit modulo) */
1684                if (do_div(start, sbi->s_stripe) == 0) {
1685                        ac->ac_found++;
1686                        ac->ac_b_ex = ex;
1687                        ext4_mb_use_best_found(ac, e4b);
1688                }
1689        } else if (max >= ac->ac_g_ex.fe_len) {
1690                BUG_ON(ex.fe_len <= 0);
1691                BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1692                BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1693                ac->ac_found++;
1694                ac->ac_b_ex = ex;
1695                ext4_mb_use_best_found(ac, e4b);
1696        } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1697                /* Sometimes, caller may want to merge even small
1698                 * number of blocks to an existing extent */
1699                BUG_ON(ex.fe_len <= 0);
1700                BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1701                BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1702                ac->ac_found++;
1703                ac->ac_b_ex = ex;
1704                ext4_mb_use_best_found(ac, e4b);
1705        }
1706        ext4_unlock_group(ac->ac_sb, group);
1707        ext4_mb_unload_buddy(e4b);
1708
1709        return 0;
1710}
1711
1712/*
1713 * The routine scans buddy structures (not bitmap!) from given order
1714 * to max order and tries to find big enough chunk to satisfy the req
1715 */
1716static noinline_for_stack
1717void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1718                                        struct ext4_buddy *e4b)
1719{
1720        struct super_block *sb = ac->ac_sb;
1721        struct ext4_group_info *grp = e4b->bd_info;
1722        void *buddy;
1723        int i;
1724        int k;
1725        int max;
1726
1727        BUG_ON(ac->ac_2order <= 0);
1728        for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1729                if (grp->bb_counters[i] == 0)
1730                        continue;
1731
1732                buddy = mb_find_buddy(e4b, i, &max);
1733                BUG_ON(buddy == NULL);
1734
1735                k = mb_find_next_zero_bit(buddy, max, 0);
1736                BUG_ON(k >= max);
1737
1738                ac->ac_found++;
1739
1740                ac->ac_b_ex.fe_len = 1 << i;
1741                ac->ac_b_ex.fe_start = k << i;
1742                ac->ac_b_ex.fe_group = e4b->bd_group;
1743
1744                ext4_mb_use_best_found(ac, e4b);
1745
1746                BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1747
1748                if (EXT4_SB(sb)->s_mb_stats)
1749                        atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1750
1751                break;
1752        }
1753}
1754
1755/*
1756 * The routine scans the group and measures all found extents.
1757 * In order to optimize scanning, caller must pass number of
1758 * free blocks in the group, so the routine can know upper limit.
1759 */
1760static noinline_for_stack
1761void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1762                                        struct ext4_buddy *e4b)
1763{
1764        struct super_block *sb = ac->ac_sb;
1765        void *bitmap = e4b->bd_bitmap;
1766        struct ext4_free_extent ex;
1767        int i;
1768        int free;
1769
1770        free = e4b->bd_info->bb_free;
1771        BUG_ON(free <= 0);
1772
1773        i = e4b->bd_info->bb_first_free;
1774
1775        while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1776                i = mb_find_next_zero_bit(bitmap,
1777                                                EXT4_CLUSTERS_PER_GROUP(sb), i);
1778                if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1779                        /*
1780                         * IF we have corrupt bitmap, we won't find any
1781                         * free blocks even though group info says we
1782                         * we have free blocks
1783                         */
1784                        ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1785                                        "%d free clusters as per "
1786                                        "group info. But bitmap says 0",
1787                                        free);
1788                        break;
1789                }
1790
1791                mb_find_extent(e4b, 0, i, ac->ac_g_ex.fe_len, &ex);
1792                BUG_ON(ex.fe_len <= 0);
1793                if (free < ex.fe_len) {
1794                        ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1795                                        "%d free clusters as per "
1796                                        "group info. But got %d blocks",
1797                                        free, ex.fe_len);
1798                        /*
1799                         * The number of free blocks differs. This mostly
1800                         * indicate that the bitmap is corrupt. So exit
1801                         * without claiming the space.
1802                         */
1803                        break;
1804                }
1805
1806                ext4_mb_measure_extent(ac, &ex, e4b);
1807
1808                i += ex.fe_len;
1809                free -= ex.fe_len;
1810        }
1811
1812        ext4_mb_check_limits(ac, e4b, 1);
1813}
1814
1815/*
1816 * This is a special case for storages like raid5
1817 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1818 */
1819static noinline_for_stack
1820void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1821                                 struct ext4_buddy *e4b)
1822{
1823        struct super_block *sb = ac->ac_sb;
1824        struct ext4_sb_info *sbi = EXT4_SB(sb);
1825        void *bitmap = e4b->bd_bitmap;
1826        struct ext4_free_extent ex;
1827        ext4_fsblk_t first_group_block;
1828        ext4_fsblk_t a;
1829        ext4_grpblk_t i;
1830        int max;
1831
1832        BUG_ON(sbi->s_stripe == 0);
1833
1834        /* find first stripe-aligned block in group */
1835        first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1836
1837        a = first_group_block + sbi->s_stripe - 1;
1838        do_div(a, sbi->s_stripe);
1839        i = (a * sbi->s_stripe) - first_group_block;
1840
1841        while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1842                if (!mb_test_bit(i, bitmap)) {
1843                        max = mb_find_extent(e4b, 0, i, sbi->s_stripe, &ex);
1844                        if (max >= sbi->s_stripe) {
1845                                ac->ac_found++;
1846                                ac->ac_b_ex = ex;
1847                                ext4_mb_use_best_found(ac, e4b);
1848                                break;
1849                        }
1850                }
1851                i += sbi->s_stripe;
1852        }
1853}
1854
1855/* This is now called BEFORE we load the buddy bitmap. */
1856static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1857                                ext4_group_t group, int cr)
1858{
1859        unsigned free, fragments;
1860        int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1861        struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1862
1863        BUG_ON(cr < 0 || cr >= 4);
1864
1865        /* We only do this if the grp has never been initialized */
1866        if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1867                int ret = ext4_mb_init_group(ac->ac_sb, group);
1868                if (ret)
1869                        return 0;
1870        }
1871
1872        free = grp->bb_free;
1873        fragments = grp->bb_fragments;
1874        if (free == 0)
1875                return 0;
1876        if (fragments == 0)
1877                return 0;
1878
1879        switch (cr) {
1880        case 0:
1881                BUG_ON(ac->ac_2order == 0);
1882
1883                if (grp->bb_largest_free_order < ac->ac_2order)
1884                        return 0;
1885
1886                /* Avoid using the first bg of a flexgroup for data files */
1887                if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1888                    (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1889                    ((group % flex_size) == 0))
1890                        return 0;
1891
1892                return 1;
1893        case 1:
1894                if ((free / fragments) >= ac->ac_g_ex.fe_len)
1895                        return 1;
1896                break;
1897        case 2:
1898                if (free >= ac->ac_g_ex.fe_len)
1899                        return 1;
1900                break;
1901        case 3:
1902                return 1;
1903        default:
1904                BUG();
1905        }
1906
1907        return 0;
1908}
1909
1910static noinline_for_stack int
1911ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1912{
1913        ext4_group_t ngroups, group, i;
1914        int cr;
1915        int err = 0;
1916        struct ext4_sb_info *sbi;
1917        struct super_block *sb;
1918        struct ext4_buddy e4b;
1919
1920        sb = ac->ac_sb;
1921        sbi = EXT4_SB(sb);
1922        ngroups = ext4_get_groups_count(sb);
1923        /* non-extent files are limited to low blocks/groups */
1924        if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1925                ngroups = sbi->s_blockfile_groups;
1926
1927        BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1928
1929        /* first, try the goal */
1930        err = ext4_mb_find_by_goal(ac, &e4b);
1931        if (err || ac->ac_status == AC_STATUS_FOUND)
1932                goto out;
1933
1934        if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1935                goto out;
1936
1937        /*
1938         * ac->ac2_order is set only if the fe_len is a power of 2
1939         * if ac2_order is set we also set criteria to 0 so that we
1940         * try exact allocation using buddy.
1941         */
1942        i = fls(ac->ac_g_ex.fe_len);
1943        ac->ac_2order = 0;
1944        /*
1945         * We search using buddy data only if the order of the request
1946         * is greater than equal to the sbi_s_mb_order2_reqs
1947         * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1948         */
1949        if (i >= sbi->s_mb_order2_reqs) {
1950                /*
1951                 * This should tell if fe_len is exactly power of 2
1952                 */
1953                if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1954                        ac->ac_2order = i - 1;
1955        }
1956
1957        /* if stream allocation is enabled, use global goal */
1958        if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1959                /* TBD: may be hot point */
1960                spin_lock(&sbi->s_md_lock);
1961                ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1962                ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1963                spin_unlock(&sbi->s_md_lock);
1964        }
1965
1966        /* Let's just scan groups to find more-less suitable blocks */
1967        cr = ac->ac_2order ? 0 : 1;
1968        /*
1969         * cr == 0 try to get exact allocation,
1970         * cr == 3  try to get anything
1971         */
1972repeat:
1973        for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1974                ac->ac_criteria = cr;
1975                /*
1976                 * searching for the right group start
1977                 * from the goal value specified
1978                 */
1979                group = ac->ac_g_ex.fe_group;
1980
1981                for (i = 0; i < ngroups; group++, i++) {
1982                        if (group == ngroups)
1983                                group = 0;
1984
1985                        /* This now checks without needing the buddy page */
1986                        if (!ext4_mb_good_group(ac, group, cr))
1987                                continue;
1988
1989                        err = ext4_mb_load_buddy(sb, group, &e4b);
1990                        if (err)
1991                                goto out;
1992
1993                        ext4_lock_group(sb, group);
1994
1995                        /*
1996                         * We need to check again after locking the
1997                         * block group
1998                         */
1999                        if (!ext4_mb_good_group(ac, group, cr)) {
2000                                ext4_unlock_group(sb, group);
2001                                ext4_mb_unload_buddy(&e4b);
2002                                continue;
2003                        }
2004
2005                        ac->ac_groups_scanned++;
2006                        if (cr == 0)
2007                                ext4_mb_simple_scan_group(ac, &e4b);
2008                        else if (cr == 1 && sbi->s_stripe &&
2009                                        !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2010                                ext4_mb_scan_aligned(ac, &e4b);
2011                        else
2012                                ext4_mb_complex_scan_group(ac, &e4b);
2013
2014                        ext4_unlock_group(sb, group);
2015                        ext4_mb_unload_buddy(&e4b);
2016
2017                        if (ac->ac_status != AC_STATUS_CONTINUE)
2018                                break;
2019                }
2020        }
2021
2022        if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2023            !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2024                /*
2025                 * We've been searching too long. Let's try to allocate
2026                 * the best chunk we've found so far
2027                 */
2028
2029                ext4_mb_try_best_found(ac, &e4b);
2030                if (ac->ac_status != AC_STATUS_FOUND) {
2031                        /*
2032                         * Someone more lucky has already allocated it.
2033                         * The only thing we can do is just take first
2034                         * found block(s)
2035                        printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2036                         */
2037                        ac->ac_b_ex.fe_group = 0;
2038                        ac->ac_b_ex.fe_start = 0;
2039                        ac->ac_b_ex.fe_len = 0;
2040                        ac->ac_status = AC_STATUS_CONTINUE;
2041                        ac->ac_flags |= EXT4_MB_HINT_FIRST;
2042                        cr = 3;
2043                        atomic_inc(&sbi->s_mb_lost_chunks);
2044                        goto repeat;
2045                }
2046        }
2047out:
2048        return err;
2049}
2050
2051static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2052{
2053        struct super_block *sb = seq->private;
2054        ext4_group_t group;
2055
2056        if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2057                return NULL;
2058        group = *pos + 1;
2059        return (void *) ((unsigned long) group);
2060}
2061
2062static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2063{
2064        struct super_block *sb = seq->private;
2065        ext4_group_t group;
2066
2067        ++*pos;
2068        if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2069                return NULL;
2070        group = *pos + 1;
2071        return (void *) ((unsigned long) group);
2072}
2073
2074static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2075{
2076        struct super_block *sb = seq->private;
2077        ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2078        int i;
2079        int err, buddy_loaded = 0;
2080        struct ext4_buddy e4b;
2081        struct ext4_group_info *grinfo;
2082        struct sg {
2083                struct ext4_group_info info;
2084                ext4_grpblk_t counters[16];
2085        } sg;
2086
2087        group--;
2088        if (group == 0)
2089                seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2090                                "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2091                                  "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2092                           "group", "free", "frags", "first",
2093                           "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2094                           "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2095
2096        i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2097                sizeof(struct ext4_group_info);
2098        grinfo = ext4_get_group_info(sb, group);
2099        /* Load the group info in memory only if not already loaded. */
2100        if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2101                err = ext4_mb_load_buddy(sb, group, &e4b);
2102                if (err) {
2103                        seq_printf(seq, "#%-5u: I/O error\n", group);
2104                        return 0;
2105                }
2106                buddy_loaded = 1;
2107        }
2108
2109        memcpy(&sg, ext4_get_group_info(sb, group), i);
2110
2111        if (buddy_loaded)
2112                ext4_mb_unload_buddy(&e4b);
2113
2114        seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2115                        sg.info.bb_fragments, sg.info.bb_first_free);
2116        for (i = 0; i <= 13; i++)
2117                seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2118                                sg.info.bb_counters[i] : 0);
2119        seq_printf(seq, " ]\n");
2120
2121        return 0;
2122}
2123
2124static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2125{
2126}
2127
2128static const struct seq_operations ext4_mb_seq_groups_ops = {
2129        .start  = ext4_mb_seq_groups_start,
2130        .next   = ext4_mb_seq_groups_next,
2131        .stop   = ext4_mb_seq_groups_stop,
2132        .show   = ext4_mb_seq_groups_show,
2133};
2134
2135static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2136{
2137        struct super_block *sb = PDE(inode)->data;
2138        int rc;
2139
2140        rc = seq_open(file, &ext4_mb_seq_groups_ops);
2141        if (rc == 0) {
2142                struct seq_file *m = file->private_data;
2143                m->private = sb;
2144        }
2145        return rc;
2146
2147}
2148
2149static const struct file_operations ext4_mb_seq_groups_fops = {
2150        .owner          = THIS_MODULE,
2151        .open           = ext4_mb_seq_groups_open,
2152        .read           = seq_read,
2153        .llseek         = seq_lseek,
2154        .release        = seq_release,
2155};
2156
2157static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2158{
2159        int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2160        struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2161
2162        BUG_ON(!cachep);
2163        return cachep;
2164}
2165
2166/* Create and initialize ext4_group_info data for the given group. */
2167int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2168                          struct ext4_group_desc *desc)
2169{
2170        int i;
2171        int metalen = 0;
2172        struct ext4_sb_info *sbi = EXT4_SB(sb);
2173        struct ext4_group_info **meta_group_info;
2174        struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2175
2176        /*
2177         * First check if this group is the first of a reserved block.
2178         * If it's true, we have to allocate a new table of pointers
2179         * to ext4_group_info structures
2180         */
2181        if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2182                metalen = sizeof(*meta_group_info) <<
2183                        EXT4_DESC_PER_BLOCK_BITS(sb);
2184                meta_group_info = kmalloc(metalen, GFP_KERNEL);
2185                if (meta_group_info == NULL) {
2186                        ext4_msg(sb, KERN_ERR, "can't allocate mem "
2187                                 "for a buddy group");
2188                        goto exit_meta_group_info;
2189                }
2190                sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2191                        meta_group_info;
2192        }
2193
2194        meta_group_info =
2195                sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2196        i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2197
2198        meta_group_info[i] = kmem_cache_alloc(cachep, GFP_KERNEL);
2199        if (meta_group_info[i] == NULL) {
2200                ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2201                goto exit_group_info;
2202        }
2203        memset(meta_group_info[i], 0, kmem_cache_size(cachep));
2204        set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2205                &(meta_group_info[i]->bb_state));
2206
2207        /*
2208         * initialize bb_free to be able to skip
2209         * empty groups without initialization
2210         */
2211        if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2212                meta_group_info[i]->bb_free =
2213                        ext4_free_clusters_after_init(sb, group, desc);
2214        } else {
2215                meta_group_info[i]->bb_free =
2216                        ext4_free_group_clusters(sb, desc);
2217        }
2218
2219        INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2220        init_rwsem(&meta_group_info[i]->alloc_sem);
2221        meta_group_info[i]->bb_free_root = RB_ROOT;
2222        meta_group_info[i]->bb_largest_free_order = -1;  /* uninit */
2223
2224#ifdef DOUBLE_CHECK
2225        {
2226                struct buffer_head *bh;
2227                meta_group_info[i]->bb_bitmap =
2228                        kmalloc(sb->s_blocksize, GFP_KERNEL);
2229                BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2230                bh = ext4_read_block_bitmap(sb, group);
2231                BUG_ON(bh == NULL);
2232                memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2233                        sb->s_blocksize);
2234                put_bh(bh);
2235        }
2236#endif
2237
2238        return 0;
2239
2240exit_group_info:
2241        /* If a meta_group_info table has been allocated, release it now */
2242        if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2243                kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2244                sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2245        }
2246exit_meta_group_info:
2247        return -ENOMEM;
2248} /* ext4_mb_add_groupinfo */
2249
2250static int ext4_mb_init_backend(struct super_block *sb)
2251{
2252        ext4_group_t ngroups = ext4_get_groups_count(sb);
2253        ext4_group_t i;
2254        struct ext4_sb_info *sbi = EXT4_SB(sb);
2255        struct ext4_super_block *es = sbi->s_es;
2256        int num_meta_group_infos;
2257        int num_meta_group_infos_max;
2258        int array_size;
2259        struct ext4_group_desc *desc;
2260        struct kmem_cache *cachep;
2261
2262        /* This is the number of blocks used by GDT */
2263        num_meta_group_infos = (ngroups + EXT4_DESC_PER_BLOCK(sb) -
2264                                1) >> EXT4_DESC_PER_BLOCK_BITS(sb);
2265
2266        /*
2267         * This is the total number of blocks used by GDT including
2268         * the number of reserved blocks for GDT.
2269         * The s_group_info array is allocated with this value
2270         * to allow a clean online resize without a complex
2271         * manipulation of pointer.
2272         * The drawback is the unused memory when no resize
2273         * occurs but it's very low in terms of pages
2274         * (see comments below)
2275         * Need to handle this properly when META_BG resizing is allowed
2276         */
2277        num_meta_group_infos_max = num_meta_group_infos +
2278                                le16_to_cpu(es->s_reserved_gdt_blocks);
2279
2280        /*
2281         * array_size is the size of s_group_info array. We round it
2282         * to the next power of two because this approximation is done
2283         * internally by kmalloc so we can have some more memory
2284         * for free here (e.g. may be used for META_BG resize).
2285         */
2286        array_size = 1;
2287        while (array_size < sizeof(*sbi->s_group_info) *
2288               num_meta_group_infos_max)
2289                array_size = array_size << 1;
2290        /* An 8TB filesystem with 64-bit pointers requires a 4096 byte
2291         * kmalloc. A 128kb malloc should suffice for a 256TB filesystem.
2292         * So a two level scheme suffices for now. */
2293        sbi->s_group_info = ext4_kvzalloc(array_size, GFP_KERNEL);
2294        if (sbi->s_group_info == NULL) {
2295                ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2296                return -ENOMEM;
2297        }
2298        sbi->s_buddy_cache = new_inode(sb);
2299        if (sbi->s_buddy_cache == NULL) {
2300                ext4_msg(sb, KERN_ERR, "can't get new inode");
2301                goto err_freesgi;
2302        }
2303        /* To avoid potentially colliding with an valid on-disk inode number,
2304         * use EXT4_BAD_INO for the buddy cache inode number.  This inode is
2305         * not in the inode hash, so it should never be found by iget(), but
2306         * this will avoid confusion if it ever shows up during debugging. */
2307        sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2308        EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2309        for (i = 0; i < ngroups; i++) {
2310                desc = ext4_get_group_desc(sb, i, NULL);
2311                if (desc == NULL) {
2312                        ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2313                        goto err_freebuddy;
2314                }
2315                if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2316                        goto err_freebuddy;
2317        }
2318
2319        return 0;
2320
2321err_freebuddy:
2322        cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2323        while (i-- > 0)
2324                kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2325        i = num_meta_group_infos;
2326        while (i-- > 0)
2327                kfree(sbi->s_group_info[i]);
2328        iput(sbi->s_buddy_cache);
2329err_freesgi:
2330        ext4_kvfree(sbi->s_group_info);
2331        return -ENOMEM;
2332}
2333
2334static void ext4_groupinfo_destroy_slabs(void)
2335{
2336        int i;
2337
2338        for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2339                if (ext4_groupinfo_caches[i])
2340                        kmem_cache_destroy(ext4_groupinfo_caches[i]);
2341                ext4_groupinfo_caches[i] = NULL;
2342        }
2343}
2344
2345static int ext4_groupinfo_create_slab(size_t size)
2346{
2347        static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2348        int slab_size;
2349        int blocksize_bits = order_base_2(size);
2350        int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2351        struct kmem_cache *cachep;
2352
2353        if (cache_index >= NR_GRPINFO_CACHES)
2354                return -EINVAL;
2355
2356        if (unlikely(cache_index < 0))
2357                cache_index = 0;
2358
2359        mutex_lock(&ext4_grpinfo_slab_create_mutex);
2360        if (ext4_groupinfo_caches[cache_index]) {
2361                mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2362                return 0;       /* Already created */
2363        }
2364
2365        slab_size = offsetof(struct ext4_group_info,
2366                                bb_counters[blocksize_bits + 2]);
2367
2368        cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2369                                        slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2370                                        NULL);
2371
2372        ext4_groupinfo_caches[cache_index] = cachep;
2373
2374        mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2375        if (!cachep) {
2376                printk(KERN_EMERG
2377                       "EXT4-fs: no memory for groupinfo slab cache\n");
2378                return -ENOMEM;
2379        }
2380
2381        return 0;
2382}
2383
2384int ext4_mb_init(struct super_block *sb)
2385{
2386        struct ext4_sb_info *sbi = EXT4_SB(sb);
2387        unsigned i, j;
2388        unsigned offset;
2389        unsigned max;
2390        int ret;
2391
2392        i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2393
2394        sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2395        if (sbi->s_mb_offsets == NULL) {
2396                ret = -ENOMEM;
2397                goto out;
2398        }
2399
2400        i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2401        sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2402        if (sbi->s_mb_maxs == NULL) {
2403                ret = -ENOMEM;
2404                goto out;
2405        }
2406
2407        ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2408        if (ret < 0)
2409                goto out;
2410
2411        /* order 0 is regular bitmap */
2412        sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2413        sbi->s_mb_offsets[0] = 0;
2414
2415        i = 1;
2416        offset = 0;
2417        max = sb->s_blocksize << 2;
2418        do {
2419                sbi->s_mb_offsets[i] = offset;
2420                sbi->s_mb_maxs[i] = max;
2421                offset += 1 << (sb->s_blocksize_bits - i);
2422                max = max >> 1;
2423                i++;
2424        } while (i <= sb->s_blocksize_bits + 1);
2425
2426        spin_lock_init(&sbi->s_md_lock);
2427        spin_lock_init(&sbi->s_bal_lock);
2428
2429        sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2430        sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2431        sbi->s_mb_stats = MB_DEFAULT_STATS;
2432        sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2433        sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2434        /*
2435         * The default group preallocation is 512, which for 4k block
2436         * sizes translates to 2 megabytes.  However for bigalloc file
2437         * systems, this is probably too big (i.e, if the cluster size
2438         * is 1 megabyte, then group preallocation size becomes half a
2439         * gigabyte!).  As a default, we will keep a two megabyte
2440         * group pralloc size for cluster sizes up to 64k, and after
2441         * that, we will force a minimum group preallocation size of
2442         * 32 clusters.  This translates to 8 megs when the cluster
2443         * size is 256k, and 32 megs when the cluster size is 1 meg,
2444         * which seems reasonable as a default.
2445         */
2446        sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2447                                       sbi->s_cluster_bits, 32);
2448        /*
2449         * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2450         * to the lowest multiple of s_stripe which is bigger than
2451         * the s_mb_group_prealloc as determined above. We want
2452         * the preallocation size to be an exact multiple of the
2453         * RAID stripe size so that preallocations don't fragment
2454         * the stripes.
2455         */
2456        if (sbi->s_stripe > 1) {
2457                sbi->s_mb_group_prealloc = roundup(
2458                        sbi->s_mb_group_prealloc, sbi->s_stripe);
2459        }
2460
2461        sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2462        if (sbi->s_locality_groups == NULL) {
2463                ret = -ENOMEM;
2464                goto out_free_groupinfo_slab;
2465        }
2466        for_each_possible_cpu(i) {
2467                struct ext4_locality_group *lg;
2468                lg = per_cpu_ptr(sbi->s_locality_groups, i);
2469                mutex_init(&lg->lg_mutex);
2470                for (j = 0; j < PREALLOC_TB_SIZE; j++)
2471                        INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2472                spin_lock_init(&lg->lg_prealloc_lock);
2473        }
2474
2475        /* init file for buddy data */
2476        ret = ext4_mb_init_backend(sb);
2477        if (ret != 0)
2478                goto out_free_locality_groups;
2479
2480        if (sbi->s_proc)
2481                proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2482                                 &ext4_mb_seq_groups_fops, sb);
2483
2484        return 0;
2485
2486out_free_locality_groups:
2487        free_percpu(sbi->s_locality_groups);
2488        sbi->s_locality_groups = NULL;
2489out_free_groupinfo_slab:
2490        ext4_groupinfo_destroy_slabs();
2491out:
2492        kfree(sbi->s_mb_offsets);
2493        sbi->s_mb_offsets = NULL;
2494        kfree(sbi->s_mb_maxs);
2495        sbi->s_mb_maxs = NULL;
2496        return ret;
2497}
2498
2499/* need to called with the ext4 group lock held */
2500static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2501{
2502        struct ext4_prealloc_space *pa;
2503        struct list_head *cur, *tmp;
2504        int count = 0;
2505
2506        list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2507                pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2508                list_del(&pa->pa_group_list);
2509                count++;
2510                kmem_cache_free(ext4_pspace_cachep, pa);
2511        }
2512        if (count)
2513                mb_debug(1, "mballoc: %u PAs left\n", count);
2514
2515}
2516
2517int ext4_mb_release(struct super_block *sb)
2518{
2519        ext4_group_t ngroups = ext4_get_groups_count(sb);
2520        ext4_group_t i;
2521        int num_meta_group_infos;
2522        struct ext4_group_info *grinfo;
2523        struct ext4_sb_info *sbi = EXT4_SB(sb);
2524        struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2525
2526        if (sbi->s_proc)
2527                remove_proc_entry("mb_groups", sbi->s_proc);
2528
2529        if (sbi->s_group_info) {
2530                for (i = 0; i < ngroups; i++) {
2531                        grinfo = ext4_get_group_info(sb, i);
2532#ifdef DOUBLE_CHECK
2533                        kfree(grinfo->bb_bitmap);
2534#endif
2535                        ext4_lock_group(sb, i);
2536                        ext4_mb_cleanup_pa(grinfo);
2537                        ext4_unlock_group(sb, i);
2538                        kmem_cache_free(cachep, grinfo);
2539                }
2540                num_meta_group_infos = (ngroups +
2541                                EXT4_DESC_PER_BLOCK(sb) - 1) >>
2542                        EXT4_DESC_PER_BLOCK_BITS(sb);
2543                for (i = 0; i < num_meta_group_infos; i++)
2544                        kfree(sbi->s_group_info[i]);
2545                ext4_kvfree(sbi->s_group_info);
2546        }
2547        kfree(sbi->s_mb_offsets);
2548        kfree(sbi->s_mb_maxs);
2549        if (sbi->s_buddy_cache)
2550                iput(sbi->s_buddy_cache);
2551        if (sbi->s_mb_stats) {
2552                ext4_msg(sb, KERN_INFO,
2553                       "mballoc: %u blocks %u reqs (%u success)",
2554                                atomic_read(&sbi->s_bal_allocated),
2555                                atomic_read(&sbi->s_bal_reqs),
2556                                atomic_read(&sbi->s_bal_success));
2557                ext4_msg(sb, KERN_INFO,
2558                      "mballoc: %u extents scanned, %u goal hits, "
2559                                "%u 2^N hits, %u breaks, %u lost",
2560                                atomic_read(&sbi->s_bal_ex_scanned),
2561                                atomic_read(&sbi->s_bal_goals),
2562                                atomic_read(&sbi->s_bal_2orders),
2563                                atomic_read(&sbi->s_bal_breaks),
2564                                atomic_read(&sbi->s_mb_lost_chunks));
2565                ext4_msg(sb, KERN_INFO,
2566                       "mballoc: %lu generated and it took %Lu",
2567                                sbi->s_mb_buddies_generated,
2568                                sbi->s_mb_generation_time);
2569                ext4_msg(sb, KERN_INFO,
2570                       "mballoc: %u preallocated, %u discarded",
2571                                atomic_read(&sbi->s_mb_preallocated),
2572                                atomic_read(&sbi->s_mb_discarded));
2573        }
2574
2575        free_percpu(sbi->s_locality_groups);
2576
2577        return 0;
2578}
2579
2580static inline int ext4_issue_discard(struct super_block *sb,
2581                ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2582{
2583        ext4_fsblk_t discard_block;
2584
2585        discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2586                         ext4_group_first_block_no(sb, block_group));
2587        count = EXT4_C2B(EXT4_SB(sb), count);
2588        trace_ext4_discard_blocks(sb,
2589                        (unsigned long long) discard_block, count);
2590        return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2591}
2592
2593/*
2594 * This function is called by the jbd2 layer once the commit has finished,
2595 * so we know we can free the blocks that were released with that commit.
2596 */
2597static void ext4_free_data_callback(struct super_block *sb,
2598                                    struct ext4_journal_cb_entry *jce,
2599                                    int rc)
2600{
2601        struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2602        struct ext4_buddy e4b;
2603        struct ext4_group_info *db;
2604        int err, count = 0, count2 = 0;
2605
2606        mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2607                 entry->efd_count, entry->efd_group, entry);
2608
2609        if (test_opt(sb, DISCARD))
2610                ext4_issue_discard(sb, entry->efd_group,
2611                                   entry->efd_start_cluster, entry->efd_count);
2612
2613        err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2614        /* we expect to find existing buddy because it's pinned */
2615        BUG_ON(err != 0);
2616
2617
2618        db = e4b.bd_info;
2619        /* there are blocks to put in buddy to make them really free */
2620        count += entry->efd_count;
2621        count2++;
2622        ext4_lock_group(sb, entry->efd_group);
2623        /* Take it out of per group rb tree */
2624        rb_erase(&entry->efd_node, &(db->bb_free_root));
2625        mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2626
2627        /*
2628         * Clear the trimmed flag for the group so that the next
2629         * ext4_trim_fs can trim it.
2630         * If the volume is mounted with -o discard, online discard
2631         * is supported and the free blocks will be trimmed online.
2632         */
2633        if (!test_opt(sb, DISCARD))
2634                EXT4_MB_GRP_CLEAR_TRIMMED(db);
2635
2636        if (!db->bb_free_root.rb_node) {
2637                /* No more items in the per group rb tree
2638                 * balance refcounts from ext4_mb_free_metadata()
2639                 */
2640                page_cache_release(e4b.bd_buddy_page);
2641                page_cache_release(e4b.bd_bitmap_page);
2642        }
2643        ext4_unlock_group(sb, entry->efd_group);
2644        kmem_cache_free(ext4_free_data_cachep, entry);
2645        ext4_mb_unload_buddy(&e4b);
2646
2647        mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2648}
2649
2650#ifdef CONFIG_EXT4_DEBUG
2651u8 mb_enable_debug __read_mostly;
2652
2653static struct dentry *debugfs_dir;
2654static struct dentry *debugfs_debug;
2655
2656static void __init ext4_create_debugfs_entry(void)
2657{
2658        debugfs_dir = debugfs_create_dir("ext4", NULL);
2659        if (debugfs_dir)
2660                debugfs_debug = debugfs_create_u8("mballoc-debug",
2661                                                  S_IRUGO | S_IWUSR,
2662                                                  debugfs_dir,
2663                                                  &mb_enable_debug);
2664}
2665
2666static void ext4_remove_debugfs_entry(void)
2667{
2668        debugfs_remove(debugfs_debug);
2669        debugfs_remove(debugfs_dir);
2670}
2671
2672#else
2673
2674static void __init ext4_create_debugfs_entry(void)
2675{
2676}
2677
2678static void ext4_remove_debugfs_entry(void)
2679{
2680}
2681
2682#endif
2683
2684int __init ext4_init_mballoc(void)
2685{
2686        ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2687                                        SLAB_RECLAIM_ACCOUNT);
2688        if (ext4_pspace_cachep == NULL)
2689                return -ENOMEM;
2690
2691        ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2692                                    SLAB_RECLAIM_ACCOUNT);
2693        if (ext4_ac_cachep == NULL) {
2694                kmem_cache_destroy(ext4_pspace_cachep);
2695                return -ENOMEM;
2696        }
2697
2698        ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2699                                           SLAB_RECLAIM_ACCOUNT);
2700        if (ext4_free_data_cachep == NULL) {
2701                kmem_cache_destroy(ext4_pspace_cachep);
2702                kmem_cache_destroy(ext4_ac_cachep);
2703                return -ENOMEM;
2704        }
2705        ext4_create_debugfs_entry();
2706        return 0;
2707}
2708
2709void ext4_exit_mballoc(void)
2710{
2711        /*
2712         * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2713         * before destroying the slab cache.
2714         */
2715        rcu_barrier();
2716        kmem_cache_destroy(ext4_pspace_cachep);
2717        kmem_cache_destroy(ext4_ac_cachep);
2718        kmem_cache_destroy(ext4_free_data_cachep);
2719        ext4_groupinfo_destroy_slabs();
2720        ext4_remove_debugfs_entry();
2721}
2722
2723
2724/*
2725 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2726 * Returns 0 if success or error code
2727 */
2728static noinline_for_stack int
2729ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2730                                handle_t *handle, unsigned int reserv_clstrs)
2731{
2732        struct buffer_head *bitmap_bh = NULL;
2733        struct ext4_group_desc *gdp;
2734        struct buffer_head *gdp_bh;
2735        struct ext4_sb_info *sbi;
2736        struct super_block *sb;
2737        ext4_fsblk_t block;
2738        int err, len;
2739
2740        BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2741        BUG_ON(ac->ac_b_ex.fe_len <= 0);
2742
2743        sb = ac->ac_sb;
2744        sbi = EXT4_SB(sb);
2745
2746        err = -EIO;
2747        bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2748        if (!bitmap_bh)
2749                goto out_err;
2750
2751        err = ext4_journal_get_write_access(handle, bitmap_bh);
2752        if (err)
2753                goto out_err;
2754
2755        err = -EIO;
2756        gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2757        if (!gdp)
2758                goto out_err;
2759
2760        ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2761                        ext4_free_group_clusters(sb, gdp));
2762
2763        err = ext4_journal_get_write_access(handle, gdp_bh);
2764        if (err)
2765                goto out_err;
2766
2767        block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2768
2769        len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2770        if (!ext4_data_block_valid(sbi, block, len)) {
2771                ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2772                           "fs metadata", block, block+len);
2773                /* File system mounted not to panic on error
2774                 * Fix the bitmap and repeat the block allocation
2775                 * We leak some of the blocks here.
2776                 */
2777                ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2778                ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2779                              ac->ac_b_ex.fe_len);
2780                ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2781                err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2782                if (!err)
2783                        err = -EAGAIN;
2784                goto out_err;
2785        }
2786
2787        ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2788#ifdef AGGRESSIVE_CHECK
2789        {
2790                int i;
2791                for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2792                        BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2793                                                bitmap_bh->b_data));
2794                }
2795        }
2796#endif
2797        ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2798                      ac->ac_b_ex.fe_len);
2799        if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2800                gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2801                ext4_free_group_clusters_set(sb, gdp,
2802                                             ext4_free_clusters_after_init(sb,
2803                                                ac->ac_b_ex.fe_group, gdp));
2804        }
2805        len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2806        ext4_free_group_clusters_set(sb, gdp, len);
2807        ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh,
2808                                   EXT4_BLOCKS_PER_GROUP(sb) / 8);
2809        ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2810
2811        ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2812        percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2813        /*
2814         * Now reduce the dirty block count also. Should not go negative
2815         */
2816        if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2817                /* release all the reserved blocks if non delalloc */
2818                percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2819                                   reserv_clstrs);
2820
2821        if (sbi->s_log_groups_per_flex) {
2822                ext4_group_t flex_group = ext4_flex_group(sbi,
2823                                                          ac->ac_b_ex.fe_group);
2824                atomic_sub(ac->ac_b_ex.fe_len,
2825                           &sbi->s_flex_groups[flex_group].free_clusters);
2826        }
2827
2828        err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2829        if (err)
2830                goto out_err;
2831        err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2832
2833out_err:
2834        brelse(bitmap_bh);
2835        return err;
2836}
2837
2838/*
2839 * here we normalize request for locality group
2840 * Group request are normalized to s_mb_group_prealloc, which goes to
2841 * s_strip if we set the same via mount option.
2842 * s_mb_group_prealloc can be configured via
2843 * /sys/fs/ext4/<partition>/mb_group_prealloc
2844 *
2845 * XXX: should we try to preallocate more than the group has now?
2846 */
2847static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
2848{
2849        struct super_block *sb = ac->ac_sb;
2850        struct ext4_locality_group *lg = ac->ac_lg;
2851
2852        BUG_ON(lg == NULL);
2853        ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
2854        mb_debug(1, "#%u: goal %u blocks for locality group\n",
2855                current->pid, ac->ac_g_ex.fe_len);
2856}
2857
2858/*
2859 * Normalization means making request better in terms of
2860 * size and alignment
2861 */
2862static noinline_for_stack void
2863ext4_mb_normalize_request(struct ext4_allocation_context *ac,
2864                                struct ext4_allocation_request *ar)
2865{
2866        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
2867        int bsbits, max;
2868        ext4_lblk_t end;
2869        loff_t size, start_off;
2870        loff_t orig_size __maybe_unused;
2871        ext4_lblk_t start;
2872        struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
2873        struct ext4_prealloc_space *pa;
2874
2875        /* do normalize only data requests, metadata requests
2876           do not need preallocation */
2877        if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
2878                return;
2879
2880        /* sometime caller may want exact blocks */
2881        if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2882                return;
2883
2884        /* caller may indicate that preallocation isn't
2885         * required (it's a tail, for example) */
2886        if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
2887                return;
2888
2889        if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
2890                ext4_mb_normalize_group_request(ac);
2891                return ;
2892        }
2893
2894        bsbits = ac->ac_sb->s_blocksize_bits;
2895
2896        /* first, let's learn actual file size
2897         * given current request is allocated */
2898        size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
2899        size = size << bsbits;
2900        if (size < i_size_read(ac->ac_inode))
2901                size = i_size_read(ac->ac_inode);
2902        orig_size = size;
2903
2904        /* max size of free chunks */
2905        max = 2 << bsbits;
2906
2907#define NRL_CHECK_SIZE(req, size, max, chunk_size)      \
2908                (req <= (size) || max <= (chunk_size))
2909
2910        /* first, try to predict filesize */
2911        /* XXX: should this table be tunable? */
2912        start_off = 0;
2913        if (size <= 16 * 1024) {
2914                size = 16 * 1024;
2915        } else if (size <= 32 * 1024) {
2916                size = 32 * 1024;
2917        } else if (size <= 64 * 1024) {
2918                size = 64 * 1024;
2919        } else if (size <= 128 * 1024) {
2920                size = 128 * 1024;
2921        } else if (size <= 256 * 1024) {
2922                size = 256 * 1024;
2923        } else if (size <= 512 * 1024) {
2924                size = 512 * 1024;
2925        } else if (size <= 1024 * 1024) {
2926                size = 1024 * 1024;
2927        } else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
2928                start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2929                                                (21 - bsbits)) << 21;
2930                size = 2 * 1024 * 1024;
2931        } else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
2932                start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2933                                                        (22 - bsbits)) << 22;
2934                size = 4 * 1024 * 1024;
2935        } else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
2936                                        (8<<20)>>bsbits, max, 8 * 1024)) {
2937                start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
2938                                                        (23 - bsbits)) << 23;
2939                size = 8 * 1024 * 1024;
2940        } else {
2941                start_off = (loff_t)ac->ac_o_ex.fe_logical << bsbits;
2942                size      = ac->ac_o_ex.fe_len << bsbits;
2943        }
2944        size = size >> bsbits;
2945        start = start_off >> bsbits;
2946
2947        /* don't cover already allocated blocks in selected range */
2948        if (ar->pleft && start <= ar->lleft) {
2949                size -= ar->lleft + 1 - start;
2950                start = ar->lleft + 1;
2951        }
2952        if (ar->pright && start + size - 1 >= ar->lright)
2953                size -= start + size - ar->lright;
2954
2955        end = start + size;
2956
2957        /* check we don't cross already preallocated blocks */
2958        rcu_read_lock();
2959        list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
2960                ext4_lblk_t pa_end;
2961
2962                if (pa->pa_deleted)
2963                        continue;
2964                spin_lock(&pa->pa_lock);
2965                if (pa->pa_deleted) {
2966                        spin_unlock(&pa->pa_lock);
2967                        continue;
2968                }
2969
2970                pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
2971                                                  pa->pa_len);
2972
2973                /* PA must not overlap original request */
2974                BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
2975                        ac->ac_o_ex.fe_logical < pa->pa_lstart));
2976
2977                /* skip PAs this normalized request doesn't overlap with */
2978                if (pa->pa_lstart >= end || pa_end <= start) {
2979                        spin_unlock(&pa->pa_lock);
2980                        continue;
2981                }
2982                BUG_ON(pa->pa_lstart <= start && pa_end >= end);
2983
2984                /* adjust start or end to be adjacent to this pa */
2985                if (pa_end <= ac->ac_o_ex.fe_logical) {
2986                        BUG_ON(pa_end < start);
2987                        start = pa_end;
2988                } else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
2989                        BUG_ON(pa->pa_lstart > end);
2990                        end = pa->pa_lstart;
2991                }
2992                spin_unlock(&pa->pa_lock);
2993        }
2994        rcu_read_unlock();
2995        size = end - start;
2996
2997        /* XXX: extra loop to check we really don't overlap preallocations */
2998        rcu_read_lock();
2999        list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3000                ext4_lblk_t pa_end;
3001
3002                spin_lock(&pa->pa_lock);
3003                if (pa->pa_deleted == 0) {
3004                        pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3005                                                          pa->pa_len);
3006                        BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3007                }
3008                spin_unlock(&pa->pa_lock);
3009        }
3010        rcu_read_unlock();
3011
3012        if (start + size <= ac->ac_o_ex.fe_logical &&
3013                        start > ac->ac_o_ex.fe_logical) {
3014                ext4_msg(ac->ac_sb, KERN_ERR,
3015                         "start %lu, size %lu, fe_logical %lu",
3016                         (unsigned long) start, (unsigned long) size,
3017                         (unsigned long) ac->ac_o_ex.fe_logical);
3018        }
3019        BUG_ON(start + size <= ac->ac_o_ex.fe_logical &&
3020                        start > ac->ac_o_ex.fe_logical);
3021        BUG_ON(size <= 0 || size > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
3022
3023        /* now prepare goal request */
3024
3025        /* XXX: is it better to align blocks WRT to logical
3026         * placement or satisfy big request as is */
3027        ac->ac_g_ex.fe_logical = start;
3028        ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3029
3030        /* define goal start in order to merge */
3031        if (ar->pright && (ar->lright == (start + size))) {
3032                /* merge to the right */
3033                ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3034                                                &ac->ac_f_ex.fe_group,
3035                                                &ac->ac_f_ex.fe_start);
3036                ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3037        }
3038        if (ar->pleft && (ar->lleft + 1 == start)) {
3039                /* merge to the left */
3040                ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3041                                                &ac->ac_f_ex.fe_group,
3042                                                &ac->ac_f_ex.fe_start);
3043                ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3044        }
3045
3046        mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3047                (unsigned) orig_size, (unsigned) start);
3048}
3049
3050static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3051{
3052        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3053
3054        if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3055                atomic_inc(&sbi->s_bal_reqs);
3056                atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3057                if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3058                        atomic_inc(&sbi->s_bal_success);
3059                atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3060                if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3061                                ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3062                        atomic_inc(&sbi->s_bal_goals);
3063                if (ac->ac_found > sbi->s_mb_max_to_scan)
3064                        atomic_inc(&sbi->s_bal_breaks);
3065        }
3066
3067        if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3068                trace_ext4_mballoc_alloc(ac);
3069        else
3070                trace_ext4_mballoc_prealloc(ac);
3071}
3072
3073/*
3074 * Called on failure; free up any blocks from the inode PA for this
3075 * context.  We don't need this for MB_GROUP_PA because we only change
3076 * pa_free in ext4_mb_release_context(), but on failure, we've already
3077 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3078 */
3079static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3080{
3081        struct ext4_prealloc_space *pa = ac->ac_pa;
3082
3083        if (pa && pa->pa_type == MB_INODE_PA)
3084                pa->pa_free += ac->ac_b_ex.fe_len;
3085}
3086
3087/*
3088 * use blocks preallocated to inode
3089 */
3090static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3091                                struct ext4_prealloc_space *pa)
3092{
3093        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3094        ext4_fsblk_t start;
3095        ext4_fsblk_t end;
3096        int len;
3097
3098        /* found preallocated blocks, use them */
3099        start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3100        end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3101                  start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3102        len = EXT4_NUM_B2C(sbi, end - start);
3103        ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3104                                        &ac->ac_b_ex.fe_start);
3105        ac->ac_b_ex.fe_len = len;
3106        ac->ac_status = AC_STATUS_FOUND;
3107        ac->ac_pa = pa;
3108
3109        BUG_ON(start < pa->pa_pstart);
3110        BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3111        BUG_ON(pa->pa_free < len);
3112        pa->pa_free -= len;
3113
3114        mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3115}
3116
3117/*
3118 * use blocks preallocated to locality group
3119 */
3120static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3121                                struct ext4_prealloc_space *pa)
3122{
3123        unsigned int len = ac->ac_o_ex.fe_len;
3124
3125        ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3126                                        &ac->ac_b_ex.fe_group,
3127                                        &ac->ac_b_ex.fe_start);
3128        ac->ac_b_ex.fe_len = len;
3129        ac->ac_status = AC_STATUS_FOUND;
3130        ac->ac_pa = pa;
3131
3132        /* we don't correct pa_pstart or pa_plen here to avoid
3133         * possible race when the group is being loaded concurrently
3134         * instead we correct pa later, after blocks are marked
3135         * in on-disk bitmap -- see ext4_mb_release_context()
3136         * Other CPUs are prevented from allocating from this pa by lg_mutex
3137         */
3138        mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3139}
3140
3141/*
3142 * Return the prealloc space that have minimal distance
3143 * from the goal block. @cpa is the prealloc
3144 * space that is having currently known minimal distance
3145 * from the goal block.
3146 */
3147static struct ext4_prealloc_space *
3148ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3149                        struct ext4_prealloc_space *pa,
3150                        struct ext4_prealloc_space *cpa)
3151{
3152        ext4_fsblk_t cur_distance, new_distance;
3153
3154        if (cpa == NULL) {
3155                atomic_inc(&pa->pa_count);
3156                return pa;
3157        }
3158        cur_distance = abs(goal_block - cpa->pa_pstart);
3159        new_distance = abs(goal_block - pa->pa_pstart);
3160
3161        if (cur_distance <= new_distance)
3162                return cpa;
3163
3164        /* drop the previous reference */
3165        atomic_dec(&cpa->pa_count);
3166        atomic_inc(&pa->pa_count);
3167        return pa;
3168}
3169
3170/*
3171 * search goal blocks in preallocated space
3172 */
3173static noinline_for_stack int
3174ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3175{
3176        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3177        int order, i;
3178        struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3179        struct ext4_locality_group *lg;
3180        struct ext4_prealloc_space *pa, *cpa = NULL;
3181        ext4_fsblk_t goal_block;
3182
3183        /* only data can be preallocated */
3184        if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3185                return 0;
3186
3187        /* first, try per-file preallocation */
3188        rcu_read_lock();
3189        list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3190
3191                /* all fields in this condition don't change,
3192                 * so we can skip locking for them */
3193                if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3194                    ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3195                                               EXT4_C2B(sbi, pa->pa_len)))
3196                        continue;
3197
3198                /* non-extent files can't have physical blocks past 2^32 */
3199                if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3200                    (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3201                     EXT4_MAX_BLOCK_FILE_PHYS))
3202                        continue;
3203
3204                /* found preallocated blocks, use them */
3205                spin_lock(&pa->pa_lock);
3206                if (pa->pa_deleted == 0 && pa->pa_free) {
3207                        atomic_inc(&pa->pa_count);
3208                        ext4_mb_use_inode_pa(ac, pa);
3209                        spin_unlock(&pa->pa_lock);
3210                        ac->ac_criteria = 10;
3211                        rcu_read_unlock();
3212                        return 1;
3213                }
3214                spin_unlock(&pa->pa_lock);
3215        }
3216        rcu_read_unlock();
3217
3218        /* can we use group allocation? */
3219        if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3220                return 0;
3221
3222        /* inode may have no locality group for some reason */
3223        lg = ac->ac_lg;
3224        if (lg == NULL)
3225                return 0;
3226        order  = fls(ac->ac_o_ex.fe_len) - 1;
3227        if (order > PREALLOC_TB_SIZE - 1)
3228                /* The max size of hash table is PREALLOC_TB_SIZE */
3229                order = PREALLOC_TB_SIZE - 1;
3230
3231        goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3232        /*
3233         * search for the prealloc space that is having
3234         * minimal distance from the goal block.
3235         */
3236        for (i = order; i < PREALLOC_TB_SIZE; i++) {
3237                rcu_read_lock();
3238                list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3239                                        pa_inode_list) {
3240                        spin_lock(&pa->pa_lock);
3241                        if (pa->pa_deleted == 0 &&
3242                                        pa->pa_free >= ac->ac_o_ex.fe_len) {
3243
3244                                cpa = ext4_mb_check_group_pa(goal_block,
3245                                                                pa, cpa);
3246                        }
3247                        spin_unlock(&pa->pa_lock);
3248                }
3249                rcu_read_unlock();
3250        }
3251        if (cpa) {
3252                ext4_mb_use_group_pa(ac, cpa);
3253                ac->ac_criteria = 20;
3254                return 1;
3255        }
3256        return 0;
3257}
3258
3259/*
3260 * the function goes through all block freed in the group
3261 * but not yet committed and marks them used in in-core bitmap.
3262 * buddy must be generated from this bitmap
3263 * Need to be called with the ext4 group lock held
3264 */
3265static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3266                                                ext4_group_t group)
3267{
3268        struct rb_node *n;
3269        struct ext4_group_info *grp;
3270        struct ext4_free_data *entry;
3271
3272        grp = ext4_get_group_info(sb, group);
3273        n = rb_first(&(grp->bb_free_root));
3274
3275        while (n) {
3276                entry = rb_entry(n, struct ext4_free_data, efd_node);
3277                ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3278                n = rb_next(n);
3279        }
3280        return;
3281}
3282
3283/*
3284 * the function goes through all preallocation in this group and marks them
3285 * used in in-core bitmap. buddy must be generated from this bitmap
3286 * Need to be called with ext4 group lock held
3287 */
3288static noinline_for_stack
3289void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3290                                        ext4_group_t group)
3291{
3292        struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3293        struct ext4_prealloc_space *pa;
3294        struct list_head *cur;
3295        ext4_group_t groupnr;
3296        ext4_grpblk_t start;
3297        int preallocated = 0;
3298        int len;
3299
3300        /* all form of preallocation discards first load group,
3301         * so the only competing code is preallocation use.
3302         * we don't need any locking here
3303         * notice we do NOT ignore preallocations with pa_deleted
3304         * otherwise we could leave used blocks available for
3305         * allocation in buddy when concurrent ext4_mb_put_pa()
3306         * is dropping preallocation
3307         */
3308        list_for_each(cur, &grp->bb_prealloc_list) {
3309                pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3310                spin_lock(&pa->pa_lock);
3311                ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3312                                             &groupnr, &start);
3313                len = pa->pa_len;
3314                spin_unlock(&pa->pa_lock);
3315                if (unlikely(len == 0))
3316                        continue;
3317                BUG_ON(groupnr != group);
3318                ext4_set_bits(bitmap, start, len);
3319                preallocated += len;
3320        }
3321        mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3322}
3323
3324static void ext4_mb_pa_callback(struct rcu_head *head)
3325{
3326        struct ext4_prealloc_space *pa;
3327        pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3328        kmem_cache_free(ext4_pspace_cachep, pa);
3329}
3330
3331/*
3332 * drops a reference to preallocated space descriptor
3333 * if this was the last reference and the space is consumed
3334 */
3335static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3336                        struct super_block *sb, struct ext4_prealloc_space *pa)
3337{
3338        ext4_group_t grp;
3339        ext4_fsblk_t grp_blk;
3340
3341        if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0)
3342                return;
3343
3344        /* in this short window concurrent discard can set pa_deleted */
3345        spin_lock(&pa->pa_lock);
3346        if (pa->pa_deleted == 1) {
3347                spin_unlock(&pa->pa_lock);
3348                return;
3349        }
3350
3351        pa->pa_deleted = 1;
3352        spin_unlock(&pa->pa_lock);
3353
3354        grp_blk = pa->pa_pstart;
3355        /*
3356         * If doing group-based preallocation, pa_pstart may be in the
3357         * next group when pa is used up
3358         */
3359        if (pa->pa_type == MB_GROUP_PA)
3360                grp_blk--;
3361
3362        ext4_get_group_no_and_offset(sb, grp_blk, &grp, NULL);
3363
3364        /*
3365         * possible race:
3366         *
3367         *  P1 (buddy init)                     P2 (regular allocation)
3368         *                                      find block B in PA
3369         *  copy on-disk bitmap to buddy
3370         *                                      mark B in on-disk bitmap
3371         *                                      drop PA from group
3372         *  mark all PAs in buddy
3373         *
3374         * thus, P1 initializes buddy with B available. to prevent this
3375         * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3376         * against that pair
3377         */
3378        ext4_lock_group(sb, grp);
3379        list_del(&pa->pa_group_list);
3380        ext4_unlock_group(sb, grp);
3381
3382        spin_lock(pa->pa_obj_lock);
3383        list_del_rcu(&pa->pa_inode_list);
3384        spin_unlock(pa->pa_obj_lock);
3385
3386        call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3387}
3388
3389/*
3390 * creates new preallocated space for given inode
3391 */
3392static noinline_for_stack int
3393ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3394{
3395        struct super_block *sb = ac->ac_sb;
3396        struct ext4_sb_info *sbi = EXT4_SB(sb);
3397        struct ext4_prealloc_space *pa;
3398        struct ext4_group_info *grp;
3399        struct ext4_inode_info *ei;
3400
3401        /* preallocate only when found space is larger then requested */
3402        BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3403        BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3404        BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3405
3406        pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3407        if (pa == NULL)
3408                return -ENOMEM;
3409
3410        if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3411                int winl;
3412                int wins;
3413                int win;
3414                int offs;
3415
3416                /* we can't allocate as much as normalizer wants.
3417                 * so, found space must get proper lstart
3418                 * to cover original request */
3419                BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3420                BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3421
3422                /* we're limited by original request in that
3423                 * logical block must be covered any way
3424                 * winl is window we can move our chunk within */
3425                winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3426
3427                /* also, we should cover whole original request */
3428                wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3429
3430                /* the smallest one defines real window */
3431                win = min(winl, wins);
3432
3433                offs = ac->ac_o_ex.fe_logical %
3434                        EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3435                if (offs && offs < win)
3436                        win = offs;
3437
3438                ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3439                        EXT4_B2C(sbi, win);
3440                BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3441                BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3442        }
3443
3444        /* preallocation can change ac_b_ex, thus we store actually
3445         * allocated blocks for history */
3446        ac->ac_f_ex = ac->ac_b_ex;
3447
3448        pa->pa_lstart = ac->ac_b_ex.fe_logical;
3449        pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3450        pa->pa_len = ac->ac_b_ex.fe_len;
3451        pa->pa_free = pa->pa_len;
3452        atomic_set(&pa->pa_count, 1);
3453        spin_lock_init(&pa->pa_lock);
3454        INIT_LIST_HEAD(&pa->pa_inode_list);
3455        INIT_LIST_HEAD(&pa->pa_group_list);
3456        pa->pa_deleted = 0;
3457        pa->pa_type = MB_INODE_PA;
3458
3459        mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3460                        pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3461        trace_ext4_mb_new_inode_pa(ac, pa);
3462
3463        ext4_mb_use_inode_pa(ac, pa);
3464        atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3465
3466        ei = EXT4_I(ac->ac_inode);
3467        grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3468
3469        pa->pa_obj_lock = &ei->i_prealloc_lock;
3470        pa->pa_inode = ac->ac_inode;
3471
3472        ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3473        list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3474        ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3475
3476        spin_lock(pa->pa_obj_lock);
3477        list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3478        spin_unlock(pa->pa_obj_lock);
3479
3480        return 0;
3481}
3482
3483/*
3484 * creates new preallocated space for locality group inodes belongs to
3485 */
3486static noinline_for_stack int
3487ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3488{
3489        struct super_block *sb = ac->ac_sb;
3490        struct ext4_locality_group *lg;
3491        struct ext4_prealloc_space *pa;
3492        struct ext4_group_info *grp;
3493
3494        /* preallocate only when found space is larger then requested */
3495        BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3496        BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3497        BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3498
3499        BUG_ON(ext4_pspace_cachep == NULL);
3500        pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3501        if (pa == NULL)
3502                return -ENOMEM;
3503
3504        /* preallocation can change ac_b_ex, thus we store actually
3505         * allocated blocks for history */
3506        ac->ac_f_ex = ac->ac_b_ex;
3507
3508        pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3509        pa->pa_lstart = pa->pa_pstart;
3510        pa->pa_len = ac->ac_b_ex.fe_len;
3511        pa->pa_free = pa->pa_len;
3512        atomic_set(&pa->pa_count, 1);
3513        spin_lock_init(&pa->pa_lock);
3514        INIT_LIST_HEAD(&pa->pa_inode_list);
3515        INIT_LIST_HEAD(&pa->pa_group_list);
3516        pa->pa_deleted = 0;
3517        pa->pa_type = MB_GROUP_PA;
3518
3519        mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3520                        pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3521        trace_ext4_mb_new_group_pa(ac, pa);
3522
3523        ext4_mb_use_group_pa(ac, pa);
3524        atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3525
3526        grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3527        lg = ac->ac_lg;
3528        BUG_ON(lg == NULL);
3529
3530        pa->pa_obj_lock = &lg->lg_prealloc_lock;
3531        pa->pa_inode = NULL;
3532
3533        ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3534        list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3535        ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3536
3537        /*
3538         * We will later add the new pa to the right bucket
3539         * after updating the pa_free in ext4_mb_release_context
3540         */
3541        return 0;
3542}
3543
3544static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3545{
3546        int err;
3547
3548        if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3549                err = ext4_mb_new_group_pa(ac);
3550        else
3551                err = ext4_mb_new_inode_pa(ac);
3552        return err;
3553}
3554
3555/*
3556 * finds all unused blocks in on-disk bitmap, frees them in
3557 * in-core bitmap and buddy.
3558 * @pa must be unlinked from inode and group lists, so that
3559 * nobody else can find/use it.
3560 * the caller MUST hold group/inode locks.
3561 * TODO: optimize the case when there are no in-core structures yet
3562 */
3563static noinline_for_stack int
3564ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3565                        struct ext4_prealloc_space *pa)
3566{
3567        struct super_block *sb = e4b->bd_sb;
3568        struct ext4_sb_info *sbi = EXT4_SB(sb);
3569        unsigned int end;
3570        unsigned int next;
3571        ext4_group_t group;
3572        ext4_grpblk_t bit;
3573        unsigned long long grp_blk_start;
3574        int err = 0;
3575        int free = 0;
3576
3577        BUG_ON(pa->pa_deleted == 0);
3578        ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3579        grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3580        BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3581        end = bit + pa->pa_len;
3582
3583        while (bit < end) {
3584                bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3585                if (bit >= end)
3586                        break;
3587                next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3588                mb_debug(1, "    free preallocated %u/%u in group %u\n",
3589                         (unsigned) ext4_group_first_block_no(sb, group) + bit,
3590                         (unsigned) next - bit, (unsigned) group);
3591                free += next - bit;
3592
3593                trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3594                trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3595                                                    EXT4_C2B(sbi, bit)),
3596                                               next - bit);
3597                mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3598                bit = next + 1;
3599        }
3600        if (free != pa->pa_free) {
3601                ext4_msg(e4b->bd_sb, KERN_CRIT,
3602                         "pa %p: logic %lu, phys. %lu, len %lu",
3603                         pa, (unsigned long) pa->pa_lstart,
3604                         (unsigned long) pa->pa_pstart,
3605                         (unsigned long) pa->pa_len);
3606                ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3607                                        free, pa->pa_free);
3608                /*
3609                 * pa is already deleted so we use the value obtained
3610                 * from the bitmap and continue.
3611                 */
3612        }
3613        atomic_add(free, &sbi->s_mb_discarded);
3614
3615        return err;
3616}
3617
3618static noinline_for_stack int
3619ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3620                                struct ext4_prealloc_space *pa)
3621{
3622        struct super_block *sb = e4b->bd_sb;
3623        ext4_group_t group;
3624        ext4_grpblk_t bit;
3625
3626        trace_ext4_mb_release_group_pa(sb, pa);
3627        BUG_ON(pa->pa_deleted == 0);
3628        ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3629        BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3630        mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3631        atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3632        trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3633
3634        return 0;
3635}
3636
3637/*
3638 * releases all preallocations in given group
3639 *
3640 * first, we need to decide discard policy:
3641 * - when do we discard
3642 *   1) ENOSPC
3643 * - how many do we discard
3644 *   1) how many requested
3645 */
3646static noinline_for_stack int
3647ext4_mb_discard_group_preallocations(struct super_block *sb,
3648                                        ext4_group_t group, int needed)
3649{
3650        struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3651        struct buffer_head *bitmap_bh = NULL;
3652        struct ext4_prealloc_space *pa, *tmp;
3653        struct list_head list;
3654        struct ext4_buddy e4b;
3655        int err;
3656        int busy = 0;
3657        int free = 0;
3658
3659        mb_debug(1, "discard preallocation for group %u\n", group);
3660
3661        if (list_empty(&grp->bb_prealloc_list))
3662                return 0;
3663
3664        bitmap_bh = ext4_read_block_bitmap(sb, group);
3665        if (bitmap_bh == NULL) {
3666                ext4_error(sb, "Error reading block bitmap for %u", group);
3667                return 0;
3668        }
3669
3670        err = ext4_mb_load_buddy(sb, group, &e4b);
3671        if (err) {
3672                ext4_error(sb, "Error loading buddy information for %u", group);
3673                put_bh(bitmap_bh);
3674                return 0;
3675        }
3676
3677        if (needed == 0)
3678                needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3679
3680        INIT_LIST_HEAD(&list);
3681repeat:
3682        ext4_lock_group(sb, group);
3683        list_for_each_entry_safe(pa, tmp,
3684                                &grp->bb_prealloc_list, pa_group_list) {
3685                spin_lock(&pa->pa_lock);
3686                if (atomic_read(&pa->pa_count)) {
3687                        spin_unlock(&pa->pa_lock);
3688                        busy = 1;
3689                        continue;
3690                }
3691                if (pa->pa_deleted) {
3692                        spin_unlock(&pa->pa_lock);
3693                        continue;
3694                }
3695
3696                /* seems this one can be freed ... */
3697                pa->pa_deleted = 1;
3698
3699                /* we can trust pa_free ... */
3700                free += pa->pa_free;
3701
3702                spin_unlock(&pa->pa_lock);
3703
3704                list_del(&pa->pa_group_list);
3705                list_add(&pa->u.pa_tmp_list, &list);
3706        }
3707
3708        /* if we still need more blocks and some PAs were used, try again */
3709        if (free < needed && busy) {
3710                busy = 0;
3711                ext4_unlock_group(sb, group);
3712                /*
3713                 * Yield the CPU here so that we don't get soft lockup
3714                 * in non preempt case.
3715                 */
3716                yield();
3717                goto repeat;
3718        }
3719
3720        /* found anything to free? */
3721        if (list_empty(&list)) {
3722                BUG_ON(free != 0);
3723                goto out;
3724        }
3725
3726        /* now free all selected PAs */
3727        list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3728
3729                /* remove from object (inode or locality group) */
3730                spin_lock(pa->pa_obj_lock);
3731                list_del_rcu(&pa->pa_inode_list);
3732                spin_unlock(pa->pa_obj_lock);
3733
3734                if (pa->pa_type == MB_GROUP_PA)
3735                        ext4_mb_release_group_pa(&e4b, pa);
3736                else
3737                        ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3738
3739                list_del(&pa->u.pa_tmp_list);
3740                call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3741        }
3742
3743out:
3744        ext4_unlock_group(sb, group);
3745        ext4_mb_unload_buddy(&e4b);
3746        put_bh(bitmap_bh);
3747        return free;
3748}
3749
3750/*
3751 * releases all non-used preallocated blocks for given inode
3752 *
3753 * It's important to discard preallocations under i_data_sem
3754 * We don't want another block to be served from the prealloc
3755 * space when we are discarding the inode prealloc space.
3756 *
3757 * FIXME!! Make sure it is valid at all the call sites
3758 */
3759void ext4_discard_preallocations(struct inode *inode)
3760{
3761        struct ext4_inode_info *ei = EXT4_I(inode);
3762        struct super_block *sb = inode->i_sb;
3763        struct buffer_head *bitmap_bh = NULL;
3764        struct ext4_prealloc_space *pa, *tmp;
3765        ext4_group_t group = 0;
3766        struct list_head list;
3767        struct ext4_buddy e4b;
3768        int err;
3769
3770        if (!S_ISREG(inode->i_mode)) {
3771                /*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3772                return;
3773        }
3774
3775        mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3776        trace_ext4_discard_preallocations(inode);
3777
3778        INIT_LIST_HEAD(&list);
3779
3780repeat:
3781        /* first, collect all pa's in the inode */
3782        spin_lock(&ei->i_prealloc_lock);
3783        while (!list_empty(&ei->i_prealloc_list)) {
3784                pa = list_entry(ei->i_prealloc_list.next,
3785                                struct ext4_prealloc_space, pa_inode_list);
3786                BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3787                spin_lock(&pa->pa_lock);
3788                if (atomic_read(&pa->pa_count)) {
3789                        /* this shouldn't happen often - nobody should
3790                         * use preallocation while we're discarding it */
3791                        spin_unlock(&pa->pa_lock);
3792                        spin_unlock(&ei->i_prealloc_lock);
3793                        ext4_msg(sb, KERN_ERR,
3794                                 "uh-oh! used pa while discarding");
3795                        WARN_ON(1);
3796                        schedule_timeout_uninterruptible(HZ);
3797                        goto repeat;
3798
3799                }
3800                if (pa->pa_deleted == 0) {
3801                        pa->pa_deleted = 1;
3802                        spin_unlock(&pa->pa_lock);
3803                        list_del_rcu(&pa->pa_inode_list);
3804                        list_add(&pa->u.pa_tmp_list, &list);
3805                        continue;
3806                }
3807
3808                /* someone is deleting pa right now */
3809                spin_unlock(&pa->pa_lock);
3810                spin_unlock(&ei->i_prealloc_lock);
3811
3812                /* we have to wait here because pa_deleted
3813                 * doesn't mean pa is already unlinked from
3814                 * the list. as we might be called from
3815                 * ->clear_inode() the inode will get freed
3816                 * and concurrent thread which is unlinking
3817                 * pa from inode's list may access already
3818                 * freed memory, bad-bad-bad */
3819
3820                /* XXX: if this happens too often, we can
3821                 * add a flag to force wait only in case
3822                 * of ->clear_inode(), but not in case of
3823                 * regular truncate */
3824                schedule_timeout_uninterruptible(HZ);
3825                goto repeat;
3826        }
3827        spin_unlock(&ei->i_prealloc_lock);
3828
3829        list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3830                BUG_ON(pa->pa_type != MB_INODE_PA);
3831                ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
3832
3833                err = ext4_mb_load_buddy(sb, group, &e4b);
3834                if (err) {
3835                        ext4_error(sb, "Error loading buddy information for %u",
3836                                        group);
3837                        continue;
3838                }
3839
3840                bitmap_bh = ext4_read_block_bitmap(sb, group);
3841                if (bitmap_bh == NULL) {
3842                        ext4_error(sb, "Error reading block bitmap for %u",
3843                                        group);
3844                        ext4_mb_unload_buddy(&e4b);
3845                        continue;
3846                }
3847
3848                ext4_lock_group(sb, group);
3849                list_del(&pa->pa_group_list);
3850                ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3851                ext4_unlock_group(sb, group);
3852
3853                ext4_mb_unload_buddy(&e4b);
3854                put_bh(bitmap_bh);
3855
3856                list_del(&pa->u.pa_tmp_list);
3857                call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3858        }
3859}
3860
3861#ifdef CONFIG_EXT4_DEBUG
3862static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3863{
3864        struct super_block *sb = ac->ac_sb;
3865        ext4_group_t ngroups, i;
3866
3867        if (!mb_enable_debug ||
3868            (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
3869                return;
3870
3871        ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
3872                        " Allocation context details:");
3873        ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
3874                        ac->ac_status, ac->ac_flags);
3875        ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
3876                        "goal %lu/%lu/%lu@%lu, "
3877                        "best %lu/%lu/%lu@%lu cr %d",
3878                        (unsigned long)ac->ac_o_ex.fe_group,
3879                        (unsigned long)ac->ac_o_ex.fe_start,
3880                        (unsigned long)ac->ac_o_ex.fe_len,
3881                        (unsigned long)ac->ac_o_ex.fe_logical,
3882                        (unsigned long)ac->ac_g_ex.fe_group,
3883                        (unsigned long)ac->ac_g_ex.fe_start,
3884                        (unsigned long)ac->ac_g_ex.fe_len,
3885                        (unsigned long)ac->ac_g_ex.fe_logical,
3886                        (unsigned long)ac->ac_b_ex.fe_group,
3887                        (unsigned long)ac->ac_b_ex.fe_start,
3888                        (unsigned long)ac->ac_b_ex.fe_len,
3889                        (unsigned long)ac->ac_b_ex.fe_logical,
3890                        (int)ac->ac_criteria);
3891        ext4_msg(ac->ac_sb, KERN_ERR, "%lu scanned, %d found",
3892                 ac->ac_ex_scanned, ac->ac_found);
3893        ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
3894        ngroups = ext4_get_groups_count(sb);
3895        for (i = 0; i < ngroups; i++) {
3896                struct ext4_group_info *grp = ext4_get_group_info(sb, i);
3897                struct ext4_prealloc_space *pa;
3898                ext4_grpblk_t start;
3899                struct list_head *cur;
3900                ext4_lock_group(sb, i);
3901                list_for_each(cur, &grp->bb_prealloc_list) {
3902                        pa = list_entry(cur, struct ext4_prealloc_space,
3903                                        pa_group_list);
3904                        spin_lock(&pa->pa_lock);
3905                        ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3906                                                     NULL, &start);
3907                        spin_unlock(&pa->pa_lock);
3908                        printk(KERN_ERR "PA:%u:%d:%u \n", i,
3909                               start, pa->pa_len);
3910                }
3911                ext4_unlock_group(sb, i);
3912
3913                if (grp->bb_free == 0)
3914                        continue;
3915                printk(KERN_ERR "%u: %d/%d \n",
3916                       i, grp->bb_free, grp->bb_fragments);
3917        }
3918        printk(KERN_ERR "\n");
3919}
3920#else
3921static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
3922{
3923        return;
3924}
3925#endif
3926
3927/*
3928 * We use locality group preallocation for small size file. The size of the
3929 * file is determined by the current size or the resulting size after
3930 * allocation which ever is larger
3931 *
3932 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
3933 */
3934static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
3935{
3936        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3937        int bsbits = ac->ac_sb->s_blocksize_bits;
3938        loff_t size, isize;
3939
3940        if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3941                return;
3942
3943        if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3944                return;
3945
3946        size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3947        isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
3948                >> bsbits;
3949
3950        if ((size == isize) &&
3951            !ext4_fs_is_busy(sbi) &&
3952            (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
3953                ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
3954                return;
3955        }
3956
3957        if (sbi->s_mb_group_prealloc <= 0) {
3958                ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3959                return;
3960        }
3961
3962        /* don't use group allocation for large files */
3963        size = max(size, isize);
3964        if (size > sbi->s_mb_stream_request) {
3965                ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
3966                return;
3967        }
3968
3969        BUG_ON(ac->ac_lg != NULL);
3970        /*
3971         * locality group prealloc space are per cpu. The reason for having
3972         * per cpu locality group is to reduce the contention between block
3973         * request from multiple CPUs.
3974         */
3975        ac->ac_lg = __this_cpu_ptr(sbi->s_locality_groups);
3976
3977        /* we're going to use group allocation */
3978        ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
3979
3980        /* serialize all allocations in the group */
3981        mutex_lock(&ac->ac_lg->lg_mutex);
3982}
3983
3984static noinline_for_stack int
3985ext4_mb_initialize_context(struct ext4_allocation_context *ac,
3986                                struct ext4_allocation_request *ar)
3987{
3988        struct super_block *sb = ar->inode->i_sb;
3989        struct ext4_sb_info *sbi = EXT4_SB(sb);
3990        struct ext4_super_block *es = sbi->s_es;
3991        ext4_group_t group;
3992        unsigned int len;
3993        ext4_fsblk_t goal;
3994        ext4_grpblk_t block;
3995
3996        /* we can't allocate > group size */
3997        len = ar->len;
3998
3999        /* just a dirty hack to filter too big requests  */
4000        if (len >= EXT4_CLUSTERS_PER_GROUP(sb) - 10)
4001                len = EXT4_CLUSTERS_PER_GROUP(sb) - 10;
4002
4003        /* start searching from the goal */
4004        goal = ar->goal;
4005        if (goal < le32_to_cpu(es->s_first_data_block) ||
4006                        goal >= ext4_blocks_count(es))
4007                goal = le32_to_cpu(es->s_first_data_block);
4008        ext4_get_group_no_and_offset(sb, goal, &group, &block);
4009
4010        /* set up allocation goals */
4011        memset(ac, 0, sizeof(struct ext4_allocation_context));
4012        ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4013        ac->ac_status = AC_STATUS_CONTINUE;
4014        ac->ac_sb = sb;
4015        ac->ac_inode = ar->inode;
4016        ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4017        ac->ac_o_ex.fe_group = group;
4018        ac->ac_o_ex.fe_start = block;
4019        ac->ac_o_ex.fe_len = len;
4020        ac->ac_g_ex = ac->ac_o_ex;
4021        ac->ac_flags = ar->flags;
4022
4023        /* we have to define context: we'll we work with a file or
4024         * locality group. this is a policy, actually */
4025        ext4_mb_group_or_file(ac);
4026
4027        mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4028                        "left: %u/%u, right %u/%u to %swritable\n",
4029                        (unsigned) ar->len, (unsigned) ar->logical,
4030                        (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4031                        (unsigned) ar->lleft, (unsigned) ar->pleft,
4032                        (unsigned) ar->lright, (unsigned) ar->pright,
4033                        atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4034        return 0;
4035
4036}
4037
4038static noinline_for_stack void
4039ext4_mb_discard_lg_preallocations(struct super_block *sb,
4040                                        struct ext4_locality_group *lg,
4041                                        int order, int total_entries)
4042{
4043        ext4_group_t group = 0;
4044        struct ext4_buddy e4b;
4045        struct list_head discard_list;
4046        struct ext4_prealloc_space *pa, *tmp;
4047
4048        mb_debug(1, "discard locality group preallocation\n");
4049
4050        INIT_LIST_HEAD(&discard_list);
4051
4052        spin_lock(&lg->lg_prealloc_lock);
4053        list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4054                                                pa_inode_list) {
4055                spin_lock(&pa->pa_lock);
4056                if (atomic_read(&pa->pa_count)) {
4057                        /*
4058                         * This is the pa that we just used
4059                         * for block allocation. So don't
4060                         * free that
4061                         */
4062                        spin_unlock(&pa->pa_lock);
4063                        continue;
4064                }
4065                if (pa->pa_deleted) {
4066                        spin_unlock(&pa->pa_lock);
4067                        continue;
4068                }
4069                /* only lg prealloc space */
4070                BUG_ON(pa->pa_type != MB_GROUP_PA);
4071
4072                /* seems this one can be freed ... */
4073                pa->pa_deleted = 1;
4074                spin_unlock(&pa->pa_lock);
4075
4076                list_del_rcu(&pa->pa_inode_list);
4077                list_add(&pa->u.pa_tmp_list, &discard_list);
4078
4079                total_entries--;
4080                if (total_entries <= 5) {
4081                        /*
4082                         * we want to keep only 5 entries
4083                         * allowing it to grow to 8. This
4084                         * mak sure we don't call discard
4085                         * soon for this list.
4086                         */
4087                        break;
4088                }
4089        }
4090        spin_unlock(&lg->lg_prealloc_lock);
4091
4092        list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4093
4094                ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4095                if (ext4_mb_load_buddy(sb, group, &e4b)) {
4096                        ext4_error(sb, "Error loading buddy information for %u",
4097                                        group);
4098                        continue;
4099                }
4100                ext4_lock_group(sb, group);
4101                list_del(&pa->pa_group_list);
4102                ext4_mb_release_group_pa(&e4b, pa);
4103                ext4_unlock_group(sb, group);
4104
4105                ext4_mb_unload_buddy(&e4b);
4106                list_del(&pa->u.pa_tmp_list);
4107                call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4108        }
4109}
4110
4111/*
4112 * We have incremented pa_count. So it cannot be freed at this
4113 * point. Also we hold lg_mutex. So no parallel allocation is
4114 * possible from this lg. That means pa_free cannot be updated.
4115 *
4116 * A parallel ext4_mb_discard_group_preallocations is possible.
4117 * which can cause the lg_prealloc_list to be updated.
4118 */
4119
4120static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4121{
4122        int order, added = 0, lg_prealloc_count = 1;
4123        struct super_block *sb = ac->ac_sb;
4124        struct ext4_locality_group *lg = ac->ac_lg;
4125        struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4126
4127        order = fls(pa->pa_free) - 1;
4128        if (order > PREALLOC_TB_SIZE - 1)
4129                /* The max size of hash table is PREALLOC_TB_SIZE */
4130                order = PREALLOC_TB_SIZE - 1;
4131        /* Add the prealloc space to lg */
4132        rcu_read_lock();
4133        list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4134                                                pa_inode_list) {
4135                spin_lock(&tmp_pa->pa_lock);
4136                if (tmp_pa->pa_deleted) {
4137                        spin_unlock(&tmp_pa->pa_lock);
4138                        continue;
4139                }
4140                if (!added && pa->pa_free < tmp_pa->pa_free) {
4141                        /* Add to the tail of the previous entry */
4142                        list_add_tail_rcu(&pa->pa_inode_list,
4143                                                &tmp_pa->pa_inode_list);
4144                        added = 1;
4145                        /*
4146                         * we want to count the total
4147                         * number of entries in the list
4148                         */
4149                }
4150                spin_unlock(&tmp_pa->pa_lock);
4151                lg_prealloc_count++;
4152        }
4153        if (!added)
4154                list_add_tail_rcu(&pa->pa_inode_list,
4155                                        &lg->lg_prealloc_list[order]);
4156        rcu_read_unlock();
4157
4158        /* Now trim the list to be not more than 8 elements */
4159        if (lg_prealloc_count > 8) {
4160                ext4_mb_discard_lg_preallocations(sb, lg,
4161                                                order, lg_prealloc_count);
4162                return;
4163        }
4164        return ;
4165}
4166
4167/*
4168 * release all resource we used in allocation
4169 */
4170static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4171{
4172        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4173        struct ext4_prealloc_space *pa = ac->ac_pa;
4174        if (pa) {
4175                if (pa->pa_type == MB_GROUP_PA) {
4176                        /* see comment in ext4_mb_use_group_pa() */
4177                        spin_lock(&pa->pa_lock);
4178                        pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4179                        pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4180                        pa->pa_free -= ac->ac_b_ex.fe_len;
4181                        pa->pa_len -= ac->ac_b_ex.fe_len;
4182                        spin_unlock(&pa->pa_lock);
4183                }
4184        }
4185        if (pa) {
4186                /*
4187                 * We want to add the pa to the right bucket.
4188                 * Remove it from the list and while adding
4189                 * make sure the list to which we are adding
4190                 * doesn't grow big.
4191                 */
4192                if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4193                        spin_lock(pa->pa_obj_lock);
4194                        list_del_rcu(&pa->pa_inode_list);
4195                        spin_unlock(pa->pa_obj_lock);
4196                        ext4_mb_add_n_trim(ac);
4197                }
4198                ext4_mb_put_pa(ac, ac->ac_sb, pa);
4199        }
4200        if (ac->ac_bitmap_page)
4201                page_cache_release(ac->ac_bitmap_page);
4202        if (ac->ac_buddy_page)
4203                page_cache_release(ac->ac_buddy_page);
4204        if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4205                mutex_unlock(&ac->ac_lg->lg_mutex);
4206        ext4_mb_collect_stats(ac);
4207        return 0;
4208}
4209
4210static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4211{
4212        ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4213        int ret;
4214        int freed = 0;
4215
4216        trace_ext4_mb_discard_preallocations(sb, needed);
4217        for (i = 0; i < ngroups && needed > 0; i++) {
4218                ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4219                freed += ret;
4220                needed -= ret;
4221        }
4222
4223        return freed;
4224}
4225
4226/*
4227 * Main entry point into mballoc to allocate blocks
4228 * it tries to use preallocation first, then falls back
4229 * to usual allocation
4230 */
4231ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4232                                struct ext4_allocation_request *ar, int *errp)
4233{
4234        int freed;
4235        struct ext4_allocation_context *ac = NULL;
4236        struct ext4_sb_info *sbi;
4237        struct super_block *sb;
4238        ext4_fsblk_t block = 0;
4239        unsigned int inquota = 0;
4240        unsigned int reserv_clstrs = 0;
4241
4242        sb = ar->inode->i_sb;
4243        sbi = EXT4_SB(sb);
4244
4245        trace_ext4_request_blocks(ar);
4246
4247        /* Allow to use superuser reservation for quota file */
4248        if (IS_NOQUOTA(ar->inode))
4249                ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4250
4251        /*
4252         * For delayed allocation, we could skip the ENOSPC and
4253         * EDQUOT check, as blocks and quotas have been already
4254         * reserved when data being copied into pagecache.
4255         */
4256        if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4257                ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4258        else {
4259                /* Without delayed allocation we need to verify
4260                 * there is enough free blocks to do block allocation
4261                 * and verify allocation doesn't exceed the quota limits.
4262                 */
4263                while (ar->len &&
4264                        ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4265
4266                        /* let others to free the space */
4267                        yield();
4268                        ar->len = ar->len >> 1;
4269                }
4270                if (!ar->len) {
4271                        *errp = -ENOSPC;
4272                        return 0;
4273                }
4274                reserv_clstrs = ar->len;
4275                if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4276                        dquot_alloc_block_nofail(ar->inode,
4277                                                 EXT4_C2B(sbi, ar->len));
4278                } else {
4279                        while (ar->len &&
4280                                dquot_alloc_block(ar->inode,
4281                                                  EXT4_C2B(sbi, ar->len))) {
4282
4283                                ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4284                                ar->len--;
4285                        }
4286                }
4287                inquota = ar->len;
4288                if (ar->len == 0) {
4289                        *errp = -EDQUOT;
4290                        goto out;
4291                }
4292        }
4293
4294        ac = kmem_cache_alloc(ext4_ac_cachep, GFP_NOFS);
4295        if (!ac) {
4296                ar->len = 0;
4297                *errp = -ENOMEM;
4298                goto out;
4299        }
4300
4301        *errp = ext4_mb_initialize_context(ac, ar);
4302        if (*errp) {
4303                ar->len = 0;
4304                goto out;
4305        }
4306
4307        ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4308        if (!ext4_mb_use_preallocated(ac)) {
4309                ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4310                ext4_mb_normalize_request(ac, ar);
4311repeat:
4312                /* allocate space in core */
4313                *errp = ext4_mb_regular_allocator(ac);
4314                if (*errp)
4315                        goto errout;
4316
4317                /* as we've just preallocated more space than
4318                 * user requested orinally, we store allocated
4319                 * space in a special descriptor */
4320                if (ac->ac_status == AC_STATUS_FOUND &&
4321                                ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4322                        ext4_mb_new_preallocation(ac);
4323        }
4324        if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4325                *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4326                if (*errp == -EAGAIN) {
4327                        /*
4328                         * drop the reference that we took
4329                         * in ext4_mb_use_best_found
4330                         */
4331                        ext4_mb_release_context(ac);
4332                        ac->ac_b_ex.fe_group = 0;
4333                        ac->ac_b_ex.fe_start = 0;
4334                        ac->ac_b_ex.fe_len = 0;
4335                        ac->ac_status = AC_STATUS_CONTINUE;
4336                        goto repeat;
4337                } else if (*errp)
4338                errout:
4339                        ext4_discard_allocated_blocks(ac);
4340                else {
4341                        block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4342                        ar->len = ac->ac_b_ex.fe_len;
4343                }
4344        } else {
4345                freed  = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4346                if (freed)
4347                        goto repeat;
4348                *errp = -ENOSPC;
4349        }
4350
4351        if (*errp) {
4352                ac->ac_b_ex.fe_len = 0;
4353                ar->len = 0;
4354                ext4_mb_show_ac(ac);
4355        }
4356        ext4_mb_release_context(ac);
4357out:
4358        if (ac)
4359                kmem_cache_free(ext4_ac_cachep, ac);
4360        if (inquota && ar->len < inquota)
4361                dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4362        if (!ar->len) {
4363                if (!ext4_test_inode_state(ar->inode,
4364                                           EXT4_STATE_DELALLOC_RESERVED))
4365                        /* release all the reserved blocks if non delalloc */
4366                        percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4367                                                reserv_clstrs);
4368        }
4369
4370        trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4371
4372        return block;
4373}
4374
4375/*
4376 * We can merge two free data extents only if the physical blocks
4377 * are contiguous, AND the extents were freed by the same transaction,
4378 * AND the blocks are associated with the same group.
4379 */
4380static int can_merge(struct ext4_free_data *entry1,
4381                        struct ext4_free_data *entry2)
4382{
4383        if ((entry1->efd_tid == entry2->efd_tid) &&
4384            (entry1->efd_group == entry2->efd_group) &&
4385            ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4386                return 1;
4387        return 0;
4388}
4389
4390static noinline_for_stack int
4391ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4392                      struct ext4_free_data *new_entry)
4393{
4394        ext4_group_t group = e4b->bd_group;
4395        ext4_grpblk_t cluster;
4396        struct ext4_free_data *entry;
4397        struct ext4_group_info *db = e4b->bd_info;
4398        struct super_block *sb = e4b->bd_sb;
4399        struct ext4_sb_info *sbi = EXT4_SB(sb);
4400        struct rb_node **n = &db->bb_free_root.rb_node, *node;
4401        struct rb_node *parent = NULL, *new_node;
4402
4403        BUG_ON(!ext4_handle_valid(handle));
4404        BUG_ON(e4b->bd_bitmap_page == NULL);
4405        BUG_ON(e4b->bd_buddy_page == NULL);
4406
4407        new_node = &new_entry->efd_node;
4408        cluster = new_entry->efd_start_cluster;
4409
4410        if (!*n) {
4411                /* first free block exent. We need to
4412                   protect buddy cache from being freed,
4413                 * otherwise we'll refresh it from
4414                 * on-disk bitmap and lose not-yet-available
4415                 * blocks */
4416                page_cache_get(e4b->bd_buddy_page);
4417                page_cache_get(e4b->bd_bitmap_page);
4418        }
4419        while (*n) {
4420                parent = *n;
4421                entry = rb_entry(parent, struct ext4_free_data, efd_node);
4422                if (cluster < entry->efd_start_cluster)
4423                        n = &(*n)->rb_left;
4424                else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4425                        n = &(*n)->rb_right;
4426                else {
4427                        ext4_grp_locked_error(sb, group, 0,
4428                                ext4_group_first_block_no(sb, group) +
4429                                EXT4_C2B(sbi, cluster),
4430                                "Block already on to-be-freed list");
4431                        return 0;
4432                }
4433        }
4434
4435        rb_link_node(new_node, parent, n);
4436        rb_insert_color(new_node, &db->bb_free_root);
4437
4438        /* Now try to see the extent can be merged to left and right */
4439        node = rb_prev(new_node);
4440        if (node) {
4441                entry = rb_entry(node, struct ext4_free_data, efd_node);
4442                if (can_merge(entry, new_entry)) {
4443                        new_entry->efd_start_cluster = entry->efd_start_cluster;
4444                        new_entry->efd_count += entry->efd_count;
4445                        rb_erase(node, &(db->bb_free_root));
4446                        ext4_journal_callback_del(handle, &entry->efd_jce);
4447                        kmem_cache_free(ext4_free_data_cachep, entry);
4448                }
4449        }
4450
4451        node = rb_next(new_node);
4452        if (node) {
4453                entry = rb_entry(node, struct ext4_free_data, efd_node);
4454                if (can_merge(new_entry, entry)) {
4455                        new_entry->efd_count += entry->efd_count;
4456                        rb_erase(node, &(db->bb_free_root));
4457                        ext4_journal_callback_del(handle, &entry->efd_jce);
4458                        kmem_cache_free(ext4_free_data_cachep, entry);
4459                }
4460        }
4461        /* Add the extent to transaction's private list */
4462        ext4_journal_callback_add(handle, ext4_free_data_callback,
4463                                  &new_entry->efd_jce);
4464        return 0;
4465}
4466
4467/**
4468 * ext4_free_blocks() -- Free given blocks and update quota
4469 * @handle:             handle for this transaction
4470 * @inode:              inode
4471 * @block:              start physical block to free
4472 * @count:              number of blocks to count
4473 * @flags:              flags used by ext4_free_blocks
4474 */
4475void ext4_free_blocks(handle_t *handle, struct inode *inode,
4476                      struct buffer_head *bh, ext4_fsblk_t block,
4477                      unsigned long count, int flags)
4478{
4479        struct buffer_head *bitmap_bh = NULL;
4480        struct super_block *sb = inode->i_sb;
4481        struct ext4_group_desc *gdp;
4482        unsigned long freed = 0;
4483        unsigned int overflow;
4484        ext4_grpblk_t bit;
4485        struct buffer_head *gd_bh;
4486        ext4_group_t block_group;
4487        struct ext4_sb_info *sbi;
4488        struct ext4_buddy e4b;
4489        unsigned int count_clusters;
4490        int err = 0;
4491        int ret;
4492
4493        if (bh) {
4494                if (block)
4495                        BUG_ON(block != bh->b_blocknr);
4496                else
4497                        block = bh->b_blocknr;
4498        }
4499
4500        sbi = EXT4_SB(sb);
4501        if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4502            !ext4_data_block_valid(sbi, block, count)) {
4503                ext4_error(sb, "Freeing blocks not in datazone - "
4504                           "block = %llu, count = %lu", block, count);
4505                goto error_return;
4506        }
4507
4508        ext4_debug("freeing block %llu\n", block);
4509        trace_ext4_free_blocks(inode, block, count, flags);
4510
4511        if (flags & EXT4_FREE_BLOCKS_FORGET) {
4512                struct buffer_head *tbh = bh;
4513                int i;
4514
4515                BUG_ON(bh && (count > 1));
4516
4517                for (i = 0; i < count; i++) {
4518                        if (!bh)
4519                                tbh = sb_find_get_block(inode->i_sb,
4520                                                        block + i);
4521                        if (unlikely(!tbh))
4522                                continue;
4523                        ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4524                                    inode, tbh, block + i);
4525                }
4526        }
4527
4528        /*
4529         * We need to make sure we don't reuse the freed block until
4530         * after the transaction is committed, which we can do by
4531         * treating the block as metadata, below.  We make an
4532         * exception if the inode is to be written in writeback mode
4533         * since writeback mode has weak data consistency guarantees.
4534         */
4535        if (!ext4_should_writeback_data(inode))
4536                flags |= EXT4_FREE_BLOCKS_METADATA;
4537
4538        /*
4539         * If the extent to be freed does not begin on a cluster
4540         * boundary, we need to deal with partial clusters at the
4541         * beginning and end of the extent.  Normally we will free
4542         * blocks at the beginning or the end unless we are explicitly
4543         * requested to avoid doing so.
4544         */
4545        overflow = block & (sbi->s_cluster_ratio - 1);
4546        if (overflow) {
4547                if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4548                        overflow = sbi->s_cluster_ratio - overflow;
4549                        block += overflow;
4550                        if (count > overflow)
4551                                count -= overflow;
4552                        else
4553                                return;
4554                } else {
4555                        block -= overflow;
4556                        count += overflow;
4557                }
4558        }
4559        overflow = count & (sbi->s_cluster_ratio - 1);
4560        if (overflow) {
4561                if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4562                        if (count > overflow)
4563                                count -= overflow;
4564                        else
4565                                return;
4566                } else
4567                        count += sbi->s_cluster_ratio - overflow;
4568        }
4569
4570do_more:
4571        overflow = 0;
4572        ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4573
4574        /*
4575         * Check to see if we are freeing blocks across a group
4576         * boundary.
4577         */
4578        if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4579                overflow = EXT4_C2B(sbi, bit) + count -
4580                        EXT4_BLOCKS_PER_GROUP(sb);
4581                count -= overflow;
4582        }
4583        count_clusters = EXT4_B2C(sbi, count);
4584        bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4585        if (!bitmap_bh) {
4586                err = -EIO;
4587                goto error_return;
4588        }
4589        gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4590        if (!gdp) {
4591                err = -EIO;
4592                goto error_return;
4593        }
4594
4595        if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4596            in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4597            in_range(block, ext4_inode_table(sb, gdp),
4598                     EXT4_SB(sb)->s_itb_per_group) ||
4599            in_range(block + count - 1, ext4_inode_table(sb, gdp),
4600                     EXT4_SB(sb)->s_itb_per_group)) {
4601
4602                ext4_error(sb, "Freeing blocks in system zone - "
4603                           "Block = %llu, count = %lu", block, count);
4604                /* err = 0. ext4_std_error should be a no op */
4605                goto error_return;
4606        }
4607
4608        BUFFER_TRACE(bitmap_bh, "getting write access");
4609        err = ext4_journal_get_write_access(handle, bitmap_bh);
4610        if (err)
4611                goto error_return;
4612
4613        /*
4614         * We are about to modify some metadata.  Call the journal APIs
4615         * to unshare ->b_data if a currently-committing transaction is
4616         * using it
4617         */
4618        BUFFER_TRACE(gd_bh, "get_write_access");
4619        err = ext4_journal_get_write_access(handle, gd_bh);
4620        if (err)
4621                goto error_return;
4622#ifdef AGGRESSIVE_CHECK
4623        {
4624                int i;
4625                for (i = 0; i < count_clusters; i++)
4626                        BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4627        }
4628#endif
4629        trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4630
4631        err = ext4_mb_load_buddy(sb, block_group, &e4b);
4632        if (err)
4633                goto error_return;
4634
4635        if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4636                struct ext4_free_data *new_entry;
4637                /*
4638                 * blocks being freed are metadata. these blocks shouldn't
4639                 * be used until this transaction is committed
4640                 */
4641                new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4642                if (!new_entry) {
4643                        ext4_mb_unload_buddy(&e4b);
4644                        err = -ENOMEM;
4645                        goto error_return;
4646                }
4647                new_entry->efd_start_cluster = bit;
4648                new_entry->efd_group = block_group;
4649                new_entry->efd_count = count_clusters;
4650                new_entry->efd_tid = handle->h_transaction->t_tid;
4651
4652                ext4_lock_group(sb, block_group);
4653                mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4654                ext4_mb_free_metadata(handle, &e4b, new_entry);
4655        } else {
4656                /* need to update group_info->bb_free and bitmap
4657                 * with group lock held. generate_buddy look at
4658                 * them with group lock_held
4659                 */
4660                ext4_lock_group(sb, block_group);
4661                mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4662                mb_free_blocks(inode, &e4b, bit, count_clusters);
4663        }
4664
4665        ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4666        ext4_free_group_clusters_set(sb, gdp, ret);
4667        ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh,
4668                                   EXT4_BLOCKS_PER_GROUP(sb) / 8);
4669        ext4_group_desc_csum_set(sb, block_group, gdp);
4670        ext4_unlock_group(sb, block_group);
4671        percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4672
4673        if (sbi->s_log_groups_per_flex) {
4674                ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4675                atomic_add(count_clusters,
4676                           &sbi->s_flex_groups[flex_group].free_clusters);
4677        }
4678
4679        ext4_mb_unload_buddy(&e4b);
4680
4681        freed += count;
4682
4683        if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4684                dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4685
4686        /* We dirtied the bitmap block */
4687        BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4688        err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4689
4690        /* And the group descriptor block */
4691        BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4692        ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4693        if (!err)
4694                err = ret;
4695
4696        if (overflow && !err) {
4697                block += count;
4698                count = overflow;
4699                put_bh(bitmap_bh);
4700                goto do_more;
4701        }
4702error_return:
4703        brelse(bitmap_bh);
4704        ext4_std_error(sb, err);
4705        return;
4706}
4707
4708/**
4709 * ext4_group_add_blocks() -- Add given blocks to an existing group
4710 * @handle:                     handle to this transaction
4711 * @sb:                         super block
4712 * @block:                      start physcial block to add to the block group
4713 * @count:                      number of blocks to free
4714 *
4715 * This marks the blocks as free in the bitmap and buddy.
4716 */
4717int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4718                         ext4_fsblk_t block, unsigned long count)
4719{
4720        struct buffer_head *bitmap_bh = NULL;
4721        struct buffer_head *gd_bh;
4722        ext4_group_t block_group;
4723        ext4_grpblk_t bit;
4724        unsigned int i;
4725        struct ext4_group_desc *desc;
4726        struct ext4_sb_info *sbi = EXT4_SB(sb);
4727        struct ext4_buddy e4b;
4728        int err = 0, ret, blk_free_count;
4729        ext4_grpblk_t blocks_freed;
4730
4731        ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4732
4733        if (count == 0)
4734                return 0;
4735
4736        ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4737        /*
4738         * Check to see if we are freeing blocks across a group
4739         * boundary.
4740         */
4741        if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4742                ext4_warning(sb, "too much blocks added to group %u\n",
4743                             block_group);
4744                err = -EINVAL;
4745                goto error_return;
4746        }
4747
4748        bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4749        if (!bitmap_bh) {
4750                err = -EIO;
4751                goto error_return;
4752        }
4753
4754        desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4755        if (!desc) {
4756                err = -EIO;
4757                goto error_return;
4758        }
4759
4760        if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4761            in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4762            in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4763            in_range(block + count - 1, ext4_inode_table(sb, desc),
4764                     sbi->s_itb_per_group)) {
4765                ext4_error(sb, "Adding blocks in system zones - "
4766                           "Block = %llu, count = %lu",
4767                           block, count);
4768                err = -EINVAL;
4769                goto error_return;
4770        }
4771
4772        BUFFER_TRACE(bitmap_bh, "getting write access");
4773        err = ext4_journal_get_write_access(handle, bitmap_bh);
4774        if (err)
4775                goto error_return;
4776
4777        /*
4778         * We are about to modify some metadata.  Call the journal APIs
4779         * to unshare ->b_data if a currently-committing transaction is
4780         * using it
4781         */
4782        BUFFER_TRACE(gd_bh, "get_write_access");
4783        err = ext4_journal_get_write_access(handle, gd_bh);
4784        if (err)
4785                goto error_return;
4786
4787        for (i = 0, blocks_freed = 0; i < count; i++) {
4788                BUFFER_TRACE(bitmap_bh, "clear bit");
4789                if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4790                        ext4_error(sb, "bit already cleared for block %llu",
4791                                   (ext4_fsblk_t)(block + i));
4792                        BUFFER_TRACE(bitmap_bh, "bit already cleared");
4793                } else {
4794                        blocks_freed++;
4795                }
4796        }
4797
4798        err = ext4_mb_load_buddy(sb, block_group, &e4b);
4799        if (err)
4800                goto error_return;
4801
4802        /*
4803         * need to update group_info->bb_free and bitmap
4804         * with group lock held. generate_buddy look at
4805         * them with group lock_held
4806         */
4807        ext4_lock_group(sb, block_group);
4808        mb_clear_bits(bitmap_bh->b_data, bit, count);
4809        mb_free_blocks(NULL, &e4b, bit, count);
4810        blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
4811        ext4_free_group_clusters_set(sb, desc, blk_free_count);
4812        ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh,
4813                                   EXT4_BLOCKS_PER_GROUP(sb) / 8);
4814        ext4_group_desc_csum_set(sb, block_group, desc);
4815        ext4_unlock_group(sb, block_group);
4816        percpu_counter_add(&sbi->s_freeclusters_counter,
4817                           EXT4_B2C(sbi, blocks_freed));
4818
4819        if (sbi->s_log_groups_per_flex) {
4820                ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4821                atomic_add(EXT4_B2C(sbi, blocks_freed),
4822                           &sbi->s_flex_groups[flex_group].free_clusters);
4823        }
4824
4825        ext4_mb_unload_buddy(&e4b);
4826
4827        /* We dirtied the bitmap block */
4828        BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4829        err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4830
4831        /* And the group descriptor block */
4832        BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4833        ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4834        if (!err)
4835                err = ret;
4836
4837error_return:
4838        brelse(bitmap_bh);
4839        ext4_std_error(sb, err);
4840        return err;
4841}
4842
4843/**
4844 * ext4_trim_extent -- function to TRIM one single free extent in the group
4845 * @sb:         super block for the file system
4846 * @start:      starting block of the free extent in the alloc. group
4847 * @count:      number of blocks to TRIM
4848 * @group:      alloc. group we are working with
4849 * @e4b:        ext4 buddy for the group
4850 *
4851 * Trim "count" blocks starting at "start" in the "group". To assure that no
4852 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4853 * be called with under the group lock.
4854 */
4855static void ext4_trim_extent(struct super_block *sb, int start, int count,
4856                             ext4_group_t group, struct ext4_buddy *e4b)
4857{
4858        struct ext4_free_extent ex;
4859
4860        trace_ext4_trim_extent(sb, group, start, count);
4861
4862        assert_spin_locked(ext4_group_lock_ptr(sb, group));
4863
4864        ex.fe_start = start;
4865        ex.fe_group = group;
4866        ex.fe_len = count;
4867
4868        /*
4869         * Mark blocks used, so no one can reuse them while
4870         * being trimmed.
4871         */
4872        mb_mark_used(e4b, &ex);
4873        ext4_unlock_group(sb, group);
4874        ext4_issue_discard(sb, group, start, count);
4875        ext4_lock_group(sb, group);
4876        mb_free_blocks(NULL, e4b, start, ex.fe_len);
4877}
4878
4879/**
4880 * ext4_trim_all_free -- function to trim all free space in alloc. group
4881 * @sb:                 super block for file system
4882 * @group:              group to be trimmed
4883 * @start:              first group block to examine
4884 * @max:                last group block to examine
4885 * @minblocks:          minimum extent block count
4886 *
4887 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4888 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4889 * the extent.
4890 *
4891 *
4892 * ext4_trim_all_free walks through group's block bitmap searching for free
4893 * extents. When the free extent is found, mark it as used in group buddy
4894 * bitmap. Then issue a TRIM command on this extent and free the extent in
4895 * the group buddy bitmap. This is done until whole group is scanned.
4896 */
4897static ext4_grpblk_t
4898ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4899                   ext4_grpblk_t start, ext4_grpblk_t max,
4900                   ext4_grpblk_t minblocks)
4901{
4902        void *bitmap;
4903        ext4_grpblk_t next, count = 0, free_count = 0;
4904        struct ext4_buddy e4b;
4905        int ret;
4906
4907        trace_ext4_trim_all_free(sb, group, start, max);
4908
4909        ret = ext4_mb_load_buddy(sb, group, &e4b);
4910        if (ret) {
4911                ext4_error(sb, "Error in loading buddy "
4912                                "information for %u", group);
4913                return ret;
4914        }
4915        bitmap = e4b.bd_bitmap;
4916
4917        ext4_lock_group(sb, group);
4918        if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4919            minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4920                goto out;
4921
4922        start = (e4b.bd_info->bb_first_free > start) ?
4923                e4b.bd_info->bb_first_free : start;
4924
4925        while (start <= max) {
4926                start = mb_find_next_zero_bit(bitmap, max + 1, start);
4927                if (start > max)
4928                        break;
4929                next = mb_find_next_bit(bitmap, max + 1, start);
4930
4931                if ((next - start) >= minblocks) {
4932                        ext4_trim_extent(sb, start,
4933                                         next - start, group, &e4b);
4934                        count += next - start;
4935                }
4936                free_count += next - start;
4937                start = next + 1;
4938
4939                if (fatal_signal_pending(current)) {
4940                        count = -ERESTARTSYS;
4941                        break;
4942                }
4943
4944                if (need_resched()) {
4945                        ext4_unlock_group(sb, group);
4946                        cond_resched();
4947                        ext4_lock_group(sb, group);
4948                }
4949
4950                if ((e4b.bd_info->bb_free - free_count) < minblocks)
4951                        break;
4952        }
4953
4954        if (!ret)
4955                EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
4956out:
4957        ext4_unlock_group(sb, group);
4958        ext4_mb_unload_buddy(&e4b);
4959
4960        ext4_debug("trimmed %d blocks in the group %d\n",
4961                count, group);
4962
4963        return count;
4964}
4965
4966/**
4967 * ext4_trim_fs() -- trim ioctl handle function
4968 * @sb:                 superblock for filesystem
4969 * @range:              fstrim_range structure
4970 *
4971 * start:       First Byte to trim
4972 * len:         number of Bytes to trim from start
4973 * minlen:      minimum extent length in Bytes
4974 * ext4_trim_fs goes through all allocation groups containing Bytes from
4975 * start to start+len. For each such a group ext4_trim_all_free function
4976 * is invoked to trim all free space.
4977 */
4978int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
4979{
4980        struct ext4_group_info *grp;
4981        ext4_group_t group, first_group, last_group;
4982        ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
4983        uint64_t start, end, minlen, trimmed = 0;
4984        ext4_fsblk_t first_data_blk =
4985                        le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
4986        ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
4987        int ret = 0;
4988
4989        start = range->start >> sb->s_blocksize_bits;
4990        end = start + (range->len >> sb->s_blocksize_bits) - 1;
4991        minlen = range->minlen >> sb->s_blocksize_bits;
4992
4993        if (unlikely(minlen > EXT4_CLUSTERS_PER_GROUP(sb)) ||
4994            unlikely(start >= max_blks))
4995                return -EINVAL;
4996        if (end >= max_blks)
4997                end = max_blks - 1;
4998        if (end <= first_data_blk)
4999                goto out;
5000        if (start < first_data_blk)
5001                start = first_data_blk;
5002
5003        /* Determine first and last group to examine based on start and end */
5004        ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5005                                     &first_group, &first_cluster);
5006        ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5007                                     &last_group, &last_cluster);
5008
5009        /* end now represents the last cluster to discard in this group */
5010        end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5011
5012        for (group = first_group; group <= last_group; group++) {
5013                grp = ext4_get_group_info(sb, group);
5014                /* We only do this if the grp has never been initialized */
5015                if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5016                        ret = ext4_mb_init_group(sb, group);
5017                        if (ret)
5018                                break;
5019                }
5020
5021                /*
5022                 * For all the groups except the last one, last cluster will
5023                 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5024                 * change it for the last group, note that last_cluster is
5025                 * already computed earlier by ext4_get_group_no_and_offset()
5026                 */
5027                if (group == last_group)
5028                        end = last_cluster;
5029
5030                if (grp->bb_free >= minlen) {
5031                        cnt = ext4_trim_all_free(sb, group, first_cluster,
5032                                                end, minlen);
5033                        if (cnt < 0) {
5034                                ret = cnt;
5035                                break;
5036                        }
5037                        trimmed += cnt;
5038                }
5039
5040                /*
5041                 * For every group except the first one, we are sure
5042                 * that the first cluster to discard will be cluster #0.
5043                 */
5044                first_cluster = 0;
5045        }
5046
5047        if (!ret)
5048                atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5049
5050out:
5051        range->len = trimmed * sb->s_blocksize;
5052        return ret;
5053}
5054
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.