fs/xfs/linux-2.6/xfs_buf.c

   1 /*
   2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would 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 the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include <linux/stddef.h>
  20 #include <linux/errno.h>
  21 #include <linux/gfp.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/init.h>
  24 #include <linux/vmalloc.h>
  25 #include <linux/bio.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/proc_fs.h>
  28 #include <linux/workqueue.h>
  29 #include <linux/percpu.h>
  30 #include <linux/blkdev.h>
  31 #include <linux/hash.h>
  32 #include <linux/kthread.h>
  33 #include <linux/migrate.h>
  34 #include <linux/backing-dev.h>
  35 #include <linux/freezer.h>
  36 #include <linux/list_sort.h>
  37
  38 #include "xfs_sb.h"
  39 #include "xfs_inum.h"
  40 #include "xfs_log.h"
  41 #include "xfs_ag.h"
  42 #include "xfs_mount.h"
  43 #include "xfs_trace.h"
  44
  45 static kmem_zone_t *xfs_buf_zone;
  46 STATIC int xfsbufd(void *);
  47 STATIC int xfsbufd_wakeup(struct shrinker *, int, gfp_t);
  48 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
  49 static struct shrinker xfs_buf_shake = {
  50         .shrink = xfsbufd_wakeup,
  51         .seeks = DEFAULT_SEEKS,
  52 };
  53
  54 static struct workqueue_struct *xfslogd_workqueue;
  55 struct workqueue_struct *xfsdatad_workqueue;
  56 struct workqueue_struct *xfsconvertd_workqueue;
  57
  58 #ifdef XFS_BUF_LOCK_TRACKING
  59 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  60 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  61 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  62 #else
  63 # define XB_SET_OWNER(bp)       do { } while (0)
  64 # define XB_CLEAR_OWNER(bp)     do { } while (0)
  65 # define XB_GET_OWNER(bp)       do { } while (0)
  66 #endif
  67
  68 #define xb_to_gfp(flags) \
  69         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
  70           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
  71
  72 #define xb_to_km(flags) \
  73          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
  74
  75 #define xfs_buf_allocate(flags) \
  76         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
  77 #define xfs_buf_deallocate(bp) \
  78         kmem_zone_free(xfs_buf_zone, (bp));
  79
  80 static inline int
  81 xfs_buf_is_vmapped(
  82         struct xfs_buf  *bp)
  83 {
  84         /*
  85          * Return true if the buffer is vmapped.
  86          *
  87          * The XBF_MAPPED flag is set if the buffer should be mapped, but the
  88          * code is clever enough to know it doesn't have to map a single page,
  89          * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
  90          */
  91         return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
  92 }
  93
  94 static inline int
  95 xfs_buf_vmap_len(
  96         struct xfs_buf  *bp)
  97 {
  98         return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
  99 }
 100
 101 /*
 102  *      Page Region interfaces.
 103  *
 104  *      For pages in filesystems where the blocksize is smaller than the
 105  *      pagesize, we use the page->private field (long) to hold a bitmap
 106  *      of uptodate regions within the page.
 107  *
 108  *      Each such region is "bytes per page / bits per long" bytes long.
 109  *
 110  *      NBPPR == number-of-bytes-per-page-region
 111  *      BTOPR == bytes-to-page-region (rounded up)
 112  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
 113  */
 114 #if (BITS_PER_LONG == 32)
 115 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
 116 #elif (BITS_PER_LONG == 64)
 117 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
 118 #else
 119 #error BITS_PER_LONG must be 32 or 64
 120 #endif
 121 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
 122 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
 123 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
 124
 125 STATIC unsigned long
 126 page_region_mask(
 127         size_t          offset,
 128         size_t          length)
 129 {
 130         unsigned long   mask;
 131         int             first, final;
 132
 133         first = BTOPR(offset);
 134         final = BTOPRT(offset + length - 1);
 135         first = min(first, final);
 136
 137         mask = ~0UL;
 138         mask <<= BITS_PER_LONG - (final - first);
 139         mask >>= BITS_PER_LONG - (final);
 140
 141         ASSERT(offset + length <= PAGE_CACHE_SIZE);
 142         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
 143
 144         return mask;
 145 }
 146
 147 STATIC void
 148 set_page_region(
 149         struct page     *page,
 150         size_t          offset,
 151         size_t          length)
 152 {
 153         set_page_private(page,
 154                 page_private(page) | page_region_mask(offset, length));
 155         if (page_private(page) == ~0UL)
 156                 SetPageUptodate(page);
 157 }
 158
 159 STATIC int
 160 test_page_region(
 161         struct page     *page,
 162         size_t          offset,
 163         size_t          length)
 164 {
 165         unsigned long   mask = page_region_mask(offset, length);
 166
 167         return (mask && (page_private(page) & mask) == mask);
 168 }
 169
 170 /*
 171  *      Internal xfs_buf_t object manipulation
 172  */
 173
 174 STATIC void
 175 _xfs_buf_initialize(
 176         xfs_buf_t               *bp,
 177         xfs_buftarg_t           *target,
 178         xfs_off_t               range_base,
 179         size_t                  range_length,
 180         xfs_buf_flags_t         flags)
 181 {
 182         /*
 183          * We don't want certain flags to appear in b_flags.
 184          */
 185         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
 186
 187         memset(bp, 0, sizeof(xfs_buf_t));
 188         atomic_set(&bp->b_hold, 1);
 189         init_completion(&bp->b_iowait);
 190         INIT_LIST_HEAD(&bp->b_list);
 191         INIT_LIST_HEAD(&bp->b_hash_list);
 192         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
 193         XB_SET_OWNER(bp);
 194         bp->b_target = target;
 195         bp->b_file_offset = range_base;
 196         /*
 197          * Set buffer_length and count_desired to the same value initially.
 198          * I/O routines should use count_desired, which will be the same in
 199          * most cases but may be reset (e.g. XFS recovery).
 200          */
 201         bp->b_buffer_length = bp->b_count_desired = range_length;
 202         bp->b_flags = flags;
 203         bp->b_bn = XFS_BUF_DADDR_NULL;
 204         atomic_set(&bp->b_pin_count, 0);
 205         init_waitqueue_head(&bp->b_waiters);
 206
 207         XFS_STATS_INC(xb_create);
 208
 209         trace_xfs_buf_init(bp, _RET_IP_);
 210 }
 211
 212 /*
 213  *      Allocate a page array capable of holding a specified number
 214  *      of pages, and point the page buf at it.
 215  */
 216 STATIC int
 217 _xfs_buf_get_pages(
 218         xfs_buf_t               *bp,
 219         int                     page_count,
 220         xfs_buf_flags_t         flags)
 221 {
 222         /* Make sure that we have a page list */
 223         if (bp->b_pages == NULL) {
 224                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
 225                 bp->b_page_count = page_count;
 226                 if (page_count <= XB_PAGES) {
 227                         bp->b_pages = bp->b_page_array;
 228                 } else {
 229                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
 230                                         page_count, xb_to_km(flags));
 231                         if (bp->b_pages == NULL)
 232                                 return -ENOMEM;
 233                 }
 234                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 235         }
 236         return 0;
 237 }
 238
 239 /*
 240  *      Frees b_pages if it was allocated.
 241  */
 242 STATIC void
 243 _xfs_buf_free_pages(
 244         xfs_buf_t       *bp)
 245 {
 246         if (bp->b_pages != bp->b_page_array) {
 247                 kmem_free(bp->b_pages);
 248                 bp->b_pages = NULL;
 249         }
 250 }
 251
 252 /*
 253  *      Releases the specified buffer.
 254  *
 255  *      The modification state of any associated pages is left unchanged.
 256  *      The buffer most not be on any hash - use xfs_buf_rele instead for
 257  *      hashed and refcounted buffers
 258  */
 259 void
 260 xfs_buf_free(
 261         xfs_buf_t               *bp)
 262 {
 263         trace_xfs_buf_free(bp, _RET_IP_);
 264
 265         ASSERT(list_empty(&bp->b_hash_list));
 266
 267         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
 268                 uint            i;
 269
 270                 if (xfs_buf_is_vmapped(bp))
 271                         vm_unmap_ram(bp->b_addr - bp->b_offset,
 272                                         bp->b_page_count);
 273
 274                 for (i = 0; i < bp->b_page_count; i++) {
 275                         struct page     *page = bp->b_pages[i];
 276
 277                         if (bp->b_flags & _XBF_PAGE_CACHE)
 278                                 ASSERT(!PagePrivate(page));
 279                         page_cache_release(page);
 280                 }
 281         }
 282         _xfs_buf_free_pages(bp);
 283         xfs_buf_deallocate(bp);
 284 }
 285
 286 /*
 287  *      Finds all pages for buffer in question and builds it's page list.
 288  */
 289 STATIC int
 290 _xfs_buf_lookup_pages(
 291         xfs_buf_t               *bp,
 292         uint                    flags)
 293 {
 294         struct address_space    *mapping = bp->b_target->bt_mapping;
 295         size_t                  blocksize = bp->b_target->bt_bsize;
 296         size_t                  size = bp->b_count_desired;
 297         size_t                  nbytes, offset;
 298         gfp_t                   gfp_mask = xb_to_gfp(flags);
 299         unsigned short          page_count, i;
 300         pgoff_t                 first;
 301         xfs_off_t               end;
 302         int                     error;
 303
 304         end = bp->b_file_offset + bp->b_buffer_length;
 305         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 306
 307         error = _xfs_buf_get_pages(bp, page_count, flags);
 308         if (unlikely(error))
 309                 return error;
 310         bp->b_flags |= _XBF_PAGE_CACHE;
 311
 312         offset = bp->b_offset;
 313         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
 314
 315         for (i = 0; i < bp->b_page_count; i++) {
 316                 struct page     *page;
 317                 uint            retries = 0;
 318
 319               retry:
 320                 page = find_or_create_page(mapping, first + i, gfp_mask);
 321                 if (unlikely(page == NULL)) {
 322                         if (flags & XBF_READ_AHEAD) {
 323                                 bp->b_page_count = i;
 324                                 for (i = 0; i < bp->b_page_count; i++)
 325                                         unlock_page(bp->b_pages[i]);
 326                                 return -ENOMEM;
 327                         }
 328
 329                         /*
 330                          * This could deadlock.
 331                          *
 332                          * But until all the XFS lowlevel code is revamped to
 333                          * handle buffer allocation failures we can't do much.
 334                          */
 335                         if (!(++retries % 100))
 336                                 printk(KERN_ERR
 337                                         "XFS: possible memory allocation "
 338                                         "deadlock in %s (mode:0x%x)\n",
 339                                         __func__, gfp_mask);
 340
 341                         XFS_STATS_INC(xb_page_retries);
 342                         xfsbufd_wakeup(NULL, 0, gfp_mask);
 343                         congestion_wait(BLK_RW_ASYNC, HZ/50);
 344                         goto retry;
 345                 }
 346
 347                 XFS_STATS_INC(xb_page_found);
 348
 349                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
 350                 size -= nbytes;
 351
 352                 ASSERT(!PagePrivate(page));
 353                 if (!PageUptodate(page)) {
 354                         page_count--;
 355                         if (blocksize >= PAGE_CACHE_SIZE) {
 356                                 if (flags & XBF_READ)
 357                                         bp->b_flags |= _XBF_PAGE_LOCKED;
 358                         } else if (!PagePrivate(page)) {
 359                                 if (test_page_region(page, offset, nbytes))
 360                                         page_count++;
 361                         }
 362                 }
 363
 364                 bp->b_pages[i] = page;
 365                 offset = 0;
 366         }
 367
 368         if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
 369                 for (i = 0; i < bp->b_page_count; i++)
 370                         unlock_page(bp->b_pages[i]);
 371         }
 372
 373         if (page_count == bp->b_page_count)
 374                 bp->b_flags |= XBF_DONE;
 375
 376         return error;
 377 }
 378
 379 /*
 380  *      Map buffer into kernel address-space if nessecary.
 381  */
 382 STATIC int
 383 _xfs_buf_map_pages(
 384         xfs_buf_t               *bp,
 385         uint                    flags)
 386 {
 387         /* A single page buffer is always mappable */
 388         if (bp->b_page_count == 1) {
 389                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 390                 bp->b_flags |= XBF_MAPPED;
 391         } else if (flags & XBF_MAPPED) {
 392                 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
 393                                         -1, PAGE_KERNEL);
 394                 if (unlikely(bp->b_addr == NULL))
 395                         return -ENOMEM;
 396                 bp->b_addr += bp->b_offset;
 397                 bp->b_flags |= XBF_MAPPED;
 398         }
 399
 400         return 0;
 401 }
 402
 403 /*
 404  *      Finding and Reading Buffers
 405  */
 406
 407 /*
 408  *      Look up, and creates if absent, a lockable buffer for
 409  *      a given range of an inode.  The buffer is returned
 410  *      locked.  If other overlapping buffers exist, they are
 411  *      released before the new buffer is created and locked,
 412  *      which may imply that this call will block until those buffers
 413  *      are unlocked.  No I/O is implied by this call.
 414  */
 415 xfs_buf_t *
 416 _xfs_buf_find(
 417         xfs_buftarg_t           *btp,   /* block device target          */
 418         xfs_off_t               ioff,   /* starting offset of range     */
 419         size_t                  isize,  /* length of range              */
 420         xfs_buf_flags_t         flags,
 421         xfs_buf_t               *new_bp)
 422 {
 423         xfs_off_t               range_base;
 424         size_t                  range_length;
 425         xfs_bufhash_t           *hash;
 426         xfs_buf_t               *bp, *n;
 427
 428         range_base = (ioff << BBSHIFT);
 429         range_length = (isize << BBSHIFT);
 430
 431         /* Check for IOs smaller than the sector size / not sector aligned */
 432         ASSERT(!(range_length < (1 << btp->bt_sshift)));
 433         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
 434
 435         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
 436
 437         spin_lock(&hash->bh_lock);
 438
 439         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
 440                 ASSERT(btp == bp->b_target);
 441                 if (bp->b_file_offset == range_base &&
 442                     bp->b_buffer_length == range_length) {
 443                         atomic_inc(&bp->b_hold);
 444                         goto found;
 445                 }
 446         }
 447
 448         /* No match found */
 449         if (new_bp) {
 450                 _xfs_buf_initialize(new_bp, btp, range_base,
 451                                 range_length, flags);
 452                 new_bp->b_hash = hash;
 453                 list_add(&new_bp->b_hash_list, &hash->bh_list);
 454         } else {
 455                 XFS_STATS_INC(xb_miss_locked);
 456         }
 457
 458         spin_unlock(&hash->bh_lock);
 459         return new_bp;
 460
 461 found:
 462         spin_unlock(&hash->bh_lock);
 463
 464         /* Attempt to get the semaphore without sleeping,
 465          * if this does not work then we need to drop the
 466          * spinlock and do a hard attempt on the semaphore.
 467          */
 468         if (down_trylock(&bp->b_sema)) {
 469                 if (!(flags & XBF_TRYLOCK)) {
 470                         /* wait for buffer ownership */
 471                         xfs_buf_lock(bp);
 472                         XFS_STATS_INC(xb_get_locked_waited);
 473                 } else {
 474                         /* We asked for a trylock and failed, no need
 475                          * to look at file offset and length here, we
 476                          * know that this buffer at least overlaps our
 477                          * buffer and is locked, therefore our buffer
 478                          * either does not exist, or is this buffer.
 479                          */
 480                         xfs_buf_rele(bp);
 481                         XFS_STATS_INC(xb_busy_locked);
 482                         return NULL;
 483                 }
 484         } else {
 485                 /* trylock worked */
 486                 XB_SET_OWNER(bp);
 487         }
 488
 489         if (bp->b_flags & XBF_STALE) {
 490                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 491                 bp->b_flags &= XBF_MAPPED;
 492         }
 493
 494         trace_xfs_buf_find(bp, flags, _RET_IP_);
 495         XFS_STATS_INC(xb_get_locked);
 496         return bp;
 497 }
 498
 499 /*
 500  *      Assembles a buffer covering the specified range.
 501  *      Storage in memory for all portions of the buffer will be allocated,
 502  *      although backing storage may not be.
 503  */
 504 xfs_buf_t *
 505 xfs_buf_get(
 506         xfs_buftarg_t           *target,/* target for buffer            */
 507         xfs_off_t               ioff,   /* starting offset of range     */
 508         size_t                  isize,  /* length of range              */
 509         xfs_buf_flags_t         flags)
 510 {
 511         xfs_buf_t               *bp, *new_bp;
 512         int                     error = 0, i;
 513
 514         new_bp = xfs_buf_allocate(flags);
 515         if (unlikely(!new_bp))
 516                 return NULL;
 517
 518         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 519         if (bp == new_bp) {
 520                 error = _xfs_buf_lookup_pages(bp, flags);
 521                 if (error)
 522                         goto no_buffer;
 523         } else {
 524                 xfs_buf_deallocate(new_bp);
 525                 if (unlikely(bp == NULL))
 526                         return NULL;
 527         }
 528
 529         for (i = 0; i < bp->b_page_count; i++)
 530                 mark_page_accessed(bp->b_pages[i]);
 531
 532         if (!(bp->b_flags & XBF_MAPPED)) {
 533                 error = _xfs_buf_map_pages(bp, flags);
 534                 if (unlikely(error)) {
 535                         printk(KERN_WARNING "%s: failed to map pages\n",
 536                                         __func__);
 537                         goto no_buffer;
 538                 }
 539         }
 540
 541         XFS_STATS_INC(xb_get);
 542
 543         /*
 544          * Always fill in the block number now, the mapped cases can do
 545          * their own overlay of this later.
 546          */
 547         bp->b_bn = ioff;
 548         bp->b_count_desired = bp->b_buffer_length;
 549
 550         trace_xfs_buf_get(bp, flags, _RET_IP_);
 551         return bp;
 552
 553  no_buffer:
 554         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 555                 xfs_buf_unlock(bp);
 556         xfs_buf_rele(bp);
 557         return NULL;
 558 }
 559
 560 STATIC int
 561 _xfs_buf_read(
 562         xfs_buf_t               *bp,
 563         xfs_buf_flags_t         flags)
 564 {
 565         int                     status;
 566
 567         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
 568         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
 569
 570         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
 571                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 572         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
 573                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 574
 575         status = xfs_buf_iorequest(bp);
 576         if (status || XFS_BUF_ISERROR(bp) || (flags & XBF_ASYNC))
 577                 return status;
 578         return xfs_buf_iowait(bp);
 579 }
 580
 581 xfs_buf_t *
 582 xfs_buf_read(
 583         xfs_buftarg_t           *target,
 584         xfs_off_t               ioff,
 585         size_t                  isize,
 586         xfs_buf_flags_t         flags)
 587 {
 588         xfs_buf_t               *bp;
 589
 590         flags |= XBF_READ;
 591
 592         bp = xfs_buf_get(target, ioff, isize, flags);
 593         if (bp) {
 594                 trace_xfs_buf_read(bp, flags, _RET_IP_);
 595
 596                 if (!XFS_BUF_ISDONE(bp)) {
 597                         XFS_STATS_INC(xb_get_read);
 598                         _xfs_buf_read(bp, flags);
 599                 } else if (flags & XBF_ASYNC) {
 600                         /*
 601                          * Read ahead call which is already satisfied,
 602                          * drop the buffer
 603                          */
 604                         goto no_buffer;
 605                 } else {
 606                         /* We do not want read in the flags */
 607                         bp->b_flags &= ~XBF_READ;
 608                 }
 609         }
 610
 611         return bp;
 612
 613  no_buffer:
 614         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 615                 xfs_buf_unlock(bp);
 616         xfs_buf_rele(bp);
 617         return NULL;
 618 }
 619
 620 /*
 621  *      If we are not low on memory then do the readahead in a deadlock
 622  *      safe manner.
 623  */
 624 void
 625 xfs_buf_readahead(
 626         xfs_buftarg_t           *target,
 627         xfs_off_t               ioff,
 628         size_t                  isize,
 629         xfs_buf_flags_t         flags)
 630 {
 631         struct backing_dev_info *bdi;
 632
 633         bdi = target->bt_mapping->backing_dev_info;
 634         if (bdi_read_congested(bdi))
 635                 return;
 636
 637         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 638         xfs_buf_read(target, ioff, isize, flags);
 639 }
 640
 641 xfs_buf_t *
 642 xfs_buf_get_empty(
 643         size_t                  len,
 644         xfs_buftarg_t           *target)
 645 {
 646         xfs_buf_t               *bp;
 647
 648         bp = xfs_buf_allocate(0);
 649         if (bp)
 650                 _xfs_buf_initialize(bp, target, 0, len, 0);
 651         return bp;
 652 }
 653
 654 static inline struct page *
 655 mem_to_page(
 656         void                    *addr)
 657 {
 658         if ((!is_vmalloc_addr(addr))) {
 659                 return virt_to_page(addr);
 660         } else {
 661                 return vmalloc_to_page(addr);
 662         }
 663 }
 664
 665 int
 666 xfs_buf_associate_memory(
 667         xfs_buf_t               *bp,
 668         void                    *mem,
 669         size_t                  len)
 670 {
 671         int                     rval;
 672         int                     i = 0;
 673         unsigned long           pageaddr;
 674         unsigned long           offset;
 675         size_t                  buflen;
 676         int                     page_count;
 677
 678         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
 679         offset = (unsigned long)mem - pageaddr;
 680         buflen = PAGE_CACHE_ALIGN(len + offset);
 681         page_count = buflen >> PAGE_CACHE_SHIFT;
 682
 683         /* Free any previous set of page pointers */
 684         if (bp->b_pages)
 685                 _xfs_buf_free_pages(bp);
 686
 687         bp->b_pages = NULL;
 688         bp->b_addr = mem;
 689
 690         rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
 691         if (rval)
 692                 return rval;
 693
 694         bp->b_offset = offset;
 695
 696         for (i = 0; i < bp->b_page_count; i++) {
 697                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 698                 pageaddr += PAGE_CACHE_SIZE;
 699         }
 700
 701         bp->b_count_desired = len;
 702         bp->b_buffer_length = buflen;
 703         bp->b_flags |= XBF_MAPPED;
 704         bp->b_flags &= ~_XBF_PAGE_LOCKED;
 705
 706         return 0;
 707 }
 708
 709 xfs_buf_t *
 710 xfs_buf_get_noaddr(
 711         size_t                  len,
 712         xfs_buftarg_t           *target)
 713 {
 714         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 715         int                     error, i;
 716         xfs_buf_t               *bp;
 717
 718         bp = xfs_buf_allocate(0);
 719         if (unlikely(bp == NULL))
 720                 goto fail;
 721         _xfs_buf_initialize(bp, target, 0, len, 0);
 722
 723         error = _xfs_buf_get_pages(bp, page_count, 0);
 724         if (error)
 725                 goto fail_free_buf;
 726
 727         for (i = 0; i < page_count; i++) {
 728                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
 729                 if (!bp->b_pages[i])
 730                         goto fail_free_mem;
 731         }
 732         bp->b_flags |= _XBF_PAGES;
 733
 734         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 735         if (unlikely(error)) {
 736                 printk(KERN_WARNING "%s: failed to map pages\n",
 737                                 __func__);
 738                 goto fail_free_mem;
 739         }
 740
 741         xfs_buf_unlock(bp);
 742
 743         trace_xfs_buf_get_noaddr(bp, _RET_IP_);
 744         return bp;
 745
 746  fail_free_mem:
 747         while (--i >= 0)
 748                 __free_page(bp->b_pages[i]);
 749         _xfs_buf_free_pages(bp);
 750  fail_free_buf:
 751         xfs_buf_deallocate(bp);
 752  fail:
 753         return NULL;
 754 }
 755
 756 /*
 757  *      Increment reference count on buffer, to hold the buffer concurrently
 758  *      with another thread which may release (free) the buffer asynchronously.
 759  *      Must hold the buffer already to call this function.
 760  */
 761 void
 762 xfs_buf_hold(
 763         xfs_buf_t               *bp)
 764 {
 765         trace_xfs_buf_hold(bp, _RET_IP_);
 766         atomic_inc(&bp->b_hold);
 767 }
 768
 769 /*
 770  *      Releases a hold on the specified buffer.  If the
 771  *      the hold count is 1, calls xfs_buf_free.
 772  */
 773 void
 774 xfs_buf_rele(
 775         xfs_buf_t               *bp)
 776 {
 777         xfs_bufhash_t           *hash = bp->b_hash;
 778
 779         trace_xfs_buf_rele(bp, _RET_IP_);
 780
 781         if (unlikely(!hash)) {
 782                 ASSERT(!bp->b_relse);
 783                 if (atomic_dec_and_test(&bp->b_hold))
 784                         xfs_buf_free(bp);
 785                 return;
 786         }
 787
 788         ASSERT(atomic_read(&bp->b_hold) > 0);
 789         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
 790                 if (bp->b_relse) {
 791                         atomic_inc(&bp->b_hold);
 792                         spin_unlock(&hash->bh_lock);
 793                         (*(bp->b_relse)) (bp);
 794                 } else if (bp->b_flags & XBF_FS_MANAGED) {
 795                         spin_unlock(&hash->bh_lock);
 796                 } else {
 797                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 798                         list_del_init(&bp->b_hash_list);
 799                         spin_unlock(&hash->bh_lock);
 800                         xfs_buf_free(bp);
 801                 }
 802         }
 803 }
 804
 805
 806 /*
 807  *      Mutual exclusion on buffers.  Locking model:
 808  *
 809  *      Buffers associated with inodes for which buffer locking
 810  *      is not enabled are not protected by semaphores, and are
 811  *      assumed to be exclusively owned by the caller.  There is a
 812  *      spinlock in the buffer, used by the caller when concurrent
 813  *      access is possible.
 814  */
 815
 816 /*
 817  *      Locks a buffer object, if it is not already locked.
 818  *      Note that this in no way locks the underlying pages, so it is only
 819  *      useful for synchronizing concurrent use of buffer objects, not for
 820  *      synchronizing independent access to the underlying pages.
 821  */
 822 int
 823 xfs_buf_cond_lock(
 824         xfs_buf_t               *bp)
 825 {
 826         int                     locked;
 827
 828         locked = down_trylock(&bp->b_sema) == 0;
 829         if (locked)
 830                 XB_SET_OWNER(bp);
 831
 832         trace_xfs_buf_cond_lock(bp, _RET_IP_);
 833         return locked ? 0 : -EBUSY;
 834 }
 835
 836 int
 837 xfs_buf_lock_value(
 838         xfs_buf_t               *bp)
 839 {
 840         return bp->b_sema.count;
 841 }
 842
 843 /*
 844  *      Locks a buffer object.
 845  *      Note that this in no way locks the underlying pages, so it is only
 846  *      useful for synchronizing concurrent use of buffer objects, not for
 847  *      synchronizing independent access to the underlying pages.
 848  *
 849  *      If we come across a stale, pinned, locked buffer, we know that we
 850  *      are being asked to lock a buffer that has been reallocated. Because
 851  *      it is pinned, we know that the log has not been pushed to disk and
 852  *      hence it will still be locked. Rather than sleeping until someone
 853  *      else pushes the log, push it ourselves before trying to get the lock.
 854  */
 855 void
 856 xfs_buf_lock(
 857         xfs_buf_t               *bp)
 858 {
 859         trace_xfs_buf_lock(bp, _RET_IP_);
 860
 861         if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 862                 xfs_log_force(bp->b_mount, 0);
 863         if (atomic_read(&bp->b_io_remaining))
 864                 blk_run_address_space(bp->b_target->bt_mapping);
 865         down(&bp->b_sema);
 866         XB_SET_OWNER(bp);
 867
 868         trace_xfs_buf_lock_done(bp, _RET_IP_);
 869 }
 870
 871 /*
 872  *      Releases the lock on the buffer object.
 873  *      If the buffer is marked delwri but is not queued, do so before we
 874  *      unlock the buffer as we need to set flags correctly.  We also need to
 875  *      take a reference for the delwri queue because the unlocker is going to
 876  *      drop their's and they don't know we just queued it.
 877  */
 878 void
 879 xfs_buf_unlock(
 880         xfs_buf_t               *bp)
 881 {
 882         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 883                 atomic_inc(&bp->b_hold);
 884                 bp->b_flags |= XBF_ASYNC;
 885                 xfs_buf_delwri_queue(bp, 0);
 886         }
 887
 888         XB_CLEAR_OWNER(bp);
 889         up(&bp->b_sema);
 890
 891         trace_xfs_buf_unlock(bp, _RET_IP_);
 892 }
 893
 894 STATIC void
 895 xfs_buf_wait_unpin(
 896         xfs_buf_t               *bp)
 897 {
 898         DECLARE_WAITQUEUE       (wait, current);
 899
 900         if (atomic_read(&bp->b_pin_count) == 0)
 901                 return;
 902
 903         add_wait_queue(&bp->b_waiters, &wait);
 904         for (;;) {
 905                 set_current_state(TASK_UNINTERRUPTIBLE);
 906                 if (atomic_read(&bp->b_pin_count) == 0)
 907                         break;
 908                 if (atomic_read(&bp->b_io_remaining))
 909                         blk_run_address_space(bp->b_target->bt_mapping);
 910                 schedule();
 911         }
 912         remove_wait_queue(&bp->b_waiters, &wait);
 913         set_current_state(TASK_RUNNING);
 914 }
 915
 916 /*
 917  *      Buffer Utility Routines
 918  */
 919
 920 STATIC void
 921 xfs_buf_iodone_work(
 922         struct work_struct      *work)
 923 {
 924         xfs_buf_t               *bp =
 925                 container_of(work, xfs_buf_t, b_iodone_work);
 926
 927         /*
 928          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
 929          * ordered flag and reissue them.  Because we can't tell the higher
 930          * layers directly that they should not issue ordered I/O anymore, they
 931          * need to check if the _XFS_BARRIER_FAILED flag was set during I/O completion.
 932          */
 933         if ((bp->b_error == EOPNOTSUPP) &&
 934             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
 935                 trace_xfs_buf_ordered_retry(bp, _RET_IP_);
 936                 bp->b_flags &= ~XBF_ORDERED;
 937                 bp->b_flags |= _XFS_BARRIER_FAILED;
 938                 xfs_buf_iorequest(bp);
 939         } else if (bp->b_iodone)
 940                 (*(bp->b_iodone))(bp);
 941         else if (bp->b_flags & XBF_ASYNC)
 942                 xfs_buf_relse(bp);
 943 }
 944
 945 void
 946 xfs_buf_ioend(
 947         xfs_buf_t               *bp,
 948         int                     schedule)
 949 {
 950         trace_xfs_buf_iodone(bp, _RET_IP_);
 951
 952         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
 953         if (bp->b_error == 0)
 954                 bp->b_flags |= XBF_DONE;
 955
 956         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
 957                 if (schedule) {
 958                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
 959                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
 960                 } else {
 961                         xfs_buf_iodone_work(&bp->b_iodone_work);
 962                 }
 963         } else {
 964                 complete(&bp->b_iowait);
 965         }
 966 }
 967
 968 void
 969 xfs_buf_ioerror(
 970         xfs_buf_t               *bp,
 971         int                     error)
 972 {
 973         ASSERT(error >= 0 && error <= 0xffff);
 974         bp->b_error = (unsigned short)error;
 975         trace_xfs_buf_ioerror(bp, error, _RET_IP_);
 976 }
 977
 978 int
 979 xfs_bwrite(
 980         struct xfs_mount        *mp,
 981         struct xfs_buf          *bp)
 982 {
 983         int                     error;
 984
 985         bp->b_mount = mp;
 986         bp->b_flags |= XBF_WRITE;
 987         bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
 988
 989         xfs_buf_delwri_dequeue(bp);
 990         xfs_bdstrat_cb(bp);
 991
 992         error = xfs_buf_iowait(bp);
 993         if (error)
 994                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
 995         xfs_buf_relse(bp);
 996         return error;
 997 }
 998
 999 void
1000 xfs_bdwrite(
1001         void                    *mp,
1002         struct xfs_buf          *bp)
1003 {
1004         trace_xfs_buf_bdwrite(bp, _RET_IP_);
1005
1006         bp->b_mount = mp;
1007
1008         bp->b_flags &= ~XBF_READ;
1009         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1010
1011         xfs_buf_delwri_queue(bp, 1);
1012 }
1013
1014 /*
1015  * Called when we want to stop a buffer from getting written or read.
1016  * We attach the EIO error, muck with its flags, and call biodone
1017  * so that the proper iodone callbacks get called.
1018  */
1019 STATIC int
1020 xfs_bioerror(
1021         xfs_buf_t *bp)
1022 {
1023 #ifdef XFSERRORDEBUG
1024         ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1025 #endif
1026
1027         /*
1028          * No need to wait until the buffer is unpinned, we aren't flushing it.
1029          */
1030         XFS_BUF_ERROR(bp, EIO);
1031
1032         /*
1033          * We're calling biodone, so delete XBF_DONE flag.
1034          */
1035         XFS_BUF_UNREAD(bp);
1036         XFS_BUF_UNDELAYWRITE(bp);
1037         XFS_BUF_UNDONE(bp);
1038         XFS_BUF_STALE(bp);
1039
1040         xfs_biodone(bp);
1041
1042         return EIO;
1043 }
1044
1045 /*
1046  * Same as xfs_bioerror, except that we are releasing the buffer
1047  * here ourselves, and avoiding the biodone call.
1048  * This is meant for userdata errors; metadata bufs come with
1049  * iodone functions attached, so that we can track down errors.
1050  */
1051 STATIC int
1052 xfs_bioerror_relse(
1053         struct xfs_buf  *bp)
1054 {
1055         int64_t         fl = XFS_BUF_BFLAGS(bp);
1056         /*
1057          * No need to wait until the buffer is unpinned.
1058          * We aren't flushing it.
1059          *
1060          * chunkhold expects B_DONE to be set, whether
1061          * we actually finish the I/O or not. We don't want to
1062          * change that interface.
1063          */
1064         XFS_BUF_UNREAD(bp);
1065         XFS_BUF_UNDELAYWRITE(bp);
1066         XFS_BUF_DONE(bp);
1067         XFS_BUF_STALE(bp);
1068         XFS_BUF_CLR_IODONE_FUNC(bp);
1069         if (!(fl & XBF_ASYNC)) {
1070                 /*
1071                  * Mark b_error and B_ERROR _both_.
1072                  * Lot's of chunkcache code assumes that.
1073                  * There's no reason to mark error for
1074                  * ASYNC buffers.
1075                  */
1076                 XFS_BUF_ERROR(bp, EIO);
1077                 XFS_BUF_FINISH_IOWAIT(bp);
1078         } else {
1079                 xfs_buf_relse(bp);
1080         }
1081
1082         return EIO;
1083 }
1084
1085
1086 /*
1087  * All xfs metadata buffers except log state machine buffers
1088  * get this attached as their b_bdstrat callback function.
1089  * This is so that we can catch a buffer
1090  * after prematurely unpinning it to forcibly shutdown the filesystem.
1091  */
1092 int
1093 xfs_bdstrat_cb(
1094         struct xfs_buf  *bp)
1095 {
1096         if (XFS_FORCED_SHUTDOWN(bp->b_mount)) {
1097                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1098                 /*
1099                  * Metadata write that didn't get logged but
1100                  * written delayed anyway. These aren't associated
1101                  * with a transaction, and can be ignored.
1102                  */
1103                 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1104                         return xfs_bioerror_relse(bp);
1105                 else
1106                         return xfs_bioerror(bp);
1107         }
1108
1109         xfs_buf_iorequest(bp);
1110         return 0;
1111 }
1112
1113 /*
1114  * Wrapper around bdstrat so that we can stop data from going to disk in case
1115  * we are shutting down the filesystem.  Typically user data goes thru this
1116  * path; one of the exceptions is the superblock.
1117  */
1118 void
1119 xfsbdstrat(
1120         struct xfs_mount        *mp,
1121         struct xfs_buf          *bp)
1122 {
1123         if (XFS_FORCED_SHUTDOWN(mp)) {
1124                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1125                 xfs_bioerror_relse(bp);
1126                 return;
1127         }
1128
1129         xfs_buf_iorequest(bp);
1130 }
1131
1132 STATIC void
1133 _xfs_buf_ioend(
1134         xfs_buf_t               *bp,
1135         int                     schedule)
1136 {
1137         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1138                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1139                 xfs_buf_ioend(bp, schedule);
1140         }
1141 }
1142
1143 STATIC void
1144 xfs_buf_bio_end_io(
1145         struct bio              *bio,
1146         int                     error)
1147 {
1148         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1149         unsigned int            blocksize = bp->b_target->bt_bsize;
1150         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1151
1152         xfs_buf_ioerror(bp, -error);
1153
1154         if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1155                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1156
1157         do {
1158                 struct page     *page = bvec->bv_page;
1159
1160                 ASSERT(!PagePrivate(page));
1161                 if (unlikely(bp->b_error)) {
1162                         if (bp->b_flags & XBF_READ)
1163                                 ClearPageUptodate(page);
1164                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1165                         SetPageUptodate(page);
1166                 } else if (!PagePrivate(page) &&
1167                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1168                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1169                 }
1170
1171                 if (--bvec >= bio->bi_io_vec)
1172                         prefetchw(&bvec->bv_page->flags);
1173
1174                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1175                         unlock_page(page);
1176         } while (bvec >= bio->bi_io_vec);
1177
1178         _xfs_buf_ioend(bp, 1);
1179         bio_put(bio);
1180 }
1181
1182 STATIC void
1183 _xfs_buf_ioapply(
1184         xfs_buf_t               *bp)
1185 {
1186         int                     rw, map_i, total_nr_pages, nr_pages;
1187         struct bio              *bio;
1188         int                     offset = bp->b_offset;
1189         int                     size = bp->b_count_desired;
1190         sector_t                sector = bp->b_bn;
1191         unsigned int            blocksize = bp->b_target->bt_bsize;
1192
1193         total_nr_pages = bp->b_page_count;
1194         map_i = 0;
1195
1196         if (bp->b_flags & XBF_ORDERED) {
1197                 ASSERT(!(bp->b_flags & XBF_READ));
1198                 rw = WRITE_BARRIER;
1199         } else if (bp->b_flags & XBF_LOG_BUFFER) {
1200                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1201                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1202                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1203         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1204                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1205                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1206                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1207         } else {
1208                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1209                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1210         }
1211
1212         /* Special code path for reading a sub page size buffer in --
1213          * we populate up the whole page, and hence the other metadata
1214          * in the same page.  This optimization is only valid when the
1215          * filesystem block size is not smaller than the page size.
1216          */
1217         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1218             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1219               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1220             (blocksize >= PAGE_CACHE_SIZE)) {
1221                 bio = bio_alloc(GFP_NOIO, 1);
1222
1223                 bio->bi_bdev = bp->b_target->bt_bdev;
1224                 bio->bi_sector = sector - (offset >> BBSHIFT);
1225                 bio->bi_end_io = xfs_buf_bio_end_io;
1226                 bio->bi_private = bp;
1227
1228                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1229                 size = 0;
1230
1231                 atomic_inc(&bp->b_io_remaining);
1232
1233                 goto submit_io;
1234         }
1235
1236 next_chunk:
1237         atomic_inc(&bp->b_io_remaining);
1238         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1239         if (nr_pages > total_nr_pages)
1240                 nr_pages = total_nr_pages;
1241
1242         bio = bio_alloc(GFP_NOIO, nr_pages);
1243         bio->bi_bdev = bp->b_target->bt_bdev;
1244         bio->bi_sector = sector;
1245         bio->bi_end_io = xfs_buf_bio_end_io;
1246         bio->bi_private = bp;
1247
1248         for (; size && nr_pages; nr_pages--, map_i++) {
1249                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1250
1251                 if (nbytes > size)
1252                         nbytes = size;
1253
1254                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1255                 if (rbytes < nbytes)
1256                         break;
1257
1258                 offset = 0;
1259                 sector += nbytes >> BBSHIFT;
1260                 size -= nbytes;
1261                 total_nr_pages--;
1262         }
1263
1264 submit_io:
1265         if (likely(bio->bi_size)) {
1266                 if (xfs_buf_is_vmapped(bp)) {
1267                         flush_kernel_vmap_range(bp->b_addr,
1268                                                 xfs_buf_vmap_len(bp));
1269                 }
1270                 submit_bio(rw, bio);
1271                 if (size)
1272                         goto next_chunk;
1273         } else {
1274                 /*
1275                  * if we get here, no pages were added to the bio. However,
1276                  * we can't just error out here - if the pages are locked then
1277                  * we have to unlock them otherwise we can hang on a later
1278                  * access to the page.
1279                  */
1280                 xfs_buf_ioerror(bp, EIO);
1281                 if (bp->b_flags & _XBF_PAGE_LOCKED) {
1282                         int i;
1283                         for (i = 0; i < bp->b_page_count; i++)
1284                                 unlock_page(bp->b_pages[i]);
1285                 }
1286                 bio_put(bio);
1287         }
1288 }
1289
1290 int
1291 xfs_buf_iorequest(
1292         xfs_buf_t               *bp)
1293 {
1294         trace_xfs_buf_iorequest(bp, _RET_IP_);
1295
1296         if (bp->b_flags & XBF_DELWRI) {
1297                 xfs_buf_delwri_queue(bp, 1);
1298                 return 0;
1299         }
1300
1301         if (bp->b_flags & XBF_WRITE) {
1302                 xfs_buf_wait_unpin(bp);
1303         }
1304
1305         xfs_buf_hold(bp);
1306
1307         /* Set the count to 1 initially, this will stop an I/O
1308          * completion callout which happens before we have started
1309          * all the I/O from calling xfs_buf_ioend too early.
1310          */
1311         atomic_set(&bp->b_io_remaining, 1);
1312         _xfs_buf_ioapply(bp);
1313         _xfs_buf_ioend(bp, 0);
1314
1315         xfs_buf_rele(bp);
1316         return 0;
1317 }
1318
1319 /*
1320  *      Waits for I/O to complete on the buffer supplied.
1321  *      It returns immediately if no I/O is pending.
1322  *      It returns the I/O error code, if any, or 0 if there was no error.
1323  */
1324 int
1325 xfs_buf_iowait(
1326         xfs_buf_t               *bp)
1327 {
1328         trace_xfs_buf_iowait(bp, _RET_IP_);
1329
1330         if (atomic_read(&bp->b_io_remaining))
1331                 blk_run_address_space(bp->b_target->bt_mapping);
1332         wait_for_completion(&bp->b_iowait);
1333
1334         trace_xfs_buf_iowait_done(bp, _RET_IP_);
1335         return bp->b_error;
1336 }
1337
1338 xfs_caddr_t
1339 xfs_buf_offset(
1340         xfs_buf_t               *bp,
1341         size_t                  offset)
1342 {
1343         struct page             *page;
1344
1345         if (bp->b_flags & XBF_MAPPED)
1346                 return XFS_BUF_PTR(bp) + offset;
1347
1348         offset += bp->b_offset;
1349         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1350         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1351 }
1352
1353 /*
1354  *      Move data into or out of a buffer.
1355  */
1356 void
1357 xfs_buf_iomove(
1358         xfs_buf_t               *bp,    /* buffer to process            */
1359         size_t                  boff,   /* starting buffer offset       */
1360         size_t                  bsize,  /* length to copy               */
1361         void                    *data,  /* data address                 */
1362         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1363 {
1364         size_t                  bend, cpoff, csize;
1365         struct page             *page;
1366
1367         bend = boff + bsize;
1368         while (boff < bend) {
1369                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1370                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1371                 csize = min_t(size_t,
1372                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1373
1374                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1375
1376                 switch (mode) {
1377                 case XBRW_ZERO:
1378                         memset(page_address(page) + cpoff, 0, csize);
1379                         break;
1380                 case XBRW_READ:
1381                         memcpy(data, page_address(page) + cpoff, csize);
1382                         break;
1383                 case XBRW_WRITE:
1384                         memcpy(page_address(page) + cpoff, data, csize);
1385                 }
1386
1387                 boff += csize;
1388                 data += csize;
1389         }
1390 }
1391
1392 /*
1393  *      Handling of buffer targets (buftargs).
1394  */
1395
1396 /*
1397  *      Wait for any bufs with callbacks that have been submitted but
1398  *      have not yet returned... walk the hash list for the target.
1399  */
1400 void
1401 xfs_wait_buftarg(
1402         xfs_buftarg_t   *btp)
1403 {
1404         xfs_buf_t       *bp, *n;
1405         xfs_bufhash_t   *hash;
1406         uint            i;
1407
1408         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1409                 hash = &btp->bt_hash[i];
1410 again:
1411                 spin_lock(&hash->bh_lock);
1412                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1413                         ASSERT(btp == bp->b_target);
1414                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1415                                 spin_unlock(&hash->bh_lock);
1416                                 /*
1417                                  * Catch superblock reference count leaks
1418                                  * immediately
1419                                  */
1420                                 BUG_ON(bp->b_bn == 0);
1421                                 delay(100);
1422                                 goto again;
1423                         }
1424                 }
1425                 spin_unlock(&hash->bh_lock);
1426         }
1427 }
1428
1429 /*
1430  *      Allocate buffer hash table for a given target.
1431  *      For devices containing metadata (i.e. not the log/realtime devices)
1432  *      we need to allocate a much larger hash table.
1433  */
1434 STATIC void
1435 xfs_alloc_bufhash(
1436         xfs_buftarg_t           *btp,
1437         int                     external)
1438 {
1439         unsigned int            i;
1440
1441         btp->bt_hashshift = external ? 3 : 12;  /* 8 or 4096 buckets */
1442         btp->bt_hash = kmem_zalloc_large((1 << btp->bt_hashshift) *
1443                                          sizeof(xfs_bufhash_t));
1444         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1445                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1446                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1447         }
1448 }
1449
1450 STATIC void
1451 xfs_free_bufhash(
1452         xfs_buftarg_t           *btp)
1453 {
1454         kmem_free_large(btp->bt_hash);
1455         btp->bt_hash = NULL;
1456 }
1457
1458 /*
1459  *      buftarg list for delwrite queue processing
1460  */
1461 static LIST_HEAD(xfs_buftarg_list);
1462 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1463
1464 STATIC void
1465 xfs_register_buftarg(
1466         xfs_buftarg_t           *btp)
1467 {
1468         spin_lock(&xfs_buftarg_lock);
1469         list_add(&btp->bt_list, &xfs_buftarg_list);
1470         spin_unlock(&xfs_buftarg_lock);
1471 }
1472
1473 STATIC void
1474 xfs_unregister_buftarg(
1475         xfs_buftarg_t           *btp)
1476 {
1477         spin_lock(&xfs_buftarg_lock);
1478         list_del(&btp->bt_list);
1479         spin_unlock(&xfs_buftarg_lock);
1480 }
1481
1482 void
1483 xfs_free_buftarg(
1484         struct xfs_mount        *mp,
1485         struct xfs_buftarg      *btp)
1486 {
1487         xfs_flush_buftarg(btp, 1);
1488         if (mp->m_flags & XFS_MOUNT_BARRIER)
1489                 xfs_blkdev_issue_flush(btp);
1490         xfs_free_bufhash(btp);
1491         iput(btp->bt_mapping->host);
1492
1493         /* Unregister the buftarg first so that we don't get a
1494          * wakeup finding a non-existent task
1495          */
1496         xfs_unregister_buftarg(btp);
1497         kthread_stop(btp->bt_task);
1498
1499         kmem_free(btp);
1500 }
1501
1502 STATIC int
1503 xfs_setsize_buftarg_flags(
1504         xfs_buftarg_t           *btp,
1505         unsigned int            blocksize,
1506         unsigned int            sectorsize,
1507         int                     verbose)
1508 {
1509         btp->bt_bsize = blocksize;
1510         btp->bt_sshift = ffs(sectorsize) - 1;
1511         btp->bt_smask = sectorsize - 1;
1512
1513         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1514                 printk(KERN_WARNING
1515                         "XFS: Cannot set_blocksize to %u on device %s\n",
1516                         sectorsize, XFS_BUFTARG_NAME(btp));
1517                 return EINVAL;
1518         }
1519
1520         if (verbose &&
1521             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1522                 printk(KERN_WARNING
1523                         "XFS: %u byte sectors in use on device %s.  "
1524                         "This is suboptimal; %u or greater is ideal.\n",
1525                         sectorsize, XFS_BUFTARG_NAME(btp),
1526                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1527         }
1528
1529         return 0;
1530 }
1531
1532 /*
1533  *      When allocating the initial buffer target we have not yet
1534  *      read in the superblock, so don't know what sized sectors
1535  *      are being used is at this early stage.  Play safe.
1536  */
1537 STATIC int
1538 xfs_setsize_buftarg_early(
1539         xfs_buftarg_t           *btp,
1540         struct block_device     *bdev)
1541 {
1542         return xfs_setsize_buftarg_flags(btp,
1543                         PAGE_CACHE_SIZE, bdev_logical_block_size(bdev), 0);
1544 }
1545
1546 int
1547 xfs_setsize_buftarg(
1548         xfs_buftarg_t           *btp,
1549         unsigned int            blocksize,
1550         unsigned int            sectorsize)
1551 {
1552         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1553 }
1554
1555 STATIC int
1556 xfs_mapping_buftarg(
1557         xfs_buftarg_t           *btp,
1558         struct block_device     *bdev)
1559 {
1560         struct backing_dev_info *bdi;
1561         struct inode            *inode;
1562         struct address_space    *mapping;
1563         static const struct address_space_operations mapping_aops = {
1564                 .sync_page = block_sync_page,
1565                 .migratepage = fail_migrate_page,
1566         };
1567
1568         inode = new_inode(bdev->bd_inode->i_sb);
1569         if (!inode) {
1570                 printk(KERN_WARNING
1571                         "XFS: Cannot allocate mapping inode for device %s\n",
1572                         XFS_BUFTARG_NAME(btp));
1573                 return ENOMEM;
1574         }
1575         inode->i_mode = S_IFBLK;
1576         inode->i_bdev = bdev;
1577         inode->i_rdev = bdev->bd_dev;
1578         bdi = blk_get_backing_dev_info(bdev);
1579         if (!bdi)
1580                 bdi = &default_backing_dev_info;
1581         mapping = &inode->i_data;
1582         mapping->a_ops = &mapping_aops;
1583         mapping->backing_dev_info = bdi;
1584         mapping_set_gfp_mask(mapping, GFP_NOFS);
1585         btp->bt_mapping = mapping;
1586         return 0;
1587 }
1588
1589 STATIC int
1590 xfs_alloc_delwrite_queue(
1591         xfs_buftarg_t           *btp,
1592         const char              *fsname)
1593 {
1594         int     error = 0;
1595
1596         INIT_LIST_HEAD(&btp->bt_list);
1597         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1598         spin_lock_init(&btp->bt_delwrite_lock);
1599         btp->bt_flags = 0;
1600         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1601         if (IS_ERR(btp->bt_task)) {
1602                 error = PTR_ERR(btp->bt_task);
1603                 goto out_error;
1604         }
1605         xfs_register_buftarg(btp);
1606 out_error:
1607         return error;
1608 }
1609
1610 xfs_buftarg_t *
1611 xfs_alloc_buftarg(
1612         struct block_device     *bdev,
1613         int                     external,
1614         const char              *fsname)
1615 {
1616         xfs_buftarg_t           *btp;
1617
1618         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1619
1620         btp->bt_dev =  bdev->bd_dev;
1621         btp->bt_bdev = bdev;
1622         if (xfs_setsize_buftarg_early(btp, bdev))
1623                 goto error;
1624         if (xfs_mapping_buftarg(btp, bdev))
1625                 goto error;
1626         if (xfs_alloc_delwrite_queue(btp, fsname))
1627                 goto error;
1628         xfs_alloc_bufhash(btp, external);
1629         return btp;
1630
1631 error:
1632         kmem_free(btp);
1633         return NULL;
1634 }
1635
1636
1637 /*
1638  *      Delayed write buffer handling
1639  */
1640 STATIC void
1641 xfs_buf_delwri_queue(
1642         xfs_buf_t               *bp,
1643         int                     unlock)
1644 {
1645         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1646         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1647
1648         trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1649
1650         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1651
1652         spin_lock(dwlk);
1653         /* If already in the queue, dequeue and place at tail */
1654         if (!list_empty(&bp->b_list)) {
1655                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1656                 if (unlock)
1657                         atomic_dec(&bp->b_hold);
1658                 list_del(&bp->b_list);
1659         }
1660
1661         if (list_empty(dwq)) {
1662                 /* start xfsbufd as it is about to have something to do */
1663                 wake_up_process(bp->b_target->bt_task);
1664         }
1665
1666         bp->b_flags |= _XBF_DELWRI_Q;
1667         list_add_tail(&bp->b_list, dwq);
1668         bp->b_queuetime = jiffies;
1669         spin_unlock(dwlk);
1670
1671         if (unlock)
1672                 xfs_buf_unlock(bp);
1673 }
1674
1675 void
1676 xfs_buf_delwri_dequeue(
1677         xfs_buf_t               *bp)
1678 {
1679         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1680         int                     dequeued = 0;
1681
1682         spin_lock(dwlk);
1683         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1684                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1685                 list_del_init(&bp->b_list);
1686                 dequeued = 1;
1687         }
1688         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1689         spin_unlock(dwlk);
1690
1691         if (dequeued)
1692                 xfs_buf_rele(bp);
1693
1694         trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1695 }
1696
1697 /*
1698  * If a delwri buffer needs to be pushed before it has aged out, then promote
1699  * it to the head of the delwri queue so that it will be flushed on the next
1700  * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1701  * than the age currently needed to flush the buffer. Hence the next time the
1702  * xfsbufd sees it is guaranteed to be considered old enough to flush.
1703  */
1704 void
1705 xfs_buf_delwri_promote(
1706         struct xfs_buf  *bp)
1707 {
1708         struct xfs_buftarg *btp = bp->b_target;
1709         long            age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1710
1711         ASSERT(bp->b_flags & XBF_DELWRI);
1712         ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1713
1714         /*
1715          * Check the buffer age before locking the delayed write queue as we
1716          * don't need to promote buffers that are already past the flush age.
1717          */
1718         if (bp->b_queuetime < jiffies - age)
1719                 return;
1720         bp->b_queuetime = jiffies - age;
1721         spin_lock(&btp->bt_delwrite_lock);
1722         list_move(&bp->b_list, &btp->bt_delwrite_queue);
1723         spin_unlock(&btp->bt_delwrite_lock);
1724 }
1725
1726 STATIC void
1727 xfs_buf_runall_queues(
1728         struct workqueue_struct *queue)
1729 {
1730         flush_workqueue(queue);
1731 }
1732
1733 STATIC int
1734 xfsbufd_wakeup(
1735         struct shrinker         *shrink,
1736         int                     priority,
1737         gfp_t                   mask)
1738 {
1739         xfs_buftarg_t           *btp;
1740
1741         spin_lock(&xfs_buftarg_lock);
1742         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1743                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1744                         continue;
1745                 if (list_empty(&btp->bt_delwrite_queue))
1746                         continue;
1747                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1748                 wake_up_process(btp->bt_task);
1749         }
1750         spin_unlock(&xfs_buftarg_lock);
1751         return 0;
1752 }
1753
1754 /*
1755  * Move as many buffers as specified to the supplied list
1756  * idicating if we skipped any buffers to prevent deadlocks.
1757  */
1758 STATIC int
1759 xfs_buf_delwri_split(
1760         xfs_buftarg_t   *target,
1761         struct list_head *list,
1762         unsigned long   age)
1763 {
1764         xfs_buf_t       *bp, *n;
1765         struct list_head *dwq = &target->bt_delwrite_queue;
1766         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1767         int             skipped = 0;
1768         int             force;
1769
1770         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1771         INIT_LIST_HEAD(list);
1772         spin_lock(dwlk);
1773         list_for_each_entry_safe(bp, n, dwq, b_list) {
1774                 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1775                 ASSERT(bp->b_flags & XBF_DELWRI);
1776
1777                 if (!XFS_BUF_ISPINNED(bp) && !xfs_buf_cond_lock(bp)) {
1778                         if (!force &&
1779                             time_before(jiffies, bp->b_queuetime + age)) {
1780                                 xfs_buf_unlock(bp);
1781                                 break;
1782                         }
1783
1784                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1785                                          _XBF_RUN_QUEUES);
1786                         bp->b_flags |= XBF_WRITE;
1787                         list_move_tail(&bp->b_list, list);
1788                 } else
1789                         skipped++;
1790         }
1791         spin_unlock(dwlk);
1792
1793         return skipped;
1794
1795 }
1796
1797 /*
1798  * Compare function is more complex than it needs to be because
1799  * the return value is only 32 bits and we are doing comparisons
1800  * on 64 bit values
1801  */
1802 static int
1803 xfs_buf_cmp(
1804         void            *priv,
1805         struct list_head *a,
1806         struct list_head *b)
1807 {
1808         struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1809         struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1810         xfs_daddr_t             diff;
1811
1812         diff = ap->b_bn - bp->b_bn;
1813         if (diff < 0)
1814                 return -1;
1815         if (diff > 0)
1816                 return 1;
1817         return 0;
1818 }
1819
1820 void
1821 xfs_buf_delwri_sort(
1822         xfs_buftarg_t   *target,
1823         struct list_head *list)
1824 {
1825         list_sort(NULL, list, xfs_buf_cmp);
1826 }
1827
1828 STATIC int
1829 xfsbufd(
1830         void            *data)
1831 {
1832         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1833
1834         current->flags |= PF_MEMALLOC;
1835
1836         set_freezable();
1837
1838         do {
1839                 long    age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1840                 long    tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1841                 int     count = 0;
1842                 struct list_head tmp;
1843
1844                 if (unlikely(freezing(current))) {
1845                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1846                         refrigerator();
1847                 } else {
1848                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1849                 }
1850
1851                 /* sleep for a long time if there is nothing to do. */
1852                 if (list_empty(&target->bt_delwrite_queue))
1853                         tout = MAX_SCHEDULE_TIMEOUT;
1854                 schedule_timeout_interruptible(tout);
1855
1856                 xfs_buf_delwri_split(target, &tmp, age);
1857                 list_sort(NULL, &tmp, xfs_buf_cmp);
1858                 while (!list_empty(&tmp)) {
1859                         struct xfs_buf *bp;
1860                         bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1861                         list_del_init(&bp->b_list);
1862                         xfs_bdstrat_cb(bp);
1863                         count++;
1864                 }
1865                 if (count)
1866                         blk_run_address_space(target->bt_mapping);
1867
1868         } while (!kthread_should_stop());
1869
1870         return 0;
1871 }
1872
1873 /*
1874  *      Go through all incore buffers, and release buffers if they belong to
1875  *      the given device. This is used in filesystem error handling to
1876  *      preserve the consistency of its metadata.
1877  */
1878 int
1879 xfs_flush_buftarg(
1880         xfs_buftarg_t   *target,
1881         int             wait)
1882 {
1883         xfs_buf_t       *bp;
1884         int             pincount = 0;
1885         LIST_HEAD(tmp_list);
1886         LIST_HEAD(wait_list);
1887
1888         xfs_buf_runall_queues(xfsconvertd_workqueue);
1889         xfs_buf_runall_queues(xfsdatad_workqueue);
1890         xfs_buf_runall_queues(xfslogd_workqueue);
1891
1892         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1893         pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1894
1895         /*
1896          * Dropped the delayed write list lock, now walk the temporary list.
1897          * All I/O is issued async and then if we need to wait for completion
1898          * we do that after issuing all the IO.
1899          */
1900         list_sort(NULL, &tmp_list, xfs_buf_cmp);
1901         while (!list_empty(&tmp_list)) {
1902                 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1903                 ASSERT(target == bp->b_target);
1904                 list_del_init(&bp->b_list);
1905                 if (wait) {
1906                         bp->b_flags &= ~XBF_ASYNC;
1907                         list_add(&bp->b_list, &wait_list);
1908                 }
1909                 xfs_bdstrat_cb(bp);
1910         }
1911
1912         if (wait) {
1913                 /* Expedite and wait for IO to complete. */
1914                 blk_run_address_space(target->bt_mapping);
1915                 while (!list_empty(&wait_list)) {
1916                         bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1917
1918                         list_del_init(&bp->b_list);
1919                         xfs_iowait(bp);
1920                         xfs_buf_relse(bp);
1921                 }
1922         }
1923
1924         return pincount;
1925 }
1926
1927 int __init
1928 xfs_buf_init(void)
1929 {
1930         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1931                                                 KM_ZONE_HWALIGN, NULL);
1932         if (!xfs_buf_zone)
1933                 goto out;
1934
1935         xfslogd_workqueue = alloc_workqueue("xfslogd",
1936                                         WQ_RESCUER | WQ_HIGHPRI, 1);
1937         if (!xfslogd_workqueue)
1938                 goto out_free_buf_zone;
1939
1940         xfsdatad_workqueue = create_workqueue("xfsdatad");
1941         if (!xfsdatad_workqueue)
1942                 goto out_destroy_xfslogd_workqueue;
1943
1944         xfsconvertd_workqueue = create_workqueue("xfsconvertd");
1945         if (!xfsconvertd_workqueue)
1946                 goto out_destroy_xfsdatad_workqueue;
1947
1948         register_shrinker(&xfs_buf_shake);
1949         return 0;
1950
1951  out_destroy_xfsdatad_workqueue:
1952         destroy_workqueue(xfsdatad_workqueue);
1953  out_destroy_xfslogd_workqueue:
1954         destroy_workqueue(xfslogd_workqueue);
1955  out_free_buf_zone:
1956         kmem_zone_destroy(xfs_buf_zone);
1957  out:
1958         return -ENOMEM;
1959 }
1960
1961 void
1962 xfs_buf_terminate(void)
1963 {
1964         unregister_shrinker(&xfs_buf_shake);
1965         destroy_workqueue(xfsconvertd_workqueue);
1966         destroy_workqueue(xfsdatad_workqueue);
1967         destroy_workqueue(xfslogd_workqueue);
1968         kmem_zone_destroy(xfs_buf_zone);
1969 }
1970
1971 #ifdef CONFIG_KDB_MODULES
1972 struct list_head *
1973 xfs_get_buftarg_list(void)
1974 {
1975         return &xfs_buftarg_list;
1976 }
1977 #endif