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/log2.h>
  28#include <linux/slab.h>
  29#include <trace/events/ext4.h>
  30
  31/*
  32 * MUSTDO:
  33 *   - test ext4_ext_search_left() and ext4_ext_search_right()
  34 *   - search for metadata in few groups
  35 *
  36 * TODO v4:
  37 *   - normalization should take into account whether file is still open
  38 *   - discard preallocations if no free space left (policy?)
  39 *   - don't normalize tails
  40 *   - quota
  41 *   - reservation for superuser
  42 *
  43 * TODO v3:
  44 *   - bitmap read-ahead (proposed by Oleg Drokin aka green)
  45 *   - track min/max extents in each group for better group selection
  46 *   - mb_mark_used() may allocate chunk right after splitting buddy
  47 *   - tree of groups sorted by number of free blocks
  48 *   - error handling
  49 */
  50
  51/*
  52 * The allocation request involve request for multiple number of blocks
  53 * near to the goal(block) value specified.
  54 *
  55 * During initialization phase of the allocator we decide to use the
  56 * group preallocation or inode preallocation depending on the size of
  57 * the file. The size of the file could be the resulting file size we
  58 * would have after allocation, or the current file size, which ever
  59 * is larger. If the size is less than sbi->s_mb_stream_request we
  60 * select to use the group preallocation. The default value of
  61 * s_mb_stream_request is 16 blocks. This can also be tuned via
  62 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
  63 * terms of number of blocks.
  64 *
  65 * The main motivation for having small file use group preallocation is to
  66 * ensure that we have small files closer together on the disk.
  67 *
  68 * First stage the allocator looks at the inode prealloc list,
  69 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
  70 * spaces for this particular inode. The inode prealloc space is
  71 * represented as:
  72 *
  73 * pa_lstart -> the logical start block for this prealloc space
  74 * pa_pstart -> the physical start block for this prealloc space
  75 * pa_len    -> length for this prealloc space (in clusters)
  76 * pa_free   ->  free space available in this prealloc space (in clusters)
  77 *
  78 * The inode preallocation space is used looking at the _logical_ start
  79 * block. If only the logical file block falls within the range of prealloc
  80 * space we will consume the particular prealloc space. This makes sure that
  81 * we have contiguous physical blocks representing the file blocks
  82 *
  83 * The important thing to be noted in case of inode prealloc space is that
  84 * we don't modify the values associated to inode prealloc space except
  85 * pa_free.
  86 *
  87 * If we are not able to find blocks in the inode prealloc space and if we
  88 * have the group allocation flag set then we look at the locality group
  89 * prealloc space. These are per CPU prealloc list represented as
  90 *
  91 * ext4_sb_info.s_locality_groups[smp_processor_id()]
  92 *
  93 * The reason for having a per cpu locality group is to reduce the contention
  94 * between CPUs. It is possible to get scheduled at this point.
  95 *
  96 * The locality group prealloc space is used looking at whether we have
  97 * enough free space (pa_free) within the prealloc space.
  98 *
  99 * If we can't allocate blocks via inode prealloc or/and locality group
 100 * prealloc then we look at the buddy cache. The buddy cache is represented
 101 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
 102 * mapped to the buddy and bitmap information regarding different
 103 * groups. The buddy information is attached to buddy cache inode so that
 104 * we can access them through the page cache. The information regarding
 105 * each group is loaded via ext4_mb_load_buddy.  The information involve
 106 * block bitmap and buddy information. The information are stored in the
 107 * inode as:
 108 *
 109 *  {                        page                        }
 110 *  [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
 111 *
 112 *
 113 * one block each for bitmap and buddy information.  So for each group we
 114 * take up 2 blocks. A page can contain blocks_per_page (PAGE_CACHE_SIZE /
 115 * blocksize) blocks.  So it can have information regarding groups_per_page
 116 * which is blocks_per_page/2
 117 *
 118 * The buddy cache inode is not stored on disk. The inode is thrown
 119 * away when the filesystem is unmounted.
 120 *
 121 * We look for count number of blocks in the buddy cache. If we were able
 122 * to locate that many free blocks we return with additional information
 123 * regarding rest of the contiguous physical block available
 124 *
 125 * Before allocating blocks via buddy cache we normalize the request
 126 * blocks. This ensure we ask for more blocks that we needed. The extra
 127 * blocks that we get after allocation is added to the respective prealloc
 128 * list. In case of inode preallocation we follow a list of heuristics
 129 * based on file size. This can be found in ext4_mb_normalize_request. If
 130 * we are doing a group prealloc we try to normalize the request to
 131 * sbi->s_mb_group_prealloc.  The default value of s_mb_group_prealloc is
 132 * dependent on the cluster size; for non-bigalloc file systems, it is
 133 * 512 blocks. This can be tuned via
 134 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
 135 * terms of number of blocks. If we have mounted the file system with -O
 136 * stripe=<value> option the group prealloc request is normalized to the
 137 * the smallest multiple of the stripe value (sbi->s_stripe) which is
 138 * greater than the default mb_group_prealloc.
 139 *
 140 * The regular allocator (using the buddy cache) supports a few tunables.
 141 *
 142 * /sys/fs/ext4/<partition>/mb_min_to_scan
 143 * /sys/fs/ext4/<partition>/mb_max_to_scan
 144 * /sys/fs/ext4/<partition>/mb_order2_req
 145 *
 146 * The regular allocator uses buddy scan only if the request len is power of
 147 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
 148 * value of s_mb_order2_reqs can be tuned via
 149 * /sys/fs/ext4/<partition>/mb_order2_req.  If the request len is equal to
 150 * stripe size (sbi->s_stripe), we try to search for contiguous block in
 151 * stripe size. This should result in better allocation on RAID setups. If
 152 * not, we search in the specific group using bitmap for best extents. The
 153 * tunable min_to_scan and max_to_scan control the behaviour here.
 154 * min_to_scan indicate how long the mballoc __must__ look for a best
 155 * extent and max_to_scan indicates how long the mballoc __can__ look for a
 156 * best extent in the found extents. Searching for the blocks starts with
 157 * the group specified as the goal value in allocation context via
 158 * ac_g_ex. Each group is first checked based on the criteria whether it
 159 * can be used for allocation. ext4_mb_good_group explains how the groups are
 160 * checked.
 161 *
 162 * Both the prealloc space are getting populated as above. So for the first
 163 * request we will hit the buddy cache which will result in this prealloc
 164 * space getting filled. The prealloc space is then later used for the
 165 * subsequent request.
 166 */
 167
 168/*
 169 * mballoc operates on the following data:
 170 *  - on-disk bitmap
 171 *  - in-core buddy (actually includes buddy and bitmap)
 172 *  - preallocation descriptors (PAs)
 173 *
 174 * there are two types of preallocations:
 175 *  - inode
 176 *    assiged to specific inode and can be used for this inode only.
 177 *    it describes part of inode's space preallocated to specific
 178 *    physical blocks. any block from that preallocated can be used
 179 *    independent. the descriptor just tracks number of blocks left
 180 *    unused. so, before taking some block from descriptor, one must
 181 *    make sure corresponded logical block isn't allocated yet. this
 182 *    also means that freeing any block within descriptor's range
 183 *    must discard all preallocated blocks.
 184 *  - locality group
 185 *    assigned to specific locality group which does not translate to
 186 *    permanent set of inodes: inode can join and leave group. space
 187 *    from this type of preallocation can be used for any inode. thus
 188 *    it's consumed from the beginning to the end.
 189 *
 190 * relation between them can be expressed as:
 191 *    in-core buddy = on-disk bitmap + preallocation descriptors
 192 *
 193 * this mean blocks mballoc considers used are:
 194 *  - allocated blocks (persistent)
 195 *  - preallocated blocks (non-persistent)
 196 *
 197 * consistency in mballoc world means that at any time a block is either
 198 * free or used in ALL structures. notice: "any time" should not be read
 199 * literally -- time is discrete and delimited by locks.
 200 *
 201 *  to keep it simple, we don't use block numbers, instead we count number of
 202 *  blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
 203 *
 204 * all operations can be expressed as:
 205 *  - init buddy:                       buddy = on-disk + PAs
 206 *  - new PA:                           buddy += N; PA = N
 207 *  - use inode PA:                     on-disk += N; PA -= N
 208 *  - discard inode PA                  buddy -= on-disk - PA; PA = 0
 209 *  - use locality group PA             on-disk += N; PA -= N
 210 *  - discard locality group PA         buddy -= PA; PA = 0
 211 *  note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
 212 *        is used in real operation because we can't know actual used
 213 *        bits from PA, only from on-disk bitmap
 214 *
 215 * if we follow this strict logic, then all operations above should be atomic.
 216 * given some of them can block, we'd have to use something like semaphores
 217 * killing performance on high-end SMP hardware. let's try to relax it using
 218 * the following knowledge:
 219 *  1) if buddy is referenced, it's already initialized
 220 *  2) while block is used in buddy and the buddy is referenced,
 221 *     nobody can re-allocate that block
 222 *  3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
 223 *     bit set and PA claims same block, it's OK. IOW, one can set bit in
 224 *     on-disk bitmap if buddy has same bit set or/and PA covers corresponded
 225 *     block
 226 *
 227 * so, now we're building a concurrency table:
 228 *  - init buddy vs.
 229 *    - new PA
 230 *      blocks for PA are allocated in the buddy, buddy must be referenced
 231 *      until PA is linked to allocation group to avoid concurrent buddy init
 232 *    - use inode PA
 233 *      we need to make sure that either on-disk bitmap or PA has uptodate data
 234 *      given (3) we care that PA-=N operation doesn't interfere with init
 235 *    - discard inode PA
 236 *      the simplest way would be to have buddy initialized by the discard
 237 *    - use locality group PA
 238 *      again PA-=N must be serialized with init
 239 *    - discard locality group PA
 240 *      the simplest way would be to have buddy initialized by the discard
 241 *  - new PA vs.
 242 *    - use inode PA
 243 *      i_data_sem serializes them
 244 *    - discard inode PA
 245 *      discard process must wait until PA isn't used by another process
 246 *    - use locality group PA
 247 *      some mutex should serialize them
 248 *    - discard locality group PA
 249 *      discard process must wait until PA isn't used by another process
 250 *  - use inode PA
 251 *    - use inode PA
 252 *      i_data_sem or another mutex should serializes them
 253 *    - discard inode PA
 254 *      discard process must wait until PA isn't used by another process
 255 *    - use locality group PA
 256 *      nothing wrong here -- they're different PAs covering different blocks
 257 *    - discard locality group PA
 258 *      discard process must wait until PA isn't used by another process
 259 *
 260 * now we're ready to make few consequences:
 261 *  - PA is referenced and while it is no discard is possible
 262 *  - PA is referenced until block isn't marked in on-disk bitmap
 263 *  - PA changes only after on-disk bitmap
 264 *  - discard must not compete with init. either init is done before
 265 *    any discard or they're serialized somehow
 266 *  - buddy init as sum of on-disk bitmap and PAs is done atomically
 267 *
 268 * a special case when we've used PA to emptiness. no need to modify buddy
 269 * in this case, but we should care about concurrent init
 270 *
 271 */
 272
 273 /*
 274 * Logic in few words:
 275 *
 276 *  - allocation:
 277 *    load group
 278 *    find blocks
 279 *    mark bits in on-disk bitmap
 280 *    release group
 281 *
 282 *  - use preallocation:
 283 *    find proper PA (per-inode or group)
 284 *    load group
 285 *    mark bits in on-disk bitmap
 286 *    release group
 287 *    release PA
 288 *
 289 *  - free:
 290 *    load group
 291 *    mark bits in on-disk bitmap
 292 *    release group
 293 *
 294 *  - discard preallocations in group:
 295 *    mark PAs deleted
 296 *    move them onto local list
 297 *    load on-disk bitmap
 298 *    load group
 299 *    remove PA from object (inode or locality group)
 300 *    mark free blocks in-core
 301 *
 302 *  - discard inode's preallocations:
 303 */
 304
 305/*
 306 * Locking rules
 307 *
 308 * Locks:
 309 *  - bitlock on a group        (group)
 310 *  - object (inode/locality)   (object)
 311 *  - per-pa lock               (pa)
 312 *
 313 * Paths:
 314 *  - new pa
 315 *    object
 316 *    group
 317 *
 318 *  - find and use pa:
 319 *    pa
 320 *
 321 *  - release consumed pa:
 322 *    pa
 323 *    group
 324 *    object
 325 *
 326 *  - generate in-core bitmap:
 327 *    group
 328 *        pa
 329 *
 330 *  - discard all for given object (inode, locality group):
 331 *    object
 332 *        pa
 333 *    group
 334 *
 335 *  - discard all for given group:
 336 *    group
 337 *        pa
 338 *    group
 339 *        object
 340 *
 341 */
 342static struct kmem_cache *ext4_pspace_cachep;
 343static struct kmem_cache *ext4_ac_cachep;
 344static struct kmem_cache *ext4_free_data_cachep;
 345
 346/* We create slab caches for groupinfo data structures based on the
 347 * superblock block size.  There will be one per mounted filesystem for
 348 * each unique s_blocksize_bits */
 349#define NR_GRPINFO_CACHES 8
 350static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
 351
 352static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
 353        "ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
 354        "ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
 355        "ext4_groupinfo_64k", "ext4_groupinfo_128k"
 356};
 357
 358static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 359                                        ext4_group_t group);
 360static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
 361                                                ext4_group_t group);
 362static void ext4_free_data_callback(struct super_block *sb,
 363                                struct ext4_journal_cb_entry *jce, int rc);
 364
 365static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
 366{
 367#if BITS_PER_LONG == 64
 368        *bit += ((unsigned long) addr & 7UL) << 3;
 369        addr = (void *) ((unsigned long) addr & ~7UL);
 370#elif BITS_PER_LONG == 32
 371        *bit += ((unsigned long) addr & 3UL) << 3;
 372        addr = (void *) ((unsigned long) addr & ~3UL);
 373#else
 374#error "how many bits you are?!"
 375#endif
 376        return addr;
 377}
 378
 379static inline int mb_test_bit(int bit, void *addr)
 380{
 381        /*
 382         * ext4_test_bit on architecture like powerpc
 383         * needs unsigned long aligned address
 384         */
 385        addr = mb_correct_addr_and_bit(&bit, addr);
 386        return ext4_test_bit(bit, addr);
 387}
 388
 389static inline void mb_set_bit(int bit, void *addr)
 390{
 391        addr = mb_correct_addr_and_bit(&bit, addr);
 392        ext4_set_bit(bit, addr);
 393}
 394
 395static inline void mb_clear_bit(int bit, void *addr)
 396{
 397        addr = mb_correct_addr_and_bit(&bit, addr);
 398        ext4_clear_bit(bit, addr);
 399}
 400
 401static inline int mb_find_next_zero_bit(void *addr, int max, int start)
 402{
 403        int fix = 0, ret, tmpmax;
 404        addr = mb_correct_addr_and_bit(&fix, addr);
 405        tmpmax = max + fix;
 406        start += fix;
 407
 408        ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
 409        if (ret > max)
 410                return max;
 411        return ret;
 412}
 413
 414static inline int mb_find_next_bit(void *addr, int max, int start)
 415{
 416        int fix = 0, ret, tmpmax;
 417        addr = mb_correct_addr_and_bit(&fix, addr);
 418        tmpmax = max + fix;
 419        start += fix;
 420
 421        ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
 422        if (ret > max)
 423                return max;
 424        return ret;
 425}
 426
 427static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
 428{
 429        char *bb;
 430
 431        BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
 432        BUG_ON(max == NULL);
 433
 434        if (order > e4b->bd_blkbits + 1) {
 435                *max = 0;
 436                return NULL;
 437        }
 438
 439        /* at order 0 we see each particular block */
 440        if (order == 0) {
 441                *max = 1 << (e4b->bd_blkbits + 3);
 442                return e4b->bd_bitmap;
 443        }
 444
 445        bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
 446        *max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
 447
 448        return bb;
 449}
 450
 451#ifdef DOUBLE_CHECK
 452static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
 453                           int first, int count)
 454{
 455        int i;
 456        struct super_block *sb = e4b->bd_sb;
 457
 458        if (unlikely(e4b->bd_info->bb_bitmap == NULL))
 459                return;
 460        assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
 461        for (i = 0; i < count; i++) {
 462                if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
 463                        ext4_fsblk_t blocknr;
 464
 465                        blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
 466                        blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
 467                        ext4_grp_locked_error(sb, e4b->bd_group,
 468                                              inode ? inode->i_ino : 0,
 469                                              blocknr,
 470                                              "freeing block already freed "
 471                                              "(bit %u)",
 472                                              first + i);
 473                }
 474                mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
 475        }
 476}
 477
 478static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
 479{
 480        int i;
 481
 482        if (unlikely(e4b->bd_info->bb_bitmap == NULL))
 483                return;
 484        assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
 485        for (i = 0; i < count; i++) {
 486                BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
 487                mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
 488        }
 489}
 490
 491static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
 492{
 493        if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
 494                unsigned char *b1, *b2;
 495                int i;
 496                b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
 497                b2 = (unsigned char *) bitmap;
 498                for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
 499                        if (b1[i] != b2[i]) {
 500                                ext4_msg(e4b->bd_sb, KERN_ERR,
 501                                         "corruption in group %u "
 502                                         "at byte %u(%u): %x in copy != %x "
 503                                         "on disk/prealloc",
 504                                         e4b->bd_group, i, i * 8, b1[i], b2[i]);
 505                                BUG();
 506                        }
 507                }
 508        }
 509}
 510
 511#else
 512static inline void mb_free_blocks_double(struct inode *inode,
 513                                struct ext4_buddy *e4b, int first, int count)
 514{
 515        return;
 516}
 517static inline void mb_mark_used_double(struct ext4_buddy *e4b,
 518                                                int first, int count)
 519{
 520        return;
 521}
 522static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
 523{
 524        return;
 525}
 526#endif
 527
 528#ifdef AGGRESSIVE_CHECK
 529
 530#define MB_CHECK_ASSERT(assert)                                         \
 531do {                                                                    \
 532        if (!(assert)) {                                                \
 533                printk(KERN_EMERG                                       \
 534                        "Assertion failure in %s() at %s:%d: \"%s\"\n", \
 535                        function, file, line, # assert);                \
 536                BUG();                                                  \
 537        }                                                               \
 538} while (0)
 539
 540static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
 541                                const char *function, int line)
 542{
 543        struct super_block *sb = e4b->bd_sb;
 544        int order = e4b->bd_blkbits + 1;
 545        int max;
 546        int max2;
 547        int i;
 548        int j;
 549        int k;
 550        int count;
 551        struct ext4_group_info *grp;
 552        int fragments = 0;
 553        int fstart;
 554        struct list_head *cur;
 555        void *buddy;
 556        void *buddy2;
 557
 558        {
 559                static int mb_check_counter;
 560                if (mb_check_counter++ % 100 != 0)
 561                        return 0;
 562        }
 563
 564        while (order > 1) {
 565                buddy = mb_find_buddy(e4b, order, &max);
 566                MB_CHECK_ASSERT(buddy);
 567                buddy2 = mb_find_buddy(e4b, order - 1, &max2);
 568                MB_CHECK_ASSERT(buddy2);
 569                MB_CHECK_ASSERT(buddy != buddy2);
 570                MB_CHECK_ASSERT(max * 2 == max2);
 571
 572                count = 0;
 573                for (i = 0; i < max; i++) {
 574
 575                        if (mb_test_bit(i, buddy)) {
 576                                /* only single bit in buddy2 may be 1 */
 577                                if (!mb_test_bit(i << 1, buddy2)) {
 578                                        MB_CHECK_ASSERT(
 579                                                mb_test_bit((i<<1)+1, buddy2));
 580                                } else if (!mb_test_bit((i << 1) + 1, buddy2)) {
 581                                        MB_CHECK_ASSERT(
 582                                                mb_test_bit(i << 1, buddy2));
 583                                }
 584                                continue;
 585                        }
 586
 587                        /* both bits in buddy2 must be 1 */
 588                        MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
 589                        MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
 590
 591                        for (j = 0; j < (1 << order); j++) {
 592                                k = (i * (1 << order)) + j;
 593                                MB_CHECK_ASSERT(
 594                                        !mb_test_bit(k, e4b->bd_bitmap));
 595                        }
 596                        count++;
 597                }
 598                MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
 599                order--;
 600        }
 601
 602        fstart = -1;
 603        buddy = mb_find_buddy(e4b, 0, &max);
 604        for (i = 0; i < max; i++) {
 605                if (!mb_test_bit(i, buddy)) {
 606                        MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
 607                        if (fstart == -1) {
 608                                fragments++;
 609                                fstart = i;
 610                        }
 611                        continue;
 612                }
 613                fstart = -1;
 614                /* check used bits only */
 615                for (j = 0; j < e4b->bd_blkbits + 1; j++) {
 616                        buddy2 = mb_find_buddy(e4b, j, &max2);
 617                        k = i >> j;
 618                        MB_CHECK_ASSERT(k < max2);
 619                        MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
 620                }
 621        }
 622        MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
 623        MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
 624
 625        grp = ext4_get_group_info(sb, e4b->bd_group);
 626        list_for_each(cur, &grp->bb_prealloc_list) {
 627                ext4_group_t groupnr;
 628                struct ext4_prealloc_space *pa;
 629                pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
 630                ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
 631                MB_CHECK_ASSERT(groupnr == e4b->bd_group);
 632                for (i = 0; i < pa->pa_len; i++)
 633                        MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
 634        }
 635        return 0;
 636}
 637#undef MB_CHECK_ASSERT
 638#define mb_check_buddy(e4b) __mb_check_buddy(e4b,       \
 639                                        __FILE__, __func__, __LINE__)
 640#else
 641#define mb_check_buddy(e4b)
 642#endif
 643
 644/*
 645 * Divide blocks started from @first with length @len into
 646 * smaller chunks with power of 2 blocks.
 647 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
 648 * then increase bb_counters[] for corresponded chunk size.
 649 */
 650static void ext4_mb_mark_free_simple(struct super_block *sb,
 651                                void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
 652                                        struct ext4_group_info *grp)
 653{
 654        struct ext4_sb_info *sbi = EXT4_SB(sb);
 655        ext4_grpblk_t min;
 656        ext4_grpblk_t max;
 657        ext4_grpblk_t chunk;
 658        unsigned short border;
 659
 660        BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
 661
 662        border = 2 << sb->s_blocksize_bits;
 663
 664        while (len > 0) {
 665                /* find how many blocks can be covered since this position */
 666                max = ffs(first | border) - 1;
 667
 668                /* find how many blocks of power 2 we need to mark */
 669                min = fls(len) - 1;
 670
 671                if (max < min)
 672                        min = max;
 673                chunk = 1 << min;
 674
 675                /* mark multiblock chunks only */
 676                grp->bb_counters[min]++;
 677                if (min > 0)
 678                        mb_clear_bit(first >> min,
 679                                     buddy + sbi->s_mb_offsets[min]);
 680
 681                len -= chunk;
 682                first += chunk;
 683        }
 684}
 685
 686/*
 687 * Cache the order of the largest free extent we have available in this block
 688 * group.
 689 */
 690static void
 691mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
 692{
 693        int i;
 694        int bits;
 695
 696        grp->bb_largest_free_order = -1; /* uninit */
 697
 698        bits = sb->s_blocksize_bits + 1;
 699        for (i = bits; i >= 0; i--) {
 700                if (grp->bb_counters[i] > 0) {
 701                        grp->bb_largest_free_order = i;
 702                        break;
 703                }
 704        }
 705}
 706
 707static noinline_for_stack
 708void ext4_mb_generate_buddy(struct super_block *sb,
 709                                void *buddy, void *bitmap, ext4_group_t group)
 710{
 711        struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 712        ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
 713        ext4_grpblk_t i = 0;
 714        ext4_grpblk_t first;
 715        ext4_grpblk_t len;
 716        unsigned free = 0;
 717        unsigned fragments = 0;
 718        unsigned long long period = get_cycles();
 719
 720        /* initialize buddy from bitmap which is aggregation
 721         * of on-disk bitmap and preallocations */
 722        i = mb_find_next_zero_bit(bitmap, max, 0);
 723        grp->bb_first_free = i;
 724        while (i < max) {
 725                fragments++;
 726                first = i;
 727                i = mb_find_next_bit(bitmap, max, i);
 728                len = i - first;
 729                free += len;
 730                if (len > 1)
 731                        ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
 732                else
 733                        grp->bb_counters[0]++;
 734                if (i < max)
 735                        i = mb_find_next_zero_bit(bitmap, max, i);
 736        }
 737        grp->bb_fragments = fragments;
 738
 739        if (free != grp->bb_free) {
 740                ext4_grp_locked_error(sb, group, 0, 0,
 741                                      "%u clusters in bitmap, %u in gd",
 742                                      free, grp->bb_free);
 743                /*
 744                 * If we intent to continue, we consider group descritor
 745                 * corrupt and update bb_free using bitmap value
 746                 */
 747                grp->bb_free = free;
 748        }
 749        mb_set_largest_free_order(sb, grp);
 750
 751        clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
 752
 753        period = get_cycles() - period;
 754        spin_lock(&EXT4_SB(sb)->s_bal_lock);
 755        EXT4_SB(sb)->s_mb_buddies_generated++;
 756        EXT4_SB(sb)->s_mb_generation_time += period;
 757        spin_unlock(&EXT4_SB(sb)->s_bal_lock);
 758}
 759
 760/* The buddy information is attached the buddy cache inode
 761 * for convenience. The information regarding each group
 762 * is loaded via ext4_mb_load_buddy. The information involve
 763 * block bitmap and buddy information. The information are
 764 * stored in the inode as
 765 *
 766 * {                        page                        }
 767 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
 768 *
 769 *
 770 * one block each for bitmap and buddy information.
 771 * So for each group we take up 2 blocks. A page can
 772 * contain blocks_per_page (PAGE_CACHE_SIZE / blocksize)  blocks.
 773 * So it can have information regarding groups_per_page which
 774 * is blocks_per_page/2
 775 *
 776 * Locking note:  This routine takes the block group lock of all groups
 777 * for this page; do not hold this lock when calling this routine!
 778 */
 779
 780static int ext4_mb_init_cache(struct page *page, char *incore)
 781{
 782        ext4_group_t ngroups;
 783        int blocksize;
 784        int blocks_per_page;
 785        int groups_per_page;
 786        int err = 0;
 787        int i;
 788        ext4_group_t first_group, group;
 789        int first_block;
 790        struct super_block *sb;
 791        struct buffer_head *bhs;
 792        struct buffer_head **bh = NULL;
 793        struct inode *inode;
 794        char *data;
 795        char *bitmap;
 796        struct ext4_group_info *grinfo;
 797
 798        mb_debug(1, "init page %lu\n", page->index);
 799
 800        inode = page->mapping->host;
 801        sb = inode->i_sb;
 802        ngroups = ext4_get_groups_count(sb);
 803        blocksize = 1 << inode->i_blkbits;
 804        blocks_per_page = PAGE_CACHE_SIZE / blocksize;
 805
 806        groups_per_page = blocks_per_page >> 1;
 807        if (groups_per_page == 0)
 808                groups_per_page = 1;
 809
 810        /* allocate buffer_heads to read bitmaps */
 811        if (groups_per_page > 1) {
 812                i = sizeof(struct buffer_head *) * groups_per_page;
 813                bh = kzalloc(i, GFP_NOFS);
 814                if (bh == NULL) {
 815                        err = -ENOMEM;
 816                        goto out;
 817                }
 818        } else
 819                bh = &bhs;
 820
 821        first_group = page->index * blocks_per_page / 2;
 822
 823        /* read all groups the page covers into the cache */
 824        for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
 825                if (group >= ngroups)
 826                        break;
 827
 828                grinfo = ext4_get_group_info(sb, group);
 829                /*
 830                 * If page is uptodate then we came here after online resize
 831                 * which added some new uninitialized group info structs, so
 832                 * we must skip all initialized uptodate buddies on the page,
 833                 * which may be currently in use by an allocating task.
 834                 */
 835                if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
 836                        bh[i] = NULL;
 837                        continue;
 838                }
 839                if (!(bh[i] = ext4_read_block_bitmap_nowait(sb, group))) {
 840                        err = -ENOMEM;
 841                        goto out;
 842                }
 843                mb_debug(1, "read bitmap for group %u\n", group);
 844        }
 845
 846        /* wait for I/O completion */
 847        for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
 848                if (bh[i] && ext4_wait_block_bitmap(sb, group, bh[i])) {
 849                        err = -EIO;
 850                        goto out;
 851                }
 852        }
 853
 854        first_block = page->index * blocks_per_page;
 855        for (i = 0; i < blocks_per_page; i++) {
 856                int group;
 857
 858                group = (first_block + i) >> 1;
 859                if (group >= ngroups)
 860                        break;
 861
 862                if (!bh[group - first_group])
 863                        /* skip initialized uptodate buddy */
 864                        continue;
 865
 866                /*
 867                 * data carry information regarding this
 868                 * particular group in the format specified
 869                 * above
 870                 *
 871                 */
 872                data = page_address(page) + (i * blocksize);
 873                bitmap = bh[group - first_group]->b_data;
 874
 875                /*
 876                 * We place the buddy block and bitmap block
 877                 * close together
 878                 */
 879                if ((first_block + i) & 1) {
 880                        /* this is block of buddy */
 881                        BUG_ON(incore == NULL);
 882                        mb_debug(1, "put buddy for group %u in page %lu/%x\n",
 883                                group, page->index, i * blocksize);
 884                        trace_ext4_mb_buddy_bitmap_load(sb, group);
 885                        grinfo = ext4_get_group_info(sb, group);
 886                        grinfo->bb_fragments = 0;
 887                        memset(grinfo->bb_counters, 0,
 888                               sizeof(*grinfo->bb_counters) *
 889                                (sb->s_blocksize_bits+2));
 890                        /*
 891                         * incore got set to the group block bitmap below
 892                         */
 893                        ext4_lock_group(sb, group);
 894                        /* init the buddy */
 895                        memset(data, 0xff, blocksize);
 896                        ext4_mb_generate_buddy(sb, data, incore, group);
 897                        ext4_unlock_group(sb, group);
 898                        incore = NULL;
 899                } else {
 900                        /* this is block of bitmap */
 901                        BUG_ON(incore != NULL);
 902                        mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
 903                                group, page->index, i * blocksize);
 904                        trace_ext4_mb_bitmap_load(sb, group);
 905
 906                        /* see comments in ext4_mb_put_pa() */
 907                        ext4_lock_group(sb, group);
 908                        memcpy(data, bitmap, blocksize);
 909
 910                        /* mark all preallocated blks used in in-core bitmap */
 911                        ext4_mb_generate_from_pa(sb, data, group);
 912                        ext4_mb_generate_from_freelist(sb, data, group);
 913                        ext4_unlock_group(sb, group);
 914
 915                        /* set incore so that the buddy information can be
 916                         * generated using this
 917                         */
 918                        incore = data;
 919                }
 920        }
 921        SetPageUptodate(page);
 922
 923out:
 924        if (bh) {
 925                for (i = 0; i < groups_per_page; i++)
 926                        brelse(bh[i]);
 927                if (bh != &bhs)
 928                        kfree(bh);
 929        }
 930        return err;
 931}
 932
 933/*
 934 * Lock the buddy and bitmap pages. This make sure other parallel init_group
 935 * on the same buddy page doesn't happen whild holding the buddy page lock.
 936 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
 937 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
 938 */
 939static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
 940                ext4_group_t group, struct ext4_buddy *e4b)
 941{
 942        struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
 943        int block, pnum, poff;
 944        int blocks_per_page;
 945        struct page *page;
 946
 947        e4b->bd_buddy_page = NULL;
 948        e4b->bd_bitmap_page = NULL;
 949
 950        blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
 951        /*
 952         * the buddy cache inode stores the block bitmap
 953         * and buddy information in consecutive blocks.
 954         * So for each group we need two blocks.
 955         */
 956        block = group * 2;
 957        pnum = block / blocks_per_page;
 958        poff = block % blocks_per_page;
 959        page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
 960        if (!page)
 961                return -EIO;
 962        BUG_ON(page->mapping != inode->i_mapping);
 963        e4b->bd_bitmap_page = page;
 964        e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
 965
 966        if (blocks_per_page >= 2) {
 967                /* buddy and bitmap are on the same page */
 968                return 0;
 969        }
 970
 971        block++;
 972        pnum = block / blocks_per_page;
 973        page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
 974        if (!page)
 975                return -EIO;
 976        BUG_ON(page->mapping != inode->i_mapping);
 977        e4b->bd_buddy_page = page;
 978        return 0;
 979}
 980
 981static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
 982{
 983        if (e4b->bd_bitmap_page) {
 984                unlock_page(e4b->bd_bitmap_page);
 985                page_cache_release(e4b->bd_bitmap_page);
 986        }
 987        if (e4b->bd_buddy_page) {
 988                unlock_page(e4b->bd_buddy_page);
 989                page_cache_release(e4b->bd_buddy_page);
 990        }
 991}
 992
 993/*
 994 * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
 995 * block group lock of all groups for this page; do not hold the BG lock when
 996 * calling this routine!
 997 */
 998static noinline_for_stack
 999int ext4_mb_init_group(struct super_block *sb, ext4_group_t group)
1000{
1001
1002        struct ext4_group_info *this_grp;
1003        struct ext4_buddy e4b;
1004        struct page *page;
1005        int ret = 0;
1006
1007        mb_debug(1, "init group %u\n", group);
1008        this_grp = ext4_get_group_info(sb, group);
1009        /*
1010         * This ensures that we don't reinit the buddy cache
1011         * page which map to the group from which we are already
1012         * allocating. If we are looking at the buddy cache we would
1013         * have taken a reference using ext4_mb_load_buddy and that
1014         * would have pinned buddy page to page cache.
1015         */
1016        ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b);
1017        if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1018                /*
1019                 * somebody initialized the group
1020                 * return without doing anything
1021                 */
1022                goto err;
1023        }
1024
1025        page = e4b.bd_bitmap_page;
1026        ret = ext4_mb_init_cache(page, NULL);
1027        if (ret)
1028                goto err;
1029        if (!PageUptodate(page)) {
1030                ret = -EIO;
1031                goto err;
1032        }
1033        mark_page_accessed(page);
1034
1035        if (e4b.bd_buddy_page == NULL) {
1036                /*
1037                 * If both the bitmap and buddy are in
1038                 * the same page we don't need to force
1039                 * init the buddy
1040                 */
1041                ret = 0;
1042                goto err;
1043        }
1044        /* init buddy cache */
1045        page = e4b.bd_buddy_page;
1046        ret = ext4_mb_init_cache(page, e4b.bd_bitmap);
1047        if (ret)
1048                goto err;
1049        if (!PageUptodate(page)) {
1050                ret = -EIO;
1051                goto err;
1052        }
1053        mark_page_accessed(page);
1054err:
1055        ext4_mb_put_buddy_page_lock(&e4b);
1056        return ret;
1057}
1058
1059/*
1060 * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1061 * block group lock of all groups for this page; do not hold the BG lock when
1062 * calling this routine!
1063 */
1064static noinline_for_stack int
1065ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1066                                        struct ext4_buddy *e4b)
1067{
1068        int blocks_per_page;
1069        int block;
1070        int pnum;
1071        int poff;
1072        struct page *page;
1073        int ret;
1074        struct ext4_group_info *grp;
1075        struct ext4_sb_info *sbi = EXT4_SB(sb);
1076        struct inode *inode = sbi->s_buddy_cache;
1077
1078        mb_debug(1, "load group %u\n", group);
1079
1080        blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1081        grp = ext4_get_group_info(sb, group);
1082
1083        e4b->bd_blkbits = sb->s_blocksize_bits;
1084        e4b->bd_info = grp;
1085        e4b->bd_sb = sb;
1086        e4b->bd_group = group;
1087        e4b->bd_buddy_page = NULL;
1088        e4b->bd_bitmap_page = NULL;
1089
1090        if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1091                /*
1092                 * we need full data about the group
1093                 * to make a good selection
1094                 */
1095                ret = ext4_mb_init_group(sb, group);
1096                if (ret)
1097                        return ret;
1098        }
1099
1100        /*
1101         * the buddy cache inode stores the block bitmap
1102         * and buddy information in consecutive blocks.
1103         * So for each group we need two blocks.
1104         */
1105        block = group * 2;
1106        pnum = block / blocks_per_page;
1107        poff = block % blocks_per_page;
1108
1109        /* we could use find_or_create_page(), but it locks page
1110         * what we'd like to avoid in fast path ... */
1111        page = find_get_page(inode->i_mapping, pnum);
1112        if (page == NULL || !PageUptodate(page)) {
1113                if (page)
1114                        /*
1115                         * drop the page reference and try
1116                         * to get the page with lock. If we
1117                         * are not uptodate that implies
1118                         * somebody just created the page but
1119                         * is yet to initialize the same. So
1120                         * wait for it to initialize.
1121                         */
1122                        page_cache_release(page);
1123                page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1124                if (page) {
1125                        BUG_ON(page->mapping != inode->i_mapping);
1126                        if (!PageUptodate(page)) {
1127                                ret = ext4_mb_init_cache(page, NULL);
1128                                if (ret) {
1129                                        unlock_page(page);
1130                                        goto err;
1131                                }
1132                                mb_cmp_bitmaps(e4b, page_address(page) +
1133                                               (poff * sb->s_blocksize));
1134                        }
1135                        unlock_page(page);
1136                }
1137        }
1138        if (page == NULL || !PageUptodate(page)) {
1139                ret = -EIO;
1140                goto err;
1141        }
1142        e4b->bd_bitmap_page = page;
1143        e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1144        mark_page_accessed(page);
1145
1146        block++;
1147        pnum = block / blocks_per_page;
1148        poff = block % blocks_per_page;
1149
1150        page = find_get_page(inode->i_mapping, pnum);
1151        if (page == NULL || !PageUptodate(page)) {
1152                if (page)
1153                        page_cache_release(page);
1154                page = find_or_create_page(inode->i_mapping, pnum, GFP_NOFS);
1155                if (page) {
1156                        BUG_ON(page->mapping != inode->i_mapping);
1157                        if (!PageUptodate(page)) {
1158                                ret = ext4_mb_init_cache(page, e4b->bd_bitmap);
1159                                if (ret) {
1160                                        unlock_page(page);
1161                                        goto err;
1162                                }
1163                        }
1164                        unlock_page(page);
1165                }
1166        }
1167        if (page == NULL || !PageUptodate(page)) {
1168                ret = -EIO;
1169                goto err;
1170        }
1171        e4b->bd_buddy_page = page;
1172        e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1173        mark_page_accessed(page);
1174
1175        BUG_ON(e4b->bd_bitmap_page == NULL);
1176        BUG_ON(e4b->bd_buddy_page == NULL);
1177
1178        return 0;
1179
1180err:
1181        if (page)
1182                page_cache_release(page);
1183        if (e4b->bd_bitmap_page)
1184                page_cache_release(e4b->bd_bitmap_page);
1185        if (e4b->bd_buddy_page)
1186                page_cache_release(e4b->bd_buddy_page);
1187        e4b->bd_buddy = NULL;
1188        e4b->bd_bitmap = NULL;
1189        return ret;
1190}
1191
1192static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1193{
1194        if (e4b->bd_bitmap_page)
1195                page_cache_release(e4b->bd_bitmap_page);
1196        if (e4b->bd_buddy_page)
1197                page_cache_release(e4b->bd_buddy_page);
1198}
1199
1200
1201static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1202{
1203        int order = 1;
1204        void *bb;
1205
1206        BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1207        BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1208
1209        bb = e4b->bd_buddy;
1210        while (order <= e4b->bd_blkbits + 1) {
1211                block = block >> 1;
1212                if (!mb_test_bit(block, bb)) {
1213                        /* this block is part of buddy of order 'order' */
1214                        return order;
1215                }
1216                bb += 1 << (e4b->bd_blkbits - order);
1217                order++;
1218        }
1219        return 0;
1220}
1221
1222static void mb_clear_bits(void *bm, int cur, int len)
1223{
1224        __u32 *addr;
1225
1226        len = cur + len;
1227        while (cur < len) {
1228                if ((cur & 31) == 0 && (len - cur) >= 32) {
1229                        /* fast path: clear whole word at once */
1230                        addr = bm + (cur >> 3);
1231                        *addr = 0;
1232                        cur += 32;
1233                        continue;
1234                }
1235                mb_clear_bit(cur, bm);
1236                cur++;
1237        }
1238}
1239
1240void ext4_set_bits(void *bm, int cur, int len)
1241{
1242        __u32 *addr;
1243
1244        len = cur + len;
1245        while (cur < len) {
1246                if ((cur & 31) == 0 && (len - cur) >= 32) {
1247                        /* fast path: set whole word at once */
1248                        addr = bm + (cur >> 3);
1249                        *addr = 0xffffffff;
1250                        cur += 32;
1251                        continue;
1252                }
1253                mb_set_bit(cur, bm);
1254                cur++;
1255        }
1256}
1257
1258static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1259                          int first, int count)
1260{
1261        int block = 0;
1262        int max = 0;
1263        int order;
1264        void *buddy;
1265        void *buddy2;
1266        struct super_block *sb = e4b->bd_sb;
1267
1268        BUG_ON(first + count > (sb->s_blocksize << 3));
1269        assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1270        mb_check_buddy(e4b);
1271        mb_free_blocks_double(inode, e4b, first, count);
1272
1273        e4b->bd_info->bb_free += count;
1274        if (first < e4b->bd_info->bb_first_free)
1275                e4b->bd_info->bb_first_free = first;
1276
1277        /* let's maintain fragments counter */
1278        if (first != 0)
1279                block = !mb_test_bit(first - 1, e4b->bd_bitmap);
1280        if (first + count < EXT4_SB(sb)->s_mb_maxs[0])
1281                max = !mb_test_bit(first + count, e4b->bd_bitmap);
1282        if (block && max)
1283                e4b->bd_info->bb_fragments--;
1284        else if (!block && !max)
1285                e4b->bd_info->bb_fragments++;
1286
1287        /* let's maintain buddy itself */
1288        while (count-- > 0) {
1289                block = first++;
1290                order = 0;
1291
1292                if (!mb_test_bit(block, e4b->bd_bitmap)) {
1293                        ext4_fsblk_t blocknr;
1294
1295                        blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1296                        blocknr += EXT4_C2B(EXT4_SB(sb), block);
1297                        ext4_grp_locked_error(sb, e4b->bd_group,
1298                                              inode ? inode->i_ino : 0,
1299                                              blocknr,
1300                                              "freeing already freed block "
1301                                              "(bit %u)", block);
1302                }
1303                mb_clear_bit(block, e4b->bd_bitmap);
1304                e4b->bd_info->bb_counters[order]++;
1305
1306                /* start of the buddy */
1307                buddy = mb_find_buddy(e4b, order, &max);
1308
1309                do {
1310                        block &= ~1UL;
1311                        if (mb_test_bit(block, buddy) ||
1312                                        mb_test_bit(block + 1, buddy))
1313                                break;
1314
1315                        /* both the buddies are free, try to coalesce them */
1316                        buddy2 = mb_find_buddy(e4b, order + 1, &max);
1317
1318                        if (!buddy2)
1319                                break;
1320
1321                        if (order > 0) {
1322                                /* for special purposes, we don't set
1323                                 * free bits in bitmap */
1324                                mb_set_bit(block, buddy);
1325                                mb_set_bit(block + 1, buddy);
1326                        }
1327                        e4b->bd_info->bb_counters[order]--;
1328                        e4b->bd_info->bb_counters[order]--;
1329
1330                        block = block >> 1;
1331                        order++;
1332                        e4b->bd_info->bb_counters[order]++;
1333
1334                        mb_clear_bit(block, buddy2);
1335                        buddy = buddy2;
1336                } while (1);
1337        }
1338        mb_set_largest_free_order(sb, e4b->bd_info);
1339        mb_check_buddy(e4b);
1340}
1341
1342static int mb_find_extent(struct ext4_buddy *e4b, int block,
1343                                int needed, struct ext4_free_extent *ex)
1344{
1345        int next = block;
1346        int max, order;
1347        void *buddy;
1348
1349        assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1350        BUG_ON(ex == NULL);
1351
1352        buddy = mb_find_buddy(e4b, 0, &max);
1353        BUG_ON(buddy == NULL);
1354        BUG_ON(block >= max);
1355        if (mb_test_bit(block, buddy)) {
1356                ex->fe_len = 0;
1357                ex->fe_start = 0;
1358                ex->fe_group = 0;
1359                return 0;
1360        }
1361
1362        /* find actual order */
1363        order = mb_find_order_for_block(e4b, block);
1364        block = block >> order;
1365
1366        ex->fe_len = 1 << order;
1367        ex->fe_start = block << order;
1368        ex->fe_group = e4b->bd_group;
1369
1370        /* calc difference from given start */
1371        next = next - ex->fe_start;
1372        ex->fe_len -= next;
1373        ex->fe_start += next;
1374
1375        while (needed > ex->fe_len &&
1376               (buddy = mb_find_buddy(e4b, order, &max))) {
1377
1378                if (block + 1 >= max)
1379                        break;
1380
1381                next = (block + 1) * (1 << order);
1382                if (mb_test_bit(next, e4b->bd_bitmap))
1383                        break;
1384
1385                order = mb_find_order_for_block(e4b, next);
1386
1387                block = next >> order;
1388                ex->fe_len += 1 << order;
1389        }
1390
1391        BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1392        return ex->fe_len;
1393}
1394
1395static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1396{
1397        int ord;
1398        int mlen = 0;
1399        int max = 0;
1400        int cur;
1401        int start = ex->fe_start;
1402        int len = ex->fe_len;
1403        unsigned ret = 0;
1404        int len0 = len;
1405        void *buddy;
1406
1407        BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1408        BUG_ON(e4b->bd_group != ex->fe_group);
1409        assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1410        mb_check_buddy(e4b);
1411        mb_mark_used_double(e4b, start, len);
1412
1413        e4b->bd_info->bb_free -= len;
1414        if (e4b->bd_info->bb_first_free == start)
1415                e4b->bd_info->bb_first_free += len;
1416
1417        /* let's maintain fragments counter */
1418        if (start != 0)
1419                mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1420        if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1421                max = !mb_test_bit(start + len, e4b->bd_bitmap);
1422        if (mlen && max)
1423                e4b->bd_info->bb_fragments++;
1424        else if (!mlen && !max)
1425                e4b->bd_info->bb_fragments--;
1426
1427        /* let's maintain buddy itself */
1428        while (len) {
1429                ord = mb_find_order_for_block(e4b, start);
1430
1431                if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1432                        /* the whole chunk may be allocated at once! */
1433                        mlen = 1 << ord;
1434                        buddy = mb_find_buddy(e4b, ord, &max);
1435                        BUG_ON((start >> ord) >= max);
1436                        mb_set_bit(start >> ord, buddy);
1437                        e4b->bd_info->bb_counters[ord]--;
1438                        start += mlen;
1439                        len -= mlen;
1440                        BUG_ON(len < 0);
1441                        continue;
1442                }
1443
1444                /* store for history */
1445                if (ret == 0)
1446                        ret = len | (ord << 16);
1447
1448                /* we have to split large buddy */
1449                BUG_ON(ord <= 0);
1450                buddy = mb_find_buddy(e4b, ord, &max);
1451                mb_set_bit(start >> ord, buddy);
1452                e4b->bd_info->bb_counters[ord]--;
1453
1454                ord--;
1455                cur = (start >> ord) & ~1U;
1456                buddy = mb_find_buddy(e4b, ord, &max);
1457                mb_clear_bit(cur, buddy);
1458                mb_clear_bit(cur + 1, buddy);
1459                e4b->bd_info->bb_counters[ord]++;
1460                e4b->bd_info->bb_counters[ord]++;
1461        }
1462        mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1463
1464        ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1465        mb_check_buddy(e4b);
1466
1467        return ret;
1468}
1469
1470/*
1471 * Must be called under group lock!
1472 */
1473static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1474                                        struct ext4_buddy *e4b)
1475{
1476        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1477        int ret;
1478
1479        BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1480        BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1481
1482        ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1483        ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1484        ret = mb_mark_used(e4b, &ac->ac_b_ex);
1485
1486        /* preallocation can change ac_b_ex, thus we store actually
1487         * allocated blocks for history */
1488        ac->ac_f_ex = ac->ac_b_ex;
1489
1490        ac->ac_status = AC_STATUS_FOUND;
1491        ac->ac_tail = ret & 0xffff;
1492        ac->ac_buddy = ret >> 16;
1493
1494        /*
1495         * take the page reference. We want the page to be pinned
1496         * so that we don't get a ext4_mb_init_cache_call for this
1497         * group until we update the bitmap. That would mean we
1498         * double allocate blocks. The reference is dropped
1499         * in ext4_mb_release_context
1500         */
1501        ac->ac_bitmap_page = e4b->bd_bitmap_page;
1502        get_page(ac->ac_bitmap_page);
1503        ac->ac_buddy_page = e4b->bd_buddy_page;
1504        get_page(ac->ac_buddy_page);
1505        /* store last allocated for subsequent stream allocation */
1506        if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1507                spin_lock(&sbi->s_md_lock);
1508                sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1509                sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1510                spin_unlock(&sbi->s_md_lock);
1511        }
1512}
1513
1514/*
1515 * regular allocator, for general purposes allocation
1516 */
1517
1518static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1519                                        struct ext4_buddy *e4b,
1520                                        int finish_group)
1521{
1522        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1523        struct ext4_free_extent *bex = &ac->ac_b_ex;
1524        struct ext4_free_extent *gex = &ac->ac_g_ex;
1525        struct ext4_free_extent ex;
1526        int max;
1527
1528        if (ac->ac_status == AC_STATUS_FOUND)
1529                return;
1530        /*
1531         * We don't want to scan for a whole year
1532         */
1533        if (ac->ac_found > sbi->s_mb_max_to_scan &&
1534                        !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1535                ac->ac_status = AC_STATUS_BREAK;
1536                return;
1537        }
1538
1539        /*
1540         * Haven't found good chunk so far, let's continue
1541         */
1542        if (bex->fe_len < gex->fe_len)
1543                return;
1544
1545        if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1546                        && bex->fe_group == e4b->bd_group) {
1547                /* recheck chunk's availability - we don't know
1548                 * when it was found (within this lock-unlock
1549                 * period or not) */
1550                max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1551                if (max >= gex->fe_len) {
1552                        ext4_mb_use_best_found(ac, e4b);
1553                        return;
1554                }
1555        }
1556}
1557
1558/*
1559 * The routine checks whether found extent is good enough. If it is,
1560 * then the extent gets marked used and flag is set to the context
1561 * to stop scanning. Otherwise, the extent is compared with the
1562 * previous found extent and if new one is better, then it's stored
1563 * in the context. Later, the best found extent will be used, if
1564 * mballoc can't find good enough extent.
1565 *
1566 * FIXME: real allocation policy is to be designed yet!
1567 */
1568static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1569                                        struct ext4_free_extent *ex,
1570                                        struct ext4_buddy *e4b)
1571{
1572        struct ext4_free_extent *bex = &ac->ac_b_ex;
1573        struct ext4_free_extent *gex = &ac->ac_g_ex;
1574
1575        BUG_ON(ex->fe_len <= 0);
1576        BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1577        BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1578        BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1579
1580        ac->ac_found++;
1581
1582        /*
1583         * The special case - take what you catch first
1584         */
1585        if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1586                *bex = *ex;
1587                ext4_mb_use_best_found(ac, e4b);
1588                return;
1589        }
1590
1591        /*
1592         * Let's check whether the chuck is good enough
1593         */
1594        if (ex->fe_len == gex->fe_len) {
1595                *bex = *ex;
1596                ext4_mb_use_best_found(ac, e4b);
1597                return;
1598        }
1599
1600        /*
1601         * If this is first found extent, just store it in the context
1602         */
1603        if (bex->fe_len == 0) {
1604                *bex = *ex;
1605                return;
1606        }
1607
1608        /*
1609         * If new found extent is better, store it in the context
1610         */
1611        if (bex->fe_len < gex->fe_len) {
1612                /* if the request isn't satisfied, any found extent
1613                 * larger than previous best one is better */
1614                if (ex->fe_len > bex->fe_len)
1615                        *bex = *ex;
1616        } else if (ex->fe_len > gex->fe_len) {
1617                /* if the request is satisfied, then we try to find
1618                 * an extent that still satisfy the request, but is
1619                 * smaller than previous one */
1620                if (ex->fe_len < bex->fe_len)
1621                        *bex = *ex;
1622        }
1623
1624        ext4_mb_check_limits(ac, e4b, 0);
1625}
1626
1627static noinline_for_stack
1628int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1629                                        struct ext4_buddy *e4b)
1630{
1631        struct ext4_free_extent ex = ac->ac_b_ex;
1632        ext4_group_t group = ex.fe_group;
1633        int max;
1634        int err;
1635
1636        BUG_ON(ex.fe_len <= 0);
1637        err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1638        if (err)
1639                return err;
1640
1641        ext4_lock_group(ac->ac_sb, group);
1642        max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1643
1644        if (max > 0) {
1645                ac->ac_b_ex = ex;
1646                ext4_mb_use_best_found(ac, e4b);
1647        }
1648
1649        ext4_unlock_group(ac->ac_sb, group);
1650        ext4_mb_unload_buddy(e4b);
1651
1652        return 0;
1653}
1654
1655static noinline_for_stack
1656int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1657                                struct ext4_buddy *e4b)
1658{
1659        ext4_group_t group = ac->ac_g_ex.fe_group;
1660        int max;
1661        int err;
1662        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1663        struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1664        struct ext4_free_extent ex;
1665
1666        if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1667                return 0;
1668        if (grp->bb_free == 0)
1669                return 0;
1670
1671        err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1672        if (err)
1673                return err;
1674
1675        ext4_lock_group(ac->ac_sb, group);
1676        max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1677                             ac->ac_g_ex.fe_len, &ex);
1678
1679        if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1680                ext4_fsblk_t start;
1681
1682                start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1683                        ex.fe_start;
1684                /* use do_div to get remainder (would be 64-bit modulo) */
1685                if (do_div(start, sbi->s_stripe) == 0) {
1686                        ac->ac_found++;
1687                        ac->ac_b_ex = ex;
1688                        ext4_mb_use_best_found(ac, e4b);
1689                }
1690        } else if (max >= ac->ac_g_ex.fe_len) {
1691                BUG_ON(ex.fe_len <= 0);
1692                BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1693                BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1694                ac->ac_found++;
1695                ac->ac_b_ex = ex;
1696                ext4_mb_use_best_found(ac, e4b);
1697        } else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1698                /* Sometimes, caller may want to merge even small
1699                 * number of blocks to an existing extent */
1700                BUG_ON(ex.fe_len <= 0);
1701                BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1702                BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1703                ac->ac_found++;
1704                ac->ac_b_ex = ex;
1705                ext4_mb_use_best_found(ac, e4b);
1706        }
1707        ext4_unlock_group(ac->ac_sb, group);
1708        ext4_mb_unload_buddy(e4b);
1709
1710        return 0;
1711}
1712
1713/*
1714 * The routine scans buddy structures (not bitmap!) from given order
1715 * to max order and tries to find big enough chunk to satisfy the req
1716 */
1717static noinline_for_stack
1718void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1719                                        struct ext4_buddy *e4b)
1720{
1721        struct super_block *sb = ac->ac_sb;
1722        struct ext4_group_info *grp = e4b->bd_info;
1723        void *buddy;
1724        int i;
1725        int k;
1726        int max;
1727
1728        BUG_ON(ac->ac_2order <= 0);
1729        for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1730                if (grp->bb_counters[i] == 0)
1731                        continue;
1732
1733                buddy = mb_find_buddy(e4b, i, &max);
1734                BUG_ON(buddy == NULL);
1735
1736                k = mb_find_next_zero_bit(buddy, max, 0);
1737                BUG_ON(k >= max);
1738
1739                ac->ac_found++;
1740
1741                ac->ac_b_ex.fe_len = 1 << i;
1742                ac->ac_b_ex.fe_start = k << i;
1743                ac->ac_b_ex.fe_group = e4b->bd_group;
1744
1745                ext4_mb_use_best_found(ac, e4b);
1746
1747                BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1748
1749                if (EXT4_SB(sb)->s_mb_stats)
1750                        atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1751
1752                break;
1753        }
1754}
1755
1756/*
1757 * The routine scans the group and measures all found extents.
1758 * In order to optimize scanning, caller must pass number of
1759 * free blocks in the group, so the routine can know upper limit.
1760 */
1761static noinline_for_stack
1762void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1763                                        struct ext4_buddy *e4b)
1764{
1765        struct super_block *sb = ac->ac_sb;
1766        void *bitmap = e4b->bd_bitmap;
1767        struct ext4_free_extent ex;
1768        int i;
1769        int free;
1770
1771        free = e4b->bd_info->bb_free;
1772        BUG_ON(free <= 0);
1773
1774        i = e4b->bd_info->bb_first_free;
1775
1776        while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1777                i = mb_find_next_zero_bit(bitmap,
1778                                                EXT4_CLUSTERS_PER_GROUP(sb), i);
1779                if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1780                        /*
1781                         * IF we have corrupt bitmap, we won't find any
1782                         * free blocks even though group info says we
1783                         * we have free blocks
1784                         */
1785                        ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1786                                        "%d free clusters as per "
1787                                        "group info. But bitmap says 0",
1788                                        free);
1789                        break;
1790                }
1791
1792                mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1793                BUG_ON(ex.fe_len <= 0);
1794                if (free < ex.fe_len) {
1795                        ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1796                                        "%d free clusters as per "
1797                                        "group info. But got %d blocks",
1798                                        free, ex.fe_len);
1799                        /*
1800                         * The number of free blocks differs. This mostly
1801                         * indicate that the bitmap is corrupt. So exit
1802                         * without claiming the space.
1803                         */
1804                        break;
1805                }
1806
1807                ext4_mb_measure_extent(ac, &ex, e4b);
1808
1809                i += ex.fe_len;
1810                free -= ex.fe_len;
1811        }
1812
1813        ext4_mb_check_limits(ac, e4b, 1);
1814}
1815
1816/*
1817 * This is a special case for storages like raid5
1818 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1819 */
1820static noinline_for_stack
1821void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1822                                 struct ext4_buddy *e4b)
1823{
1824        struct super_block *sb = ac->ac_sb;
1825        struct ext4_sb_info *sbi = EXT4_SB(sb);
1826        void *bitmap = e4b->bd_bitmap;
1827        struct ext4_free_extent ex;
1828        ext4_fsblk_t first_group_block;
1829        ext4_fsblk_t a;
1830        ext4_grpblk_t i;
1831        int max;
1832
1833        BUG_ON(sbi->s_stripe == 0);
1834
1835        /* find first stripe-aligned block in group */
1836        first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
1837
1838        a = first_group_block + sbi->s_stripe - 1;
1839        do_div(a, sbi->s_stripe);
1840        i = (a * sbi->s_stripe) - first_group_block;
1841
1842        while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
1843                if (!mb_test_bit(i, bitmap)) {
1844                        max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
1845                        if (max >= sbi->s_stripe) {
1846                                ac->ac_found++;
1847                                ac->ac_b_ex = ex;
1848                                ext4_mb_use_best_found(ac, e4b);
1849                                break;
1850                        }
1851                }
1852                i += sbi->s_stripe;
1853        }
1854}
1855
1856/* This is now called BEFORE we load the buddy bitmap. */
1857static int ext4_mb_good_group(struct ext4_allocation_context *ac,
1858                                ext4_group_t group, int cr)
1859{
1860        unsigned free, fragments;
1861        int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
1862        struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1863
1864        BUG_ON(cr < 0 || cr >= 4);
1865
1866        free = grp->bb_free;
1867        if (free == 0)
1868                return 0;
1869        if (cr <= 2 && free < ac->ac_g_ex.fe_len)
1870                return 0;
1871
1872        /* We only do this if the grp has never been initialized */
1873        if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1874                int ret = ext4_mb_init_group(ac->ac_sb, group);
1875                if (ret)
1876                        return 0;
1877        }
1878
1879        fragments = grp->bb_fragments;
1880        if (fragments == 0)
1881                return 0;
1882
1883        switch (cr) {
1884        case 0:
1885                BUG_ON(ac->ac_2order == 0);
1886
1887                if (grp->bb_largest_free_order < ac->ac_2order)
1888                        return 0;
1889
1890                /* Avoid using the first bg of a flexgroup for data files */
1891                if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
1892                    (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
1893                    ((group % flex_size) == 0))
1894                        return 0;
1895
1896                return 1;
1897        case 1:
1898                if ((free / fragments) >= ac->ac_g_ex.fe_len)
1899                        return 1;
1900                break;
1901        case 2:
1902                if (free >= ac->ac_g_ex.fe_len)
1903                        return 1;
1904                break;
1905        case 3:
1906                return 1;
1907        default:
1908                BUG();
1909        }
1910
1911        return 0;
1912}
1913
1914static noinline_for_stack int
1915ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
1916{
1917        ext4_group_t ngroups, group, i;
1918        int cr;
1919        int err = 0;
1920        struct ext4_sb_info *sbi;
1921        struct super_block *sb;
1922        struct ext4_buddy e4b;
1923
1924        sb = ac->ac_sb;
1925        sbi = EXT4_SB(sb);
1926        ngroups = ext4_get_groups_count(sb);
1927        /* non-extent files are limited to low blocks/groups */
1928        if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
1929                ngroups = sbi->s_blockfile_groups;
1930
1931        BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1932
1933        /* first, try the goal */
1934        err = ext4_mb_find_by_goal(ac, &e4b);
1935        if (err || ac->ac_status == AC_STATUS_FOUND)
1936                goto out;
1937
1938        if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
1939                goto out;
1940
1941        /*
1942         * ac->ac2_order is set only if the fe_len is a power of 2
1943         * if ac2_order is set we also set criteria to 0 so that we
1944         * try exact allocation using buddy.
1945         */
1946        i = fls(ac->ac_g_ex.fe_len);
1947        ac->ac_2order = 0;
1948        /*
1949         * We search using buddy data only if the order of the request
1950         * is greater than equal to the sbi_s_mb_order2_reqs
1951         * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
1952         */
1953        if (i >= sbi->s_mb_order2_reqs) {
1954                /*
1955                 * This should tell if fe_len is exactly power of 2
1956                 */
1957                if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
1958                        ac->ac_2order = i - 1;
1959        }
1960
1961        /* if stream allocation is enabled, use global goal */
1962        if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1963                /* TBD: may be hot point */
1964                spin_lock(&sbi->s_md_lock);
1965                ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
1966                ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
1967                spin_unlock(&sbi->s_md_lock);
1968        }
1969
1970        /* Let's just scan groups to find more-less suitable blocks */
1971        cr = ac->ac_2order ? 0 : 1;
1972        /*
1973         * cr == 0 try to get exact allocation,
1974         * cr == 3  try to get anything
1975         */
1976repeat:
1977        for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
1978                ac->ac_criteria = cr;
1979                /*
1980                 * searching for the right group start
1981                 * from the goal value specified
1982                 */
1983                group = ac->ac_g_ex.fe_group;
1984
1985                for (i = 0; i < ngroups; group++, i++) {
1986                        if (group == ngroups)
1987                                group = 0;
1988
1989                        /* This now checks without needing the buddy page */
1990                        if (!ext4_mb_good_group(ac, group, cr))
1991                                continue;
1992
1993                        err = ext4_mb_load_buddy(sb, group, &e4b);
1994                        if (err)
1995                                goto out;
1996
1997                        ext4_lock_group(sb, group);
1998
1999                        /*
2000                         * We need to check again after locking the
2001                         * block group
2002                         */
2003                        if (!ext4_mb_good_group(ac, group, cr)) {
2004                                ext4_unlock_group(sb, group);
2005                                ext4_mb_unload_buddy(&e4b);
2006                                continue;
2007                        }
2008
2009                        ac->ac_groups_scanned++;
2010                        if (cr == 0)
2011                                ext4_mb_simple_scan_group(ac, &e4b);
2012                        else if (cr == 1 && sbi->s_stripe &&
2013                                        !(ac->ac_g_ex.fe_len % sbi->s_stripe))
2014                                ext4_mb_scan_aligned(ac, &e4b);
2015                        else
2016                                ext4_mb_complex_scan_group(ac, &e4b);
2017
2018                        ext4_unlock_group(sb, group);
2019                        ext4_mb_unload_buddy(&e4b);
2020
2021                        if (ac->ac_status != AC_STATUS_CONTINUE)
2022                                break;
2023                }
2024        }
2025
2026        if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2027            !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2028                /*
2029                 * We've been searching too long. Let's try to allocate
2030                 * the best chunk we've found so far
2031                 */
2032
2033                ext4_mb_try_best_found(ac, &e4b);
2034                if (ac->ac_status != AC_STATUS_FOUND) {
2035                        /*
2036                         * Someone more lucky has already allocated it.
2037                         * The only thing we can do is just take first
2038                         * found block(s)
2039                        printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2040                         */
2041                        ac->ac_b_ex.fe_group = 0;
2042                        ac->ac_b_ex.fe_start = 0;
2043                        ac->ac_b_ex.fe_len = 0;
2044                        ac->ac_status = AC_STATUS_CONTINUE;
2045                        ac->ac_flags |= EXT4_MB_HINT_FIRST;
2046                        cr = 3;
2047                        atomic_inc(&sbi->s_mb_lost_chunks);
2048                        goto repeat;
2049                }
2050        }
2051out:
2052        return err;
2053}
2054
2055static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2056{
2057        struct super_block *sb = seq->private;
2058        ext4_group_t group;
2059
2060        if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2061                return NULL;
2062        group = *pos + 1;
2063        return (void *) ((unsigned long) group);
2064}
2065
2066static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2067{
2068        struct super_block *sb = seq->private;
2069        ext4_group_t group;
2070
2071        ++*pos;
2072        if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2073                return NULL;
2074        group = *pos + 1;
2075        return (void *) ((unsigned long) group);
2076}
2077
2078static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2079{
2080        struct super_block *sb = seq->private;
2081        ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2082        int i;
2083        int err, buddy_loaded = 0;
2084        struct ext4_buddy e4b;
2085        struct ext4_group_info *grinfo;
2086        struct sg {
2087                struct ext4_group_info info;
2088                ext4_grpblk_t counters[16];
2089        } sg;
2090
2091        group--;
2092        if (group == 0)
2093                seq_printf(seq, "#%-5s: %-5s %-5s %-5s "
2094                                "[ %-5s %-5s %-5s %-5s %-5s %-5s %-5s "
2095                                  "%-5s %-5s %-5s %-5s %-5s %-5s %-5s ]\n",
2096                           "group", "free", "frags", "first",
2097                           "2^0", "2^1", "2^2", "2^3", "2^4", "2^5", "2^6",
2098                           "2^7", "2^8", "2^9", "2^10", "2^11", "2^12", "2^13");
2099
2100        i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2101                sizeof(struct ext4_group_info);
2102        grinfo = ext4_get_group_info(sb, group);
2103        /* Load the group info in memory only if not already loaded. */
2104        if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2105                err = ext4_mb_load_buddy(sb, group, &e4b);
2106                if (err) {
2107                        seq_printf(seq, "#%-5u: I/O error\n", group);
2108                        return 0;
2109                }
2110                buddy_loaded = 1;
2111        }
2112
2113        memcpy(&sg, ext4_get_group_info(sb, group), i);
2114
2115        if (buddy_loaded)
2116                ext4_mb_unload_buddy(&e4b);
2117
2118        seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2119                        sg.info.bb_fragments, sg.info.bb_first_free);
2120        for (i = 0; i <= 13; i++)
2121                seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2122                                sg.info.bb_counters[i] : 0);
2123        seq_printf(seq, " ]\n");
2124
2125        return 0;
2126}
2127
2128static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2129{
2130}
2131
2132static const struct seq_operations ext4_mb_seq_groups_ops = {
2133        .start  = ext4_mb_seq_groups_start,
2134        .next   = ext4_mb_seq_groups_next,
2135        .stop   = ext4_mb_seq_groups_stop,
2136        .show   = ext4_mb_seq_groups_show,
2137};
2138
2139static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2140{
2141        struct super_block *sb = PDE(inode)->data;
2142        int rc;
2143
2144        rc = seq_open(file, &ext4_mb_seq_groups_ops);
2145        if (rc == 0) {
2146                struct seq_file *m = file->private_data;
2147                m->private = sb;
2148        }
2149        return rc;
2150
2151}
2152
2153static const struct file_operations ext4_mb_seq_groups_fops = {
2154        .owner          = THIS_MODULE,
2155        .open           = ext4_mb_seq_groups_open,
2156        .read           = seq_read,
2157        .llseek         = seq_lseek,
2158        .release        = seq_release,
2159};
2160
2161static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2162{
2163        int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2164        struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2165
2166        BUG_ON(!cachep);
2167        return cachep;
2168}
2169
2170/*
2171 * Allocate the top-level s_group_info array for the specified number
2172 * of groups
2173 */
2174int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2175{
2176        struct ext4_sb_info *sbi = EXT4_SB(sb);
2177        unsigned size;
2178        struct ext4_group_info ***new_groupinfo;
2179
2180        size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2181                EXT4_DESC_PER_BLOCK_BITS(sb);
2182        if (size <= sbi->s_group_info_size)
2183                return 0;
2184
2185        size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2186        new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2187        if (!new_groupinfo) {
2188                ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2189                return -ENOMEM;
2190        }
2191        if (sbi->s_group_info) {
2192                memcpy(new_groupinfo, sbi->s_group_info,
2193                       sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2194                ext4_kvfree(sbi->s_group_info);
2195        }
2196        sbi->s_group_info = new_groupinfo;
2197        sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2198        ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 
2199                   sbi->s_group_info_size);
2200        return 0;
2201}
2202
2203/* Create and initialize ext4_group_info data for the given group. */
2204int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2205                          struct ext4_group_desc *desc)
2206{
2207        int i;
2208        int metalen = 0;
2209        struct ext4_sb_info *sbi = EXT4_SB(sb);
2210        struct ext4_group_info **meta_group_info;
2211        struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2212
2213        /*
2214         * First check if this group is the first of a reserved block.
2215         * If it's true, we have to allocate a new table of pointers
2216         * to ext4_group_info structures
2217         */
2218        if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2219                metalen = sizeof(*meta_group_info) <<
2220                        EXT4_DESC_PER_BLOCK_BITS(sb);
2221                meta_group_info = kmalloc(metalen, GFP_KERNEL);
2222                if (meta_group_info == NULL) {
2223                        ext4_msg(sb, KERN_ERR, "can't allocate mem "
2224                                 "for a buddy group");
2225                        goto exit_meta_group_info;
2226                }
2227                sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2228                        meta_group_info;
2229        }
2230
2231        meta_group_info =
2232                sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2233        i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2234
2235        meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_KERNEL);
2236        if (meta_group_info[i] == NULL) {
2237                ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2238                goto exit_group_info;
2239        }
2240        set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2241                &(meta_group_info[i]->bb_state));
2242
2243        /*
2244         * initialize bb_free to be able to skip
2245         * empty groups without initialization
2246         */
2247        if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2248                meta_group_info[i]->bb_free =
2249                        ext4_free_clusters_after_init(sb, group, desc);
2250        } else {
2251                meta_group_info[i]->bb_free =
2252                        ext4_free_group_clusters(sb, desc);
2253        }
2254
2255        INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2256        init_rwsem(&meta_group_info[i]->alloc_sem);
2257        meta_group_info[i]->bb_free_root = RB_ROOT;
2258        meta_group_info[i]->bb_largest_free_order = -1;  /* uninit */
2259
2260#ifdef DOUBLE_CHECK
2261        {
2262                struct buffer_head *bh;
2263                meta_group_info[i]->bb_bitmap =
2264                        kmalloc(sb->s_blocksize, GFP_KERNEL);
2265                BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2266                bh = ext4_read_block_bitmap(sb, group);
2267                BUG_ON(bh == NULL);
2268                memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2269                        sb->s_blocksize);
2270                put_bh(bh);
2271        }
2272#endif
2273
2274        return 0;
2275
2276exit_group_info:
2277        /* If a meta_group_info table has been allocated, release it now */
2278        if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2279                kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2280                sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2281        }
2282exit_meta_group_info:
2283        return -ENOMEM;
2284} /* ext4_mb_add_groupinfo */
2285
2286static int ext4_mb_init_backend(struct super_block *sb)
2287{
2288        ext4_group_t ngroups = ext4_get_groups_count(sb);
2289        ext4_group_t i;
2290        struct ext4_sb_info *sbi = EXT4_SB(sb);
2291        int err;
2292        struct ext4_group_desc *desc;
2293        struct kmem_cache *cachep;
2294
2295        err = ext4_mb_alloc_groupinfo(sb, ngroups);
2296        if (err)
2297                return err;
2298
2299        sbi->s_buddy_cache = new_inode(sb);
2300        if (sbi->s_buddy_cache == NULL) {
2301                ext4_msg(sb, KERN_ERR, "can't get new inode");
2302                goto err_freesgi;
2303        }
2304        /* To avoid potentially colliding with an valid on-disk inode number,
2305         * use EXT4_BAD_INO for the buddy cache inode number.  This inode is
2306         * not in the inode hash, so it should never be found by iget(), but
2307         * this will avoid confusion if it ever shows up during debugging. */
2308        sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2309        EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2310        for (i = 0; i < ngroups; i++) {
2311                desc = ext4_get_group_desc(sb, i, NULL);
2312                if (desc == NULL) {
2313                        ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2314                        goto err_freebuddy;
2315                }
2316                if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2317                        goto err_freebuddy;
2318        }
2319
2320        return 0;
2321
2322err_freebuddy:
2323        cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2324        while (i-- > 0)
2325                kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2326        i = sbi->s_group_info_size;
2327        while (i-- > 0)
2328                kfree(sbi->s_group_info[i]);
2329        iput(sbi->s_buddy_cache);
2330err_freesgi:
2331        ext4_kvfree(sbi->s_group_info);
2332        return -ENOMEM;
2333}
2334
2335static void ext4_groupinfo_destroy_slabs(void)
2336{
2337        int i;
2338
2339        for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2340                if (ext4_groupinfo_caches[i])
2341                        kmem_cache_destroy(ext4_groupinfo_caches[i]);
2342                ext4_groupinfo_caches[i] = NULL;
2343        }
2344}
2345
2346static int ext4_groupinfo_create_slab(size_t size)
2347{
2348        static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2349        int slab_size;
2350        int blocksize_bits = order_base_2(size);
2351        int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2352        struct kmem_cache *cachep;
2353
2354        if (cache_index >= NR_GRPINFO_CACHES)
2355                return -EINVAL;
2356
2357        if (unlikely(cache_index < 0))
2358                cache_index = 0;
2359
2360        mutex_lock(&ext4_grpinfo_slab_create_mutex);
2361        if (ext4_groupinfo_caches[cache_index]) {
2362                mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2363                return 0;       /* Already created */
2364        }
2365
2366        slab_size = offsetof(struct ext4_group_info,
2367                                bb_counters[blocksize_bits + 2]);
2368
2369        cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2370                                        slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2371                                        NULL);
2372
2373        ext4_groupinfo_caches[cache_index] = cachep;
2374
2375        mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2376        if (!cachep) {
2377                printk(KERN_EMERG
2378                       "EXT4-fs: no memory for groupinfo slab cache\n");
2379                return -ENOMEM;
2380        }
2381
2382        return 0;
2383}
2384
2385int ext4_mb_init(struct super_block *sb)
2386{
2387        struct ext4_sb_info *sbi = EXT4_SB(sb);
2388        unsigned i, j;
2389        unsigned offset;
2390        unsigned max;
2391        int ret;
2392
2393        i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2394
2395        sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2396        if (sbi->s_mb_offsets == NULL) {
2397                ret = -ENOMEM;
2398                goto out;
2399        }
2400
2401        i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2402        sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2403        if (sbi->s_mb_maxs == NULL) {
2404                ret = -ENOMEM;
2405                goto out;
2406        }
2407
2408        ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2409        if (ret < 0)
2410                goto out;
2411
2412        /* order 0 is regular bitmap */
2413        sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2414        sbi->s_mb_offsets[0] = 0;
2415
2416        i = 1;
2417        offset = 0;
2418        max = sb->s_blocksize << 2;
2419        do {
2420                sbi->s_mb_offsets[i] = offset;
2421                sbi->s_mb_maxs[i] = max;
2422                offset += 1 << (sb->s_blocksize_bits - i);
2423                max = max >> 1;
2424                i++;
2425        } while (i <= sb->s_blocksize_bits + 1);
2426
2427        spin_lock_init(&sbi->s_md_lock);
2428        spin_lock_init(&sbi->s_bal_lock);
2429
2430        sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2431        sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2432        sbi->s_mb_stats = MB_DEFAULT_STATS;
2433        sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2434        sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2435        /*
2436         * The default group preallocation is 512, which for 4k block
2437         * sizes translates to 2 megabytes.  However for bigalloc file
2438         * systems, this is probably too big (i.e, if the cluster size
2439         * is 1 megabyte, then group preallocation size becomes half a
2440         * gigabyte!).  As a default, we will keep a two megabyte
2441         * group pralloc size for cluster sizes up to 64k, and after
2442         * that, we will force a minimum group preallocation size of
2443         * 32 clusters.  This translates to 8 megs when the cluster
2444         * size is 256k, and 32 megs when the cluster size is 1 meg,
2445         * which seems reasonable as a default.
2446         */
2447        sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2448                                       sbi->s_cluster_bits, 32);
2449        /*
2450         * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2451         * to the lowest multiple of s_stripe which is bigger than
2452         * the s_mb_group_prealloc as determined above. We want
2453         * the preallocation size to be an exact multiple of the
2454         * RAID stripe size so that preallocations don't fragment
2455         * the stripes.
2456         */
2457        if (sbi->s_stripe > 1) {
2458                sbi->s_mb_group_prealloc = roundup(
2459                        sbi->s_mb_group_prealloc, sbi->s_stripe);
2460        }
2461
2462        sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2463        if (sbi->s_locality_groups == NULL) {
2464                ret = -ENOMEM;
2465                goto out_free_groupinfo_slab;
2466        }
2467        for_each_possible_cpu(i) {
2468                struct ext4_locality_group *lg;
2469                lg = per_cpu_ptr(sbi->s_locality_groups, i);
2470                mutex_init(&lg->lg_mutex);
2471                for (j = 0; j < PREALLOC_TB_SIZE; j++)
2472                        INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2473                spin_lock_init(&lg->lg_prealloc_lock);
2474        }
2475
2476        /* init file for buddy data */
2477        ret = ext4_mb_init_backend(sb);
2478        if (ret != 0)
2479                goto out_free_locality_groups;
2480
2481        if (sbi->s_proc)
2482                proc_create_data("mb_groups", S_IRUGO, sbi->s_proc,
2483                                 &ext4_mb_seq_groups_fops, sb);
2484
2485        return 0;
2486
2487out_free_locality_groups:
2488        free_percpu(sbi->s_locality_groups);
2489        sbi->s_locality_groups = NULL;
2490out_free_groupinfo_slab:
2491        ext4_groupinfo_destroy_slabs();
2492out:
2493        kfree(sbi->s_mb_offsets);
2494        sbi->s_mb_offsets = NULL;
2495        kfree(sbi->s_mb_maxs);
2496        sbi->s_mb_maxs = NULL;
2497        return ret;
2498}
2499
2500/* need to called with the ext4 group lock held */
2501static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2502{
2503        struct ext4_prealloc_space *pa;
2504        struct list_head *cur, *tmp;
2505        int count = 0;
2506
2507        list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2508                pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2509                list_del(&pa->pa_group_list);
2510                count++;
2511                kmem_cache_free(ext4_pspace_cachep, pa);
2512        }
2513        if (count)
2514                mb_debug(1, "mballoc: %u PAs left\n", count);
2515
2516}
2517
2518int ext4_mb_release(struct super_block *sb)
2519{
2520        ext4_group_t ngroups = ext4_get_groups_count(sb);
2521        ext4_group_t i;
2522        int num_meta_group_infos;
2523        struct ext4_group_info *grinfo;
2524        struct ext4_sb_info *sbi = EXT4_SB(sb);
2525        struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2526
2527        if (sbi->s_proc)
2528                remove_proc_entry("mb_groups", sbi->s_proc);
2529
2530        if (sbi->s_group_info) {
2531                for (i = 0; i < ngroups; i++) {
2532                        grinfo = ext4_get_group_info(sb, i);
2533#ifdef DOUBLE_CHECK
2534                        kfree(grinfo->bb_bitmap);
2535#endif
2536                        ext4_lock_group(sb, i);
2537                        ext4_mb_cleanup_pa(grinfo);
2538                        ext4_unlock_group(sb, i);
2539                        kmem_cache_free(cachep, grinfo);
2540                }
2541                num_meta_group_infos = (ngroups +
2542                                EXT4_DESC_PER_BLOCK(sb) - 1) >>
2543                        EXT4_DESC_PER_BLOCK_BITS(sb);
2544                for (i = 0; i < num_meta_group_infos; i++)
2545                        kfree(sbi->s_group_info[i]);
2546                ext4_kvfree(sbi->s_group_info);
2547        }
2548        kfree(sbi->s_mb_offsets);
2549        kfree(sbi->s_mb_maxs);
2550        if (sbi->s_buddy_cache)
2551                iput(sbi->s_buddy_cache);
2552        if (sbi->s_mb_stats) {
2553                ext4_msg(sb, KERN_INFO,
2554                       "mballoc: %u blocks %u reqs (%u success)",
2555                                atomic_read(&sbi->s_bal_allocated),
2556                                atomic_read(&sbi->s_bal_reqs),
2557                                atomic_read(&sbi->s_bal_success));
2558                ext4_msg(sb, KERN_INFO,
2559                      "mballoc: %u extents scanned, %u goal hits, "
2560                                "%u 2^N hits, %u breaks, %u lost",
2561                                atomic_read(&sbi->s_bal_ex_scanned),
2562                                atomic_read(&sbi->s_bal_goals),
2563                                atomic_read(&sbi->s_bal_2orders),
2564                                atomic_read(&sbi->s_bal_breaks),
2565                                atomic_read(&sbi->s_mb_lost_chunks));
2566                ext4_msg(sb, KERN_INFO,
2567                       "mballoc: %lu generated and it took %Lu",
2568                                sbi->s_mb_buddies_generated,
2569                                sbi->s_mb_generation_time);
2570                ext4_msg(sb, KERN_INFO,
2571                       "mballoc: %u preallocated, %u discarded",
2572                                atomic_read(&sbi->s_mb_preallocated),
2573                                atomic_read(&sbi->s_mb_discarded));
2574        }
2575
2576        free_percpu(sbi->s_locality_groups);
2577
2578        return 0;
2579}
2580
2581static inline int ext4_issue_discard(struct super_block *sb,
2582                ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2583{
2584        ext4_fsblk_t discard_block;
2585
2586        discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2587                         ext4_group_first_block_no(sb, block_group));
2588        count = EXT4_C2B(EXT4_SB(sb), count);
2589        trace_ext4_discard_blocks(sb,
2590                        (unsigned long long) discard_block, count);
2591        return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2592}
2593
2594/*
2595 * This function is called by the jbd2 layer once the commit has finished,
2596 * so we know we can free the blocks that were released with that commit.
2597 */
2598static void ext4_free_data_callback(struct super_block *sb,
2599                                    struct ext4_journal_cb_entry *jce,
2600                                    int rc)
2601{
2602        struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2603        struct ext4_buddy e4b;
2604        struct ext4_group_info *db;
2605        int err, count = 0, count2 = 0;
2606
2607        mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2608                 entry->efd_count, entry->efd_group, entry);
2609
2610        if (test_opt(sb, DISCARD))
2611                ext4_issue_discard(sb, entry->efd_group,
2612                                   entry->efd_start_cluster, entry->efd_count);
2613
2614        err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2615        /* we expect to find existing buddy because it's pinned */
2616        BUG_ON(err != 0);
2617
2618
2619        db = e4b.bd_info;
2620        /* there are blocks to put in buddy to make them really free */
2621        count += entry->efd_count;
2622        count2++;
2623        ext4_lock_group(sb, entry->efd_group);
2624        /* Take it out of per group rb tree */
2625        rb_erase(&entry->efd_node, &(db->bb_free_root));
2626        mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2627
2628        /*
2629         * Clear the trimmed flag for the group so that the next
2630         * ext4_trim_fs can trim it.
2631         * If the volume is mounted with -o discard, online discard
2632         * is supported and the free blocks will be trimmed online.
2633         */
2634        if (!test_opt(sb, DISCARD))
2635                EXT4_MB_GRP_CLEAR_TRIMMED(db);
2636
2637        if (!db->bb_free_root.rb_node) {
2638                /* No more items in the per group rb tree
2639                 * balance refcounts from ext4_mb_free_metadata()
2640                 */
2641                page_cache_release(e4b.bd_buddy_page);
2642                page_cache_release(e4b.bd_bitmap_page);
2643        }
2644        ext4_unlock_group(sb, entry->efd_group);
2645        kmem_cache_free(ext4_free_data_cachep, entry);
2646        ext4_mb_unload_buddy(&e4b);
2647
2648        mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2649}
2650
2651#ifdef CONFIG_EXT4_DEBUG
2652u8 mb_enable_debug __read_mostly;
2653
2654static struct dentry *debugfs_dir;
2655static struct dentry *debugfs_debug;
2656
2657static void __init ext4_create_debugfs_entry(void)
2658{
2659        debugfs_dir = debugfs_create_dir("ext4", NULL);
2660        if (debugfs_dir)
2661                debugfs_debug = debugfs_create_u8("mballoc-debug",
2662                                                  S_IRUGO | S_IWUSR,
2663                                                  debugfs_dir,
2664                                                  &mb_enable_debug);
2665}
2666
2667static void ext4_remove_debugfs_entry(void)
2668{
2669        debugfs_remove(debugfs_debug);
2670        debugfs_remove(debugfs_dir);
2671}
2672
2673#else
2674
2675static void __init ext4_create_debugfs_entry(void)
2676{
2677}
2678
2679static void ext4_remove_debugfs_entry(void)
2680{
2681}
2682
2683#endif
2684
2685int __init ext4_init_mballoc(void)
2686{
2687        ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2688                                        SLAB_RECLAIM_ACCOUNT);
2689        if (ext4_pspace_cachep == NULL)
2690                return -ENOMEM;
2691
2692        ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2693                                    SLAB_RECLAIM_ACCOUNT);
2694        if (ext4_ac_cachep == NULL) {
2695                kmem_cache_destroy(ext4_pspace_cachep);
2696                return -ENOMEM;
2697        }
2698
2699        ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2700                                           SLAB_RECLAIM_ACCOUNT);
2701        if (ext4_free_data_cachep == NULL) {
2702                kmem_cache_destroy(ext4_pspace_cachep);
2703                kmem_cache_destroy(ext4_ac_cachep);
2704                return -ENOMEM;
2705        }
2706        ext4_create_debugfs_entry();
2707        return 0;
2708}
2709
2710void ext4_exit_mballoc(void)
2711{
2712        /*
2713         * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2714         * before destroying the slab cache.
2715         */
2716        rcu_barrier();
2717        kmem_cache_destroy(ext4_pspace_cachep);
2718        kmem_cache_destroy(ext4_ac_cachep);
2719        kmem_cache_destroy(ext4_free_data_cachep);
2720        ext4_groupinfo_destroy_slabs();
2721        ext4_remove_debugfs_entry();
2722}
2723
2724
2725/*
2726 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2727 * Returns 0 if success or error code
2728 */
2729static noinline_for_stack int
2730ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2731                                handle_t *handle, unsigned int reserv_clstrs)
2732{
2733        struct buffer_head *bitmap_bh = NULL;
2734        struct ext4_group_desc *gdp;
2735        struct buffer_head *gdp_bh;
2736        struct ext4_sb_info *sbi;
2737        struct super_block *sb;
2738        ext4_fsblk_t block;
2739        int err, len;
2740
2741        BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2742        BUG_ON(ac->ac_b_ex.fe_len <= 0);
2743
2744        sb = ac->ac_sb;
2745        sbi = EXT4_SB(sb);
2746
2747        err = -EIO;
2748        bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2749        if (!bitmap_bh)
2750                goto out_err;
2751
2752        err = ext4_journal_get_write_access(handle, bitmap_bh);
2753        if (err)
2754                goto out_err;
2755
2756        err = -EIO;
2757        gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2758        if (!gdp)
2759                goto out_err;
2760
2761        ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2762                        ext4_free_group_clusters(sb, gdp));
2763
2764        err = ext4_journal_get_write_access(handle, gdp_bh);
2765        if (err)
2766                goto out_err;
2767
2768        block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2769
2770        len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2771        if (!ext4_data_block_valid(sbi, block, len)) {
2772                ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2773                           "fs metadata", block, block+len);
2774                /* File system mounted not to panic on error
2775                 * Fix the bitmap and repeat the block allocation
2776                 * We leak some of the blocks here.
2777                 */
2778                ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2779                ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2780                              ac->ac_b_ex.fe_len);
2781                ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2782                err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2783                if (!err)
2784                        err = -EAGAIN;
2785                goto out_err;
2786        }
2787
2788        ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2789#ifdef AGGRESSIVE_CHECK
2790        {
2791                int i;
2792                for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2793                        BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2794                                                bitmap_bh->b_data));
2795                }
2796        }
2797#endif
2798        ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2799                      ac->ac_b_ex.fe_len);
2800        if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2801                gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2802                ext4_free_group_clusters_set(sb, gdp,
2803                                             ext4_free_clusters_after_init(sb,
2804                                                ac->ac_b_ex.fe_group, gdp));
2805        }
2806        len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2807        ext4_free_group_clusters_set(sb, gdp, len);
2808        ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
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        ac->ac_b_ex.fe_logical = ar->logical & ~(sbi->s_cluster_ratio - 1);
4012        ac->ac_status = AC_STATUS_CONTINUE;
4013        ac->ac_sb = sb;
4014        ac->ac_inode = ar->inode;
4015        ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4016        ac->ac_o_ex.fe_group = group;
4017        ac->ac_o_ex.fe_start = block;
4018        ac->ac_o_ex.fe_len = len;
4019        ac->ac_g_ex = ac->ac_o_ex;
4020        ac->ac_flags = ar->flags;
4021
4022        /* we have to define context: we'll we work with a file or
4023         * locality group. this is a policy, actually */
4024        ext4_mb_group_or_file(ac);
4025
4026        mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4027                        "left: %u/%u, right %u/%u to %swritable\n",
4028                        (unsigned) ar->len, (unsigned) ar->logical,
4029                        (unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4030                        (unsigned) ar->lleft, (unsigned) ar->pleft,
4031                        (unsigned) ar->lright, (unsigned) ar->pright,
4032                        atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4033        return 0;
4034
4035}
4036
4037static noinline_for_stack void
4038ext4_mb_discard_lg_preallocations(struct super_block *sb,
4039                                        struct ext4_locality_group *lg,
4040                                        int order, int total_entries)
4041{
4042        ext4_group_t group = 0;
4043        struct ext4_buddy e4b;
4044        struct list_head discard_list;
4045        struct ext4_prealloc_space *pa, *tmp;
4046
4047        mb_debug(1, "discard locality group preallocation\n");
4048
4049        INIT_LIST_HEAD(&discard_list);
4050
4051        spin_lock(&lg->lg_prealloc_lock);
4052        list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4053                                                pa_inode_list) {
4054                spin_lock(&pa->pa_lock);
4055                if (atomic_read(&pa->pa_count)) {
4056                        /*
4057                         * This is the pa that we just used
4058                         * for block allocation. So don't
4059                         * free that
4060                         */
4061                        spin_unlock(&pa->pa_lock);
4062                        continue;
4063                }
4064                if (pa->pa_deleted) {
4065                        spin_unlock(&pa->pa_lock);
4066                        continue;
4067                }
4068                /* only lg prealloc space */
4069                BUG_ON(pa->pa_type != MB_GROUP_PA);
4070
4071                /* seems this one can be freed ... */
4072                pa->pa_deleted = 1;
4073                spin_unlock(&pa->pa_lock);
4074
4075                list_del_rcu(&pa->pa_inode_list);
4076                list_add(&pa->u.pa_tmp_list, &discard_list);
4077
4078                total_entries--;
4079                if (total_entries <= 5) {
4080                        /*
4081                         * we want to keep only 5 entries
4082                         * allowing it to grow to 8. This
4083                         * mak sure we don't call discard
4084                         * soon for this list.
4085                         */
4086                        break;
4087                }
4088        }
4089        spin_unlock(&lg->lg_prealloc_lock);
4090
4091        list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4092
4093                ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, NULL);
4094                if (ext4_mb_load_buddy(sb, group, &e4b)) {
4095                        ext4_error(sb, "Error loading buddy information for %u",
4096                                        group);
4097                        continue;
4098                }
4099                ext4_lock_group(sb, group);
4100                list_del(&pa->pa_group_list);
4101                ext4_mb_release_group_pa(&e4b, pa);
4102                ext4_unlock_group(sb, group);
4103
4104                ext4_mb_unload_buddy(&e4b);
4105                list_del(&pa->u.pa_tmp_list);
4106                call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4107        }
4108}
4109
4110/*
4111 * We have incremented pa_count. So it cannot be freed at this
4112 * point. Also we hold lg_mutex. So no parallel allocation is
4113 * possible from this lg. That means pa_free cannot be updated.
4114 *
4115 * A parallel ext4_mb_discard_group_preallocations is possible.
4116 * which can cause the lg_prealloc_list to be updated.
4117 */
4118
4119static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4120{
4121        int order, added = 0, lg_prealloc_count = 1;
4122        struct super_block *sb = ac->ac_sb;
4123        struct ext4_locality_group *lg = ac->ac_lg;
4124        struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4125
4126        order = fls(pa->pa_free) - 1;
4127        if (order > PREALLOC_TB_SIZE - 1)
4128                /* The max size of hash table is PREALLOC_TB_SIZE */
4129                order = PREALLOC_TB_SIZE - 1;
4130        /* Add the prealloc space to lg */
4131        rcu_read_lock();
4132        list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4133                                                pa_inode_list) {
4134                spin_lock(&tmp_pa->pa_lock);
4135                if (tmp_pa->pa_deleted) {
4136                        spin_unlock(&tmp_pa->pa_lock);
4137                        continue;
4138                }
4139                if (!added && pa->pa_free < tmp_pa->pa_free) {
4140                        /* Add to the tail of the previous entry */
4141                        list_add_tail_rcu(&pa->pa_inode_list,
4142                                                &tmp_pa->pa_inode_list);
4143                        added = 1;
4144                        /*
4145                         * we want to count the total
4146                         * number of entries in the list
4147                         */
4148                }
4149                spin_unlock(&tmp_pa->pa_lock);
4150                lg_prealloc_count++;
4151        }
4152        if (!added)
4153                list_add_tail_rcu(&pa->pa_inode_list,
4154                                        &lg->lg_prealloc_list[order]);
4155        rcu_read_unlock();
4156
4157        /* Now trim the list to be not more than 8 elements */
4158        if (lg_prealloc_count > 8) {
4159                ext4_mb_discard_lg_preallocations(sb, lg,
4160                                                order, lg_prealloc_count);
4161                return;
4162        }
4163        return ;
4164}
4165
4166/*
4167 * release all resource we used in allocation
4168 */
4169static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4170{
4171        struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4172        struct ext4_prealloc_space *pa = ac->ac_pa;
4173        if (pa) {
4174                if (pa->pa_type == MB_GROUP_PA) {
4175                        /* see comment in ext4_mb_use_group_pa() */
4176                        spin_lock(&pa->pa_lock);
4177                        pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4178                        pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4179                        pa->pa_free -= ac->ac_b_ex.fe_len;
4180                        pa->pa_len -= ac->ac_b_ex.fe_len;
4181                        spin_unlock(&pa->pa_lock);
4182                }
4183        }
4184        if (pa) {
4185                /*
4186                 * We want to add the pa to the right bucket.
4187                 * Remove it from the list and while adding
4188                 * make sure the list to which we are adding
4189                 * doesn't grow big.
4190                 */
4191                if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4192                        spin_lock(pa->pa_obj_lock);
4193                        list_del_rcu(&pa->pa_inode_list);
4194                        spin_unlock(pa->pa_obj_lock);
4195                        ext4_mb_add_n_trim(ac);
4196                }
4197                ext4_mb_put_pa(ac, ac->ac_sb, pa);
4198        }
4199        if (ac->ac_bitmap_page)
4200                page_cache_release(ac->ac_bitmap_page);
4201        if (ac->ac_buddy_page)
4202                page_cache_release(ac->ac_buddy_page);
4203        if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4204                mutex_unlock(&ac->ac_lg->lg_mutex);
4205        ext4_mb_collect_stats(ac);
4206        return 0;
4207}
4208
4209static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4210{
4211        ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4212        int ret;
4213        int freed = 0;
4214
4215        trace_ext4_mb_discard_preallocations(sb, needed);
4216        for (i = 0; i < ngroups && needed > 0; i++) {
4217                ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4218                freed += ret;
4219                needed -= ret;
4220        }
4221
4222        return freed;
4223}
4224
4225/*
4226 * Main entry point into mballoc to allocate blocks
4227 * it tries to use preallocation first, then falls back
4228 * to usual allocation
4229 */
4230ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4231                                struct ext4_allocation_request *ar, int *errp)
4232{
4233        int freed;
4234        struct ext4_allocation_context *ac = NULL;
4235        struct ext4_sb_info *sbi;
4236        struct super_block *sb;
4237        ext4_fsblk_t block = 0;
4238        unsigned int inquota = 0;
4239        unsigned int reserv_clstrs = 0;
4240
4241        sb = ar->inode->i_sb;
4242        sbi = EXT4_SB(sb);
4243
4244        trace_ext4_request_blocks(ar);
4245
4246        /* Allow to use superuser reservation for quota file */
4247        if (IS_NOQUOTA(ar->inode))
4248                ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4249
4250        /*
4251         * For delayed allocation, we could skip the ENOSPC and
4252         * EDQUOT check, as blocks and quotas have been already
4253         * reserved when data being copied into pagecache.
4254         */
4255        if (ext4_test_inode_state(ar->inode, EXT4_STATE_DELALLOC_RESERVED))
4256                ar->flags |= EXT4_MB_DELALLOC_RESERVED;
4257        else {
4258                /* Without delayed allocation we need to verify
4259                 * there is enough free blocks to do block allocation
4260                 * and verify allocation doesn't exceed the quota limits.
4261                 */
4262                while (ar->len &&
4263                        ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4264
4265                        /* let others to free the space */
4266                        yield();
4267                        ar->len = ar->len >> 1;
4268                }
4269                if (!ar->len) {
4270                        *errp = -ENOSPC;
4271                        return 0;
4272                }
4273                reserv_clstrs = ar->len;
4274                if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4275                        dquot_alloc_block_nofail(ar->inode,
4276                                                 EXT4_C2B(sbi, ar->len));
4277                } else {
4278                        while (ar->len &&
4279                                dquot_alloc_block(ar->inode,
4280                                                  EXT4_C2B(sbi, ar->len))) {
4281
4282                                ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4283                                ar->len--;
4284                        }
4285                }
4286                inquota = ar->len;
4287                if (ar->len == 0) {
4288                        *errp = -EDQUOT;
4289                        goto out;
4290                }
4291        }
4292
4293        ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4294        if (!ac) {
4295                ar->len = 0;
4296                *errp = -ENOMEM;
4297                goto out;
4298        }
4299
4300        *errp = ext4_mb_initialize_context(ac, ar);
4301        if (*errp) {
4302                ar->len = 0;
4303                goto out;
4304        }
4305
4306        ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4307        if (!ext4_mb_use_preallocated(ac)) {
4308                ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4309                ext4_mb_normalize_request(ac, ar);
4310repeat:
4311                /* allocate space in core */
4312                *errp = ext4_mb_regular_allocator(ac);
4313                if (*errp)
4314                        goto errout;
4315
4316                /* as we've just preallocated more space than
4317                 * user requested orinally, we store allocated
4318                 * space in a special descriptor */
4319                if (ac->ac_status == AC_STATUS_FOUND &&
4320                                ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4321                        ext4_mb_new_preallocation(ac);
4322        }
4323        if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4324                *errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4325                if (*errp == -EAGAIN) {
4326                        /*
4327                         * drop the reference that we took
4328                         * in ext4_mb_use_best_found
4329                         */
4330                        ext4_mb_release_context(ac);
4331                        ac->ac_b_ex.fe_group = 0;
4332                        ac->ac_b_ex.fe_start = 0;
4333                        ac->ac_b_ex.fe_len = 0;
4334                        ac->ac_status = AC_STATUS_CONTINUE;
4335                        goto repeat;
4336                } else if (*errp)
4337                errout:
4338                        ext4_discard_allocated_blocks(ac);
4339                else {
4340                        block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4341                        ar->len = ac->ac_b_ex.fe_len;
4342                }
4343        } else {
4344                freed  = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4345                if (freed)
4346                        goto repeat;
4347                *errp = -ENOSPC;
4348        }
4349
4350        if (*errp) {
4351                ac->ac_b_ex.fe_len = 0;
4352                ar->len = 0;
4353                ext4_mb_show_ac(ac);
4354        }
4355        ext4_mb_release_context(ac);
4356out:
4357        if (ac)
4358                kmem_cache_free(ext4_ac_cachep, ac);
4359        if (inquota && ar->len < inquota)
4360                dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4361        if (!ar->len) {
4362                if (!ext4_test_inode_state(ar->inode,
4363                                           EXT4_STATE_DELALLOC_RESERVED))
4364                        /* release all the reserved blocks if non delalloc */
4365                        percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4366                                                reserv_clstrs);
4367        }
4368
4369        trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4370
4371        return block;
4372}
4373
4374/*
4375 * We can merge two free data extents only if the physical blocks
4376 * are contiguous, AND the extents were freed by the same transaction,
4377 * AND the blocks are associated with the same group.
4378 */
4379static int can_merge(struct ext4_free_data *entry1,
4380                        struct ext4_free_data *entry2)
4381{
4382        if ((entry1->efd_tid == entry2->efd_tid) &&
4383            (entry1->efd_group == entry2->efd_group) &&
4384            ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4385                return 1;
4386        return 0;
4387}
4388
4389static noinline_for_stack int
4390ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4391                      struct ext4_free_data *new_entry)
4392{
4393        ext4_group_t group = e4b->bd_group;
4394        ext4_grpblk_t cluster;
4395        struct ext4_free_data *entry;
4396        struct ext4_group_info *db = e4b->bd_info;
4397        struct super_block *sb = e4b->bd_sb;
4398        struct ext4_sb_info *sbi = EXT4_SB(sb);
4399        struct rb_node **n = &db->bb_free_root.rb_node, *node;
4400        struct rb_node *parent = NULL, *new_node;
4401
4402        BUG_ON(!ext4_handle_valid(handle));
4403        BUG_ON(e4b->bd_bitmap_page == NULL);
4404        BUG_ON(e4b->bd_buddy_page == NULL);
4405
4406        new_node = &new_entry->efd_node;
4407        cluster = new_entry->efd_start_cluster;
4408
4409        if (!*n) {
4410                /* first free block exent. We need to
4411                   protect buddy cache from being freed,
4412                 * otherwise we'll refresh it from
4413                 * on-disk bitmap and lose not-yet-available
4414                 * blocks */
4415                page_cache_get(e4b->bd_buddy_page);
4416                page_cache_get(e4b->bd_bitmap_page);
4417        }
4418        while (*n) {
4419                parent = *n;
4420                entry = rb_entry(parent, struct ext4_free_data, efd_node);
4421                if (cluster < entry->efd_start_cluster)
4422                        n = &(*n)->rb_left;
4423                else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4424                        n = &(*n)->rb_right;
4425                else {
4426                        ext4_grp_locked_error(sb, group, 0,
4427                                ext4_group_first_block_no(sb, group) +
4428                                EXT4_C2B(sbi, cluster),
4429                                "Block already on to-be-freed list");
4430                        return 0;
4431                }
4432        }
4433
4434        rb_link_node(new_node, parent, n);
4435        rb_insert_color(new_node, &db->bb_free_root);
4436
4437        /* Now try to see the extent can be merged to left and right */
4438        node = rb_prev(new_node);
4439        if (node) {
4440                entry = rb_entry(node, struct ext4_free_data, efd_node);
4441                if (can_merge(entry, new_entry)) {
4442                        new_entry->efd_start_cluster = entry->efd_start_cluster;
4443                        new_entry->efd_count += entry->efd_count;
4444                        rb_erase(node, &(db->bb_free_root));
4445                        ext4_journal_callback_del(handle, &entry->efd_jce);
4446                        kmem_cache_free(ext4_free_data_cachep, entry);
4447                }
4448        }
4449
4450        node = rb_next(new_node);
4451        if (node) {
4452                entry = rb_entry(node, struct ext4_free_data, efd_node);
4453                if (can_merge(new_entry, entry)) {
4454                        new_entry->efd_count += entry->efd_count;
4455                        rb_erase(node, &(db->bb_free_root));
4456                        ext4_journal_callback_del(handle, &entry->efd_jce);
4457                        kmem_cache_free(ext4_free_data_cachep, entry);
4458                }
4459        }
4460        /* Add the extent to transaction's private list */
4461        ext4_journal_callback_add(handle, ext4_free_data_callback,
4462                                  &new_entry->efd_jce);
4463        return 0;
4464}
4465
4466/**
4467 * ext4_free_blocks() -- Free given blocks and update quota
4468 * @handle:             handle for this transaction
4469 * @inode:              inode
4470 * @block:              start physical block to free
4471 * @count:              number of blocks to count
4472 * @flags:              flags used by ext4_free_blocks
4473 */
4474void ext4_free_blocks(handle_t *handle, struct inode *inode,
4475                      struct buffer_head *bh, ext4_fsblk_t block,
4476                      unsigned long count, int flags)
4477{
4478        struct buffer_head *bitmap_bh = NULL;
4479        struct super_block *sb = inode->i_sb;
4480        struct ext4_group_desc *gdp;
4481        unsigned long freed = 0;
4482        unsigned int overflow;
4483        ext4_grpblk_t bit;
4484        struct buffer_head *gd_bh;
4485        ext4_group_t block_group;
4486        struct ext4_sb_info *sbi;
4487        struct ext4_buddy e4b;
4488        unsigned int count_clusters;
4489        int err = 0;
4490        int ret;
4491
4492        if (bh) {
4493                if (block)
4494                        BUG_ON(block != bh->b_blocknr);
4495                else
4496                        block = bh->b_blocknr;
4497        }
4498
4499        sbi = EXT4_SB(sb);
4500        if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4501            !ext4_data_block_valid(sbi, block, count)) {
4502                ext4_error(sb, "Freeing blocks not in datazone - "
4503                           "block = %llu, count = %lu", block, count);
4504                goto error_return;
4505        }
4506
4507        ext4_debug("freeing block %llu\n", block);
4508        trace_ext4_free_blocks(inode, block, count, flags);
4509
4510        if (flags & EXT4_FREE_BLOCKS_FORGET) {
4511                struct buffer_head *tbh = bh;
4512                int i;
4513
4514                BUG_ON(bh && (count > 1));
4515
4516                for (i = 0; i < count; i++) {
4517                        if (!bh)
4518                                tbh = sb_find_get_block(inode->i_sb,
4519                                                        block + i);
4520                        if (unlikely(!tbh))
4521                                continue;
4522                        ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4523                                    inode, tbh, block + i);
4524                }
4525        }
4526
4527        /*
4528         * We need to make sure we don't reuse the freed block until
4529         * after the transaction is committed, which we can do by
4530         * treating the block as metadata, below.  We make an
4531         * exception if the inode is to be written in writeback mode
4532         * since writeback mode has weak data consistency guarantees.
4533         */
4534        if (!ext4_should_writeback_data(inode))
4535                flags |= EXT4_FREE_BLOCKS_METADATA;
4536
4537        /*
4538         * If the extent to be freed does not begin on a cluster
4539         * boundary, we need to deal with partial clusters at the
4540         * beginning and end of the extent.  Normally we will free
4541         * blocks at the beginning or the end unless we are explicitly
4542         * requested to avoid doing so.
4543         */
4544        overflow = block & (sbi->s_cluster_ratio - 1);
4545        if (overflow) {
4546                if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4547                        overflow = sbi->s_cluster_ratio - overflow;
4548                        block += overflow;
4549                        if (count > overflow)
4550                                count -= overflow;
4551                        else
4552                                return;
4553                } else {
4554                        block -= overflow;
4555                        count += overflow;
4556                }
4557        }
4558        overflow = count & (sbi->s_cluster_ratio - 1);
4559        if (overflow) {
4560                if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4561                        if (count > overflow)
4562                                count -= overflow;
4563                        else
4564                                return;
4565                } else
4566                        count += sbi->s_cluster_ratio - overflow;
4567        }
4568
4569do_more:
4570        overflow = 0;
4571        ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4572
4573        /*
4574         * Check to see if we are freeing blocks across a group
4575         * boundary.
4576         */
4577        if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4578                overflow = EXT4_C2B(sbi, bit) + count -
4579                        EXT4_BLOCKS_PER_GROUP(sb);
4580                count -= overflow;
4581        }
4582        count_clusters = EXT4_B2C(sbi, count);
4583        bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4584        if (!bitmap_bh) {
4585                err = -EIO;
4586                goto error_return;
4587        }
4588        gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4589        if (!gdp) {
4590                err = -EIO;
4591                goto error_return;
4592        }
4593
4594        if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4595            in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4596            in_range(block, ext4_inode_table(sb, gdp),
4597                     EXT4_SB(sb)->s_itb_per_group) ||
4598            in_range(block + count - 1, ext4_inode_table(sb, gdp),
4599                     EXT4_SB(sb)->s_itb_per_group)) {
4600
4601                ext4_error(sb, "Freeing blocks in system zone - "
4602                           "Block = %llu, count = %lu", block, count);
4603                /* err = 0. ext4_std_error should be a no op */
4604                goto error_return;
4605        }
4606
4607        BUFFER_TRACE(bitmap_bh, "getting write access");
4608        err = ext4_journal_get_write_access(handle, bitmap_bh);
4609        if (err)
4610                goto error_return;
4611
4612        /*
4613         * We are about to modify some metadata.  Call the journal APIs
4614         * to unshare ->b_data if a currently-committing transaction is
4615         * using it
4616         */
4617        BUFFER_TRACE(gd_bh, "get_write_access");
4618        err = ext4_journal_get_write_access(handle, gd_bh);
4619        if (err)
4620                goto error_return;
4621#ifdef AGGRESSIVE_CHECK
4622        {
4623                int i;
4624                for (i = 0; i < count_clusters; i++)
4625                        BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4626        }
4627#endif
4628        trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4629
4630        err = ext4_mb_load_buddy(sb, block_group, &e4b);
4631        if (err)
4632                goto error_return;
4633
4634        if ((flags & EXT4_FREE_BLOCKS_METADATA) && ext4_handle_valid(handle)) {
4635                struct ext4_free_data *new_entry;
4636                /*
4637                 * blocks being freed are metadata. these blocks shouldn't
4638                 * be used until this transaction is committed
4639                 */
4640                new_entry = kmem_cache_alloc(ext4_free_data_cachep, GFP_NOFS);
4641                if (!new_entry) {
4642                        ext4_mb_unload_buddy(&e4b);
4643                        err = -ENOMEM;
4644                        goto error_return;
4645                }
4646                new_entry->efd_start_cluster = bit;
4647                new_entry->efd_group = block_group;
4648                new_entry->efd_count = count_clusters;
4649                new_entry->efd_tid = handle->h_transaction->t_tid;
4650
4651                ext4_lock_group(sb, block_group);
4652                mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4653                ext4_mb_free_metadata(handle, &e4b, new_entry);
4654        } else {
4655                /* need to update group_info->bb_free and bitmap
4656                 * with group lock held. generate_buddy look at
4657                 * them with group lock_held
4658                 */
4659                if (test_opt(sb, DISCARD))
4660                        ext4_issue_discard(sb, block_group, bit, count);
4661                ext4_lock_group(sb, block_group);
4662                mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4663                mb_free_blocks(inode, &e4b, bit, count_clusters);
4664        }
4665
4666        ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4667        ext4_free_group_clusters_set(sb, gdp, ret);
4668        ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
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 physical 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_group_desc_csum_set(sb, block_group, desc);
4814        ext4_unlock_group(sb, block_group);
4815        percpu_counter_add(&sbi->s_freeclusters_counter,
4816                           EXT4_B2C(sbi, blocks_freed));
4817
4818        if (sbi->s_log_groups_per_flex) {
4819                ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4820                atomic_add(EXT4_B2C(sbi, blocks_freed),
4821                           &sbi->s_flex_groups[flex_group].free_clusters);
4822        }
4823
4824        ext4_mb_unload_buddy(&e4b);
4825
4826        /* We dirtied the bitmap block */
4827        BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4828        err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4829
4830        /* And the group descriptor block */
4831        BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4832        ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4833        if (!err)
4834                err = ret;
4835
4836error_return:
4837        brelse(bitmap_bh);
4838        ext4_std_error(sb, err);
4839        return err;
4840}
4841
4842/**
4843 * ext4_trim_extent -- function to TRIM one single free extent in the group
4844 * @sb:         super block for the file system
4845 * @start:      starting block of the free extent in the alloc. group
4846 * @count:      number of blocks to TRIM
4847 * @group:      alloc. group we are working with
4848 * @e4b:        ext4 buddy for the group
4849 *
4850 * Trim "count" blocks starting at "start" in the "group". To assure that no
4851 * one will allocate those blocks, mark it as used in buddy bitmap. This must
4852 * be called with under the group lock.
4853 */
4854static void ext4_trim_extent(struct super_block *sb, int start, int count,
4855                             ext4_group_t group, struct ext4_buddy *e4b)
4856{
4857        struct ext4_free_extent ex;
4858
4859        trace_ext4_trim_extent(sb, group, start, count);
4860
4861        assert_spin_locked(ext4_group_lock_ptr(sb, group));
4862
4863        ex.fe_start = start;
4864        ex.fe_group = group;
4865        ex.fe_len = count;
4866
4867        /*
4868         * Mark blocks used, so no one can reuse them while
4869         * being trimmed.
4870         */
4871        mb_mark_used(e4b, &ex);
4872        ext4_unlock_group(sb, group);
4873        ext4_issue_discard(sb, group, start, count);
4874        ext4_lock_group(sb, group);
4875        mb_free_blocks(NULL, e4b, start, ex.fe_len);
4876}
4877
4878/**
4879 * ext4_trim_all_free -- function to trim all free space in alloc. group
4880 * @sb:                 super block for file system
4881 * @group:              group to be trimmed
4882 * @start:              first group block to examine
4883 * @max:                last group block to examine
4884 * @minblocks:          minimum extent block count
4885 *
4886 * ext4_trim_all_free walks through group's buddy bitmap searching for free
4887 * extents. When the free block is found, ext4_trim_extent is called to TRIM
4888 * the extent.
4889 *
4890 *
4891 * ext4_trim_all_free walks through group's block bitmap searching for free
4892 * extents. When the free extent is found, mark it as used in group buddy
4893 * bitmap. Then issue a TRIM command on this extent and free the extent in
4894 * the group buddy bitmap. This is done until whole group is scanned.
4895 */
4896static ext4_grpblk_t
4897ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
4898                   ext4_grpblk_t start, ext4_grpblk_t max,
4899                   ext4_grpblk_t minblocks)
4900{
4901        void *bitmap;
4902        ext4_grpblk_t next, count = 0, free_count = 0;
4903        struct ext4_buddy e4b;
4904        int ret;
4905
4906        trace_ext4_trim_all_free(sb, group, start, max);
4907
4908        ret = ext4_mb_load_buddy(sb, group, &e4b);
4909        if (ret) {
4910                ext4_error(sb, "Error in loading buddy "
4911                                "information for %u", group);
4912                return ret;
4913        }
4914        bitmap = e4b.bd_bitmap;
4915
4916        ext4_lock_group(sb, group);
4917        if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
4918            minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
4919                goto out;
4920
4921        start = (e4b.bd_info->bb_first_free > start) ?
4922                e4b.bd_info->bb_first_free : start;
4923
4924        while (start <= max) {
4925                start = mb_find_next_zero_bit(bitmap, max + 1, start);
4926                if (start > max)
4927                        break;
4928                next = mb_find_next_bit(bitmap, max + 1, start);
4929
4930                if ((next - start) >= minblocks) {
4931                        ext4_trim_extent(sb, start,
4932                                         next - start, group, &e4b);
4933                        count += next - start;
4934                }
4935                free_count += next - start;
4936                start = next + 1;
4937
4938                if (fatal_signal_pending(current)) {
4939                        count = -ERESTARTSYS;
4940                        break;
4941                }
4942
4943                if (need_resched()) {
4944                        ext4_unlock_group(sb, group);
4945                        cond_resched();
4946                        ext4_lock_group(sb, group);
4947                }
4948
4949                if ((e4b.bd_info->bb_free - free_count) < minblocks)
4950                        break;
4951        }
4952
4953        if (!ret)
4954                EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
4955out:
4956        ext4_unlock_group(sb, group);
4957        ext4_mb_unload_buddy(&e4b);
4958
4959        ext4_debug("trimmed %d blocks in the group %d\n",
4960                count, group);
4961
4962        return count;
4963}
4964
4965/**
4966 * ext4_trim_fs() -- trim ioctl handle function
4967 * @sb:                 superblock for filesystem
4968 * @range:              fstrim_range structure
4969 *
4970 * start:       First Byte to trim
4971 * len:         number of Bytes to trim from start
4972 * minlen:      minimum extent length in Bytes
4973 * ext4_trim_fs goes through all allocation groups containing Bytes from
4974 * start to start+len. For each such a group ext4_trim_all_free function
4975 * is invoked to trim all free space.
4976 */
4977int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
4978{
4979        struct ext4_group_info *grp;
4980        ext4_group_t group, first_group, last_group;
4981        ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
4982        uint64_t start, end, minlen, trimmed = 0;
4983        ext4_fsblk_t first_data_blk =
4984                        le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
4985        ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
4986        int ret = 0;
4987
4988        start = range->start >> sb->s_blocksize_bits;
4989        end = start + (range->len >> sb->s_blocksize_bits) - 1;
4990        minlen = EXT4_NUM_B2C(EXT4_SB(sb),
4991                              range->minlen >> sb->s_blocksize_bits);
4992
4993        if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
4994            start >= max_blks ||
4995            range->len < sb->s_blocksize)
4996                return -EINVAL;
4997        if (end >= max_blks)
4998                end = max_blks - 1;
4999        if (end <= first_data_blk)
5000                goto out;
5001        if (start < first_data_blk)
5002                start = first_data_blk;
5003
5004        /* Determine first and last group to examine based on start and end */
5005        ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5006                                     &first_group, &first_cluster);
5007        ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5008                                     &last_group, &last_cluster);
5009
5010        /* end now represents the last cluster to discard in this group */
5011        end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5012
5013        for (group = first_group; group <= last_group; group++) {
5014                grp = ext4_get_group_info(sb, group);
5015                /* We only do this if the grp has never been initialized */
5016                if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5017                        ret = ext4_mb_init_group(sb, group);
5018                        if (ret)
5019                                break;
5020                }
5021
5022                /*
5023                 * For all the groups except the last one, last cluster will
5024                 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5025                 * change it for the last group, note that last_cluster is
5026                 * already computed earlier by ext4_get_group_no_and_offset()
5027                 */
5028                if (group == last_group)
5029                        end = last_cluster;
5030
5031                if (grp->bb_free >= minlen) {
5032                        cnt = ext4_trim_all_free(sb, group, first_cluster,
5033                                                end, minlen);
5034                        if (cnt < 0) {
5035                                ret = cnt;
5036                                break;
5037                        }
5038                        trimmed += cnt;
5039                }
5040
5041                /*
5042                 * For every group except the first one, we are sure
5043                 * that the first cluster to discard will be cluster #0.
5044                 */
5045                first_cluster = 0;
5046        }
5047
5048        if (!ret)
5049                atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5050
5051out:
5052        range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5053        return ret;
5054}
5055
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.