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                         /*
 444                          * If we look at something, bring it to the
 445                          * front of the list for next time.
 446                          */
 447                         atomic_inc(&bp->b_hold);
 448                         list_move(&bp->b_hash_list, &hash->bh_list);
 449                         goto found;
 450                 }
 451         }
 452
 453         /* No match found */
 454         if (new_bp) {
 455                 _xfs_buf_initialize(new_bp, btp, range_base,
 456                                 range_length, flags);
 457                 new_bp->b_hash = hash;
 458                 list_add(&new_bp->b_hash_list, &hash->bh_list);
 459         } else {
 460                 XFS_STATS_INC(xb_miss_locked);
 461         }
 462
 463         spin_unlock(&hash->bh_lock);
 464         return new_bp;
 465
 466 found:
 467         spin_unlock(&hash->bh_lock);
 468
 469         /* Attempt to get the semaphore without sleeping,
 470          * if this does not work then we need to drop the
 471          * spinlock and do a hard attempt on the semaphore.
 472          */
 473         if (down_trylock(&bp->b_sema)) {
 474                 if (!(flags & XBF_TRYLOCK)) {
 475                         /* wait for buffer ownership */
 476                         xfs_buf_lock(bp);
 477                         XFS_STATS_INC(xb_get_locked_waited);
 478                 } else {
 479                         /* We asked for a trylock and failed, no need
 480                          * to look at file offset and length here, we
 481                          * know that this buffer at least overlaps our
 482                          * buffer and is locked, therefore our buffer
 483                          * either does not exist, or is this buffer.
 484                          */
 485                         xfs_buf_rele(bp);
 486                         XFS_STATS_INC(xb_busy_locked);
 487                         return NULL;
 488                 }
 489         } else {
 490                 /* trylock worked */
 491                 XB_SET_OWNER(bp);
 492         }
 493
 494         if (bp->b_flags & XBF_STALE) {
 495                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 496                 bp->b_flags &= XBF_MAPPED;
 497         }
 498
 499         trace_xfs_buf_find(bp, flags, _RET_IP_);
 500         XFS_STATS_INC(xb_get_locked);
 501         return bp;
 502 }
 503
 504 /*
 505  *      Assembles a buffer covering the specified range.
 506  *      Storage in memory for all portions of the buffer will be allocated,
 507  *      although backing storage may not be.
 508  */
 509 xfs_buf_t *
 510 xfs_buf_get(
 511         xfs_buftarg_t           *target,/* target for buffer            */
 512         xfs_off_t               ioff,   /* starting offset of range     */
 513         size_t                  isize,  /* length of range              */
 514         xfs_buf_flags_t         flags)
 515 {
 516         xfs_buf_t               *bp, *new_bp;
 517         int                     error = 0, i;
 518
 519         new_bp = xfs_buf_allocate(flags);
 520         if (unlikely(!new_bp))
 521                 return NULL;
 522
 523         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 524         if (bp == new_bp) {
 525                 error = _xfs_buf_lookup_pages(bp, flags);
 526                 if (error)
 527                         goto no_buffer;
 528         } else {
 529                 xfs_buf_deallocate(new_bp);
 530                 if (unlikely(bp == NULL))
 531                         return NULL;
 532         }
 533
 534         for (i = 0; i < bp->b_page_count; i++)
 535                 mark_page_accessed(bp->b_pages[i]);
 536
 537         if (!(bp->b_flags & XBF_MAPPED)) {
 538                 error = _xfs_buf_map_pages(bp, flags);
 539                 if (unlikely(error)) {
 540                         printk(KERN_WARNING "%s: failed to map pages\n",
 541                                         __func__);
 542                         goto no_buffer;
 543                 }
 544         }
 545
 546         XFS_STATS_INC(xb_get);
 547
 548         /*
 549          * Always fill in the block number now, the mapped cases can do
 550          * their own overlay of this later.
 551          */
 552         bp->b_bn = ioff;
 553         bp->b_count_desired = bp->b_buffer_length;
 554
 555         trace_xfs_buf_get(bp, flags, _RET_IP_);
 556         return bp;
 557
 558  no_buffer:
 559         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 560                 xfs_buf_unlock(bp);
 561         xfs_buf_rele(bp);
 562         return NULL;
 563 }
 564
 565 STATIC int
 566 _xfs_buf_read(
 567         xfs_buf_t               *bp,
 568         xfs_buf_flags_t         flags)
 569 {
 570         int                     status;
 571
 572         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
 573         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
 574
 575         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
 576                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 577         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
 578                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 579
 580         status = xfs_buf_iorequest(bp);
 581         if (status || XFS_BUF_ISERROR(bp) || (flags & XBF_ASYNC))
 582                 return status;
 583         return xfs_buf_iowait(bp);
 584 }
 585
 586 xfs_buf_t *
 587 xfs_buf_read(
 588         xfs_buftarg_t           *target,
 589         xfs_off_t               ioff,
 590         size_t                  isize,
 591         xfs_buf_flags_t         flags)
 592 {
 593         xfs_buf_t               *bp;
 594
 595         flags |= XBF_READ;
 596
 597         bp = xfs_buf_get(target, ioff, isize, flags);
 598         if (bp) {
 599                 trace_xfs_buf_read(bp, flags, _RET_IP_);
 600
 601                 if (!XFS_BUF_ISDONE(bp)) {
 602                         XFS_STATS_INC(xb_get_read);
 603                         _xfs_buf_read(bp, flags);
 604                 } else if (flags & XBF_ASYNC) {
 605                         /*
 606                          * Read ahead call which is already satisfied,
 607                          * drop the buffer
 608                          */
 609                         goto no_buffer;
 610                 } else {
 611                         /* We do not want read in the flags */
 612                         bp->b_flags &= ~XBF_READ;
 613                 }
 614         }
 615
 616         return bp;
 617
 618  no_buffer:
 619         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 620                 xfs_buf_unlock(bp);
 621         xfs_buf_rele(bp);
 622         return NULL;
 623 }
 624
 625 /*
 626  *      If we are not low on memory then do the readahead in a deadlock
 627  *      safe manner.
 628  */
 629 void
 630 xfs_buf_readahead(
 631         xfs_buftarg_t           *target,
 632         xfs_off_t               ioff,
 633         size_t                  isize,
 634         xfs_buf_flags_t         flags)
 635 {
 636         struct backing_dev_info *bdi;
 637
 638         bdi = target->bt_mapping->backing_dev_info;
 639         if (bdi_read_congested(bdi))
 640                 return;
 641
 642         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 643         xfs_buf_read(target, ioff, isize, flags);
 644 }
 645
 646 xfs_buf_t *
 647 xfs_buf_get_empty(
 648         size_t                  len,
 649         xfs_buftarg_t           *target)
 650 {
 651         xfs_buf_t               *bp;
 652
 653         bp = xfs_buf_allocate(0);
 654         if (bp)
 655                 _xfs_buf_initialize(bp, target, 0, len, 0);
 656         return bp;
 657 }
 658
 659 static inline struct page *
 660 mem_to_page(
 661         void                    *addr)
 662 {
 663         if ((!is_vmalloc_addr(addr))) {
 664                 return virt_to_page(addr);
 665         } else {
 666                 return vmalloc_to_page(addr);
 667         }
 668 }
 669
 670 int
 671 xfs_buf_associate_memory(
 672         xfs_buf_t               *bp,
 673         void                    *mem,
 674         size_t                  len)
 675 {
 676         int                     rval;
 677         int                     i = 0;
 678         unsigned long           pageaddr;
 679         unsigned long           offset;
 680         size_t                  buflen;
 681         int                     page_count;
 682
 683         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
 684         offset = (unsigned long)mem - pageaddr;
 685         buflen = PAGE_CACHE_ALIGN(len + offset);
 686         page_count = buflen >> PAGE_CACHE_SHIFT;
 687
 688         /* Free any previous set of page pointers */
 689         if (bp->b_pages)
 690                 _xfs_buf_free_pages(bp);
 691
 692         bp->b_pages = NULL;
 693         bp->b_addr = mem;
 694
 695         rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
 696         if (rval)
 697                 return rval;
 698
 699         bp->b_offset = offset;
 700
 701         for (i = 0; i < bp->b_page_count; i++) {
 702                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 703                 pageaddr += PAGE_CACHE_SIZE;
 704         }
 705
 706         bp->b_count_desired = len;
 707         bp->b_buffer_length = buflen;
 708         bp->b_flags |= XBF_MAPPED;
 709         bp->b_flags &= ~_XBF_PAGE_LOCKED;
 710
 711         return 0;
 712 }
 713
 714 xfs_buf_t *
 715 xfs_buf_get_noaddr(
 716         size_t                  len,
 717         xfs_buftarg_t           *target)
 718 {
 719         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 720         int                     error, i;
 721         xfs_buf_t               *bp;
 722
 723         bp = xfs_buf_allocate(0);
 724         if (unlikely(bp == NULL))
 725                 goto fail;
 726         _xfs_buf_initialize(bp, target, 0, len, 0);
 727
 728         error = _xfs_buf_get_pages(bp, page_count, 0);
 729         if (error)
 730                 goto fail_free_buf;
 731
 732         for (i = 0; i < page_count; i++) {
 733                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
 734                 if (!bp->b_pages[i])
 735                         goto fail_free_mem;
 736         }
 737         bp->b_flags |= _XBF_PAGES;
 738
 739         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 740         if (unlikely(error)) {
 741                 printk(KERN_WARNING "%s: failed to map pages\n",
 742                                 __func__);
 743                 goto fail_free_mem;
 744         }
 745
 746         xfs_buf_unlock(bp);
 747
 748         trace_xfs_buf_get_noaddr(bp, _RET_IP_);
 749         return bp;
 750
 751  fail_free_mem:
 752         while (--i >= 0)
 753                 __free_page(bp->b_pages[i]);
 754         _xfs_buf_free_pages(bp);
 755  fail_free_buf:
 756         xfs_buf_deallocate(bp);
 757  fail:
 758         return NULL;
 759 }
 760
 761 /*
 762  *      Increment reference count on buffer, to hold the buffer concurrently
 763  *      with another thread which may release (free) the buffer asynchronously.
 764  *      Must hold the buffer already to call this function.
 765  */
 766 void
 767 xfs_buf_hold(
 768         xfs_buf_t               *bp)
 769 {
 770         trace_xfs_buf_hold(bp, _RET_IP_);
 771         atomic_inc(&bp->b_hold);
 772 }
 773
 774 /*
 775  *      Releases a hold on the specified buffer.  If the
 776  *      the hold count is 1, calls xfs_buf_free.
 777  */
 778 void
 779 xfs_buf_rele(
 780         xfs_buf_t               *bp)
 781 {
 782         xfs_bufhash_t           *hash = bp->b_hash;
 783
 784         trace_xfs_buf_rele(bp, _RET_IP_);
 785
 786         if (unlikely(!hash)) {
 787                 ASSERT(!bp->b_relse);
 788                 if (atomic_dec_and_test(&bp->b_hold))
 789                         xfs_buf_free(bp);
 790                 return;
 791         }
 792
 793         ASSERT(atomic_read(&bp->b_hold) > 0);
 794         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
 795                 if (bp->b_relse) {
 796                         atomic_inc(&bp->b_hold);
 797                         spin_unlock(&hash->bh_lock);
 798                         (*(bp->b_relse)) (bp);
 799                 } else if (bp->b_flags & XBF_FS_MANAGED) {
 800                         spin_unlock(&hash->bh_lock);
 801                 } else {
 802                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 803                         list_del_init(&bp->b_hash_list);
 804                         spin_unlock(&hash->bh_lock);
 805                         xfs_buf_free(bp);
 806                 }
 807         }
 808 }
 809
 810
 811 /*
 812  *      Mutual exclusion on buffers.  Locking model:
 813  *
 814  *      Buffers associated with inodes for which buffer locking
 815  *      is not enabled are not protected by semaphores, and are
 816  *      assumed to be exclusively owned by the caller.  There is a
 817  *      spinlock in the buffer, used by the caller when concurrent
 818  *      access is possible.
 819  */
 820
 821 /*
 822  *      Locks a buffer object, if it is not already locked.
 823  *      Note that this in no way locks the underlying pages, so it is only
 824  *      useful for synchronizing concurrent use of buffer objects, not for
 825  *      synchronizing independent access to the underlying pages.
 826  */
 827 int
 828 xfs_buf_cond_lock(
 829         xfs_buf_t               *bp)
 830 {
 831         int                     locked;
 832
 833         locked = down_trylock(&bp->b_sema) == 0;
 834         if (locked)
 835                 XB_SET_OWNER(bp);
 836
 837         trace_xfs_buf_cond_lock(bp, _RET_IP_);
 838         return locked ? 0 : -EBUSY;
 839 }
 840
 841 int
 842 xfs_buf_lock_value(
 843         xfs_buf_t               *bp)
 844 {
 845         return bp->b_sema.count;
 846 }
 847
 848 /*
 849  *      Locks a buffer object.
 850  *      Note that this in no way locks the underlying pages, so it is only
 851  *      useful for synchronizing concurrent use of buffer objects, not for
 852  *      synchronizing independent access to the underlying pages.
 853  *
 854  *      If we come across a stale, pinned, locked buffer, we know that we
 855  *      are being asked to lock a buffer that has been reallocated. Because
 856  *      it is pinned, we know that the log has not been pushed to disk and
 857  *      hence it will still be locked. Rather than sleeping until someone
 858  *      else pushes the log, push it ourselves before trying to get the lock.
 859  */
 860 void
 861 xfs_buf_lock(
 862         xfs_buf_t               *bp)
 863 {
 864         trace_xfs_buf_lock(bp, _RET_IP_);
 865
 866         if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 867                 xfs_log_force(bp->b_mount, 0);
 868         if (atomic_read(&bp->b_io_remaining))
 869                 blk_run_address_space(bp->b_target->bt_mapping);
 870         down(&bp->b_sema);
 871         XB_SET_OWNER(bp);
 872
 873         trace_xfs_buf_lock_done(bp, _RET_IP_);
 874 }
 875
 876 /*
 877  *      Releases the lock on the buffer object.
 878  *      If the buffer is marked delwri but is not queued, do so before we
 879  *      unlock the buffer as we need to set flags correctly.  We also need to
 880  *      take a reference for the delwri queue because the unlocker is going to
 881  *      drop their's and they don't know we just queued it.
 882  */
 883 void
 884 xfs_buf_unlock(
 885         xfs_buf_t               *bp)
 886 {
 887         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 888                 atomic_inc(&bp->b_hold);
 889                 bp->b_flags |= XBF_ASYNC;
 890                 xfs_buf_delwri_queue(bp, 0);
 891         }
 892
 893         XB_CLEAR_OWNER(bp);
 894         up(&bp->b_sema);
 895
 896         trace_xfs_buf_unlock(bp, _RET_IP_);
 897 }
 898
 899 STATIC void
 900 xfs_buf_wait_unpin(
 901         xfs_buf_t               *bp)
 902 {
 903         DECLARE_WAITQUEUE       (wait, current);
 904
 905         if (atomic_read(&bp->b_pin_count) == 0)
 906                 return;
 907
 908         add_wait_queue(&bp->b_waiters, &wait);
 909         for (;;) {
 910                 set_current_state(TASK_UNINTERRUPTIBLE);
 911                 if (atomic_read(&bp->b_pin_count) == 0)
 912                         break;
 913                 if (atomic_read(&bp->b_io_remaining))
 914                         blk_run_address_space(bp->b_target->bt_mapping);
 915                 schedule();
 916         }
 917         remove_wait_queue(&bp->b_waiters, &wait);
 918         set_current_state(TASK_RUNNING);
 919 }
 920
 921 /*
 922  *      Buffer Utility Routines
 923  */
 924
 925 STATIC void
 926 xfs_buf_iodone_work(
 927         struct work_struct      *work)
 928 {
 929         xfs_buf_t               *bp =
 930                 container_of(work, xfs_buf_t, b_iodone_work);
 931
 932         if (bp->b_iodone)
 933                 (*(bp->b_iodone))(bp);
 934         else if (bp->b_flags & XBF_ASYNC)
 935                 xfs_buf_relse(bp);
 936 }
 937
 938 void
 939 xfs_buf_ioend(
 940         xfs_buf_t               *bp,
 941         int                     schedule)
 942 {
 943         trace_xfs_buf_iodone(bp, _RET_IP_);
 944
 945         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
 946         if (bp->b_error == 0)
 947                 bp->b_flags |= XBF_DONE;
 948
 949         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
 950                 if (schedule) {
 951                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
 952                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
 953                 } else {
 954                         xfs_buf_iodone_work(&bp->b_iodone_work);
 955                 }
 956         } else {
 957                 complete(&bp->b_iowait);
 958         }
 959 }
 960
 961 void
 962 xfs_buf_ioerror(
 963         xfs_buf_t               *bp,
 964         int                     error)
 965 {
 966         ASSERT(error >= 0 && error <= 0xffff);
 967         bp->b_error = (unsigned short)error;
 968         trace_xfs_buf_ioerror(bp, error, _RET_IP_);
 969 }
 970
 971 int
 972 xfs_bwrite(
 973         struct xfs_mount        *mp,
 974         struct xfs_buf          *bp)
 975 {
 976         int                     error;
 977
 978         bp->b_mount = mp;
 979         bp->b_flags |= XBF_WRITE;
 980         bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
 981
 982         xfs_buf_delwri_dequeue(bp);
 983         xfs_bdstrat_cb(bp);
 984
 985         error = xfs_buf_iowait(bp);
 986         if (error)
 987                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
 988         xfs_buf_relse(bp);
 989         return error;
 990 }
 991
 992 void
 993 xfs_bdwrite(
 994         void                    *mp,
 995         struct xfs_buf          *bp)
 996 {
 997         trace_xfs_buf_bdwrite(bp, _RET_IP_);
 998
 999         bp->b_mount = mp;
1000
1001         bp->b_flags &= ~XBF_READ;
1002         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1003
1004         xfs_buf_delwri_queue(bp, 1);
1005 }
1006
1007 /*
1008  * Called when we want to stop a buffer from getting written or read.
1009  * We attach the EIO error, muck with its flags, and call biodone
1010  * so that the proper iodone callbacks get called.
1011  */
1012 STATIC int
1013 xfs_bioerror(
1014         xfs_buf_t *bp)
1015 {
1016 #ifdef XFSERRORDEBUG
1017         ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1018 #endif
1019
1020         /*
1021          * No need to wait until the buffer is unpinned, we aren't flushing it.
1022          */
1023         XFS_BUF_ERROR(bp, EIO);
1024
1025         /*
1026          * We're calling biodone, so delete XBF_DONE flag.
1027          */
1028         XFS_BUF_UNREAD(bp);
1029         XFS_BUF_UNDELAYWRITE(bp);
1030         XFS_BUF_UNDONE(bp);
1031         XFS_BUF_STALE(bp);
1032
1033         xfs_biodone(bp);
1034
1035         return EIO;
1036 }
1037
1038 /*
1039  * Same as xfs_bioerror, except that we are releasing the buffer
1040  * here ourselves, and avoiding the biodone call.
1041  * This is meant for userdata errors; metadata bufs come with
1042  * iodone functions attached, so that we can track down errors.
1043  */
1044 STATIC int
1045 xfs_bioerror_relse(
1046         struct xfs_buf  *bp)
1047 {
1048         int64_t         fl = XFS_BUF_BFLAGS(bp);
1049         /*
1050          * No need to wait until the buffer is unpinned.
1051          * We aren't flushing it.
1052          *
1053          * chunkhold expects B_DONE to be set, whether
1054          * we actually finish the I/O or not. We don't want to
1055          * change that interface.
1056          */
1057         XFS_BUF_UNREAD(bp);
1058         XFS_BUF_UNDELAYWRITE(bp);
1059         XFS_BUF_DONE(bp);
1060         XFS_BUF_STALE(bp);
1061         XFS_BUF_CLR_IODONE_FUNC(bp);
1062         if (!(fl & XBF_ASYNC)) {
1063                 /*
1064                  * Mark b_error and B_ERROR _both_.
1065                  * Lot's of chunkcache code assumes that.
1066                  * There's no reason to mark error for
1067                  * ASYNC buffers.
1068                  */
1069                 XFS_BUF_ERROR(bp, EIO);
1070                 XFS_BUF_FINISH_IOWAIT(bp);
1071         } else {
1072                 xfs_buf_relse(bp);
1073         }
1074
1075         return EIO;
1076 }
1077
1078
1079 /*
1080  * All xfs metadata buffers except log state machine buffers
1081  * get this attached as their b_bdstrat callback function.
1082  * This is so that we can catch a buffer
1083  * after prematurely unpinning it to forcibly shutdown the filesystem.
1084  */
1085 int
1086 xfs_bdstrat_cb(
1087         struct xfs_buf  *bp)
1088 {
1089         if (XFS_FORCED_SHUTDOWN(bp->b_mount)) {
1090                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1091                 /*
1092                  * Metadata write that didn't get logged but
1093                  * written delayed anyway. These aren't associated
1094                  * with a transaction, and can be ignored.
1095                  */
1096                 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1097                         return xfs_bioerror_relse(bp);
1098                 else
1099                         return xfs_bioerror(bp);
1100         }
1101
1102         xfs_buf_iorequest(bp);
1103         return 0;
1104 }
1105
1106 /*
1107  * Wrapper around bdstrat so that we can stop data from going to disk in case
1108  * we are shutting down the filesystem.  Typically user data goes thru this
1109  * path; one of the exceptions is the superblock.
1110  */
1111 void
1112 xfsbdstrat(
1113         struct xfs_mount        *mp,
1114         struct xfs_buf          *bp)
1115 {
1116         if (XFS_FORCED_SHUTDOWN(mp)) {
1117                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1118                 xfs_bioerror_relse(bp);
1119                 return;
1120         }
1121
1122         xfs_buf_iorequest(bp);
1123 }
1124
1125 STATIC void
1126 _xfs_buf_ioend(
1127         xfs_buf_t               *bp,
1128         int                     schedule)
1129 {
1130         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1131                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1132                 xfs_buf_ioend(bp, schedule);
1133         }
1134 }
1135
1136 STATIC void
1137 xfs_buf_bio_end_io(
1138         struct bio              *bio,
1139         int                     error)
1140 {
1141         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1142         unsigned int            blocksize = bp->b_target->bt_bsize;
1143         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1144
1145         xfs_buf_ioerror(bp, -error);
1146
1147         if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1148                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1149
1150         do {
1151                 struct page     *page = bvec->bv_page;
1152
1153                 ASSERT(!PagePrivate(page));
1154                 if (unlikely(bp->b_error)) {
1155                         if (bp->b_flags & XBF_READ)
1156                                 ClearPageUptodate(page);
1157                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1158                         SetPageUptodate(page);
1159                 } else if (!PagePrivate(page) &&
1160                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1161                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1162                 }
1163
1164                 if (--bvec >= bio->bi_io_vec)
1165                         prefetchw(&bvec->bv_page->flags);
1166
1167                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1168                         unlock_page(page);
1169         } while (bvec >= bio->bi_io_vec);
1170
1171         _xfs_buf_ioend(bp, 1);
1172         bio_put(bio);
1173 }
1174
1175 STATIC void
1176 _xfs_buf_ioapply(
1177         xfs_buf_t               *bp)
1178 {
1179         int                     rw, map_i, total_nr_pages, nr_pages;
1180         struct bio              *bio;
1181         int                     offset = bp->b_offset;
1182         int                     size = bp->b_count_desired;
1183         sector_t                sector = bp->b_bn;
1184         unsigned int            blocksize = bp->b_target->bt_bsize;
1185
1186         total_nr_pages = bp->b_page_count;
1187         map_i = 0;
1188
1189         if (bp->b_flags & XBF_ORDERED) {
1190                 ASSERT(!(bp->b_flags & XBF_READ));
1191                 rw = WRITE_FLUSH_FUA;
1192         } else if (bp->b_flags & XBF_LOG_BUFFER) {
1193                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1194                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1195                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1196         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1197                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1198                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1199                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1200         } else {
1201                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1202                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1203         }
1204
1205         /* Special code path for reading a sub page size buffer in --
1206          * we populate up the whole page, and hence the other metadata
1207          * in the same page.  This optimization is only valid when the
1208          * filesystem block size is not smaller than the page size.
1209          */
1210         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1211             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1212               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1213             (blocksize >= PAGE_CACHE_SIZE)) {
1214                 bio = bio_alloc(GFP_NOIO, 1);
1215
1216                 bio->bi_bdev = bp->b_target->bt_bdev;
1217                 bio->bi_sector = sector - (offset >> BBSHIFT);
1218                 bio->bi_end_io = xfs_buf_bio_end_io;
1219                 bio->bi_private = bp;
1220
1221                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1222                 size = 0;
1223
1224                 atomic_inc(&bp->b_io_remaining);
1225
1226                 goto submit_io;
1227         }
1228
1229 next_chunk:
1230         atomic_inc(&bp->b_io_remaining);
1231         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1232         if (nr_pages > total_nr_pages)
1233                 nr_pages = total_nr_pages;
1234
1235         bio = bio_alloc(GFP_NOIO, nr_pages);
1236         bio->bi_bdev = bp->b_target->bt_bdev;
1237         bio->bi_sector = sector;
1238         bio->bi_end_io = xfs_buf_bio_end_io;
1239         bio->bi_private = bp;
1240
1241         for (; size && nr_pages; nr_pages--, map_i++) {
1242                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1243
1244                 if (nbytes > size)
1245                         nbytes = size;
1246
1247                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1248                 if (rbytes < nbytes)
1249                         break;
1250
1251                 offset = 0;
1252                 sector += nbytes >> BBSHIFT;
1253                 size -= nbytes;
1254                 total_nr_pages--;
1255         }
1256
1257 submit_io:
1258         if (likely(bio->bi_size)) {
1259                 if (xfs_buf_is_vmapped(bp)) {
1260                         flush_kernel_vmap_range(bp->b_addr,
1261                                                 xfs_buf_vmap_len(bp));
1262                 }
1263                 submit_bio(rw, bio);
1264                 if (size)
1265                         goto next_chunk;
1266         } else {
1267                 /*
1268                  * if we get here, no pages were added to the bio. However,
1269                  * we can't just error out here - if the pages are locked then
1270                  * we have to unlock them otherwise we can hang on a later
1271                  * access to the page.
1272                  */
1273                 xfs_buf_ioerror(bp, EIO);
1274                 if (bp->b_flags & _XBF_PAGE_LOCKED) {
1275                         int i;
1276                         for (i = 0; i < bp->b_page_count; i++)
1277                                 unlock_page(bp->b_pages[i]);
1278                 }
1279                 bio_put(bio);
1280         }
1281 }
1282
1283 int
1284 xfs_buf_iorequest(
1285         xfs_buf_t               *bp)
1286 {
1287         trace_xfs_buf_iorequest(bp, _RET_IP_);
1288
1289         if (bp->b_flags & XBF_DELWRI) {
1290                 xfs_buf_delwri_queue(bp, 1);
1291                 return 0;
1292         }
1293
1294         if (bp->b_flags & XBF_WRITE) {
1295                 xfs_buf_wait_unpin(bp);
1296         }
1297
1298         xfs_buf_hold(bp);
1299
1300         /* Set the count to 1 initially, this will stop an I/O
1301          * completion callout which happens before we have started
1302          * all the I/O from calling xfs_buf_ioend too early.
1303          */
1304         atomic_set(&bp->b_io_remaining, 1);
1305         _xfs_buf_ioapply(bp);
1306         _xfs_buf_ioend(bp, 0);
1307
1308         xfs_buf_rele(bp);
1309         return 0;
1310 }
1311
1312 /*
1313  *      Waits for I/O to complete on the buffer supplied.
1314  *      It returns immediately if no I/O is pending.
1315  *      It returns the I/O error code, if any, or 0 if there was no error.
1316  */
1317 int
1318 xfs_buf_iowait(
1319         xfs_buf_t               *bp)
1320 {
1321         trace_xfs_buf_iowait(bp, _RET_IP_);
1322
1323         if (atomic_read(&bp->b_io_remaining))
1324                 blk_run_address_space(bp->b_target->bt_mapping);
1325         wait_for_completion(&bp->b_iowait);
1326
1327         trace_xfs_buf_iowait_done(bp, _RET_IP_);
1328         return bp->b_error;
1329 }
1330
1331 xfs_caddr_t
1332 xfs_buf_offset(
1333         xfs_buf_t               *bp,
1334         size_t                  offset)
1335 {
1336         struct page             *page;
1337
1338         if (bp->b_flags & XBF_MAPPED)
1339                 return XFS_BUF_PTR(bp) + offset;
1340
1341         offset += bp->b_offset;
1342         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1343         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1344 }
1345
1346 /*
1347  *      Move data into or out of a buffer.
1348  */
1349 void
1350 xfs_buf_iomove(
1351         xfs_buf_t               *bp,    /* buffer to process            */
1352         size_t                  boff,   /* starting buffer offset       */
1353         size_t                  bsize,  /* length to copy               */
1354         void                    *data,  /* data address                 */
1355         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1356 {
1357         size_t                  bend, cpoff, csize;
1358         struct page             *page;
1359
1360         bend = boff + bsize;
1361         while (boff < bend) {
1362                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1363                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1364                 csize = min_t(size_t,
1365                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1366
1367                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1368
1369                 switch (mode) {
1370                 case XBRW_ZERO:
1371                         memset(page_address(page) + cpoff, 0, csize);
1372                         break;
1373                 case XBRW_READ:
1374                         memcpy(data, page_address(page) + cpoff, csize);
1375                         break;
1376                 case XBRW_WRITE:
1377                         memcpy(page_address(page) + cpoff, data, csize);
1378                 }
1379
1380                 boff += csize;
1381                 data += csize;
1382         }
1383 }
1384
1385 /*
1386  *      Handling of buffer targets (buftargs).
1387  */
1388
1389 /*
1390  *      Wait for any bufs with callbacks that have been submitted but
1391  *      have not yet returned... walk the hash list for the target.
1392  */
1393 void
1394 xfs_wait_buftarg(
1395         xfs_buftarg_t   *btp)
1396 {
1397         xfs_buf_t       *bp, *n;
1398         xfs_bufhash_t   *hash;
1399         uint            i;
1400
1401         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1402                 hash = &btp->bt_hash[i];
1403 again:
1404                 spin_lock(&hash->bh_lock);
1405                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1406                         ASSERT(btp == bp->b_target);
1407                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1408                                 spin_unlock(&hash->bh_lock);
1409                                 /*
1410                                  * Catch superblock reference count leaks
1411                                  * immediately
1412                                  */
1413                                 BUG_ON(bp->b_bn == 0);
1414                                 delay(100);
1415                                 goto again;
1416                         }
1417                 }
1418                 spin_unlock(&hash->bh_lock);
1419         }
1420 }
1421
1422 /*
1423  *      Allocate buffer hash table for a given target.
1424  *      For devices containing metadata (i.e. not the log/realtime devices)
1425  *      we need to allocate a much larger hash table.
1426  */
1427 STATIC void
1428 xfs_alloc_bufhash(
1429         xfs_buftarg_t           *btp,
1430         int                     external)
1431 {
1432         unsigned int            i;
1433
1434         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1435         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1436         btp->bt_hash = kmem_zalloc_large((1 << btp->bt_hashshift) *
1437                                          sizeof(xfs_bufhash_t));
1438         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1439                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1440                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1441         }
1442 }
1443
1444 STATIC void
1445 xfs_free_bufhash(
1446         xfs_buftarg_t           *btp)
1447 {
1448         kmem_free_large(btp->bt_hash);
1449         btp->bt_hash = NULL;
1450 }
1451
1452 /*
1453  *      buftarg list for delwrite queue processing
1454  */
1455 static LIST_HEAD(xfs_buftarg_list);
1456 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1457
1458 STATIC void
1459 xfs_register_buftarg(
1460         xfs_buftarg_t           *btp)
1461 {
1462         spin_lock(&xfs_buftarg_lock);
1463         list_add(&btp->bt_list, &xfs_buftarg_list);
1464         spin_unlock(&xfs_buftarg_lock);
1465 }
1466
1467 STATIC void
1468 xfs_unregister_buftarg(
1469         xfs_buftarg_t           *btp)
1470 {
1471         spin_lock(&xfs_buftarg_lock);
1472         list_del(&btp->bt_list);
1473         spin_unlock(&xfs_buftarg_lock);
1474 }
1475
1476 void
1477 xfs_free_buftarg(
1478         struct xfs_mount        *mp,
1479         struct xfs_buftarg      *btp)
1480 {
1481         xfs_flush_buftarg(btp, 1);
1482         if (mp->m_flags & XFS_MOUNT_BARRIER)
1483                 xfs_blkdev_issue_flush(btp);
1484         xfs_free_bufhash(btp);
1485         iput(btp->bt_mapping->host);
1486
1487         /* Unregister the buftarg first so that we don't get a
1488          * wakeup finding a non-existent task
1489          */
1490         xfs_unregister_buftarg(btp);
1491         kthread_stop(btp->bt_task);
1492
1493         kmem_free(btp);
1494 }
1495
1496 STATIC int
1497 xfs_setsize_buftarg_flags(
1498         xfs_buftarg_t           *btp,
1499         unsigned int            blocksize,
1500         unsigned int            sectorsize,
1501         int                     verbose)
1502 {
1503         btp->bt_bsize = blocksize;
1504         btp->bt_sshift = ffs(sectorsize) - 1;
1505         btp->bt_smask = sectorsize - 1;
1506
1507         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1508                 printk(KERN_WARNING
1509                         "XFS: Cannot set_blocksize to %u on device %s\n",
1510                         sectorsize, XFS_BUFTARG_NAME(btp));
1511                 return EINVAL;
1512         }
1513
1514         if (verbose &&
1515             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1516                 printk(KERN_WARNING
1517                         "XFS: %u byte sectors in use on device %s.  "
1518                         "This is suboptimal; %u or greater is ideal.\n",
1519                         sectorsize, XFS_BUFTARG_NAME(btp),
1520                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1521         }
1522
1523         return 0;
1524 }
1525
1526 /*
1527  *      When allocating the initial buffer target we have not yet
1528  *      read in the superblock, so don't know what sized sectors
1529  *      are being used is at this early stage.  Play safe.
1530  */
1531 STATIC int
1532 xfs_setsize_buftarg_early(
1533         xfs_buftarg_t           *btp,
1534         struct block_device     *bdev)
1535 {
1536         return xfs_setsize_buftarg_flags(btp,
1537                         PAGE_CACHE_SIZE, bdev_logical_block_size(bdev), 0);
1538 }
1539
1540 int
1541 xfs_setsize_buftarg(
1542         xfs_buftarg_t           *btp,
1543         unsigned int            blocksize,
1544         unsigned int            sectorsize)
1545 {
1546         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1547 }
1548
1549 STATIC int
1550 xfs_mapping_buftarg(
1551         xfs_buftarg_t           *btp,
1552         struct block_device     *bdev)
1553 {
1554         struct backing_dev_info *bdi;
1555         struct inode            *inode;
1556         struct address_space    *mapping;
1557         static const struct address_space_operations mapping_aops = {
1558                 .sync_page = block_sync_page,
1559                 .migratepage = fail_migrate_page,
1560         };
1561
1562         inode = new_inode(bdev->bd_inode->i_sb);
1563         if (!inode) {
1564                 printk(KERN_WARNING
1565                         "XFS: Cannot allocate mapping inode for device %s\n",
1566                         XFS_BUFTARG_NAME(btp));
1567                 return ENOMEM;
1568         }
1569         inode->i_mode = S_IFBLK;
1570         inode->i_bdev = bdev;
1571         inode->i_rdev = bdev->bd_dev;
1572         bdi = blk_get_backing_dev_info(bdev);
1573         if (!bdi)
1574                 bdi = &default_backing_dev_info;
1575         mapping = &inode->i_data;
1576         mapping->a_ops = &mapping_aops;
1577         mapping->backing_dev_info = bdi;
1578         mapping_set_gfp_mask(mapping, GFP_NOFS);
1579         btp->bt_mapping = mapping;
1580         return 0;
1581 }
1582
1583 STATIC int
1584 xfs_alloc_delwrite_queue(
1585         xfs_buftarg_t           *btp,
1586         const char              *fsname)
1587 {
1588         int     error = 0;
1589
1590         INIT_LIST_HEAD(&btp->bt_list);
1591         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1592         spin_lock_init(&btp->bt_delwrite_lock);
1593         btp->bt_flags = 0;
1594         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1595         if (IS_ERR(btp->bt_task)) {
1596                 error = PTR_ERR(btp->bt_task);
1597                 goto out_error;
1598         }
1599         xfs_register_buftarg(btp);
1600 out_error:
1601         return error;
1602 }
1603
1604 xfs_buftarg_t *
1605 xfs_alloc_buftarg(
1606         struct block_device     *bdev,
1607         int                     external,
1608         const char              *fsname)
1609 {
1610         xfs_buftarg_t           *btp;
1611
1612         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1613
1614         btp->bt_dev =  bdev->bd_dev;
1615         btp->bt_bdev = bdev;
1616         if (xfs_setsize_buftarg_early(btp, bdev))
1617                 goto error;
1618         if (xfs_mapping_buftarg(btp, bdev))
1619                 goto error;
1620         if (xfs_alloc_delwrite_queue(btp, fsname))
1621                 goto error;
1622         xfs_alloc_bufhash(btp, external);
1623         return btp;
1624
1625 error:
1626         kmem_free(btp);
1627         return NULL;
1628 }
1629
1630
1631 /*
1632  *      Delayed write buffer handling
1633  */
1634 STATIC void
1635 xfs_buf_delwri_queue(
1636         xfs_buf_t               *bp,
1637         int                     unlock)
1638 {
1639         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1640         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1641
1642         trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1643
1644         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1645
1646         spin_lock(dwlk);
1647         /* If already in the queue, dequeue and place at tail */
1648         if (!list_empty(&bp->b_list)) {
1649                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1650                 if (unlock)
1651                         atomic_dec(&bp->b_hold);
1652                 list_del(&bp->b_list);
1653         }
1654
1655         if (list_empty(dwq)) {
1656                 /* start xfsbufd as it is about to have something to do */
1657                 wake_up_process(bp->b_target->bt_task);
1658         }
1659
1660         bp->b_flags |= _XBF_DELWRI_Q;
1661         list_add_tail(&bp->b_list, dwq);
1662         bp->b_queuetime = jiffies;
1663         spin_unlock(dwlk);
1664
1665         if (unlock)
1666                 xfs_buf_unlock(bp);
1667 }
1668
1669 void
1670 xfs_buf_delwri_dequeue(
1671         xfs_buf_t               *bp)
1672 {
1673         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1674         int                     dequeued = 0;
1675
1676         spin_lock(dwlk);
1677         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1678                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1679                 list_del_init(&bp->b_list);
1680                 dequeued = 1;
1681         }
1682         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1683         spin_unlock(dwlk);
1684
1685         if (dequeued)
1686                 xfs_buf_rele(bp);
1687
1688         trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1689 }
1690
1691 /*
1692  * If a delwri buffer needs to be pushed before it has aged out, then promote
1693  * it to the head of the delwri queue so that it will be flushed on the next
1694  * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1695  * than the age currently needed to flush the buffer. Hence the next time the
1696  * xfsbufd sees it is guaranteed to be considered old enough to flush.
1697  */
1698 void
1699 xfs_buf_delwri_promote(
1700         struct xfs_buf  *bp)
1701 {
1702         struct xfs_buftarg *btp = bp->b_target;
1703         long            age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1704
1705         ASSERT(bp->b_flags & XBF_DELWRI);
1706         ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1707
1708         /*
1709          * Check the buffer age before locking the delayed write queue as we
1710          * don't need to promote buffers that are already past the flush age.
1711          */
1712         if (bp->b_queuetime < jiffies - age)
1713                 return;
1714         bp->b_queuetime = jiffies - age;
1715         spin_lock(&btp->bt_delwrite_lock);
1716         list_move(&bp->b_list, &btp->bt_delwrite_queue);
1717         spin_unlock(&btp->bt_delwrite_lock);
1718 }
1719
1720 STATIC void
1721 xfs_buf_runall_queues(
1722         struct workqueue_struct *queue)
1723 {
1724         flush_workqueue(queue);
1725 }
1726
1727 STATIC int
1728 xfsbufd_wakeup(
1729         struct shrinker         *shrink,
1730         int                     priority,
1731         gfp_t                   mask)
1732 {
1733         xfs_buftarg_t           *btp;
1734
1735         spin_lock(&xfs_buftarg_lock);
1736         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1737                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1738                         continue;
1739                 if (list_empty(&btp->bt_delwrite_queue))
1740                         continue;
1741                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1742                 wake_up_process(btp->bt_task);
1743         }
1744         spin_unlock(&xfs_buftarg_lock);
1745         return 0;
1746 }
1747
1748 /*
1749  * Move as many buffers as specified to the supplied list
1750  * idicating if we skipped any buffers to prevent deadlocks.
1751  */
1752 STATIC int
1753 xfs_buf_delwri_split(
1754         xfs_buftarg_t   *target,
1755         struct list_head *list,
1756         unsigned long   age)
1757 {
1758         xfs_buf_t       *bp, *n;
1759         struct list_head *dwq = &target->bt_delwrite_queue;
1760         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1761         int             skipped = 0;
1762         int             force;
1763
1764         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1765         INIT_LIST_HEAD(list);
1766         spin_lock(dwlk);
1767         list_for_each_entry_safe(bp, n, dwq, b_list) {
1768                 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1769                 ASSERT(bp->b_flags & XBF_DELWRI);
1770
1771                 if (!XFS_BUF_ISPINNED(bp) && !xfs_buf_cond_lock(bp)) {
1772                         if (!force &&
1773                             time_before(jiffies, bp->b_queuetime + age)) {
1774                                 xfs_buf_unlock(bp);
1775                                 break;
1776                         }
1777
1778                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1779                                          _XBF_RUN_QUEUES);
1780                         bp->b_flags |= XBF_WRITE;
1781                         list_move_tail(&bp->b_list, list);
1782                 } else
1783                         skipped++;
1784         }
1785         spin_unlock(dwlk);
1786
1787         return skipped;
1788
1789 }
1790
1791 /*
1792  * Compare function is more complex than it needs to be because
1793  * the return value is only 32 bits and we are doing comparisons
1794  * on 64 bit values
1795  */
1796 static int
1797 xfs_buf_cmp(
1798         void            *priv,
1799         struct list_head *a,
1800         struct list_head *b)
1801 {
1802         struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1803         struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1804         xfs_daddr_t             diff;
1805
1806         diff = ap->b_bn - bp->b_bn;
1807         if (diff < 0)
1808                 return -1;
1809         if (diff > 0)
1810                 return 1;
1811         return 0;
1812 }
1813
1814 void
1815 xfs_buf_delwri_sort(
1816         xfs_buftarg_t   *target,
1817         struct list_head *list)
1818 {
1819         list_sort(NULL, list, xfs_buf_cmp);
1820 }
1821
1822 STATIC int
1823 xfsbufd(
1824         void            *data)
1825 {
1826         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1827
1828         current->flags |= PF_MEMALLOC;
1829
1830         set_freezable();
1831
1832         do {
1833                 long    age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1834                 long    tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1835                 int     count = 0;
1836                 struct list_head tmp;
1837
1838                 if (unlikely(freezing(current))) {
1839                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1840                         refrigerator();
1841                 } else {
1842                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1843                 }
1844
1845                 /* sleep for a long time if there is nothing to do. */
1846                 if (list_empty(&target->bt_delwrite_queue))
1847                         tout = MAX_SCHEDULE_TIMEOUT;
1848                 schedule_timeout_interruptible(tout);
1849
1850                 xfs_buf_delwri_split(target, &tmp, age);
1851                 list_sort(NULL, &tmp, xfs_buf_cmp);
1852                 while (!list_empty(&tmp)) {
1853                         struct xfs_buf *bp;
1854                         bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1855                         list_del_init(&bp->b_list);
1856                         xfs_bdstrat_cb(bp);
1857                         count++;
1858                 }
1859                 if (count)
1860                         blk_run_address_space(target->bt_mapping);
1861
1862         } while (!kthread_should_stop());
1863
1864         return 0;
1865 }
1866
1867 /*
1868  *      Go through all incore buffers, and release buffers if they belong to
1869  *      the given device. This is used in filesystem error handling to
1870  *      preserve the consistency of its metadata.
1871  */
1872 int
1873 xfs_flush_buftarg(
1874         xfs_buftarg_t   *target,
1875         int             wait)
1876 {
1877         xfs_buf_t       *bp;
1878         int             pincount = 0;
1879         LIST_HEAD(tmp_list);
1880         LIST_HEAD(wait_list);
1881
1882         xfs_buf_runall_queues(xfsconvertd_workqueue);
1883         xfs_buf_runall_queues(xfsdatad_workqueue);
1884         xfs_buf_runall_queues(xfslogd_workqueue);
1885
1886         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1887         pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1888
1889         /*
1890          * Dropped the delayed write list lock, now walk the temporary list.
1891          * All I/O is issued async and then if we need to wait for completion
1892          * we do that after issuing all the IO.
1893          */
1894         list_sort(NULL, &tmp_list, xfs_buf_cmp);
1895         while (!list_empty(&tmp_list)) {
1896                 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1897                 ASSERT(target == bp->b_target);
1898                 list_del_init(&bp->b_list);
1899                 if (wait) {
1900                         bp->b_flags &= ~XBF_ASYNC;
1901                         list_add(&bp->b_list, &wait_list);
1902                 }
1903                 xfs_bdstrat_cb(bp);
1904         }
1905
1906         if (wait) {
1907                 /* Expedite and wait for IO to complete. */
1908                 blk_run_address_space(target->bt_mapping);
1909                 while (!list_empty(&wait_list)) {
1910                         bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1911
1912                         list_del_init(&bp->b_list);
1913                         xfs_iowait(bp);
1914                         xfs_buf_relse(bp);
1915                 }
1916         }
1917
1918         return pincount;
1919 }
1920
1921 int __init
1922 xfs_buf_init(void)
1923 {
1924         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1925                                                 KM_ZONE_HWALIGN, NULL);
1926         if (!xfs_buf_zone)
1927                 goto out;
1928
1929         xfslogd_workqueue = create_workqueue("xfslogd");
1930         if (!xfslogd_workqueue)
1931                 goto out_free_buf_zone;
1932
1933         xfsdatad_workqueue = create_workqueue("xfsdatad");
1934         if (!xfsdatad_workqueue)
1935                 goto out_destroy_xfslogd_workqueue;
1936
1937         xfsconvertd_workqueue = create_workqueue("xfsconvertd");
1938         if (!xfsconvertd_workqueue)
1939                 goto out_destroy_xfsdatad_workqueue;
1940
1941         register_shrinker(&xfs_buf_shake);
1942         return 0;
1943
1944  out_destroy_xfsdatad_workqueue:
1945         destroy_workqueue(xfsdatad_workqueue);
1946  out_destroy_xfslogd_workqueue:
1947         destroy_workqueue(xfslogd_workqueue);
1948  out_free_buf_zone:
1949         kmem_zone_destroy(xfs_buf_zone);
1950  out:
1951         return -ENOMEM;
1952 }
1953
1954 void
1955 xfs_buf_terminate(void)
1956 {
1957         unregister_shrinker(&xfs_buf_shake);
1958         destroy_workqueue(xfsconvertd_workqueue);
1959         destroy_workqueue(xfsdatad_workqueue);
1960         destroy_workqueue(xfslogd_workqueue);
1961         kmem_zone_destroy(xfs_buf_zone);
1962 }
1963
1964 #ifdef CONFIG_KDB_MODULES
1965 struct list_head *
1966 xfs_get_buftarg_list(void)
1967 {
1968         return &xfs_buftarg_list;
1969 }
1970 #endif