Linux Audio

Check our new training course

Loading...
v4.6
   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 License
  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/log2.h>
  27#include <linux/module.h>
  28#include <linux/slab.h>
  29#include <linux/backing-dev.h>
  30#include <trace/events/ext4.h>
  31
  32#ifdef CONFIG_EXT4_DEBUG
  33ushort ext4_mballoc_debug __read_mostly;
  34
  35module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
  36MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
  37#endif
  38
  39/*
  40 * MUSTDO:
  41 *   - test ext4_ext_search_left() and ext4_ext_search_right()
  42 *   - search for metadata in few groups
  43 *
  44 * TODO v4:
  45 *   - normalization should take into account whether file is still open
  46 *   - discard preallocations if no free space left (policy?)
  47 *   - don't normalize tails
  48 *   - quota
  49 *   - reservation for superuser
  50 *
  51 * TODO v3:
  52 *   - bitmap read-ahead (proposed by Oleg Drokin aka green)
  53 *   - track min/max extents in each group for better group selection
  54 *   - mb_mark_used() may allocate chunk right after splitting buddy
  55 *   - tree of groups sorted by number of free blocks
  56 *   - error handling
  57 */
  58
  59/*
  60 * The allocation request involve request for multiple number of blocks
  61 * near to the goal(block) value specified.
  62 *
  63 * During initialization phase of the allocator we decide to use the
  64 * group preallocation or inode preallocation depending on the size of
  65 * the file. The size of the file could be the resulting file size we
  66 * would have after allocation, or the current file size, which ever
  67 * is larger. If the size is less than sbi->s_mb_stream_request we
  68 * select to use the group preallocation. The default value of
  69 * s_mb_stream_request is 16 blocks. This can also be tuned via
  70 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
  71 * terms of number of blocks.
  72 *
  73 * The main motivation for having small file use group preallocation is to
  74 * ensure that we have small files closer together on the disk.
  75 *
  76 * First stage the allocator looks at the inode prealloc list,
  77 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
  78 * spaces for this particular inode. The inode prealloc space is
  79 * represented as:
  80 *
  81 * pa_lstart -> the logical start block for this prealloc space
  82 * pa_pstart -> the physical start block for this prealloc space
  83 * pa_len    -> length for this prealloc space (in clusters)
  84 * pa_free   ->  free space available in this prealloc space (in clusters)
  85 *
  86 * The inode preallocation space is used looking at the _logical_ start
  87 * block. If only the logical file block falls within the range of prealloc
  88 * space we will consume the particular prealloc space. This makes sure that
  89 * we have contiguous physical blocks representing the file blocks
  90 *
  91 * The important thing to be noted in case of inode prealloc space is that
  92 * we don't modify the values associated to inode prealloc space except
  93 * pa_free.
  94 *
  95 * If we are not able to find blocks in the inode prealloc space and if we
  96 * have the group allocation flag set then we look at the locality group
  97 * prealloc space. These are per CPU prealloc list represented as
  98 *
  99 * ext4_sb_info.s_locality_groups[smp_processor_id()]
 100 *
 101 * The reason for having a per cpu locality group is to reduce the contention
 102 * between CPUs. It is possible to get scheduled at this point.
 103 *
 104 * The locality group prealloc space is used looking at whether we have
 105 * enough free space (pa_free) within the prealloc space.
 106 *
 107 * If we can't allocate blocks via inode prealloc or/and locality group
 108 * prealloc then we look at the buddy cache. The buddy cache is represented
 109 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
 110 * mapped to the buddy and bitmap information regarding different
 111 * groups. The buddy information is attached to buddy cache inode so that
 112 * we can access them through the page cache. The information regarding
 113 * each group is loaded via ext4_mb_load_buddy.  The information involve
 114 * block bitmap and buddy information. The information are stored in the
 115 * inode as:
 116 *
 117 *  {                        page                        }
 118 *  [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
 119 *
 120 *
 121 * one block each for bitmap and buddy information.  So for each group we
 122 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
 123 * blocksize) blocks.  So it can have information regarding groups_per_page
 124 * which is blocks_per_page/2
 125 *
 126 * The buddy cache inode is not stored on disk. The inode is thrown
 127 * away when the filesystem is unmounted.
 128 *
 129 * We look for count number of blocks in the buddy cache. If we were able
 130 * to locate that many free blocks we return with additional information
 131 * regarding rest of the contiguous physical block available
 132 *
 133 * Before allocating blocks via buddy cache we normalize the request
 134 * blocks. This ensure we ask for more blocks that we needed. The extra
 135 * blocks that we get after allocation is added to the respective prealloc
 136 * list. In case of inode preallocation we follow a list of heuristics
 137 * based on file size. This can be found in ext4_mb_normalize_request. If
 138 * we are doing a group prealloc we try to normalize the request to
 139 * sbi->s_mb_group_prealloc.  The default value of s_mb_group_prealloc is
 140 * dependent on the cluster size; for non-bigalloc file systems, it is
 141 * 512 blocks. This can be tuned via
 142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
 143 * terms of number of blocks. If we have mounted the file system with -O
 144 * stripe=<value> option the group prealloc request is normalized to the
 145 * the smallest multiple of the stripe value (sbi->s_stripe) which is
 146 * greater than the default mb_group_prealloc.
 147 *
 148 * The regular allocator (using the buddy cache) supports a few tunables.
 149 *
 150 * /sys/fs/ext4/<partition>/mb_min_to_scan
 151 * /sys/fs/ext4/<partition>/mb_max_to_scan
 152 * /sys/fs/ext4/<partition>/mb_order2_req
 153 *
 154 * The regular allocator uses buddy scan only if the request len is power of
 155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
 156 * value of s_mb_order2_reqs can be tuned via
 157 * /sys/fs/ext4/<partition>/mb_order2_req.  If the request len is equal to
 158 * stripe size (sbi->s_stripe), we try to search for contiguous block in
 159 * stripe size. This should result in better allocation on RAID setups. If
 160 * not, we search in the specific group using bitmap for best extents. The
 161 * tunable min_to_scan and max_to_scan control the behaviour here.
 162 * min_to_scan indicate how long the mballoc __must__ look for a best
 163 * extent and max_to_scan indicates how long the mballoc __can__ look for a
 164 * best extent in the found extents. Searching for the blocks starts with
 165 * the group specified as the goal value in allocation context via
 166 * ac_g_ex. Each group is first checked based on the criteria whether it
 167 * can be used for allocation. ext4_mb_good_group explains how the groups are
 168 * checked.
 169 *
 170 * Both the prealloc space are getting populated as above. So for the first
 171 * request we will hit the buddy cache which will result in this prealloc
 172 * space getting filled. The prealloc space is then later used for the
 173 * subsequent request.
 174 */
 175
 176/*
 177 * mballoc operates on the following data:
 178 *  - on-disk bitmap
 179 *  - in-core buddy (actually includes buddy and bitmap)
 180 *  - preallocation descriptors (PAs)
 181 *
 182 * there are two types of preallocations:
 183 *  - inode
 184 *    assiged to specific inode and can be used for this inode only.
 185 *    it describes part of inode's space preallocated to specific
 186 *    physical blocks. any block from that preallocated can be used
 187 *    independent. the descriptor just tracks number of blocks left
 188 *    unused. so, before taking some block from descriptor, one must
 189 *    make sure corresponded logical block isn't allocated yet. this
 190 *    also means that freeing any block within descriptor's range
 191 *    must discard all preallocated blocks.
 192 *  - locality group
 193 *    assigned to specific locality group which does not translate to
 194 *    permanent set of inodes: inode can join and leave group. space
 195 *    from this type of preallocation can be used for any inode. thus
 196 *    it's consumed from the beginning to the end.
 197 *
 198 * relation between them can be expressed as:
 199 *    in-core buddy = on-disk bitmap + preallocation descriptors
 200 *
 201 * this mean blocks mballoc considers used are:
 202 *  - allocated blocks (persistent)
 203 *  - preallocated blocks (non-persistent)
 204 *
 205 * consistency in mballoc world means that at any time a block is either
 206 * free or used in ALL structures. notice: "any time" should not be read
 207 * literally -- time is discrete and delimited by locks.
 208 *
 209 *  to keep it simple, we don't use block numbers, instead we count number of
 210 *  blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
 211 *
 212 * all operations can be expressed as:
 213 *  - init buddy:			buddy = on-disk + PAs
 214 *  - new PA:				buddy += N; PA = N
 215 *  - use inode PA:			on-disk += N; PA -= N
 216 *  - discard inode PA			buddy -= on-disk - PA; PA = 0
 217 *  - use locality group PA		on-disk += N; PA -= N
 218 *  - discard locality group PA		buddy -= PA; PA = 0
 219 *  note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
 220 *        is used in real operation because we can't know actual used
 221 *        bits from PA, only from on-disk bitmap
 222 *
 223 * if we follow this strict logic, then all operations above should be atomic.
 224 * given some of them can block, we'd have to use something like semaphores
 225 * killing performance on high-end SMP hardware. let's try to relax it using
 226 * the following knowledge:
 227 *  1) if buddy is referenced, it's already initialized
 228 *  2) while block is used in buddy and the buddy is referenced,
 229 *     nobody can re-allocate that block
 230 *  3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
 231 *     bit set and PA claims same block, it's OK. IOW, one can set bit in
 232 *     on-disk bitmap if buddy has same bit set or/and PA covers corresponded
 233 *     block
 234 *
 235 * so, now we're building a concurrency table:
 236 *  - init buddy vs.
 237 *    - new PA
 238 *      blocks for PA are allocated in the buddy, buddy must be referenced
 239 *      until PA is linked to allocation group to avoid concurrent buddy init
 240 *    - use inode PA
 241 *      we need to make sure that either on-disk bitmap or PA has uptodate data
 242 *      given (3) we care that PA-=N operation doesn't interfere with init
 243 *    - discard inode PA
 244 *      the simplest way would be to have buddy initialized by the discard
 245 *    - use locality group PA
 246 *      again PA-=N must be serialized with init
 247 *    - discard locality group PA
 248 *      the simplest way would be to have buddy initialized by the discard
 249 *  - new PA vs.
 250 *    - use inode PA
 251 *      i_data_sem serializes them
 252 *    - discard inode PA
 253 *      discard process must wait until PA isn't used by another process
 254 *    - use locality group PA
 255 *      some mutex should serialize them
 256 *    - discard locality group PA
 257 *      discard process must wait until PA isn't used by another process
 258 *  - use inode PA
 259 *    - use inode PA
 260 *      i_data_sem or another mutex should serializes them
 261 *    - discard inode PA
 262 *      discard process must wait until PA isn't used by another process
 263 *    - use locality group PA
 264 *      nothing wrong here -- they're different PAs covering different blocks
 265 *    - discard locality group PA
 266 *      discard process must wait until PA isn't used by another process
 267 *
 268 * now we're ready to make few consequences:
 269 *  - PA is referenced and while it is no discard is possible
 270 *  - PA is referenced until block isn't marked in on-disk bitmap
 271 *  - PA changes only after on-disk bitmap
 272 *  - discard must not compete with init. either init is done before
 273 *    any discard or they're serialized somehow
 274 *  - buddy init as sum of on-disk bitmap and PAs is done atomically
 275 *
 276 * a special case when we've used PA to emptiness. no need to modify buddy
 277 * in this case, but we should care about concurrent init
 278 *
 279 */
 280
 281 /*
 282 * Logic in few words:
 283 *
 284 *  - allocation:
 285 *    load group
 286 *    find blocks
 287 *    mark bits in on-disk bitmap
 288 *    release group
 289 *
 290 *  - use preallocation:
 291 *    find proper PA (per-inode or group)
 292 *    load group
 293 *    mark bits in on-disk bitmap
 294 *    release group
 295 *    release PA
 296 *
 297 *  - free:
 298 *    load group
 299 *    mark bits in on-disk bitmap
 300 *    release group
 301 *
 302 *  - discard preallocations in group:
 303 *    mark PAs deleted
 304 *    move them onto local list
 305 *    load on-disk bitmap
 306 *    load group
 307 *    remove PA from object (inode or locality group)
 308 *    mark free blocks in-core
 309 *
 310 *  - discard inode's preallocations:
 311 */
 312
 313/*
 314 * Locking rules
 315 *
 316 * Locks:
 317 *  - bitlock on a group	(group)
 318 *  - object (inode/locality)	(object)
 319 *  - per-pa lock		(pa)
 320 *
 321 * Paths:
 322 *  - new pa
 323 *    object
 324 *    group
 325 *
 326 *  - find and use pa:
 327 *    pa
 328 *
 329 *  - release consumed pa:
 330 *    pa
 331 *    group
 332 *    object
 333 *
 334 *  - generate in-core bitmap:
 335 *    group
 336 *        pa
 337 *
 338 *  - discard all for given object (inode, locality group):
 339 *    object
 340 *        pa
 341 *    group
 342 *
 343 *  - discard all for given group:
 344 *    group
 345 *        pa
 346 *    group
 347 *        object
 348 *
 349 */
 350static struct kmem_cache *ext4_pspace_cachep;
 351static struct kmem_cache *ext4_ac_cachep;
 352static struct kmem_cache *ext4_free_data_cachep;
 353
 354/* We create slab caches for groupinfo data structures based on the
 355 * superblock block size.  There will be one per mounted filesystem for
 356 * each unique s_blocksize_bits */
 357#define NR_GRPINFO_CACHES 8
 358static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
 359
 360static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
 361	"ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
 362	"ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
 363	"ext4_groupinfo_64k", "ext4_groupinfo_128k"
 364};
 365
 366static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 367					ext4_group_t group);
 368static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
 369						ext4_group_t group);
 370static void ext4_free_data_callback(struct super_block *sb,
 371				struct ext4_journal_cb_entry *jce, int rc);
 372
 373static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
 374{
 375#if BITS_PER_LONG == 64
 376	*bit += ((unsigned long) addr & 7UL) << 3;
 377	addr = (void *) ((unsigned long) addr & ~7UL);
 378#elif BITS_PER_LONG == 32
 379	*bit += ((unsigned long) addr & 3UL) << 3;
 380	addr = (void *) ((unsigned long) addr & ~3UL);
 381#else
 382#error "how many bits you are?!"
 383#endif
 384	return addr;
 385}
 386
 387static inline int mb_test_bit(int bit, void *addr)
 388{
 389	/*
 390	 * ext4_test_bit on architecture like powerpc
 391	 * needs unsigned long aligned address
 392	 */
 393	addr = mb_correct_addr_and_bit(&bit, addr);
 394	return ext4_test_bit(bit, addr);
 395}
 396
 397static inline void mb_set_bit(int bit, void *addr)
 398{
 399	addr = mb_correct_addr_and_bit(&bit, addr);
 400	ext4_set_bit(bit, addr);
 401}
 402
 403static inline void mb_clear_bit(int bit, void *addr)
 404{
 405	addr = mb_correct_addr_and_bit(&bit, addr);
 406	ext4_clear_bit(bit, addr);
 407}
 408
 409static inline int mb_test_and_clear_bit(int bit, void *addr)
 410{
 411	addr = mb_correct_addr_and_bit(&bit, addr);
 412	return ext4_test_and_clear_bit(bit, addr);
 413}
 414
 415static inline int mb_find_next_zero_bit(void *addr, int max, int start)
 416{
 417	int fix = 0, ret, tmpmax;
 418	addr = mb_correct_addr_and_bit(&fix, addr);
 419	tmpmax = max + fix;
 420	start += fix;
 421
 422	ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
 423	if (ret > max)
 424		return max;
 425	return ret;
 426}
 427
 428static inline int mb_find_next_bit(void *addr, int max, int start)
 429{
 430	int fix = 0, ret, tmpmax;
 431	addr = mb_correct_addr_and_bit(&fix, addr);
 432	tmpmax = max + fix;
 433	start += fix;
 434
 435	ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
 436	if (ret > max)
 437		return max;
 438	return ret;
 439}
 440
 441static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
 442{
 443	char *bb;
 444
 445	BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
 446	BUG_ON(max == NULL);
 447
 448	if (order > e4b->bd_blkbits + 1) {
 449		*max = 0;
 450		return NULL;
 451	}
 452
 453	/* at order 0 we see each particular block */
 454	if (order == 0) {
 455		*max = 1 << (e4b->bd_blkbits + 3);
 456		return e4b->bd_bitmap;
 457	}
 458
 459	bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
 460	*max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
 461
 462	return bb;
 463}
 464
 465#ifdef DOUBLE_CHECK
 466static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
 467			   int first, int count)
 468{
 469	int i;
 470	struct super_block *sb = e4b->bd_sb;
 471
 472	if (unlikely(e4b->bd_info->bb_bitmap == NULL))
 473		return;
 474	assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
 475	for (i = 0; i < count; i++) {
 476		if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
 477			ext4_fsblk_t blocknr;
 478
 479			blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
 480			blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
 481			ext4_grp_locked_error(sb, e4b->bd_group,
 482					      inode ? inode->i_ino : 0,
 483					      blocknr,
 484					      "freeing block already freed "
 485					      "(bit %u)",
 486					      first + i);
 487		}
 488		mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
 489	}
 490}
 491
 492static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
 493{
 494	int i;
 495
 496	if (unlikely(e4b->bd_info->bb_bitmap == NULL))
 497		return;
 498	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
 499	for (i = 0; i < count; i++) {
 500		BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
 501		mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
 502	}
 503}
 504
 505static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
 506{
 507	if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
 508		unsigned char *b1, *b2;
 509		int i;
 510		b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
 511		b2 = (unsigned char *) bitmap;
 512		for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
 513			if (b1[i] != b2[i]) {
 514				ext4_msg(e4b->bd_sb, KERN_ERR,
 515					 "corruption in group %u "
 516					 "at byte %u(%u): %x in copy != %x "
 517					 "on disk/prealloc",
 518					 e4b->bd_group, i, i * 8, b1[i], b2[i]);
 519				BUG();
 520			}
 521		}
 522	}
 523}
 524
 525#else
 526static inline void mb_free_blocks_double(struct inode *inode,
 527				struct ext4_buddy *e4b, int first, int count)
 528{
 529	return;
 530}
 531static inline void mb_mark_used_double(struct ext4_buddy *e4b,
 532						int first, int count)
 533{
 534	return;
 535}
 536static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
 537{
 538	return;
 539}
 540#endif
 541
 542#ifdef AGGRESSIVE_CHECK
 543
 544#define MB_CHECK_ASSERT(assert)						\
 545do {									\
 546	if (!(assert)) {						\
 547		printk(KERN_EMERG					\
 548			"Assertion failure in %s() at %s:%d: \"%s\"\n",	\
 549			function, file, line, # assert);		\
 550		BUG();							\
 551	}								\
 552} while (0)
 553
 554static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
 555				const char *function, int line)
 556{
 557	struct super_block *sb = e4b->bd_sb;
 558	int order = e4b->bd_blkbits + 1;
 559	int max;
 560	int max2;
 561	int i;
 562	int j;
 563	int k;
 564	int count;
 565	struct ext4_group_info *grp;
 566	int fragments = 0;
 567	int fstart;
 568	struct list_head *cur;
 569	void *buddy;
 570	void *buddy2;
 571
 572	{
 573		static int mb_check_counter;
 574		if (mb_check_counter++ % 100 != 0)
 575			return 0;
 576	}
 577
 578	while (order > 1) {
 579		buddy = mb_find_buddy(e4b, order, &max);
 580		MB_CHECK_ASSERT(buddy);
 581		buddy2 = mb_find_buddy(e4b, order - 1, &max2);
 582		MB_CHECK_ASSERT(buddy2);
 583		MB_CHECK_ASSERT(buddy != buddy2);
 584		MB_CHECK_ASSERT(max * 2 == max2);
 585
 586		count = 0;
 587		for (i = 0; i < max; i++) {
 588
 589			if (mb_test_bit(i, buddy)) {
 590				/* only single bit in buddy2 may be 1 */
 591				if (!mb_test_bit(i << 1, buddy2)) {
 592					MB_CHECK_ASSERT(
 593						mb_test_bit((i<<1)+1, buddy2));
 594				} else if (!mb_test_bit((i << 1) + 1, buddy2)) {
 595					MB_CHECK_ASSERT(
 596						mb_test_bit(i << 1, buddy2));
 597				}
 598				continue;
 599			}
 600
 601			/* both bits in buddy2 must be 1 */
 602			MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
 603			MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
 604
 605			for (j = 0; j < (1 << order); j++) {
 606				k = (i * (1 << order)) + j;
 607				MB_CHECK_ASSERT(
 608					!mb_test_bit(k, e4b->bd_bitmap));
 609			}
 610			count++;
 611		}
 612		MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
 613		order--;
 614	}
 615
 616	fstart = -1;
 617	buddy = mb_find_buddy(e4b, 0, &max);
 618	for (i = 0; i < max; i++) {
 619		if (!mb_test_bit(i, buddy)) {
 620			MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
 621			if (fstart == -1) {
 622				fragments++;
 623				fstart = i;
 624			}
 625			continue;
 626		}
 627		fstart = -1;
 628		/* check used bits only */
 629		for (j = 0; j < e4b->bd_blkbits + 1; j++) {
 630			buddy2 = mb_find_buddy(e4b, j, &max2);
 631			k = i >> j;
 632			MB_CHECK_ASSERT(k < max2);
 633			MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
 634		}
 635	}
 636	MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
 637	MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
 638
 639	grp = ext4_get_group_info(sb, e4b->bd_group);
 640	list_for_each(cur, &grp->bb_prealloc_list) {
 641		ext4_group_t groupnr;
 642		struct ext4_prealloc_space *pa;
 643		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
 644		ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
 645		MB_CHECK_ASSERT(groupnr == e4b->bd_group);
 646		for (i = 0; i < pa->pa_len; i++)
 647			MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
 648	}
 649	return 0;
 650}
 651#undef MB_CHECK_ASSERT
 652#define mb_check_buddy(e4b) __mb_check_buddy(e4b,	\
 653					__FILE__, __func__, __LINE__)
 654#else
 655#define mb_check_buddy(e4b)
 656#endif
 657
 658/*
 659 * Divide blocks started from @first with length @len into
 660 * smaller chunks with power of 2 blocks.
 661 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
 662 * then increase bb_counters[] for corresponded chunk size.
 663 */
 664static void ext4_mb_mark_free_simple(struct super_block *sb,
 665				void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
 666					struct ext4_group_info *grp)
 667{
 668	struct ext4_sb_info *sbi = EXT4_SB(sb);
 669	ext4_grpblk_t min;
 670	ext4_grpblk_t max;
 671	ext4_grpblk_t chunk;
 672	unsigned short border;
 673
 674	BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
 675
 676	border = 2 << sb->s_blocksize_bits;
 677
 678	while (len > 0) {
 679		/* find how many blocks can be covered since this position */
 680		max = ffs(first | border) - 1;
 681
 682		/* find how many blocks of power 2 we need to mark */
 683		min = fls(len) - 1;
 684
 685		if (max < min)
 686			min = max;
 687		chunk = 1 << min;
 688
 689		/* mark multiblock chunks only */
 690		grp->bb_counters[min]++;
 691		if (min > 0)
 692			mb_clear_bit(first >> min,
 693				     buddy + sbi->s_mb_offsets[min]);
 694
 695		len -= chunk;
 696		first += chunk;
 697	}
 698}
 699
 700/*
 701 * Cache the order of the largest free extent we have available in this block
 702 * group.
 703 */
 704static void
 705mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
 706{
 707	int i;
 708	int bits;
 709
 710	grp->bb_largest_free_order = -1; /* uninit */
 711
 712	bits = sb->s_blocksize_bits + 1;
 713	for (i = bits; i >= 0; i--) {
 714		if (grp->bb_counters[i] > 0) {
 715			grp->bb_largest_free_order = i;
 716			break;
 717		}
 718	}
 719}
 720
 721static noinline_for_stack
 722void ext4_mb_generate_buddy(struct super_block *sb,
 723				void *buddy, void *bitmap, ext4_group_t group)
 724{
 725	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 726	struct ext4_sb_info *sbi = EXT4_SB(sb);
 727	ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
 728	ext4_grpblk_t i = 0;
 729	ext4_grpblk_t first;
 730	ext4_grpblk_t len;
 731	unsigned free = 0;
 732	unsigned fragments = 0;
 733	unsigned long long period = get_cycles();
 734
 735	/* initialize buddy from bitmap which is aggregation
 736	 * of on-disk bitmap and preallocations */
 737	i = mb_find_next_zero_bit(bitmap, max, 0);
 738	grp->bb_first_free = i;
 739	while (i < max) {
 740		fragments++;
 741		first = i;
 742		i = mb_find_next_bit(bitmap, max, i);
 743		len = i - first;
 744		free += len;
 745		if (len > 1)
 746			ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
 747		else
 748			grp->bb_counters[0]++;
 749		if (i < max)
 750			i = mb_find_next_zero_bit(bitmap, max, i);
 751	}
 752	grp->bb_fragments = fragments;
 753
 754	if (free != grp->bb_free) {
 755		ext4_grp_locked_error(sb, group, 0, 0,
 756				      "block bitmap and bg descriptor "
 757				      "inconsistent: %u vs %u free clusters",
 758				      free, grp->bb_free);
 759		/*
 760		 * If we intend to continue, we consider group descriptor
 761		 * corrupt and update bb_free using bitmap value
 762		 */
 763		grp->bb_free = free;
 764		if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
 765			percpu_counter_sub(&sbi->s_freeclusters_counter,
 766					   grp->bb_free);
 767		set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
 768	}
 769	mb_set_largest_free_order(sb, grp);
 770
 771	clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
 772
 773	period = get_cycles() - period;
 774	spin_lock(&EXT4_SB(sb)->s_bal_lock);
 775	EXT4_SB(sb)->s_mb_buddies_generated++;
 776	EXT4_SB(sb)->s_mb_generation_time += period;
 777	spin_unlock(&EXT4_SB(sb)->s_bal_lock);
 778}
 779
 780static void mb_regenerate_buddy(struct ext4_buddy *e4b)
 781{
 782	int count;
 783	int order = 1;
 784	void *buddy;
 785
 786	while ((buddy = mb_find_buddy(e4b, order++, &count))) {
 787		ext4_set_bits(buddy, 0, count);
 788	}
 789	e4b->bd_info->bb_fragments = 0;
 790	memset(e4b->bd_info->bb_counters, 0,
 791		sizeof(*e4b->bd_info->bb_counters) *
 792		(e4b->bd_sb->s_blocksize_bits + 2));
 793
 794	ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
 795		e4b->bd_bitmap, e4b->bd_group);
 796}
 797
 798/* The buddy information is attached the buddy cache inode
 799 * for convenience. The information regarding each group
 800 * is loaded via ext4_mb_load_buddy. The information involve
 801 * block bitmap and buddy information. The information are
 802 * stored in the inode as
 803 *
 804 * {                        page                        }
 805 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
 806 *
 807 *
 808 * one block each for bitmap and buddy information.
 809 * So for each group we take up 2 blocks. A page can
 810 * contain blocks_per_page (PAGE_SIZE / blocksize)  blocks.
 811 * So it can have information regarding groups_per_page which
 812 * is blocks_per_page/2
 813 *
 814 * Locking note:  This routine takes the block group lock of all groups
 815 * for this page; do not hold this lock when calling this routine!
 816 */
 817
 818static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
 819{
 820	ext4_group_t ngroups;
 821	int blocksize;
 822	int blocks_per_page;
 823	int groups_per_page;
 824	int err = 0;
 825	int i;
 826	ext4_group_t first_group, group;
 827	int first_block;
 828	struct super_block *sb;
 829	struct buffer_head *bhs;
 830	struct buffer_head **bh = NULL;
 831	struct inode *inode;
 832	char *data;
 833	char *bitmap;
 834	struct ext4_group_info *grinfo;
 835
 836	mb_debug(1, "init page %lu\n", page->index);
 837
 838	inode = page->mapping->host;
 839	sb = inode->i_sb;
 840	ngroups = ext4_get_groups_count(sb);
 841	blocksize = 1 << inode->i_blkbits;
 842	blocks_per_page = PAGE_SIZE / blocksize;
 843
 844	groups_per_page = blocks_per_page >> 1;
 845	if (groups_per_page == 0)
 846		groups_per_page = 1;
 847
 848	/* allocate buffer_heads to read bitmaps */
 849	if (groups_per_page > 1) {
 850		i = sizeof(struct buffer_head *) * groups_per_page;
 851		bh = kzalloc(i, gfp);
 852		if (bh == NULL) {
 853			err = -ENOMEM;
 854			goto out;
 855		}
 856	} else
 857		bh = &bhs;
 858
 859	first_group = page->index * blocks_per_page / 2;
 860
 861	/* read all groups the page covers into the cache */
 862	for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
 863		if (group >= ngroups)
 864			break;
 865
 866		grinfo = ext4_get_group_info(sb, group);
 867		/*
 868		 * If page is uptodate then we came here after online resize
 869		 * which added some new uninitialized group info structs, so
 870		 * we must skip all initialized uptodate buddies on the page,
 871		 * which may be currently in use by an allocating task.
 872		 */
 873		if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
 874			bh[i] = NULL;
 875			continue;
 876		}
 877		bh[i] = ext4_read_block_bitmap_nowait(sb, group);
 878		if (IS_ERR(bh[i])) {
 879			err = PTR_ERR(bh[i]);
 880			bh[i] = NULL;
 881			goto out;
 882		}
 883		mb_debug(1, "read bitmap for group %u\n", group);
 884	}
 885
 886	/* wait for I/O completion */
 887	for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
 888		int err2;
 889
 890		if (!bh[i])
 891			continue;
 892		err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
 893		if (!err)
 894			err = err2;
 895	}
 896
 897	first_block = page->index * blocks_per_page;
 898	for (i = 0; i < blocks_per_page; i++) {
 899		group = (first_block + i) >> 1;
 900		if (group >= ngroups)
 901			break;
 902
 903		if (!bh[group - first_group])
 904			/* skip initialized uptodate buddy */
 905			continue;
 906
 907		if (!buffer_verified(bh[group - first_group]))
 908			/* Skip faulty bitmaps */
 909			continue;
 910		err = 0;
 911
 912		/*
 913		 * data carry information regarding this
 914		 * particular group in the format specified
 915		 * above
 916		 *
 917		 */
 918		data = page_address(page) + (i * blocksize);
 919		bitmap = bh[group - first_group]->b_data;
 920
 921		/*
 922		 * We place the buddy block and bitmap block
 923		 * close together
 924		 */
 925		if ((first_block + i) & 1) {
 926			/* this is block of buddy */
 927			BUG_ON(incore == NULL);
 928			mb_debug(1, "put buddy for group %u in page %lu/%x\n",
 929				group, page->index, i * blocksize);
 930			trace_ext4_mb_buddy_bitmap_load(sb, group);
 931			grinfo = ext4_get_group_info(sb, group);
 932			grinfo->bb_fragments = 0;
 933			memset(grinfo->bb_counters, 0,
 934			       sizeof(*grinfo->bb_counters) *
 935				(sb->s_blocksize_bits+2));
 936			/*
 937			 * incore got set to the group block bitmap below
 938			 */
 939			ext4_lock_group(sb, group);
 940			/* init the buddy */
 941			memset(data, 0xff, blocksize);
 942			ext4_mb_generate_buddy(sb, data, incore, group);
 943			ext4_unlock_group(sb, group);
 944			incore = NULL;
 945		} else {
 946			/* this is block of bitmap */
 947			BUG_ON(incore != NULL);
 948			mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
 949				group, page->index, i * blocksize);
 950			trace_ext4_mb_bitmap_load(sb, group);
 951
 952			/* see comments in ext4_mb_put_pa() */
 953			ext4_lock_group(sb, group);
 954			memcpy(data, bitmap, blocksize);
 955
 956			/* mark all preallocated blks used in in-core bitmap */
 957			ext4_mb_generate_from_pa(sb, data, group);
 958			ext4_mb_generate_from_freelist(sb, data, group);
 959			ext4_unlock_group(sb, group);
 960
 961			/* set incore so that the buddy information can be
 962			 * generated using this
 963			 */
 964			incore = data;
 965		}
 966	}
 967	SetPageUptodate(page);
 968
 969out:
 970	if (bh) {
 971		for (i = 0; i < groups_per_page; i++)
 972			brelse(bh[i]);
 973		if (bh != &bhs)
 974			kfree(bh);
 975	}
 976	return err;
 977}
 978
 979/*
 980 * Lock the buddy and bitmap pages. This make sure other parallel init_group
 981 * on the same buddy page doesn't happen whild holding the buddy page lock.
 982 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
 983 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
 984 */
 985static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
 986		ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
 987{
 988	struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
 989	int block, pnum, poff;
 990	int blocks_per_page;
 991	struct page *page;
 992
 993	e4b->bd_buddy_page = NULL;
 994	e4b->bd_bitmap_page = NULL;
 995
 996	blocks_per_page = PAGE_SIZE / sb->s_blocksize;
 997	/*
 998	 * the buddy cache inode stores the block bitmap
 999	 * and buddy information in consecutive blocks.
1000	 * So for each group we need two blocks.
1001	 */
1002	block = group * 2;
1003	pnum = block / blocks_per_page;
1004	poff = block % blocks_per_page;
1005	page = find_or_create_page(inode->i_mapping, pnum, gfp);
1006	if (!page)
1007		return -ENOMEM;
1008	BUG_ON(page->mapping != inode->i_mapping);
1009	e4b->bd_bitmap_page = page;
1010	e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1011
1012	if (blocks_per_page >= 2) {
1013		/* buddy and bitmap are on the same page */
1014		return 0;
1015	}
1016
1017	block++;
1018	pnum = block / blocks_per_page;
1019	page = find_or_create_page(inode->i_mapping, pnum, gfp);
1020	if (!page)
1021		return -ENOMEM;
1022	BUG_ON(page->mapping != inode->i_mapping);
1023	e4b->bd_buddy_page = page;
1024	return 0;
1025}
1026
1027static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1028{
1029	if (e4b->bd_bitmap_page) {
1030		unlock_page(e4b->bd_bitmap_page);
1031		put_page(e4b->bd_bitmap_page);
1032	}
1033	if (e4b->bd_buddy_page) {
1034		unlock_page(e4b->bd_buddy_page);
1035		put_page(e4b->bd_buddy_page);
1036	}
1037}
1038
1039/*
1040 * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1041 * block group lock of all groups for this page; do not hold the BG lock when
1042 * calling this routine!
1043 */
1044static noinline_for_stack
1045int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1046{
1047
1048	struct ext4_group_info *this_grp;
1049	struct ext4_buddy e4b;
1050	struct page *page;
1051	int ret = 0;
1052
1053	might_sleep();
1054	mb_debug(1, "init group %u\n", group);
1055	this_grp = ext4_get_group_info(sb, group);
1056	/*
1057	 * This ensures that we don't reinit the buddy cache
1058	 * page which map to the group from which we are already
1059	 * allocating. If we are looking at the buddy cache we would
1060	 * have taken a reference using ext4_mb_load_buddy and that
1061	 * would have pinned buddy page to page cache.
1062	 * The call to ext4_mb_get_buddy_page_lock will mark the
1063	 * page accessed.
1064	 */
1065	ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1066	if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1067		/*
1068		 * somebody initialized the group
1069		 * return without doing anything
1070		 */
1071		goto err;
1072	}
1073
1074	page = e4b.bd_bitmap_page;
1075	ret = ext4_mb_init_cache(page, NULL, gfp);
1076	if (ret)
1077		goto err;
1078	if (!PageUptodate(page)) {
1079		ret = -EIO;
1080		goto err;
1081	}
1082
1083	if (e4b.bd_buddy_page == NULL) {
1084		/*
1085		 * If both the bitmap and buddy are in
1086		 * the same page we don't need to force
1087		 * init the buddy
1088		 */
1089		ret = 0;
1090		goto err;
1091	}
1092	/* init buddy cache */
1093	page = e4b.bd_buddy_page;
1094	ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1095	if (ret)
1096		goto err;
1097	if (!PageUptodate(page)) {
1098		ret = -EIO;
1099		goto err;
1100	}
1101err:
1102	ext4_mb_put_buddy_page_lock(&e4b);
1103	return ret;
1104}
1105
1106/*
1107 * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1108 * block group lock of all groups for this page; do not hold the BG lock when
1109 * calling this routine!
1110 */
1111static noinline_for_stack int
1112ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1113		       struct ext4_buddy *e4b, gfp_t gfp)
1114{
1115	int blocks_per_page;
1116	int block;
1117	int pnum;
1118	int poff;
1119	struct page *page;
1120	int ret;
1121	struct ext4_group_info *grp;
1122	struct ext4_sb_info *sbi = EXT4_SB(sb);
1123	struct inode *inode = sbi->s_buddy_cache;
1124
1125	might_sleep();
1126	mb_debug(1, "load group %u\n", group);
1127
1128	blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1129	grp = ext4_get_group_info(sb, group);
1130
1131	e4b->bd_blkbits = sb->s_blocksize_bits;
1132	e4b->bd_info = grp;
1133	e4b->bd_sb = sb;
1134	e4b->bd_group = group;
1135	e4b->bd_buddy_page = NULL;
1136	e4b->bd_bitmap_page = NULL;
1137
1138	if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1139		/*
1140		 * we need full data about the group
1141		 * to make a good selection
1142		 */
1143		ret = ext4_mb_init_group(sb, group, gfp);
1144		if (ret)
1145			return ret;
1146	}
1147
1148	/*
1149	 * the buddy cache inode stores the block bitmap
1150	 * and buddy information in consecutive blocks.
1151	 * So for each group we need two blocks.
1152	 */
1153	block = group * 2;
1154	pnum = block / blocks_per_page;
1155	poff = block % blocks_per_page;
1156
1157	/* we could use find_or_create_page(), but it locks page
1158	 * what we'd like to avoid in fast path ... */
1159	page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1160	if (page == NULL || !PageUptodate(page)) {
1161		if (page)
1162			/*
1163			 * drop the page reference and try
1164			 * to get the page with lock. If we
1165			 * are not uptodate that implies
1166			 * somebody just created the page but
1167			 * is yet to initialize the same. So
1168			 * wait for it to initialize.
1169			 */
1170			put_page(page);
1171		page = find_or_create_page(inode->i_mapping, pnum, gfp);
1172		if (page) {
1173			BUG_ON(page->mapping != inode->i_mapping);
1174			if (!PageUptodate(page)) {
1175				ret = ext4_mb_init_cache(page, NULL, gfp);
1176				if (ret) {
1177					unlock_page(page);
1178					goto err;
1179				}
1180				mb_cmp_bitmaps(e4b, page_address(page) +
1181					       (poff * sb->s_blocksize));
1182			}
1183			unlock_page(page);
1184		}
1185	}
1186	if (page == NULL) {
1187		ret = -ENOMEM;
1188		goto err;
1189	}
1190	if (!PageUptodate(page)) {
1191		ret = -EIO;
1192		goto err;
1193	}
1194
1195	/* Pages marked accessed already */
1196	e4b->bd_bitmap_page = page;
1197	e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1198
1199	block++;
1200	pnum = block / blocks_per_page;
1201	poff = block % blocks_per_page;
1202
1203	page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1204	if (page == NULL || !PageUptodate(page)) {
1205		if (page)
1206			put_page(page);
1207		page = find_or_create_page(inode->i_mapping, pnum, gfp);
1208		if (page) {
1209			BUG_ON(page->mapping != inode->i_mapping);
1210			if (!PageUptodate(page)) {
1211				ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1212							 gfp);
1213				if (ret) {
1214					unlock_page(page);
1215					goto err;
1216				}
1217			}
1218			unlock_page(page);
1219		}
1220	}
1221	if (page == NULL) {
1222		ret = -ENOMEM;
1223		goto err;
1224	}
1225	if (!PageUptodate(page)) {
1226		ret = -EIO;
1227		goto err;
1228	}
1229
1230	/* Pages marked accessed already */
1231	e4b->bd_buddy_page = page;
1232	e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1233
1234	BUG_ON(e4b->bd_bitmap_page == NULL);
1235	BUG_ON(e4b->bd_buddy_page == NULL);
1236
1237	return 0;
1238
1239err:
1240	if (page)
1241		put_page(page);
1242	if (e4b->bd_bitmap_page)
1243		put_page(e4b->bd_bitmap_page);
1244	if (e4b->bd_buddy_page)
1245		put_page(e4b->bd_buddy_page);
1246	e4b->bd_buddy = NULL;
1247	e4b->bd_bitmap = NULL;
1248	return ret;
1249}
1250
1251static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1252			      struct ext4_buddy *e4b)
1253{
1254	return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1255}
1256
1257static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1258{
1259	if (e4b->bd_bitmap_page)
1260		put_page(e4b->bd_bitmap_page);
1261	if (e4b->bd_buddy_page)
1262		put_page(e4b->bd_buddy_page);
1263}
1264
1265
1266static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1267{
1268	int order = 1;
 
1269	void *bb;
1270
1271	BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1272	BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1273
1274	bb = e4b->bd_buddy;
1275	while (order <= e4b->bd_blkbits + 1) {
1276		block = block >> 1;
1277		if (!mb_test_bit(block, bb)) {
1278			/* this block is part of buddy of order 'order' */
1279			return order;
1280		}
1281		bb += 1 << (e4b->bd_blkbits - order);
 
1282		order++;
1283	}
1284	return 0;
1285}
1286
1287static void mb_clear_bits(void *bm, int cur, int len)
1288{
1289	__u32 *addr;
1290
1291	len = cur + len;
1292	while (cur < len) {
1293		if ((cur & 31) == 0 && (len - cur) >= 32) {
1294			/* fast path: clear whole word at once */
1295			addr = bm + (cur >> 3);
1296			*addr = 0;
1297			cur += 32;
1298			continue;
1299		}
1300		mb_clear_bit(cur, bm);
1301		cur++;
1302	}
1303}
1304
1305/* clear bits in given range
1306 * will return first found zero bit if any, -1 otherwise
1307 */
1308static int mb_test_and_clear_bits(void *bm, int cur, int len)
1309{
1310	__u32 *addr;
1311	int zero_bit = -1;
1312
1313	len = cur + len;
1314	while (cur < len) {
1315		if ((cur & 31) == 0 && (len - cur) >= 32) {
1316			/* fast path: clear whole word at once */
1317			addr = bm + (cur >> 3);
1318			if (*addr != (__u32)(-1) && zero_bit == -1)
1319				zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1320			*addr = 0;
1321			cur += 32;
1322			continue;
1323		}
1324		if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1325			zero_bit = cur;
1326		cur++;
1327	}
1328
1329	return zero_bit;
1330}
1331
1332void ext4_set_bits(void *bm, int cur, int len)
1333{
1334	__u32 *addr;
1335
1336	len = cur + len;
1337	while (cur < len) {
1338		if ((cur & 31) == 0 && (len - cur) >= 32) {
1339			/* fast path: set whole word at once */
1340			addr = bm + (cur >> 3);
1341			*addr = 0xffffffff;
1342			cur += 32;
1343			continue;
1344		}
1345		mb_set_bit(cur, bm);
1346		cur++;
1347	}
1348}
1349
1350/*
1351 * _________________________________________________________________ */
1352
1353static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1354{
1355	if (mb_test_bit(*bit + side, bitmap)) {
1356		mb_clear_bit(*bit, bitmap);
1357		(*bit) -= side;
1358		return 1;
1359	}
1360	else {
1361		(*bit) += side;
1362		mb_set_bit(*bit, bitmap);
1363		return -1;
1364	}
1365}
1366
1367static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1368{
1369	int max;
1370	int order = 1;
1371	void *buddy = mb_find_buddy(e4b, order, &max);
1372
1373	while (buddy) {
1374		void *buddy2;
1375
1376		/* Bits in range [first; last] are known to be set since
1377		 * corresponding blocks were allocated. Bits in range
1378		 * (first; last) will stay set because they form buddies on
1379		 * upper layer. We just deal with borders if they don't
1380		 * align with upper layer and then go up.
1381		 * Releasing entire group is all about clearing
1382		 * single bit of highest order buddy.
1383		 */
1384
1385		/* Example:
1386		 * ---------------------------------
1387		 * |   1   |   1   |   1   |   1   |
1388		 * ---------------------------------
1389		 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1390		 * ---------------------------------
1391		 *   0   1   2   3   4   5   6   7
1392		 *      \_____________________/
1393		 *
1394		 * Neither [1] nor [6] is aligned to above layer.
1395		 * Left neighbour [0] is free, so mark it busy,
1396		 * decrease bb_counters and extend range to
1397		 * [0; 6]
1398		 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1399		 * mark [6] free, increase bb_counters and shrink range to
1400		 * [0; 5].
1401		 * Then shift range to [0; 2], go up and do the same.
1402		 */
1403
1404
1405		if (first & 1)
1406			e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1407		if (!(last & 1))
1408			e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1409		if (first > last)
1410			break;
1411		order++;
1412
1413		if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1414			mb_clear_bits(buddy, first, last - first + 1);
1415			e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1416			break;
1417		}
1418		first >>= 1;
1419		last >>= 1;
1420		buddy = buddy2;
1421	}
1422}
1423
1424static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1425			   int first, int count)
1426{
1427	int left_is_free = 0;
1428	int right_is_free = 0;
1429	int block;
1430	int last = first + count - 1;
1431	struct super_block *sb = e4b->bd_sb;
1432
1433	if (WARN_ON(count == 0))
1434		return;
1435	BUG_ON(last >= (sb->s_blocksize << 3));
1436	assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1437	/* Don't bother if the block group is corrupt. */
1438	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1439		return;
1440
1441	mb_check_buddy(e4b);
1442	mb_free_blocks_double(inode, e4b, first, count);
1443
1444	e4b->bd_info->bb_free += count;
1445	if (first < e4b->bd_info->bb_first_free)
1446		e4b->bd_info->bb_first_free = first;
1447
1448	/* access memory sequentially: check left neighbour,
1449	 * clear range and then check right neighbour
1450	 */
1451	if (first != 0)
1452		left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1453	block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1454	if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1455		right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1456
1457	if (unlikely(block != -1)) {
1458		struct ext4_sb_info *sbi = EXT4_SB(sb);
1459		ext4_fsblk_t blocknr;
1460
1461		blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1462		blocknr += EXT4_C2B(EXT4_SB(sb), block);
1463		ext4_grp_locked_error(sb, e4b->bd_group,
1464				      inode ? inode->i_ino : 0,
1465				      blocknr,
1466				      "freeing already freed block "
1467				      "(bit %u); block bitmap corrupt.",
1468				      block);
1469		if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1470			percpu_counter_sub(&sbi->s_freeclusters_counter,
1471					   e4b->bd_info->bb_free);
1472		/* Mark the block group as corrupt. */
1473		set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1474			&e4b->bd_info->bb_state);
1475		mb_regenerate_buddy(e4b);
1476		goto done;
1477	}
1478
1479	/* let's maintain fragments counter */
1480	if (left_is_free && right_is_free)
1481		e4b->bd_info->bb_fragments--;
1482	else if (!left_is_free && !right_is_free)
1483		e4b->bd_info->bb_fragments++;
1484
1485	/* buddy[0] == bd_bitmap is a special case, so handle
1486	 * it right away and let mb_buddy_mark_free stay free of
1487	 * zero order checks.
1488	 * Check if neighbours are to be coaleasced,
1489	 * adjust bitmap bb_counters and borders appropriately.
1490	 */
1491	if (first & 1) {
1492		first += !left_is_free;
1493		e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1494	}
1495	if (!(last & 1)) {
1496		last -= !right_is_free;
1497		e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1498	}
1499
1500	if (first <= last)
1501		mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1502
1503done:
1504	mb_set_largest_free_order(sb, e4b->bd_info);
1505	mb_check_buddy(e4b);
1506}
1507
1508static int mb_find_extent(struct ext4_buddy *e4b, int block,
1509				int needed, struct ext4_free_extent *ex)
1510{
1511	int next = block;
1512	int max, order;
1513	void *buddy;
1514
1515	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1516	BUG_ON(ex == NULL);
1517
1518	buddy = mb_find_buddy(e4b, 0, &max);
1519	BUG_ON(buddy == NULL);
1520	BUG_ON(block >= max);
1521	if (mb_test_bit(block, buddy)) {
1522		ex->fe_len = 0;
1523		ex->fe_start = 0;
1524		ex->fe_group = 0;
1525		return 0;
1526	}
1527
1528	/* find actual order */
1529	order = mb_find_order_for_block(e4b, block);
1530	block = block >> order;
1531
1532	ex->fe_len = 1 << order;
1533	ex->fe_start = block << order;
1534	ex->fe_group = e4b->bd_group;
1535
1536	/* calc difference from given start */
1537	next = next - ex->fe_start;
1538	ex->fe_len -= next;
1539	ex->fe_start += next;
1540
1541	while (needed > ex->fe_len &&
1542	       mb_find_buddy(e4b, order, &max)) {
1543
1544		if (block + 1 >= max)
1545			break;
1546
1547		next = (block + 1) * (1 << order);
1548		if (mb_test_bit(next, e4b->bd_bitmap))
1549			break;
1550
1551		order = mb_find_order_for_block(e4b, next);
1552
1553		block = next >> order;
1554		ex->fe_len += 1 << order;
1555	}
1556
1557	BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1558	return ex->fe_len;
1559}
1560
1561static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1562{
1563	int ord;
1564	int mlen = 0;
1565	int max = 0;
1566	int cur;
1567	int start = ex->fe_start;
1568	int len = ex->fe_len;
1569	unsigned ret = 0;
1570	int len0 = len;
1571	void *buddy;
1572
1573	BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1574	BUG_ON(e4b->bd_group != ex->fe_group);
1575	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1576	mb_check_buddy(e4b);
1577	mb_mark_used_double(e4b, start, len);
1578
1579	e4b->bd_info->bb_free -= len;
1580	if (e4b->bd_info->bb_first_free == start)
1581		e4b->bd_info->bb_first_free += len;
1582
1583	/* let's maintain fragments counter */
1584	if (start != 0)
1585		mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1586	if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1587		max = !mb_test_bit(start + len, e4b->bd_bitmap);
1588	if (mlen && max)
1589		e4b->bd_info->bb_fragments++;
1590	else if (!mlen && !max)
1591		e4b->bd_info->bb_fragments--;
1592
1593	/* let's maintain buddy itself */
1594	while (len) {
1595		ord = mb_find_order_for_block(e4b, start);
1596
1597		if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1598			/* the whole chunk may be allocated at once! */
1599			mlen = 1 << ord;
1600			buddy = mb_find_buddy(e4b, ord, &max);
1601			BUG_ON((start >> ord) >= max);
1602			mb_set_bit(start >> ord, buddy);
1603			e4b->bd_info->bb_counters[ord]--;
1604			start += mlen;
1605			len -= mlen;
1606			BUG_ON(len < 0);
1607			continue;
1608		}
1609
1610		/* store for history */
1611		if (ret == 0)
1612			ret = len | (ord << 16);
1613
1614		/* we have to split large buddy */
1615		BUG_ON(ord <= 0);
1616		buddy = mb_find_buddy(e4b, ord, &max);
1617		mb_set_bit(start >> ord, buddy);
1618		e4b->bd_info->bb_counters[ord]--;
1619
1620		ord--;
1621		cur = (start >> ord) & ~1U;
1622		buddy = mb_find_buddy(e4b, ord, &max);
1623		mb_clear_bit(cur, buddy);
1624		mb_clear_bit(cur + 1, buddy);
1625		e4b->bd_info->bb_counters[ord]++;
1626		e4b->bd_info->bb_counters[ord]++;
1627	}
1628	mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1629
1630	ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1631	mb_check_buddy(e4b);
1632
1633	return ret;
1634}
1635
1636/*
1637 * Must be called under group lock!
1638 */
1639static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1640					struct ext4_buddy *e4b)
1641{
1642	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1643	int ret;
1644
1645	BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1646	BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1647
1648	ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1649	ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1650	ret = mb_mark_used(e4b, &ac->ac_b_ex);
1651
1652	/* preallocation can change ac_b_ex, thus we store actually
1653	 * allocated blocks for history */
1654	ac->ac_f_ex = ac->ac_b_ex;
1655
1656	ac->ac_status = AC_STATUS_FOUND;
1657	ac->ac_tail = ret & 0xffff;
1658	ac->ac_buddy = ret >> 16;
1659
1660	/*
1661	 * take the page reference. We want the page to be pinned
1662	 * so that we don't get a ext4_mb_init_cache_call for this
1663	 * group until we update the bitmap. That would mean we
1664	 * double allocate blocks. The reference is dropped
1665	 * in ext4_mb_release_context
1666	 */
1667	ac->ac_bitmap_page = e4b->bd_bitmap_page;
1668	get_page(ac->ac_bitmap_page);
1669	ac->ac_buddy_page = e4b->bd_buddy_page;
1670	get_page(ac->ac_buddy_page);
1671	/* store last allocated for subsequent stream allocation */
1672	if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1673		spin_lock(&sbi->s_md_lock);
1674		sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1675		sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1676		spin_unlock(&sbi->s_md_lock);
1677	}
1678}
1679
1680/*
1681 * regular allocator, for general purposes allocation
1682 */
1683
1684static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1685					struct ext4_buddy *e4b,
1686					int finish_group)
1687{
1688	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1689	struct ext4_free_extent *bex = &ac->ac_b_ex;
1690	struct ext4_free_extent *gex = &ac->ac_g_ex;
1691	struct ext4_free_extent ex;
1692	int max;
1693
1694	if (ac->ac_status == AC_STATUS_FOUND)
1695		return;
1696	/*
1697	 * We don't want to scan for a whole year
1698	 */
1699	if (ac->ac_found > sbi->s_mb_max_to_scan &&
1700			!(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1701		ac->ac_status = AC_STATUS_BREAK;
1702		return;
1703	}
1704
1705	/*
1706	 * Haven't found good chunk so far, let's continue
1707	 */
1708	if (bex->fe_len < gex->fe_len)
1709		return;
1710
1711	if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1712			&& bex->fe_group == e4b->bd_group) {
1713		/* recheck chunk's availability - we don't know
1714		 * when it was found (within this lock-unlock
1715		 * period or not) */
1716		max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1717		if (max >= gex->fe_len) {
1718			ext4_mb_use_best_found(ac, e4b);
1719			return;
1720		}
1721	}
1722}
1723
1724/*
1725 * The routine checks whether found extent is good enough. If it is,
1726 * then the extent gets marked used and flag is set to the context
1727 * to stop scanning. Otherwise, the extent is compared with the
1728 * previous found extent and if new one is better, then it's stored
1729 * in the context. Later, the best found extent will be used, if
1730 * mballoc can't find good enough extent.
1731 *
1732 * FIXME: real allocation policy is to be designed yet!
1733 */
1734static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1735					struct ext4_free_extent *ex,
1736					struct ext4_buddy *e4b)
1737{
1738	struct ext4_free_extent *bex = &ac->ac_b_ex;
1739	struct ext4_free_extent *gex = &ac->ac_g_ex;
1740
1741	BUG_ON(ex->fe_len <= 0);
1742	BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1743	BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1744	BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1745
1746	ac->ac_found++;
1747
1748	/*
1749	 * The special case - take what you catch first
1750	 */
1751	if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1752		*bex = *ex;
1753		ext4_mb_use_best_found(ac, e4b);
1754		return;
1755	}
1756
1757	/*
1758	 * Let's check whether the chuck is good enough
1759	 */
1760	if (ex->fe_len == gex->fe_len) {
1761		*bex = *ex;
1762		ext4_mb_use_best_found(ac, e4b);
1763		return;
1764	}
1765
1766	/*
1767	 * If this is first found extent, just store it in the context
1768	 */
1769	if (bex->fe_len == 0) {
1770		*bex = *ex;
1771		return;
1772	}
1773
1774	/*
1775	 * If new found extent is better, store it in the context
1776	 */
1777	if (bex->fe_len < gex->fe_len) {
1778		/* if the request isn't satisfied, any found extent
1779		 * larger than previous best one is better */
1780		if (ex->fe_len > bex->fe_len)
1781			*bex = *ex;
1782	} else if (ex->fe_len > gex->fe_len) {
1783		/* if the request is satisfied, then we try to find
1784		 * an extent that still satisfy the request, but is
1785		 * smaller than previous one */
1786		if (ex->fe_len < bex->fe_len)
1787			*bex = *ex;
1788	}
1789
1790	ext4_mb_check_limits(ac, e4b, 0);
1791}
1792
1793static noinline_for_stack
1794int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1795					struct ext4_buddy *e4b)
1796{
1797	struct ext4_free_extent ex = ac->ac_b_ex;
1798	ext4_group_t group = ex.fe_group;
1799	int max;
1800	int err;
1801
1802	BUG_ON(ex.fe_len <= 0);
1803	err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1804	if (err)
1805		return err;
1806
1807	ext4_lock_group(ac->ac_sb, group);
1808	max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1809
1810	if (max > 0) {
1811		ac->ac_b_ex = ex;
1812		ext4_mb_use_best_found(ac, e4b);
1813	}
1814
1815	ext4_unlock_group(ac->ac_sb, group);
1816	ext4_mb_unload_buddy(e4b);
1817
1818	return 0;
1819}
1820
1821static noinline_for_stack
1822int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1823				struct ext4_buddy *e4b)
1824{
1825	ext4_group_t group = ac->ac_g_ex.fe_group;
1826	int max;
1827	int err;
1828	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1829	struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1830	struct ext4_free_extent ex;
1831
1832	if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1833		return 0;
1834	if (grp->bb_free == 0)
1835		return 0;
1836
1837	err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1838	if (err)
1839		return err;
1840
1841	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1842		ext4_mb_unload_buddy(e4b);
1843		return 0;
1844	}
1845
1846	ext4_lock_group(ac->ac_sb, group);
1847	max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1848			     ac->ac_g_ex.fe_len, &ex);
1849	ex.fe_logical = 0xDEADFA11; /* debug value */
1850
1851	if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1852		ext4_fsblk_t start;
1853
1854		start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1855			ex.fe_start;
1856		/* use do_div to get remainder (would be 64-bit modulo) */
1857		if (do_div(start, sbi->s_stripe) == 0) {
1858			ac->ac_found++;
1859			ac->ac_b_ex = ex;
1860			ext4_mb_use_best_found(ac, e4b);
1861		}
1862	} else if (max >= ac->ac_g_ex.fe_len) {
1863		BUG_ON(ex.fe_len <= 0);
1864		BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1865		BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1866		ac->ac_found++;
1867		ac->ac_b_ex = ex;
1868		ext4_mb_use_best_found(ac, e4b);
1869	} else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1870		/* Sometimes, caller may want to merge even small
1871		 * number of blocks to an existing extent */
1872		BUG_ON(ex.fe_len <= 0);
1873		BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1874		BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1875		ac->ac_found++;
1876		ac->ac_b_ex = ex;
1877		ext4_mb_use_best_found(ac, e4b);
1878	}
1879	ext4_unlock_group(ac->ac_sb, group);
1880	ext4_mb_unload_buddy(e4b);
1881
1882	return 0;
1883}
1884
1885/*
1886 * The routine scans buddy structures (not bitmap!) from given order
1887 * to max order and tries to find big enough chunk to satisfy the req
1888 */
1889static noinline_for_stack
1890void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1891					struct ext4_buddy *e4b)
1892{
1893	struct super_block *sb = ac->ac_sb;
1894	struct ext4_group_info *grp = e4b->bd_info;
1895	void *buddy;
1896	int i;
1897	int k;
1898	int max;
1899
1900	BUG_ON(ac->ac_2order <= 0);
1901	for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1902		if (grp->bb_counters[i] == 0)
1903			continue;
1904
1905		buddy = mb_find_buddy(e4b, i, &max);
1906		BUG_ON(buddy == NULL);
1907
1908		k = mb_find_next_zero_bit(buddy, max, 0);
1909		BUG_ON(k >= max);
1910
1911		ac->ac_found++;
1912
1913		ac->ac_b_ex.fe_len = 1 << i;
1914		ac->ac_b_ex.fe_start = k << i;
1915		ac->ac_b_ex.fe_group = e4b->bd_group;
1916
1917		ext4_mb_use_best_found(ac, e4b);
1918
1919		BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1920
1921		if (EXT4_SB(sb)->s_mb_stats)
1922			atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1923
1924		break;
1925	}
1926}
1927
1928/*
1929 * The routine scans the group and measures all found extents.
1930 * In order to optimize scanning, caller must pass number of
1931 * free blocks in the group, so the routine can know upper limit.
1932 */
1933static noinline_for_stack
1934void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1935					struct ext4_buddy *e4b)
1936{
1937	struct super_block *sb = ac->ac_sb;
1938	void *bitmap = e4b->bd_bitmap;
1939	struct ext4_free_extent ex;
1940	int i;
1941	int free;
1942
1943	free = e4b->bd_info->bb_free;
1944	BUG_ON(free <= 0);
1945
1946	i = e4b->bd_info->bb_first_free;
1947
1948	while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1949		i = mb_find_next_zero_bit(bitmap,
1950						EXT4_CLUSTERS_PER_GROUP(sb), i);
1951		if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1952			/*
1953			 * IF we have corrupt bitmap, we won't find any
1954			 * free blocks even though group info says we
1955			 * we have free blocks
1956			 */
1957			ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1958					"%d free clusters as per "
1959					"group info. But bitmap says 0",
1960					free);
1961			break;
1962		}
1963
1964		mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1965		BUG_ON(ex.fe_len <= 0);
1966		if (free < ex.fe_len) {
1967			ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1968					"%d free clusters as per "
1969					"group info. But got %d blocks",
1970					free, ex.fe_len);
1971			/*
1972			 * The number of free blocks differs. This mostly
1973			 * indicate that the bitmap is corrupt. So exit
1974			 * without claiming the space.
1975			 */
1976			break;
1977		}
1978		ex.fe_logical = 0xDEADC0DE; /* debug value */
1979		ext4_mb_measure_extent(ac, &ex, e4b);
1980
1981		i += ex.fe_len;
1982		free -= ex.fe_len;
1983	}
1984
1985	ext4_mb_check_limits(ac, e4b, 1);
1986}
1987
1988/*
1989 * This is a special case for storages like raid5
1990 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1991 */
1992static noinline_for_stack
1993void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1994				 struct ext4_buddy *e4b)
1995{
1996	struct super_block *sb = ac->ac_sb;
1997	struct ext4_sb_info *sbi = EXT4_SB(sb);
1998	void *bitmap = e4b->bd_bitmap;
1999	struct ext4_free_extent ex;
2000	ext4_fsblk_t first_group_block;
2001	ext4_fsblk_t a;
2002	ext4_grpblk_t i;
2003	int max;
2004
2005	BUG_ON(sbi->s_stripe == 0);
2006
2007	/* find first stripe-aligned block in group */
2008	first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2009
2010	a = first_group_block + sbi->s_stripe - 1;
2011	do_div(a, sbi->s_stripe);
2012	i = (a * sbi->s_stripe) - first_group_block;
2013
2014	while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2015		if (!mb_test_bit(i, bitmap)) {
2016			max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2017			if (max >= sbi->s_stripe) {
2018				ac->ac_found++;
2019				ex.fe_logical = 0xDEADF00D; /* debug value */
2020				ac->ac_b_ex = ex;
2021				ext4_mb_use_best_found(ac, e4b);
2022				break;
2023			}
2024		}
2025		i += sbi->s_stripe;
2026	}
2027}
2028
2029/*
2030 * This is now called BEFORE we load the buddy bitmap.
2031 * Returns either 1 or 0 indicating that the group is either suitable
2032 * for the allocation or not. In addition it can also return negative
2033 * error code when something goes wrong.
2034 */
2035static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2036				ext4_group_t group, int cr)
2037{
2038	unsigned free, fragments;
2039	int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2040	struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2041
2042	BUG_ON(cr < 0 || cr >= 4);
2043
2044	free = grp->bb_free;
2045	if (free == 0)
2046		return 0;
2047	if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2048		return 0;
2049
2050	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2051		return 0;
2052
2053	/* We only do this if the grp has never been initialized */
2054	if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2055		int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS);
2056		if (ret)
2057			return ret;
2058	}
2059
2060	fragments = grp->bb_fragments;
2061	if (fragments == 0)
2062		return 0;
2063
2064	switch (cr) {
2065	case 0:
2066		BUG_ON(ac->ac_2order == 0);
2067
2068		/* Avoid using the first bg of a flexgroup for data files */
2069		if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2070		    (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2071		    ((group % flex_size) == 0))
2072			return 0;
2073
2074		if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2075		    (free / fragments) >= ac->ac_g_ex.fe_len)
2076			return 1;
2077
2078		if (grp->bb_largest_free_order < ac->ac_2order)
2079			return 0;
2080
2081		return 1;
2082	case 1:
2083		if ((free / fragments) >= ac->ac_g_ex.fe_len)
2084			return 1;
2085		break;
2086	case 2:
2087		if (free >= ac->ac_g_ex.fe_len)
2088			return 1;
2089		break;
2090	case 3:
2091		return 1;
2092	default:
2093		BUG();
2094	}
2095
2096	return 0;
2097}
2098
2099static noinline_for_stack int
2100ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2101{
2102	ext4_group_t ngroups, group, i;
2103	int cr;
2104	int err = 0, first_err = 0;
2105	struct ext4_sb_info *sbi;
2106	struct super_block *sb;
2107	struct ext4_buddy e4b;
2108
2109	sb = ac->ac_sb;
2110	sbi = EXT4_SB(sb);
2111	ngroups = ext4_get_groups_count(sb);
2112	/* non-extent files are limited to low blocks/groups */
2113	if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2114		ngroups = sbi->s_blockfile_groups;
2115
2116	BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2117
2118	/* first, try the goal */
2119	err = ext4_mb_find_by_goal(ac, &e4b);
2120	if (err || ac->ac_status == AC_STATUS_FOUND)
2121		goto out;
2122
2123	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2124		goto out;
2125
2126	/*
2127	 * ac->ac2_order is set only if the fe_len is a power of 2
2128	 * if ac2_order is set we also set criteria to 0 so that we
2129	 * try exact allocation using buddy.
2130	 */
2131	i = fls(ac->ac_g_ex.fe_len);
2132	ac->ac_2order = 0;
2133	/*
2134	 * We search using buddy data only if the order of the request
2135	 * is greater than equal to the sbi_s_mb_order2_reqs
2136	 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2137	 */
2138	if (i >= sbi->s_mb_order2_reqs) {
2139		/*
2140		 * This should tell if fe_len is exactly power of 2
2141		 */
2142		if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2143			ac->ac_2order = i - 1;
2144	}
2145
2146	/* if stream allocation is enabled, use global goal */
2147	if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2148		/* TBD: may be hot point */
2149		spin_lock(&sbi->s_md_lock);
2150		ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2151		ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2152		spin_unlock(&sbi->s_md_lock);
2153	}
2154
2155	/* Let's just scan groups to find more-less suitable blocks */
2156	cr = ac->ac_2order ? 0 : 1;
2157	/*
2158	 * cr == 0 try to get exact allocation,
2159	 * cr == 3  try to get anything
2160	 */
2161repeat:
2162	for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2163		ac->ac_criteria = cr;
2164		/*
2165		 * searching for the right group start
2166		 * from the goal value specified
2167		 */
2168		group = ac->ac_g_ex.fe_group;
2169
2170		for (i = 0; i < ngroups; group++, i++) {
2171			int ret = 0;
2172			cond_resched();
2173			/*
2174			 * Artificially restricted ngroups for non-extent
2175			 * files makes group > ngroups possible on first loop.
2176			 */
2177			if (group >= ngroups)
2178				group = 0;
2179
2180			/* This now checks without needing the buddy page */
2181			ret = ext4_mb_good_group(ac, group, cr);
2182			if (ret <= 0) {
2183				if (!first_err)
2184					first_err = ret;
2185				continue;
2186			}
2187
2188			err = ext4_mb_load_buddy(sb, group, &e4b);
2189			if (err)
2190				goto out;
2191
2192			ext4_lock_group(sb, group);
2193
2194			/*
2195			 * We need to check again after locking the
2196			 * block group
2197			 */
2198			ret = ext4_mb_good_group(ac, group, cr);
2199			if (ret <= 0) {
2200				ext4_unlock_group(sb, group);
2201				ext4_mb_unload_buddy(&e4b);
2202				if (!first_err)
2203					first_err = ret;
2204				continue;
2205			}
2206
2207			ac->ac_groups_scanned++;
2208			if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2209				ext4_mb_simple_scan_group(ac, &e4b);
2210			else if (cr == 1 && sbi->s_stripe &&
2211					!(ac->ac_g_ex.fe_len % sbi->s_stripe))
2212				ext4_mb_scan_aligned(ac, &e4b);
2213			else
2214				ext4_mb_complex_scan_group(ac, &e4b);
2215
2216			ext4_unlock_group(sb, group);
2217			ext4_mb_unload_buddy(&e4b);
2218
2219			if (ac->ac_status != AC_STATUS_CONTINUE)
2220				break;
2221		}
2222	}
2223
2224	if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2225	    !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2226		/*
2227		 * We've been searching too long. Let's try to allocate
2228		 * the best chunk we've found so far
2229		 */
2230
2231		ext4_mb_try_best_found(ac, &e4b);
2232		if (ac->ac_status != AC_STATUS_FOUND) {
2233			/*
2234			 * Someone more lucky has already allocated it.
2235			 * The only thing we can do is just take first
2236			 * found block(s)
2237			printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2238			 */
2239			ac->ac_b_ex.fe_group = 0;
2240			ac->ac_b_ex.fe_start = 0;
2241			ac->ac_b_ex.fe_len = 0;
2242			ac->ac_status = AC_STATUS_CONTINUE;
2243			ac->ac_flags |= EXT4_MB_HINT_FIRST;
2244			cr = 3;
2245			atomic_inc(&sbi->s_mb_lost_chunks);
2246			goto repeat;
2247		}
2248	}
2249out:
2250	if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2251		err = first_err;
2252	return err;
2253}
2254
2255static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2256{
2257	struct super_block *sb = seq->private;
2258	ext4_group_t group;
2259
2260	if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2261		return NULL;
2262	group = *pos + 1;
2263	return (void *) ((unsigned long) group);
2264}
2265
2266static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2267{
2268	struct super_block *sb = seq->private;
2269	ext4_group_t group;
2270
2271	++*pos;
2272	if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2273		return NULL;
2274	group = *pos + 1;
2275	return (void *) ((unsigned long) group);
2276}
2277
2278static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2279{
2280	struct super_block *sb = seq->private;
2281	ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2282	int i;
2283	int err, buddy_loaded = 0;
2284	struct ext4_buddy e4b;
2285	struct ext4_group_info *grinfo;
2286	struct sg {
2287		struct ext4_group_info info;
2288		ext4_grpblk_t counters[16];
2289	} sg;
2290
2291	group--;
2292	if (group == 0)
2293		seq_puts(seq, "#group: free  frags first ["
2294			      " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
2295			      " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
2296
2297	i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2298		sizeof(struct ext4_group_info);
2299	grinfo = ext4_get_group_info(sb, group);
2300	/* Load the group info in memory only if not already loaded. */
2301	if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2302		err = ext4_mb_load_buddy(sb, group, &e4b);
2303		if (err) {
2304			seq_printf(seq, "#%-5u: I/O error\n", group);
2305			return 0;
2306		}
2307		buddy_loaded = 1;
2308	}
2309
2310	memcpy(&sg, ext4_get_group_info(sb, group), i);
2311
2312	if (buddy_loaded)
2313		ext4_mb_unload_buddy(&e4b);
2314
2315	seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2316			sg.info.bb_fragments, sg.info.bb_first_free);
2317	for (i = 0; i <= 13; i++)
2318		seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2319				sg.info.bb_counters[i] : 0);
2320	seq_printf(seq, " ]\n");
2321
2322	return 0;
2323}
2324
2325static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2326{
2327}
2328
2329static const struct seq_operations ext4_mb_seq_groups_ops = {
2330	.start  = ext4_mb_seq_groups_start,
2331	.next   = ext4_mb_seq_groups_next,
2332	.stop   = ext4_mb_seq_groups_stop,
2333	.show   = ext4_mb_seq_groups_show,
2334};
2335
2336static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2337{
2338	struct super_block *sb = PDE_DATA(inode);
2339	int rc;
2340
2341	rc = seq_open(file, &ext4_mb_seq_groups_ops);
2342	if (rc == 0) {
2343		struct seq_file *m = file->private_data;
2344		m->private = sb;
2345	}
2346	return rc;
2347
2348}
2349
2350const struct file_operations ext4_seq_mb_groups_fops = {
2351	.owner		= THIS_MODULE,
2352	.open		= ext4_mb_seq_groups_open,
2353	.read		= seq_read,
2354	.llseek		= seq_lseek,
2355	.release	= seq_release,
2356};
2357
2358static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2359{
2360	int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2361	struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2362
2363	BUG_ON(!cachep);
2364	return cachep;
2365}
2366
2367/*
2368 * Allocate the top-level s_group_info array for the specified number
2369 * of groups
2370 */
2371int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2372{
2373	struct ext4_sb_info *sbi = EXT4_SB(sb);
2374	unsigned size;
2375	struct ext4_group_info ***new_groupinfo;
2376
2377	size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2378		EXT4_DESC_PER_BLOCK_BITS(sb);
2379	if (size <= sbi->s_group_info_size)
2380		return 0;
2381
2382	size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2383	new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2384	if (!new_groupinfo) {
2385		ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2386		return -ENOMEM;
2387	}
2388	if (sbi->s_group_info) {
2389		memcpy(new_groupinfo, sbi->s_group_info,
2390		       sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2391		kvfree(sbi->s_group_info);
2392	}
2393	sbi->s_group_info = new_groupinfo;
2394	sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2395	ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 
2396		   sbi->s_group_info_size);
2397	return 0;
2398}
2399
2400/* Create and initialize ext4_group_info data for the given group. */
2401int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2402			  struct ext4_group_desc *desc)
2403{
2404	int i;
2405	int metalen = 0;
2406	struct ext4_sb_info *sbi = EXT4_SB(sb);
2407	struct ext4_group_info **meta_group_info;
2408	struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2409
2410	/*
2411	 * First check if this group is the first of a reserved block.
2412	 * If it's true, we have to allocate a new table of pointers
2413	 * to ext4_group_info structures
2414	 */
2415	if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2416		metalen = sizeof(*meta_group_info) <<
2417			EXT4_DESC_PER_BLOCK_BITS(sb);
2418		meta_group_info = kmalloc(metalen, GFP_NOFS);
2419		if (meta_group_info == NULL) {
2420			ext4_msg(sb, KERN_ERR, "can't allocate mem "
2421				 "for a buddy group");
2422			goto exit_meta_group_info;
2423		}
2424		sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2425			meta_group_info;
2426	}
2427
2428	meta_group_info =
2429		sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2430	i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2431
2432	meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2433	if (meta_group_info[i] == NULL) {
2434		ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2435		goto exit_group_info;
2436	}
2437	set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2438		&(meta_group_info[i]->bb_state));
2439
2440	/*
2441	 * initialize bb_free to be able to skip
2442	 * empty groups without initialization
2443	 */
2444	if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2445		meta_group_info[i]->bb_free =
2446			ext4_free_clusters_after_init(sb, group, desc);
2447	} else {
2448		meta_group_info[i]->bb_free =
2449			ext4_free_group_clusters(sb, desc);
2450	}
2451
2452	INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2453	init_rwsem(&meta_group_info[i]->alloc_sem);
2454	meta_group_info[i]->bb_free_root = RB_ROOT;
2455	meta_group_info[i]->bb_largest_free_order = -1;  /* uninit */
2456
2457#ifdef DOUBLE_CHECK
2458	{
2459		struct buffer_head *bh;
2460		meta_group_info[i]->bb_bitmap =
2461			kmalloc(sb->s_blocksize, GFP_NOFS);
2462		BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2463		bh = ext4_read_block_bitmap(sb, group);
2464		BUG_ON(IS_ERR_OR_NULL(bh));
2465		memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2466			sb->s_blocksize);
2467		put_bh(bh);
2468	}
2469#endif
2470
2471	return 0;
2472
2473exit_group_info:
2474	/* If a meta_group_info table has been allocated, release it now */
2475	if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2476		kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2477		sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2478	}
2479exit_meta_group_info:
2480	return -ENOMEM;
2481} /* ext4_mb_add_groupinfo */
2482
2483static int ext4_mb_init_backend(struct super_block *sb)
2484{
2485	ext4_group_t ngroups = ext4_get_groups_count(sb);
2486	ext4_group_t i;
2487	struct ext4_sb_info *sbi = EXT4_SB(sb);
2488	int err;
2489	struct ext4_group_desc *desc;
2490	struct kmem_cache *cachep;
2491
2492	err = ext4_mb_alloc_groupinfo(sb, ngroups);
2493	if (err)
2494		return err;
2495
2496	sbi->s_buddy_cache = new_inode(sb);
2497	if (sbi->s_buddy_cache == NULL) {
2498		ext4_msg(sb, KERN_ERR, "can't get new inode");
2499		goto err_freesgi;
2500	}
2501	/* To avoid potentially colliding with an valid on-disk inode number,
2502	 * use EXT4_BAD_INO for the buddy cache inode number.  This inode is
2503	 * not in the inode hash, so it should never be found by iget(), but
2504	 * this will avoid confusion if it ever shows up during debugging. */
2505	sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2506	EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2507	for (i = 0; i < ngroups; i++) {
2508		desc = ext4_get_group_desc(sb, i, NULL);
2509		if (desc == NULL) {
2510			ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2511			goto err_freebuddy;
2512		}
2513		if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2514			goto err_freebuddy;
2515	}
2516
2517	return 0;
2518
2519err_freebuddy:
2520	cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2521	while (i-- > 0)
2522		kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2523	i = sbi->s_group_info_size;
2524	while (i-- > 0)
2525		kfree(sbi->s_group_info[i]);
2526	iput(sbi->s_buddy_cache);
2527err_freesgi:
2528	kvfree(sbi->s_group_info);
2529	return -ENOMEM;
2530}
2531
2532static void ext4_groupinfo_destroy_slabs(void)
2533{
2534	int i;
2535
2536	for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2537		if (ext4_groupinfo_caches[i])
2538			kmem_cache_destroy(ext4_groupinfo_caches[i]);
2539		ext4_groupinfo_caches[i] = NULL;
2540	}
2541}
2542
2543static int ext4_groupinfo_create_slab(size_t size)
2544{
2545	static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2546	int slab_size;
2547	int blocksize_bits = order_base_2(size);
2548	int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2549	struct kmem_cache *cachep;
2550
2551	if (cache_index >= NR_GRPINFO_CACHES)
2552		return -EINVAL;
2553
2554	if (unlikely(cache_index < 0))
2555		cache_index = 0;
2556
2557	mutex_lock(&ext4_grpinfo_slab_create_mutex);
2558	if (ext4_groupinfo_caches[cache_index]) {
2559		mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2560		return 0;	/* Already created */
2561	}
2562
2563	slab_size = offsetof(struct ext4_group_info,
2564				bb_counters[blocksize_bits + 2]);
2565
2566	cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2567					slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2568					NULL);
2569
2570	ext4_groupinfo_caches[cache_index] = cachep;
2571
2572	mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2573	if (!cachep) {
2574		printk(KERN_EMERG
2575		       "EXT4-fs: no memory for groupinfo slab cache\n");
2576		return -ENOMEM;
2577	}
2578
2579	return 0;
2580}
2581
2582int ext4_mb_init(struct super_block *sb)
2583{
2584	struct ext4_sb_info *sbi = EXT4_SB(sb);
2585	unsigned i, j;
2586	unsigned offset;
2587	unsigned max;
2588	int ret;
2589
2590	i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2591
2592	sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2593	if (sbi->s_mb_offsets == NULL) {
2594		ret = -ENOMEM;
2595		goto out;
2596	}
2597
2598	i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2599	sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2600	if (sbi->s_mb_maxs == NULL) {
2601		ret = -ENOMEM;
2602		goto out;
2603	}
2604
2605	ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2606	if (ret < 0)
2607		goto out;
2608
2609	/* order 0 is regular bitmap */
2610	sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2611	sbi->s_mb_offsets[0] = 0;
2612
2613	i = 1;
2614	offset = 0;
 
2615	max = sb->s_blocksize << 2;
2616	do {
2617		sbi->s_mb_offsets[i] = offset;
2618		sbi->s_mb_maxs[i] = max;
2619		offset += 1 << (sb->s_blocksize_bits - i);
 
2620		max = max >> 1;
2621		i++;
2622	} while (i <= sb->s_blocksize_bits + 1);
2623
2624	spin_lock_init(&sbi->s_md_lock);
2625	spin_lock_init(&sbi->s_bal_lock);
 
2626
2627	sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2628	sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2629	sbi->s_mb_stats = MB_DEFAULT_STATS;
2630	sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2631	sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2632	/*
2633	 * The default group preallocation is 512, which for 4k block
2634	 * sizes translates to 2 megabytes.  However for bigalloc file
2635	 * systems, this is probably too big (i.e, if the cluster size
2636	 * is 1 megabyte, then group preallocation size becomes half a
2637	 * gigabyte!).  As a default, we will keep a two megabyte
2638	 * group pralloc size for cluster sizes up to 64k, and after
2639	 * that, we will force a minimum group preallocation size of
2640	 * 32 clusters.  This translates to 8 megs when the cluster
2641	 * size is 256k, and 32 megs when the cluster size is 1 meg,
2642	 * which seems reasonable as a default.
2643	 */
2644	sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2645				       sbi->s_cluster_bits, 32);
2646	/*
2647	 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2648	 * to the lowest multiple of s_stripe which is bigger than
2649	 * the s_mb_group_prealloc as determined above. We want
2650	 * the preallocation size to be an exact multiple of the
2651	 * RAID stripe size so that preallocations don't fragment
2652	 * the stripes.
2653	 */
2654	if (sbi->s_stripe > 1) {
2655		sbi->s_mb_group_prealloc = roundup(
2656			sbi->s_mb_group_prealloc, sbi->s_stripe);
2657	}
2658
2659	sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2660	if (sbi->s_locality_groups == NULL) {
2661		ret = -ENOMEM;
2662		goto out;
2663	}
2664	for_each_possible_cpu(i) {
2665		struct ext4_locality_group *lg;
2666		lg = per_cpu_ptr(sbi->s_locality_groups, i);
2667		mutex_init(&lg->lg_mutex);
2668		for (j = 0; j < PREALLOC_TB_SIZE; j++)
2669			INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2670		spin_lock_init(&lg->lg_prealloc_lock);
2671	}
2672
2673	/* init file for buddy data */
2674	ret = ext4_mb_init_backend(sb);
2675	if (ret != 0)
2676		goto out_free_locality_groups;
2677
2678	return 0;
2679
2680out_free_locality_groups:
2681	free_percpu(sbi->s_locality_groups);
2682	sbi->s_locality_groups = NULL;
2683out:
2684	kfree(sbi->s_mb_offsets);
2685	sbi->s_mb_offsets = NULL;
2686	kfree(sbi->s_mb_maxs);
2687	sbi->s_mb_maxs = NULL;
2688	return ret;
2689}
2690
2691/* need to called with the ext4 group lock held */
2692static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2693{
2694	struct ext4_prealloc_space *pa;
2695	struct list_head *cur, *tmp;
2696	int count = 0;
2697
2698	list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2699		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2700		list_del(&pa->pa_group_list);
2701		count++;
2702		kmem_cache_free(ext4_pspace_cachep, pa);
2703	}
2704	if (count)
2705		mb_debug(1, "mballoc: %u PAs left\n", count);
2706
2707}
2708
2709int ext4_mb_release(struct super_block *sb)
2710{
2711	ext4_group_t ngroups = ext4_get_groups_count(sb);
2712	ext4_group_t i;
2713	int num_meta_group_infos;
2714	struct ext4_group_info *grinfo;
2715	struct ext4_sb_info *sbi = EXT4_SB(sb);
2716	struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2717
2718	if (sbi->s_group_info) {
2719		for (i = 0; i < ngroups; i++) {
2720			grinfo = ext4_get_group_info(sb, i);
2721#ifdef DOUBLE_CHECK
2722			kfree(grinfo->bb_bitmap);
2723#endif
2724			ext4_lock_group(sb, i);
2725			ext4_mb_cleanup_pa(grinfo);
2726			ext4_unlock_group(sb, i);
2727			kmem_cache_free(cachep, grinfo);
2728		}
2729		num_meta_group_infos = (ngroups +
2730				EXT4_DESC_PER_BLOCK(sb) - 1) >>
2731			EXT4_DESC_PER_BLOCK_BITS(sb);
2732		for (i = 0; i < num_meta_group_infos; i++)
2733			kfree(sbi->s_group_info[i]);
2734		kvfree(sbi->s_group_info);
2735	}
2736	kfree(sbi->s_mb_offsets);
2737	kfree(sbi->s_mb_maxs);
2738	iput(sbi->s_buddy_cache);
2739	if (sbi->s_mb_stats) {
2740		ext4_msg(sb, KERN_INFO,
2741		       "mballoc: %u blocks %u reqs (%u success)",
2742				atomic_read(&sbi->s_bal_allocated),
2743				atomic_read(&sbi->s_bal_reqs),
2744				atomic_read(&sbi->s_bal_success));
2745		ext4_msg(sb, KERN_INFO,
2746		      "mballoc: %u extents scanned, %u goal hits, "
2747				"%u 2^N hits, %u breaks, %u lost",
2748				atomic_read(&sbi->s_bal_ex_scanned),
2749				atomic_read(&sbi->s_bal_goals),
2750				atomic_read(&sbi->s_bal_2orders),
2751				atomic_read(&sbi->s_bal_breaks),
2752				atomic_read(&sbi->s_mb_lost_chunks));
2753		ext4_msg(sb, KERN_INFO,
2754		       "mballoc: %lu generated and it took %Lu",
2755				sbi->s_mb_buddies_generated,
2756				sbi->s_mb_generation_time);
2757		ext4_msg(sb, KERN_INFO,
2758		       "mballoc: %u preallocated, %u discarded",
2759				atomic_read(&sbi->s_mb_preallocated),
2760				atomic_read(&sbi->s_mb_discarded));
2761	}
2762
2763	free_percpu(sbi->s_locality_groups);
2764
2765	return 0;
2766}
2767
2768static inline int ext4_issue_discard(struct super_block *sb,
2769		ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2770{
2771	ext4_fsblk_t discard_block;
2772
2773	discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2774			 ext4_group_first_block_no(sb, block_group));
2775	count = EXT4_C2B(EXT4_SB(sb), count);
2776	trace_ext4_discard_blocks(sb,
2777			(unsigned long long) discard_block, count);
2778	return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2779}
2780
2781/*
2782 * This function is called by the jbd2 layer once the commit has finished,
2783 * so we know we can free the blocks that were released with that commit.
2784 */
2785static void ext4_free_data_callback(struct super_block *sb,
2786				    struct ext4_journal_cb_entry *jce,
2787				    int rc)
2788{
2789	struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2790	struct ext4_buddy e4b;
2791	struct ext4_group_info *db;
2792	int err, count = 0, count2 = 0;
2793
2794	mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2795		 entry->efd_count, entry->efd_group, entry);
2796
2797	if (test_opt(sb, DISCARD)) {
2798		err = ext4_issue_discard(sb, entry->efd_group,
2799					 entry->efd_start_cluster,
2800					 entry->efd_count);
2801		if (err && err != -EOPNOTSUPP)
2802			ext4_msg(sb, KERN_WARNING, "discard request in"
2803				 " group:%d block:%d count:%d failed"
2804				 " with %d", entry->efd_group,
2805				 entry->efd_start_cluster,
2806				 entry->efd_count, err);
2807	}
2808
2809	err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2810	/* we expect to find existing buddy because it's pinned */
2811	BUG_ON(err != 0);
2812
 
 
 
2813
2814	db = e4b.bd_info;
2815	/* there are blocks to put in buddy to make them really free */
2816	count += entry->efd_count;
2817	count2++;
2818	ext4_lock_group(sb, entry->efd_group);
2819	/* Take it out of per group rb tree */
2820	rb_erase(&entry->efd_node, &(db->bb_free_root));
2821	mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2822
2823	/*
2824	 * Clear the trimmed flag for the group so that the next
2825	 * ext4_trim_fs can trim it.
2826	 * If the volume is mounted with -o discard, online discard
2827	 * is supported and the free blocks will be trimmed online.
2828	 */
2829	if (!test_opt(sb, DISCARD))
2830		EXT4_MB_GRP_CLEAR_TRIMMED(db);
2831
2832	if (!db->bb_free_root.rb_node) {
2833		/* No more items in the per group rb tree
2834		 * balance refcounts from ext4_mb_free_metadata()
2835		 */
2836		put_page(e4b.bd_buddy_page);
2837		put_page(e4b.bd_bitmap_page);
2838	}
2839	ext4_unlock_group(sb, entry->efd_group);
2840	kmem_cache_free(ext4_free_data_cachep, entry);
2841	ext4_mb_unload_buddy(&e4b);
2842
2843	mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2844}
2845
2846int __init ext4_init_mballoc(void)
2847{
2848	ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2849					SLAB_RECLAIM_ACCOUNT);
2850	if (ext4_pspace_cachep == NULL)
2851		return -ENOMEM;
2852
2853	ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2854				    SLAB_RECLAIM_ACCOUNT);
2855	if (ext4_ac_cachep == NULL) {
2856		kmem_cache_destroy(ext4_pspace_cachep);
2857		return -ENOMEM;
2858	}
2859
2860	ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2861					   SLAB_RECLAIM_ACCOUNT);
2862	if (ext4_free_data_cachep == NULL) {
2863		kmem_cache_destroy(ext4_pspace_cachep);
2864		kmem_cache_destroy(ext4_ac_cachep);
2865		return -ENOMEM;
2866	}
2867	return 0;
2868}
2869
2870void ext4_exit_mballoc(void)
2871{
2872	/*
2873	 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2874	 * before destroying the slab cache.
2875	 */
2876	rcu_barrier();
2877	kmem_cache_destroy(ext4_pspace_cachep);
2878	kmem_cache_destroy(ext4_ac_cachep);
2879	kmem_cache_destroy(ext4_free_data_cachep);
2880	ext4_groupinfo_destroy_slabs();
2881}
2882
2883
2884/*
2885 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2886 * Returns 0 if success or error code
2887 */
2888static noinline_for_stack int
2889ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2890				handle_t *handle, unsigned int reserv_clstrs)
2891{
2892	struct buffer_head *bitmap_bh = NULL;
2893	struct ext4_group_desc *gdp;
2894	struct buffer_head *gdp_bh;
2895	struct ext4_sb_info *sbi;
2896	struct super_block *sb;
2897	ext4_fsblk_t block;
2898	int err, len;
2899
2900	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2901	BUG_ON(ac->ac_b_ex.fe_len <= 0);
2902
2903	sb = ac->ac_sb;
2904	sbi = EXT4_SB(sb);
2905
2906	bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2907	if (IS_ERR(bitmap_bh)) {
2908		err = PTR_ERR(bitmap_bh);
2909		bitmap_bh = NULL;
2910		goto out_err;
2911	}
2912
2913	BUFFER_TRACE(bitmap_bh, "getting write access");
2914	err = ext4_journal_get_write_access(handle, bitmap_bh);
2915	if (err)
2916		goto out_err;
2917
2918	err = -EIO;
2919	gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2920	if (!gdp)
2921		goto out_err;
2922
2923	ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2924			ext4_free_group_clusters(sb, gdp));
2925
2926	BUFFER_TRACE(gdp_bh, "get_write_access");
2927	err = ext4_journal_get_write_access(handle, gdp_bh);
2928	if (err)
2929		goto out_err;
2930
2931	block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2932
2933	len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2934	if (!ext4_data_block_valid(sbi, block, len)) {
2935		ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2936			   "fs metadata", block, block+len);
2937		/* File system mounted not to panic on error
2938		 * Fix the bitmap and repeat the block allocation
2939		 * We leak some of the blocks here.
2940		 */
2941		ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2942		ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2943			      ac->ac_b_ex.fe_len);
2944		ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2945		err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2946		if (!err)
2947			err = -EAGAIN;
2948		goto out_err;
2949	}
2950
2951	ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2952#ifdef AGGRESSIVE_CHECK
2953	{
2954		int i;
2955		for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2956			BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2957						bitmap_bh->b_data));
2958		}
2959	}
2960#endif
2961	ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2962		      ac->ac_b_ex.fe_len);
2963	if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2964		gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2965		ext4_free_group_clusters_set(sb, gdp,
2966					     ext4_free_clusters_after_init(sb,
2967						ac->ac_b_ex.fe_group, gdp));
2968	}
2969	len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2970	ext4_free_group_clusters_set(sb, gdp, len);
2971	ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2972	ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2973
2974	ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2975	percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2976	/*
2977	 * Now reduce the dirty block count also. Should not go negative
2978	 */
2979	if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2980		/* release all the reserved blocks if non delalloc */
2981		percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2982				   reserv_clstrs);
2983
2984	if (sbi->s_log_groups_per_flex) {
2985		ext4_group_t flex_group = ext4_flex_group(sbi,
2986							  ac->ac_b_ex.fe_group);
2987		atomic64_sub(ac->ac_b_ex.fe_len,
2988			     &sbi->s_flex_groups[flex_group].free_clusters);
2989	}
2990
2991	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2992	if (err)
2993		goto out_err;
2994	err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
2995
2996out_err:
2997	brelse(bitmap_bh);
2998	return err;
2999}
3000
3001/*
3002 * here we normalize request for locality group
3003 * Group request are normalized to s_mb_group_prealloc, which goes to
3004 * s_strip if we set the same via mount option.
3005 * s_mb_group_prealloc can be configured via
3006 * /sys/fs/ext4/<partition>/mb_group_prealloc
3007 *
3008 * XXX: should we try to preallocate more than the group has now?
3009 */
3010static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3011{
3012	struct super_block *sb = ac->ac_sb;
3013	struct ext4_locality_group *lg = ac->ac_lg;
3014
3015	BUG_ON(lg == NULL);
3016	ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3017	mb_debug(1, "#%u: goal %u blocks for locality group\n",
3018		current->pid, ac->ac_g_ex.fe_len);
3019}
3020
3021/*
3022 * Normalization means making request better in terms of
3023 * size and alignment
3024 */
3025static noinline_for_stack void
3026ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3027				struct ext4_allocation_request *ar)
3028{
3029	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3030	int bsbits, max;
3031	ext4_lblk_t end;
3032	loff_t size, start_off;
3033	loff_t orig_size __maybe_unused;
3034	ext4_lblk_t start;
3035	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3036	struct ext4_prealloc_space *pa;
3037
3038	/* do normalize only data requests, metadata requests
3039	   do not need preallocation */
3040	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3041		return;
3042
3043	/* sometime caller may want exact blocks */
3044	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3045		return;
3046
3047	/* caller may indicate that preallocation isn't
3048	 * required (it's a tail, for example) */
3049	if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3050		return;
3051
3052	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3053		ext4_mb_normalize_group_request(ac);
3054		return ;
3055	}
3056
3057	bsbits = ac->ac_sb->s_blocksize_bits;
3058
3059	/* first, let's learn actual file size
3060	 * given current request is allocated */
3061	size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3062	size = size << bsbits;
3063	if (size < i_size_read(ac->ac_inode))
3064		size = i_size_read(ac->ac_inode);
3065	orig_size = size;
3066
3067	/* max size of free chunks */
3068	max = 2 << bsbits;
3069
3070#define NRL_CHECK_SIZE(req, size, max, chunk_size)	\
3071		(req <= (size) || max <= (chunk_size))
3072
3073	/* first, try to predict filesize */
3074	/* XXX: should this table be tunable? */
3075	start_off = 0;
3076	if (size <= 16 * 1024) {
3077		size = 16 * 1024;
3078	} else if (size <= 32 * 1024) {
3079		size = 32 * 1024;
3080	} else if (size <= 64 * 1024) {
3081		size = 64 * 1024;
3082	} else if (size <= 128 * 1024) {
3083		size = 128 * 1024;
3084	} else if (size <= 256 * 1024) {
3085		size = 256 * 1024;
3086	} else if (size <= 512 * 1024) {
3087		size = 512 * 1024;
3088	} else if (size <= 1024 * 1024) {
3089		size = 1024 * 1024;
3090	} else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3091		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3092						(21 - bsbits)) << 21;
3093		size = 2 * 1024 * 1024;
3094	} else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3095		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3096							(22 - bsbits)) << 22;
3097		size = 4 * 1024 * 1024;
3098	} else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3099					(8<<20)>>bsbits, max, 8 * 1024)) {
3100		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3101							(23 - bsbits)) << 23;
3102		size = 8 * 1024 * 1024;
3103	} else {
3104		start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3105		size	  = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3106					      ac->ac_o_ex.fe_len) << bsbits;
3107	}
3108	size = size >> bsbits;
3109	start = start_off >> bsbits;
3110
3111	/* don't cover already allocated blocks in selected range */
3112	if (ar->pleft && start <= ar->lleft) {
3113		size -= ar->lleft + 1 - start;
3114		start = ar->lleft + 1;
3115	}
3116	if (ar->pright && start + size - 1 >= ar->lright)
3117		size -= start + size - ar->lright;
3118
 
 
 
 
 
 
 
3119	end = start + size;
3120
3121	/* check we don't cross already preallocated blocks */
3122	rcu_read_lock();
3123	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3124		ext4_lblk_t pa_end;
3125
3126		if (pa->pa_deleted)
3127			continue;
3128		spin_lock(&pa->pa_lock);
3129		if (pa->pa_deleted) {
3130			spin_unlock(&pa->pa_lock);
3131			continue;
3132		}
3133
3134		pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3135						  pa->pa_len);
3136
3137		/* PA must not overlap original request */
3138		BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3139			ac->ac_o_ex.fe_logical < pa->pa_lstart));
3140
3141		/* skip PAs this normalized request doesn't overlap with */
3142		if (pa->pa_lstart >= end || pa_end <= start) {
3143			spin_unlock(&pa->pa_lock);
3144			continue;
3145		}
3146		BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3147
3148		/* adjust start or end to be adjacent to this pa */
3149		if (pa_end <= ac->ac_o_ex.fe_logical) {
3150			BUG_ON(pa_end < start);
3151			start = pa_end;
3152		} else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3153			BUG_ON(pa->pa_lstart > end);
3154			end = pa->pa_lstart;
3155		}
3156		spin_unlock(&pa->pa_lock);
3157	}
3158	rcu_read_unlock();
3159	size = end - start;
3160
3161	/* XXX: extra loop to check we really don't overlap preallocations */
3162	rcu_read_lock();
3163	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3164		ext4_lblk_t pa_end;
3165
3166		spin_lock(&pa->pa_lock);
3167		if (pa->pa_deleted == 0) {
3168			pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3169							  pa->pa_len);
3170			BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3171		}
3172		spin_unlock(&pa->pa_lock);
3173	}
3174	rcu_read_unlock();
3175
3176	if (start + size <= ac->ac_o_ex.fe_logical &&
3177			start > ac->ac_o_ex.fe_logical) {
3178		ext4_msg(ac->ac_sb, KERN_ERR,
3179			 "start %lu, size %lu, fe_logical %lu",
3180			 (unsigned long) start, (unsigned long) size,
3181			 (unsigned long) ac->ac_o_ex.fe_logical);
3182		BUG();
3183	}
3184	BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3185
3186	/* now prepare goal request */
3187
3188	/* XXX: is it better to align blocks WRT to logical
3189	 * placement or satisfy big request as is */
3190	ac->ac_g_ex.fe_logical = start;
3191	ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3192
3193	/* define goal start in order to merge */
3194	if (ar->pright && (ar->lright == (start + size))) {
3195		/* merge to the right */
3196		ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3197						&ac->ac_f_ex.fe_group,
3198						&ac->ac_f_ex.fe_start);
3199		ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3200	}
3201	if (ar->pleft && (ar->lleft + 1 == start)) {
3202		/* merge to the left */
3203		ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3204						&ac->ac_f_ex.fe_group,
3205						&ac->ac_f_ex.fe_start);
3206		ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3207	}
3208
3209	mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3210		(unsigned) orig_size, (unsigned) start);
3211}
3212
3213static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3214{
3215	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3216
3217	if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3218		atomic_inc(&sbi->s_bal_reqs);
3219		atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3220		if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3221			atomic_inc(&sbi->s_bal_success);
3222		atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3223		if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3224				ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3225			atomic_inc(&sbi->s_bal_goals);
3226		if (ac->ac_found > sbi->s_mb_max_to_scan)
3227			atomic_inc(&sbi->s_bal_breaks);
3228	}
3229
3230	if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3231		trace_ext4_mballoc_alloc(ac);
3232	else
3233		trace_ext4_mballoc_prealloc(ac);
3234}
3235
3236/*
3237 * Called on failure; free up any blocks from the inode PA for this
3238 * context.  We don't need this for MB_GROUP_PA because we only change
3239 * pa_free in ext4_mb_release_context(), but on failure, we've already
3240 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3241 */
3242static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3243{
3244	struct ext4_prealloc_space *pa = ac->ac_pa;
3245	struct ext4_buddy e4b;
3246	int err;
3247
3248	if (pa == NULL) {
3249		if (ac->ac_f_ex.fe_len == 0)
3250			return;
3251		err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3252		if (err) {
3253			/*
3254			 * This should never happen since we pin the
3255			 * pages in the ext4_allocation_context so
3256			 * ext4_mb_load_buddy() should never fail.
3257			 */
3258			WARN(1, "mb_load_buddy failed (%d)", err);
3259			return;
3260		}
3261		ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3262		mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3263			       ac->ac_f_ex.fe_len);
3264		ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3265		ext4_mb_unload_buddy(&e4b);
3266		return;
3267	}
3268	if (pa->pa_type == MB_INODE_PA)
3269		pa->pa_free += ac->ac_b_ex.fe_len;
3270}
3271
3272/*
3273 * use blocks preallocated to inode
3274 */
3275static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3276				struct ext4_prealloc_space *pa)
3277{
3278	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3279	ext4_fsblk_t start;
3280	ext4_fsblk_t end;
3281	int len;
3282
3283	/* found preallocated blocks, use them */
3284	start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3285	end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3286		  start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3287	len = EXT4_NUM_B2C(sbi, end - start);
3288	ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3289					&ac->ac_b_ex.fe_start);
3290	ac->ac_b_ex.fe_len = len;
3291	ac->ac_status = AC_STATUS_FOUND;
3292	ac->ac_pa = pa;
3293
3294	BUG_ON(start < pa->pa_pstart);
3295	BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3296	BUG_ON(pa->pa_free < len);
3297	pa->pa_free -= len;
3298
3299	mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3300}
3301
3302/*
3303 * use blocks preallocated to locality group
3304 */
3305static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3306				struct ext4_prealloc_space *pa)
3307{
3308	unsigned int len = ac->ac_o_ex.fe_len;
3309
3310	ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3311					&ac->ac_b_ex.fe_group,
3312					&ac->ac_b_ex.fe_start);
3313	ac->ac_b_ex.fe_len = len;
3314	ac->ac_status = AC_STATUS_FOUND;
3315	ac->ac_pa = pa;
3316
3317	/* we don't correct pa_pstart or pa_plen here to avoid
3318	 * possible race when the group is being loaded concurrently
3319	 * instead we correct pa later, after blocks are marked
3320	 * in on-disk bitmap -- see ext4_mb_release_context()
3321	 * Other CPUs are prevented from allocating from this pa by lg_mutex
3322	 */
3323	mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3324}
3325
3326/*
3327 * Return the prealloc space that have minimal distance
3328 * from the goal block. @cpa is the prealloc
3329 * space that is having currently known minimal distance
3330 * from the goal block.
3331 */
3332static struct ext4_prealloc_space *
3333ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3334			struct ext4_prealloc_space *pa,
3335			struct ext4_prealloc_space *cpa)
3336{
3337	ext4_fsblk_t cur_distance, new_distance;
3338
3339	if (cpa == NULL) {
3340		atomic_inc(&pa->pa_count);
3341		return pa;
3342	}
3343	cur_distance = abs(goal_block - cpa->pa_pstart);
3344	new_distance = abs(goal_block - pa->pa_pstart);
3345
3346	if (cur_distance <= new_distance)
3347		return cpa;
3348
3349	/* drop the previous reference */
3350	atomic_dec(&cpa->pa_count);
3351	atomic_inc(&pa->pa_count);
3352	return pa;
3353}
3354
3355/*
3356 * search goal blocks in preallocated space
3357 */
3358static noinline_for_stack int
3359ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3360{
3361	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3362	int order, i;
3363	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3364	struct ext4_locality_group *lg;
3365	struct ext4_prealloc_space *pa, *cpa = NULL;
3366	ext4_fsblk_t goal_block;
3367
3368	/* only data can be preallocated */
3369	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3370		return 0;
3371
3372	/* first, try per-file preallocation */
3373	rcu_read_lock();
3374	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3375
3376		/* all fields in this condition don't change,
3377		 * so we can skip locking for them */
3378		if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3379		    ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3380					       EXT4_C2B(sbi, pa->pa_len)))
3381			continue;
3382
3383		/* non-extent files can't have physical blocks past 2^32 */
3384		if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3385		    (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3386		     EXT4_MAX_BLOCK_FILE_PHYS))
3387			continue;
3388
3389		/* found preallocated blocks, use them */
3390		spin_lock(&pa->pa_lock);
3391		if (pa->pa_deleted == 0 && pa->pa_free) {
3392			atomic_inc(&pa->pa_count);
3393			ext4_mb_use_inode_pa(ac, pa);
3394			spin_unlock(&pa->pa_lock);
3395			ac->ac_criteria = 10;
3396			rcu_read_unlock();
3397			return 1;
3398		}
3399		spin_unlock(&pa->pa_lock);
3400	}
3401	rcu_read_unlock();
3402
3403	/* can we use group allocation? */
3404	if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3405		return 0;
3406
3407	/* inode may have no locality group for some reason */
3408	lg = ac->ac_lg;
3409	if (lg == NULL)
3410		return 0;
3411	order  = fls(ac->ac_o_ex.fe_len) - 1;
3412	if (order > PREALLOC_TB_SIZE - 1)
3413		/* The max size of hash table is PREALLOC_TB_SIZE */
3414		order = PREALLOC_TB_SIZE - 1;
3415
3416	goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3417	/*
3418	 * search for the prealloc space that is having
3419	 * minimal distance from the goal block.
3420	 */
3421	for (i = order; i < PREALLOC_TB_SIZE; i++) {
3422		rcu_read_lock();
3423		list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3424					pa_inode_list) {
3425			spin_lock(&pa->pa_lock);
3426			if (pa->pa_deleted == 0 &&
3427					pa->pa_free >= ac->ac_o_ex.fe_len) {
3428
3429				cpa = ext4_mb_check_group_pa(goal_block,
3430								pa, cpa);
3431			}
3432			spin_unlock(&pa->pa_lock);
3433		}
3434		rcu_read_unlock();
3435	}
3436	if (cpa) {
3437		ext4_mb_use_group_pa(ac, cpa);
3438		ac->ac_criteria = 20;
3439		return 1;
3440	}
3441	return 0;
3442}
3443
3444/*
3445 * the function goes through all block freed in the group
3446 * but not yet committed and marks them used in in-core bitmap.
3447 * buddy must be generated from this bitmap
3448 * Need to be called with the ext4 group lock held
3449 */
3450static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3451						ext4_group_t group)
3452{
3453	struct rb_node *n;
3454	struct ext4_group_info *grp;
3455	struct ext4_free_data *entry;
3456
3457	grp = ext4_get_group_info(sb, group);
3458	n = rb_first(&(grp->bb_free_root));
3459
3460	while (n) {
3461		entry = rb_entry(n, struct ext4_free_data, efd_node);
3462		ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3463		n = rb_next(n);
3464	}
3465	return;
3466}
3467
3468/*
3469 * the function goes through all preallocation in this group and marks them
3470 * used in in-core bitmap. buddy must be generated from this bitmap
3471 * Need to be called with ext4 group lock held
3472 */
3473static noinline_for_stack
3474void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3475					ext4_group_t group)
3476{
3477	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3478	struct ext4_prealloc_space *pa;
3479	struct list_head *cur;
3480	ext4_group_t groupnr;
3481	ext4_grpblk_t start;
3482	int preallocated = 0;
3483	int len;
3484
3485	/* all form of preallocation discards first load group,
3486	 * so the only competing code is preallocation use.
3487	 * we don't need any locking here
3488	 * notice we do NOT ignore preallocations with pa_deleted
3489	 * otherwise we could leave used blocks available for
3490	 * allocation in buddy when concurrent ext4_mb_put_pa()
3491	 * is dropping preallocation
3492	 */
3493	list_for_each(cur, &grp->bb_prealloc_list) {
3494		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3495		spin_lock(&pa->pa_lock);
3496		ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3497					     &groupnr, &start);
3498		len = pa->pa_len;
3499		spin_unlock(&pa->pa_lock);
3500		if (unlikely(len == 0))
3501			continue;
3502		BUG_ON(groupnr != group);
3503		ext4_set_bits(bitmap, start, len);
3504		preallocated += len;
3505	}
3506	mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3507}
3508
3509static void ext4_mb_pa_callback(struct rcu_head *head)
3510{
3511	struct ext4_prealloc_space *pa;
3512	pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3513
3514	BUG_ON(atomic_read(&pa->pa_count));
3515	BUG_ON(pa->pa_deleted == 0);
3516	kmem_cache_free(ext4_pspace_cachep, pa);
3517}
3518
3519/*
3520 * drops a reference to preallocated space descriptor
3521 * if this was the last reference and the space is consumed
3522 */
3523static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3524			struct super_block *sb, struct ext4_prealloc_space *pa)
3525{
3526	ext4_group_t grp;
3527	ext4_fsblk_t grp_blk;
3528
3529	/* in this short window concurrent discard can set pa_deleted */
3530	spin_lock(&pa->pa_lock);
3531	if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3532		spin_unlock(&pa->pa_lock);
3533		return;
3534	}
3535
3536	if (pa->pa_deleted == 1) {
3537		spin_unlock(&pa->pa_lock);
3538		return;
3539	}
3540
3541	pa->pa_deleted = 1;
3542	spin_unlock(&pa->pa_lock);
3543
3544	grp_blk = pa->pa_pstart;
3545	/*
3546	 * If doing group-based preallocation, pa_pstart may be in the
3547	 * next group when pa is used up
3548	 */
3549	if (pa->pa_type == MB_GROUP_PA)
3550		grp_blk--;
3551
3552	grp = ext4_get_group_number(sb, grp_blk);
3553
3554	/*
3555	 * possible race:
3556	 *
3557	 *  P1 (buddy init)			P2 (regular allocation)
3558	 *					find block B in PA
3559	 *  copy on-disk bitmap to buddy
3560	 *  					mark B in on-disk bitmap
3561	 *					drop PA from group
3562	 *  mark all PAs in buddy
3563	 *
3564	 * thus, P1 initializes buddy with B available. to prevent this
3565	 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3566	 * against that pair
3567	 */
3568	ext4_lock_group(sb, grp);
3569	list_del(&pa->pa_group_list);
3570	ext4_unlock_group(sb, grp);
3571
3572	spin_lock(pa->pa_obj_lock);
3573	list_del_rcu(&pa->pa_inode_list);
3574	spin_unlock(pa->pa_obj_lock);
3575
3576	call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3577}
3578
3579/*
3580 * creates new preallocated space for given inode
3581 */
3582static noinline_for_stack int
3583ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3584{
3585	struct super_block *sb = ac->ac_sb;
3586	struct ext4_sb_info *sbi = EXT4_SB(sb);
3587	struct ext4_prealloc_space *pa;
3588	struct ext4_group_info *grp;
3589	struct ext4_inode_info *ei;
3590
3591	/* preallocate only when found space is larger then requested */
3592	BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3593	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3594	BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3595
3596	pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3597	if (pa == NULL)
3598		return -ENOMEM;
3599
3600	if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3601		int winl;
3602		int wins;
3603		int win;
3604		int offs;
3605
3606		/* we can't allocate as much as normalizer wants.
3607		 * so, found space must get proper lstart
3608		 * to cover original request */
3609		BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3610		BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3611
3612		/* we're limited by original request in that
3613		 * logical block must be covered any way
3614		 * winl is window we can move our chunk within */
3615		winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3616
3617		/* also, we should cover whole original request */
3618		wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3619
3620		/* the smallest one defines real window */
3621		win = min(winl, wins);
3622
3623		offs = ac->ac_o_ex.fe_logical %
3624			EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3625		if (offs && offs < win)
3626			win = offs;
3627
3628		ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3629			EXT4_NUM_B2C(sbi, win);
3630		BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3631		BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3632	}
3633
3634	/* preallocation can change ac_b_ex, thus we store actually
3635	 * allocated blocks for history */
3636	ac->ac_f_ex = ac->ac_b_ex;
3637
3638	pa->pa_lstart = ac->ac_b_ex.fe_logical;
3639	pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3640	pa->pa_len = ac->ac_b_ex.fe_len;
3641	pa->pa_free = pa->pa_len;
3642	atomic_set(&pa->pa_count, 1);
3643	spin_lock_init(&pa->pa_lock);
3644	INIT_LIST_HEAD(&pa->pa_inode_list);
3645	INIT_LIST_HEAD(&pa->pa_group_list);
3646	pa->pa_deleted = 0;
3647	pa->pa_type = MB_INODE_PA;
3648
3649	mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3650			pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3651	trace_ext4_mb_new_inode_pa(ac, pa);
3652
3653	ext4_mb_use_inode_pa(ac, pa);
3654	atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3655
3656	ei = EXT4_I(ac->ac_inode);
3657	grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3658
3659	pa->pa_obj_lock = &ei->i_prealloc_lock;
3660	pa->pa_inode = ac->ac_inode;
3661
3662	ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3663	list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3664	ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3665
3666	spin_lock(pa->pa_obj_lock);
3667	list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3668	spin_unlock(pa->pa_obj_lock);
3669
3670	return 0;
3671}
3672
3673/*
3674 * creates new preallocated space for locality group inodes belongs to
3675 */
3676static noinline_for_stack int
3677ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3678{
3679	struct super_block *sb = ac->ac_sb;
3680	struct ext4_locality_group *lg;
3681	struct ext4_prealloc_space *pa;
3682	struct ext4_group_info *grp;
3683
3684	/* preallocate only when found space is larger then requested */
3685	BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3686	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3687	BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3688
3689	BUG_ON(ext4_pspace_cachep == NULL);
3690	pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3691	if (pa == NULL)
3692		return -ENOMEM;
3693
3694	/* preallocation can change ac_b_ex, thus we store actually
3695	 * allocated blocks for history */
3696	ac->ac_f_ex = ac->ac_b_ex;
3697
3698	pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3699	pa->pa_lstart = pa->pa_pstart;
3700	pa->pa_len = ac->ac_b_ex.fe_len;
3701	pa->pa_free = pa->pa_len;
3702	atomic_set(&pa->pa_count, 1);
3703	spin_lock_init(&pa->pa_lock);
3704	INIT_LIST_HEAD(&pa->pa_inode_list);
3705	INIT_LIST_HEAD(&pa->pa_group_list);
3706	pa->pa_deleted = 0;
3707	pa->pa_type = MB_GROUP_PA;
3708
3709	mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3710			pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3711	trace_ext4_mb_new_group_pa(ac, pa);
3712
3713	ext4_mb_use_group_pa(ac, pa);
3714	atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3715
3716	grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3717	lg = ac->ac_lg;
3718	BUG_ON(lg == NULL);
3719
3720	pa->pa_obj_lock = &lg->lg_prealloc_lock;
3721	pa->pa_inode = NULL;
3722
3723	ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3724	list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3725	ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3726
3727	/*
3728	 * We will later add the new pa to the right bucket
3729	 * after updating the pa_free in ext4_mb_release_context
3730	 */
3731	return 0;
3732}
3733
3734static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3735{
3736	int err;
3737
3738	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3739		err = ext4_mb_new_group_pa(ac);
3740	else
3741		err = ext4_mb_new_inode_pa(ac);
3742	return err;
3743}
3744
3745/*
3746 * finds all unused blocks in on-disk bitmap, frees them in
3747 * in-core bitmap and buddy.
3748 * @pa must be unlinked from inode and group lists, so that
3749 * nobody else can find/use it.
3750 * the caller MUST hold group/inode locks.
3751 * TODO: optimize the case when there are no in-core structures yet
3752 */
3753static noinline_for_stack int
3754ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3755			struct ext4_prealloc_space *pa)
3756{
3757	struct super_block *sb = e4b->bd_sb;
3758	struct ext4_sb_info *sbi = EXT4_SB(sb);
3759	unsigned int end;
3760	unsigned int next;
3761	ext4_group_t group;
3762	ext4_grpblk_t bit;
3763	unsigned long long grp_blk_start;
3764	int err = 0;
3765	int free = 0;
3766
3767	BUG_ON(pa->pa_deleted == 0);
3768	ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3769	grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3770	BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3771	end = bit + pa->pa_len;
3772
3773	while (bit < end) {
3774		bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3775		if (bit >= end)
3776			break;
3777		next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3778		mb_debug(1, "    free preallocated %u/%u in group %u\n",
3779			 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3780			 (unsigned) next - bit, (unsigned) group);
3781		free += next - bit;
3782
3783		trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3784		trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3785						    EXT4_C2B(sbi, bit)),
3786					       next - bit);
3787		mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3788		bit = next + 1;
3789	}
3790	if (free != pa->pa_free) {
3791		ext4_msg(e4b->bd_sb, KERN_CRIT,
3792			 "pa %p: logic %lu, phys. %lu, len %lu",
3793			 pa, (unsigned long) pa->pa_lstart,
3794			 (unsigned long) pa->pa_pstart,
3795			 (unsigned long) pa->pa_len);
3796		ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3797					free, pa->pa_free);
3798		/*
3799		 * pa is already deleted so we use the value obtained
3800		 * from the bitmap and continue.
3801		 */
3802	}
3803	atomic_add(free, &sbi->s_mb_discarded);
3804
3805	return err;
3806}
3807
3808static noinline_for_stack int
3809ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3810				struct ext4_prealloc_space *pa)
3811{
3812	struct super_block *sb = e4b->bd_sb;
3813	ext4_group_t group;
3814	ext4_grpblk_t bit;
3815
3816	trace_ext4_mb_release_group_pa(sb, pa);
3817	BUG_ON(pa->pa_deleted == 0);
3818	ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3819	BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3820	mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3821	atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3822	trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3823
3824	return 0;
3825}
3826
3827/*
3828 * releases all preallocations in given group
3829 *
3830 * first, we need to decide discard policy:
3831 * - when do we discard
3832 *   1) ENOSPC
3833 * - how many do we discard
3834 *   1) how many requested
3835 */
3836static noinline_for_stack int
3837ext4_mb_discard_group_preallocations(struct super_block *sb,
3838					ext4_group_t group, int needed)
3839{
3840	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3841	struct buffer_head *bitmap_bh = NULL;
3842	struct ext4_prealloc_space *pa, *tmp;
3843	struct list_head list;
3844	struct ext4_buddy e4b;
3845	int err;
3846	int busy = 0;
3847	int free = 0;
3848
3849	mb_debug(1, "discard preallocation for group %u\n", group);
3850
3851	if (list_empty(&grp->bb_prealloc_list))
3852		return 0;
3853
3854	bitmap_bh = ext4_read_block_bitmap(sb, group);
3855	if (IS_ERR(bitmap_bh)) {
3856		err = PTR_ERR(bitmap_bh);
3857		ext4_error(sb, "Error %d reading block bitmap for %u",
3858			   err, group);
3859		return 0;
3860	}
3861
3862	err = ext4_mb_load_buddy(sb, group, &e4b);
3863	if (err) {
3864		ext4_error(sb, "Error loading buddy information for %u", group);
3865		put_bh(bitmap_bh);
3866		return 0;
3867	}
3868
3869	if (needed == 0)
3870		needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3871
3872	INIT_LIST_HEAD(&list);
3873repeat:
3874	ext4_lock_group(sb, group);
3875	list_for_each_entry_safe(pa, tmp,
3876				&grp->bb_prealloc_list, pa_group_list) {
3877		spin_lock(&pa->pa_lock);
3878		if (atomic_read(&pa->pa_count)) {
3879			spin_unlock(&pa->pa_lock);
3880			busy = 1;
3881			continue;
3882		}
3883		if (pa->pa_deleted) {
3884			spin_unlock(&pa->pa_lock);
3885			continue;
3886		}
3887
3888		/* seems this one can be freed ... */
3889		pa->pa_deleted = 1;
3890
3891		/* we can trust pa_free ... */
3892		free += pa->pa_free;
3893
3894		spin_unlock(&pa->pa_lock);
3895
3896		list_del(&pa->pa_group_list);
3897		list_add(&pa->u.pa_tmp_list, &list);
3898	}
3899
3900	/* if we still need more blocks and some PAs were used, try again */
3901	if (free < needed && busy) {
3902		busy = 0;
3903		ext4_unlock_group(sb, group);
3904		cond_resched();
3905		goto repeat;
3906	}
3907
3908	/* found anything to free? */
3909	if (list_empty(&list)) {
3910		BUG_ON(free != 0);
3911		goto out;
3912	}
3913
3914	/* now free all selected PAs */
3915	list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3916
3917		/* remove from object (inode or locality group) */
3918		spin_lock(pa->pa_obj_lock);
3919		list_del_rcu(&pa->pa_inode_list);
3920		spin_unlock(pa->pa_obj_lock);
3921
3922		if (pa->pa_type == MB_GROUP_PA)
3923			ext4_mb_release_group_pa(&e4b, pa);
3924		else
3925			ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3926
3927		list_del(&pa->u.pa_tmp_list);
3928		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3929	}
3930
3931out:
3932	ext4_unlock_group(sb, group);
3933	ext4_mb_unload_buddy(&e4b);
3934	put_bh(bitmap_bh);
3935	return free;
3936}
3937
3938/*
3939 * releases all non-used preallocated blocks for given inode
3940 *
3941 * It's important to discard preallocations under i_data_sem
3942 * We don't want another block to be served from the prealloc
3943 * space when we are discarding the inode prealloc space.
3944 *
3945 * FIXME!! Make sure it is valid at all the call sites
3946 */
3947void ext4_discard_preallocations(struct inode *inode)
3948{
3949	struct ext4_inode_info *ei = EXT4_I(inode);
3950	struct super_block *sb = inode->i_sb;
3951	struct buffer_head *bitmap_bh = NULL;
3952	struct ext4_prealloc_space *pa, *tmp;
3953	ext4_group_t group = 0;
3954	struct list_head list;
3955	struct ext4_buddy e4b;
3956	int err;
3957
3958	if (!S_ISREG(inode->i_mode)) {
3959		/*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3960		return;
3961	}
3962
3963	mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3964	trace_ext4_discard_preallocations(inode);
3965
3966	INIT_LIST_HEAD(&list);
3967
3968repeat:
3969	/* first, collect all pa's in the inode */
3970	spin_lock(&ei->i_prealloc_lock);
3971	while (!list_empty(&ei->i_prealloc_list)) {
3972		pa = list_entry(ei->i_prealloc_list.next,
3973				struct ext4_prealloc_space, pa_inode_list);
3974		BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3975		spin_lock(&pa->pa_lock);
3976		if (atomic_read(&pa->pa_count)) {
3977			/* this shouldn't happen often - nobody should
3978			 * use preallocation while we're discarding it */
3979			spin_unlock(&pa->pa_lock);
3980			spin_unlock(&ei->i_prealloc_lock);
3981			ext4_msg(sb, KERN_ERR,
3982				 "uh-oh! used pa while discarding");
3983			WARN_ON(1);
3984			schedule_timeout_uninterruptible(HZ);
3985			goto repeat;
3986
3987		}
3988		if (pa->pa_deleted == 0) {
3989			pa->pa_deleted = 1;
3990			spin_unlock(&pa->pa_lock);
3991			list_del_rcu(&pa->pa_inode_list);
3992			list_add(&pa->u.pa_tmp_list, &list);
3993			continue;
3994		}
3995
3996		/* someone is deleting pa right now */
3997		spin_unlock(&pa->pa_lock);
3998		spin_unlock(&ei->i_prealloc_lock);
3999
4000		/* we have to wait here because pa_deleted
4001		 * doesn't mean pa is already unlinked from
4002		 * the list. as we might be called from
4003		 * ->clear_inode() the inode will get freed
4004		 * and concurrent thread which is unlinking
4005		 * pa from inode's list may access already
4006		 * freed memory, bad-bad-bad */
4007
4008		/* XXX: if this happens too often, we can
4009		 * add a flag to force wait only in case
4010		 * of ->clear_inode(), but not in case of
4011		 * regular truncate */
4012		schedule_timeout_uninterruptible(HZ);
4013		goto repeat;
4014	}
4015	spin_unlock(&ei->i_prealloc_lock);
4016
4017	list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4018		BUG_ON(pa->pa_type != MB_INODE_PA);
4019		group = ext4_get_group_number(sb, pa->pa_pstart);
4020
4021		err = ext4_mb_load_buddy(sb, group, &e4b);
4022		if (err) {
4023			ext4_error(sb, "Error loading buddy information for %u",
4024					group);
4025			continue;
4026		}
4027
4028		bitmap_bh = ext4_read_block_bitmap(sb, group);
4029		if (IS_ERR(bitmap_bh)) {
4030			err = PTR_ERR(bitmap_bh);
4031			ext4_error(sb, "Error %d reading block bitmap for %u",
4032					err, group);
4033			ext4_mb_unload_buddy(&e4b);
4034			continue;
4035		}
4036
4037		ext4_lock_group(sb, group);
4038		list_del(&pa->pa_group_list);
4039		ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4040		ext4_unlock_group(sb, group);
4041
4042		ext4_mb_unload_buddy(&e4b);
4043		put_bh(bitmap_bh);
4044
4045		list_del(&pa->u.pa_tmp_list);
4046		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4047	}
4048}
4049
4050#ifdef CONFIG_EXT4_DEBUG
4051static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4052{
4053	struct super_block *sb = ac->ac_sb;
4054	ext4_group_t ngroups, i;
4055
4056	if (!ext4_mballoc_debug ||
4057	    (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4058		return;
4059
4060	ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4061			" Allocation context details:");
4062	ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4063			ac->ac_status, ac->ac_flags);
4064	ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4065		 	"goal %lu/%lu/%lu@%lu, "
4066			"best %lu/%lu/%lu@%lu cr %d",
4067			(unsigned long)ac->ac_o_ex.fe_group,
4068			(unsigned long)ac->ac_o_ex.fe_start,
4069			(unsigned long)ac->ac_o_ex.fe_len,
4070			(unsigned long)ac->ac_o_ex.fe_logical,
4071			(unsigned long)ac->ac_g_ex.fe_group,
4072			(unsigned long)ac->ac_g_ex.fe_start,
4073			(unsigned long)ac->ac_g_ex.fe_len,
4074			(unsigned long)ac->ac_g_ex.fe_logical,
4075			(unsigned long)ac->ac_b_ex.fe_group,
4076			(unsigned long)ac->ac_b_ex.fe_start,
4077			(unsigned long)ac->ac_b_ex.fe_len,
4078			(unsigned long)ac->ac_b_ex.fe_logical,
4079			(int)ac->ac_criteria);
4080	ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4081	ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4082	ngroups = ext4_get_groups_count(sb);
4083	for (i = 0; i < ngroups; i++) {
4084		struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4085		struct ext4_prealloc_space *pa;
4086		ext4_grpblk_t start;
4087		struct list_head *cur;
4088		ext4_lock_group(sb, i);
4089		list_for_each(cur, &grp->bb_prealloc_list) {
4090			pa = list_entry(cur, struct ext4_prealloc_space,
4091					pa_group_list);
4092			spin_lock(&pa->pa_lock);
4093			ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4094						     NULL, &start);
4095			spin_unlock(&pa->pa_lock);
4096			printk(KERN_ERR "PA:%u:%d:%u \n", i,
4097			       start, pa->pa_len);
4098		}
4099		ext4_unlock_group(sb, i);
4100
4101		if (grp->bb_free == 0)
4102			continue;
4103		printk(KERN_ERR "%u: %d/%d \n",
4104		       i, grp->bb_free, grp->bb_fragments);
4105	}
4106	printk(KERN_ERR "\n");
4107}
4108#else
4109static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4110{
4111	return;
4112}
4113#endif
4114
4115/*
4116 * We use locality group preallocation for small size file. The size of the
4117 * file is determined by the current size or the resulting size after
4118 * allocation which ever is larger
4119 *
4120 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4121 */
4122static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4123{
4124	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4125	int bsbits = ac->ac_sb->s_blocksize_bits;
4126	loff_t size, isize;
4127
4128	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4129		return;
4130
4131	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4132		return;
4133
4134	size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4135	isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4136		>> bsbits;
4137
4138	if ((size == isize) &&
4139	    !ext4_fs_is_busy(sbi) &&
4140	    (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4141		ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4142		return;
4143	}
4144
4145	if (sbi->s_mb_group_prealloc <= 0) {
4146		ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4147		return;
4148	}
4149
4150	/* don't use group allocation for large files */
4151	size = max(size, isize);
4152	if (size > sbi->s_mb_stream_request) {
4153		ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4154		return;
4155	}
4156
4157	BUG_ON(ac->ac_lg != NULL);
4158	/*
4159	 * locality group prealloc space are per cpu. The reason for having
4160	 * per cpu locality group is to reduce the contention between block
4161	 * request from multiple CPUs.
4162	 */
4163	ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4164
4165	/* we're going to use group allocation */
4166	ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4167
4168	/* serialize all allocations in the group */
4169	mutex_lock(&ac->ac_lg->lg_mutex);
4170}
4171
4172static noinline_for_stack int
4173ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4174				struct ext4_allocation_request *ar)
4175{
4176	struct super_block *sb = ar->inode->i_sb;
4177	struct ext4_sb_info *sbi = EXT4_SB(sb);
4178	struct ext4_super_block *es = sbi->s_es;
4179	ext4_group_t group;
4180	unsigned int len;
4181	ext4_fsblk_t goal;
4182	ext4_grpblk_t block;
4183
4184	/* we can't allocate > group size */
4185	len = ar->len;
4186
4187	/* just a dirty hack to filter too big requests  */
4188	if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4189		len = EXT4_CLUSTERS_PER_GROUP(sb);
4190
4191	/* start searching from the goal */
4192	goal = ar->goal;
4193	if (goal < le32_to_cpu(es->s_first_data_block) ||
4194			goal >= ext4_blocks_count(es))
4195		goal = le32_to_cpu(es->s_first_data_block);
4196	ext4_get_group_no_and_offset(sb, goal, &group, &block);
4197
4198	/* set up allocation goals */
4199	ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4200	ac->ac_status = AC_STATUS_CONTINUE;
4201	ac->ac_sb = sb;
4202	ac->ac_inode = ar->inode;
4203	ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4204	ac->ac_o_ex.fe_group = group;
4205	ac->ac_o_ex.fe_start = block;
4206	ac->ac_o_ex.fe_len = len;
4207	ac->ac_g_ex = ac->ac_o_ex;
4208	ac->ac_flags = ar->flags;
4209
4210	/* we have to define context: we'll we work with a file or
4211	 * locality group. this is a policy, actually */
4212	ext4_mb_group_or_file(ac);
4213
4214	mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4215			"left: %u/%u, right %u/%u to %swritable\n",
4216			(unsigned) ar->len, (unsigned) ar->logical,
4217			(unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4218			(unsigned) ar->lleft, (unsigned) ar->pleft,
4219			(unsigned) ar->lright, (unsigned) ar->pright,
4220			atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4221	return 0;
4222
4223}
4224
4225static noinline_for_stack void
4226ext4_mb_discard_lg_preallocations(struct super_block *sb,
4227					struct ext4_locality_group *lg,
4228					int order, int total_entries)
4229{
4230	ext4_group_t group = 0;
4231	struct ext4_buddy e4b;
4232	struct list_head discard_list;
4233	struct ext4_prealloc_space *pa, *tmp;
4234
4235	mb_debug(1, "discard locality group preallocation\n");
4236
4237	INIT_LIST_HEAD(&discard_list);
4238
4239	spin_lock(&lg->lg_prealloc_lock);
4240	list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4241						pa_inode_list) {
4242		spin_lock(&pa->pa_lock);
4243		if (atomic_read(&pa->pa_count)) {
4244			/*
4245			 * This is the pa that we just used
4246			 * for block allocation. So don't
4247			 * free that
4248			 */
4249			spin_unlock(&pa->pa_lock);
4250			continue;
4251		}
4252		if (pa->pa_deleted) {
4253			spin_unlock(&pa->pa_lock);
4254			continue;
4255		}
4256		/* only lg prealloc space */
4257		BUG_ON(pa->pa_type != MB_GROUP_PA);
4258
4259		/* seems this one can be freed ... */
4260		pa->pa_deleted = 1;
4261		spin_unlock(&pa->pa_lock);
4262
4263		list_del_rcu(&pa->pa_inode_list);
4264		list_add(&pa->u.pa_tmp_list, &discard_list);
4265
4266		total_entries--;
4267		if (total_entries <= 5) {
4268			/*
4269			 * we want to keep only 5 entries
4270			 * allowing it to grow to 8. This
4271			 * mak sure we don't call discard
4272			 * soon for this list.
4273			 */
4274			break;
4275		}
4276	}
4277	spin_unlock(&lg->lg_prealloc_lock);
4278
4279	list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4280
4281		group = ext4_get_group_number(sb, pa->pa_pstart);
4282		if (ext4_mb_load_buddy(sb, group, &e4b)) {
4283			ext4_error(sb, "Error loading buddy information for %u",
4284					group);
4285			continue;
4286		}
4287		ext4_lock_group(sb, group);
4288		list_del(&pa->pa_group_list);
4289		ext4_mb_release_group_pa(&e4b, pa);
4290		ext4_unlock_group(sb, group);
4291
4292		ext4_mb_unload_buddy(&e4b);
4293		list_del(&pa->u.pa_tmp_list);
4294		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4295	}
4296}
4297
4298/*
4299 * We have incremented pa_count. So it cannot be freed at this
4300 * point. Also we hold lg_mutex. So no parallel allocation is
4301 * possible from this lg. That means pa_free cannot be updated.
4302 *
4303 * A parallel ext4_mb_discard_group_preallocations is possible.
4304 * which can cause the lg_prealloc_list to be updated.
4305 */
4306
4307static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4308{
4309	int order, added = 0, lg_prealloc_count = 1;
4310	struct super_block *sb = ac->ac_sb;
4311	struct ext4_locality_group *lg = ac->ac_lg;
4312	struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4313
4314	order = fls(pa->pa_free) - 1;
4315	if (order > PREALLOC_TB_SIZE - 1)
4316		/* The max size of hash table is PREALLOC_TB_SIZE */
4317		order = PREALLOC_TB_SIZE - 1;
4318	/* Add the prealloc space to lg */
4319	spin_lock(&lg->lg_prealloc_lock);
4320	list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4321						pa_inode_list) {
4322		spin_lock(&tmp_pa->pa_lock);
4323		if (tmp_pa->pa_deleted) {
4324			spin_unlock(&tmp_pa->pa_lock);
4325			continue;
4326		}
4327		if (!added && pa->pa_free < tmp_pa->pa_free) {
4328			/* Add to the tail of the previous entry */
4329			list_add_tail_rcu(&pa->pa_inode_list,
4330						&tmp_pa->pa_inode_list);
4331			added = 1;
4332			/*
4333			 * we want to count the total
4334			 * number of entries in the list
4335			 */
4336		}
4337		spin_unlock(&tmp_pa->pa_lock);
4338		lg_prealloc_count++;
4339	}
4340	if (!added)
4341		list_add_tail_rcu(&pa->pa_inode_list,
4342					&lg->lg_prealloc_list[order]);
4343	spin_unlock(&lg->lg_prealloc_lock);
4344
4345	/* Now trim the list to be not more than 8 elements */
4346	if (lg_prealloc_count > 8) {
4347		ext4_mb_discard_lg_preallocations(sb, lg,
4348						  order, lg_prealloc_count);
4349		return;
4350	}
4351	return ;
4352}
4353
4354/*
4355 * release all resource we used in allocation
4356 */
4357static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4358{
4359	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4360	struct ext4_prealloc_space *pa = ac->ac_pa;
4361	if (pa) {
4362		if (pa->pa_type == MB_GROUP_PA) {
4363			/* see comment in ext4_mb_use_group_pa() */
4364			spin_lock(&pa->pa_lock);
4365			pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4366			pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4367			pa->pa_free -= ac->ac_b_ex.fe_len;
4368			pa->pa_len -= ac->ac_b_ex.fe_len;
4369			spin_unlock(&pa->pa_lock);
4370		}
4371	}
4372	if (pa) {
4373		/*
4374		 * We want to add the pa to the right bucket.
4375		 * Remove it from the list and while adding
4376		 * make sure the list to which we are adding
4377		 * doesn't grow big.
4378		 */
4379		if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4380			spin_lock(pa->pa_obj_lock);
4381			list_del_rcu(&pa->pa_inode_list);
4382			spin_unlock(pa->pa_obj_lock);
4383			ext4_mb_add_n_trim(ac);
4384		}
4385		ext4_mb_put_pa(ac, ac->ac_sb, pa);
4386	}
4387	if (ac->ac_bitmap_page)
4388		put_page(ac->ac_bitmap_page);
4389	if (ac->ac_buddy_page)
4390		put_page(ac->ac_buddy_page);
4391	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4392		mutex_unlock(&ac->ac_lg->lg_mutex);
4393	ext4_mb_collect_stats(ac);
4394	return 0;
4395}
4396
4397static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4398{
4399	ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4400	int ret;
4401	int freed = 0;
4402
4403	trace_ext4_mb_discard_preallocations(sb, needed);
4404	for (i = 0; i < ngroups && needed > 0; i++) {
4405		ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4406		freed += ret;
4407		needed -= ret;
4408	}
4409
4410	return freed;
4411}
4412
4413/*
4414 * Main entry point into mballoc to allocate blocks
4415 * it tries to use preallocation first, then falls back
4416 * to usual allocation
4417 */
4418ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4419				struct ext4_allocation_request *ar, int *errp)
4420{
4421	int freed;
4422	struct ext4_allocation_context *ac = NULL;
4423	struct ext4_sb_info *sbi;
4424	struct super_block *sb;
4425	ext4_fsblk_t block = 0;
4426	unsigned int inquota = 0;
4427	unsigned int reserv_clstrs = 0;
4428
4429	might_sleep();
4430	sb = ar->inode->i_sb;
4431	sbi = EXT4_SB(sb);
4432
4433	trace_ext4_request_blocks(ar);
4434
4435	/* Allow to use superuser reservation for quota file */
4436	if (IS_NOQUOTA(ar->inode))
4437		ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4438
4439	if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4440		/* Without delayed allocation we need to verify
4441		 * there is enough free blocks to do block allocation
4442		 * and verify allocation doesn't exceed the quota limits.
4443		 */
4444		while (ar->len &&
4445			ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4446
4447			/* let others to free the space */
4448			cond_resched();
4449			ar->len = ar->len >> 1;
4450		}
4451		if (!ar->len) {
4452			*errp = -ENOSPC;
4453			return 0;
4454		}
4455		reserv_clstrs = ar->len;
4456		if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4457			dquot_alloc_block_nofail(ar->inode,
4458						 EXT4_C2B(sbi, ar->len));
4459		} else {
4460			while (ar->len &&
4461				dquot_alloc_block(ar->inode,
4462						  EXT4_C2B(sbi, ar->len))) {
4463
4464				ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4465				ar->len--;
4466			}
4467		}
4468		inquota = ar->len;
4469		if (ar->len == 0) {
4470			*errp = -EDQUOT;
4471			goto out;
4472		}
4473	}
4474
4475	ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4476	if (!ac) {
4477		ar->len = 0;
4478		*errp = -ENOMEM;
4479		goto out;
4480	}
4481
4482	*errp = ext4_mb_initialize_context(ac, ar);
4483	if (*errp) {
4484		ar->len = 0;
4485		goto out;
4486	}
4487
4488	ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4489	if (!ext4_mb_use_preallocated(ac)) {
4490		ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4491		ext4_mb_normalize_request(ac, ar);
4492repeat:
4493		/* allocate space in core */
4494		*errp = ext4_mb_regular_allocator(ac);
4495		if (*errp)
4496			goto discard_and_exit;
4497
4498		/* as we've just preallocated more space than
4499		 * user requested originally, we store allocated
4500		 * space in a special descriptor */
4501		if (ac->ac_status == AC_STATUS_FOUND &&
4502		    ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4503			*errp = ext4_mb_new_preallocation(ac);
4504		if (*errp) {
4505		discard_and_exit:
4506			ext4_discard_allocated_blocks(ac);
4507			goto errout;
4508		}
4509	}
4510	if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4511		*errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4512		if (*errp == -EAGAIN) {
4513			/*
4514			 * drop the reference that we took
4515			 * in ext4_mb_use_best_found
4516			 */
4517			ext4_mb_release_context(ac);
4518			ac->ac_b_ex.fe_group = 0;
4519			ac->ac_b_ex.fe_start = 0;
4520			ac->ac_b_ex.fe_len = 0;
4521			ac->ac_status = AC_STATUS_CONTINUE;
4522			goto repeat;
4523		} else if (*errp) {
4524			ext4_discard_allocated_blocks(ac);
4525			goto errout;
4526		} else {
4527			block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4528			ar->len = ac->ac_b_ex.fe_len;
4529		}
4530	} else {
4531		freed  = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4532		if (freed)
4533			goto repeat;
4534		*errp = -ENOSPC;
4535	}
4536
4537errout:
4538	if (*errp) {
4539		ac->ac_b_ex.fe_len = 0;
4540		ar->len = 0;
4541		ext4_mb_show_ac(ac);
4542	}
4543	ext4_mb_release_context(ac);
4544out:
4545	if (ac)
4546		kmem_cache_free(ext4_ac_cachep, ac);
4547	if (inquota && ar->len < inquota)
4548		dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4549	if (!ar->len) {
4550		if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4551			/* release all the reserved blocks if non delalloc */
4552			percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4553						reserv_clstrs);
4554	}
4555
4556	trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4557
4558	return block;
4559}
4560
4561/*
4562 * We can merge two free data extents only if the physical blocks
4563 * are contiguous, AND the extents were freed by the same transaction,
4564 * AND the blocks are associated with the same group.
4565 */
4566static int can_merge(struct ext4_free_data *entry1,
4567			struct ext4_free_data *entry2)
4568{
4569	if ((entry1->efd_tid == entry2->efd_tid) &&
4570	    (entry1->efd_group == entry2->efd_group) &&
4571	    ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4572		return 1;
4573	return 0;
4574}
4575
4576static noinline_for_stack int
4577ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4578		      struct ext4_free_data *new_entry)
4579{
4580	ext4_group_t group = e4b->bd_group;
4581	ext4_grpblk_t cluster;
 
4582	struct ext4_free_data *entry;
4583	struct ext4_group_info *db = e4b->bd_info;
4584	struct super_block *sb = e4b->bd_sb;
4585	struct ext4_sb_info *sbi = EXT4_SB(sb);
4586	struct rb_node **n = &db->bb_free_root.rb_node, *node;
4587	struct rb_node *parent = NULL, *new_node;
4588
4589	BUG_ON(!ext4_handle_valid(handle));
4590	BUG_ON(e4b->bd_bitmap_page == NULL);
4591	BUG_ON(e4b->bd_buddy_page == NULL);
4592
4593	new_node = &new_entry->efd_node;
4594	cluster = new_entry->efd_start_cluster;
4595
4596	if (!*n) {
4597		/* first free block exent. We need to
4598		   protect buddy cache from being freed,
4599		 * otherwise we'll refresh it from
4600		 * on-disk bitmap and lose not-yet-available
4601		 * blocks */
4602		get_page(e4b->bd_buddy_page);
4603		get_page(e4b->bd_bitmap_page);
4604	}
4605	while (*n) {
4606		parent = *n;
4607		entry = rb_entry(parent, struct ext4_free_data, efd_node);
4608		if (cluster < entry->efd_start_cluster)
4609			n = &(*n)->rb_left;
4610		else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4611			n = &(*n)->rb_right;
4612		else {
4613			ext4_grp_locked_error(sb, group, 0,
4614				ext4_group_first_block_no(sb, group) +
4615				EXT4_C2B(sbi, cluster),
4616				"Block already on to-be-freed list");
4617			return 0;
4618		}
4619	}
4620
4621	rb_link_node(new_node, parent, n);
4622	rb_insert_color(new_node, &db->bb_free_root);
4623
4624	/* Now try to see the extent can be merged to left and right */
4625	node = rb_prev(new_node);
4626	if (node) {
4627		entry = rb_entry(node, struct ext4_free_data, efd_node);
4628		if (can_merge(entry, new_entry) &&
4629		    ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4630			new_entry->efd_start_cluster = entry->efd_start_cluster;
4631			new_entry->efd_count += entry->efd_count;
4632			rb_erase(node, &(db->bb_free_root));
4633			kmem_cache_free(ext4_free_data_cachep, entry);
4634		}
4635	}
4636
4637	node = rb_next(new_node);
4638	if (node) {
4639		entry = rb_entry(node, struct ext4_free_data, efd_node);
4640		if (can_merge(new_entry, entry) &&
4641		    ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4642			new_entry->efd_count += entry->efd_count;
4643			rb_erase(node, &(db->bb_free_root));
4644			kmem_cache_free(ext4_free_data_cachep, entry);
4645		}
4646	}
4647	/* Add the extent to transaction's private list */
4648	ext4_journal_callback_add(handle, ext4_free_data_callback,
4649				  &new_entry->efd_jce);
 
 
 
4650	return 0;
4651}
4652
4653/**
4654 * ext4_free_blocks() -- Free given blocks and update quota
4655 * @handle:		handle for this transaction
4656 * @inode:		inode
4657 * @block:		start physical block to free
4658 * @count:		number of blocks to count
4659 * @flags:		flags used by ext4_free_blocks
4660 */
4661void ext4_free_blocks(handle_t *handle, struct inode *inode,
4662		      struct buffer_head *bh, ext4_fsblk_t block,
4663		      unsigned long count, int flags)
4664{
4665	struct buffer_head *bitmap_bh = NULL;
4666	struct super_block *sb = inode->i_sb;
4667	struct ext4_group_desc *gdp;
4668	unsigned int overflow;
4669	ext4_grpblk_t bit;
4670	struct buffer_head *gd_bh;
4671	ext4_group_t block_group;
4672	struct ext4_sb_info *sbi;
4673	struct ext4_buddy e4b;
4674	unsigned int count_clusters;
4675	int err = 0;
4676	int ret;
4677
4678	might_sleep();
4679	if (bh) {
4680		if (block)
4681			BUG_ON(block != bh->b_blocknr);
4682		else
4683			block = bh->b_blocknr;
4684	}
4685
4686	sbi = EXT4_SB(sb);
4687	if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4688	    !ext4_data_block_valid(sbi, block, count)) {
4689		ext4_error(sb, "Freeing blocks not in datazone - "
4690			   "block = %llu, count = %lu", block, count);
4691		goto error_return;
4692	}
4693
4694	ext4_debug("freeing block %llu\n", block);
4695	trace_ext4_free_blocks(inode, block, count, flags);
4696
4697	if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4698		BUG_ON(count > 1);
4699
4700		ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4701			    inode, bh, block);
4702	}
4703
4704	/*
4705	 * If the extent to be freed does not begin on a cluster
4706	 * boundary, we need to deal with partial clusters at the
4707	 * beginning and end of the extent.  Normally we will free
4708	 * blocks at the beginning or the end unless we are explicitly
4709	 * requested to avoid doing so.
4710	 */
4711	overflow = EXT4_PBLK_COFF(sbi, block);
4712	if (overflow) {
4713		if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4714			overflow = sbi->s_cluster_ratio - overflow;
4715			block += overflow;
4716			if (count > overflow)
4717				count -= overflow;
4718			else
4719				return;
4720		} else {
4721			block -= overflow;
4722			count += overflow;
4723		}
4724	}
4725	overflow = EXT4_LBLK_COFF(sbi, count);
4726	if (overflow) {
4727		if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4728			if (count > overflow)
4729				count -= overflow;
4730			else
4731				return;
4732		} else
4733			count += sbi->s_cluster_ratio - overflow;
4734	}
4735
4736	if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4737		int i;
4738		int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
4739
4740		for (i = 0; i < count; i++) {
4741			cond_resched();
4742			if (is_metadata)
4743				bh = sb_find_get_block(inode->i_sb, block + i);
4744			ext4_forget(handle, is_metadata, inode, bh, block + i);
4745		}
4746	}
4747
4748do_more:
4749	overflow = 0;
4750	ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4751
4752	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4753			ext4_get_group_info(sb, block_group))))
4754		return;
4755
4756	/*
4757	 * Check to see if we are freeing blocks across a group
4758	 * boundary.
4759	 */
4760	if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4761		overflow = EXT4_C2B(sbi, bit) + count -
4762			EXT4_BLOCKS_PER_GROUP(sb);
4763		count -= overflow;
4764	}
4765	count_clusters = EXT4_NUM_B2C(sbi, count);
4766	bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4767	if (IS_ERR(bitmap_bh)) {
4768		err = PTR_ERR(bitmap_bh);
4769		bitmap_bh = NULL;
4770		goto error_return;
4771	}
4772	gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4773	if (!gdp) {
4774		err = -EIO;
4775		goto error_return;
4776	}
4777
4778	if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4779	    in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4780	    in_range(block, ext4_inode_table(sb, gdp),
4781		     EXT4_SB(sb)->s_itb_per_group) ||
4782	    in_range(block + count - 1, ext4_inode_table(sb, gdp),
4783		     EXT4_SB(sb)->s_itb_per_group)) {
4784
4785		ext4_error(sb, "Freeing blocks in system zone - "
4786			   "Block = %llu, count = %lu", block, count);
4787		/* err = 0. ext4_std_error should be a no op */
4788		goto error_return;
4789	}
4790
4791	BUFFER_TRACE(bitmap_bh, "getting write access");
4792	err = ext4_journal_get_write_access(handle, bitmap_bh);
4793	if (err)
4794		goto error_return;
4795
4796	/*
4797	 * We are about to modify some metadata.  Call the journal APIs
4798	 * to unshare ->b_data if a currently-committing transaction is
4799	 * using it
4800	 */
4801	BUFFER_TRACE(gd_bh, "get_write_access");
4802	err = ext4_journal_get_write_access(handle, gd_bh);
4803	if (err)
4804		goto error_return;
4805#ifdef AGGRESSIVE_CHECK
4806	{
4807		int i;
4808		for (i = 0; i < count_clusters; i++)
4809			BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4810	}
4811#endif
4812	trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4813
4814	/* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4815	err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
4816				     GFP_NOFS|__GFP_NOFAIL);
4817	if (err)
4818		goto error_return;
4819
4820	/*
4821	 * We need to make sure we don't reuse the freed block until after the
4822	 * transaction is committed. We make an exception if the inode is to be
4823	 * written in writeback mode since writeback mode has weak data
4824	 * consistency guarantees.
4825	 */
4826	if (ext4_handle_valid(handle) &&
4827	    ((flags & EXT4_FREE_BLOCKS_METADATA) ||
4828	     !ext4_should_writeback_data(inode))) {
4829		struct ext4_free_data *new_entry;
4830		/*
4831		 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4832		 * to fail.
4833		 */
4834		new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4835				GFP_NOFS|__GFP_NOFAIL);
4836		new_entry->efd_start_cluster = bit;
4837		new_entry->efd_group = block_group;
4838		new_entry->efd_count = count_clusters;
4839		new_entry->efd_tid = handle->h_transaction->t_tid;
4840
4841		ext4_lock_group(sb, block_group);
4842		mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4843		ext4_mb_free_metadata(handle, &e4b, new_entry);
4844	} else {
4845		/* need to update group_info->bb_free and bitmap
4846		 * with group lock held. generate_buddy look at
4847		 * them with group lock_held
4848		 */
4849		if (test_opt(sb, DISCARD)) {
4850			err = ext4_issue_discard(sb, block_group, bit, count);
4851			if (err && err != -EOPNOTSUPP)
4852				ext4_msg(sb, KERN_WARNING, "discard request in"
4853					 " group:%d block:%d count:%lu failed"
4854					 " with %d", block_group, bit, count,
4855					 err);
4856		} else
4857			EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4858
4859		ext4_lock_group(sb, block_group);
4860		mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4861		mb_free_blocks(inode, &e4b, bit, count_clusters);
4862	}
4863
4864	ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4865	ext4_free_group_clusters_set(sb, gdp, ret);
4866	ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4867	ext4_group_desc_csum_set(sb, block_group, gdp);
4868	ext4_unlock_group(sb, block_group);
4869
4870	if (sbi->s_log_groups_per_flex) {
4871		ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4872		atomic64_add(count_clusters,
4873			     &sbi->s_flex_groups[flex_group].free_clusters);
4874	}
4875
4876	if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4877		dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4878	percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4879
4880	ext4_mb_unload_buddy(&e4b);
4881
4882	/* We dirtied the bitmap block */
4883	BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4884	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4885
4886	/* And the group descriptor block */
4887	BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4888	ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4889	if (!err)
4890		err = ret;
4891
4892	if (overflow && !err) {
4893		block += count;
4894		count = overflow;
4895		put_bh(bitmap_bh);
4896		goto do_more;
4897	}
4898error_return:
4899	brelse(bitmap_bh);
4900	ext4_std_error(sb, err);
4901	return;
4902}
4903
4904/**
4905 * ext4_group_add_blocks() -- Add given blocks to an existing group
4906 * @handle:			handle to this transaction
4907 * @sb:				super block
4908 * @block:			start physical block to add to the block group
4909 * @count:			number of blocks to free
4910 *
4911 * This marks the blocks as free in the bitmap and buddy.
4912 */
4913int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4914			 ext4_fsblk_t block, unsigned long count)
4915{
4916	struct buffer_head *bitmap_bh = NULL;
4917	struct buffer_head *gd_bh;
4918	ext4_group_t block_group;
4919	ext4_grpblk_t bit;
4920	unsigned int i;
4921	struct ext4_group_desc *desc;
4922	struct ext4_sb_info *sbi = EXT4_SB(sb);
4923	struct ext4_buddy e4b;
4924	int err = 0, ret, blk_free_count;
4925	ext4_grpblk_t blocks_freed;
4926
4927	ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4928
4929	if (count == 0)
4930		return 0;
4931
4932	ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4933	/*
4934	 * Check to see if we are freeing blocks across a group
4935	 * boundary.
4936	 */
4937	if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4938		ext4_warning(sb, "too much blocks added to group %u\n",
4939			     block_group);
4940		err = -EINVAL;
4941		goto error_return;
4942	}
4943
4944	bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4945	if (IS_ERR(bitmap_bh)) {
4946		err = PTR_ERR(bitmap_bh);
4947		bitmap_bh = NULL;
4948		goto error_return;
4949	}
4950
4951	desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4952	if (!desc) {
4953		err = -EIO;
4954		goto error_return;
4955	}
4956
4957	if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4958	    in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4959	    in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4960	    in_range(block + count - 1, ext4_inode_table(sb, desc),
4961		     sbi->s_itb_per_group)) {
4962		ext4_error(sb, "Adding blocks in system zones - "
4963			   "Block = %llu, count = %lu",
4964			   block, count);
4965		err = -EINVAL;
4966		goto error_return;
4967	}
4968
4969	BUFFER_TRACE(bitmap_bh, "getting write access");
4970	err = ext4_journal_get_write_access(handle, bitmap_bh);
4971	if (err)
4972		goto error_return;
4973
4974	/*
4975	 * We are about to modify some metadata.  Call the journal APIs
4976	 * to unshare ->b_data if a currently-committing transaction is
4977	 * using it
4978	 */
4979	BUFFER_TRACE(gd_bh, "get_write_access");
4980	err = ext4_journal_get_write_access(handle, gd_bh);
4981	if (err)
4982		goto error_return;
4983
4984	for (i = 0, blocks_freed = 0; i < count; i++) {
4985		BUFFER_TRACE(bitmap_bh, "clear bit");
4986		if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4987			ext4_error(sb, "bit already cleared for block %llu",
4988				   (ext4_fsblk_t)(block + i));
4989			BUFFER_TRACE(bitmap_bh, "bit already cleared");
4990		} else {
4991			blocks_freed++;
4992		}
4993	}
4994
4995	err = ext4_mb_load_buddy(sb, block_group, &e4b);
4996	if (err)
4997		goto error_return;
4998
4999	/*
5000	 * need to update group_info->bb_free and bitmap
5001	 * with group lock held. generate_buddy look at
5002	 * them with group lock_held
5003	 */
5004	ext4_lock_group(sb, block_group);
5005	mb_clear_bits(bitmap_bh->b_data, bit, count);
5006	mb_free_blocks(NULL, &e4b, bit, count);
5007	blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5008	ext4_free_group_clusters_set(sb, desc, blk_free_count);
5009	ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5010	ext4_group_desc_csum_set(sb, block_group, desc);
5011	ext4_unlock_group(sb, block_group);
5012	percpu_counter_add(&sbi->s_freeclusters_counter,
5013			   EXT4_NUM_B2C(sbi, blocks_freed));
5014
5015	if (sbi->s_log_groups_per_flex) {
5016		ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5017		atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5018			     &sbi->s_flex_groups[flex_group].free_clusters);
5019	}
5020
5021	ext4_mb_unload_buddy(&e4b);
5022
5023	/* We dirtied the bitmap block */
5024	BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5025	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5026
5027	/* And the group descriptor block */
5028	BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5029	ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5030	if (!err)
5031		err = ret;
5032
5033error_return:
5034	brelse(bitmap_bh);
5035	ext4_std_error(sb, err);
5036	return err;
5037}
5038
5039/**
5040 * ext4_trim_extent -- function to TRIM one single free extent in the group
5041 * @sb:		super block for the file system
5042 * @start:	starting block of the free extent in the alloc. group
5043 * @count:	number of blocks to TRIM
5044 * @group:	alloc. group we are working with
5045 * @e4b:	ext4 buddy for the group
5046 *
5047 * Trim "count" blocks starting at "start" in the "group". To assure that no
5048 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5049 * be called with under the group lock.
5050 */
5051static int ext4_trim_extent(struct super_block *sb, int start, int count,
5052			     ext4_group_t group, struct ext4_buddy *e4b)
5053__releases(bitlock)
5054__acquires(bitlock)
5055{
5056	struct ext4_free_extent ex;
5057	int ret = 0;
5058
5059	trace_ext4_trim_extent(sb, group, start, count);
5060
5061	assert_spin_locked(ext4_group_lock_ptr(sb, group));
5062
5063	ex.fe_start = start;
5064	ex.fe_group = group;
5065	ex.fe_len = count;
5066
5067	/*
5068	 * Mark blocks used, so no one can reuse them while
5069	 * being trimmed.
5070	 */
5071	mb_mark_used(e4b, &ex);
5072	ext4_unlock_group(sb, group);
5073	ret = ext4_issue_discard(sb, group, start, count);
5074	ext4_lock_group(sb, group);
5075	mb_free_blocks(NULL, e4b, start, ex.fe_len);
5076	return ret;
5077}
5078
5079/**
5080 * ext4_trim_all_free -- function to trim all free space in alloc. group
5081 * @sb:			super block for file system
5082 * @group:		group to be trimmed
5083 * @start:		first group block to examine
5084 * @max:		last group block to examine
5085 * @minblocks:		minimum extent block count
5086 *
5087 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5088 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5089 * the extent.
5090 *
5091 *
5092 * ext4_trim_all_free walks through group's block bitmap searching for free
5093 * extents. When the free extent is found, mark it as used in group buddy
5094 * bitmap. Then issue a TRIM command on this extent and free the extent in
5095 * the group buddy bitmap. This is done until whole group is scanned.
5096 */
5097static ext4_grpblk_t
5098ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5099		   ext4_grpblk_t start, ext4_grpblk_t max,
5100		   ext4_grpblk_t minblocks)
5101{
5102	void *bitmap;
5103	ext4_grpblk_t next, count = 0, free_count = 0;
5104	struct ext4_buddy e4b;
5105	int ret = 0;
5106
5107	trace_ext4_trim_all_free(sb, group, start, max);
5108
5109	ret = ext4_mb_load_buddy(sb, group, &e4b);
5110	if (ret) {
5111		ext4_error(sb, "Error in loading buddy "
5112				"information for %u", group);
5113		return ret;
5114	}
5115	bitmap = e4b.bd_bitmap;
5116
5117	ext4_lock_group(sb, group);
5118	if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5119	    minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5120		goto out;
5121
5122	start = (e4b.bd_info->bb_first_free > start) ?
5123		e4b.bd_info->bb_first_free : start;
5124
5125	while (start <= max) {
5126		start = mb_find_next_zero_bit(bitmap, max + 1, start);
5127		if (start > max)
5128			break;
5129		next = mb_find_next_bit(bitmap, max + 1, start);
5130
5131		if ((next - start) >= minblocks) {
5132			ret = ext4_trim_extent(sb, start,
5133					       next - start, group, &e4b);
5134			if (ret && ret != -EOPNOTSUPP)
5135				break;
5136			ret = 0;
5137			count += next - start;
5138		}
5139		free_count += next - start;
5140		start = next + 1;
5141
5142		if (fatal_signal_pending(current)) {
5143			count = -ERESTARTSYS;
5144			break;
5145		}
5146
5147		if (need_resched()) {
5148			ext4_unlock_group(sb, group);
5149			cond_resched();
5150			ext4_lock_group(sb, group);
5151		}
5152
5153		if ((e4b.bd_info->bb_free - free_count) < minblocks)
5154			break;
5155	}
5156
5157	if (!ret) {
5158		ret = count;
5159		EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5160	}
5161out:
5162	ext4_unlock_group(sb, group);
5163	ext4_mb_unload_buddy(&e4b);
5164
5165	ext4_debug("trimmed %d blocks in the group %d\n",
5166		count, group);
5167
5168	return ret;
5169}
5170
5171/**
5172 * ext4_trim_fs() -- trim ioctl handle function
5173 * @sb:			superblock for filesystem
5174 * @range:		fstrim_range structure
5175 *
5176 * start:	First Byte to trim
5177 * len:		number of Bytes to trim from start
5178 * minlen:	minimum extent length in Bytes
5179 * ext4_trim_fs goes through all allocation groups containing Bytes from
5180 * start to start+len. For each such a group ext4_trim_all_free function
5181 * is invoked to trim all free space.
5182 */
5183int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5184{
5185	struct ext4_group_info *grp;
5186	ext4_group_t group, first_group, last_group;
5187	ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5188	uint64_t start, end, minlen, trimmed = 0;
5189	ext4_fsblk_t first_data_blk =
5190			le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5191	ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5192	int ret = 0;
5193
5194	start = range->start >> sb->s_blocksize_bits;
5195	end = start + (range->len >> sb->s_blocksize_bits) - 1;
5196	minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5197			      range->minlen >> sb->s_blocksize_bits);
5198
5199	if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5200	    start >= max_blks ||
5201	    range->len < sb->s_blocksize)
5202		return -EINVAL;
5203	if (end >= max_blks)
5204		end = max_blks - 1;
5205	if (end <= first_data_blk)
5206		goto out;
5207	if (start < first_data_blk)
5208		start = first_data_blk;
5209
5210	/* Determine first and last group to examine based on start and end */
5211	ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5212				     &first_group, &first_cluster);
5213	ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5214				     &last_group, &last_cluster);
5215
5216	/* end now represents the last cluster to discard in this group */
5217	end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5218
5219	for (group = first_group; group <= last_group; group++) {
5220		grp = ext4_get_group_info(sb, group);
5221		/* We only do this if the grp has never been initialized */
5222		if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5223			ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5224			if (ret)
5225				break;
5226		}
5227
5228		/*
5229		 * For all the groups except the last one, last cluster will
5230		 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5231		 * change it for the last group, note that last_cluster is
5232		 * already computed earlier by ext4_get_group_no_and_offset()
5233		 */
5234		if (group == last_group)
5235			end = last_cluster;
5236
5237		if (grp->bb_free >= minlen) {
5238			cnt = ext4_trim_all_free(sb, group, first_cluster,
5239						end, minlen);
5240			if (cnt < 0) {
5241				ret = cnt;
5242				break;
5243			}
5244			trimmed += cnt;
5245		}
5246
5247		/*
5248		 * For every group except the first one, we are sure
5249		 * that the first cluster to discard will be cluster #0.
5250		 */
5251		first_cluster = 0;
5252	}
5253
5254	if (!ret)
5255		atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5256
5257out:
5258	range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5259	return ret;
5260}
v4.10.11
   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 License
  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/log2.h>
  27#include <linux/module.h>
  28#include <linux/slab.h>
  29#include <linux/backing-dev.h>
  30#include <trace/events/ext4.h>
  31
  32#ifdef CONFIG_EXT4_DEBUG
  33ushort ext4_mballoc_debug __read_mostly;
  34
  35module_param_named(mballoc_debug, ext4_mballoc_debug, ushort, 0644);
  36MODULE_PARM_DESC(mballoc_debug, "Debugging level for ext4's mballoc");
  37#endif
  38
  39/*
  40 * MUSTDO:
  41 *   - test ext4_ext_search_left() and ext4_ext_search_right()
  42 *   - search for metadata in few groups
  43 *
  44 * TODO v4:
  45 *   - normalization should take into account whether file is still open
  46 *   - discard preallocations if no free space left (policy?)
  47 *   - don't normalize tails
  48 *   - quota
  49 *   - reservation for superuser
  50 *
  51 * TODO v3:
  52 *   - bitmap read-ahead (proposed by Oleg Drokin aka green)
  53 *   - track min/max extents in each group for better group selection
  54 *   - mb_mark_used() may allocate chunk right after splitting buddy
  55 *   - tree of groups sorted by number of free blocks
  56 *   - error handling
  57 */
  58
  59/*
  60 * The allocation request involve request for multiple number of blocks
  61 * near to the goal(block) value specified.
  62 *
  63 * During initialization phase of the allocator we decide to use the
  64 * group preallocation or inode preallocation depending on the size of
  65 * the file. The size of the file could be the resulting file size we
  66 * would have after allocation, or the current file size, which ever
  67 * is larger. If the size is less than sbi->s_mb_stream_request we
  68 * select to use the group preallocation. The default value of
  69 * s_mb_stream_request is 16 blocks. This can also be tuned via
  70 * /sys/fs/ext4/<partition>/mb_stream_req. The value is represented in
  71 * terms of number of blocks.
  72 *
  73 * The main motivation for having small file use group preallocation is to
  74 * ensure that we have small files closer together on the disk.
  75 *
  76 * First stage the allocator looks at the inode prealloc list,
  77 * ext4_inode_info->i_prealloc_list, which contains list of prealloc
  78 * spaces for this particular inode. The inode prealloc space is
  79 * represented as:
  80 *
  81 * pa_lstart -> the logical start block for this prealloc space
  82 * pa_pstart -> the physical start block for this prealloc space
  83 * pa_len    -> length for this prealloc space (in clusters)
  84 * pa_free   ->  free space available in this prealloc space (in clusters)
  85 *
  86 * The inode preallocation space is used looking at the _logical_ start
  87 * block. If only the logical file block falls within the range of prealloc
  88 * space we will consume the particular prealloc space. This makes sure that
  89 * we have contiguous physical blocks representing the file blocks
  90 *
  91 * The important thing to be noted in case of inode prealloc space is that
  92 * we don't modify the values associated to inode prealloc space except
  93 * pa_free.
  94 *
  95 * If we are not able to find blocks in the inode prealloc space and if we
  96 * have the group allocation flag set then we look at the locality group
  97 * prealloc space. These are per CPU prealloc list represented as
  98 *
  99 * ext4_sb_info.s_locality_groups[smp_processor_id()]
 100 *
 101 * The reason for having a per cpu locality group is to reduce the contention
 102 * between CPUs. It is possible to get scheduled at this point.
 103 *
 104 * The locality group prealloc space is used looking at whether we have
 105 * enough free space (pa_free) within the prealloc space.
 106 *
 107 * If we can't allocate blocks via inode prealloc or/and locality group
 108 * prealloc then we look at the buddy cache. The buddy cache is represented
 109 * by ext4_sb_info.s_buddy_cache (struct inode) whose file offset gets
 110 * mapped to the buddy and bitmap information regarding different
 111 * groups. The buddy information is attached to buddy cache inode so that
 112 * we can access them through the page cache. The information regarding
 113 * each group is loaded via ext4_mb_load_buddy.  The information involve
 114 * block bitmap and buddy information. The information are stored in the
 115 * inode as:
 116 *
 117 *  {                        page                        }
 118 *  [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
 119 *
 120 *
 121 * one block each for bitmap and buddy information.  So for each group we
 122 * take up 2 blocks. A page can contain blocks_per_page (PAGE_SIZE /
 123 * blocksize) blocks.  So it can have information regarding groups_per_page
 124 * which is blocks_per_page/2
 125 *
 126 * The buddy cache inode is not stored on disk. The inode is thrown
 127 * away when the filesystem is unmounted.
 128 *
 129 * We look for count number of blocks in the buddy cache. If we were able
 130 * to locate that many free blocks we return with additional information
 131 * regarding rest of the contiguous physical block available
 132 *
 133 * Before allocating blocks via buddy cache we normalize the request
 134 * blocks. This ensure we ask for more blocks that we needed. The extra
 135 * blocks that we get after allocation is added to the respective prealloc
 136 * list. In case of inode preallocation we follow a list of heuristics
 137 * based on file size. This can be found in ext4_mb_normalize_request. If
 138 * we are doing a group prealloc we try to normalize the request to
 139 * sbi->s_mb_group_prealloc.  The default value of s_mb_group_prealloc is
 140 * dependent on the cluster size; for non-bigalloc file systems, it is
 141 * 512 blocks. This can be tuned via
 142 * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
 143 * terms of number of blocks. If we have mounted the file system with -O
 144 * stripe=<value> option the group prealloc request is normalized to the
 145 * the smallest multiple of the stripe value (sbi->s_stripe) which is
 146 * greater than the default mb_group_prealloc.
 147 *
 148 * The regular allocator (using the buddy cache) supports a few tunables.
 149 *
 150 * /sys/fs/ext4/<partition>/mb_min_to_scan
 151 * /sys/fs/ext4/<partition>/mb_max_to_scan
 152 * /sys/fs/ext4/<partition>/mb_order2_req
 153 *
 154 * The regular allocator uses buddy scan only if the request len is power of
 155 * 2 blocks and the order of allocation is >= sbi->s_mb_order2_reqs. The
 156 * value of s_mb_order2_reqs can be tuned via
 157 * /sys/fs/ext4/<partition>/mb_order2_req.  If the request len is equal to
 158 * stripe size (sbi->s_stripe), we try to search for contiguous block in
 159 * stripe size. This should result in better allocation on RAID setups. If
 160 * not, we search in the specific group using bitmap for best extents. The
 161 * tunable min_to_scan and max_to_scan control the behaviour here.
 162 * min_to_scan indicate how long the mballoc __must__ look for a best
 163 * extent and max_to_scan indicates how long the mballoc __can__ look for a
 164 * best extent in the found extents. Searching for the blocks starts with
 165 * the group specified as the goal value in allocation context via
 166 * ac_g_ex. Each group is first checked based on the criteria whether it
 167 * can be used for allocation. ext4_mb_good_group explains how the groups are
 168 * checked.
 169 *
 170 * Both the prealloc space are getting populated as above. So for the first
 171 * request we will hit the buddy cache which will result in this prealloc
 172 * space getting filled. The prealloc space is then later used for the
 173 * subsequent request.
 174 */
 175
 176/*
 177 * mballoc operates on the following data:
 178 *  - on-disk bitmap
 179 *  - in-core buddy (actually includes buddy and bitmap)
 180 *  - preallocation descriptors (PAs)
 181 *
 182 * there are two types of preallocations:
 183 *  - inode
 184 *    assiged to specific inode and can be used for this inode only.
 185 *    it describes part of inode's space preallocated to specific
 186 *    physical blocks. any block from that preallocated can be used
 187 *    independent. the descriptor just tracks number of blocks left
 188 *    unused. so, before taking some block from descriptor, one must
 189 *    make sure corresponded logical block isn't allocated yet. this
 190 *    also means that freeing any block within descriptor's range
 191 *    must discard all preallocated blocks.
 192 *  - locality group
 193 *    assigned to specific locality group which does not translate to
 194 *    permanent set of inodes: inode can join and leave group. space
 195 *    from this type of preallocation can be used for any inode. thus
 196 *    it's consumed from the beginning to the end.
 197 *
 198 * relation between them can be expressed as:
 199 *    in-core buddy = on-disk bitmap + preallocation descriptors
 200 *
 201 * this mean blocks mballoc considers used are:
 202 *  - allocated blocks (persistent)
 203 *  - preallocated blocks (non-persistent)
 204 *
 205 * consistency in mballoc world means that at any time a block is either
 206 * free or used in ALL structures. notice: "any time" should not be read
 207 * literally -- time is discrete and delimited by locks.
 208 *
 209 *  to keep it simple, we don't use block numbers, instead we count number of
 210 *  blocks: how many blocks marked used/free in on-disk bitmap, buddy and PA.
 211 *
 212 * all operations can be expressed as:
 213 *  - init buddy:			buddy = on-disk + PAs
 214 *  - new PA:				buddy += N; PA = N
 215 *  - use inode PA:			on-disk += N; PA -= N
 216 *  - discard inode PA			buddy -= on-disk - PA; PA = 0
 217 *  - use locality group PA		on-disk += N; PA -= N
 218 *  - discard locality group PA		buddy -= PA; PA = 0
 219 *  note: 'buddy -= on-disk - PA' is used to show that on-disk bitmap
 220 *        is used in real operation because we can't know actual used
 221 *        bits from PA, only from on-disk bitmap
 222 *
 223 * if we follow this strict logic, then all operations above should be atomic.
 224 * given some of them can block, we'd have to use something like semaphores
 225 * killing performance on high-end SMP hardware. let's try to relax it using
 226 * the following knowledge:
 227 *  1) if buddy is referenced, it's already initialized
 228 *  2) while block is used in buddy and the buddy is referenced,
 229 *     nobody can re-allocate that block
 230 *  3) we work on bitmaps and '+' actually means 'set bits'. if on-disk has
 231 *     bit set and PA claims same block, it's OK. IOW, one can set bit in
 232 *     on-disk bitmap if buddy has same bit set or/and PA covers corresponded
 233 *     block
 234 *
 235 * so, now we're building a concurrency table:
 236 *  - init buddy vs.
 237 *    - new PA
 238 *      blocks for PA are allocated in the buddy, buddy must be referenced
 239 *      until PA is linked to allocation group to avoid concurrent buddy init
 240 *    - use inode PA
 241 *      we need to make sure that either on-disk bitmap or PA has uptodate data
 242 *      given (3) we care that PA-=N operation doesn't interfere with init
 243 *    - discard inode PA
 244 *      the simplest way would be to have buddy initialized by the discard
 245 *    - use locality group PA
 246 *      again PA-=N must be serialized with init
 247 *    - discard locality group PA
 248 *      the simplest way would be to have buddy initialized by the discard
 249 *  - new PA vs.
 250 *    - use inode PA
 251 *      i_data_sem serializes them
 252 *    - discard inode PA
 253 *      discard process must wait until PA isn't used by another process
 254 *    - use locality group PA
 255 *      some mutex should serialize them
 256 *    - discard locality group PA
 257 *      discard process must wait until PA isn't used by another process
 258 *  - use inode PA
 259 *    - use inode PA
 260 *      i_data_sem or another mutex should serializes them
 261 *    - discard inode PA
 262 *      discard process must wait until PA isn't used by another process
 263 *    - use locality group PA
 264 *      nothing wrong here -- they're different PAs covering different blocks
 265 *    - discard locality group PA
 266 *      discard process must wait until PA isn't used by another process
 267 *
 268 * now we're ready to make few consequences:
 269 *  - PA is referenced and while it is no discard is possible
 270 *  - PA is referenced until block isn't marked in on-disk bitmap
 271 *  - PA changes only after on-disk bitmap
 272 *  - discard must not compete with init. either init is done before
 273 *    any discard or they're serialized somehow
 274 *  - buddy init as sum of on-disk bitmap and PAs is done atomically
 275 *
 276 * a special case when we've used PA to emptiness. no need to modify buddy
 277 * in this case, but we should care about concurrent init
 278 *
 279 */
 280
 281 /*
 282 * Logic in few words:
 283 *
 284 *  - allocation:
 285 *    load group
 286 *    find blocks
 287 *    mark bits in on-disk bitmap
 288 *    release group
 289 *
 290 *  - use preallocation:
 291 *    find proper PA (per-inode or group)
 292 *    load group
 293 *    mark bits in on-disk bitmap
 294 *    release group
 295 *    release PA
 296 *
 297 *  - free:
 298 *    load group
 299 *    mark bits in on-disk bitmap
 300 *    release group
 301 *
 302 *  - discard preallocations in group:
 303 *    mark PAs deleted
 304 *    move them onto local list
 305 *    load on-disk bitmap
 306 *    load group
 307 *    remove PA from object (inode or locality group)
 308 *    mark free blocks in-core
 309 *
 310 *  - discard inode's preallocations:
 311 */
 312
 313/*
 314 * Locking rules
 315 *
 316 * Locks:
 317 *  - bitlock on a group	(group)
 318 *  - object (inode/locality)	(object)
 319 *  - per-pa lock		(pa)
 320 *
 321 * Paths:
 322 *  - new pa
 323 *    object
 324 *    group
 325 *
 326 *  - find and use pa:
 327 *    pa
 328 *
 329 *  - release consumed pa:
 330 *    pa
 331 *    group
 332 *    object
 333 *
 334 *  - generate in-core bitmap:
 335 *    group
 336 *        pa
 337 *
 338 *  - discard all for given object (inode, locality group):
 339 *    object
 340 *        pa
 341 *    group
 342 *
 343 *  - discard all for given group:
 344 *    group
 345 *        pa
 346 *    group
 347 *        object
 348 *
 349 */
 350static struct kmem_cache *ext4_pspace_cachep;
 351static struct kmem_cache *ext4_ac_cachep;
 352static struct kmem_cache *ext4_free_data_cachep;
 353
 354/* We create slab caches for groupinfo data structures based on the
 355 * superblock block size.  There will be one per mounted filesystem for
 356 * each unique s_blocksize_bits */
 357#define NR_GRPINFO_CACHES 8
 358static struct kmem_cache *ext4_groupinfo_caches[NR_GRPINFO_CACHES];
 359
 360static const char *ext4_groupinfo_slab_names[NR_GRPINFO_CACHES] = {
 361	"ext4_groupinfo_1k", "ext4_groupinfo_2k", "ext4_groupinfo_4k",
 362	"ext4_groupinfo_8k", "ext4_groupinfo_16k", "ext4_groupinfo_32k",
 363	"ext4_groupinfo_64k", "ext4_groupinfo_128k"
 364};
 365
 366static void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
 367					ext4_group_t group);
 368static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
 369						ext4_group_t group);
 370static void ext4_free_data_callback(struct super_block *sb,
 371				struct ext4_journal_cb_entry *jce, int rc);
 372
 373static inline void *mb_correct_addr_and_bit(int *bit, void *addr)
 374{
 375#if BITS_PER_LONG == 64
 376	*bit += ((unsigned long) addr & 7UL) << 3;
 377	addr = (void *) ((unsigned long) addr & ~7UL);
 378#elif BITS_PER_LONG == 32
 379	*bit += ((unsigned long) addr & 3UL) << 3;
 380	addr = (void *) ((unsigned long) addr & ~3UL);
 381#else
 382#error "how many bits you are?!"
 383#endif
 384	return addr;
 385}
 386
 387static inline int mb_test_bit(int bit, void *addr)
 388{
 389	/*
 390	 * ext4_test_bit on architecture like powerpc
 391	 * needs unsigned long aligned address
 392	 */
 393	addr = mb_correct_addr_and_bit(&bit, addr);
 394	return ext4_test_bit(bit, addr);
 395}
 396
 397static inline void mb_set_bit(int bit, void *addr)
 398{
 399	addr = mb_correct_addr_and_bit(&bit, addr);
 400	ext4_set_bit(bit, addr);
 401}
 402
 403static inline void mb_clear_bit(int bit, void *addr)
 404{
 405	addr = mb_correct_addr_and_bit(&bit, addr);
 406	ext4_clear_bit(bit, addr);
 407}
 408
 409static inline int mb_test_and_clear_bit(int bit, void *addr)
 410{
 411	addr = mb_correct_addr_and_bit(&bit, addr);
 412	return ext4_test_and_clear_bit(bit, addr);
 413}
 414
 415static inline int mb_find_next_zero_bit(void *addr, int max, int start)
 416{
 417	int fix = 0, ret, tmpmax;
 418	addr = mb_correct_addr_and_bit(&fix, addr);
 419	tmpmax = max + fix;
 420	start += fix;
 421
 422	ret = ext4_find_next_zero_bit(addr, tmpmax, start) - fix;
 423	if (ret > max)
 424		return max;
 425	return ret;
 426}
 427
 428static inline int mb_find_next_bit(void *addr, int max, int start)
 429{
 430	int fix = 0, ret, tmpmax;
 431	addr = mb_correct_addr_and_bit(&fix, addr);
 432	tmpmax = max + fix;
 433	start += fix;
 434
 435	ret = ext4_find_next_bit(addr, tmpmax, start) - fix;
 436	if (ret > max)
 437		return max;
 438	return ret;
 439}
 440
 441static void *mb_find_buddy(struct ext4_buddy *e4b, int order, int *max)
 442{
 443	char *bb;
 444
 445	BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
 446	BUG_ON(max == NULL);
 447
 448	if (order > e4b->bd_blkbits + 1) {
 449		*max = 0;
 450		return NULL;
 451	}
 452
 453	/* at order 0 we see each particular block */
 454	if (order == 0) {
 455		*max = 1 << (e4b->bd_blkbits + 3);
 456		return e4b->bd_bitmap;
 457	}
 458
 459	bb = e4b->bd_buddy + EXT4_SB(e4b->bd_sb)->s_mb_offsets[order];
 460	*max = EXT4_SB(e4b->bd_sb)->s_mb_maxs[order];
 461
 462	return bb;
 463}
 464
 465#ifdef DOUBLE_CHECK
 466static void mb_free_blocks_double(struct inode *inode, struct ext4_buddy *e4b,
 467			   int first, int count)
 468{
 469	int i;
 470	struct super_block *sb = e4b->bd_sb;
 471
 472	if (unlikely(e4b->bd_info->bb_bitmap == NULL))
 473		return;
 474	assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
 475	for (i = 0; i < count; i++) {
 476		if (!mb_test_bit(first + i, e4b->bd_info->bb_bitmap)) {
 477			ext4_fsblk_t blocknr;
 478
 479			blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
 480			blocknr += EXT4_C2B(EXT4_SB(sb), first + i);
 481			ext4_grp_locked_error(sb, e4b->bd_group,
 482					      inode ? inode->i_ino : 0,
 483					      blocknr,
 484					      "freeing block already freed "
 485					      "(bit %u)",
 486					      first + i);
 487		}
 488		mb_clear_bit(first + i, e4b->bd_info->bb_bitmap);
 489	}
 490}
 491
 492static void mb_mark_used_double(struct ext4_buddy *e4b, int first, int count)
 493{
 494	int i;
 495
 496	if (unlikely(e4b->bd_info->bb_bitmap == NULL))
 497		return;
 498	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
 499	for (i = 0; i < count; i++) {
 500		BUG_ON(mb_test_bit(first + i, e4b->bd_info->bb_bitmap));
 501		mb_set_bit(first + i, e4b->bd_info->bb_bitmap);
 502	}
 503}
 504
 505static void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
 506{
 507	if (memcmp(e4b->bd_info->bb_bitmap, bitmap, e4b->bd_sb->s_blocksize)) {
 508		unsigned char *b1, *b2;
 509		int i;
 510		b1 = (unsigned char *) e4b->bd_info->bb_bitmap;
 511		b2 = (unsigned char *) bitmap;
 512		for (i = 0; i < e4b->bd_sb->s_blocksize; i++) {
 513			if (b1[i] != b2[i]) {
 514				ext4_msg(e4b->bd_sb, KERN_ERR,
 515					 "corruption in group %u "
 516					 "at byte %u(%u): %x in copy != %x "
 517					 "on disk/prealloc",
 518					 e4b->bd_group, i, i * 8, b1[i], b2[i]);
 519				BUG();
 520			}
 521		}
 522	}
 523}
 524
 525#else
 526static inline void mb_free_blocks_double(struct inode *inode,
 527				struct ext4_buddy *e4b, int first, int count)
 528{
 529	return;
 530}
 531static inline void mb_mark_used_double(struct ext4_buddy *e4b,
 532						int first, int count)
 533{
 534	return;
 535}
 536static inline void mb_cmp_bitmaps(struct ext4_buddy *e4b, void *bitmap)
 537{
 538	return;
 539}
 540#endif
 541
 542#ifdef AGGRESSIVE_CHECK
 543
 544#define MB_CHECK_ASSERT(assert)						\
 545do {									\
 546	if (!(assert)) {						\
 547		printk(KERN_EMERG					\
 548			"Assertion failure in %s() at %s:%d: \"%s\"\n",	\
 549			function, file, line, # assert);		\
 550		BUG();							\
 551	}								\
 552} while (0)
 553
 554static int __mb_check_buddy(struct ext4_buddy *e4b, char *file,
 555				const char *function, int line)
 556{
 557	struct super_block *sb = e4b->bd_sb;
 558	int order = e4b->bd_blkbits + 1;
 559	int max;
 560	int max2;
 561	int i;
 562	int j;
 563	int k;
 564	int count;
 565	struct ext4_group_info *grp;
 566	int fragments = 0;
 567	int fstart;
 568	struct list_head *cur;
 569	void *buddy;
 570	void *buddy2;
 571
 572	{
 573		static int mb_check_counter;
 574		if (mb_check_counter++ % 100 != 0)
 575			return 0;
 576	}
 577
 578	while (order > 1) {
 579		buddy = mb_find_buddy(e4b, order, &max);
 580		MB_CHECK_ASSERT(buddy);
 581		buddy2 = mb_find_buddy(e4b, order - 1, &max2);
 582		MB_CHECK_ASSERT(buddy2);
 583		MB_CHECK_ASSERT(buddy != buddy2);
 584		MB_CHECK_ASSERT(max * 2 == max2);
 585
 586		count = 0;
 587		for (i = 0; i < max; i++) {
 588
 589			if (mb_test_bit(i, buddy)) {
 590				/* only single bit in buddy2 may be 1 */
 591				if (!mb_test_bit(i << 1, buddy2)) {
 592					MB_CHECK_ASSERT(
 593						mb_test_bit((i<<1)+1, buddy2));
 594				} else if (!mb_test_bit((i << 1) + 1, buddy2)) {
 595					MB_CHECK_ASSERT(
 596						mb_test_bit(i << 1, buddy2));
 597				}
 598				continue;
 599			}
 600
 601			/* both bits in buddy2 must be 1 */
 602			MB_CHECK_ASSERT(mb_test_bit(i << 1, buddy2));
 603			MB_CHECK_ASSERT(mb_test_bit((i << 1) + 1, buddy2));
 604
 605			for (j = 0; j < (1 << order); j++) {
 606				k = (i * (1 << order)) + j;
 607				MB_CHECK_ASSERT(
 608					!mb_test_bit(k, e4b->bd_bitmap));
 609			}
 610			count++;
 611		}
 612		MB_CHECK_ASSERT(e4b->bd_info->bb_counters[order] == count);
 613		order--;
 614	}
 615
 616	fstart = -1;
 617	buddy = mb_find_buddy(e4b, 0, &max);
 618	for (i = 0; i < max; i++) {
 619		if (!mb_test_bit(i, buddy)) {
 620			MB_CHECK_ASSERT(i >= e4b->bd_info->bb_first_free);
 621			if (fstart == -1) {
 622				fragments++;
 623				fstart = i;
 624			}
 625			continue;
 626		}
 627		fstart = -1;
 628		/* check used bits only */
 629		for (j = 0; j < e4b->bd_blkbits + 1; j++) {
 630			buddy2 = mb_find_buddy(e4b, j, &max2);
 631			k = i >> j;
 632			MB_CHECK_ASSERT(k < max2);
 633			MB_CHECK_ASSERT(mb_test_bit(k, buddy2));
 634		}
 635	}
 636	MB_CHECK_ASSERT(!EXT4_MB_GRP_NEED_INIT(e4b->bd_info));
 637	MB_CHECK_ASSERT(e4b->bd_info->bb_fragments == fragments);
 638
 639	grp = ext4_get_group_info(sb, e4b->bd_group);
 640	list_for_each(cur, &grp->bb_prealloc_list) {
 641		ext4_group_t groupnr;
 642		struct ext4_prealloc_space *pa;
 643		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
 644		ext4_get_group_no_and_offset(sb, pa->pa_pstart, &groupnr, &k);
 645		MB_CHECK_ASSERT(groupnr == e4b->bd_group);
 646		for (i = 0; i < pa->pa_len; i++)
 647			MB_CHECK_ASSERT(mb_test_bit(k + i, buddy));
 648	}
 649	return 0;
 650}
 651#undef MB_CHECK_ASSERT
 652#define mb_check_buddy(e4b) __mb_check_buddy(e4b,	\
 653					__FILE__, __func__, __LINE__)
 654#else
 655#define mb_check_buddy(e4b)
 656#endif
 657
 658/*
 659 * Divide blocks started from @first with length @len into
 660 * smaller chunks with power of 2 blocks.
 661 * Clear the bits in bitmap which the blocks of the chunk(s) covered,
 662 * then increase bb_counters[] for corresponded chunk size.
 663 */
 664static void ext4_mb_mark_free_simple(struct super_block *sb,
 665				void *buddy, ext4_grpblk_t first, ext4_grpblk_t len,
 666					struct ext4_group_info *grp)
 667{
 668	struct ext4_sb_info *sbi = EXT4_SB(sb);
 669	ext4_grpblk_t min;
 670	ext4_grpblk_t max;
 671	ext4_grpblk_t chunk;
 672	unsigned int border;
 673
 674	BUG_ON(len > EXT4_CLUSTERS_PER_GROUP(sb));
 675
 676	border = 2 << sb->s_blocksize_bits;
 677
 678	while (len > 0) {
 679		/* find how many blocks can be covered since this position */
 680		max = ffs(first | border) - 1;
 681
 682		/* find how many blocks of power 2 we need to mark */
 683		min = fls(len) - 1;
 684
 685		if (max < min)
 686			min = max;
 687		chunk = 1 << min;
 688
 689		/* mark multiblock chunks only */
 690		grp->bb_counters[min]++;
 691		if (min > 0)
 692			mb_clear_bit(first >> min,
 693				     buddy + sbi->s_mb_offsets[min]);
 694
 695		len -= chunk;
 696		first += chunk;
 697	}
 698}
 699
 700/*
 701 * Cache the order of the largest free extent we have available in this block
 702 * group.
 703 */
 704static void
 705mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp)
 706{
 707	int i;
 708	int bits;
 709
 710	grp->bb_largest_free_order = -1; /* uninit */
 711
 712	bits = sb->s_blocksize_bits + 1;
 713	for (i = bits; i >= 0; i--) {
 714		if (grp->bb_counters[i] > 0) {
 715			grp->bb_largest_free_order = i;
 716			break;
 717		}
 718	}
 719}
 720
 721static noinline_for_stack
 722void ext4_mb_generate_buddy(struct super_block *sb,
 723				void *buddy, void *bitmap, ext4_group_t group)
 724{
 725	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
 726	struct ext4_sb_info *sbi = EXT4_SB(sb);
 727	ext4_grpblk_t max = EXT4_CLUSTERS_PER_GROUP(sb);
 728	ext4_grpblk_t i = 0;
 729	ext4_grpblk_t first;
 730	ext4_grpblk_t len;
 731	unsigned free = 0;
 732	unsigned fragments = 0;
 733	unsigned long long period = get_cycles();
 734
 735	/* initialize buddy from bitmap which is aggregation
 736	 * of on-disk bitmap and preallocations */
 737	i = mb_find_next_zero_bit(bitmap, max, 0);
 738	grp->bb_first_free = i;
 739	while (i < max) {
 740		fragments++;
 741		first = i;
 742		i = mb_find_next_bit(bitmap, max, i);
 743		len = i - first;
 744		free += len;
 745		if (len > 1)
 746			ext4_mb_mark_free_simple(sb, buddy, first, len, grp);
 747		else
 748			grp->bb_counters[0]++;
 749		if (i < max)
 750			i = mb_find_next_zero_bit(bitmap, max, i);
 751	}
 752	grp->bb_fragments = fragments;
 753
 754	if (free != grp->bb_free) {
 755		ext4_grp_locked_error(sb, group, 0, 0,
 756				      "block bitmap and bg descriptor "
 757				      "inconsistent: %u vs %u free clusters",
 758				      free, grp->bb_free);
 759		/*
 760		 * If we intend to continue, we consider group descriptor
 761		 * corrupt and update bb_free using bitmap value
 762		 */
 763		grp->bb_free = free;
 764		if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
 765			percpu_counter_sub(&sbi->s_freeclusters_counter,
 766					   grp->bb_free);
 767		set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
 768	}
 769	mb_set_largest_free_order(sb, grp);
 770
 771	clear_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &(grp->bb_state));
 772
 773	period = get_cycles() - period;
 774	spin_lock(&EXT4_SB(sb)->s_bal_lock);
 775	EXT4_SB(sb)->s_mb_buddies_generated++;
 776	EXT4_SB(sb)->s_mb_generation_time += period;
 777	spin_unlock(&EXT4_SB(sb)->s_bal_lock);
 778}
 779
 780static void mb_regenerate_buddy(struct ext4_buddy *e4b)
 781{
 782	int count;
 783	int order = 1;
 784	void *buddy;
 785
 786	while ((buddy = mb_find_buddy(e4b, order++, &count))) {
 787		ext4_set_bits(buddy, 0, count);
 788	}
 789	e4b->bd_info->bb_fragments = 0;
 790	memset(e4b->bd_info->bb_counters, 0,
 791		sizeof(*e4b->bd_info->bb_counters) *
 792		(e4b->bd_sb->s_blocksize_bits + 2));
 793
 794	ext4_mb_generate_buddy(e4b->bd_sb, e4b->bd_buddy,
 795		e4b->bd_bitmap, e4b->bd_group);
 796}
 797
 798/* The buddy information is attached the buddy cache inode
 799 * for convenience. The information regarding each group
 800 * is loaded via ext4_mb_load_buddy. The information involve
 801 * block bitmap and buddy information. The information are
 802 * stored in the inode as
 803 *
 804 * {                        page                        }
 805 * [ group 0 bitmap][ group 0 buddy] [group 1][ group 1]...
 806 *
 807 *
 808 * one block each for bitmap and buddy information.
 809 * So for each group we take up 2 blocks. A page can
 810 * contain blocks_per_page (PAGE_SIZE / blocksize)  blocks.
 811 * So it can have information regarding groups_per_page which
 812 * is blocks_per_page/2
 813 *
 814 * Locking note:  This routine takes the block group lock of all groups
 815 * for this page; do not hold this lock when calling this routine!
 816 */
 817
 818static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
 819{
 820	ext4_group_t ngroups;
 821	int blocksize;
 822	int blocks_per_page;
 823	int groups_per_page;
 824	int err = 0;
 825	int i;
 826	ext4_group_t first_group, group;
 827	int first_block;
 828	struct super_block *sb;
 829	struct buffer_head *bhs;
 830	struct buffer_head **bh = NULL;
 831	struct inode *inode;
 832	char *data;
 833	char *bitmap;
 834	struct ext4_group_info *grinfo;
 835
 836	mb_debug(1, "init page %lu\n", page->index);
 837
 838	inode = page->mapping->host;
 839	sb = inode->i_sb;
 840	ngroups = ext4_get_groups_count(sb);
 841	blocksize = 1 << inode->i_blkbits;
 842	blocks_per_page = PAGE_SIZE / blocksize;
 843
 844	groups_per_page = blocks_per_page >> 1;
 845	if (groups_per_page == 0)
 846		groups_per_page = 1;
 847
 848	/* allocate buffer_heads to read bitmaps */
 849	if (groups_per_page > 1) {
 850		i = sizeof(struct buffer_head *) * groups_per_page;
 851		bh = kzalloc(i, gfp);
 852		if (bh == NULL) {
 853			err = -ENOMEM;
 854			goto out;
 855		}
 856	} else
 857		bh = &bhs;
 858
 859	first_group = page->index * blocks_per_page / 2;
 860
 861	/* read all groups the page covers into the cache */
 862	for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
 863		if (group >= ngroups)
 864			break;
 865
 866		grinfo = ext4_get_group_info(sb, group);
 867		/*
 868		 * If page is uptodate then we came here after online resize
 869		 * which added some new uninitialized group info structs, so
 870		 * we must skip all initialized uptodate buddies on the page,
 871		 * which may be currently in use by an allocating task.
 872		 */
 873		if (PageUptodate(page) && !EXT4_MB_GRP_NEED_INIT(grinfo)) {
 874			bh[i] = NULL;
 875			continue;
 876		}
 877		bh[i] = ext4_read_block_bitmap_nowait(sb, group);
 878		if (IS_ERR(bh[i])) {
 879			err = PTR_ERR(bh[i]);
 880			bh[i] = NULL;
 881			goto out;
 882		}
 883		mb_debug(1, "read bitmap for group %u\n", group);
 884	}
 885
 886	/* wait for I/O completion */
 887	for (i = 0, group = first_group; i < groups_per_page; i++, group++) {
 888		int err2;
 889
 890		if (!bh[i])
 891			continue;
 892		err2 = ext4_wait_block_bitmap(sb, group, bh[i]);
 893		if (!err)
 894			err = err2;
 895	}
 896
 897	first_block = page->index * blocks_per_page;
 898	for (i = 0; i < blocks_per_page; i++) {
 899		group = (first_block + i) >> 1;
 900		if (group >= ngroups)
 901			break;
 902
 903		if (!bh[group - first_group])
 904			/* skip initialized uptodate buddy */
 905			continue;
 906
 907		if (!buffer_verified(bh[group - first_group]))
 908			/* Skip faulty bitmaps */
 909			continue;
 910		err = 0;
 911
 912		/*
 913		 * data carry information regarding this
 914		 * particular group in the format specified
 915		 * above
 916		 *
 917		 */
 918		data = page_address(page) + (i * blocksize);
 919		bitmap = bh[group - first_group]->b_data;
 920
 921		/*
 922		 * We place the buddy block and bitmap block
 923		 * close together
 924		 */
 925		if ((first_block + i) & 1) {
 926			/* this is block of buddy */
 927			BUG_ON(incore == NULL);
 928			mb_debug(1, "put buddy for group %u in page %lu/%x\n",
 929				group, page->index, i * blocksize);
 930			trace_ext4_mb_buddy_bitmap_load(sb, group);
 931			grinfo = ext4_get_group_info(sb, group);
 932			grinfo->bb_fragments = 0;
 933			memset(grinfo->bb_counters, 0,
 934			       sizeof(*grinfo->bb_counters) *
 935				(sb->s_blocksize_bits+2));
 936			/*
 937			 * incore got set to the group block bitmap below
 938			 */
 939			ext4_lock_group(sb, group);
 940			/* init the buddy */
 941			memset(data, 0xff, blocksize);
 942			ext4_mb_generate_buddy(sb, data, incore, group);
 943			ext4_unlock_group(sb, group);
 944			incore = NULL;
 945		} else {
 946			/* this is block of bitmap */
 947			BUG_ON(incore != NULL);
 948			mb_debug(1, "put bitmap for group %u in page %lu/%x\n",
 949				group, page->index, i * blocksize);
 950			trace_ext4_mb_bitmap_load(sb, group);
 951
 952			/* see comments in ext4_mb_put_pa() */
 953			ext4_lock_group(sb, group);
 954			memcpy(data, bitmap, blocksize);
 955
 956			/* mark all preallocated blks used in in-core bitmap */
 957			ext4_mb_generate_from_pa(sb, data, group);
 958			ext4_mb_generate_from_freelist(sb, data, group);
 959			ext4_unlock_group(sb, group);
 960
 961			/* set incore so that the buddy information can be
 962			 * generated using this
 963			 */
 964			incore = data;
 965		}
 966	}
 967	SetPageUptodate(page);
 968
 969out:
 970	if (bh) {
 971		for (i = 0; i < groups_per_page; i++)
 972			brelse(bh[i]);
 973		if (bh != &bhs)
 974			kfree(bh);
 975	}
 976	return err;
 977}
 978
 979/*
 980 * Lock the buddy and bitmap pages. This make sure other parallel init_group
 981 * on the same buddy page doesn't happen whild holding the buddy page lock.
 982 * Return locked buddy and bitmap pages on e4b struct. If buddy and bitmap
 983 * are on the same page e4b->bd_buddy_page is NULL and return value is 0.
 984 */
 985static int ext4_mb_get_buddy_page_lock(struct super_block *sb,
 986		ext4_group_t group, struct ext4_buddy *e4b, gfp_t gfp)
 987{
 988	struct inode *inode = EXT4_SB(sb)->s_buddy_cache;
 989	int block, pnum, poff;
 990	int blocks_per_page;
 991	struct page *page;
 992
 993	e4b->bd_buddy_page = NULL;
 994	e4b->bd_bitmap_page = NULL;
 995
 996	blocks_per_page = PAGE_SIZE / sb->s_blocksize;
 997	/*
 998	 * the buddy cache inode stores the block bitmap
 999	 * and buddy information in consecutive blocks.
1000	 * So for each group we need two blocks.
1001	 */
1002	block = group * 2;
1003	pnum = block / blocks_per_page;
1004	poff = block % blocks_per_page;
1005	page = find_or_create_page(inode->i_mapping, pnum, gfp);
1006	if (!page)
1007		return -ENOMEM;
1008	BUG_ON(page->mapping != inode->i_mapping);
1009	e4b->bd_bitmap_page = page;
1010	e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1011
1012	if (blocks_per_page >= 2) {
1013		/* buddy and bitmap are on the same page */
1014		return 0;
1015	}
1016
1017	block++;
1018	pnum = block / blocks_per_page;
1019	page = find_or_create_page(inode->i_mapping, pnum, gfp);
1020	if (!page)
1021		return -ENOMEM;
1022	BUG_ON(page->mapping != inode->i_mapping);
1023	e4b->bd_buddy_page = page;
1024	return 0;
1025}
1026
1027static void ext4_mb_put_buddy_page_lock(struct ext4_buddy *e4b)
1028{
1029	if (e4b->bd_bitmap_page) {
1030		unlock_page(e4b->bd_bitmap_page);
1031		put_page(e4b->bd_bitmap_page);
1032	}
1033	if (e4b->bd_buddy_page) {
1034		unlock_page(e4b->bd_buddy_page);
1035		put_page(e4b->bd_buddy_page);
1036	}
1037}
1038
1039/*
1040 * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1041 * block group lock of all groups for this page; do not hold the BG lock when
1042 * calling this routine!
1043 */
1044static noinline_for_stack
1045int ext4_mb_init_group(struct super_block *sb, ext4_group_t group, gfp_t gfp)
1046{
1047
1048	struct ext4_group_info *this_grp;
1049	struct ext4_buddy e4b;
1050	struct page *page;
1051	int ret = 0;
1052
1053	might_sleep();
1054	mb_debug(1, "init group %u\n", group);
1055	this_grp = ext4_get_group_info(sb, group);
1056	/*
1057	 * This ensures that we don't reinit the buddy cache
1058	 * page which map to the group from which we are already
1059	 * allocating. If we are looking at the buddy cache we would
1060	 * have taken a reference using ext4_mb_load_buddy and that
1061	 * would have pinned buddy page to page cache.
1062	 * The call to ext4_mb_get_buddy_page_lock will mark the
1063	 * page accessed.
1064	 */
1065	ret = ext4_mb_get_buddy_page_lock(sb, group, &e4b, gfp);
1066	if (ret || !EXT4_MB_GRP_NEED_INIT(this_grp)) {
1067		/*
1068		 * somebody initialized the group
1069		 * return without doing anything
1070		 */
1071		goto err;
1072	}
1073
1074	page = e4b.bd_bitmap_page;
1075	ret = ext4_mb_init_cache(page, NULL, gfp);
1076	if (ret)
1077		goto err;
1078	if (!PageUptodate(page)) {
1079		ret = -EIO;
1080		goto err;
1081	}
1082
1083	if (e4b.bd_buddy_page == NULL) {
1084		/*
1085		 * If both the bitmap and buddy are in
1086		 * the same page we don't need to force
1087		 * init the buddy
1088		 */
1089		ret = 0;
1090		goto err;
1091	}
1092	/* init buddy cache */
1093	page = e4b.bd_buddy_page;
1094	ret = ext4_mb_init_cache(page, e4b.bd_bitmap, gfp);
1095	if (ret)
1096		goto err;
1097	if (!PageUptodate(page)) {
1098		ret = -EIO;
1099		goto err;
1100	}
1101err:
1102	ext4_mb_put_buddy_page_lock(&e4b);
1103	return ret;
1104}
1105
1106/*
1107 * Locking note:  This routine calls ext4_mb_init_cache(), which takes the
1108 * block group lock of all groups for this page; do not hold the BG lock when
1109 * calling this routine!
1110 */
1111static noinline_for_stack int
1112ext4_mb_load_buddy_gfp(struct super_block *sb, ext4_group_t group,
1113		       struct ext4_buddy *e4b, gfp_t gfp)
1114{
1115	int blocks_per_page;
1116	int block;
1117	int pnum;
1118	int poff;
1119	struct page *page;
1120	int ret;
1121	struct ext4_group_info *grp;
1122	struct ext4_sb_info *sbi = EXT4_SB(sb);
1123	struct inode *inode = sbi->s_buddy_cache;
1124
1125	might_sleep();
1126	mb_debug(1, "load group %u\n", group);
1127
1128	blocks_per_page = PAGE_SIZE / sb->s_blocksize;
1129	grp = ext4_get_group_info(sb, group);
1130
1131	e4b->bd_blkbits = sb->s_blocksize_bits;
1132	e4b->bd_info = grp;
1133	e4b->bd_sb = sb;
1134	e4b->bd_group = group;
1135	e4b->bd_buddy_page = NULL;
1136	e4b->bd_bitmap_page = NULL;
1137
1138	if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
1139		/*
1140		 * we need full data about the group
1141		 * to make a good selection
1142		 */
1143		ret = ext4_mb_init_group(sb, group, gfp);
1144		if (ret)
1145			return ret;
1146	}
1147
1148	/*
1149	 * the buddy cache inode stores the block bitmap
1150	 * and buddy information in consecutive blocks.
1151	 * So for each group we need two blocks.
1152	 */
1153	block = group * 2;
1154	pnum = block / blocks_per_page;
1155	poff = block % blocks_per_page;
1156
1157	/* we could use find_or_create_page(), but it locks page
1158	 * what we'd like to avoid in fast path ... */
1159	page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1160	if (page == NULL || !PageUptodate(page)) {
1161		if (page)
1162			/*
1163			 * drop the page reference and try
1164			 * to get the page with lock. If we
1165			 * are not uptodate that implies
1166			 * somebody just created the page but
1167			 * is yet to initialize the same. So
1168			 * wait for it to initialize.
1169			 */
1170			put_page(page);
1171		page = find_or_create_page(inode->i_mapping, pnum, gfp);
1172		if (page) {
1173			BUG_ON(page->mapping != inode->i_mapping);
1174			if (!PageUptodate(page)) {
1175				ret = ext4_mb_init_cache(page, NULL, gfp);
1176				if (ret) {
1177					unlock_page(page);
1178					goto err;
1179				}
1180				mb_cmp_bitmaps(e4b, page_address(page) +
1181					       (poff * sb->s_blocksize));
1182			}
1183			unlock_page(page);
1184		}
1185	}
1186	if (page == NULL) {
1187		ret = -ENOMEM;
1188		goto err;
1189	}
1190	if (!PageUptodate(page)) {
1191		ret = -EIO;
1192		goto err;
1193	}
1194
1195	/* Pages marked accessed already */
1196	e4b->bd_bitmap_page = page;
1197	e4b->bd_bitmap = page_address(page) + (poff * sb->s_blocksize);
1198
1199	block++;
1200	pnum = block / blocks_per_page;
1201	poff = block % blocks_per_page;
1202
1203	page = find_get_page_flags(inode->i_mapping, pnum, FGP_ACCESSED);
1204	if (page == NULL || !PageUptodate(page)) {
1205		if (page)
1206			put_page(page);
1207		page = find_or_create_page(inode->i_mapping, pnum, gfp);
1208		if (page) {
1209			BUG_ON(page->mapping != inode->i_mapping);
1210			if (!PageUptodate(page)) {
1211				ret = ext4_mb_init_cache(page, e4b->bd_bitmap,
1212							 gfp);
1213				if (ret) {
1214					unlock_page(page);
1215					goto err;
1216				}
1217			}
1218			unlock_page(page);
1219		}
1220	}
1221	if (page == NULL) {
1222		ret = -ENOMEM;
1223		goto err;
1224	}
1225	if (!PageUptodate(page)) {
1226		ret = -EIO;
1227		goto err;
1228	}
1229
1230	/* Pages marked accessed already */
1231	e4b->bd_buddy_page = page;
1232	e4b->bd_buddy = page_address(page) + (poff * sb->s_blocksize);
1233
1234	BUG_ON(e4b->bd_bitmap_page == NULL);
1235	BUG_ON(e4b->bd_buddy_page == NULL);
1236
1237	return 0;
1238
1239err:
1240	if (page)
1241		put_page(page);
1242	if (e4b->bd_bitmap_page)
1243		put_page(e4b->bd_bitmap_page);
1244	if (e4b->bd_buddy_page)
1245		put_page(e4b->bd_buddy_page);
1246	e4b->bd_buddy = NULL;
1247	e4b->bd_bitmap = NULL;
1248	return ret;
1249}
1250
1251static int ext4_mb_load_buddy(struct super_block *sb, ext4_group_t group,
1252			      struct ext4_buddy *e4b)
1253{
1254	return ext4_mb_load_buddy_gfp(sb, group, e4b, GFP_NOFS);
1255}
1256
1257static void ext4_mb_unload_buddy(struct ext4_buddy *e4b)
1258{
1259	if (e4b->bd_bitmap_page)
1260		put_page(e4b->bd_bitmap_page);
1261	if (e4b->bd_buddy_page)
1262		put_page(e4b->bd_buddy_page);
1263}
1264
1265
1266static int mb_find_order_for_block(struct ext4_buddy *e4b, int block)
1267{
1268	int order = 1;
1269	int bb_incr = 1 << (e4b->bd_blkbits - 1);
1270	void *bb;
1271
1272	BUG_ON(e4b->bd_bitmap == e4b->bd_buddy);
1273	BUG_ON(block >= (1 << (e4b->bd_blkbits + 3)));
1274
1275	bb = e4b->bd_buddy;
1276	while (order <= e4b->bd_blkbits + 1) {
1277		block = block >> 1;
1278		if (!mb_test_bit(block, bb)) {
1279			/* this block is part of buddy of order 'order' */
1280			return order;
1281		}
1282		bb += bb_incr;
1283		bb_incr >>= 1;
1284		order++;
1285	}
1286	return 0;
1287}
1288
1289static void mb_clear_bits(void *bm, int cur, int len)
1290{
1291	__u32 *addr;
1292
1293	len = cur + len;
1294	while (cur < len) {
1295		if ((cur & 31) == 0 && (len - cur) >= 32) {
1296			/* fast path: clear whole word at once */
1297			addr = bm + (cur >> 3);
1298			*addr = 0;
1299			cur += 32;
1300			continue;
1301		}
1302		mb_clear_bit(cur, bm);
1303		cur++;
1304	}
1305}
1306
1307/* clear bits in given range
1308 * will return first found zero bit if any, -1 otherwise
1309 */
1310static int mb_test_and_clear_bits(void *bm, int cur, int len)
1311{
1312	__u32 *addr;
1313	int zero_bit = -1;
1314
1315	len = cur + len;
1316	while (cur < len) {
1317		if ((cur & 31) == 0 && (len - cur) >= 32) {
1318			/* fast path: clear whole word at once */
1319			addr = bm + (cur >> 3);
1320			if (*addr != (__u32)(-1) && zero_bit == -1)
1321				zero_bit = cur + mb_find_next_zero_bit(addr, 32, 0);
1322			*addr = 0;
1323			cur += 32;
1324			continue;
1325		}
1326		if (!mb_test_and_clear_bit(cur, bm) && zero_bit == -1)
1327			zero_bit = cur;
1328		cur++;
1329	}
1330
1331	return zero_bit;
1332}
1333
1334void ext4_set_bits(void *bm, int cur, int len)
1335{
1336	__u32 *addr;
1337
1338	len = cur + len;
1339	while (cur < len) {
1340		if ((cur & 31) == 0 && (len - cur) >= 32) {
1341			/* fast path: set whole word at once */
1342			addr = bm + (cur >> 3);
1343			*addr = 0xffffffff;
1344			cur += 32;
1345			continue;
1346		}
1347		mb_set_bit(cur, bm);
1348		cur++;
1349	}
1350}
1351
1352/*
1353 * _________________________________________________________________ */
1354
1355static inline int mb_buddy_adjust_border(int* bit, void* bitmap, int side)
1356{
1357	if (mb_test_bit(*bit + side, bitmap)) {
1358		mb_clear_bit(*bit, bitmap);
1359		(*bit) -= side;
1360		return 1;
1361	}
1362	else {
1363		(*bit) += side;
1364		mb_set_bit(*bit, bitmap);
1365		return -1;
1366	}
1367}
1368
1369static void mb_buddy_mark_free(struct ext4_buddy *e4b, int first, int last)
1370{
1371	int max;
1372	int order = 1;
1373	void *buddy = mb_find_buddy(e4b, order, &max);
1374
1375	while (buddy) {
1376		void *buddy2;
1377
1378		/* Bits in range [first; last] are known to be set since
1379		 * corresponding blocks were allocated. Bits in range
1380		 * (first; last) will stay set because they form buddies on
1381		 * upper layer. We just deal with borders if they don't
1382		 * align with upper layer and then go up.
1383		 * Releasing entire group is all about clearing
1384		 * single bit of highest order buddy.
1385		 */
1386
1387		/* Example:
1388		 * ---------------------------------
1389		 * |   1   |   1   |   1   |   1   |
1390		 * ---------------------------------
1391		 * | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
1392		 * ---------------------------------
1393		 *   0   1   2   3   4   5   6   7
1394		 *      \_____________________/
1395		 *
1396		 * Neither [1] nor [6] is aligned to above layer.
1397		 * Left neighbour [0] is free, so mark it busy,
1398		 * decrease bb_counters and extend range to
1399		 * [0; 6]
1400		 * Right neighbour [7] is busy. It can't be coaleasced with [6], so
1401		 * mark [6] free, increase bb_counters and shrink range to
1402		 * [0; 5].
1403		 * Then shift range to [0; 2], go up and do the same.
1404		 */
1405
1406
1407		if (first & 1)
1408			e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&first, buddy, -1);
1409		if (!(last & 1))
1410			e4b->bd_info->bb_counters[order] += mb_buddy_adjust_border(&last, buddy, 1);
1411		if (first > last)
1412			break;
1413		order++;
1414
1415		if (first == last || !(buddy2 = mb_find_buddy(e4b, order, &max))) {
1416			mb_clear_bits(buddy, first, last - first + 1);
1417			e4b->bd_info->bb_counters[order - 1] += last - first + 1;
1418			break;
1419		}
1420		first >>= 1;
1421		last >>= 1;
1422		buddy = buddy2;
1423	}
1424}
1425
1426static void mb_free_blocks(struct inode *inode, struct ext4_buddy *e4b,
1427			   int first, int count)
1428{
1429	int left_is_free = 0;
1430	int right_is_free = 0;
1431	int block;
1432	int last = first + count - 1;
1433	struct super_block *sb = e4b->bd_sb;
1434
1435	if (WARN_ON(count == 0))
1436		return;
1437	BUG_ON(last >= (sb->s_blocksize << 3));
1438	assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
1439	/* Don't bother if the block group is corrupt. */
1440	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info)))
1441		return;
1442
1443	mb_check_buddy(e4b);
1444	mb_free_blocks_double(inode, e4b, first, count);
1445
1446	e4b->bd_info->bb_free += count;
1447	if (first < e4b->bd_info->bb_first_free)
1448		e4b->bd_info->bb_first_free = first;
1449
1450	/* access memory sequentially: check left neighbour,
1451	 * clear range and then check right neighbour
1452	 */
1453	if (first != 0)
1454		left_is_free = !mb_test_bit(first - 1, e4b->bd_bitmap);
1455	block = mb_test_and_clear_bits(e4b->bd_bitmap, first, count);
1456	if (last + 1 < EXT4_SB(sb)->s_mb_maxs[0])
1457		right_is_free = !mb_test_bit(last + 1, e4b->bd_bitmap);
1458
1459	if (unlikely(block != -1)) {
1460		struct ext4_sb_info *sbi = EXT4_SB(sb);
1461		ext4_fsblk_t blocknr;
1462
1463		blocknr = ext4_group_first_block_no(sb, e4b->bd_group);
1464		blocknr += EXT4_C2B(EXT4_SB(sb), block);
1465		ext4_grp_locked_error(sb, e4b->bd_group,
1466				      inode ? inode->i_ino : 0,
1467				      blocknr,
1468				      "freeing already freed block "
1469				      "(bit %u); block bitmap corrupt.",
1470				      block);
1471		if (!EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))
1472			percpu_counter_sub(&sbi->s_freeclusters_counter,
1473					   e4b->bd_info->bb_free);
1474		/* Mark the block group as corrupt. */
1475		set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT,
1476			&e4b->bd_info->bb_state);
1477		mb_regenerate_buddy(e4b);
1478		goto done;
1479	}
1480
1481	/* let's maintain fragments counter */
1482	if (left_is_free && right_is_free)
1483		e4b->bd_info->bb_fragments--;
1484	else if (!left_is_free && !right_is_free)
1485		e4b->bd_info->bb_fragments++;
1486
1487	/* buddy[0] == bd_bitmap is a special case, so handle
1488	 * it right away and let mb_buddy_mark_free stay free of
1489	 * zero order checks.
1490	 * Check if neighbours are to be coaleasced,
1491	 * adjust bitmap bb_counters and borders appropriately.
1492	 */
1493	if (first & 1) {
1494		first += !left_is_free;
1495		e4b->bd_info->bb_counters[0] += left_is_free ? -1 : 1;
1496	}
1497	if (!(last & 1)) {
1498		last -= !right_is_free;
1499		e4b->bd_info->bb_counters[0] += right_is_free ? -1 : 1;
1500	}
1501
1502	if (first <= last)
1503		mb_buddy_mark_free(e4b, first >> 1, last >> 1);
1504
1505done:
1506	mb_set_largest_free_order(sb, e4b->bd_info);
1507	mb_check_buddy(e4b);
1508}
1509
1510static int mb_find_extent(struct ext4_buddy *e4b, int block,
1511				int needed, struct ext4_free_extent *ex)
1512{
1513	int next = block;
1514	int max, order;
1515	void *buddy;
1516
1517	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1518	BUG_ON(ex == NULL);
1519
1520	buddy = mb_find_buddy(e4b, 0, &max);
1521	BUG_ON(buddy == NULL);
1522	BUG_ON(block >= max);
1523	if (mb_test_bit(block, buddy)) {
1524		ex->fe_len = 0;
1525		ex->fe_start = 0;
1526		ex->fe_group = 0;
1527		return 0;
1528	}
1529
1530	/* find actual order */
1531	order = mb_find_order_for_block(e4b, block);
1532	block = block >> order;
1533
1534	ex->fe_len = 1 << order;
1535	ex->fe_start = block << order;
1536	ex->fe_group = e4b->bd_group;
1537
1538	/* calc difference from given start */
1539	next = next - ex->fe_start;
1540	ex->fe_len -= next;
1541	ex->fe_start += next;
1542
1543	while (needed > ex->fe_len &&
1544	       mb_find_buddy(e4b, order, &max)) {
1545
1546		if (block + 1 >= max)
1547			break;
1548
1549		next = (block + 1) * (1 << order);
1550		if (mb_test_bit(next, e4b->bd_bitmap))
1551			break;
1552
1553		order = mb_find_order_for_block(e4b, next);
1554
1555		block = next >> order;
1556		ex->fe_len += 1 << order;
1557	}
1558
1559	BUG_ON(ex->fe_start + ex->fe_len > (1 << (e4b->bd_blkbits + 3)));
1560	return ex->fe_len;
1561}
1562
1563static int mb_mark_used(struct ext4_buddy *e4b, struct ext4_free_extent *ex)
1564{
1565	int ord;
1566	int mlen = 0;
1567	int max = 0;
1568	int cur;
1569	int start = ex->fe_start;
1570	int len = ex->fe_len;
1571	unsigned ret = 0;
1572	int len0 = len;
1573	void *buddy;
1574
1575	BUG_ON(start + len > (e4b->bd_sb->s_blocksize << 3));
1576	BUG_ON(e4b->bd_group != ex->fe_group);
1577	assert_spin_locked(ext4_group_lock_ptr(e4b->bd_sb, e4b->bd_group));
1578	mb_check_buddy(e4b);
1579	mb_mark_used_double(e4b, start, len);
1580
1581	e4b->bd_info->bb_free -= len;
1582	if (e4b->bd_info->bb_first_free == start)
1583		e4b->bd_info->bb_first_free += len;
1584
1585	/* let's maintain fragments counter */
1586	if (start != 0)
1587		mlen = !mb_test_bit(start - 1, e4b->bd_bitmap);
1588	if (start + len < EXT4_SB(e4b->bd_sb)->s_mb_maxs[0])
1589		max = !mb_test_bit(start + len, e4b->bd_bitmap);
1590	if (mlen && max)
1591		e4b->bd_info->bb_fragments++;
1592	else if (!mlen && !max)
1593		e4b->bd_info->bb_fragments--;
1594
1595	/* let's maintain buddy itself */
1596	while (len) {
1597		ord = mb_find_order_for_block(e4b, start);
1598
1599		if (((start >> ord) << ord) == start && len >= (1 << ord)) {
1600			/* the whole chunk may be allocated at once! */
1601			mlen = 1 << ord;
1602			buddy = mb_find_buddy(e4b, ord, &max);
1603			BUG_ON((start >> ord) >= max);
1604			mb_set_bit(start >> ord, buddy);
1605			e4b->bd_info->bb_counters[ord]--;
1606			start += mlen;
1607			len -= mlen;
1608			BUG_ON(len < 0);
1609			continue;
1610		}
1611
1612		/* store for history */
1613		if (ret == 0)
1614			ret = len | (ord << 16);
1615
1616		/* we have to split large buddy */
1617		BUG_ON(ord <= 0);
1618		buddy = mb_find_buddy(e4b, ord, &max);
1619		mb_set_bit(start >> ord, buddy);
1620		e4b->bd_info->bb_counters[ord]--;
1621
1622		ord--;
1623		cur = (start >> ord) & ~1U;
1624		buddy = mb_find_buddy(e4b, ord, &max);
1625		mb_clear_bit(cur, buddy);
1626		mb_clear_bit(cur + 1, buddy);
1627		e4b->bd_info->bb_counters[ord]++;
1628		e4b->bd_info->bb_counters[ord]++;
1629	}
1630	mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
1631
1632	ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
1633	mb_check_buddy(e4b);
1634
1635	return ret;
1636}
1637
1638/*
1639 * Must be called under group lock!
1640 */
1641static void ext4_mb_use_best_found(struct ext4_allocation_context *ac,
1642					struct ext4_buddy *e4b)
1643{
1644	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1645	int ret;
1646
1647	BUG_ON(ac->ac_b_ex.fe_group != e4b->bd_group);
1648	BUG_ON(ac->ac_status == AC_STATUS_FOUND);
1649
1650	ac->ac_b_ex.fe_len = min(ac->ac_b_ex.fe_len, ac->ac_g_ex.fe_len);
1651	ac->ac_b_ex.fe_logical = ac->ac_g_ex.fe_logical;
1652	ret = mb_mark_used(e4b, &ac->ac_b_ex);
1653
1654	/* preallocation can change ac_b_ex, thus we store actually
1655	 * allocated blocks for history */
1656	ac->ac_f_ex = ac->ac_b_ex;
1657
1658	ac->ac_status = AC_STATUS_FOUND;
1659	ac->ac_tail = ret & 0xffff;
1660	ac->ac_buddy = ret >> 16;
1661
1662	/*
1663	 * take the page reference. We want the page to be pinned
1664	 * so that we don't get a ext4_mb_init_cache_call for this
1665	 * group until we update the bitmap. That would mean we
1666	 * double allocate blocks. The reference is dropped
1667	 * in ext4_mb_release_context
1668	 */
1669	ac->ac_bitmap_page = e4b->bd_bitmap_page;
1670	get_page(ac->ac_bitmap_page);
1671	ac->ac_buddy_page = e4b->bd_buddy_page;
1672	get_page(ac->ac_buddy_page);
1673	/* store last allocated for subsequent stream allocation */
1674	if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
1675		spin_lock(&sbi->s_md_lock);
1676		sbi->s_mb_last_group = ac->ac_f_ex.fe_group;
1677		sbi->s_mb_last_start = ac->ac_f_ex.fe_start;
1678		spin_unlock(&sbi->s_md_lock);
1679	}
1680}
1681
1682/*
1683 * regular allocator, for general purposes allocation
1684 */
1685
1686static void ext4_mb_check_limits(struct ext4_allocation_context *ac,
1687					struct ext4_buddy *e4b,
1688					int finish_group)
1689{
1690	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1691	struct ext4_free_extent *bex = &ac->ac_b_ex;
1692	struct ext4_free_extent *gex = &ac->ac_g_ex;
1693	struct ext4_free_extent ex;
1694	int max;
1695
1696	if (ac->ac_status == AC_STATUS_FOUND)
1697		return;
1698	/*
1699	 * We don't want to scan for a whole year
1700	 */
1701	if (ac->ac_found > sbi->s_mb_max_to_scan &&
1702			!(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1703		ac->ac_status = AC_STATUS_BREAK;
1704		return;
1705	}
1706
1707	/*
1708	 * Haven't found good chunk so far, let's continue
1709	 */
1710	if (bex->fe_len < gex->fe_len)
1711		return;
1712
1713	if ((finish_group || ac->ac_found > sbi->s_mb_min_to_scan)
1714			&& bex->fe_group == e4b->bd_group) {
1715		/* recheck chunk's availability - we don't know
1716		 * when it was found (within this lock-unlock
1717		 * period or not) */
1718		max = mb_find_extent(e4b, bex->fe_start, gex->fe_len, &ex);
1719		if (max >= gex->fe_len) {
1720			ext4_mb_use_best_found(ac, e4b);
1721			return;
1722		}
1723	}
1724}
1725
1726/*
1727 * The routine checks whether found extent is good enough. If it is,
1728 * then the extent gets marked used and flag is set to the context
1729 * to stop scanning. Otherwise, the extent is compared with the
1730 * previous found extent and if new one is better, then it's stored
1731 * in the context. Later, the best found extent will be used, if
1732 * mballoc can't find good enough extent.
1733 *
1734 * FIXME: real allocation policy is to be designed yet!
1735 */
1736static void ext4_mb_measure_extent(struct ext4_allocation_context *ac,
1737					struct ext4_free_extent *ex,
1738					struct ext4_buddy *e4b)
1739{
1740	struct ext4_free_extent *bex = &ac->ac_b_ex;
1741	struct ext4_free_extent *gex = &ac->ac_g_ex;
1742
1743	BUG_ON(ex->fe_len <= 0);
1744	BUG_ON(ex->fe_len > EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1745	BUG_ON(ex->fe_start >= EXT4_CLUSTERS_PER_GROUP(ac->ac_sb));
1746	BUG_ON(ac->ac_status != AC_STATUS_CONTINUE);
1747
1748	ac->ac_found++;
1749
1750	/*
1751	 * The special case - take what you catch first
1752	 */
1753	if (unlikely(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
1754		*bex = *ex;
1755		ext4_mb_use_best_found(ac, e4b);
1756		return;
1757	}
1758
1759	/*
1760	 * Let's check whether the chuck is good enough
1761	 */
1762	if (ex->fe_len == gex->fe_len) {
1763		*bex = *ex;
1764		ext4_mb_use_best_found(ac, e4b);
1765		return;
1766	}
1767
1768	/*
1769	 * If this is first found extent, just store it in the context
1770	 */
1771	if (bex->fe_len == 0) {
1772		*bex = *ex;
1773		return;
1774	}
1775
1776	/*
1777	 * If new found extent is better, store it in the context
1778	 */
1779	if (bex->fe_len < gex->fe_len) {
1780		/* if the request isn't satisfied, any found extent
1781		 * larger than previous best one is better */
1782		if (ex->fe_len > bex->fe_len)
1783			*bex = *ex;
1784	} else if (ex->fe_len > gex->fe_len) {
1785		/* if the request is satisfied, then we try to find
1786		 * an extent that still satisfy the request, but is
1787		 * smaller than previous one */
1788		if (ex->fe_len < bex->fe_len)
1789			*bex = *ex;
1790	}
1791
1792	ext4_mb_check_limits(ac, e4b, 0);
1793}
1794
1795static noinline_for_stack
1796int ext4_mb_try_best_found(struct ext4_allocation_context *ac,
1797					struct ext4_buddy *e4b)
1798{
1799	struct ext4_free_extent ex = ac->ac_b_ex;
1800	ext4_group_t group = ex.fe_group;
1801	int max;
1802	int err;
1803
1804	BUG_ON(ex.fe_len <= 0);
1805	err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1806	if (err)
1807		return err;
1808
1809	ext4_lock_group(ac->ac_sb, group);
1810	max = mb_find_extent(e4b, ex.fe_start, ex.fe_len, &ex);
1811
1812	if (max > 0) {
1813		ac->ac_b_ex = ex;
1814		ext4_mb_use_best_found(ac, e4b);
1815	}
1816
1817	ext4_unlock_group(ac->ac_sb, group);
1818	ext4_mb_unload_buddy(e4b);
1819
1820	return 0;
1821}
1822
1823static noinline_for_stack
1824int ext4_mb_find_by_goal(struct ext4_allocation_context *ac,
1825				struct ext4_buddy *e4b)
1826{
1827	ext4_group_t group = ac->ac_g_ex.fe_group;
1828	int max;
1829	int err;
1830	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
1831	struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
1832	struct ext4_free_extent ex;
1833
1834	if (!(ac->ac_flags & EXT4_MB_HINT_TRY_GOAL))
1835		return 0;
1836	if (grp->bb_free == 0)
1837		return 0;
1838
1839	err = ext4_mb_load_buddy(ac->ac_sb, group, e4b);
1840	if (err)
1841		return err;
1842
1843	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(e4b->bd_info))) {
1844		ext4_mb_unload_buddy(e4b);
1845		return 0;
1846	}
1847
1848	ext4_lock_group(ac->ac_sb, group);
1849	max = mb_find_extent(e4b, ac->ac_g_ex.fe_start,
1850			     ac->ac_g_ex.fe_len, &ex);
1851	ex.fe_logical = 0xDEADFA11; /* debug value */
1852
1853	if (max >= ac->ac_g_ex.fe_len && ac->ac_g_ex.fe_len == sbi->s_stripe) {
1854		ext4_fsblk_t start;
1855
1856		start = ext4_group_first_block_no(ac->ac_sb, e4b->bd_group) +
1857			ex.fe_start;
1858		/* use do_div to get remainder (would be 64-bit modulo) */
1859		if (do_div(start, sbi->s_stripe) == 0) {
1860			ac->ac_found++;
1861			ac->ac_b_ex = ex;
1862			ext4_mb_use_best_found(ac, e4b);
1863		}
1864	} else if (max >= ac->ac_g_ex.fe_len) {
1865		BUG_ON(ex.fe_len <= 0);
1866		BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1867		BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1868		ac->ac_found++;
1869		ac->ac_b_ex = ex;
1870		ext4_mb_use_best_found(ac, e4b);
1871	} else if (max > 0 && (ac->ac_flags & EXT4_MB_HINT_MERGE)) {
1872		/* Sometimes, caller may want to merge even small
1873		 * number of blocks to an existing extent */
1874		BUG_ON(ex.fe_len <= 0);
1875		BUG_ON(ex.fe_group != ac->ac_g_ex.fe_group);
1876		BUG_ON(ex.fe_start != ac->ac_g_ex.fe_start);
1877		ac->ac_found++;
1878		ac->ac_b_ex = ex;
1879		ext4_mb_use_best_found(ac, e4b);
1880	}
1881	ext4_unlock_group(ac->ac_sb, group);
1882	ext4_mb_unload_buddy(e4b);
1883
1884	return 0;
1885}
1886
1887/*
1888 * The routine scans buddy structures (not bitmap!) from given order
1889 * to max order and tries to find big enough chunk to satisfy the req
1890 */
1891static noinline_for_stack
1892void ext4_mb_simple_scan_group(struct ext4_allocation_context *ac,
1893					struct ext4_buddy *e4b)
1894{
1895	struct super_block *sb = ac->ac_sb;
1896	struct ext4_group_info *grp = e4b->bd_info;
1897	void *buddy;
1898	int i;
1899	int k;
1900	int max;
1901
1902	BUG_ON(ac->ac_2order <= 0);
1903	for (i = ac->ac_2order; i <= sb->s_blocksize_bits + 1; i++) {
1904		if (grp->bb_counters[i] == 0)
1905			continue;
1906
1907		buddy = mb_find_buddy(e4b, i, &max);
1908		BUG_ON(buddy == NULL);
1909
1910		k = mb_find_next_zero_bit(buddy, max, 0);
1911		BUG_ON(k >= max);
1912
1913		ac->ac_found++;
1914
1915		ac->ac_b_ex.fe_len = 1 << i;
1916		ac->ac_b_ex.fe_start = k << i;
1917		ac->ac_b_ex.fe_group = e4b->bd_group;
1918
1919		ext4_mb_use_best_found(ac, e4b);
1920
1921		BUG_ON(ac->ac_b_ex.fe_len != ac->ac_g_ex.fe_len);
1922
1923		if (EXT4_SB(sb)->s_mb_stats)
1924			atomic_inc(&EXT4_SB(sb)->s_bal_2orders);
1925
1926		break;
1927	}
1928}
1929
1930/*
1931 * The routine scans the group and measures all found extents.
1932 * In order to optimize scanning, caller must pass number of
1933 * free blocks in the group, so the routine can know upper limit.
1934 */
1935static noinline_for_stack
1936void ext4_mb_complex_scan_group(struct ext4_allocation_context *ac,
1937					struct ext4_buddy *e4b)
1938{
1939	struct super_block *sb = ac->ac_sb;
1940	void *bitmap = e4b->bd_bitmap;
1941	struct ext4_free_extent ex;
1942	int i;
1943	int free;
1944
1945	free = e4b->bd_info->bb_free;
1946	BUG_ON(free <= 0);
1947
1948	i = e4b->bd_info->bb_first_free;
1949
1950	while (free && ac->ac_status == AC_STATUS_CONTINUE) {
1951		i = mb_find_next_zero_bit(bitmap,
1952						EXT4_CLUSTERS_PER_GROUP(sb), i);
1953		if (i >= EXT4_CLUSTERS_PER_GROUP(sb)) {
1954			/*
1955			 * IF we have corrupt bitmap, we won't find any
1956			 * free blocks even though group info says we
1957			 * we have free blocks
1958			 */
1959			ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1960					"%d free clusters as per "
1961					"group info. But bitmap says 0",
1962					free);
1963			break;
1964		}
1965
1966		mb_find_extent(e4b, i, ac->ac_g_ex.fe_len, &ex);
1967		BUG_ON(ex.fe_len <= 0);
1968		if (free < ex.fe_len) {
1969			ext4_grp_locked_error(sb, e4b->bd_group, 0, 0,
1970					"%d free clusters as per "
1971					"group info. But got %d blocks",
1972					free, ex.fe_len);
1973			/*
1974			 * The number of free blocks differs. This mostly
1975			 * indicate that the bitmap is corrupt. So exit
1976			 * without claiming the space.
1977			 */
1978			break;
1979		}
1980		ex.fe_logical = 0xDEADC0DE; /* debug value */
1981		ext4_mb_measure_extent(ac, &ex, e4b);
1982
1983		i += ex.fe_len;
1984		free -= ex.fe_len;
1985	}
1986
1987	ext4_mb_check_limits(ac, e4b, 1);
1988}
1989
1990/*
1991 * This is a special case for storages like raid5
1992 * we try to find stripe-aligned chunks for stripe-size-multiple requests
1993 */
1994static noinline_for_stack
1995void ext4_mb_scan_aligned(struct ext4_allocation_context *ac,
1996				 struct ext4_buddy *e4b)
1997{
1998	struct super_block *sb = ac->ac_sb;
1999	struct ext4_sb_info *sbi = EXT4_SB(sb);
2000	void *bitmap = e4b->bd_bitmap;
2001	struct ext4_free_extent ex;
2002	ext4_fsblk_t first_group_block;
2003	ext4_fsblk_t a;
2004	ext4_grpblk_t i;
2005	int max;
2006
2007	BUG_ON(sbi->s_stripe == 0);
2008
2009	/* find first stripe-aligned block in group */
2010	first_group_block = ext4_group_first_block_no(sb, e4b->bd_group);
2011
2012	a = first_group_block + sbi->s_stripe - 1;
2013	do_div(a, sbi->s_stripe);
2014	i = (a * sbi->s_stripe) - first_group_block;
2015
2016	while (i < EXT4_CLUSTERS_PER_GROUP(sb)) {
2017		if (!mb_test_bit(i, bitmap)) {
2018			max = mb_find_extent(e4b, i, sbi->s_stripe, &ex);
2019			if (max >= sbi->s_stripe) {
2020				ac->ac_found++;
2021				ex.fe_logical = 0xDEADF00D; /* debug value */
2022				ac->ac_b_ex = ex;
2023				ext4_mb_use_best_found(ac, e4b);
2024				break;
2025			}
2026		}
2027		i += sbi->s_stripe;
2028	}
2029}
2030
2031/*
2032 * This is now called BEFORE we load the buddy bitmap.
2033 * Returns either 1 or 0 indicating that the group is either suitable
2034 * for the allocation or not. In addition it can also return negative
2035 * error code when something goes wrong.
2036 */
2037static int ext4_mb_good_group(struct ext4_allocation_context *ac,
2038				ext4_group_t group, int cr)
2039{
2040	unsigned free, fragments;
2041	int flex_size = ext4_flex_bg_size(EXT4_SB(ac->ac_sb));
2042	struct ext4_group_info *grp = ext4_get_group_info(ac->ac_sb, group);
2043
2044	BUG_ON(cr < 0 || cr >= 4);
2045
2046	free = grp->bb_free;
2047	if (free == 0)
2048		return 0;
2049	if (cr <= 2 && free < ac->ac_g_ex.fe_len)
2050		return 0;
2051
2052	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(grp)))
2053		return 0;
2054
2055	/* We only do this if the grp has never been initialized */
2056	if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
2057		int ret = ext4_mb_init_group(ac->ac_sb, group, GFP_NOFS);
2058		if (ret)
2059			return ret;
2060	}
2061
2062	fragments = grp->bb_fragments;
2063	if (fragments == 0)
2064		return 0;
2065
2066	switch (cr) {
2067	case 0:
2068		BUG_ON(ac->ac_2order == 0);
2069
2070		/* Avoid using the first bg of a flexgroup for data files */
2071		if ((ac->ac_flags & EXT4_MB_HINT_DATA) &&
2072		    (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) &&
2073		    ((group % flex_size) == 0))
2074			return 0;
2075
2076		if ((ac->ac_2order > ac->ac_sb->s_blocksize_bits+1) ||
2077		    (free / fragments) >= ac->ac_g_ex.fe_len)
2078			return 1;
2079
2080		if (grp->bb_largest_free_order < ac->ac_2order)
2081			return 0;
2082
2083		return 1;
2084	case 1:
2085		if ((free / fragments) >= ac->ac_g_ex.fe_len)
2086			return 1;
2087		break;
2088	case 2:
2089		if (free >= ac->ac_g_ex.fe_len)
2090			return 1;
2091		break;
2092	case 3:
2093		return 1;
2094	default:
2095		BUG();
2096	}
2097
2098	return 0;
2099}
2100
2101static noinline_for_stack int
2102ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
2103{
2104	ext4_group_t ngroups, group, i;
2105	int cr;
2106	int err = 0, first_err = 0;
2107	struct ext4_sb_info *sbi;
2108	struct super_block *sb;
2109	struct ext4_buddy e4b;
2110
2111	sb = ac->ac_sb;
2112	sbi = EXT4_SB(sb);
2113	ngroups = ext4_get_groups_count(sb);
2114	/* non-extent files are limited to low blocks/groups */
2115	if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)))
2116		ngroups = sbi->s_blockfile_groups;
2117
2118	BUG_ON(ac->ac_status == AC_STATUS_FOUND);
2119
2120	/* first, try the goal */
2121	err = ext4_mb_find_by_goal(ac, &e4b);
2122	if (err || ac->ac_status == AC_STATUS_FOUND)
2123		goto out;
2124
2125	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
2126		goto out;
2127
2128	/*
2129	 * ac->ac2_order is set only if the fe_len is a power of 2
2130	 * if ac2_order is set we also set criteria to 0 so that we
2131	 * try exact allocation using buddy.
2132	 */
2133	i = fls(ac->ac_g_ex.fe_len);
2134	ac->ac_2order = 0;
2135	/*
2136	 * We search using buddy data only if the order of the request
2137	 * is greater than equal to the sbi_s_mb_order2_reqs
2138	 * You can tune it via /sys/fs/ext4/<partition>/mb_order2_req
2139	 */
2140	if (i >= sbi->s_mb_order2_reqs) {
2141		/*
2142		 * This should tell if fe_len is exactly power of 2
2143		 */
2144		if ((ac->ac_g_ex.fe_len & (~(1 << (i - 1)))) == 0)
2145			ac->ac_2order = i - 1;
2146	}
2147
2148	/* if stream allocation is enabled, use global goal */
2149	if (ac->ac_flags & EXT4_MB_STREAM_ALLOC) {
2150		/* TBD: may be hot point */
2151		spin_lock(&sbi->s_md_lock);
2152		ac->ac_g_ex.fe_group = sbi->s_mb_last_group;
2153		ac->ac_g_ex.fe_start = sbi->s_mb_last_start;
2154		spin_unlock(&sbi->s_md_lock);
2155	}
2156
2157	/* Let's just scan groups to find more-less suitable blocks */
2158	cr = ac->ac_2order ? 0 : 1;
2159	/*
2160	 * cr == 0 try to get exact allocation,
2161	 * cr == 3  try to get anything
2162	 */
2163repeat:
2164	for (; cr < 4 && ac->ac_status == AC_STATUS_CONTINUE; cr++) {
2165		ac->ac_criteria = cr;
2166		/*
2167		 * searching for the right group start
2168		 * from the goal value specified
2169		 */
2170		group = ac->ac_g_ex.fe_group;
2171
2172		for (i = 0; i < ngroups; group++, i++) {
2173			int ret = 0;
2174			cond_resched();
2175			/*
2176			 * Artificially restricted ngroups for non-extent
2177			 * files makes group > ngroups possible on first loop.
2178			 */
2179			if (group >= ngroups)
2180				group = 0;
2181
2182			/* This now checks without needing the buddy page */
2183			ret = ext4_mb_good_group(ac, group, cr);
2184			if (ret <= 0) {
2185				if (!first_err)
2186					first_err = ret;
2187				continue;
2188			}
2189
2190			err = ext4_mb_load_buddy(sb, group, &e4b);
2191			if (err)
2192				goto out;
2193
2194			ext4_lock_group(sb, group);
2195
2196			/*
2197			 * We need to check again after locking the
2198			 * block group
2199			 */
2200			ret = ext4_mb_good_group(ac, group, cr);
2201			if (ret <= 0) {
2202				ext4_unlock_group(sb, group);
2203				ext4_mb_unload_buddy(&e4b);
2204				if (!first_err)
2205					first_err = ret;
2206				continue;
2207			}
2208
2209			ac->ac_groups_scanned++;
2210			if (cr == 0 && ac->ac_2order < sb->s_blocksize_bits+2)
2211				ext4_mb_simple_scan_group(ac, &e4b);
2212			else if (cr == 1 && sbi->s_stripe &&
2213					!(ac->ac_g_ex.fe_len % sbi->s_stripe))
2214				ext4_mb_scan_aligned(ac, &e4b);
2215			else
2216				ext4_mb_complex_scan_group(ac, &e4b);
2217
2218			ext4_unlock_group(sb, group);
2219			ext4_mb_unload_buddy(&e4b);
2220
2221			if (ac->ac_status != AC_STATUS_CONTINUE)
2222				break;
2223		}
2224	}
2225
2226	if (ac->ac_b_ex.fe_len > 0 && ac->ac_status != AC_STATUS_FOUND &&
2227	    !(ac->ac_flags & EXT4_MB_HINT_FIRST)) {
2228		/*
2229		 * We've been searching too long. Let's try to allocate
2230		 * the best chunk we've found so far
2231		 */
2232
2233		ext4_mb_try_best_found(ac, &e4b);
2234		if (ac->ac_status != AC_STATUS_FOUND) {
2235			/*
2236			 * Someone more lucky has already allocated it.
2237			 * The only thing we can do is just take first
2238			 * found block(s)
2239			printk(KERN_DEBUG "EXT4-fs: someone won our chunk\n");
2240			 */
2241			ac->ac_b_ex.fe_group = 0;
2242			ac->ac_b_ex.fe_start = 0;
2243			ac->ac_b_ex.fe_len = 0;
2244			ac->ac_status = AC_STATUS_CONTINUE;
2245			ac->ac_flags |= EXT4_MB_HINT_FIRST;
2246			cr = 3;
2247			atomic_inc(&sbi->s_mb_lost_chunks);
2248			goto repeat;
2249		}
2250	}
2251out:
2252	if (!err && ac->ac_status != AC_STATUS_FOUND && first_err)
2253		err = first_err;
2254	return err;
2255}
2256
2257static void *ext4_mb_seq_groups_start(struct seq_file *seq, loff_t *pos)
2258{
2259	struct super_block *sb = seq->private;
2260	ext4_group_t group;
2261
2262	if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2263		return NULL;
2264	group = *pos + 1;
2265	return (void *) ((unsigned long) group);
2266}
2267
2268static void *ext4_mb_seq_groups_next(struct seq_file *seq, void *v, loff_t *pos)
2269{
2270	struct super_block *sb = seq->private;
2271	ext4_group_t group;
2272
2273	++*pos;
2274	if (*pos < 0 || *pos >= ext4_get_groups_count(sb))
2275		return NULL;
2276	group = *pos + 1;
2277	return (void *) ((unsigned long) group);
2278}
2279
2280static int ext4_mb_seq_groups_show(struct seq_file *seq, void *v)
2281{
2282	struct super_block *sb = seq->private;
2283	ext4_group_t group = (ext4_group_t) ((unsigned long) v);
2284	int i;
2285	int err, buddy_loaded = 0;
2286	struct ext4_buddy e4b;
2287	struct ext4_group_info *grinfo;
2288	struct sg {
2289		struct ext4_group_info info;
2290		ext4_grpblk_t counters[EXT4_MAX_BLOCK_LOG_SIZE + 2];
2291	} sg;
2292
2293	group--;
2294	if (group == 0)
2295		seq_puts(seq, "#group: free  frags first ["
2296			      " 2^0   2^1   2^2   2^3   2^4   2^5   2^6  "
2297			      " 2^7   2^8   2^9   2^10  2^11  2^12  2^13  ]\n");
2298
2299	i = (sb->s_blocksize_bits + 2) * sizeof(sg.info.bb_counters[0]) +
2300		sizeof(struct ext4_group_info);
2301	grinfo = ext4_get_group_info(sb, group);
2302	/* Load the group info in memory only if not already loaded. */
2303	if (unlikely(EXT4_MB_GRP_NEED_INIT(grinfo))) {
2304		err = ext4_mb_load_buddy(sb, group, &e4b);
2305		if (err) {
2306			seq_printf(seq, "#%-5u: I/O error\n", group);
2307			return 0;
2308		}
2309		buddy_loaded = 1;
2310	}
2311
2312	memcpy(&sg, ext4_get_group_info(sb, group), i);
2313
2314	if (buddy_loaded)
2315		ext4_mb_unload_buddy(&e4b);
2316
2317	seq_printf(seq, "#%-5u: %-5u %-5u %-5u [", group, sg.info.bb_free,
2318			sg.info.bb_fragments, sg.info.bb_first_free);
2319	for (i = 0; i <= 13; i++)
2320		seq_printf(seq, " %-5u", i <= sb->s_blocksize_bits + 1 ?
2321				sg.info.bb_counters[i] : 0);
2322	seq_printf(seq, " ]\n");
2323
2324	return 0;
2325}
2326
2327static void ext4_mb_seq_groups_stop(struct seq_file *seq, void *v)
2328{
2329}
2330
2331static const struct seq_operations ext4_mb_seq_groups_ops = {
2332	.start  = ext4_mb_seq_groups_start,
2333	.next   = ext4_mb_seq_groups_next,
2334	.stop   = ext4_mb_seq_groups_stop,
2335	.show   = ext4_mb_seq_groups_show,
2336};
2337
2338static int ext4_mb_seq_groups_open(struct inode *inode, struct file *file)
2339{
2340	struct super_block *sb = PDE_DATA(inode);
2341	int rc;
2342
2343	rc = seq_open(file, &ext4_mb_seq_groups_ops);
2344	if (rc == 0) {
2345		struct seq_file *m = file->private_data;
2346		m->private = sb;
2347	}
2348	return rc;
2349
2350}
2351
2352const struct file_operations ext4_seq_mb_groups_fops = {
 
2353	.open		= ext4_mb_seq_groups_open,
2354	.read		= seq_read,
2355	.llseek		= seq_lseek,
2356	.release	= seq_release,
2357};
2358
2359static struct kmem_cache *get_groupinfo_cache(int blocksize_bits)
2360{
2361	int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2362	struct kmem_cache *cachep = ext4_groupinfo_caches[cache_index];
2363
2364	BUG_ON(!cachep);
2365	return cachep;
2366}
2367
2368/*
2369 * Allocate the top-level s_group_info array for the specified number
2370 * of groups
2371 */
2372int ext4_mb_alloc_groupinfo(struct super_block *sb, ext4_group_t ngroups)
2373{
2374	struct ext4_sb_info *sbi = EXT4_SB(sb);
2375	unsigned size;
2376	struct ext4_group_info ***new_groupinfo;
2377
2378	size = (ngroups + EXT4_DESC_PER_BLOCK(sb) - 1) >>
2379		EXT4_DESC_PER_BLOCK_BITS(sb);
2380	if (size <= sbi->s_group_info_size)
2381		return 0;
2382
2383	size = roundup_pow_of_two(sizeof(*sbi->s_group_info) * size);
2384	new_groupinfo = ext4_kvzalloc(size, GFP_KERNEL);
2385	if (!new_groupinfo) {
2386		ext4_msg(sb, KERN_ERR, "can't allocate buddy meta group");
2387		return -ENOMEM;
2388	}
2389	if (sbi->s_group_info) {
2390		memcpy(new_groupinfo, sbi->s_group_info,
2391		       sbi->s_group_info_size * sizeof(*sbi->s_group_info));
2392		kvfree(sbi->s_group_info);
2393	}
2394	sbi->s_group_info = new_groupinfo;
2395	sbi->s_group_info_size = size / sizeof(*sbi->s_group_info);
2396	ext4_debug("allocated s_groupinfo array for %d meta_bg's\n", 
2397		   sbi->s_group_info_size);
2398	return 0;
2399}
2400
2401/* Create and initialize ext4_group_info data for the given group. */
2402int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group,
2403			  struct ext4_group_desc *desc)
2404{
2405	int i;
2406	int metalen = 0;
2407	struct ext4_sb_info *sbi = EXT4_SB(sb);
2408	struct ext4_group_info **meta_group_info;
2409	struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2410
2411	/*
2412	 * First check if this group is the first of a reserved block.
2413	 * If it's true, we have to allocate a new table of pointers
2414	 * to ext4_group_info structures
2415	 */
2416	if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2417		metalen = sizeof(*meta_group_info) <<
2418			EXT4_DESC_PER_BLOCK_BITS(sb);
2419		meta_group_info = kmalloc(metalen, GFP_NOFS);
2420		if (meta_group_info == NULL) {
2421			ext4_msg(sb, KERN_ERR, "can't allocate mem "
2422				 "for a buddy group");
2423			goto exit_meta_group_info;
2424		}
2425		sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] =
2426			meta_group_info;
2427	}
2428
2429	meta_group_info =
2430		sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)];
2431	i = group & (EXT4_DESC_PER_BLOCK(sb) - 1);
2432
2433	meta_group_info[i] = kmem_cache_zalloc(cachep, GFP_NOFS);
2434	if (meta_group_info[i] == NULL) {
2435		ext4_msg(sb, KERN_ERR, "can't allocate buddy mem");
2436		goto exit_group_info;
2437	}
2438	set_bit(EXT4_GROUP_INFO_NEED_INIT_BIT,
2439		&(meta_group_info[i]->bb_state));
2440
2441	/*
2442	 * initialize bb_free to be able to skip
2443	 * empty groups without initialization
2444	 */
2445	if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2446		meta_group_info[i]->bb_free =
2447			ext4_free_clusters_after_init(sb, group, desc);
2448	} else {
2449		meta_group_info[i]->bb_free =
2450			ext4_free_group_clusters(sb, desc);
2451	}
2452
2453	INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
2454	init_rwsem(&meta_group_info[i]->alloc_sem);
2455	meta_group_info[i]->bb_free_root = RB_ROOT;
2456	meta_group_info[i]->bb_largest_free_order = -1;  /* uninit */
2457
2458#ifdef DOUBLE_CHECK
2459	{
2460		struct buffer_head *bh;
2461		meta_group_info[i]->bb_bitmap =
2462			kmalloc(sb->s_blocksize, GFP_NOFS);
2463		BUG_ON(meta_group_info[i]->bb_bitmap == NULL);
2464		bh = ext4_read_block_bitmap(sb, group);
2465		BUG_ON(IS_ERR_OR_NULL(bh));
2466		memcpy(meta_group_info[i]->bb_bitmap, bh->b_data,
2467			sb->s_blocksize);
2468		put_bh(bh);
2469	}
2470#endif
2471
2472	return 0;
2473
2474exit_group_info:
2475	/* If a meta_group_info table has been allocated, release it now */
2476	if (group % EXT4_DESC_PER_BLOCK(sb) == 0) {
2477		kfree(sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)]);
2478		sbi->s_group_info[group >> EXT4_DESC_PER_BLOCK_BITS(sb)] = NULL;
2479	}
2480exit_meta_group_info:
2481	return -ENOMEM;
2482} /* ext4_mb_add_groupinfo */
2483
2484static int ext4_mb_init_backend(struct super_block *sb)
2485{
2486	ext4_group_t ngroups = ext4_get_groups_count(sb);
2487	ext4_group_t i;
2488	struct ext4_sb_info *sbi = EXT4_SB(sb);
2489	int err;
2490	struct ext4_group_desc *desc;
2491	struct kmem_cache *cachep;
2492
2493	err = ext4_mb_alloc_groupinfo(sb, ngroups);
2494	if (err)
2495		return err;
2496
2497	sbi->s_buddy_cache = new_inode(sb);
2498	if (sbi->s_buddy_cache == NULL) {
2499		ext4_msg(sb, KERN_ERR, "can't get new inode");
2500		goto err_freesgi;
2501	}
2502	/* To avoid potentially colliding with an valid on-disk inode number,
2503	 * use EXT4_BAD_INO for the buddy cache inode number.  This inode is
2504	 * not in the inode hash, so it should never be found by iget(), but
2505	 * this will avoid confusion if it ever shows up during debugging. */
2506	sbi->s_buddy_cache->i_ino = EXT4_BAD_INO;
2507	EXT4_I(sbi->s_buddy_cache)->i_disksize = 0;
2508	for (i = 0; i < ngroups; i++) {
2509		desc = ext4_get_group_desc(sb, i, NULL);
2510		if (desc == NULL) {
2511			ext4_msg(sb, KERN_ERR, "can't read descriptor %u", i);
2512			goto err_freebuddy;
2513		}
2514		if (ext4_mb_add_groupinfo(sb, i, desc) != 0)
2515			goto err_freebuddy;
2516	}
2517
2518	return 0;
2519
2520err_freebuddy:
2521	cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2522	while (i-- > 0)
2523		kmem_cache_free(cachep, ext4_get_group_info(sb, i));
2524	i = sbi->s_group_info_size;
2525	while (i-- > 0)
2526		kfree(sbi->s_group_info[i]);
2527	iput(sbi->s_buddy_cache);
2528err_freesgi:
2529	kvfree(sbi->s_group_info);
2530	return -ENOMEM;
2531}
2532
2533static void ext4_groupinfo_destroy_slabs(void)
2534{
2535	int i;
2536
2537	for (i = 0; i < NR_GRPINFO_CACHES; i++) {
2538		if (ext4_groupinfo_caches[i])
2539			kmem_cache_destroy(ext4_groupinfo_caches[i]);
2540		ext4_groupinfo_caches[i] = NULL;
2541	}
2542}
2543
2544static int ext4_groupinfo_create_slab(size_t size)
2545{
2546	static DEFINE_MUTEX(ext4_grpinfo_slab_create_mutex);
2547	int slab_size;
2548	int blocksize_bits = order_base_2(size);
2549	int cache_index = blocksize_bits - EXT4_MIN_BLOCK_LOG_SIZE;
2550	struct kmem_cache *cachep;
2551
2552	if (cache_index >= NR_GRPINFO_CACHES)
2553		return -EINVAL;
2554
2555	if (unlikely(cache_index < 0))
2556		cache_index = 0;
2557
2558	mutex_lock(&ext4_grpinfo_slab_create_mutex);
2559	if (ext4_groupinfo_caches[cache_index]) {
2560		mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2561		return 0;	/* Already created */
2562	}
2563
2564	slab_size = offsetof(struct ext4_group_info,
2565				bb_counters[blocksize_bits + 2]);
2566
2567	cachep = kmem_cache_create(ext4_groupinfo_slab_names[cache_index],
2568					slab_size, 0, SLAB_RECLAIM_ACCOUNT,
2569					NULL);
2570
2571	ext4_groupinfo_caches[cache_index] = cachep;
2572
2573	mutex_unlock(&ext4_grpinfo_slab_create_mutex);
2574	if (!cachep) {
2575		printk(KERN_EMERG
2576		       "EXT4-fs: no memory for groupinfo slab cache\n");
2577		return -ENOMEM;
2578	}
2579
2580	return 0;
2581}
2582
2583int ext4_mb_init(struct super_block *sb)
2584{
2585	struct ext4_sb_info *sbi = EXT4_SB(sb);
2586	unsigned i, j;
2587	unsigned offset, offset_incr;
2588	unsigned max;
2589	int ret;
2590
2591	i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_offsets);
2592
2593	sbi->s_mb_offsets = kmalloc(i, GFP_KERNEL);
2594	if (sbi->s_mb_offsets == NULL) {
2595		ret = -ENOMEM;
2596		goto out;
2597	}
2598
2599	i = (sb->s_blocksize_bits + 2) * sizeof(*sbi->s_mb_maxs);
2600	sbi->s_mb_maxs = kmalloc(i, GFP_KERNEL);
2601	if (sbi->s_mb_maxs == NULL) {
2602		ret = -ENOMEM;
2603		goto out;
2604	}
2605
2606	ret = ext4_groupinfo_create_slab(sb->s_blocksize);
2607	if (ret < 0)
2608		goto out;
2609
2610	/* order 0 is regular bitmap */
2611	sbi->s_mb_maxs[0] = sb->s_blocksize << 3;
2612	sbi->s_mb_offsets[0] = 0;
2613
2614	i = 1;
2615	offset = 0;
2616	offset_incr = 1 << (sb->s_blocksize_bits - 1);
2617	max = sb->s_blocksize << 2;
2618	do {
2619		sbi->s_mb_offsets[i] = offset;
2620		sbi->s_mb_maxs[i] = max;
2621		offset += offset_incr;
2622		offset_incr = offset_incr >> 1;
2623		max = max >> 1;
2624		i++;
2625	} while (i <= sb->s_blocksize_bits + 1);
2626
2627	spin_lock_init(&sbi->s_md_lock);
2628	spin_lock_init(&sbi->s_bal_lock);
2629	sbi->s_mb_free_pending = 0;
2630
2631	sbi->s_mb_max_to_scan = MB_DEFAULT_MAX_TO_SCAN;
2632	sbi->s_mb_min_to_scan = MB_DEFAULT_MIN_TO_SCAN;
2633	sbi->s_mb_stats = MB_DEFAULT_STATS;
2634	sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
2635	sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
2636	/*
2637	 * The default group preallocation is 512, which for 4k block
2638	 * sizes translates to 2 megabytes.  However for bigalloc file
2639	 * systems, this is probably too big (i.e, if the cluster size
2640	 * is 1 megabyte, then group preallocation size becomes half a
2641	 * gigabyte!).  As a default, we will keep a two megabyte
2642	 * group pralloc size for cluster sizes up to 64k, and after
2643	 * that, we will force a minimum group preallocation size of
2644	 * 32 clusters.  This translates to 8 megs when the cluster
2645	 * size is 256k, and 32 megs when the cluster size is 1 meg,
2646	 * which seems reasonable as a default.
2647	 */
2648	sbi->s_mb_group_prealloc = max(MB_DEFAULT_GROUP_PREALLOC >>
2649				       sbi->s_cluster_bits, 32);
2650	/*
2651	 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
2652	 * to the lowest multiple of s_stripe which is bigger than
2653	 * the s_mb_group_prealloc as determined above. We want
2654	 * the preallocation size to be an exact multiple of the
2655	 * RAID stripe size so that preallocations don't fragment
2656	 * the stripes.
2657	 */
2658	if (sbi->s_stripe > 1) {
2659		sbi->s_mb_group_prealloc = roundup(
2660			sbi->s_mb_group_prealloc, sbi->s_stripe);
2661	}
2662
2663	sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
2664	if (sbi->s_locality_groups == NULL) {
2665		ret = -ENOMEM;
2666		goto out;
2667	}
2668	for_each_possible_cpu(i) {
2669		struct ext4_locality_group *lg;
2670		lg = per_cpu_ptr(sbi->s_locality_groups, i);
2671		mutex_init(&lg->lg_mutex);
2672		for (j = 0; j < PREALLOC_TB_SIZE; j++)
2673			INIT_LIST_HEAD(&lg->lg_prealloc_list[j]);
2674		spin_lock_init(&lg->lg_prealloc_lock);
2675	}
2676
2677	/* init file for buddy data */
2678	ret = ext4_mb_init_backend(sb);
2679	if (ret != 0)
2680		goto out_free_locality_groups;
2681
2682	return 0;
2683
2684out_free_locality_groups:
2685	free_percpu(sbi->s_locality_groups);
2686	sbi->s_locality_groups = NULL;
2687out:
2688	kfree(sbi->s_mb_offsets);
2689	sbi->s_mb_offsets = NULL;
2690	kfree(sbi->s_mb_maxs);
2691	sbi->s_mb_maxs = NULL;
2692	return ret;
2693}
2694
2695/* need to called with the ext4 group lock held */
2696static void ext4_mb_cleanup_pa(struct ext4_group_info *grp)
2697{
2698	struct ext4_prealloc_space *pa;
2699	struct list_head *cur, *tmp;
2700	int count = 0;
2701
2702	list_for_each_safe(cur, tmp, &grp->bb_prealloc_list) {
2703		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
2704		list_del(&pa->pa_group_list);
2705		count++;
2706		kmem_cache_free(ext4_pspace_cachep, pa);
2707	}
2708	if (count)
2709		mb_debug(1, "mballoc: %u PAs left\n", count);
2710
2711}
2712
2713int ext4_mb_release(struct super_block *sb)
2714{
2715	ext4_group_t ngroups = ext4_get_groups_count(sb);
2716	ext4_group_t i;
2717	int num_meta_group_infos;
2718	struct ext4_group_info *grinfo;
2719	struct ext4_sb_info *sbi = EXT4_SB(sb);
2720	struct kmem_cache *cachep = get_groupinfo_cache(sb->s_blocksize_bits);
2721
2722	if (sbi->s_group_info) {
2723		for (i = 0; i < ngroups; i++) {
2724			grinfo = ext4_get_group_info(sb, i);
2725#ifdef DOUBLE_CHECK
2726			kfree(grinfo->bb_bitmap);
2727#endif
2728			ext4_lock_group(sb, i);
2729			ext4_mb_cleanup_pa(grinfo);
2730			ext4_unlock_group(sb, i);
2731			kmem_cache_free(cachep, grinfo);
2732		}
2733		num_meta_group_infos = (ngroups +
2734				EXT4_DESC_PER_BLOCK(sb) - 1) >>
2735			EXT4_DESC_PER_BLOCK_BITS(sb);
2736		for (i = 0; i < num_meta_group_infos; i++)
2737			kfree(sbi->s_group_info[i]);
2738		kvfree(sbi->s_group_info);
2739	}
2740	kfree(sbi->s_mb_offsets);
2741	kfree(sbi->s_mb_maxs);
2742	iput(sbi->s_buddy_cache);
2743	if (sbi->s_mb_stats) {
2744		ext4_msg(sb, KERN_INFO,
2745		       "mballoc: %u blocks %u reqs (%u success)",
2746				atomic_read(&sbi->s_bal_allocated),
2747				atomic_read(&sbi->s_bal_reqs),
2748				atomic_read(&sbi->s_bal_success));
2749		ext4_msg(sb, KERN_INFO,
2750		      "mballoc: %u extents scanned, %u goal hits, "
2751				"%u 2^N hits, %u breaks, %u lost",
2752				atomic_read(&sbi->s_bal_ex_scanned),
2753				atomic_read(&sbi->s_bal_goals),
2754				atomic_read(&sbi->s_bal_2orders),
2755				atomic_read(&sbi->s_bal_breaks),
2756				atomic_read(&sbi->s_mb_lost_chunks));
2757		ext4_msg(sb, KERN_INFO,
2758		       "mballoc: %lu generated and it took %Lu",
2759				sbi->s_mb_buddies_generated,
2760				sbi->s_mb_generation_time);
2761		ext4_msg(sb, KERN_INFO,
2762		       "mballoc: %u preallocated, %u discarded",
2763				atomic_read(&sbi->s_mb_preallocated),
2764				atomic_read(&sbi->s_mb_discarded));
2765	}
2766
2767	free_percpu(sbi->s_locality_groups);
2768
2769	return 0;
2770}
2771
2772static inline int ext4_issue_discard(struct super_block *sb,
2773		ext4_group_t block_group, ext4_grpblk_t cluster, int count)
2774{
2775	ext4_fsblk_t discard_block;
2776
2777	discard_block = (EXT4_C2B(EXT4_SB(sb), cluster) +
2778			 ext4_group_first_block_no(sb, block_group));
2779	count = EXT4_C2B(EXT4_SB(sb), count);
2780	trace_ext4_discard_blocks(sb,
2781			(unsigned long long) discard_block, count);
2782	return sb_issue_discard(sb, discard_block, count, GFP_NOFS, 0);
2783}
2784
2785/*
2786 * This function is called by the jbd2 layer once the commit has finished,
2787 * so we know we can free the blocks that were released with that commit.
2788 */
2789static void ext4_free_data_callback(struct super_block *sb,
2790				    struct ext4_journal_cb_entry *jce,
2791				    int rc)
2792{
2793	struct ext4_free_data *entry = (struct ext4_free_data *)jce;
2794	struct ext4_buddy e4b;
2795	struct ext4_group_info *db;
2796	int err, count = 0, count2 = 0;
2797
2798	mb_debug(1, "gonna free %u blocks in group %u (0x%p):",
2799		 entry->efd_count, entry->efd_group, entry);
2800
2801	if (test_opt(sb, DISCARD)) {
2802		err = ext4_issue_discard(sb, entry->efd_group,
2803					 entry->efd_start_cluster,
2804					 entry->efd_count);
2805		if (err && err != -EOPNOTSUPP)
2806			ext4_msg(sb, KERN_WARNING, "discard request in"
2807				 " group:%d block:%d count:%d failed"
2808				 " with %d", entry->efd_group,
2809				 entry->efd_start_cluster,
2810				 entry->efd_count, err);
2811	}
2812
2813	err = ext4_mb_load_buddy(sb, entry->efd_group, &e4b);
2814	/* we expect to find existing buddy because it's pinned */
2815	BUG_ON(err != 0);
2816
2817	spin_lock(&EXT4_SB(sb)->s_md_lock);
2818	EXT4_SB(sb)->s_mb_free_pending -= entry->efd_count;
2819	spin_unlock(&EXT4_SB(sb)->s_md_lock);
2820
2821	db = e4b.bd_info;
2822	/* there are blocks to put in buddy to make them really free */
2823	count += entry->efd_count;
2824	count2++;
2825	ext4_lock_group(sb, entry->efd_group);
2826	/* Take it out of per group rb tree */
2827	rb_erase(&entry->efd_node, &(db->bb_free_root));
2828	mb_free_blocks(NULL, &e4b, entry->efd_start_cluster, entry->efd_count);
2829
2830	/*
2831	 * Clear the trimmed flag for the group so that the next
2832	 * ext4_trim_fs can trim it.
2833	 * If the volume is mounted with -o discard, online discard
2834	 * is supported and the free blocks will be trimmed online.
2835	 */
2836	if (!test_opt(sb, DISCARD))
2837		EXT4_MB_GRP_CLEAR_TRIMMED(db);
2838
2839	if (!db->bb_free_root.rb_node) {
2840		/* No more items in the per group rb tree
2841		 * balance refcounts from ext4_mb_free_metadata()
2842		 */
2843		put_page(e4b.bd_buddy_page);
2844		put_page(e4b.bd_bitmap_page);
2845	}
2846	ext4_unlock_group(sb, entry->efd_group);
2847	kmem_cache_free(ext4_free_data_cachep, entry);
2848	ext4_mb_unload_buddy(&e4b);
2849
2850	mb_debug(1, "freed %u blocks in %u structures\n", count, count2);
2851}
2852
2853int __init ext4_init_mballoc(void)
2854{
2855	ext4_pspace_cachep = KMEM_CACHE(ext4_prealloc_space,
2856					SLAB_RECLAIM_ACCOUNT);
2857	if (ext4_pspace_cachep == NULL)
2858		return -ENOMEM;
2859
2860	ext4_ac_cachep = KMEM_CACHE(ext4_allocation_context,
2861				    SLAB_RECLAIM_ACCOUNT);
2862	if (ext4_ac_cachep == NULL) {
2863		kmem_cache_destroy(ext4_pspace_cachep);
2864		return -ENOMEM;
2865	}
2866
2867	ext4_free_data_cachep = KMEM_CACHE(ext4_free_data,
2868					   SLAB_RECLAIM_ACCOUNT);
2869	if (ext4_free_data_cachep == NULL) {
2870		kmem_cache_destroy(ext4_pspace_cachep);
2871		kmem_cache_destroy(ext4_ac_cachep);
2872		return -ENOMEM;
2873	}
2874	return 0;
2875}
2876
2877void ext4_exit_mballoc(void)
2878{
2879	/*
2880	 * Wait for completion of call_rcu()'s on ext4_pspace_cachep
2881	 * before destroying the slab cache.
2882	 */
2883	rcu_barrier();
2884	kmem_cache_destroy(ext4_pspace_cachep);
2885	kmem_cache_destroy(ext4_ac_cachep);
2886	kmem_cache_destroy(ext4_free_data_cachep);
2887	ext4_groupinfo_destroy_slabs();
2888}
2889
2890
2891/*
2892 * Check quota and mark chosen space (ac->ac_b_ex) non-free in bitmaps
2893 * Returns 0 if success or error code
2894 */
2895static noinline_for_stack int
2896ext4_mb_mark_diskspace_used(struct ext4_allocation_context *ac,
2897				handle_t *handle, unsigned int reserv_clstrs)
2898{
2899	struct buffer_head *bitmap_bh = NULL;
2900	struct ext4_group_desc *gdp;
2901	struct buffer_head *gdp_bh;
2902	struct ext4_sb_info *sbi;
2903	struct super_block *sb;
2904	ext4_fsblk_t block;
2905	int err, len;
2906
2907	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
2908	BUG_ON(ac->ac_b_ex.fe_len <= 0);
2909
2910	sb = ac->ac_sb;
2911	sbi = EXT4_SB(sb);
2912
2913	bitmap_bh = ext4_read_block_bitmap(sb, ac->ac_b_ex.fe_group);
2914	if (IS_ERR(bitmap_bh)) {
2915		err = PTR_ERR(bitmap_bh);
2916		bitmap_bh = NULL;
2917		goto out_err;
2918	}
2919
2920	BUFFER_TRACE(bitmap_bh, "getting write access");
2921	err = ext4_journal_get_write_access(handle, bitmap_bh);
2922	if (err)
2923		goto out_err;
2924
2925	err = -EIO;
2926	gdp = ext4_get_group_desc(sb, ac->ac_b_ex.fe_group, &gdp_bh);
2927	if (!gdp)
2928		goto out_err;
2929
2930	ext4_debug("using block group %u(%d)\n", ac->ac_b_ex.fe_group,
2931			ext4_free_group_clusters(sb, gdp));
2932
2933	BUFFER_TRACE(gdp_bh, "get_write_access");
2934	err = ext4_journal_get_write_access(handle, gdp_bh);
2935	if (err)
2936		goto out_err;
2937
2938	block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
2939
2940	len = EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
2941	if (!ext4_data_block_valid(sbi, block, len)) {
2942		ext4_error(sb, "Allocating blocks %llu-%llu which overlap "
2943			   "fs metadata", block, block+len);
2944		/* File system mounted not to panic on error
2945		 * Fix the bitmap and return EFSCORRUPTED
2946		 * We leak some of the blocks here.
2947		 */
2948		ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2949		ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2950			      ac->ac_b_ex.fe_len);
2951		ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2952		err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2953		if (!err)
2954			err = -EFSCORRUPTED;
2955		goto out_err;
2956	}
2957
2958	ext4_lock_group(sb, ac->ac_b_ex.fe_group);
2959#ifdef AGGRESSIVE_CHECK
2960	{
2961		int i;
2962		for (i = 0; i < ac->ac_b_ex.fe_len; i++) {
2963			BUG_ON(mb_test_bit(ac->ac_b_ex.fe_start + i,
2964						bitmap_bh->b_data));
2965		}
2966	}
2967#endif
2968	ext4_set_bits(bitmap_bh->b_data, ac->ac_b_ex.fe_start,
2969		      ac->ac_b_ex.fe_len);
2970	if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
2971		gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
2972		ext4_free_group_clusters_set(sb, gdp,
2973					     ext4_free_clusters_after_init(sb,
2974						ac->ac_b_ex.fe_group, gdp));
2975	}
2976	len = ext4_free_group_clusters(sb, gdp) - ac->ac_b_ex.fe_len;
2977	ext4_free_group_clusters_set(sb, gdp, len);
2978	ext4_block_bitmap_csum_set(sb, ac->ac_b_ex.fe_group, gdp, bitmap_bh);
2979	ext4_group_desc_csum_set(sb, ac->ac_b_ex.fe_group, gdp);
2980
2981	ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
2982	percpu_counter_sub(&sbi->s_freeclusters_counter, ac->ac_b_ex.fe_len);
2983	/*
2984	 * Now reduce the dirty block count also. Should not go negative
2985	 */
2986	if (!(ac->ac_flags & EXT4_MB_DELALLOC_RESERVED))
2987		/* release all the reserved blocks if non delalloc */
2988		percpu_counter_sub(&sbi->s_dirtyclusters_counter,
2989				   reserv_clstrs);
2990
2991	if (sbi->s_log_groups_per_flex) {
2992		ext4_group_t flex_group = ext4_flex_group(sbi,
2993							  ac->ac_b_ex.fe_group);
2994		atomic64_sub(ac->ac_b_ex.fe_len,
2995			     &sbi->s_flex_groups[flex_group].free_clusters);
2996	}
2997
2998	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
2999	if (err)
3000		goto out_err;
3001	err = ext4_handle_dirty_metadata(handle, NULL, gdp_bh);
3002
3003out_err:
3004	brelse(bitmap_bh);
3005	return err;
3006}
3007
3008/*
3009 * here we normalize request for locality group
3010 * Group request are normalized to s_mb_group_prealloc, which goes to
3011 * s_strip if we set the same via mount option.
3012 * s_mb_group_prealloc can be configured via
3013 * /sys/fs/ext4/<partition>/mb_group_prealloc
3014 *
3015 * XXX: should we try to preallocate more than the group has now?
3016 */
3017static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
3018{
3019	struct super_block *sb = ac->ac_sb;
3020	struct ext4_locality_group *lg = ac->ac_lg;
3021
3022	BUG_ON(lg == NULL);
3023	ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
3024	mb_debug(1, "#%u: goal %u blocks for locality group\n",
3025		current->pid, ac->ac_g_ex.fe_len);
3026}
3027
3028/*
3029 * Normalization means making request better in terms of
3030 * size and alignment
3031 */
3032static noinline_for_stack void
3033ext4_mb_normalize_request(struct ext4_allocation_context *ac,
3034				struct ext4_allocation_request *ar)
3035{
3036	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3037	int bsbits, max;
3038	ext4_lblk_t end;
3039	loff_t size, start_off;
3040	loff_t orig_size __maybe_unused;
3041	ext4_lblk_t start;
3042	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3043	struct ext4_prealloc_space *pa;
3044
3045	/* do normalize only data requests, metadata requests
3046	   do not need preallocation */
3047	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3048		return;
3049
3050	/* sometime caller may want exact blocks */
3051	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
3052		return;
3053
3054	/* caller may indicate that preallocation isn't
3055	 * required (it's a tail, for example) */
3056	if (ac->ac_flags & EXT4_MB_HINT_NOPREALLOC)
3057		return;
3058
3059	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC) {
3060		ext4_mb_normalize_group_request(ac);
3061		return ;
3062	}
3063
3064	bsbits = ac->ac_sb->s_blocksize_bits;
3065
3066	/* first, let's learn actual file size
3067	 * given current request is allocated */
3068	size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
3069	size = size << bsbits;
3070	if (size < i_size_read(ac->ac_inode))
3071		size = i_size_read(ac->ac_inode);
3072	orig_size = size;
3073
3074	/* max size of free chunks */
3075	max = 2 << bsbits;
3076
3077#define NRL_CHECK_SIZE(req, size, max, chunk_size)	\
3078		(req <= (size) || max <= (chunk_size))
3079
3080	/* first, try to predict filesize */
3081	/* XXX: should this table be tunable? */
3082	start_off = 0;
3083	if (size <= 16 * 1024) {
3084		size = 16 * 1024;
3085	} else if (size <= 32 * 1024) {
3086		size = 32 * 1024;
3087	} else if (size <= 64 * 1024) {
3088		size = 64 * 1024;
3089	} else if (size <= 128 * 1024) {
3090		size = 128 * 1024;
3091	} else if (size <= 256 * 1024) {
3092		size = 256 * 1024;
3093	} else if (size <= 512 * 1024) {
3094		size = 512 * 1024;
3095	} else if (size <= 1024 * 1024) {
3096		size = 1024 * 1024;
3097	} else if (NRL_CHECK_SIZE(size, 4 * 1024 * 1024, max, 2 * 1024)) {
3098		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3099						(21 - bsbits)) << 21;
3100		size = 2 * 1024 * 1024;
3101	} else if (NRL_CHECK_SIZE(size, 8 * 1024 * 1024, max, 4 * 1024)) {
3102		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3103							(22 - bsbits)) << 22;
3104		size = 4 * 1024 * 1024;
3105	} else if (NRL_CHECK_SIZE(ac->ac_o_ex.fe_len,
3106					(8<<20)>>bsbits, max, 8 * 1024)) {
3107		start_off = ((loff_t)ac->ac_o_ex.fe_logical >>
3108							(23 - bsbits)) << 23;
3109		size = 8 * 1024 * 1024;
3110	} else {
3111		start_off = (loff_t) ac->ac_o_ex.fe_logical << bsbits;
3112		size	  = (loff_t) EXT4_C2B(EXT4_SB(ac->ac_sb),
3113					      ac->ac_o_ex.fe_len) << bsbits;
3114	}
3115	size = size >> bsbits;
3116	start = start_off >> bsbits;
3117
3118	/* don't cover already allocated blocks in selected range */
3119	if (ar->pleft && start <= ar->lleft) {
3120		size -= ar->lleft + 1 - start;
3121		start = ar->lleft + 1;
3122	}
3123	if (ar->pright && start + size - 1 >= ar->lright)
3124		size -= start + size - ar->lright;
3125
3126	/*
3127	 * Trim allocation request for filesystems with artificially small
3128	 * groups.
3129	 */
3130	if (size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb))
3131		size = EXT4_BLOCKS_PER_GROUP(ac->ac_sb);
3132
3133	end = start + size;
3134
3135	/* check we don't cross already preallocated blocks */
3136	rcu_read_lock();
3137	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3138		ext4_lblk_t pa_end;
3139
3140		if (pa->pa_deleted)
3141			continue;
3142		spin_lock(&pa->pa_lock);
3143		if (pa->pa_deleted) {
3144			spin_unlock(&pa->pa_lock);
3145			continue;
3146		}
3147
3148		pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3149						  pa->pa_len);
3150
3151		/* PA must not overlap original request */
3152		BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
3153			ac->ac_o_ex.fe_logical < pa->pa_lstart));
3154
3155		/* skip PAs this normalized request doesn't overlap with */
3156		if (pa->pa_lstart >= end || pa_end <= start) {
3157			spin_unlock(&pa->pa_lock);
3158			continue;
3159		}
3160		BUG_ON(pa->pa_lstart <= start && pa_end >= end);
3161
3162		/* adjust start or end to be adjacent to this pa */
3163		if (pa_end <= ac->ac_o_ex.fe_logical) {
3164			BUG_ON(pa_end < start);
3165			start = pa_end;
3166		} else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
3167			BUG_ON(pa->pa_lstart > end);
3168			end = pa->pa_lstart;
3169		}
3170		spin_unlock(&pa->pa_lock);
3171	}
3172	rcu_read_unlock();
3173	size = end - start;
3174
3175	/* XXX: extra loop to check we really don't overlap preallocations */
3176	rcu_read_lock();
3177	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3178		ext4_lblk_t pa_end;
3179
3180		spin_lock(&pa->pa_lock);
3181		if (pa->pa_deleted == 0) {
3182			pa_end = pa->pa_lstart + EXT4_C2B(EXT4_SB(ac->ac_sb),
3183							  pa->pa_len);
3184			BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
3185		}
3186		spin_unlock(&pa->pa_lock);
3187	}
3188	rcu_read_unlock();
3189
3190	if (start + size <= ac->ac_o_ex.fe_logical &&
3191			start > ac->ac_o_ex.fe_logical) {
3192		ext4_msg(ac->ac_sb, KERN_ERR,
3193			 "start %lu, size %lu, fe_logical %lu",
3194			 (unsigned long) start, (unsigned long) size,
3195			 (unsigned long) ac->ac_o_ex.fe_logical);
3196		BUG();
3197	}
3198	BUG_ON(size <= 0 || size > EXT4_BLOCKS_PER_GROUP(ac->ac_sb));
3199
3200	/* now prepare goal request */
3201
3202	/* XXX: is it better to align blocks WRT to logical
3203	 * placement or satisfy big request as is */
3204	ac->ac_g_ex.fe_logical = start;
3205	ac->ac_g_ex.fe_len = EXT4_NUM_B2C(sbi, size);
3206
3207	/* define goal start in order to merge */
3208	if (ar->pright && (ar->lright == (start + size))) {
3209		/* merge to the right */
3210		ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size,
3211						&ac->ac_f_ex.fe_group,
3212						&ac->ac_f_ex.fe_start);
3213		ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3214	}
3215	if (ar->pleft && (ar->lleft + 1 == start)) {
3216		/* merge to the left */
3217		ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1,
3218						&ac->ac_f_ex.fe_group,
3219						&ac->ac_f_ex.fe_start);
3220		ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL;
3221	}
3222
3223	mb_debug(1, "goal: %u(was %u) blocks at %u\n", (unsigned) size,
3224		(unsigned) orig_size, (unsigned) start);
3225}
3226
3227static void ext4_mb_collect_stats(struct ext4_allocation_context *ac)
3228{
3229	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3230
3231	if (sbi->s_mb_stats && ac->ac_g_ex.fe_len > 1) {
3232		atomic_inc(&sbi->s_bal_reqs);
3233		atomic_add(ac->ac_b_ex.fe_len, &sbi->s_bal_allocated);
3234		if (ac->ac_b_ex.fe_len >= ac->ac_o_ex.fe_len)
3235			atomic_inc(&sbi->s_bal_success);
3236		atomic_add(ac->ac_found, &sbi->s_bal_ex_scanned);
3237		if (ac->ac_g_ex.fe_start == ac->ac_b_ex.fe_start &&
3238				ac->ac_g_ex.fe_group == ac->ac_b_ex.fe_group)
3239			atomic_inc(&sbi->s_bal_goals);
3240		if (ac->ac_found > sbi->s_mb_max_to_scan)
3241			atomic_inc(&sbi->s_bal_breaks);
3242	}
3243
3244	if (ac->ac_op == EXT4_MB_HISTORY_ALLOC)
3245		trace_ext4_mballoc_alloc(ac);
3246	else
3247		trace_ext4_mballoc_prealloc(ac);
3248}
3249
3250/*
3251 * Called on failure; free up any blocks from the inode PA for this
3252 * context.  We don't need this for MB_GROUP_PA because we only change
3253 * pa_free in ext4_mb_release_context(), but on failure, we've already
3254 * zeroed out ac->ac_b_ex.fe_len, so group_pa->pa_free is not changed.
3255 */
3256static void ext4_discard_allocated_blocks(struct ext4_allocation_context *ac)
3257{
3258	struct ext4_prealloc_space *pa = ac->ac_pa;
3259	struct ext4_buddy e4b;
3260	int err;
3261
3262	if (pa == NULL) {
3263		if (ac->ac_f_ex.fe_len == 0)
3264			return;
3265		err = ext4_mb_load_buddy(ac->ac_sb, ac->ac_f_ex.fe_group, &e4b);
3266		if (err) {
3267			/*
3268			 * This should never happen since we pin the
3269			 * pages in the ext4_allocation_context so
3270			 * ext4_mb_load_buddy() should never fail.
3271			 */
3272			WARN(1, "mb_load_buddy failed (%d)", err);
3273			return;
3274		}
3275		ext4_lock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3276		mb_free_blocks(ac->ac_inode, &e4b, ac->ac_f_ex.fe_start,
3277			       ac->ac_f_ex.fe_len);
3278		ext4_unlock_group(ac->ac_sb, ac->ac_f_ex.fe_group);
3279		ext4_mb_unload_buddy(&e4b);
3280		return;
3281	}
3282	if (pa->pa_type == MB_INODE_PA)
3283		pa->pa_free += ac->ac_b_ex.fe_len;
3284}
3285
3286/*
3287 * use blocks preallocated to inode
3288 */
3289static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
3290				struct ext4_prealloc_space *pa)
3291{
3292	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3293	ext4_fsblk_t start;
3294	ext4_fsblk_t end;
3295	int len;
3296
3297	/* found preallocated blocks, use them */
3298	start = pa->pa_pstart + (ac->ac_o_ex.fe_logical - pa->pa_lstart);
3299	end = min(pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len),
3300		  start + EXT4_C2B(sbi, ac->ac_o_ex.fe_len));
3301	len = EXT4_NUM_B2C(sbi, end - start);
3302	ext4_get_group_no_and_offset(ac->ac_sb, start, &ac->ac_b_ex.fe_group,
3303					&ac->ac_b_ex.fe_start);
3304	ac->ac_b_ex.fe_len = len;
3305	ac->ac_status = AC_STATUS_FOUND;
3306	ac->ac_pa = pa;
3307
3308	BUG_ON(start < pa->pa_pstart);
3309	BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
3310	BUG_ON(pa->pa_free < len);
3311	pa->pa_free -= len;
3312
3313	mb_debug(1, "use %llu/%u from inode pa %p\n", start, len, pa);
3314}
3315
3316/*
3317 * use blocks preallocated to locality group
3318 */
3319static void ext4_mb_use_group_pa(struct ext4_allocation_context *ac,
3320				struct ext4_prealloc_space *pa)
3321{
3322	unsigned int len = ac->ac_o_ex.fe_len;
3323
3324	ext4_get_group_no_and_offset(ac->ac_sb, pa->pa_pstart,
3325					&ac->ac_b_ex.fe_group,
3326					&ac->ac_b_ex.fe_start);
3327	ac->ac_b_ex.fe_len = len;
3328	ac->ac_status = AC_STATUS_FOUND;
3329	ac->ac_pa = pa;
3330
3331	/* we don't correct pa_pstart or pa_plen here to avoid
3332	 * possible race when the group is being loaded concurrently
3333	 * instead we correct pa later, after blocks are marked
3334	 * in on-disk bitmap -- see ext4_mb_release_context()
3335	 * Other CPUs are prevented from allocating from this pa by lg_mutex
3336	 */
3337	mb_debug(1, "use %u/%u from group pa %p\n", pa->pa_lstart-len, len, pa);
3338}
3339
3340/*
3341 * Return the prealloc space that have minimal distance
3342 * from the goal block. @cpa is the prealloc
3343 * space that is having currently known minimal distance
3344 * from the goal block.
3345 */
3346static struct ext4_prealloc_space *
3347ext4_mb_check_group_pa(ext4_fsblk_t goal_block,
3348			struct ext4_prealloc_space *pa,
3349			struct ext4_prealloc_space *cpa)
3350{
3351	ext4_fsblk_t cur_distance, new_distance;
3352
3353	if (cpa == NULL) {
3354		atomic_inc(&pa->pa_count);
3355		return pa;
3356	}
3357	cur_distance = abs(goal_block - cpa->pa_pstart);
3358	new_distance = abs(goal_block - pa->pa_pstart);
3359
3360	if (cur_distance <= new_distance)
3361		return cpa;
3362
3363	/* drop the previous reference */
3364	atomic_dec(&cpa->pa_count);
3365	atomic_inc(&pa->pa_count);
3366	return pa;
3367}
3368
3369/*
3370 * search goal blocks in preallocated space
3371 */
3372static noinline_for_stack int
3373ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
3374{
3375	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
3376	int order, i;
3377	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
3378	struct ext4_locality_group *lg;
3379	struct ext4_prealloc_space *pa, *cpa = NULL;
3380	ext4_fsblk_t goal_block;
3381
3382	/* only data can be preallocated */
3383	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
3384		return 0;
3385
3386	/* first, try per-file preallocation */
3387	rcu_read_lock();
3388	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
3389
3390		/* all fields in this condition don't change,
3391		 * so we can skip locking for them */
3392		if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
3393		    ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
3394					       EXT4_C2B(sbi, pa->pa_len)))
3395			continue;
3396
3397		/* non-extent files can't have physical blocks past 2^32 */
3398		if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
3399		    (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
3400		     EXT4_MAX_BLOCK_FILE_PHYS))
3401			continue;
3402
3403		/* found preallocated blocks, use them */
3404		spin_lock(&pa->pa_lock);
3405		if (pa->pa_deleted == 0 && pa->pa_free) {
3406			atomic_inc(&pa->pa_count);
3407			ext4_mb_use_inode_pa(ac, pa);
3408			spin_unlock(&pa->pa_lock);
3409			ac->ac_criteria = 10;
3410			rcu_read_unlock();
3411			return 1;
3412		}
3413		spin_unlock(&pa->pa_lock);
3414	}
3415	rcu_read_unlock();
3416
3417	/* can we use group allocation? */
3418	if (!(ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC))
3419		return 0;
3420
3421	/* inode may have no locality group for some reason */
3422	lg = ac->ac_lg;
3423	if (lg == NULL)
3424		return 0;
3425	order  = fls(ac->ac_o_ex.fe_len) - 1;
3426	if (order > PREALLOC_TB_SIZE - 1)
3427		/* The max size of hash table is PREALLOC_TB_SIZE */
3428		order = PREALLOC_TB_SIZE - 1;
3429
3430	goal_block = ext4_grp_offs_to_block(ac->ac_sb, &ac->ac_g_ex);
3431	/*
3432	 * search for the prealloc space that is having
3433	 * minimal distance from the goal block.
3434	 */
3435	for (i = order; i < PREALLOC_TB_SIZE; i++) {
3436		rcu_read_lock();
3437		list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
3438					pa_inode_list) {
3439			spin_lock(&pa->pa_lock);
3440			if (pa->pa_deleted == 0 &&
3441					pa->pa_free >= ac->ac_o_ex.fe_len) {
3442
3443				cpa = ext4_mb_check_group_pa(goal_block,
3444								pa, cpa);
3445			}
3446			spin_unlock(&pa->pa_lock);
3447		}
3448		rcu_read_unlock();
3449	}
3450	if (cpa) {
3451		ext4_mb_use_group_pa(ac, cpa);
3452		ac->ac_criteria = 20;
3453		return 1;
3454	}
3455	return 0;
3456}
3457
3458/*
3459 * the function goes through all block freed in the group
3460 * but not yet committed and marks them used in in-core bitmap.
3461 * buddy must be generated from this bitmap
3462 * Need to be called with the ext4 group lock held
3463 */
3464static void ext4_mb_generate_from_freelist(struct super_block *sb, void *bitmap,
3465						ext4_group_t group)
3466{
3467	struct rb_node *n;
3468	struct ext4_group_info *grp;
3469	struct ext4_free_data *entry;
3470
3471	grp = ext4_get_group_info(sb, group);
3472	n = rb_first(&(grp->bb_free_root));
3473
3474	while (n) {
3475		entry = rb_entry(n, struct ext4_free_data, efd_node);
3476		ext4_set_bits(bitmap, entry->efd_start_cluster, entry->efd_count);
3477		n = rb_next(n);
3478	}
3479	return;
3480}
3481
3482/*
3483 * the function goes through all preallocation in this group and marks them
3484 * used in in-core bitmap. buddy must be generated from this bitmap
3485 * Need to be called with ext4 group lock held
3486 */
3487static noinline_for_stack
3488void ext4_mb_generate_from_pa(struct super_block *sb, void *bitmap,
3489					ext4_group_t group)
3490{
3491	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3492	struct ext4_prealloc_space *pa;
3493	struct list_head *cur;
3494	ext4_group_t groupnr;
3495	ext4_grpblk_t start;
3496	int preallocated = 0;
3497	int len;
3498
3499	/* all form of preallocation discards first load group,
3500	 * so the only competing code is preallocation use.
3501	 * we don't need any locking here
3502	 * notice we do NOT ignore preallocations with pa_deleted
3503	 * otherwise we could leave used blocks available for
3504	 * allocation in buddy when concurrent ext4_mb_put_pa()
3505	 * is dropping preallocation
3506	 */
3507	list_for_each(cur, &grp->bb_prealloc_list) {
3508		pa = list_entry(cur, struct ext4_prealloc_space, pa_group_list);
3509		spin_lock(&pa->pa_lock);
3510		ext4_get_group_no_and_offset(sb, pa->pa_pstart,
3511					     &groupnr, &start);
3512		len = pa->pa_len;
3513		spin_unlock(&pa->pa_lock);
3514		if (unlikely(len == 0))
3515			continue;
3516		BUG_ON(groupnr != group);
3517		ext4_set_bits(bitmap, start, len);
3518		preallocated += len;
3519	}
3520	mb_debug(1, "prellocated %u for group %u\n", preallocated, group);
3521}
3522
3523static void ext4_mb_pa_callback(struct rcu_head *head)
3524{
3525	struct ext4_prealloc_space *pa;
3526	pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
3527
3528	BUG_ON(atomic_read(&pa->pa_count));
3529	BUG_ON(pa->pa_deleted == 0);
3530	kmem_cache_free(ext4_pspace_cachep, pa);
3531}
3532
3533/*
3534 * drops a reference to preallocated space descriptor
3535 * if this was the last reference and the space is consumed
3536 */
3537static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
3538			struct super_block *sb, struct ext4_prealloc_space *pa)
3539{
3540	ext4_group_t grp;
3541	ext4_fsblk_t grp_blk;
3542
3543	/* in this short window concurrent discard can set pa_deleted */
3544	spin_lock(&pa->pa_lock);
3545	if (!atomic_dec_and_test(&pa->pa_count) || pa->pa_free != 0) {
3546		spin_unlock(&pa->pa_lock);
3547		return;
3548	}
3549
3550	if (pa->pa_deleted == 1) {
3551		spin_unlock(&pa->pa_lock);
3552		return;
3553	}
3554
3555	pa->pa_deleted = 1;
3556	spin_unlock(&pa->pa_lock);
3557
3558	grp_blk = pa->pa_pstart;
3559	/*
3560	 * If doing group-based preallocation, pa_pstart may be in the
3561	 * next group when pa is used up
3562	 */
3563	if (pa->pa_type == MB_GROUP_PA)
3564		grp_blk--;
3565
3566	grp = ext4_get_group_number(sb, grp_blk);
3567
3568	/*
3569	 * possible race:
3570	 *
3571	 *  P1 (buddy init)			P2 (regular allocation)
3572	 *					find block B in PA
3573	 *  copy on-disk bitmap to buddy
3574	 *  					mark B in on-disk bitmap
3575	 *					drop PA from group
3576	 *  mark all PAs in buddy
3577	 *
3578	 * thus, P1 initializes buddy with B available. to prevent this
3579	 * we make "copy" and "mark all PAs" atomic and serialize "drop PA"
3580	 * against that pair
3581	 */
3582	ext4_lock_group(sb, grp);
3583	list_del(&pa->pa_group_list);
3584	ext4_unlock_group(sb, grp);
3585
3586	spin_lock(pa->pa_obj_lock);
3587	list_del_rcu(&pa->pa_inode_list);
3588	spin_unlock(pa->pa_obj_lock);
3589
3590	call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3591}
3592
3593/*
3594 * creates new preallocated space for given inode
3595 */
3596static noinline_for_stack int
3597ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
3598{
3599	struct super_block *sb = ac->ac_sb;
3600	struct ext4_sb_info *sbi = EXT4_SB(sb);
3601	struct ext4_prealloc_space *pa;
3602	struct ext4_group_info *grp;
3603	struct ext4_inode_info *ei;
3604
3605	/* preallocate only when found space is larger then requested */
3606	BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3607	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3608	BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3609
3610	pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3611	if (pa == NULL)
3612		return -ENOMEM;
3613
3614	if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
3615		int winl;
3616		int wins;
3617		int win;
3618		int offs;
3619
3620		/* we can't allocate as much as normalizer wants.
3621		 * so, found space must get proper lstart
3622		 * to cover original request */
3623		BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
3624		BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
3625
3626		/* we're limited by original request in that
3627		 * logical block must be covered any way
3628		 * winl is window we can move our chunk within */
3629		winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
3630
3631		/* also, we should cover whole original request */
3632		wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
3633
3634		/* the smallest one defines real window */
3635		win = min(winl, wins);
3636
3637		offs = ac->ac_o_ex.fe_logical %
3638			EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
3639		if (offs && offs < win)
3640			win = offs;
3641
3642		ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
3643			EXT4_NUM_B2C(sbi, win);
3644		BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
3645		BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
3646	}
3647
3648	/* preallocation can change ac_b_ex, thus we store actually
3649	 * allocated blocks for history */
3650	ac->ac_f_ex = ac->ac_b_ex;
3651
3652	pa->pa_lstart = ac->ac_b_ex.fe_logical;
3653	pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3654	pa->pa_len = ac->ac_b_ex.fe_len;
3655	pa->pa_free = pa->pa_len;
3656	atomic_set(&pa->pa_count, 1);
3657	spin_lock_init(&pa->pa_lock);
3658	INIT_LIST_HEAD(&pa->pa_inode_list);
3659	INIT_LIST_HEAD(&pa->pa_group_list);
3660	pa->pa_deleted = 0;
3661	pa->pa_type = MB_INODE_PA;
3662
3663	mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
3664			pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3665	trace_ext4_mb_new_inode_pa(ac, pa);
3666
3667	ext4_mb_use_inode_pa(ac, pa);
3668	atomic_add(pa->pa_free, &sbi->s_mb_preallocated);
3669
3670	ei = EXT4_I(ac->ac_inode);
3671	grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3672
3673	pa->pa_obj_lock = &ei->i_prealloc_lock;
3674	pa->pa_inode = ac->ac_inode;
3675
3676	ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3677	list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3678	ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3679
3680	spin_lock(pa->pa_obj_lock);
3681	list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
3682	spin_unlock(pa->pa_obj_lock);
3683
3684	return 0;
3685}
3686
3687/*
3688 * creates new preallocated space for locality group inodes belongs to
3689 */
3690static noinline_for_stack int
3691ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
3692{
3693	struct super_block *sb = ac->ac_sb;
3694	struct ext4_locality_group *lg;
3695	struct ext4_prealloc_space *pa;
3696	struct ext4_group_info *grp;
3697
3698	/* preallocate only when found space is larger then requested */
3699	BUG_ON(ac->ac_o_ex.fe_len >= ac->ac_b_ex.fe_len);
3700	BUG_ON(ac->ac_status != AC_STATUS_FOUND);
3701	BUG_ON(!S_ISREG(ac->ac_inode->i_mode));
3702
3703	BUG_ON(ext4_pspace_cachep == NULL);
3704	pa = kmem_cache_alloc(ext4_pspace_cachep, GFP_NOFS);
3705	if (pa == NULL)
3706		return -ENOMEM;
3707
3708	/* preallocation can change ac_b_ex, thus we store actually
3709	 * allocated blocks for history */
3710	ac->ac_f_ex = ac->ac_b_ex;
3711
3712	pa->pa_pstart = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
3713	pa->pa_lstart = pa->pa_pstart;
3714	pa->pa_len = ac->ac_b_ex.fe_len;
3715	pa->pa_free = pa->pa_len;
3716	atomic_set(&pa->pa_count, 1);
3717	spin_lock_init(&pa->pa_lock);
3718	INIT_LIST_HEAD(&pa->pa_inode_list);
3719	INIT_LIST_HEAD(&pa->pa_group_list);
3720	pa->pa_deleted = 0;
3721	pa->pa_type = MB_GROUP_PA;
3722
3723	mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
3724			pa->pa_pstart, pa->pa_len, pa->pa_lstart);
3725	trace_ext4_mb_new_group_pa(ac, pa);
3726
3727	ext4_mb_use_group_pa(ac, pa);
3728	atomic_add(pa->pa_free, &EXT4_SB(sb)->s_mb_preallocated);
3729
3730	grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
3731	lg = ac->ac_lg;
3732	BUG_ON(lg == NULL);
3733
3734	pa->pa_obj_lock = &lg->lg_prealloc_lock;
3735	pa->pa_inode = NULL;
3736
3737	ext4_lock_group(sb, ac->ac_b_ex.fe_group);
3738	list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
3739	ext4_unlock_group(sb, ac->ac_b_ex.fe_group);
3740
3741	/*
3742	 * We will later add the new pa to the right bucket
3743	 * after updating the pa_free in ext4_mb_release_context
3744	 */
3745	return 0;
3746}
3747
3748static int ext4_mb_new_preallocation(struct ext4_allocation_context *ac)
3749{
3750	int err;
3751
3752	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
3753		err = ext4_mb_new_group_pa(ac);
3754	else
3755		err = ext4_mb_new_inode_pa(ac);
3756	return err;
3757}
3758
3759/*
3760 * finds all unused blocks in on-disk bitmap, frees them in
3761 * in-core bitmap and buddy.
3762 * @pa must be unlinked from inode and group lists, so that
3763 * nobody else can find/use it.
3764 * the caller MUST hold group/inode locks.
3765 * TODO: optimize the case when there are no in-core structures yet
3766 */
3767static noinline_for_stack int
3768ext4_mb_release_inode_pa(struct ext4_buddy *e4b, struct buffer_head *bitmap_bh,
3769			struct ext4_prealloc_space *pa)
3770{
3771	struct super_block *sb = e4b->bd_sb;
3772	struct ext4_sb_info *sbi = EXT4_SB(sb);
3773	unsigned int end;
3774	unsigned int next;
3775	ext4_group_t group;
3776	ext4_grpblk_t bit;
3777	unsigned long long grp_blk_start;
3778	int err = 0;
3779	int free = 0;
3780
3781	BUG_ON(pa->pa_deleted == 0);
3782	ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3783	grp_blk_start = pa->pa_pstart - EXT4_C2B(sbi, bit);
3784	BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3785	end = bit + pa->pa_len;
3786
3787	while (bit < end) {
3788		bit = mb_find_next_zero_bit(bitmap_bh->b_data, end, bit);
3789		if (bit >= end)
3790			break;
3791		next = mb_find_next_bit(bitmap_bh->b_data, end, bit);
3792		mb_debug(1, "    free preallocated %u/%u in group %u\n",
3793			 (unsigned) ext4_group_first_block_no(sb, group) + bit,
3794			 (unsigned) next - bit, (unsigned) group);
3795		free += next - bit;
3796
3797		trace_ext4_mballoc_discard(sb, NULL, group, bit, next - bit);
3798		trace_ext4_mb_release_inode_pa(pa, (grp_blk_start +
3799						    EXT4_C2B(sbi, bit)),
3800					       next - bit);
3801		mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
3802		bit = next + 1;
3803	}
3804	if (free != pa->pa_free) {
3805		ext4_msg(e4b->bd_sb, KERN_CRIT,
3806			 "pa %p: logic %lu, phys. %lu, len %lu",
3807			 pa, (unsigned long) pa->pa_lstart,
3808			 (unsigned long) pa->pa_pstart,
3809			 (unsigned long) pa->pa_len);
3810		ext4_grp_locked_error(sb, group, 0, 0, "free %u, pa_free %u",
3811					free, pa->pa_free);
3812		/*
3813		 * pa is already deleted so we use the value obtained
3814		 * from the bitmap and continue.
3815		 */
3816	}
3817	atomic_add(free, &sbi->s_mb_discarded);
3818
3819	return err;
3820}
3821
3822static noinline_for_stack int
3823ext4_mb_release_group_pa(struct ext4_buddy *e4b,
3824				struct ext4_prealloc_space *pa)
3825{
3826	struct super_block *sb = e4b->bd_sb;
3827	ext4_group_t group;
3828	ext4_grpblk_t bit;
3829
3830	trace_ext4_mb_release_group_pa(sb, pa);
3831	BUG_ON(pa->pa_deleted == 0);
3832	ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
3833	BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
3834	mb_free_blocks(pa->pa_inode, e4b, bit, pa->pa_len);
3835	atomic_add(pa->pa_len, &EXT4_SB(sb)->s_mb_discarded);
3836	trace_ext4_mballoc_discard(sb, NULL, group, bit, pa->pa_len);
3837
3838	return 0;
3839}
3840
3841/*
3842 * releases all preallocations in given group
3843 *
3844 * first, we need to decide discard policy:
3845 * - when do we discard
3846 *   1) ENOSPC
3847 * - how many do we discard
3848 *   1) how many requested
3849 */
3850static noinline_for_stack int
3851ext4_mb_discard_group_preallocations(struct super_block *sb,
3852					ext4_group_t group, int needed)
3853{
3854	struct ext4_group_info *grp = ext4_get_group_info(sb, group);
3855	struct buffer_head *bitmap_bh = NULL;
3856	struct ext4_prealloc_space *pa, *tmp;
3857	struct list_head list;
3858	struct ext4_buddy e4b;
3859	int err;
3860	int busy = 0;
3861	int free = 0;
3862
3863	mb_debug(1, "discard preallocation for group %u\n", group);
3864
3865	if (list_empty(&grp->bb_prealloc_list))
3866		return 0;
3867
3868	bitmap_bh = ext4_read_block_bitmap(sb, group);
3869	if (IS_ERR(bitmap_bh)) {
3870		err = PTR_ERR(bitmap_bh);
3871		ext4_error(sb, "Error %d reading block bitmap for %u",
3872			   err, group);
3873		return 0;
3874	}
3875
3876	err = ext4_mb_load_buddy(sb, group, &e4b);
3877	if (err) {
3878		ext4_error(sb, "Error loading buddy information for %u", group);
3879		put_bh(bitmap_bh);
3880		return 0;
3881	}
3882
3883	if (needed == 0)
3884		needed = EXT4_CLUSTERS_PER_GROUP(sb) + 1;
3885
3886	INIT_LIST_HEAD(&list);
3887repeat:
3888	ext4_lock_group(sb, group);
3889	list_for_each_entry_safe(pa, tmp,
3890				&grp->bb_prealloc_list, pa_group_list) {
3891		spin_lock(&pa->pa_lock);
3892		if (atomic_read(&pa->pa_count)) {
3893			spin_unlock(&pa->pa_lock);
3894			busy = 1;
3895			continue;
3896		}
3897		if (pa->pa_deleted) {
3898			spin_unlock(&pa->pa_lock);
3899			continue;
3900		}
3901
3902		/* seems this one can be freed ... */
3903		pa->pa_deleted = 1;
3904
3905		/* we can trust pa_free ... */
3906		free += pa->pa_free;
3907
3908		spin_unlock(&pa->pa_lock);
3909
3910		list_del(&pa->pa_group_list);
3911		list_add(&pa->u.pa_tmp_list, &list);
3912	}
3913
3914	/* if we still need more blocks and some PAs were used, try again */
3915	if (free < needed && busy) {
3916		busy = 0;
3917		ext4_unlock_group(sb, group);
3918		cond_resched();
3919		goto repeat;
3920	}
3921
3922	/* found anything to free? */
3923	if (list_empty(&list)) {
3924		BUG_ON(free != 0);
3925		goto out;
3926	}
3927
3928	/* now free all selected PAs */
3929	list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
3930
3931		/* remove from object (inode or locality group) */
3932		spin_lock(pa->pa_obj_lock);
3933		list_del_rcu(&pa->pa_inode_list);
3934		spin_unlock(pa->pa_obj_lock);
3935
3936		if (pa->pa_type == MB_GROUP_PA)
3937			ext4_mb_release_group_pa(&e4b, pa);
3938		else
3939			ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
3940
3941		list_del(&pa->u.pa_tmp_list);
3942		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
3943	}
3944
3945out:
3946	ext4_unlock_group(sb, group);
3947	ext4_mb_unload_buddy(&e4b);
3948	put_bh(bitmap_bh);
3949	return free;
3950}
3951
3952/*
3953 * releases all non-used preallocated blocks for given inode
3954 *
3955 * It's important to discard preallocations under i_data_sem
3956 * We don't want another block to be served from the prealloc
3957 * space when we are discarding the inode prealloc space.
3958 *
3959 * FIXME!! Make sure it is valid at all the call sites
3960 */
3961void ext4_discard_preallocations(struct inode *inode)
3962{
3963	struct ext4_inode_info *ei = EXT4_I(inode);
3964	struct super_block *sb = inode->i_sb;
3965	struct buffer_head *bitmap_bh = NULL;
3966	struct ext4_prealloc_space *pa, *tmp;
3967	ext4_group_t group = 0;
3968	struct list_head list;
3969	struct ext4_buddy e4b;
3970	int err;
3971
3972	if (!S_ISREG(inode->i_mode)) {
3973		/*BUG_ON(!list_empty(&ei->i_prealloc_list));*/
3974		return;
3975	}
3976
3977	mb_debug(1, "discard preallocation for inode %lu\n", inode->i_ino);
3978	trace_ext4_discard_preallocations(inode);
3979
3980	INIT_LIST_HEAD(&list);
3981
3982repeat:
3983	/* first, collect all pa's in the inode */
3984	spin_lock(&ei->i_prealloc_lock);
3985	while (!list_empty(&ei->i_prealloc_list)) {
3986		pa = list_entry(ei->i_prealloc_list.next,
3987				struct ext4_prealloc_space, pa_inode_list);
3988		BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
3989		spin_lock(&pa->pa_lock);
3990		if (atomic_read(&pa->pa_count)) {
3991			/* this shouldn't happen often - nobody should
3992			 * use preallocation while we're discarding it */
3993			spin_unlock(&pa->pa_lock);
3994			spin_unlock(&ei->i_prealloc_lock);
3995			ext4_msg(sb, KERN_ERR,
3996				 "uh-oh! used pa while discarding");
3997			WARN_ON(1);
3998			schedule_timeout_uninterruptible(HZ);
3999			goto repeat;
4000
4001		}
4002		if (pa->pa_deleted == 0) {
4003			pa->pa_deleted = 1;
4004			spin_unlock(&pa->pa_lock);
4005			list_del_rcu(&pa->pa_inode_list);
4006			list_add(&pa->u.pa_tmp_list, &list);
4007			continue;
4008		}
4009
4010		/* someone is deleting pa right now */
4011		spin_unlock(&pa->pa_lock);
4012		spin_unlock(&ei->i_prealloc_lock);
4013
4014		/* we have to wait here because pa_deleted
4015		 * doesn't mean pa is already unlinked from
4016		 * the list. as we might be called from
4017		 * ->clear_inode() the inode will get freed
4018		 * and concurrent thread which is unlinking
4019		 * pa from inode's list may access already
4020		 * freed memory, bad-bad-bad */
4021
4022		/* XXX: if this happens too often, we can
4023		 * add a flag to force wait only in case
4024		 * of ->clear_inode(), but not in case of
4025		 * regular truncate */
4026		schedule_timeout_uninterruptible(HZ);
4027		goto repeat;
4028	}
4029	spin_unlock(&ei->i_prealloc_lock);
4030
4031	list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
4032		BUG_ON(pa->pa_type != MB_INODE_PA);
4033		group = ext4_get_group_number(sb, pa->pa_pstart);
4034
4035		err = ext4_mb_load_buddy(sb, group, &e4b);
4036		if (err) {
4037			ext4_error(sb, "Error loading buddy information for %u",
4038					group);
4039			continue;
4040		}
4041
4042		bitmap_bh = ext4_read_block_bitmap(sb, group);
4043		if (IS_ERR(bitmap_bh)) {
4044			err = PTR_ERR(bitmap_bh);
4045			ext4_error(sb, "Error %d reading block bitmap for %u",
4046					err, group);
4047			ext4_mb_unload_buddy(&e4b);
4048			continue;
4049		}
4050
4051		ext4_lock_group(sb, group);
4052		list_del(&pa->pa_group_list);
4053		ext4_mb_release_inode_pa(&e4b, bitmap_bh, pa);
4054		ext4_unlock_group(sb, group);
4055
4056		ext4_mb_unload_buddy(&e4b);
4057		put_bh(bitmap_bh);
4058
4059		list_del(&pa->u.pa_tmp_list);
4060		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4061	}
4062}
4063
4064#ifdef CONFIG_EXT4_DEBUG
4065static void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4066{
4067	struct super_block *sb = ac->ac_sb;
4068	ext4_group_t ngroups, i;
4069
4070	if (!ext4_mballoc_debug ||
4071	    (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED))
4072		return;
4073
4074	ext4_msg(ac->ac_sb, KERN_ERR, "Can't allocate:"
4075			" Allocation context details:");
4076	ext4_msg(ac->ac_sb, KERN_ERR, "status %d flags %d",
4077			ac->ac_status, ac->ac_flags);
4078	ext4_msg(ac->ac_sb, KERN_ERR, "orig %lu/%lu/%lu@%lu, "
4079		 	"goal %lu/%lu/%lu@%lu, "
4080			"best %lu/%lu/%lu@%lu cr %d",
4081			(unsigned long)ac->ac_o_ex.fe_group,
4082			(unsigned long)ac->ac_o_ex.fe_start,
4083			(unsigned long)ac->ac_o_ex.fe_len,
4084			(unsigned long)ac->ac_o_ex.fe_logical,
4085			(unsigned long)ac->ac_g_ex.fe_group,
4086			(unsigned long)ac->ac_g_ex.fe_start,
4087			(unsigned long)ac->ac_g_ex.fe_len,
4088			(unsigned long)ac->ac_g_ex.fe_logical,
4089			(unsigned long)ac->ac_b_ex.fe_group,
4090			(unsigned long)ac->ac_b_ex.fe_start,
4091			(unsigned long)ac->ac_b_ex.fe_len,
4092			(unsigned long)ac->ac_b_ex.fe_logical,
4093			(int)ac->ac_criteria);
4094	ext4_msg(ac->ac_sb, KERN_ERR, "%d found", ac->ac_found);
4095	ext4_msg(ac->ac_sb, KERN_ERR, "groups: ");
4096	ngroups = ext4_get_groups_count(sb);
4097	for (i = 0; i < ngroups; i++) {
4098		struct ext4_group_info *grp = ext4_get_group_info(sb, i);
4099		struct ext4_prealloc_space *pa;
4100		ext4_grpblk_t start;
4101		struct list_head *cur;
4102		ext4_lock_group(sb, i);
4103		list_for_each(cur, &grp->bb_prealloc_list) {
4104			pa = list_entry(cur, struct ext4_prealloc_space,
4105					pa_group_list);
4106			spin_lock(&pa->pa_lock);
4107			ext4_get_group_no_and_offset(sb, pa->pa_pstart,
4108						     NULL, &start);
4109			spin_unlock(&pa->pa_lock);
4110			printk(KERN_ERR "PA:%u:%d:%u \n", i,
4111			       start, pa->pa_len);
4112		}
4113		ext4_unlock_group(sb, i);
4114
4115		if (grp->bb_free == 0)
4116			continue;
4117		printk(KERN_ERR "%u: %d/%d \n",
4118		       i, grp->bb_free, grp->bb_fragments);
4119	}
4120	printk(KERN_ERR "\n");
4121}
4122#else
4123static inline void ext4_mb_show_ac(struct ext4_allocation_context *ac)
4124{
4125	return;
4126}
4127#endif
4128
4129/*
4130 * We use locality group preallocation for small size file. The size of the
4131 * file is determined by the current size or the resulting size after
4132 * allocation which ever is larger
4133 *
4134 * One can tune this size via /sys/fs/ext4/<partition>/mb_stream_req
4135 */
4136static void ext4_mb_group_or_file(struct ext4_allocation_context *ac)
4137{
4138	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4139	int bsbits = ac->ac_sb->s_blocksize_bits;
4140	loff_t size, isize;
4141
4142	if (!(ac->ac_flags & EXT4_MB_HINT_DATA))
4143		return;
4144
4145	if (unlikely(ac->ac_flags & EXT4_MB_HINT_GOAL_ONLY))
4146		return;
4147
4148	size = ac->ac_o_ex.fe_logical + EXT4_C2B(sbi, ac->ac_o_ex.fe_len);
4149	isize = (i_size_read(ac->ac_inode) + ac->ac_sb->s_blocksize - 1)
4150		>> bsbits;
4151
4152	if ((size == isize) &&
4153	    !ext4_fs_is_busy(sbi) &&
4154	    (atomic_read(&ac->ac_inode->i_writecount) == 0)) {
4155		ac->ac_flags |= EXT4_MB_HINT_NOPREALLOC;
4156		return;
4157	}
4158
4159	if (sbi->s_mb_group_prealloc <= 0) {
4160		ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4161		return;
4162	}
4163
4164	/* don't use group allocation for large files */
4165	size = max(size, isize);
4166	if (size > sbi->s_mb_stream_request) {
4167		ac->ac_flags |= EXT4_MB_STREAM_ALLOC;
4168		return;
4169	}
4170
4171	BUG_ON(ac->ac_lg != NULL);
4172	/*
4173	 * locality group prealloc space are per cpu. The reason for having
4174	 * per cpu locality group is to reduce the contention between block
4175	 * request from multiple CPUs.
4176	 */
4177	ac->ac_lg = raw_cpu_ptr(sbi->s_locality_groups);
4178
4179	/* we're going to use group allocation */
4180	ac->ac_flags |= EXT4_MB_HINT_GROUP_ALLOC;
4181
4182	/* serialize all allocations in the group */
4183	mutex_lock(&ac->ac_lg->lg_mutex);
4184}
4185
4186static noinline_for_stack int
4187ext4_mb_initialize_context(struct ext4_allocation_context *ac,
4188				struct ext4_allocation_request *ar)
4189{
4190	struct super_block *sb = ar->inode->i_sb;
4191	struct ext4_sb_info *sbi = EXT4_SB(sb);
4192	struct ext4_super_block *es = sbi->s_es;
4193	ext4_group_t group;
4194	unsigned int len;
4195	ext4_fsblk_t goal;
4196	ext4_grpblk_t block;
4197
4198	/* we can't allocate > group size */
4199	len = ar->len;
4200
4201	/* just a dirty hack to filter too big requests  */
4202	if (len >= EXT4_CLUSTERS_PER_GROUP(sb))
4203		len = EXT4_CLUSTERS_PER_GROUP(sb);
4204
4205	/* start searching from the goal */
4206	goal = ar->goal;
4207	if (goal < le32_to_cpu(es->s_first_data_block) ||
4208			goal >= ext4_blocks_count(es))
4209		goal = le32_to_cpu(es->s_first_data_block);
4210	ext4_get_group_no_and_offset(sb, goal, &group, &block);
4211
4212	/* set up allocation goals */
4213	ac->ac_b_ex.fe_logical = EXT4_LBLK_CMASK(sbi, ar->logical);
4214	ac->ac_status = AC_STATUS_CONTINUE;
4215	ac->ac_sb = sb;
4216	ac->ac_inode = ar->inode;
4217	ac->ac_o_ex.fe_logical = ac->ac_b_ex.fe_logical;
4218	ac->ac_o_ex.fe_group = group;
4219	ac->ac_o_ex.fe_start = block;
4220	ac->ac_o_ex.fe_len = len;
4221	ac->ac_g_ex = ac->ac_o_ex;
4222	ac->ac_flags = ar->flags;
4223
4224	/* we have to define context: we'll we work with a file or
4225	 * locality group. this is a policy, actually */
4226	ext4_mb_group_or_file(ac);
4227
4228	mb_debug(1, "init ac: %u blocks @ %u, goal %u, flags %x, 2^%d, "
4229			"left: %u/%u, right %u/%u to %swritable\n",
4230			(unsigned) ar->len, (unsigned) ar->logical,
4231			(unsigned) ar->goal, ac->ac_flags, ac->ac_2order,
4232			(unsigned) ar->lleft, (unsigned) ar->pleft,
4233			(unsigned) ar->lright, (unsigned) ar->pright,
4234			atomic_read(&ar->inode->i_writecount) ? "" : "non-");
4235	return 0;
4236
4237}
4238
4239static noinline_for_stack void
4240ext4_mb_discard_lg_preallocations(struct super_block *sb,
4241					struct ext4_locality_group *lg,
4242					int order, int total_entries)
4243{
4244	ext4_group_t group = 0;
4245	struct ext4_buddy e4b;
4246	struct list_head discard_list;
4247	struct ext4_prealloc_space *pa, *tmp;
4248
4249	mb_debug(1, "discard locality group preallocation\n");
4250
4251	INIT_LIST_HEAD(&discard_list);
4252
4253	spin_lock(&lg->lg_prealloc_lock);
4254	list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
4255						pa_inode_list) {
4256		spin_lock(&pa->pa_lock);
4257		if (atomic_read(&pa->pa_count)) {
4258			/*
4259			 * This is the pa that we just used
4260			 * for block allocation. So don't
4261			 * free that
4262			 */
4263			spin_unlock(&pa->pa_lock);
4264			continue;
4265		}
4266		if (pa->pa_deleted) {
4267			spin_unlock(&pa->pa_lock);
4268			continue;
4269		}
4270		/* only lg prealloc space */
4271		BUG_ON(pa->pa_type != MB_GROUP_PA);
4272
4273		/* seems this one can be freed ... */
4274		pa->pa_deleted = 1;
4275		spin_unlock(&pa->pa_lock);
4276
4277		list_del_rcu(&pa->pa_inode_list);
4278		list_add(&pa->u.pa_tmp_list, &discard_list);
4279
4280		total_entries--;
4281		if (total_entries <= 5) {
4282			/*
4283			 * we want to keep only 5 entries
4284			 * allowing it to grow to 8. This
4285			 * mak sure we don't call discard
4286			 * soon for this list.
4287			 */
4288			break;
4289		}
4290	}
4291	spin_unlock(&lg->lg_prealloc_lock);
4292
4293	list_for_each_entry_safe(pa, tmp, &discard_list, u.pa_tmp_list) {
4294
4295		group = ext4_get_group_number(sb, pa->pa_pstart);
4296		if (ext4_mb_load_buddy(sb, group, &e4b)) {
4297			ext4_error(sb, "Error loading buddy information for %u",
4298					group);
4299			continue;
4300		}
4301		ext4_lock_group(sb, group);
4302		list_del(&pa->pa_group_list);
4303		ext4_mb_release_group_pa(&e4b, pa);
4304		ext4_unlock_group(sb, group);
4305
4306		ext4_mb_unload_buddy(&e4b);
4307		list_del(&pa->u.pa_tmp_list);
4308		call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
4309	}
4310}
4311
4312/*
4313 * We have incremented pa_count. So it cannot be freed at this
4314 * point. Also we hold lg_mutex. So no parallel allocation is
4315 * possible from this lg. That means pa_free cannot be updated.
4316 *
4317 * A parallel ext4_mb_discard_group_preallocations is possible.
4318 * which can cause the lg_prealloc_list to be updated.
4319 */
4320
4321static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
4322{
4323	int order, added = 0, lg_prealloc_count = 1;
4324	struct super_block *sb = ac->ac_sb;
4325	struct ext4_locality_group *lg = ac->ac_lg;
4326	struct ext4_prealloc_space *tmp_pa, *pa = ac->ac_pa;
4327
4328	order = fls(pa->pa_free) - 1;
4329	if (order > PREALLOC_TB_SIZE - 1)
4330		/* The max size of hash table is PREALLOC_TB_SIZE */
4331		order = PREALLOC_TB_SIZE - 1;
4332	/* Add the prealloc space to lg */
4333	spin_lock(&lg->lg_prealloc_lock);
4334	list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
4335						pa_inode_list) {
4336		spin_lock(&tmp_pa->pa_lock);
4337		if (tmp_pa->pa_deleted) {
4338			spin_unlock(&tmp_pa->pa_lock);
4339			continue;
4340		}
4341		if (!added && pa->pa_free < tmp_pa->pa_free) {
4342			/* Add to the tail of the previous entry */
4343			list_add_tail_rcu(&pa->pa_inode_list,
4344						&tmp_pa->pa_inode_list);
4345			added = 1;
4346			/*
4347			 * we want to count the total
4348			 * number of entries in the list
4349			 */
4350		}
4351		spin_unlock(&tmp_pa->pa_lock);
4352		lg_prealloc_count++;
4353	}
4354	if (!added)
4355		list_add_tail_rcu(&pa->pa_inode_list,
4356					&lg->lg_prealloc_list[order]);
4357	spin_unlock(&lg->lg_prealloc_lock);
4358
4359	/* Now trim the list to be not more than 8 elements */
4360	if (lg_prealloc_count > 8) {
4361		ext4_mb_discard_lg_preallocations(sb, lg,
4362						  order, lg_prealloc_count);
4363		return;
4364	}
4365	return ;
4366}
4367
4368/*
4369 * release all resource we used in allocation
4370 */
4371static int ext4_mb_release_context(struct ext4_allocation_context *ac)
4372{
4373	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
4374	struct ext4_prealloc_space *pa = ac->ac_pa;
4375	if (pa) {
4376		if (pa->pa_type == MB_GROUP_PA) {
4377			/* see comment in ext4_mb_use_group_pa() */
4378			spin_lock(&pa->pa_lock);
4379			pa->pa_pstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4380			pa->pa_lstart += EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
4381			pa->pa_free -= ac->ac_b_ex.fe_len;
4382			pa->pa_len -= ac->ac_b_ex.fe_len;
4383			spin_unlock(&pa->pa_lock);
4384		}
4385	}
4386	if (pa) {
4387		/*
4388		 * We want to add the pa to the right bucket.
4389		 * Remove it from the list and while adding
4390		 * make sure the list to which we are adding
4391		 * doesn't grow big.
4392		 */
4393		if ((pa->pa_type == MB_GROUP_PA) && likely(pa->pa_free)) {
4394			spin_lock(pa->pa_obj_lock);
4395			list_del_rcu(&pa->pa_inode_list);
4396			spin_unlock(pa->pa_obj_lock);
4397			ext4_mb_add_n_trim(ac);
4398		}
4399		ext4_mb_put_pa(ac, ac->ac_sb, pa);
4400	}
4401	if (ac->ac_bitmap_page)
4402		put_page(ac->ac_bitmap_page);
4403	if (ac->ac_buddy_page)
4404		put_page(ac->ac_buddy_page);
4405	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
4406		mutex_unlock(&ac->ac_lg->lg_mutex);
4407	ext4_mb_collect_stats(ac);
4408	return 0;
4409}
4410
4411static int ext4_mb_discard_preallocations(struct super_block *sb, int needed)
4412{
4413	ext4_group_t i, ngroups = ext4_get_groups_count(sb);
4414	int ret;
4415	int freed = 0;
4416
4417	trace_ext4_mb_discard_preallocations(sb, needed);
4418	for (i = 0; i < ngroups && needed > 0; i++) {
4419		ret = ext4_mb_discard_group_preallocations(sb, i, needed);
4420		freed += ret;
4421		needed -= ret;
4422	}
4423
4424	return freed;
4425}
4426
4427/*
4428 * Main entry point into mballoc to allocate blocks
4429 * it tries to use preallocation first, then falls back
4430 * to usual allocation
4431 */
4432ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
4433				struct ext4_allocation_request *ar, int *errp)
4434{
4435	int freed;
4436	struct ext4_allocation_context *ac = NULL;
4437	struct ext4_sb_info *sbi;
4438	struct super_block *sb;
4439	ext4_fsblk_t block = 0;
4440	unsigned int inquota = 0;
4441	unsigned int reserv_clstrs = 0;
4442
4443	might_sleep();
4444	sb = ar->inode->i_sb;
4445	sbi = EXT4_SB(sb);
4446
4447	trace_ext4_request_blocks(ar);
4448
4449	/* Allow to use superuser reservation for quota file */
4450	if (IS_NOQUOTA(ar->inode))
4451		ar->flags |= EXT4_MB_USE_ROOT_BLOCKS;
4452
4453	if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0) {
4454		/* Without delayed allocation we need to verify
4455		 * there is enough free blocks to do block allocation
4456		 * and verify allocation doesn't exceed the quota limits.
4457		 */
4458		while (ar->len &&
4459			ext4_claim_free_clusters(sbi, ar->len, ar->flags)) {
4460
4461			/* let others to free the space */
4462			cond_resched();
4463			ar->len = ar->len >> 1;
4464		}
4465		if (!ar->len) {
4466			*errp = -ENOSPC;
4467			return 0;
4468		}
4469		reserv_clstrs = ar->len;
4470		if (ar->flags & EXT4_MB_USE_ROOT_BLOCKS) {
4471			dquot_alloc_block_nofail(ar->inode,
4472						 EXT4_C2B(sbi, ar->len));
4473		} else {
4474			while (ar->len &&
4475				dquot_alloc_block(ar->inode,
4476						  EXT4_C2B(sbi, ar->len))) {
4477
4478				ar->flags |= EXT4_MB_HINT_NOPREALLOC;
4479				ar->len--;
4480			}
4481		}
4482		inquota = ar->len;
4483		if (ar->len == 0) {
4484			*errp = -EDQUOT;
4485			goto out;
4486		}
4487	}
4488
4489	ac = kmem_cache_zalloc(ext4_ac_cachep, GFP_NOFS);
4490	if (!ac) {
4491		ar->len = 0;
4492		*errp = -ENOMEM;
4493		goto out;
4494	}
4495
4496	*errp = ext4_mb_initialize_context(ac, ar);
4497	if (*errp) {
4498		ar->len = 0;
4499		goto out;
4500	}
4501
4502	ac->ac_op = EXT4_MB_HISTORY_PREALLOC;
4503	if (!ext4_mb_use_preallocated(ac)) {
4504		ac->ac_op = EXT4_MB_HISTORY_ALLOC;
4505		ext4_mb_normalize_request(ac, ar);
4506repeat:
4507		/* allocate space in core */
4508		*errp = ext4_mb_regular_allocator(ac);
4509		if (*errp)
4510			goto discard_and_exit;
4511
4512		/* as we've just preallocated more space than
4513		 * user requested originally, we store allocated
4514		 * space in a special descriptor */
4515		if (ac->ac_status == AC_STATUS_FOUND &&
4516		    ac->ac_o_ex.fe_len < ac->ac_b_ex.fe_len)
4517			*errp = ext4_mb_new_preallocation(ac);
4518		if (*errp) {
4519		discard_and_exit:
4520			ext4_discard_allocated_blocks(ac);
4521			goto errout;
4522		}
4523	}
4524	if (likely(ac->ac_status == AC_STATUS_FOUND)) {
4525		*errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);
4526		if (*errp) {
 
 
 
 
 
 
 
 
 
 
 
4527			ext4_discard_allocated_blocks(ac);
4528			goto errout;
4529		} else {
4530			block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
4531			ar->len = ac->ac_b_ex.fe_len;
4532		}
4533	} else {
4534		freed  = ext4_mb_discard_preallocations(sb, ac->ac_o_ex.fe_len);
4535		if (freed)
4536			goto repeat;
4537		*errp = -ENOSPC;
4538	}
4539
4540errout:
4541	if (*errp) {
4542		ac->ac_b_ex.fe_len = 0;
4543		ar->len = 0;
4544		ext4_mb_show_ac(ac);
4545	}
4546	ext4_mb_release_context(ac);
4547out:
4548	if (ac)
4549		kmem_cache_free(ext4_ac_cachep, ac);
4550	if (inquota && ar->len < inquota)
4551		dquot_free_block(ar->inode, EXT4_C2B(sbi, inquota - ar->len));
4552	if (!ar->len) {
4553		if ((ar->flags & EXT4_MB_DELALLOC_RESERVED) == 0)
4554			/* release all the reserved blocks if non delalloc */
4555			percpu_counter_sub(&sbi->s_dirtyclusters_counter,
4556						reserv_clstrs);
4557	}
4558
4559	trace_ext4_allocate_blocks(ar, (unsigned long long)block);
4560
4561	return block;
4562}
4563
4564/*
4565 * We can merge two free data extents only if the physical blocks
4566 * are contiguous, AND the extents were freed by the same transaction,
4567 * AND the blocks are associated with the same group.
4568 */
4569static int can_merge(struct ext4_free_data *entry1,
4570			struct ext4_free_data *entry2)
4571{
4572	if ((entry1->efd_tid == entry2->efd_tid) &&
4573	    (entry1->efd_group == entry2->efd_group) &&
4574	    ((entry1->efd_start_cluster + entry1->efd_count) == entry2->efd_start_cluster))
4575		return 1;
4576	return 0;
4577}
4578
4579static noinline_for_stack int
4580ext4_mb_free_metadata(handle_t *handle, struct ext4_buddy *e4b,
4581		      struct ext4_free_data *new_entry)
4582{
4583	ext4_group_t group = e4b->bd_group;
4584	ext4_grpblk_t cluster;
4585	ext4_grpblk_t clusters = new_entry->efd_count;
4586	struct ext4_free_data *entry;
4587	struct ext4_group_info *db = e4b->bd_info;
4588	struct super_block *sb = e4b->bd_sb;
4589	struct ext4_sb_info *sbi = EXT4_SB(sb);
4590	struct rb_node **n = &db->bb_free_root.rb_node, *node;
4591	struct rb_node *parent = NULL, *new_node;
4592
4593	BUG_ON(!ext4_handle_valid(handle));
4594	BUG_ON(e4b->bd_bitmap_page == NULL);
4595	BUG_ON(e4b->bd_buddy_page == NULL);
4596
4597	new_node = &new_entry->efd_node;
4598	cluster = new_entry->efd_start_cluster;
4599
4600	if (!*n) {
4601		/* first free block exent. We need to
4602		   protect buddy cache from being freed,
4603		 * otherwise we'll refresh it from
4604		 * on-disk bitmap and lose not-yet-available
4605		 * blocks */
4606		get_page(e4b->bd_buddy_page);
4607		get_page(e4b->bd_bitmap_page);
4608	}
4609	while (*n) {
4610		parent = *n;
4611		entry = rb_entry(parent, struct ext4_free_data, efd_node);
4612		if (cluster < entry->efd_start_cluster)
4613			n = &(*n)->rb_left;
4614		else if (cluster >= (entry->efd_start_cluster + entry->efd_count))
4615			n = &(*n)->rb_right;
4616		else {
4617			ext4_grp_locked_error(sb, group, 0,
4618				ext4_group_first_block_no(sb, group) +
4619				EXT4_C2B(sbi, cluster),
4620				"Block already on to-be-freed list");
4621			return 0;
4622		}
4623	}
4624
4625	rb_link_node(new_node, parent, n);
4626	rb_insert_color(new_node, &db->bb_free_root);
4627
4628	/* Now try to see the extent can be merged to left and right */
4629	node = rb_prev(new_node);
4630	if (node) {
4631		entry = rb_entry(node, struct ext4_free_data, efd_node);
4632		if (can_merge(entry, new_entry) &&
4633		    ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4634			new_entry->efd_start_cluster = entry->efd_start_cluster;
4635			new_entry->efd_count += entry->efd_count;
4636			rb_erase(node, &(db->bb_free_root));
4637			kmem_cache_free(ext4_free_data_cachep, entry);
4638		}
4639	}
4640
4641	node = rb_next(new_node);
4642	if (node) {
4643		entry = rb_entry(node, struct ext4_free_data, efd_node);
4644		if (can_merge(new_entry, entry) &&
4645		    ext4_journal_callback_try_del(handle, &entry->efd_jce)) {
4646			new_entry->efd_count += entry->efd_count;
4647			rb_erase(node, &(db->bb_free_root));
4648			kmem_cache_free(ext4_free_data_cachep, entry);
4649		}
4650	}
4651	/* Add the extent to transaction's private list */
4652	new_entry->efd_jce.jce_func = ext4_free_data_callback;
4653	spin_lock(&sbi->s_md_lock);
4654	_ext4_journal_callback_add(handle, &new_entry->efd_jce);
4655	sbi->s_mb_free_pending += clusters;
4656	spin_unlock(&sbi->s_md_lock);
4657	return 0;
4658}
4659
4660/**
4661 * ext4_free_blocks() -- Free given blocks and update quota
4662 * @handle:		handle for this transaction
4663 * @inode:		inode
4664 * @block:		start physical block to free
4665 * @count:		number of blocks to count
4666 * @flags:		flags used by ext4_free_blocks
4667 */
4668void ext4_free_blocks(handle_t *handle, struct inode *inode,
4669		      struct buffer_head *bh, ext4_fsblk_t block,
4670		      unsigned long count, int flags)
4671{
4672	struct buffer_head *bitmap_bh = NULL;
4673	struct super_block *sb = inode->i_sb;
4674	struct ext4_group_desc *gdp;
4675	unsigned int overflow;
4676	ext4_grpblk_t bit;
4677	struct buffer_head *gd_bh;
4678	ext4_group_t block_group;
4679	struct ext4_sb_info *sbi;
4680	struct ext4_buddy e4b;
4681	unsigned int count_clusters;
4682	int err = 0;
4683	int ret;
4684
4685	might_sleep();
4686	if (bh) {
4687		if (block)
4688			BUG_ON(block != bh->b_blocknr);
4689		else
4690			block = bh->b_blocknr;
4691	}
4692
4693	sbi = EXT4_SB(sb);
4694	if (!(flags & EXT4_FREE_BLOCKS_VALIDATED) &&
4695	    !ext4_data_block_valid(sbi, block, count)) {
4696		ext4_error(sb, "Freeing blocks not in datazone - "
4697			   "block = %llu, count = %lu", block, count);
4698		goto error_return;
4699	}
4700
4701	ext4_debug("freeing block %llu\n", block);
4702	trace_ext4_free_blocks(inode, block, count, flags);
4703
4704	if (bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4705		BUG_ON(count > 1);
4706
4707		ext4_forget(handle, flags & EXT4_FREE_BLOCKS_METADATA,
4708			    inode, bh, block);
4709	}
4710
4711	/*
4712	 * If the extent to be freed does not begin on a cluster
4713	 * boundary, we need to deal with partial clusters at the
4714	 * beginning and end of the extent.  Normally we will free
4715	 * blocks at the beginning or the end unless we are explicitly
4716	 * requested to avoid doing so.
4717	 */
4718	overflow = EXT4_PBLK_COFF(sbi, block);
4719	if (overflow) {
4720		if (flags & EXT4_FREE_BLOCKS_NOFREE_FIRST_CLUSTER) {
4721			overflow = sbi->s_cluster_ratio - overflow;
4722			block += overflow;
4723			if (count > overflow)
4724				count -= overflow;
4725			else
4726				return;
4727		} else {
4728			block -= overflow;
4729			count += overflow;
4730		}
4731	}
4732	overflow = EXT4_LBLK_COFF(sbi, count);
4733	if (overflow) {
4734		if (flags & EXT4_FREE_BLOCKS_NOFREE_LAST_CLUSTER) {
4735			if (count > overflow)
4736				count -= overflow;
4737			else
4738				return;
4739		} else
4740			count += sbi->s_cluster_ratio - overflow;
4741	}
4742
4743	if (!bh && (flags & EXT4_FREE_BLOCKS_FORGET)) {
4744		int i;
4745		int is_metadata = flags & EXT4_FREE_BLOCKS_METADATA;
4746
4747		for (i = 0; i < count; i++) {
4748			cond_resched();
4749			if (is_metadata)
4750				bh = sb_find_get_block(inode->i_sb, block + i);
4751			ext4_forget(handle, is_metadata, inode, bh, block + i);
4752		}
4753	}
4754
4755do_more:
4756	overflow = 0;
4757	ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4758
4759	if (unlikely(EXT4_MB_GRP_BBITMAP_CORRUPT(
4760			ext4_get_group_info(sb, block_group))))
4761		return;
4762
4763	/*
4764	 * Check to see if we are freeing blocks across a group
4765	 * boundary.
4766	 */
4767	if (EXT4_C2B(sbi, bit) + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4768		overflow = EXT4_C2B(sbi, bit) + count -
4769			EXT4_BLOCKS_PER_GROUP(sb);
4770		count -= overflow;
4771	}
4772	count_clusters = EXT4_NUM_B2C(sbi, count);
4773	bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4774	if (IS_ERR(bitmap_bh)) {
4775		err = PTR_ERR(bitmap_bh);
4776		bitmap_bh = NULL;
4777		goto error_return;
4778	}
4779	gdp = ext4_get_group_desc(sb, block_group, &gd_bh);
4780	if (!gdp) {
4781		err = -EIO;
4782		goto error_return;
4783	}
4784
4785	if (in_range(ext4_block_bitmap(sb, gdp), block, count) ||
4786	    in_range(ext4_inode_bitmap(sb, gdp), block, count) ||
4787	    in_range(block, ext4_inode_table(sb, gdp),
4788		     EXT4_SB(sb)->s_itb_per_group) ||
4789	    in_range(block + count - 1, ext4_inode_table(sb, gdp),
4790		     EXT4_SB(sb)->s_itb_per_group)) {
4791
4792		ext4_error(sb, "Freeing blocks in system zone - "
4793			   "Block = %llu, count = %lu", block, count);
4794		/* err = 0. ext4_std_error should be a no op */
4795		goto error_return;
4796	}
4797
4798	BUFFER_TRACE(bitmap_bh, "getting write access");
4799	err = ext4_journal_get_write_access(handle, bitmap_bh);
4800	if (err)
4801		goto error_return;
4802
4803	/*
4804	 * We are about to modify some metadata.  Call the journal APIs
4805	 * to unshare ->b_data if a currently-committing transaction is
4806	 * using it
4807	 */
4808	BUFFER_TRACE(gd_bh, "get_write_access");
4809	err = ext4_journal_get_write_access(handle, gd_bh);
4810	if (err)
4811		goto error_return;
4812#ifdef AGGRESSIVE_CHECK
4813	{
4814		int i;
4815		for (i = 0; i < count_clusters; i++)
4816			BUG_ON(!mb_test_bit(bit + i, bitmap_bh->b_data));
4817	}
4818#endif
4819	trace_ext4_mballoc_free(sb, inode, block_group, bit, count_clusters);
4820
4821	/* __GFP_NOFAIL: retry infinitely, ignore TIF_MEMDIE and memcg limit. */
4822	err = ext4_mb_load_buddy_gfp(sb, block_group, &e4b,
4823				     GFP_NOFS|__GFP_NOFAIL);
4824	if (err)
4825		goto error_return;
4826
4827	/*
4828	 * We need to make sure we don't reuse the freed block until after the
4829	 * transaction is committed. We make an exception if the inode is to be
4830	 * written in writeback mode since writeback mode has weak data
4831	 * consistency guarantees.
4832	 */
4833	if (ext4_handle_valid(handle) &&
4834	    ((flags & EXT4_FREE_BLOCKS_METADATA) ||
4835	     !ext4_should_writeback_data(inode))) {
4836		struct ext4_free_data *new_entry;
4837		/*
4838		 * We use __GFP_NOFAIL because ext4_free_blocks() is not allowed
4839		 * to fail.
4840		 */
4841		new_entry = kmem_cache_alloc(ext4_free_data_cachep,
4842				GFP_NOFS|__GFP_NOFAIL);
4843		new_entry->efd_start_cluster = bit;
4844		new_entry->efd_group = block_group;
4845		new_entry->efd_count = count_clusters;
4846		new_entry->efd_tid = handle->h_transaction->t_tid;
4847
4848		ext4_lock_group(sb, block_group);
4849		mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4850		ext4_mb_free_metadata(handle, &e4b, new_entry);
4851	} else {
4852		/* need to update group_info->bb_free and bitmap
4853		 * with group lock held. generate_buddy look at
4854		 * them with group lock_held
4855		 */
4856		if (test_opt(sb, DISCARD)) {
4857			err = ext4_issue_discard(sb, block_group, bit, count);
4858			if (err && err != -EOPNOTSUPP)
4859				ext4_msg(sb, KERN_WARNING, "discard request in"
4860					 " group:%d block:%d count:%lu failed"
4861					 " with %d", block_group, bit, count,
4862					 err);
4863		} else
4864			EXT4_MB_GRP_CLEAR_TRIMMED(e4b.bd_info);
4865
4866		ext4_lock_group(sb, block_group);
4867		mb_clear_bits(bitmap_bh->b_data, bit, count_clusters);
4868		mb_free_blocks(inode, &e4b, bit, count_clusters);
4869	}
4870
4871	ret = ext4_free_group_clusters(sb, gdp) + count_clusters;
4872	ext4_free_group_clusters_set(sb, gdp, ret);
4873	ext4_block_bitmap_csum_set(sb, block_group, gdp, bitmap_bh);
4874	ext4_group_desc_csum_set(sb, block_group, gdp);
4875	ext4_unlock_group(sb, block_group);
4876
4877	if (sbi->s_log_groups_per_flex) {
4878		ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
4879		atomic64_add(count_clusters,
4880			     &sbi->s_flex_groups[flex_group].free_clusters);
4881	}
4882
4883	if (!(flags & EXT4_FREE_BLOCKS_NO_QUOT_UPDATE))
4884		dquot_free_block(inode, EXT4_C2B(sbi, count_clusters));
4885	percpu_counter_add(&sbi->s_freeclusters_counter, count_clusters);
4886
4887	ext4_mb_unload_buddy(&e4b);
4888
4889	/* We dirtied the bitmap block */
4890	BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
4891	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
4892
4893	/* And the group descriptor block */
4894	BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
4895	ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
4896	if (!err)
4897		err = ret;
4898
4899	if (overflow && !err) {
4900		block += count;
4901		count = overflow;
4902		put_bh(bitmap_bh);
4903		goto do_more;
4904	}
4905error_return:
4906	brelse(bitmap_bh);
4907	ext4_std_error(sb, err);
4908	return;
4909}
4910
4911/**
4912 * ext4_group_add_blocks() -- Add given blocks to an existing group
4913 * @handle:			handle to this transaction
4914 * @sb:				super block
4915 * @block:			start physical block to add to the block group
4916 * @count:			number of blocks to free
4917 *
4918 * This marks the blocks as free in the bitmap and buddy.
4919 */
4920int ext4_group_add_blocks(handle_t *handle, struct super_block *sb,
4921			 ext4_fsblk_t block, unsigned long count)
4922{
4923	struct buffer_head *bitmap_bh = NULL;
4924	struct buffer_head *gd_bh;
4925	ext4_group_t block_group;
4926	ext4_grpblk_t bit;
4927	unsigned int i;
4928	struct ext4_group_desc *desc;
4929	struct ext4_sb_info *sbi = EXT4_SB(sb);
4930	struct ext4_buddy e4b;
4931	int err = 0, ret, blk_free_count;
4932	ext4_grpblk_t blocks_freed;
4933
4934	ext4_debug("Adding block(s) %llu-%llu\n", block, block + count - 1);
4935
4936	if (count == 0)
4937		return 0;
4938
4939	ext4_get_group_no_and_offset(sb, block, &block_group, &bit);
4940	/*
4941	 * Check to see if we are freeing blocks across a group
4942	 * boundary.
4943	 */
4944	if (bit + count > EXT4_BLOCKS_PER_GROUP(sb)) {
4945		ext4_warning(sb, "too much blocks added to group %u",
4946			     block_group);
4947		err = -EINVAL;
4948		goto error_return;
4949	}
4950
4951	bitmap_bh = ext4_read_block_bitmap(sb, block_group);
4952	if (IS_ERR(bitmap_bh)) {
4953		err = PTR_ERR(bitmap_bh);
4954		bitmap_bh = NULL;
4955		goto error_return;
4956	}
4957
4958	desc = ext4_get_group_desc(sb, block_group, &gd_bh);
4959	if (!desc) {
4960		err = -EIO;
4961		goto error_return;
4962	}
4963
4964	if (in_range(ext4_block_bitmap(sb, desc), block, count) ||
4965	    in_range(ext4_inode_bitmap(sb, desc), block, count) ||
4966	    in_range(block, ext4_inode_table(sb, desc), sbi->s_itb_per_group) ||
4967	    in_range(block + count - 1, ext4_inode_table(sb, desc),
4968		     sbi->s_itb_per_group)) {
4969		ext4_error(sb, "Adding blocks in system zones - "
4970			   "Block = %llu, count = %lu",
4971			   block, count);
4972		err = -EINVAL;
4973		goto error_return;
4974	}
4975
4976	BUFFER_TRACE(bitmap_bh, "getting write access");
4977	err = ext4_journal_get_write_access(handle, bitmap_bh);
4978	if (err)
4979		goto error_return;
4980
4981	/*
4982	 * We are about to modify some metadata.  Call the journal APIs
4983	 * to unshare ->b_data if a currently-committing transaction is
4984	 * using it
4985	 */
4986	BUFFER_TRACE(gd_bh, "get_write_access");
4987	err = ext4_journal_get_write_access(handle, gd_bh);
4988	if (err)
4989		goto error_return;
4990
4991	for (i = 0, blocks_freed = 0; i < count; i++) {
4992		BUFFER_TRACE(bitmap_bh, "clear bit");
4993		if (!mb_test_bit(bit + i, bitmap_bh->b_data)) {
4994			ext4_error(sb, "bit already cleared for block %llu",
4995				   (ext4_fsblk_t)(block + i));
4996			BUFFER_TRACE(bitmap_bh, "bit already cleared");
4997		} else {
4998			blocks_freed++;
4999		}
5000	}
5001
5002	err = ext4_mb_load_buddy(sb, block_group, &e4b);
5003	if (err)
5004		goto error_return;
5005
5006	/*
5007	 * need to update group_info->bb_free and bitmap
5008	 * with group lock held. generate_buddy look at
5009	 * them with group lock_held
5010	 */
5011	ext4_lock_group(sb, block_group);
5012	mb_clear_bits(bitmap_bh->b_data, bit, count);
5013	mb_free_blocks(NULL, &e4b, bit, count);
5014	blk_free_count = blocks_freed + ext4_free_group_clusters(sb, desc);
5015	ext4_free_group_clusters_set(sb, desc, blk_free_count);
5016	ext4_block_bitmap_csum_set(sb, block_group, desc, bitmap_bh);
5017	ext4_group_desc_csum_set(sb, block_group, desc);
5018	ext4_unlock_group(sb, block_group);
5019	percpu_counter_add(&sbi->s_freeclusters_counter,
5020			   EXT4_NUM_B2C(sbi, blocks_freed));
5021
5022	if (sbi->s_log_groups_per_flex) {
5023		ext4_group_t flex_group = ext4_flex_group(sbi, block_group);
5024		atomic64_add(EXT4_NUM_B2C(sbi, blocks_freed),
5025			     &sbi->s_flex_groups[flex_group].free_clusters);
5026	}
5027
5028	ext4_mb_unload_buddy(&e4b);
5029
5030	/* We dirtied the bitmap block */
5031	BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
5032	err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
5033
5034	/* And the group descriptor block */
5035	BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
5036	ret = ext4_handle_dirty_metadata(handle, NULL, gd_bh);
5037	if (!err)
5038		err = ret;
5039
5040error_return:
5041	brelse(bitmap_bh);
5042	ext4_std_error(sb, err);
5043	return err;
5044}
5045
5046/**
5047 * ext4_trim_extent -- function to TRIM one single free extent in the group
5048 * @sb:		super block for the file system
5049 * @start:	starting block of the free extent in the alloc. group
5050 * @count:	number of blocks to TRIM
5051 * @group:	alloc. group we are working with
5052 * @e4b:	ext4 buddy for the group
5053 *
5054 * Trim "count" blocks starting at "start" in the "group". To assure that no
5055 * one will allocate those blocks, mark it as used in buddy bitmap. This must
5056 * be called with under the group lock.
5057 */
5058static int ext4_trim_extent(struct super_block *sb, int start, int count,
5059			     ext4_group_t group, struct ext4_buddy *e4b)
5060__releases(bitlock)
5061__acquires(bitlock)
5062{
5063	struct ext4_free_extent ex;
5064	int ret = 0;
5065
5066	trace_ext4_trim_extent(sb, group, start, count);
5067
5068	assert_spin_locked(ext4_group_lock_ptr(sb, group));
5069
5070	ex.fe_start = start;
5071	ex.fe_group = group;
5072	ex.fe_len = count;
5073
5074	/*
5075	 * Mark blocks used, so no one can reuse them while
5076	 * being trimmed.
5077	 */
5078	mb_mark_used(e4b, &ex);
5079	ext4_unlock_group(sb, group);
5080	ret = ext4_issue_discard(sb, group, start, count);
5081	ext4_lock_group(sb, group);
5082	mb_free_blocks(NULL, e4b, start, ex.fe_len);
5083	return ret;
5084}
5085
5086/**
5087 * ext4_trim_all_free -- function to trim all free space in alloc. group
5088 * @sb:			super block for file system
5089 * @group:		group to be trimmed
5090 * @start:		first group block to examine
5091 * @max:		last group block to examine
5092 * @minblocks:		minimum extent block count
5093 *
5094 * ext4_trim_all_free walks through group's buddy bitmap searching for free
5095 * extents. When the free block is found, ext4_trim_extent is called to TRIM
5096 * the extent.
5097 *
5098 *
5099 * ext4_trim_all_free walks through group's block bitmap searching for free
5100 * extents. When the free extent is found, mark it as used in group buddy
5101 * bitmap. Then issue a TRIM command on this extent and free the extent in
5102 * the group buddy bitmap. This is done until whole group is scanned.
5103 */
5104static ext4_grpblk_t
5105ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
5106		   ext4_grpblk_t start, ext4_grpblk_t max,
5107		   ext4_grpblk_t minblocks)
5108{
5109	void *bitmap;
5110	ext4_grpblk_t next, count = 0, free_count = 0;
5111	struct ext4_buddy e4b;
5112	int ret = 0;
5113
5114	trace_ext4_trim_all_free(sb, group, start, max);
5115
5116	ret = ext4_mb_load_buddy(sb, group, &e4b);
5117	if (ret) {
5118		ext4_error(sb, "Error in loading buddy "
5119				"information for %u", group);
5120		return ret;
5121	}
5122	bitmap = e4b.bd_bitmap;
5123
5124	ext4_lock_group(sb, group);
5125	if (EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) &&
5126	    minblocks >= atomic_read(&EXT4_SB(sb)->s_last_trim_minblks))
5127		goto out;
5128
5129	start = (e4b.bd_info->bb_first_free > start) ?
5130		e4b.bd_info->bb_first_free : start;
5131
5132	while (start <= max) {
5133		start = mb_find_next_zero_bit(bitmap, max + 1, start);
5134		if (start > max)
5135			break;
5136		next = mb_find_next_bit(bitmap, max + 1, start);
5137
5138		if ((next - start) >= minblocks) {
5139			ret = ext4_trim_extent(sb, start,
5140					       next - start, group, &e4b);
5141			if (ret && ret != -EOPNOTSUPP)
5142				break;
5143			ret = 0;
5144			count += next - start;
5145		}
5146		free_count += next - start;
5147		start = next + 1;
5148
5149		if (fatal_signal_pending(current)) {
5150			count = -ERESTARTSYS;
5151			break;
5152		}
5153
5154		if (need_resched()) {
5155			ext4_unlock_group(sb, group);
5156			cond_resched();
5157			ext4_lock_group(sb, group);
5158		}
5159
5160		if ((e4b.bd_info->bb_free - free_count) < minblocks)
5161			break;
5162	}
5163
5164	if (!ret) {
5165		ret = count;
5166		EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);
5167	}
5168out:
5169	ext4_unlock_group(sb, group);
5170	ext4_mb_unload_buddy(&e4b);
5171
5172	ext4_debug("trimmed %d blocks in the group %d\n",
5173		count, group);
5174
5175	return ret;
5176}
5177
5178/**
5179 * ext4_trim_fs() -- trim ioctl handle function
5180 * @sb:			superblock for filesystem
5181 * @range:		fstrim_range structure
5182 *
5183 * start:	First Byte to trim
5184 * len:		number of Bytes to trim from start
5185 * minlen:	minimum extent length in Bytes
5186 * ext4_trim_fs goes through all allocation groups containing Bytes from
5187 * start to start+len. For each such a group ext4_trim_all_free function
5188 * is invoked to trim all free space.
5189 */
5190int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
5191{
5192	struct ext4_group_info *grp;
5193	ext4_group_t group, first_group, last_group;
5194	ext4_grpblk_t cnt = 0, first_cluster, last_cluster;
5195	uint64_t start, end, minlen, trimmed = 0;
5196	ext4_fsblk_t first_data_blk =
5197			le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block);
5198	ext4_fsblk_t max_blks = ext4_blocks_count(EXT4_SB(sb)->s_es);
5199	int ret = 0;
5200
5201	start = range->start >> sb->s_blocksize_bits;
5202	end = start + (range->len >> sb->s_blocksize_bits) - 1;
5203	minlen = EXT4_NUM_B2C(EXT4_SB(sb),
5204			      range->minlen >> sb->s_blocksize_bits);
5205
5206	if (minlen > EXT4_CLUSTERS_PER_GROUP(sb) ||
5207	    start >= max_blks ||
5208	    range->len < sb->s_blocksize)
5209		return -EINVAL;
5210	if (end >= max_blks)
5211		end = max_blks - 1;
5212	if (end <= first_data_blk)
5213		goto out;
5214	if (start < first_data_blk)
5215		start = first_data_blk;
5216
5217	/* Determine first and last group to examine based on start and end */
5218	ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) start,
5219				     &first_group, &first_cluster);
5220	ext4_get_group_no_and_offset(sb, (ext4_fsblk_t) end,
5221				     &last_group, &last_cluster);
5222
5223	/* end now represents the last cluster to discard in this group */
5224	end = EXT4_CLUSTERS_PER_GROUP(sb) - 1;
5225
5226	for (group = first_group; group <= last_group; group++) {
5227		grp = ext4_get_group_info(sb, group);
5228		/* We only do this if the grp has never been initialized */
5229		if (unlikely(EXT4_MB_GRP_NEED_INIT(grp))) {
5230			ret = ext4_mb_init_group(sb, group, GFP_NOFS);
5231			if (ret)
5232				break;
5233		}
5234
5235		/*
5236		 * For all the groups except the last one, last cluster will
5237		 * always be EXT4_CLUSTERS_PER_GROUP(sb)-1, so we only need to
5238		 * change it for the last group, note that last_cluster is
5239		 * already computed earlier by ext4_get_group_no_and_offset()
5240		 */
5241		if (group == last_group)
5242			end = last_cluster;
5243
5244		if (grp->bb_free >= minlen) {
5245			cnt = ext4_trim_all_free(sb, group, first_cluster,
5246						end, minlen);
5247			if (cnt < 0) {
5248				ret = cnt;
5249				break;
5250			}
5251			trimmed += cnt;
5252		}
5253
5254		/*
5255		 * For every group except the first one, we are sure
5256		 * that the first cluster to discard will be cluster #0.
5257		 */
5258		first_cluster = 0;
5259	}
5260
5261	if (!ret)
5262		atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
5263
5264out:
5265	range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;
5266	return ret;
5267}