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         RB_CLEAR_NODE(&bp->b_rbnode);
 192         sema_init(&bp->b_sema, 0); /* 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         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
 266                 uint            i;
 267
 268                 if (xfs_buf_is_vmapped(bp))
 269                         vm_unmap_ram(bp->b_addr - bp->b_offset,
 270                                         bp->b_page_count);
 271
 272                 for (i = 0; i < bp->b_page_count; i++) {
 273                         struct page     *page = bp->b_pages[i];
 274
 275                         if (bp->b_flags & _XBF_PAGE_CACHE)
 276                                 ASSERT(!PagePrivate(page));
 277                         page_cache_release(page);
 278                 }
 279         }
 280         _xfs_buf_free_pages(bp);
 281         xfs_buf_deallocate(bp);
 282 }
 283
 284 /*
 285  *      Finds all pages for buffer in question and builds it's page list.
 286  */
 287 STATIC int
 288 _xfs_buf_lookup_pages(
 289         xfs_buf_t               *bp,
 290         uint                    flags)
 291 {
 292         struct address_space    *mapping = bp->b_target->bt_mapping;
 293         size_t                  blocksize = bp->b_target->bt_bsize;
 294         size_t                  size = bp->b_count_desired;
 295         size_t                  nbytes, offset;
 296         gfp_t                   gfp_mask = xb_to_gfp(flags);
 297         unsigned short          page_count, i;
 298         pgoff_t                 first;
 299         xfs_off_t               end;
 300         int                     error;
 301
 302         end = bp->b_file_offset + bp->b_buffer_length;
 303         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 304
 305         error = _xfs_buf_get_pages(bp, page_count, flags);
 306         if (unlikely(error))
 307                 return error;
 308         bp->b_flags |= _XBF_PAGE_CACHE;
 309
 310         offset = bp->b_offset;
 311         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
 312
 313         for (i = 0; i < bp->b_page_count; i++) {
 314                 struct page     *page;
 315                 uint            retries = 0;
 316
 317               retry:
 318                 page = find_or_create_page(mapping, first + i, gfp_mask);
 319                 if (unlikely(page == NULL)) {
 320                         if (flags & XBF_READ_AHEAD) {
 321                                 bp->b_page_count = i;
 322                                 for (i = 0; i < bp->b_page_count; i++)
 323                                         unlock_page(bp->b_pages[i]);
 324                                 return -ENOMEM;
 325                         }
 326
 327                         /*
 328                          * This could deadlock.
 329                          *
 330                          * But until all the XFS lowlevel code is revamped to
 331                          * handle buffer allocation failures we can't do much.
 332                          */
 333                         if (!(++retries % 100))
 334                                 printk(KERN_ERR
 335                                         "XFS: possible memory allocation "
 336                                         "deadlock in %s (mode:0x%x)\n",
 337                                         __func__, gfp_mask);
 338
 339                         XFS_STATS_INC(xb_page_retries);
 340                         xfsbufd_wakeup(NULL, 0, gfp_mask);
 341                         congestion_wait(BLK_RW_ASYNC, HZ/50);
 342                         goto retry;
 343                 }
 344
 345                 XFS_STATS_INC(xb_page_found);
 346
 347                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
 348                 size -= nbytes;
 349
 350                 ASSERT(!PagePrivate(page));
 351                 if (!PageUptodate(page)) {
 352                         page_count--;
 353                         if (blocksize >= PAGE_CACHE_SIZE) {
 354                                 if (flags & XBF_READ)
 355                                         bp->b_flags |= _XBF_PAGE_LOCKED;
 356                         } else if (!PagePrivate(page)) {
 357                                 if (test_page_region(page, offset, nbytes))
 358                                         page_count++;
 359                         }
 360                 }
 361
 362                 bp->b_pages[i] = page;
 363                 offset = 0;
 364         }
 365
 366         if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
 367                 for (i = 0; i < bp->b_page_count; i++)
 368                         unlock_page(bp->b_pages[i]);
 369         }
 370
 371         if (page_count == bp->b_page_count)
 372                 bp->b_flags |= XBF_DONE;
 373
 374         return error;
 375 }
 376
 377 /*
 378  *      Map buffer into kernel address-space if nessecary.
 379  */
 380 STATIC int
 381 _xfs_buf_map_pages(
 382         xfs_buf_t               *bp,
 383         uint                    flags)
 384 {
 385         /* A single page buffer is always mappable */
 386         if (bp->b_page_count == 1) {
 387                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 388                 bp->b_flags |= XBF_MAPPED;
 389         } else if (flags & XBF_MAPPED) {
 390                 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
 391                                         -1, PAGE_KERNEL);
 392                 if (unlikely(bp->b_addr == NULL))
 393                         return -ENOMEM;
 394                 bp->b_addr += bp->b_offset;
 395                 bp->b_flags |= XBF_MAPPED;
 396         }
 397
 398         return 0;
 399 }
 400
 401 /*
 402  *      Finding and Reading Buffers
 403  */
 404
 405 /*
 406  *      Look up, and creates if absent, a lockable buffer for
 407  *      a given range of an inode.  The buffer is returned
 408  *      locked.  If other overlapping buffers exist, they are
 409  *      released before the new buffer is created and locked,
 410  *      which may imply that this call will block until those buffers
 411  *      are unlocked.  No I/O is implied by this call.
 412  */
 413 xfs_buf_t *
 414 _xfs_buf_find(
 415         xfs_buftarg_t           *btp,   /* block device target          */
 416         xfs_off_t               ioff,   /* starting offset of range     */
 417         size_t                  isize,  /* length of range              */
 418         xfs_buf_flags_t         flags,
 419         xfs_buf_t               *new_bp)
 420 {
 421         xfs_off_t               range_base;
 422         size_t                  range_length;
 423         struct xfs_perag        *pag;
 424         struct rb_node          **rbp;
 425         struct rb_node          *parent;
 426         xfs_buf_t               *bp;
 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         /* get tree root */
 436         pag = xfs_perag_get(btp->bt_mount,
 437                                 xfs_daddr_to_agno(btp->bt_mount, ioff));
 438
 439         /* walk tree */
 440         spin_lock(&pag->pag_buf_lock);
 441         rbp = &pag->pag_buf_tree.rb_node;
 442         parent = NULL;
 443         bp = NULL;
 444         while (*rbp) {
 445                 parent = *rbp;
 446                 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
 447
 448                 if (range_base < bp->b_file_offset)
 449                         rbp = &(*rbp)->rb_left;
 450                 else if (range_base > bp->b_file_offset)
 451                         rbp = &(*rbp)->rb_right;
 452                 else {
 453                         /*
 454                          * found a block offset match. If the range doesn't
 455                          * match, the only way this is allowed is if the buffer
 456                          * in the cache is stale and the transaction that made
 457                          * it stale has not yet committed. i.e. we are
 458                          * reallocating a busy extent. Skip this buffer and
 459                          * continue searching to the right for an exact match.
 460                          */
 461                         if (bp->b_buffer_length != range_length) {
 462                                 ASSERT(bp->b_flags & XBF_STALE);
 463                                 rbp = &(*rbp)->rb_right;
 464                                 continue;
 465                         }
 466                         atomic_inc(&bp->b_hold);
 467                         goto found;
 468                 }
 469         }
 470
 471         /* No match found */
 472         if (new_bp) {
 473                 _xfs_buf_initialize(new_bp, btp, range_base,
 474                                 range_length, flags);
 475                 rb_link_node(&new_bp->b_rbnode, parent, rbp);
 476                 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
 477                 /* the buffer keeps the perag reference until it is freed */
 478                 new_bp->b_pag = pag;
 479                 spin_unlock(&pag->pag_buf_lock);
 480         } else {
 481                 XFS_STATS_INC(xb_miss_locked);
 482                 spin_unlock(&pag->pag_buf_lock);
 483                 xfs_perag_put(pag);
 484         }
 485         return new_bp;
 486
 487 found:
 488         spin_unlock(&pag->pag_buf_lock);
 489         xfs_perag_put(pag);
 490
 491         /* Attempt to get the semaphore without sleeping,
 492          * if this does not work then we need to drop the
 493          * spinlock and do a hard attempt on the semaphore.
 494          */
 495         if (down_trylock(&bp->b_sema)) {
 496                 if (!(flags & XBF_TRYLOCK)) {
 497                         /* wait for buffer ownership */
 498                         xfs_buf_lock(bp);
 499                         XFS_STATS_INC(xb_get_locked_waited);
 500                 } else {
 501                         /* We asked for a trylock and failed, no need
 502                          * to look at file offset and length here, we
 503                          * know that this buffer at least overlaps our
 504                          * buffer and is locked, therefore our buffer
 505                          * either does not exist, or is this buffer.
 506                          */
 507                         xfs_buf_rele(bp);
 508                         XFS_STATS_INC(xb_busy_locked);
 509                         return NULL;
 510                 }
 511         } else {
 512                 /* trylock worked */
 513                 XB_SET_OWNER(bp);
 514         }
 515
 516         if (bp->b_flags & XBF_STALE) {
 517                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 518                 bp->b_flags &= XBF_MAPPED;
 519         }
 520
 521         trace_xfs_buf_find(bp, flags, _RET_IP_);
 522         XFS_STATS_INC(xb_get_locked);
 523         return bp;
 524 }
 525
 526 /*
 527  *      Assembles a buffer covering the specified range.
 528  *      Storage in memory for all portions of the buffer will be allocated,
 529  *      although backing storage may not be.
 530  */
 531 xfs_buf_t *
 532 xfs_buf_get(
 533         xfs_buftarg_t           *target,/* target for buffer            */
 534         xfs_off_t               ioff,   /* starting offset of range     */
 535         size_t                  isize,  /* length of range              */
 536         xfs_buf_flags_t         flags)
 537 {
 538         xfs_buf_t               *bp, *new_bp;
 539         int                     error = 0, i;
 540
 541         new_bp = xfs_buf_allocate(flags);
 542         if (unlikely(!new_bp))
 543                 return NULL;
 544
 545         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 546         if (bp == new_bp) {
 547                 error = _xfs_buf_lookup_pages(bp, flags);
 548                 if (error)
 549                         goto no_buffer;
 550         } else {
 551                 xfs_buf_deallocate(new_bp);
 552                 if (unlikely(bp == NULL))
 553                         return NULL;
 554         }
 555
 556         for (i = 0; i < bp->b_page_count; i++)
 557                 mark_page_accessed(bp->b_pages[i]);
 558
 559         if (!(bp->b_flags & XBF_MAPPED)) {
 560                 error = _xfs_buf_map_pages(bp, flags);
 561                 if (unlikely(error)) {
 562                         printk(KERN_WARNING "%s: failed to map pages\n",
 563                                         __func__);
 564                         goto no_buffer;
 565                 }
 566         }
 567
 568         XFS_STATS_INC(xb_get);
 569
 570         /*
 571          * Always fill in the block number now, the mapped cases can do
 572          * their own overlay of this later.
 573          */
 574         bp->b_bn = ioff;
 575         bp->b_count_desired = bp->b_buffer_length;
 576
 577         trace_xfs_buf_get(bp, flags, _RET_IP_);
 578         return bp;
 579
 580  no_buffer:
 581         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 582                 xfs_buf_unlock(bp);
 583         xfs_buf_rele(bp);
 584         return NULL;
 585 }
 586
 587 STATIC int
 588 _xfs_buf_read(
 589         xfs_buf_t               *bp,
 590         xfs_buf_flags_t         flags)
 591 {
 592         int                     status;
 593
 594         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
 595         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
 596
 597         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
 598                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 599         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
 600                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 601
 602         status = xfs_buf_iorequest(bp);
 603         if (status || XFS_BUF_ISERROR(bp) || (flags & XBF_ASYNC))
 604                 return status;
 605         return xfs_buf_iowait(bp);
 606 }
 607
 608 xfs_buf_t *
 609 xfs_buf_read(
 610         xfs_buftarg_t           *target,
 611         xfs_off_t               ioff,
 612         size_t                  isize,
 613         xfs_buf_flags_t         flags)
 614 {
 615         xfs_buf_t               *bp;
 616
 617         flags |= XBF_READ;
 618
 619         bp = xfs_buf_get(target, ioff, isize, flags);
 620         if (bp) {
 621                 trace_xfs_buf_read(bp, flags, _RET_IP_);
 622
 623                 if (!XFS_BUF_ISDONE(bp)) {
 624                         XFS_STATS_INC(xb_get_read);
 625                         _xfs_buf_read(bp, flags);
 626                 } else if (flags & XBF_ASYNC) {
 627                         /*
 628                          * Read ahead call which is already satisfied,
 629                          * drop the buffer
 630                          */
 631                         goto no_buffer;
 632                 } else {
 633                         /* We do not want read in the flags */
 634                         bp->b_flags &= ~XBF_READ;
 635                 }
 636         }
 637
 638         return bp;
 639
 640  no_buffer:
 641         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 642                 xfs_buf_unlock(bp);
 643         xfs_buf_rele(bp);
 644         return NULL;
 645 }
 646
 647 /*
 648  *      If we are not low on memory then do the readahead in a deadlock
 649  *      safe manner.
 650  */
 651 void
 652 xfs_buf_readahead(
 653         xfs_buftarg_t           *target,
 654         xfs_off_t               ioff,
 655         size_t                  isize)
 656 {
 657         struct backing_dev_info *bdi;
 658
 659         bdi = target->bt_mapping->backing_dev_info;
 660         if (bdi_read_congested(bdi))
 661                 return;
 662
 663         xfs_buf_read(target, ioff, isize,
 664                      XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
 665 }
 666
 667 /*
 668  * Read an uncached buffer from disk. Allocates and returns a locked
 669  * buffer containing the disk contents or nothing.
 670  */
 671 struct xfs_buf *
 672 xfs_buf_read_uncached(
 673         struct xfs_mount        *mp,
 674         struct xfs_buftarg      *target,
 675         xfs_daddr_t             daddr,
 676         size_t                  length,
 677         int                     flags)
 678 {
 679         xfs_buf_t               *bp;
 680         int                     error;
 681
 682         bp = xfs_buf_get_uncached(target, length, flags);
 683         if (!bp)
 684                 return NULL;
 685
 686         /* set up the buffer for a read IO */
 687         xfs_buf_lock(bp);
 688         XFS_BUF_SET_ADDR(bp, daddr);
 689         XFS_BUF_READ(bp);
 690         XFS_BUF_BUSY(bp);
 691
 692         xfsbdstrat(mp, bp);
 693         error = xfs_buf_iowait(bp);
 694         if (error || bp->b_error) {
 695                 xfs_buf_relse(bp);
 696                 return NULL;
 697         }
 698         return bp;
 699 }
 700
 701 xfs_buf_t *
 702 xfs_buf_get_empty(
 703         size_t                  len,
 704         xfs_buftarg_t           *target)
 705 {
 706         xfs_buf_t               *bp;
 707
 708         bp = xfs_buf_allocate(0);
 709         if (bp)
 710                 _xfs_buf_initialize(bp, target, 0, len, 0);
 711         return bp;
 712 }
 713
 714 static inline struct page *
 715 mem_to_page(
 716         void                    *addr)
 717 {
 718         if ((!is_vmalloc_addr(addr))) {
 719                 return virt_to_page(addr);
 720         } else {
 721                 return vmalloc_to_page(addr);
 722         }
 723 }
 724
 725 int
 726 xfs_buf_associate_memory(
 727         xfs_buf_t               *bp,
 728         void                    *mem,
 729         size_t                  len)
 730 {
 731         int                     rval;
 732         int                     i = 0;
 733         unsigned long           pageaddr;
 734         unsigned long           offset;
 735         size_t                  buflen;
 736         int                     page_count;
 737
 738         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
 739         offset = (unsigned long)mem - pageaddr;
 740         buflen = PAGE_CACHE_ALIGN(len + offset);
 741         page_count = buflen >> PAGE_CACHE_SHIFT;
 742
 743         /* Free any previous set of page pointers */
 744         if (bp->b_pages)
 745                 _xfs_buf_free_pages(bp);
 746
 747         bp->b_pages = NULL;
 748         bp->b_addr = mem;
 749
 750         rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
 751         if (rval)
 752                 return rval;
 753
 754         bp->b_offset = offset;
 755
 756         for (i = 0; i < bp->b_page_count; i++) {
 757                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 758                 pageaddr += PAGE_CACHE_SIZE;
 759         }
 760
 761         bp->b_count_desired = len;
 762         bp->b_buffer_length = buflen;
 763         bp->b_flags |= XBF_MAPPED;
 764         bp->b_flags &= ~_XBF_PAGE_LOCKED;
 765
 766         return 0;
 767 }
 768
 769 xfs_buf_t *
 770 xfs_buf_get_uncached(
 771         struct xfs_buftarg      *target,
 772         size_t                  len,
 773         int                     flags)
 774 {
 775         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 776         int                     error, i;
 777         xfs_buf_t               *bp;
 778
 779         bp = xfs_buf_allocate(0);
 780         if (unlikely(bp == NULL))
 781                 goto fail;
 782         _xfs_buf_initialize(bp, target, 0, len, 0);
 783
 784         error = _xfs_buf_get_pages(bp, page_count, 0);
 785         if (error)
 786                 goto fail_free_buf;
 787
 788         for (i = 0; i < page_count; i++) {
 789                 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
 790                 if (!bp->b_pages[i])
 791                         goto fail_free_mem;
 792         }
 793         bp->b_flags |= _XBF_PAGES;
 794
 795         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 796         if (unlikely(error)) {
 797                 printk(KERN_WARNING "%s: failed to map pages\n",
 798                                 __func__);
 799                 goto fail_free_mem;
 800         }
 801
 802         xfs_buf_unlock(bp);
 803
 804         trace_xfs_buf_get_uncached(bp, _RET_IP_);
 805         return bp;
 806
 807  fail_free_mem:
 808         while (--i >= 0)
 809                 __free_page(bp->b_pages[i]);
 810         _xfs_buf_free_pages(bp);
 811  fail_free_buf:
 812         xfs_buf_deallocate(bp);
 813  fail:
 814         return NULL;
 815 }
 816
 817 /*
 818  *      Increment reference count on buffer, to hold the buffer concurrently
 819  *      with another thread which may release (free) the buffer asynchronously.
 820  *      Must hold the buffer already to call this function.
 821  */
 822 void
 823 xfs_buf_hold(
 824         xfs_buf_t               *bp)
 825 {
 826         trace_xfs_buf_hold(bp, _RET_IP_);
 827         atomic_inc(&bp->b_hold);
 828 }
 829
 830 /*
 831  *      Releases a hold on the specified buffer.  If the
 832  *      the hold count is 1, calls xfs_buf_free.
 833  */
 834 void
 835 xfs_buf_rele(
 836         xfs_buf_t               *bp)
 837 {
 838         struct xfs_perag        *pag = bp->b_pag;
 839
 840         trace_xfs_buf_rele(bp, _RET_IP_);
 841
 842         if (!pag) {
 843                 ASSERT(!bp->b_relse);
 844                 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
 845                 if (atomic_dec_and_test(&bp->b_hold))
 846                         xfs_buf_free(bp);
 847                 return;
 848         }
 849
 850         ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
 851         ASSERT(atomic_read(&bp->b_hold) > 0);
 852         if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
 853                 if (bp->b_relse) {
 854                         atomic_inc(&bp->b_hold);
 855                         spin_unlock(&pag->pag_buf_lock);
 856                         bp->b_relse(bp);
 857                 } else {
 858                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 859                         rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
 860                         spin_unlock(&pag->pag_buf_lock);
 861                         xfs_perag_put(pag);
 862                         xfs_buf_free(bp);
 863                 }
 864         }
 865 }
 866
 867
 868 /*
 869  *      Mutual exclusion on buffers.  Locking model:
 870  *
 871  *      Buffers associated with inodes for which buffer locking
 872  *      is not enabled are not protected by semaphores, and are
 873  *      assumed to be exclusively owned by the caller.  There is a
 874  *      spinlock in the buffer, used by the caller when concurrent
 875  *      access is possible.
 876  */
 877
 878 /*
 879  *      Locks a buffer object, if it is not already locked.
 880  *      Note that this in no way locks the underlying pages, so it is only
 881  *      useful for synchronizing concurrent use of buffer objects, not for
 882  *      synchronizing independent access to the underlying pages.
 883  */
 884 int
 885 xfs_buf_cond_lock(
 886         xfs_buf_t               *bp)
 887 {
 888         int                     locked;
 889
 890         locked = down_trylock(&bp->b_sema) == 0;
 891         if (locked)
 892                 XB_SET_OWNER(bp);
 893
 894         trace_xfs_buf_cond_lock(bp, _RET_IP_);
 895         return locked ? 0 : -EBUSY;
 896 }
 897
 898 int
 899 xfs_buf_lock_value(
 900         xfs_buf_t               *bp)
 901 {
 902         return bp->b_sema.count;
 903 }
 904
 905 /*
 906  *      Locks a buffer object.
 907  *      Note that this in no way locks the underlying pages, so it is only
 908  *      useful for synchronizing concurrent use of buffer objects, not for
 909  *      synchronizing independent access to the underlying pages.
 910  *
 911  *      If we come across a stale, pinned, locked buffer, we know that we
 912  *      are being asked to lock a buffer that has been reallocated. Because
 913  *      it is pinned, we know that the log has not been pushed to disk and
 914  *      hence it will still be locked. Rather than sleeping until someone
 915  *      else pushes the log, push it ourselves before trying to get the lock.
 916  */
 917 void
 918 xfs_buf_lock(
 919         xfs_buf_t               *bp)
 920 {
 921         trace_xfs_buf_lock(bp, _RET_IP_);
 922
 923         if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 924                 xfs_log_force(bp->b_target->bt_mount, 0);
 925         if (atomic_read(&bp->b_io_remaining))
 926                 blk_run_address_space(bp->b_target->bt_mapping);
 927         down(&bp->b_sema);
 928         XB_SET_OWNER(bp);
 929
 930         trace_xfs_buf_lock_done(bp, _RET_IP_);
 931 }
 932
 933 /*
 934  *      Releases the lock on the buffer object.
 935  *      If the buffer is marked delwri but is not queued, do so before we
 936  *      unlock the buffer as we need to set flags correctly.  We also need to
 937  *      take a reference for the delwri queue because the unlocker is going to
 938  *      drop their's and they don't know we just queued it.
 939  */
 940 void
 941 xfs_buf_unlock(
 942         xfs_buf_t               *bp)
 943 {
 944         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 945                 atomic_inc(&bp->b_hold);
 946                 bp->b_flags |= XBF_ASYNC;
 947                 xfs_buf_delwri_queue(bp, 0);
 948         }
 949
 950         XB_CLEAR_OWNER(bp);
 951         up(&bp->b_sema);
 952
 953         trace_xfs_buf_unlock(bp, _RET_IP_);
 954 }
 955
 956 STATIC void
 957 xfs_buf_wait_unpin(
 958         xfs_buf_t               *bp)
 959 {
 960         DECLARE_WAITQUEUE       (wait, current);
 961
 962         if (atomic_read(&bp->b_pin_count) == 0)
 963                 return;
 964
 965         add_wait_queue(&bp->b_waiters, &wait);
 966         for (;;) {
 967                 set_current_state(TASK_UNINTERRUPTIBLE);
 968                 if (atomic_read(&bp->b_pin_count) == 0)
 969                         break;
 970                 if (atomic_read(&bp->b_io_remaining))
 971                         blk_run_address_space(bp->b_target->bt_mapping);
 972                 schedule();
 973         }
 974         remove_wait_queue(&bp->b_waiters, &wait);
 975         set_current_state(TASK_RUNNING);
 976 }
 977
 978 /*
 979  *      Buffer Utility Routines
 980  */
 981
 982 STATIC void
 983 xfs_buf_iodone_work(
 984         struct work_struct      *work)
 985 {
 986         xfs_buf_t               *bp =
 987                 container_of(work, xfs_buf_t, b_iodone_work);
 988
 989         if (bp->b_iodone)
 990                 (*(bp->b_iodone))(bp);
 991         else if (bp->b_flags & XBF_ASYNC)
 992                 xfs_buf_relse(bp);
 993 }
 994
 995 void
 996 xfs_buf_ioend(
 997         xfs_buf_t               *bp,
 998         int                     schedule)
 999 {
1000         trace_xfs_buf_iodone(bp, _RET_IP_);
1001
1002         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1003         if (bp->b_error == 0)
1004                 bp->b_flags |= XBF_DONE;
1005
1006         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1007                 if (schedule) {
1008                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1009                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1010                 } else {
1011                         xfs_buf_iodone_work(&bp->b_iodone_work);
1012                 }
1013         } else {
1014                 complete(&bp->b_iowait);
1015         }
1016 }
1017
1018 void
1019 xfs_buf_ioerror(
1020         xfs_buf_t               *bp,
1021         int                     error)
1022 {
1023         ASSERT(error >= 0 && error <= 0xffff);
1024         bp->b_error = (unsigned short)error;
1025         trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1026 }
1027
1028 int
1029 xfs_bwrite(
1030         struct xfs_mount        *mp,
1031         struct xfs_buf          *bp)
1032 {
1033         int                     error;
1034
1035         bp->b_flags |= XBF_WRITE;
1036         bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1037
1038         xfs_buf_delwri_dequeue(bp);
1039         xfs_bdstrat_cb(bp);
1040
1041         error = xfs_buf_iowait(bp);
1042         if (error)
1043                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1044         xfs_buf_relse(bp);
1045         return error;
1046 }
1047
1048 void
1049 xfs_bdwrite(
1050         void                    *mp,
1051         struct xfs_buf          *bp)
1052 {
1053         trace_xfs_buf_bdwrite(bp, _RET_IP_);
1054
1055         bp->b_flags &= ~XBF_READ;
1056         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1057
1058         xfs_buf_delwri_queue(bp, 1);
1059 }
1060
1061 /*
1062  * Called when we want to stop a buffer from getting written or read.
1063  * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1064  * so that the proper iodone callbacks get called.
1065  */
1066 STATIC int
1067 xfs_bioerror(
1068         xfs_buf_t *bp)
1069 {
1070 #ifdef XFSERRORDEBUG
1071         ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1072 #endif
1073
1074         /*
1075          * No need to wait until the buffer is unpinned, we aren't flushing it.
1076          */
1077         XFS_BUF_ERROR(bp, EIO);
1078
1079         /*
1080          * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1081          */
1082         XFS_BUF_UNREAD(bp);
1083         XFS_BUF_UNDELAYWRITE(bp);
1084         XFS_BUF_UNDONE(bp);
1085         XFS_BUF_STALE(bp);
1086
1087         xfs_buf_ioend(bp, 0);
1088
1089         return EIO;
1090 }
1091
1092 /*
1093  * Same as xfs_bioerror, except that we are releasing the buffer
1094  * here ourselves, and avoiding the xfs_buf_ioend call.
1095  * This is meant for userdata errors; metadata bufs come with
1096  * iodone functions attached, so that we can track down errors.
1097  */
1098 STATIC int
1099 xfs_bioerror_relse(
1100         struct xfs_buf  *bp)
1101 {
1102         int64_t         fl = XFS_BUF_BFLAGS(bp);
1103         /*
1104          * No need to wait until the buffer is unpinned.
1105          * We aren't flushing it.
1106          *
1107          * chunkhold expects B_DONE to be set, whether
1108          * we actually finish the I/O or not. We don't want to
1109          * change that interface.
1110          */
1111         XFS_BUF_UNREAD(bp);
1112         XFS_BUF_UNDELAYWRITE(bp);
1113         XFS_BUF_DONE(bp);
1114         XFS_BUF_STALE(bp);
1115         XFS_BUF_CLR_IODONE_FUNC(bp);
1116         if (!(fl & XBF_ASYNC)) {
1117                 /*
1118                  * Mark b_error and B_ERROR _both_.
1119                  * Lot's of chunkcache code assumes that.
1120                  * There's no reason to mark error for
1121                  * ASYNC buffers.
1122                  */
1123                 XFS_BUF_ERROR(bp, EIO);
1124                 XFS_BUF_FINISH_IOWAIT(bp);
1125         } else {
1126                 xfs_buf_relse(bp);
1127         }
1128
1129         return EIO;
1130 }
1131
1132
1133 /*
1134  * All xfs metadata buffers except log state machine buffers
1135  * get this attached as their b_bdstrat callback function.
1136  * This is so that we can catch a buffer
1137  * after prematurely unpinning it to forcibly shutdown the filesystem.
1138  */
1139 int
1140 xfs_bdstrat_cb(
1141         struct xfs_buf  *bp)
1142 {
1143         if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1144                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1145                 /*
1146                  * Metadata write that didn't get logged but
1147                  * written delayed anyway. These aren't associated
1148                  * with a transaction, and can be ignored.
1149                  */
1150                 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1151                         return xfs_bioerror_relse(bp);
1152                 else
1153                         return xfs_bioerror(bp);
1154         }
1155
1156         xfs_buf_iorequest(bp);
1157         return 0;
1158 }
1159
1160 /*
1161  * Wrapper around bdstrat so that we can stop data from going to disk in case
1162  * we are shutting down the filesystem.  Typically user data goes thru this
1163  * path; one of the exceptions is the superblock.
1164  */
1165 void
1166 xfsbdstrat(
1167         struct xfs_mount        *mp,
1168         struct xfs_buf          *bp)
1169 {
1170         if (XFS_FORCED_SHUTDOWN(mp)) {
1171                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1172                 xfs_bioerror_relse(bp);
1173                 return;
1174         }
1175
1176         xfs_buf_iorequest(bp);
1177 }
1178
1179 STATIC void
1180 _xfs_buf_ioend(
1181         xfs_buf_t               *bp,
1182         int                     schedule)
1183 {
1184         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1185                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1186                 xfs_buf_ioend(bp, schedule);
1187         }
1188 }
1189
1190 STATIC void
1191 xfs_buf_bio_end_io(
1192         struct bio              *bio,
1193         int                     error)
1194 {
1195         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1196         unsigned int            blocksize = bp->b_target->bt_bsize;
1197         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1198
1199         xfs_buf_ioerror(bp, -error);
1200
1201         if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1202                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1203
1204         do {
1205                 struct page     *page = bvec->bv_page;
1206
1207                 ASSERT(!PagePrivate(page));
1208                 if (unlikely(bp->b_error)) {
1209                         if (bp->b_flags & XBF_READ)
1210                                 ClearPageUptodate(page);
1211                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1212                         SetPageUptodate(page);
1213                 } else if (!PagePrivate(page) &&
1214                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1215                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1216                 }
1217
1218                 if (--bvec >= bio->bi_io_vec)
1219                         prefetchw(&bvec->bv_page->flags);
1220
1221                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1222                         unlock_page(page);
1223         } while (bvec >= bio->bi_io_vec);
1224
1225         _xfs_buf_ioend(bp, 1);
1226         bio_put(bio);
1227 }
1228
1229 STATIC void
1230 _xfs_buf_ioapply(
1231         xfs_buf_t               *bp)
1232 {
1233         int                     rw, map_i, total_nr_pages, nr_pages;
1234         struct bio              *bio;
1235         int                     offset = bp->b_offset;
1236         int                     size = bp->b_count_desired;
1237         sector_t                sector = bp->b_bn;
1238         unsigned int            blocksize = bp->b_target->bt_bsize;
1239
1240         total_nr_pages = bp->b_page_count;
1241         map_i = 0;
1242
1243         if (bp->b_flags & XBF_ORDERED) {
1244                 ASSERT(!(bp->b_flags & XBF_READ));
1245                 rw = WRITE_FLUSH_FUA;
1246         } else if (bp->b_flags & XBF_LOG_BUFFER) {
1247                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1248                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1249                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1250         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1251                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1252                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1253                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1254         } else {
1255                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1256                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1257         }
1258
1259         /* Special code path for reading a sub page size buffer in --
1260          * we populate up the whole page, and hence the other metadata
1261          * in the same page.  This optimization is only valid when the
1262          * filesystem block size is not smaller than the page size.
1263          */
1264         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1265             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1266               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1267             (blocksize >= PAGE_CACHE_SIZE)) {
1268                 bio = bio_alloc(GFP_NOIO, 1);
1269
1270                 bio->bi_bdev = bp->b_target->bt_bdev;
1271                 bio->bi_sector = sector - (offset >> BBSHIFT);
1272                 bio->bi_end_io = xfs_buf_bio_end_io;
1273                 bio->bi_private = bp;
1274
1275                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1276                 size = 0;
1277
1278                 atomic_inc(&bp->b_io_remaining);
1279
1280                 goto submit_io;
1281         }
1282
1283 next_chunk:
1284         atomic_inc(&bp->b_io_remaining);
1285         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1286         if (nr_pages > total_nr_pages)
1287                 nr_pages = total_nr_pages;
1288
1289         bio = bio_alloc(GFP_NOIO, nr_pages);
1290         bio->bi_bdev = bp->b_target->bt_bdev;
1291         bio->bi_sector = sector;
1292         bio->bi_end_io = xfs_buf_bio_end_io;
1293         bio->bi_private = bp;
1294
1295         for (; size && nr_pages; nr_pages--, map_i++) {
1296                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1297
1298                 if (nbytes > size)
1299                         nbytes = size;
1300
1301                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1302                 if (rbytes < nbytes)
1303                         break;
1304
1305                 offset = 0;
1306                 sector += nbytes >> BBSHIFT;
1307                 size -= nbytes;
1308                 total_nr_pages--;
1309         }
1310
1311 submit_io:
1312         if (likely(bio->bi_size)) {
1313                 if (xfs_buf_is_vmapped(bp)) {
1314                         flush_kernel_vmap_range(bp->b_addr,
1315                                                 xfs_buf_vmap_len(bp));
1316                 }
1317                 submit_bio(rw, bio);
1318                 if (size)
1319                         goto next_chunk;
1320         } else {
1321                 /*
1322                  * if we get here, no pages were added to the bio. However,
1323                  * we can't just error out here - if the pages are locked then
1324                  * we have to unlock them otherwise we can hang on a later
1325                  * access to the page.
1326                  */
1327                 xfs_buf_ioerror(bp, EIO);
1328                 if (bp->b_flags & _XBF_PAGE_LOCKED) {
1329                         int i;
1330                         for (i = 0; i < bp->b_page_count; i++)
1331                                 unlock_page(bp->b_pages[i]);
1332                 }
1333                 bio_put(bio);
1334         }
1335 }
1336
1337 int
1338 xfs_buf_iorequest(
1339         xfs_buf_t               *bp)
1340 {
1341         trace_xfs_buf_iorequest(bp, _RET_IP_);
1342
1343         if (bp->b_flags & XBF_DELWRI) {
1344                 xfs_buf_delwri_queue(bp, 1);
1345                 return 0;
1346         }
1347
1348         if (bp->b_flags & XBF_WRITE) {
1349                 xfs_buf_wait_unpin(bp);
1350         }
1351
1352         xfs_buf_hold(bp);
1353
1354         /* Set the count to 1 initially, this will stop an I/O
1355          * completion callout which happens before we have started
1356          * all the I/O from calling xfs_buf_ioend too early.
1357          */
1358         atomic_set(&bp->b_io_remaining, 1);
1359         _xfs_buf_ioapply(bp);
1360         _xfs_buf_ioend(bp, 0);
1361
1362         xfs_buf_rele(bp);
1363         return 0;
1364 }
1365
1366 /*
1367  *      Waits for I/O to complete on the buffer supplied.
1368  *      It returns immediately if no I/O is pending.
1369  *      It returns the I/O error code, if any, or 0 if there was no error.
1370  */
1371 int
1372 xfs_buf_iowait(
1373         xfs_buf_t               *bp)
1374 {
1375         trace_xfs_buf_iowait(bp, _RET_IP_);
1376
1377         if (atomic_read(&bp->b_io_remaining))
1378                 blk_run_address_space(bp->b_target->bt_mapping);
1379         wait_for_completion(&bp->b_iowait);
1380
1381         trace_xfs_buf_iowait_done(bp, _RET_IP_);
1382         return bp->b_error;
1383 }
1384
1385 xfs_caddr_t
1386 xfs_buf_offset(
1387         xfs_buf_t               *bp,
1388         size_t                  offset)
1389 {
1390         struct page             *page;
1391
1392         if (bp->b_flags & XBF_MAPPED)
1393                 return XFS_BUF_PTR(bp) + offset;
1394
1395         offset += bp->b_offset;
1396         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1397         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1398 }
1399
1400 /*
1401  *      Move data into or out of a buffer.
1402  */
1403 void
1404 xfs_buf_iomove(
1405         xfs_buf_t               *bp,    /* buffer to process            */
1406         size_t                  boff,   /* starting buffer offset       */
1407         size_t                  bsize,  /* length to copy               */
1408         void                    *data,  /* data address                 */
1409         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1410 {
1411         size_t                  bend, cpoff, csize;
1412         struct page             *page;
1413
1414         bend = boff + bsize;
1415         while (boff < bend) {
1416                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1417                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1418                 csize = min_t(size_t,
1419                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1420
1421                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1422
1423                 switch (mode) {
1424                 case XBRW_ZERO:
1425                         memset(page_address(page) + cpoff, 0, csize);
1426                         break;
1427                 case XBRW_READ:
1428                         memcpy(data, page_address(page) + cpoff, csize);
1429                         break;
1430                 case XBRW_WRITE:
1431                         memcpy(page_address(page) + cpoff, data, csize);
1432                 }
1433
1434                 boff += csize;
1435                 data += csize;
1436         }
1437 }
1438
1439 /*
1440  *      Handling of buffer targets (buftargs).
1441  */
1442
1443 /*
1444  *      Wait for any bufs with callbacks that have been submitted but
1445  *      have not yet returned... walk the hash list for the target.
1446  */
1447 void
1448 xfs_wait_buftarg(
1449         struct xfs_buftarg      *btp)
1450 {
1451         struct xfs_perag        *pag;
1452         uint                    i;
1453
1454         for (i = 0; i < btp->bt_mount->m_sb.sb_agcount; i++) {
1455                 pag = xfs_perag_get(btp->bt_mount, i);
1456                 spin_lock(&pag->pag_buf_lock);
1457                 while (rb_first(&pag->pag_buf_tree)) {
1458                         spin_unlock(&pag->pag_buf_lock);
1459                         delay(100);
1460                         spin_lock(&pag->pag_buf_lock);
1461                 }
1462                 spin_unlock(&pag->pag_buf_lock);
1463                 xfs_perag_put(pag);
1464         }
1465 }
1466
1467 /*
1468  *      buftarg list for delwrite queue processing
1469  */
1470 static LIST_HEAD(xfs_buftarg_list);
1471 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1472
1473 STATIC void
1474 xfs_register_buftarg(
1475         xfs_buftarg_t           *btp)
1476 {
1477         spin_lock(&xfs_buftarg_lock);
1478         list_add(&btp->bt_list, &xfs_buftarg_list);
1479         spin_unlock(&xfs_buftarg_lock);
1480 }
1481
1482 STATIC void
1483 xfs_unregister_buftarg(
1484         xfs_buftarg_t           *btp)
1485 {
1486         spin_lock(&xfs_buftarg_lock);
1487         list_del(&btp->bt_list);
1488         spin_unlock(&xfs_buftarg_lock);
1489 }
1490
1491 void
1492 xfs_free_buftarg(
1493         struct xfs_mount        *mp,
1494         struct xfs_buftarg      *btp)
1495 {
1496         xfs_flush_buftarg(btp, 1);
1497         if (mp->m_flags & XFS_MOUNT_BARRIER)
1498                 xfs_blkdev_issue_flush(btp);
1499         iput(btp->bt_mapping->host);
1500
1501         /* Unregister the buftarg first so that we don't get a
1502          * wakeup finding a non-existent task
1503          */
1504         xfs_unregister_buftarg(btp);
1505         kthread_stop(btp->bt_task);
1506
1507         kmem_free(btp);
1508 }
1509
1510 STATIC int
1511 xfs_setsize_buftarg_flags(
1512         xfs_buftarg_t           *btp,
1513         unsigned int            blocksize,
1514         unsigned int            sectorsize,
1515         int                     verbose)
1516 {
1517         btp->bt_bsize = blocksize;
1518         btp->bt_sshift = ffs(sectorsize) - 1;
1519         btp->bt_smask = sectorsize - 1;
1520
1521         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1522                 printk(KERN_WARNING
1523                         "XFS: Cannot set_blocksize to %u on device %s\n",
1524                         sectorsize, XFS_BUFTARG_NAME(btp));
1525                 return EINVAL;
1526         }
1527
1528         if (verbose &&
1529             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1530                 printk(KERN_WARNING
1531                         "XFS: %u byte sectors in use on device %s.  "
1532                         "This is suboptimal; %u or greater is ideal.\n",
1533                         sectorsize, XFS_BUFTARG_NAME(btp),
1534                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1535         }
1536
1537         return 0;
1538 }
1539
1540 /*
1541  *      When allocating the initial buffer target we have not yet
1542  *      read in the superblock, so don't know what sized sectors
1543  *      are being used is at this early stage.  Play safe.
1544  */
1545 STATIC int
1546 xfs_setsize_buftarg_early(
1547         xfs_buftarg_t           *btp,
1548         struct block_device     *bdev)
1549 {
1550         return xfs_setsize_buftarg_flags(btp,
1551                         PAGE_CACHE_SIZE, bdev_logical_block_size(bdev), 0);
1552 }
1553
1554 int
1555 xfs_setsize_buftarg(
1556         xfs_buftarg_t           *btp,
1557         unsigned int            blocksize,
1558         unsigned int            sectorsize)
1559 {
1560         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1561 }
1562
1563 STATIC int
1564 xfs_mapping_buftarg(
1565         xfs_buftarg_t           *btp,
1566         struct block_device     *bdev)
1567 {
1568         struct backing_dev_info *bdi;
1569         struct inode            *inode;
1570         struct address_space    *mapping;
1571         static const struct address_space_operations mapping_aops = {
1572                 .sync_page = block_sync_page,
1573                 .migratepage = fail_migrate_page,
1574         };
1575
1576         inode = new_inode(bdev->bd_inode->i_sb);
1577         if (!inode) {
1578                 printk(KERN_WARNING
1579                         "XFS: Cannot allocate mapping inode for device %s\n",
1580                         XFS_BUFTARG_NAME(btp));
1581                 return ENOMEM;
1582         }
1583         inode->i_ino = get_next_ino();
1584         inode->i_mode = S_IFBLK;
1585         inode->i_bdev = bdev;
1586         inode->i_rdev = bdev->bd_dev;
1587         bdi = blk_get_backing_dev_info(bdev);
1588         if (!bdi)
1589                 bdi = &default_backing_dev_info;
1590         mapping = &inode->i_data;
1591         mapping->a_ops = &mapping_aops;
1592         mapping->backing_dev_info = bdi;
1593         mapping_set_gfp_mask(mapping, GFP_NOFS);
1594         btp->bt_mapping = mapping;
1595         return 0;
1596 }
1597
1598 STATIC int
1599 xfs_alloc_delwrite_queue(
1600         xfs_buftarg_t           *btp,
1601         const char              *fsname)
1602 {
1603         int     error = 0;
1604
1605         INIT_LIST_HEAD(&btp->bt_list);
1606         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1607         spin_lock_init(&btp->bt_delwrite_lock);
1608         btp->bt_flags = 0;
1609         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1610         if (IS_ERR(btp->bt_task)) {
1611                 error = PTR_ERR(btp->bt_task);
1612                 goto out_error;
1613         }
1614         xfs_register_buftarg(btp);
1615 out_error:
1616         return error;
1617 }
1618
1619 xfs_buftarg_t *
1620 xfs_alloc_buftarg(
1621         struct xfs_mount        *mp,
1622         struct block_device     *bdev,
1623         int                     external,
1624         const char              *fsname)
1625 {
1626         xfs_buftarg_t           *btp;
1627
1628         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1629
1630         btp->bt_mount = mp;
1631         btp->bt_dev =  bdev->bd_dev;
1632         btp->bt_bdev = bdev;
1633         if (xfs_setsize_buftarg_early(btp, bdev))
1634                 goto error;
1635         if (xfs_mapping_buftarg(btp, bdev))
1636                 goto error;
1637         if (xfs_alloc_delwrite_queue(btp, fsname))
1638                 goto error;
1639         return btp;
1640
1641 error:
1642         kmem_free(btp);
1643         return NULL;
1644 }
1645
1646
1647 /*
1648  *      Delayed write buffer handling
1649  */
1650 STATIC void
1651 xfs_buf_delwri_queue(
1652         xfs_buf_t               *bp,
1653         int                     unlock)
1654 {
1655         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1656         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1657
1658         trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1659
1660         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1661
1662         spin_lock(dwlk);
1663         /* If already in the queue, dequeue and place at tail */
1664         if (!list_empty(&bp->b_list)) {
1665                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1666                 if (unlock)
1667                         atomic_dec(&bp->b_hold);
1668                 list_del(&bp->b_list);
1669         }
1670
1671         if (list_empty(dwq)) {
1672                 /* start xfsbufd as it is about to have something to do */
1673                 wake_up_process(bp->b_target->bt_task);
1674         }
1675
1676         bp->b_flags |= _XBF_DELWRI_Q;
1677         list_add_tail(&bp->b_list, dwq);
1678         bp->b_queuetime = jiffies;
1679         spin_unlock(dwlk);
1680
1681         if (unlock)
1682                 xfs_buf_unlock(bp);
1683 }
1684
1685 void
1686 xfs_buf_delwri_dequeue(
1687         xfs_buf_t               *bp)
1688 {
1689         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1690         int                     dequeued = 0;
1691
1692         spin_lock(dwlk);
1693         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1694                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1695                 list_del_init(&bp->b_list);
1696                 dequeued = 1;
1697         }
1698         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1699         spin_unlock(dwlk);
1700
1701         if (dequeued)
1702                 xfs_buf_rele(bp);
1703
1704         trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1705 }
1706
1707 /*
1708  * If a delwri buffer needs to be pushed before it has aged out, then promote
1709  * it to the head of the delwri queue so that it will be flushed on the next
1710  * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1711  * than the age currently needed to flush the buffer. Hence the next time the
1712  * xfsbufd sees it is guaranteed to be considered old enough to flush.
1713  */
1714 void
1715 xfs_buf_delwri_promote(
1716         struct xfs_buf  *bp)
1717 {
1718         struct xfs_buftarg *btp = bp->b_target;
1719         long            age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1720
1721         ASSERT(bp->b_flags & XBF_DELWRI);
1722         ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1723
1724         /*
1725          * Check the buffer age before locking the delayed write queue as we
1726          * don't need to promote buffers that are already past the flush age.
1727          */
1728         if (bp->b_queuetime < jiffies - age)
1729                 return;
1730         bp->b_queuetime = jiffies - age;
1731         spin_lock(&btp->bt_delwrite_lock);
1732         list_move(&bp->b_list, &btp->bt_delwrite_queue);
1733         spin_unlock(&btp->bt_delwrite_lock);
1734 }
1735
1736 STATIC void
1737 xfs_buf_runall_queues(
1738         struct workqueue_struct *queue)
1739 {
1740         flush_workqueue(queue);
1741 }
1742
1743 STATIC int
1744 xfsbufd_wakeup(
1745         struct shrinker         *shrink,
1746         int                     priority,
1747         gfp_t                   mask)
1748 {
1749         xfs_buftarg_t           *btp;
1750
1751         spin_lock(&xfs_buftarg_lock);
1752         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1753                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1754                         continue;
1755                 if (list_empty(&btp->bt_delwrite_queue))
1756                         continue;
1757                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1758                 wake_up_process(btp->bt_task);
1759         }
1760         spin_unlock(&xfs_buftarg_lock);
1761         return 0;
1762 }
1763
1764 /*
1765  * Move as many buffers as specified to the supplied list
1766  * idicating if we skipped any buffers to prevent deadlocks.
1767  */
1768 STATIC int
1769 xfs_buf_delwri_split(
1770         xfs_buftarg_t   *target,
1771         struct list_head *list,
1772         unsigned long   age)
1773 {
1774         xfs_buf_t       *bp, *n;
1775         struct list_head *dwq = &target->bt_delwrite_queue;
1776         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1777         int             skipped = 0;
1778         int             force;
1779
1780         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1781         INIT_LIST_HEAD(list);
1782         spin_lock(dwlk);
1783         list_for_each_entry_safe(bp, n, dwq, b_list) {
1784                 ASSERT(bp->b_flags & XBF_DELWRI);
1785
1786                 if (!XFS_BUF_ISPINNED(bp) && !xfs_buf_cond_lock(bp)) {
1787                         if (!force &&
1788                             time_before(jiffies, bp->b_queuetime + age)) {
1789                                 xfs_buf_unlock(bp);
1790                                 break;
1791                         }
1792
1793                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1794                                          _XBF_RUN_QUEUES);
1795                         bp->b_flags |= XBF_WRITE;
1796                         list_move_tail(&bp->b_list, list);
1797                         trace_xfs_buf_delwri_split(bp, _RET_IP_);
1798                 } else
1799                         skipped++;
1800         }
1801         spin_unlock(dwlk);
1802
1803         return skipped;
1804
1805 }
1806
1807 /*
1808  * Compare function is more complex than it needs to be because
1809  * the return value is only 32 bits and we are doing comparisons
1810  * on 64 bit values
1811  */
1812 static int
1813 xfs_buf_cmp(
1814         void            *priv,
1815         struct list_head *a,
1816         struct list_head *b)
1817 {
1818         struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1819         struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1820         xfs_daddr_t             diff;
1821
1822         diff = ap->b_bn - bp->b_bn;
1823         if (diff < 0)
1824                 return -1;
1825         if (diff > 0)
1826                 return 1;
1827         return 0;
1828 }
1829
1830 void
1831 xfs_buf_delwri_sort(
1832         xfs_buftarg_t   *target,
1833         struct list_head *list)
1834 {
1835         list_sort(NULL, list, xfs_buf_cmp);
1836 }
1837
1838 STATIC int
1839 xfsbufd(
1840         void            *data)
1841 {
1842         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1843
1844         current->flags |= PF_MEMALLOC;
1845
1846         set_freezable();
1847
1848         do {
1849                 long    age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1850                 long    tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1851                 int     count = 0;
1852                 struct list_head tmp;
1853
1854                 if (unlikely(freezing(current))) {
1855                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1856                         refrigerator();
1857                 } else {
1858                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1859                 }
1860
1861                 /* sleep for a long time if there is nothing to do. */
1862                 if (list_empty(&target->bt_delwrite_queue))
1863                         tout = MAX_SCHEDULE_TIMEOUT;
1864                 schedule_timeout_interruptible(tout);
1865
1866                 xfs_buf_delwri_split(target, &tmp, age);
1867                 list_sort(NULL, &tmp, xfs_buf_cmp);
1868                 while (!list_empty(&tmp)) {
1869                         struct xfs_buf *bp;
1870                         bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1871                         list_del_init(&bp->b_list);
1872                         xfs_bdstrat_cb(bp);
1873                         count++;
1874                 }
1875                 if (count)
1876                         blk_run_address_space(target->bt_mapping);
1877
1878         } while (!kthread_should_stop());
1879
1880         return 0;
1881 }
1882
1883 /*
1884  *      Go through all incore buffers, and release buffers if they belong to
1885  *      the given device. This is used in filesystem error handling to
1886  *      preserve the consistency of its metadata.
1887  */
1888 int
1889 xfs_flush_buftarg(
1890         xfs_buftarg_t   *target,
1891         int             wait)
1892 {
1893         xfs_buf_t       *bp;
1894         int             pincount = 0;
1895         LIST_HEAD(tmp_list);
1896         LIST_HEAD(wait_list);
1897
1898         xfs_buf_runall_queues(xfsconvertd_workqueue);
1899         xfs_buf_runall_queues(xfsdatad_workqueue);
1900         xfs_buf_runall_queues(xfslogd_workqueue);
1901
1902         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1903         pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1904
1905         /*
1906          * Dropped the delayed write list lock, now walk the temporary list.
1907          * All I/O is issued async and then if we need to wait for completion
1908          * we do that after issuing all the IO.
1909          */
1910         list_sort(NULL, &tmp_list, xfs_buf_cmp);
1911         while (!list_empty(&tmp_list)) {
1912                 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1913                 ASSERT(target == bp->b_target);
1914                 list_del_init(&bp->b_list);
1915                 if (wait) {
1916                         bp->b_flags &= ~XBF_ASYNC;
1917                         list_add(&bp->b_list, &wait_list);
1918                 }
1919                 xfs_bdstrat_cb(bp);
1920         }
1921
1922         if (wait) {
1923                 /* Expedite and wait for IO to complete. */
1924                 blk_run_address_space(target->bt_mapping);
1925                 while (!list_empty(&wait_list)) {
1926                         bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1927
1928                         list_del_init(&bp->b_list);
1929                         xfs_buf_iowait(bp);
1930                         xfs_buf_relse(bp);
1931                 }
1932         }
1933
1934         return pincount;
1935 }
1936
1937 int __init
1938 xfs_buf_init(void)
1939 {
1940         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1941                                                 KM_ZONE_HWALIGN, NULL);
1942         if (!xfs_buf_zone)
1943                 goto out;
1944
1945         xfslogd_workqueue = alloc_workqueue("xfslogd",
1946                                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1947         if (!xfslogd_workqueue)
1948                 goto out_free_buf_zone;
1949
1950         xfsdatad_workqueue = create_workqueue("xfsdatad");
1951         if (!xfsdatad_workqueue)
1952                 goto out_destroy_xfslogd_workqueue;
1953
1954         xfsconvertd_workqueue = create_workqueue("xfsconvertd");
1955         if (!xfsconvertd_workqueue)
1956                 goto out_destroy_xfsdatad_workqueue;
1957
1958         register_shrinker(&xfs_buf_shake);
1959         return 0;
1960
1961  out_destroy_xfsdatad_workqueue:
1962         destroy_workqueue(xfsdatad_workqueue);
1963  out_destroy_xfslogd_workqueue:
1964         destroy_workqueue(xfslogd_workqueue);
1965  out_free_buf_zone:
1966         kmem_zone_destroy(xfs_buf_zone);
1967  out:
1968         return -ENOMEM;
1969 }
1970
1971 void
1972 xfs_buf_terminate(void)
1973 {
1974         unregister_shrinker(&xfs_buf_shake);
1975         destroy_workqueue(xfsconvertd_workqueue);
1976         destroy_workqueue(xfsdatad_workqueue);
1977         destroy_workqueue(xfslogd_workqueue);
1978         kmem_zone_destroy(xfs_buf_zone);
1979 }
1980
1981 #ifdef CONFIG_KDB_MODULES
1982 struct list_head *
1983 xfs_get_buftarg_list(void)
1984 {
1985         return &xfs_buftarg_list;
1986 }
1987 #endif