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/slab.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
  37 STATIC kmem_zone_t *xfs_buf_zone;
  38 STATIC kmem_shaker_t xfs_buf_shake;
  39 STATIC int xfsbufd(void *);
  40 STATIC int xfsbufd_wakeup(int, gfp_t);
  41 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
  42
  43 STATIC struct workqueue_struct *xfslogd_workqueue;
  44 struct workqueue_struct *xfsdatad_workqueue;
  45
  46 #ifdef XFS_BUF_TRACE
  47 void
  48 xfs_buf_trace(
  49         xfs_buf_t       *bp,
  50         char            *id,
  51         void            *data,
  52         void            *ra)
  53 {
  54         ktrace_enter(xfs_buf_trace_buf,
  55                 bp, id,
  56                 (void *)(unsigned long)bp->b_flags,
  57                 (void *)(unsigned long)bp->b_hold.counter,
  58                 (void *)(unsigned long)bp->b_sema.count.counter,
  59                 (void *)current,
  60                 data, ra,
  61                 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
  62                 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
  63                 (void *)(unsigned long)bp->b_buffer_length,
  64                 NULL, NULL, NULL, NULL, NULL);
  65 }
  66 ktrace_t *xfs_buf_trace_buf;
  67 #define XFS_BUF_TRACE_SIZE      4096
  68 #define XB_TRACE(bp, id, data)  \
  69         xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
  70 #else
  71 #define XB_TRACE(bp, id, data)  do { } while (0)
  72 #endif
  73
  74 #ifdef XFS_BUF_LOCK_TRACKING
  75 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  76 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  77 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  78 #else
  79 # define XB_SET_OWNER(bp)       do { } while (0)
  80 # define XB_CLEAR_OWNER(bp)     do { } while (0)
  81 # define XB_GET_OWNER(bp)       do { } while (0)
  82 #endif
  83
  84 #define xb_to_gfp(flags) \
  85         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
  86           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
  87
  88 #define xb_to_km(flags) \
  89          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
  90
  91 #define xfs_buf_allocate(flags) \
  92         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
  93 #define xfs_buf_deallocate(bp) \
  94         kmem_zone_free(xfs_buf_zone, (bp));
  95
  96 /*
  97  *      Page Region interfaces.
  98  *
  99  *      For pages in filesystems where the blocksize is smaller than the
 100  *      pagesize, we use the page->private field (long) to hold a bitmap
 101  *      of uptodate regions within the page.
 102  *
 103  *      Each such region is "bytes per page / bits per long" bytes long.
 104  *
 105  *      NBPPR == number-of-bytes-per-page-region
 106  *      BTOPR == bytes-to-page-region (rounded up)
 107  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
 108  */
 109 #if (BITS_PER_LONG == 32)
 110 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
 111 #elif (BITS_PER_LONG == 64)
 112 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
 113 #else
 114 #error BITS_PER_LONG must be 32 or 64
 115 #endif
 116 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
 117 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
 118 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
 119
 120 STATIC unsigned long
 121 page_region_mask(
 122         size_t          offset,
 123         size_t          length)
 124 {
 125         unsigned long   mask;
 126         int             first, final;
 127
 128         first = BTOPR(offset);
 129         final = BTOPRT(offset + length - 1);
 130         first = min(first, final);
 131
 132         mask = ~0UL;
 133         mask <<= BITS_PER_LONG - (final - first);
 134         mask >>= BITS_PER_LONG - (final);
 135
 136         ASSERT(offset + length <= PAGE_CACHE_SIZE);
 137         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
 138
 139         return mask;
 140 }
 141
 142 STATIC inline void
 143 set_page_region(
 144         struct page     *page,
 145         size_t          offset,
 146         size_t          length)
 147 {
 148         set_page_private(page,
 149                 page_private(page) | page_region_mask(offset, length));
 150         if (page_private(page) == ~0UL)
 151                 SetPageUptodate(page);
 152 }
 153
 154 STATIC inline int
 155 test_page_region(
 156         struct page     *page,
 157         size_t          offset,
 158         size_t          length)
 159 {
 160         unsigned long   mask = page_region_mask(offset, length);
 161
 162         return (mask && (page_private(page) & mask) == mask);
 163 }
 164
 165 /*
 166  *      Mapping of multi-page buffers into contiguous virtual space
 167  */
 168
 169 typedef struct a_list {
 170         void            *vm_addr;
 171         struct a_list   *next;
 172 } a_list_t;
 173
 174 STATIC a_list_t         *as_free_head;
 175 STATIC int              as_list_len;
 176 STATIC DEFINE_SPINLOCK(as_lock);
 177
 178 /*
 179  *      Try to batch vunmaps because they are costly.
 180  */
 181 STATIC void
 182 free_address(
 183         void            *addr)
 184 {
 185         a_list_t        *aentry;
 186
 187         aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
 188         if (likely(aentry)) {
 189                 spin_lock(&as_lock);
 190                 aentry->next = as_free_head;
 191                 aentry->vm_addr = addr;
 192                 as_free_head = aentry;
 193                 as_list_len++;
 194                 spin_unlock(&as_lock);
 195         } else {
 196                 vunmap(addr);
 197         }
 198 }
 199
 200 STATIC void
 201 purge_addresses(void)
 202 {
 203         a_list_t        *aentry, *old;
 204
 205         if (as_free_head == NULL)
 206                 return;
 207
 208         spin_lock(&as_lock);
 209         aentry = as_free_head;
 210         as_free_head = NULL;
 211         as_list_len = 0;
 212         spin_unlock(&as_lock);
 213
 214         while ((old = aentry) != NULL) {
 215                 vunmap(aentry->vm_addr);
 216                 aentry = aentry->next;
 217                 kfree(old);
 218         }
 219 }
 220
 221 /*
 222  *      Internal xfs_buf_t object manipulation
 223  */
 224
 225 STATIC void
 226 _xfs_buf_initialize(
 227         xfs_buf_t               *bp,
 228         xfs_buftarg_t           *target,
 229         xfs_off_t               range_base,
 230         size_t                  range_length,
 231         xfs_buf_flags_t         flags)
 232 {
 233         /*
 234          * We don't want certain flags to appear in b_flags.
 235          */
 236         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
 237
 238         memset(bp, 0, sizeof(xfs_buf_t));
 239         atomic_set(&bp->b_hold, 1);
 240         init_MUTEX_LOCKED(&bp->b_iodonesema);
 241         INIT_LIST_HEAD(&bp->b_list);
 242         INIT_LIST_HEAD(&bp->b_hash_list);
 243         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
 244         XB_SET_OWNER(bp);
 245         bp->b_target = target;
 246         bp->b_file_offset = range_base;
 247         /*
 248          * Set buffer_length and count_desired to the same value initially.
 249          * I/O routines should use count_desired, which will be the same in
 250          * most cases but may be reset (e.g. XFS recovery).
 251          */
 252         bp->b_buffer_length = bp->b_count_desired = range_length;
 253         bp->b_flags = flags;
 254         bp->b_bn = XFS_BUF_DADDR_NULL;
 255         atomic_set(&bp->b_pin_count, 0);
 256         init_waitqueue_head(&bp->b_waiters);
 257
 258         XFS_STATS_INC(xb_create);
 259         XB_TRACE(bp, "initialize", target);
 260 }
 261
 262 /*
 263  *      Allocate a page array capable of holding a specified number
 264  *      of pages, and point the page buf at it.
 265  */
 266 STATIC int
 267 _xfs_buf_get_pages(
 268         xfs_buf_t               *bp,
 269         int                     page_count,
 270         xfs_buf_flags_t         flags)
 271 {
 272         /* Make sure that we have a page list */
 273         if (bp->b_pages == NULL) {
 274                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
 275                 bp->b_page_count = page_count;
 276                 if (page_count <= XB_PAGES) {
 277                         bp->b_pages = bp->b_page_array;
 278                 } else {
 279                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
 280                                         page_count, xb_to_km(flags));
 281                         if (bp->b_pages == NULL)
 282                                 return -ENOMEM;
 283                 }
 284                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 285         }
 286         return 0;
 287 }
 288
 289 /*
 290  *      Frees b_pages if it was allocated.
 291  */
 292 STATIC void
 293 _xfs_buf_free_pages(
 294         xfs_buf_t       *bp)
 295 {
 296         if (bp->b_pages != bp->b_page_array) {
 297                 kmem_free(bp->b_pages,
 298                           bp->b_page_count * sizeof(struct page *));
 299         }
 300 }
 301
 302 /*
 303  *      Releases the specified buffer.
 304  *
 305  *      The modification state of any associated pages is left unchanged.
 306  *      The buffer most not be on any hash - use xfs_buf_rele instead for
 307  *      hashed and refcounted buffers
 308  */
 309 void
 310 xfs_buf_free(
 311         xfs_buf_t               *bp)
 312 {
 313         XB_TRACE(bp, "free", 0);
 314
 315         ASSERT(list_empty(&bp->b_hash_list));
 316
 317         if (bp->b_flags & _XBF_PAGE_CACHE) {
 318                 uint            i;
 319
 320                 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
 321                         free_address(bp->b_addr - bp->b_offset);
 322
 323                 for (i = 0; i < bp->b_page_count; i++) {
 324                         struct page     *page = bp->b_pages[i];
 325
 326                         ASSERT(!PagePrivate(page));
 327                         page_cache_release(page);
 328                 }
 329                 _xfs_buf_free_pages(bp);
 330         } else if (bp->b_flags & _XBF_KMEM_ALLOC) {
 331                  /*
 332                   * XXX(hch): bp->b_count_desired might be incorrect (see
 333                   * xfs_buf_associate_memory for details), but fortunately
 334                   * the Linux version of kmem_free ignores the len argument..
 335                   */
 336                 kmem_free(bp->b_addr, bp->b_count_desired);
 337                 _xfs_buf_free_pages(bp);
 338         }
 339
 340         xfs_buf_deallocate(bp);
 341 }
 342
 343 /*
 344  *      Finds all pages for buffer in question and builds it's page list.
 345  */
 346 STATIC int
 347 _xfs_buf_lookup_pages(
 348         xfs_buf_t               *bp,
 349         uint                    flags)
 350 {
 351         struct address_space    *mapping = bp->b_target->bt_mapping;
 352         size_t                  blocksize = bp->b_target->bt_bsize;
 353         size_t                  size = bp->b_count_desired;
 354         size_t                  nbytes, offset;
 355         gfp_t                   gfp_mask = xb_to_gfp(flags);
 356         unsigned short          page_count, i;
 357         pgoff_t                 first;
 358         xfs_off_t               end;
 359         int                     error;
 360
 361         end = bp->b_file_offset + bp->b_buffer_length;
 362         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 363
 364         error = _xfs_buf_get_pages(bp, page_count, flags);
 365         if (unlikely(error))
 366                 return error;
 367         bp->b_flags |= _XBF_PAGE_CACHE;
 368
 369         offset = bp->b_offset;
 370         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
 371
 372         for (i = 0; i < bp->b_page_count; i++) {
 373                 struct page     *page;
 374                 uint            retries = 0;
 375
 376               retry:
 377                 page = find_or_create_page(mapping, first + i, gfp_mask);
 378                 if (unlikely(page == NULL)) {
 379                         if (flags & XBF_READ_AHEAD) {
 380                                 bp->b_page_count = i;
 381                                 for (i = 0; i < bp->b_page_count; i++)
 382                                         unlock_page(bp->b_pages[i]);
 383                                 return -ENOMEM;
 384                         }
 385
 386                         /*
 387                          * This could deadlock.
 388                          *
 389                          * But until all the XFS lowlevel code is revamped to
 390                          * handle buffer allocation failures we can't do much.
 391                          */
 392                         if (!(++retries % 100))
 393                                 printk(KERN_ERR
 394                                         "XFS: possible memory allocation "
 395                                         "deadlock in %s (mode:0x%x)\n",
 396                                         __FUNCTION__, gfp_mask);
 397
 398                         XFS_STATS_INC(xb_page_retries);
 399                         xfsbufd_wakeup(0, gfp_mask);
 400                         congestion_wait(WRITE, HZ/50);
 401                         goto retry;
 402                 }
 403
 404                 XFS_STATS_INC(xb_page_found);
 405
 406                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
 407                 size -= nbytes;
 408
 409                 ASSERT(!PagePrivate(page));
 410                 if (!PageUptodate(page)) {
 411                         page_count--;
 412                         if (blocksize >= PAGE_CACHE_SIZE) {
 413                                 if (flags & XBF_READ)
 414                                         bp->b_locked = 1;
 415                         } else if (!PagePrivate(page)) {
 416                                 if (test_page_region(page, offset, nbytes))
 417                                         page_count++;
 418                         }
 419                 }
 420
 421                 bp->b_pages[i] = page;
 422                 offset = 0;
 423         }
 424
 425         if (!bp->b_locked) {
 426                 for (i = 0; i < bp->b_page_count; i++)
 427                         unlock_page(bp->b_pages[i]);
 428         }
 429
 430         if (page_count == bp->b_page_count)
 431                 bp->b_flags |= XBF_DONE;
 432
 433         XB_TRACE(bp, "lookup_pages", (long)page_count);
 434         return error;
 435 }
 436
 437 /*
 438  *      Map buffer into kernel address-space if nessecary.
 439  */
 440 STATIC int
 441 _xfs_buf_map_pages(
 442         xfs_buf_t               *bp,
 443         uint                    flags)
 444 {
 445         /* A single page buffer is always mappable */
 446         if (bp->b_page_count == 1) {
 447                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 448                 bp->b_flags |= XBF_MAPPED;
 449         } else if (flags & XBF_MAPPED) {
 450                 if (as_list_len > 64)
 451                         purge_addresses();
 452                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
 453                                         VM_MAP, PAGE_KERNEL);
 454                 if (unlikely(bp->b_addr == NULL))
 455                         return -ENOMEM;
 456                 bp->b_addr += bp->b_offset;
 457                 bp->b_flags |= XBF_MAPPED;
 458         }
 459
 460         return 0;
 461 }
 462
 463 /*
 464  *      Finding and Reading Buffers
 465  */
 466
 467 /*
 468  *      Look up, and creates if absent, a lockable buffer for
 469  *      a given range of an inode.  The buffer is returned
 470  *      locked.  If other overlapping buffers exist, they are
 471  *      released before the new buffer is created and locked,
 472  *      which may imply that this call will block until those buffers
 473  *      are unlocked.  No I/O is implied by this call.
 474  */
 475 xfs_buf_t *
 476 _xfs_buf_find(
 477         xfs_buftarg_t           *btp,   /* block device target          */
 478         xfs_off_t               ioff,   /* starting offset of range     */
 479         size_t                  isize,  /* length of range              */
 480         xfs_buf_flags_t         flags,
 481         xfs_buf_t               *new_bp)
 482 {
 483         xfs_off_t               range_base;
 484         size_t                  range_length;
 485         xfs_bufhash_t           *hash;
 486         xfs_buf_t               *bp, *n;
 487
 488         range_base = (ioff << BBSHIFT);
 489         range_length = (isize << BBSHIFT);
 490
 491         /* Check for IOs smaller than the sector size / not sector aligned */
 492         ASSERT(!(range_length < (1 << btp->bt_sshift)));
 493         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
 494
 495         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
 496
 497         spin_lock(&hash->bh_lock);
 498
 499         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
 500                 ASSERT(btp == bp->b_target);
 501                 if (bp->b_file_offset == range_base &&
 502                     bp->b_buffer_length == range_length) {
 503                         /*
 504                          * If we look at something, bring it to the
 505                          * front of the list for next time.
 506                          */
 507                         atomic_inc(&bp->b_hold);
 508                         list_move(&bp->b_hash_list, &hash->bh_list);
 509                         goto found;
 510                 }
 511         }
 512
 513         /* No match found */
 514         if (new_bp) {
 515                 _xfs_buf_initialize(new_bp, btp, range_base,
 516                                 range_length, flags);
 517                 new_bp->b_hash = hash;
 518                 list_add(&new_bp->b_hash_list, &hash->bh_list);
 519         } else {
 520                 XFS_STATS_INC(xb_miss_locked);
 521         }
 522
 523         spin_unlock(&hash->bh_lock);
 524         return new_bp;
 525
 526 found:
 527         spin_unlock(&hash->bh_lock);
 528
 529         /* Attempt to get the semaphore without sleeping,
 530          * if this does not work then we need to drop the
 531          * spinlock and do a hard attempt on the semaphore.
 532          */
 533         if (down_trylock(&bp->b_sema)) {
 534                 if (!(flags & XBF_TRYLOCK)) {
 535                         /* wait for buffer ownership */
 536                         XB_TRACE(bp, "get_lock", 0);
 537                         xfs_buf_lock(bp);
 538                         XFS_STATS_INC(xb_get_locked_waited);
 539                 } else {
 540                         /* We asked for a trylock and failed, no need
 541                          * to look at file offset and length here, we
 542                          * know that this buffer at least overlaps our
 543                          * buffer and is locked, therefore our buffer
 544                          * either does not exist, or is this buffer.
 545                          */
 546                         xfs_buf_rele(bp);
 547                         XFS_STATS_INC(xb_busy_locked);
 548                         return NULL;
 549                 }
 550         } else {
 551                 /* trylock worked */
 552                 XB_SET_OWNER(bp);
 553         }
 554
 555         if (bp->b_flags & XBF_STALE) {
 556                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 557                 bp->b_flags &= XBF_MAPPED;
 558         }
 559         XB_TRACE(bp, "got_lock", 0);
 560         XFS_STATS_INC(xb_get_locked);
 561         return bp;
 562 }
 563
 564 /*
 565  *      Assembles a buffer covering the specified range.
 566  *      Storage in memory for all portions of the buffer will be allocated,
 567  *      although backing storage may not be.
 568  */
 569 xfs_buf_t *
 570 xfs_buf_get_flags(
 571         xfs_buftarg_t           *target,/* target for buffer            */
 572         xfs_off_t               ioff,   /* starting offset of range     */
 573         size_t                  isize,  /* length of range              */
 574         xfs_buf_flags_t         flags)
 575 {
 576         xfs_buf_t               *bp, *new_bp;
 577         int                     error = 0, i;
 578
 579         new_bp = xfs_buf_allocate(flags);
 580         if (unlikely(!new_bp))
 581                 return NULL;
 582
 583         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 584         if (bp == new_bp) {
 585                 error = _xfs_buf_lookup_pages(bp, flags);
 586                 if (error)
 587                         goto no_buffer;
 588         } else {
 589                 xfs_buf_deallocate(new_bp);
 590                 if (unlikely(bp == NULL))
 591                         return NULL;
 592         }
 593
 594         for (i = 0; i < bp->b_page_count; i++)
 595                 mark_page_accessed(bp->b_pages[i]);
 596
 597         if (!(bp->b_flags & XBF_MAPPED)) {
 598                 error = _xfs_buf_map_pages(bp, flags);
 599                 if (unlikely(error)) {
 600                         printk(KERN_WARNING "%s: failed to map pages\n",
 601                                         __FUNCTION__);
 602                         goto no_buffer;
 603                 }
 604         }
 605
 606         XFS_STATS_INC(xb_get);
 607
 608         /*
 609          * Always fill in the block number now, the mapped cases can do
 610          * their own overlay of this later.
 611          */
 612         bp->b_bn = ioff;
 613         bp->b_count_desired = bp->b_buffer_length;
 614
 615         XB_TRACE(bp, "get", (unsigned long)flags);
 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 xfs_buf_t *
 626 xfs_buf_read_flags(
 627         xfs_buftarg_t           *target,
 628         xfs_off_t               ioff,
 629         size_t                  isize,
 630         xfs_buf_flags_t         flags)
 631 {
 632         xfs_buf_t               *bp;
 633
 634         flags |= XBF_READ;
 635
 636         bp = xfs_buf_get_flags(target, ioff, isize, flags);
 637         if (bp) {
 638                 if (!XFS_BUF_ISDONE(bp)) {
 639                         XB_TRACE(bp, "read", (unsigned long)flags);
 640                         XFS_STATS_INC(xb_get_read);
 641                         xfs_buf_iostart(bp, flags);
 642                 } else if (flags & XBF_ASYNC) {
 643                         XB_TRACE(bp, "read_async", (unsigned long)flags);
 644                         /*
 645                          * Read ahead call which is already satisfied,
 646                          * drop the buffer
 647                          */
 648                         goto no_buffer;
 649                 } else {
 650                         XB_TRACE(bp, "read_done", (unsigned long)flags);
 651                         /* We do not want read in the flags */
 652                         bp->b_flags &= ~XBF_READ;
 653                 }
 654         }
 655
 656         return bp;
 657
 658  no_buffer:
 659         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 660                 xfs_buf_unlock(bp);
 661         xfs_buf_rele(bp);
 662         return NULL;
 663 }
 664
 665 /*
 666  *      If we are not low on memory then do the readahead in a deadlock
 667  *      safe manner.
 668  */
 669 void
 670 xfs_buf_readahead(
 671         xfs_buftarg_t           *target,
 672         xfs_off_t               ioff,
 673         size_t                  isize,
 674         xfs_buf_flags_t         flags)
 675 {
 676         struct backing_dev_info *bdi;
 677
 678         bdi = target->bt_mapping->backing_dev_info;
 679         if (bdi_read_congested(bdi))
 680                 return;
 681
 682         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 683         xfs_buf_read_flags(target, ioff, isize, flags);
 684 }
 685
 686 xfs_buf_t *
 687 xfs_buf_get_empty(
 688         size_t                  len,
 689         xfs_buftarg_t           *target)
 690 {
 691         xfs_buf_t               *bp;
 692
 693         bp = xfs_buf_allocate(0);
 694         if (bp)
 695                 _xfs_buf_initialize(bp, target, 0, len, 0);
 696         return bp;
 697 }
 698
 699 static inline struct page *
 700 mem_to_page(
 701         void                    *addr)
 702 {
 703         if (((unsigned long)addr < VMALLOC_START) ||
 704             ((unsigned long)addr >= VMALLOC_END)) {
 705                 return virt_to_page(addr);
 706         } else {
 707                 return vmalloc_to_page(addr);
 708         }
 709 }
 710
 711 int
 712 xfs_buf_associate_memory(
 713         xfs_buf_t               *bp,
 714         void                    *mem,
 715         size_t                  len)
 716 {
 717         int                     rval;
 718         int                     i = 0;
 719         size_t                  ptr;
 720         size_t                  end, end_cur;
 721         off_t                   offset;
 722         int                     page_count;
 723
 724         page_count = PAGE_CACHE_ALIGN(len) >> PAGE_CACHE_SHIFT;
 725         offset = (off_t) mem - ((off_t)mem & PAGE_CACHE_MASK);
 726         if (offset && (len > PAGE_CACHE_SIZE))
 727                 page_count++;
 728
 729         /* Free any previous set of page pointers */
 730         if (bp->b_pages)
 731                 _xfs_buf_free_pages(bp);
 732
 733         bp->b_pages = NULL;
 734         bp->b_addr = mem;
 735
 736         rval = _xfs_buf_get_pages(bp, page_count, 0);
 737         if (rval)
 738                 return rval;
 739
 740         bp->b_offset = offset;
 741         ptr = (size_t) mem & PAGE_CACHE_MASK;
 742         end = PAGE_CACHE_ALIGN((size_t) mem + len);
 743         end_cur = end;
 744         /* set up first page */
 745         bp->b_pages[0] = mem_to_page(mem);
 746
 747         ptr += PAGE_CACHE_SIZE;
 748         bp->b_page_count = ++i;
 749         while (ptr < end) {
 750                 bp->b_pages[i] = mem_to_page((void *)ptr);
 751                 bp->b_page_count = ++i;
 752                 ptr += PAGE_CACHE_SIZE;
 753         }
 754         bp->b_locked = 0;
 755
 756         bp->b_count_desired = bp->b_buffer_length = len;
 757         bp->b_flags |= XBF_MAPPED;
 758
 759         return 0;
 760 }
 761
 762 xfs_buf_t *
 763 xfs_buf_get_noaddr(
 764         size_t                  len,
 765         xfs_buftarg_t           *target)
 766 {
 767         size_t                  malloc_len = len;
 768         xfs_buf_t               *bp;
 769         void                    *data;
 770         int                     error;
 771
 772         bp = xfs_buf_allocate(0);
 773         if (unlikely(bp == NULL))
 774                 goto fail;
 775         _xfs_buf_initialize(bp, target, 0, len, 0);
 776
 777  try_again:
 778         data = kmem_alloc(malloc_len, KM_SLEEP | KM_MAYFAIL | KM_LARGE);
 779         if (unlikely(data == NULL))
 780                 goto fail_free_buf;
 781
 782         /* check whether alignment matches.. */
 783         if ((__psunsigned_t)data !=
 784             ((__psunsigned_t)data & ~target->bt_smask)) {
 785                 /* .. else double the size and try again */
 786                 kmem_free(data, malloc_len);
 787                 malloc_len <<= 1;
 788                 goto try_again;
 789         }
 790
 791         error = xfs_buf_associate_memory(bp, data, len);
 792         if (error)
 793                 goto fail_free_mem;
 794         bp->b_flags |= _XBF_KMEM_ALLOC;
 795
 796         xfs_buf_unlock(bp);
 797
 798         XB_TRACE(bp, "no_daddr", data);
 799         return bp;
 800  fail_free_mem:
 801         kmem_free(data, malloc_len);
 802  fail_free_buf:
 803         xfs_buf_free(bp);
 804  fail:
 805         return NULL;
 806 }
 807
 808 /*
 809  *      Increment reference count on buffer, to hold the buffer concurrently
 810  *      with another thread which may release (free) the buffer asynchronously.
 811  *      Must hold the buffer already to call this function.
 812  */
 813 void
 814 xfs_buf_hold(
 815         xfs_buf_t               *bp)
 816 {
 817         atomic_inc(&bp->b_hold);
 818         XB_TRACE(bp, "hold", 0);
 819 }
 820
 821 /*
 822  *      Releases a hold on the specified buffer.  If the
 823  *      the hold count is 1, calls xfs_buf_free.
 824  */
 825 void
 826 xfs_buf_rele(
 827         xfs_buf_t               *bp)
 828 {
 829         xfs_bufhash_t           *hash = bp->b_hash;
 830
 831         XB_TRACE(bp, "rele", bp->b_relse);
 832
 833         if (unlikely(!hash)) {
 834                 ASSERT(!bp->b_relse);
 835                 if (atomic_dec_and_test(&bp->b_hold))
 836                         xfs_buf_free(bp);
 837                 return;
 838         }
 839
 840         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
 841                 if (bp->b_relse) {
 842                         atomic_inc(&bp->b_hold);
 843                         spin_unlock(&hash->bh_lock);
 844                         (*(bp->b_relse)) (bp);
 845                 } else if (bp->b_flags & XBF_FS_MANAGED) {
 846                         spin_unlock(&hash->bh_lock);
 847                 } else {
 848                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 849                         list_del_init(&bp->b_hash_list);
 850                         spin_unlock(&hash->bh_lock);
 851                         xfs_buf_free(bp);
 852                 }
 853         } else {
 854                 /*
 855                  * Catch reference count leaks
 856                  */
 857                 ASSERT(atomic_read(&bp->b_hold) >= 0);
 858         }
 859 }
 860
 861
 862 /*
 863  *      Mutual exclusion on buffers.  Locking model:
 864  *
 865  *      Buffers associated with inodes for which buffer locking
 866  *      is not enabled are not protected by semaphores, and are
 867  *      assumed to be exclusively owned by the caller.  There is a
 868  *      spinlock in the buffer, used by the caller when concurrent
 869  *      access is possible.
 870  */
 871
 872 /*
 873  *      Locks a buffer object, if it is not already locked.
 874  *      Note that this in no way locks the underlying pages, so it is only
 875  *      useful for synchronizing concurrent use of buffer objects, not for
 876  *      synchronizing independent access to the underlying pages.
 877  */
 878 int
 879 xfs_buf_cond_lock(
 880         xfs_buf_t               *bp)
 881 {
 882         int                     locked;
 883
 884         locked = down_trylock(&bp->b_sema) == 0;
 885         if (locked) {
 886                 XB_SET_OWNER(bp);
 887         }
 888         XB_TRACE(bp, "cond_lock", (long)locked);
 889         return locked ? 0 : -EBUSY;
 890 }
 891
 892 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
 893 int
 894 xfs_buf_lock_value(
 895         xfs_buf_t               *bp)
 896 {
 897         return atomic_read(&bp->b_sema.count);
 898 }
 899 #endif
 900
 901 /*
 902  *      Locks a buffer object.
 903  *      Note that this in no way locks the underlying pages, so it is only
 904  *      useful for synchronizing concurrent use of buffer objects, not for
 905  *      synchronizing independent access to the underlying pages.
 906  */
 907 void
 908 xfs_buf_lock(
 909         xfs_buf_t               *bp)
 910 {
 911         XB_TRACE(bp, "lock", 0);
 912         if (atomic_read(&bp->b_io_remaining))
 913                 blk_run_address_space(bp->b_target->bt_mapping);
 914         down(&bp->b_sema);
 915         XB_SET_OWNER(bp);
 916         XB_TRACE(bp, "locked", 0);
 917 }
 918
 919 /*
 920  *      Releases the lock on the buffer object.
 921  *      If the buffer is marked delwri but is not queued, do so before we
 922  *      unlock the buffer as we need to set flags correctly.  We also need to
 923  *      take a reference for the delwri queue because the unlocker is going to
 924  *      drop their's and they don't know we just queued it.
 925  */
 926 void
 927 xfs_buf_unlock(
 928         xfs_buf_t               *bp)
 929 {
 930         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 931                 atomic_inc(&bp->b_hold);
 932                 bp->b_flags |= XBF_ASYNC;
 933                 xfs_buf_delwri_queue(bp, 0);
 934         }
 935
 936         XB_CLEAR_OWNER(bp);
 937         up(&bp->b_sema);
 938         XB_TRACE(bp, "unlock", 0);
 939 }
 940
 941
 942 /*
 943  *      Pinning Buffer Storage in Memory
 944  *      Ensure that no attempt to force a buffer to disk will succeed.
 945  */
 946 void
 947 xfs_buf_pin(
 948         xfs_buf_t               *bp)
 949 {
 950         atomic_inc(&bp->b_pin_count);
 951         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
 952 }
 953
 954 void
 955 xfs_buf_unpin(
 956         xfs_buf_t               *bp)
 957 {
 958         if (atomic_dec_and_test(&bp->b_pin_count))
 959                 wake_up_all(&bp->b_waiters);
 960         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
 961 }
 962
 963 int
 964 xfs_buf_ispin(
 965         xfs_buf_t               *bp)
 966 {
 967         return atomic_read(&bp->b_pin_count);
 968 }
 969
 970 STATIC void
 971 xfs_buf_wait_unpin(
 972         xfs_buf_t               *bp)
 973 {
 974         DECLARE_WAITQUEUE       (wait, current);
 975
 976         if (atomic_read(&bp->b_pin_count) == 0)
 977                 return;
 978
 979         add_wait_queue(&bp->b_waiters, &wait);
 980         for (;;) {
 981                 set_current_state(TASK_UNINTERRUPTIBLE);
 982                 if (atomic_read(&bp->b_pin_count) == 0)
 983                         break;
 984                 if (atomic_read(&bp->b_io_remaining))
 985                         blk_run_address_space(bp->b_target->bt_mapping);
 986                 schedule();
 987         }
 988         remove_wait_queue(&bp->b_waiters, &wait);
 989         set_current_state(TASK_RUNNING);
 990 }
 991
 992 /*
 993  *      Buffer Utility Routines
 994  */
 995
 996 STATIC void
 997 xfs_buf_iodone_work(
 998         struct work_struct      *work)
 999 {
1000         xfs_buf_t               *bp =
1001                 container_of(work, xfs_buf_t, b_iodone_work);
1002
1003         if (bp->b_iodone)
1004                 (*(bp->b_iodone))(bp);
1005         else if (bp->b_flags & XBF_ASYNC)
1006                 xfs_buf_relse(bp);
1007 }
1008
1009 void
1010 xfs_buf_ioend(
1011         xfs_buf_t               *bp,
1012         int                     schedule)
1013 {
1014         bp->b_flags &= ~(XBF_READ | XBF_WRITE);
1015         if (bp->b_error == 0)
1016                 bp->b_flags |= XBF_DONE;
1017
1018         XB_TRACE(bp, "iodone", bp->b_iodone);
1019
1020         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1021                 if (schedule) {
1022                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1023                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1024                 } else {
1025                         xfs_buf_iodone_work(&bp->b_iodone_work);
1026                 }
1027         } else {
1028                 up(&bp->b_iodonesema);
1029         }
1030 }
1031
1032 void
1033 xfs_buf_ioerror(
1034         xfs_buf_t               *bp,
1035         int                     error)
1036 {
1037         ASSERT(error >= 0 && error <= 0xffff);
1038         bp->b_error = (unsigned short)error;
1039         XB_TRACE(bp, "ioerror", (unsigned long)error);
1040 }
1041
1042 /*
1043  *      Initiate I/O on a buffer, based on the flags supplied.
1044  *      The b_iodone routine in the buffer supplied will only be called
1045  *      when all of the subsidiary I/O requests, if any, have been completed.
1046  */
1047 int
1048 xfs_buf_iostart(
1049         xfs_buf_t               *bp,
1050         xfs_buf_flags_t         flags)
1051 {
1052         int                     status = 0;
1053
1054         XB_TRACE(bp, "iostart", (unsigned long)flags);
1055
1056         if (flags & XBF_DELWRI) {
1057                 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1058                 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1059                 xfs_buf_delwri_queue(bp, 1);
1060                 return status;
1061         }
1062
1063         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1064                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1065         bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1066                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1067
1068         BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1069
1070         /* For writes allow an alternate strategy routine to precede
1071          * the actual I/O request (which may not be issued at all in
1072          * a shutdown situation, for example).
1073          */
1074         status = (flags & XBF_WRITE) ?
1075                 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1076
1077         /* Wait for I/O if we are not an async request.
1078          * Note: async I/O request completion will release the buffer,
1079          * and that can already be done by this point.  So using the
1080          * buffer pointer from here on, after async I/O, is invalid.
1081          */
1082         if (!status && !(flags & XBF_ASYNC))
1083                 status = xfs_buf_iowait(bp);
1084
1085         return status;
1086 }
1087
1088 STATIC __inline__ int
1089 _xfs_buf_iolocked(
1090         xfs_buf_t               *bp)
1091 {
1092         ASSERT(bp->b_flags & (XBF_READ | XBF_WRITE));
1093         if (bp->b_flags & XBF_READ)
1094                 return bp->b_locked;
1095         return 0;
1096 }
1097
1098 STATIC __inline__ void
1099 _xfs_buf_ioend(
1100         xfs_buf_t               *bp,
1101         int                     schedule)
1102 {
1103         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1104                 bp->b_locked = 0;
1105                 xfs_buf_ioend(bp, schedule);
1106         }
1107 }
1108
1109 STATIC int
1110 xfs_buf_bio_end_io(
1111         struct bio              *bio,
1112         unsigned int            bytes_done,
1113         int                     error)
1114 {
1115         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1116         unsigned int            blocksize = bp->b_target->bt_bsize;
1117         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1118
1119         if (bio->bi_size)
1120                 return 1;
1121
1122         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1123                 bp->b_error = EIO;
1124
1125         do {
1126                 struct page     *page = bvec->bv_page;
1127
1128                 ASSERT(!PagePrivate(page));
1129                 if (unlikely(bp->b_error)) {
1130                         if (bp->b_flags & XBF_READ)
1131                                 ClearPageUptodate(page);
1132                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1133                         SetPageUptodate(page);
1134                 } else if (!PagePrivate(page) &&
1135                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1136                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1137                 }
1138
1139                 if (--bvec >= bio->bi_io_vec)
1140                         prefetchw(&bvec->bv_page->flags);
1141
1142                 if (_xfs_buf_iolocked(bp)) {
1143                         unlock_page(page);
1144                 }
1145         } while (bvec >= bio->bi_io_vec);
1146
1147         _xfs_buf_ioend(bp, 1);
1148         bio_put(bio);
1149         return 0;
1150 }
1151
1152 STATIC void
1153 _xfs_buf_ioapply(
1154         xfs_buf_t               *bp)
1155 {
1156         int                     i, rw, map_i, total_nr_pages, nr_pages;
1157         struct bio              *bio;
1158         int                     offset = bp->b_offset;
1159         int                     size = bp->b_count_desired;
1160         sector_t                sector = bp->b_bn;
1161         unsigned int            blocksize = bp->b_target->bt_bsize;
1162         int                     locking = _xfs_buf_iolocked(bp);
1163
1164         total_nr_pages = bp->b_page_count;
1165         map_i = 0;
1166
1167         if (bp->b_flags & XBF_ORDERED) {
1168                 ASSERT(!(bp->b_flags & XBF_READ));
1169                 rw = WRITE_BARRIER;
1170         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1171                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1172                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1173                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1174         } else {
1175                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1176                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1177         }
1178
1179         /* Special code path for reading a sub page size buffer in --
1180          * we populate up the whole page, and hence the other metadata
1181          * in the same page.  This optimization is only valid when the
1182          * filesystem block size is not smaller than the page size.
1183          */
1184         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1185             (bp->b_flags & XBF_READ) && locking &&
1186             (blocksize >= PAGE_CACHE_SIZE)) {
1187                 bio = bio_alloc(GFP_NOIO, 1);
1188
1189                 bio->bi_bdev = bp->b_target->bt_bdev;
1190                 bio->bi_sector = sector - (offset >> BBSHIFT);
1191                 bio->bi_end_io = xfs_buf_bio_end_io;
1192                 bio->bi_private = bp;
1193
1194                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1195                 size = 0;
1196
1197                 atomic_inc(&bp->b_io_remaining);
1198
1199                 goto submit_io;
1200         }
1201
1202         /* Lock down the pages which we need to for the request */
1203         if (locking && (bp->b_flags & XBF_WRITE) && (bp->b_locked == 0)) {
1204                 for (i = 0; size; i++) {
1205                         int             nbytes = PAGE_CACHE_SIZE - offset;
1206                         struct page     *page = bp->b_pages[i];
1207
1208                         if (nbytes > size)
1209                                 nbytes = size;
1210
1211                         lock_page(page);
1212
1213                         size -= nbytes;
1214                         offset = 0;
1215                 }
1216                 offset = bp->b_offset;
1217                 size = bp->b_count_desired;
1218         }
1219
1220 next_chunk:
1221         atomic_inc(&bp->b_io_remaining);
1222         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1223         if (nr_pages > total_nr_pages)
1224                 nr_pages = total_nr_pages;
1225
1226         bio = bio_alloc(GFP_NOIO, nr_pages);
1227         bio->bi_bdev = bp->b_target->bt_bdev;
1228         bio->bi_sector = sector;
1229         bio->bi_end_io = xfs_buf_bio_end_io;
1230         bio->bi_private = bp;
1231
1232         for (; size && nr_pages; nr_pages--, map_i++) {
1233                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1234
1235                 if (nbytes > size)
1236                         nbytes = size;
1237
1238                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1239                 if (rbytes < nbytes)
1240                         break;
1241
1242                 offset = 0;
1243                 sector += nbytes >> BBSHIFT;
1244                 size -= nbytes;
1245                 total_nr_pages--;
1246         }
1247
1248 submit_io:
1249         if (likely(bio->bi_size)) {
1250                 submit_bio(rw, bio);
1251                 if (size)
1252                         goto next_chunk;
1253         } else {
1254                 bio_put(bio);
1255                 xfs_buf_ioerror(bp, EIO);
1256         }
1257 }
1258
1259 int
1260 xfs_buf_iorequest(
1261         xfs_buf_t               *bp)
1262 {
1263         XB_TRACE(bp, "iorequest", 0);
1264
1265         if (bp->b_flags & XBF_DELWRI) {
1266                 xfs_buf_delwri_queue(bp, 1);
1267                 return 0;
1268         }
1269
1270         if (bp->b_flags & XBF_WRITE) {
1271                 xfs_buf_wait_unpin(bp);
1272         }
1273
1274         xfs_buf_hold(bp);
1275
1276         /* Set the count to 1 initially, this will stop an I/O
1277          * completion callout which happens before we have started
1278          * all the I/O from calling xfs_buf_ioend too early.
1279          */
1280         atomic_set(&bp->b_io_remaining, 1);
1281         _xfs_buf_ioapply(bp);
1282         _xfs_buf_ioend(bp, 0);
1283
1284         xfs_buf_rele(bp);
1285         return 0;
1286 }
1287
1288 /*
1289  *      Waits for I/O to complete on the buffer supplied.
1290  *      It returns immediately if no I/O is pending.
1291  *      It returns the I/O error code, if any, or 0 if there was no error.
1292  */
1293 int
1294 xfs_buf_iowait(
1295         xfs_buf_t               *bp)
1296 {
1297         XB_TRACE(bp, "iowait", 0);
1298         if (atomic_read(&bp->b_io_remaining))
1299                 blk_run_address_space(bp->b_target->bt_mapping);
1300         down(&bp->b_iodonesema);
1301         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1302         return bp->b_error;
1303 }
1304
1305 xfs_caddr_t
1306 xfs_buf_offset(
1307         xfs_buf_t               *bp,
1308         size_t                  offset)
1309 {
1310         struct page             *page;
1311
1312         if (bp->b_flags & XBF_MAPPED)
1313                 return XFS_BUF_PTR(bp) + offset;
1314
1315         offset += bp->b_offset;
1316         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1317         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1318 }
1319
1320 /*
1321  *      Move data into or out of a buffer.
1322  */
1323 void
1324 xfs_buf_iomove(
1325         xfs_buf_t               *bp,    /* buffer to process            */
1326         size_t                  boff,   /* starting buffer offset       */
1327         size_t                  bsize,  /* length to copy               */
1328         caddr_t                 data,   /* data address                 */
1329         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1330 {
1331         size_t                  bend, cpoff, csize;
1332         struct page             *page;
1333
1334         bend = boff + bsize;
1335         while (boff < bend) {
1336                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1337                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1338                 csize = min_t(size_t,
1339                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1340
1341                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1342
1343                 switch (mode) {
1344                 case XBRW_ZERO:
1345                         memset(page_address(page) + cpoff, 0, csize);
1346                         break;
1347                 case XBRW_READ:
1348                         memcpy(data, page_address(page) + cpoff, csize);
1349                         break;
1350                 case XBRW_WRITE:
1351                         memcpy(page_address(page) + cpoff, data, csize);
1352                 }
1353
1354                 boff += csize;
1355                 data += csize;
1356         }
1357 }
1358
1359 /*
1360  *      Handling of buffer targets (buftargs).
1361  */
1362
1363 /*
1364  *      Wait for any bufs with callbacks that have been submitted but
1365  *      have not yet returned... walk the hash list for the target.
1366  */
1367 void
1368 xfs_wait_buftarg(
1369         xfs_buftarg_t   *btp)
1370 {
1371         xfs_buf_t       *bp, *n;
1372         xfs_bufhash_t   *hash;
1373         uint            i;
1374
1375         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1376                 hash = &btp->bt_hash[i];
1377 again:
1378                 spin_lock(&hash->bh_lock);
1379                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1380                         ASSERT(btp == bp->b_target);
1381                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1382                                 spin_unlock(&hash->bh_lock);
1383                                 /*
1384                                  * Catch superblock reference count leaks
1385                                  * immediately
1386                                  */
1387                                 BUG_ON(bp->b_bn == 0);
1388                                 delay(100);
1389                                 goto again;
1390                         }
1391                 }
1392                 spin_unlock(&hash->bh_lock);
1393         }
1394 }
1395
1396 /*
1397  *      Allocate buffer hash table for a given target.
1398  *      For devices containing metadata (i.e. not the log/realtime devices)
1399  *      we need to allocate a much larger hash table.
1400  */
1401 STATIC void
1402 xfs_alloc_bufhash(
1403         xfs_buftarg_t           *btp,
1404         int                     external)
1405 {
1406         unsigned int            i;
1407
1408         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1409         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1410         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1411                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1412         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1413                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1414                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1415         }
1416 }
1417
1418 STATIC void
1419 xfs_free_bufhash(
1420         xfs_buftarg_t           *btp)
1421 {
1422         kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1423         btp->bt_hash = NULL;
1424 }
1425
1426 /*
1427  *      buftarg list for delwrite queue processing
1428  */
1429 STATIC LIST_HEAD(xfs_buftarg_list);
1430 STATIC DEFINE_SPINLOCK(xfs_buftarg_lock);
1431
1432 STATIC void
1433 xfs_register_buftarg(
1434         xfs_buftarg_t           *btp)
1435 {
1436         spin_lock(&xfs_buftarg_lock);
1437         list_add(&btp->bt_list, &xfs_buftarg_list);
1438         spin_unlock(&xfs_buftarg_lock);
1439 }
1440
1441 STATIC void
1442 xfs_unregister_buftarg(
1443         xfs_buftarg_t           *btp)
1444 {
1445         spin_lock(&xfs_buftarg_lock);
1446         list_del(&btp->bt_list);
1447         spin_unlock(&xfs_buftarg_lock);
1448 }
1449
1450 void
1451 xfs_free_buftarg(
1452         xfs_buftarg_t           *btp,
1453         int                     external)
1454 {
1455         xfs_flush_buftarg(btp, 1);
1456         if (external)
1457                 xfs_blkdev_put(btp->bt_bdev);
1458         xfs_free_bufhash(btp);
1459         iput(btp->bt_mapping->host);
1460
1461         /* Unregister the buftarg first so that we don't get a
1462          * wakeup finding a non-existent task
1463          */
1464         xfs_unregister_buftarg(btp);
1465         kthread_stop(btp->bt_task);
1466
1467         kmem_free(btp, sizeof(*btp));
1468 }
1469
1470 STATIC int
1471 xfs_setsize_buftarg_flags(
1472         xfs_buftarg_t           *btp,
1473         unsigned int            blocksize,
1474         unsigned int            sectorsize,
1475         int                     verbose)
1476 {
1477         btp->bt_bsize = blocksize;
1478         btp->bt_sshift = ffs(sectorsize) - 1;
1479         btp->bt_smask = sectorsize - 1;
1480
1481         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1482                 printk(KERN_WARNING
1483                         "XFS: Cannot set_blocksize to %u on device %s\n",
1484                         sectorsize, XFS_BUFTARG_NAME(btp));
1485                 return EINVAL;
1486         }
1487
1488         if (verbose &&
1489             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1490                 printk(KERN_WARNING
1491                         "XFS: %u byte sectors in use on device %s.  "
1492                         "This is suboptimal; %u or greater is ideal.\n",
1493                         sectorsize, XFS_BUFTARG_NAME(btp),
1494                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1495         }
1496
1497         return 0;
1498 }
1499
1500 /*
1501  *      When allocating the initial buffer target we have not yet
1502  *      read in the superblock, so don't know what sized sectors
1503  *      are being used is at this early stage.  Play safe.
1504  */
1505 STATIC int
1506 xfs_setsize_buftarg_early(
1507         xfs_buftarg_t           *btp,
1508         struct block_device     *bdev)
1509 {
1510         return xfs_setsize_buftarg_flags(btp,
1511                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1512 }
1513
1514 int
1515 xfs_setsize_buftarg(
1516         xfs_buftarg_t           *btp,
1517         unsigned int            blocksize,
1518         unsigned int            sectorsize)
1519 {
1520         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1521 }
1522
1523 STATIC int
1524 xfs_mapping_buftarg(
1525         xfs_buftarg_t           *btp,
1526         struct block_device     *bdev)
1527 {
1528         struct backing_dev_info *bdi;
1529         struct inode            *inode;
1530         struct address_space    *mapping;
1531         static const struct address_space_operations mapping_aops = {
1532                 .sync_page = block_sync_page,
1533                 .migratepage = fail_migrate_page,
1534         };
1535
1536         inode = new_inode(bdev->bd_inode->i_sb);
1537         if (!inode) {
1538                 printk(KERN_WARNING
1539                         "XFS: Cannot allocate mapping inode for device %s\n",
1540                         XFS_BUFTARG_NAME(btp));
1541                 return ENOMEM;
1542         }
1543         inode->i_mode = S_IFBLK;
1544         inode->i_bdev = bdev;
1545         inode->i_rdev = bdev->bd_dev;
1546         bdi = blk_get_backing_dev_info(bdev);
1547         if (!bdi)
1548                 bdi = &default_backing_dev_info;
1549         mapping = &inode->i_data;
1550         mapping->a_ops = &mapping_aops;
1551         mapping->backing_dev_info = bdi;
1552         mapping_set_gfp_mask(mapping, GFP_NOFS);
1553         btp->bt_mapping = mapping;
1554         return 0;
1555 }
1556
1557 STATIC int
1558 xfs_alloc_delwrite_queue(
1559         xfs_buftarg_t           *btp)
1560 {
1561         int     error = 0;
1562
1563         INIT_LIST_HEAD(&btp->bt_list);
1564         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1565         spinlock_init(&btp->bt_delwrite_lock, "delwri_lock");
1566         btp->bt_flags = 0;
1567         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1568         if (IS_ERR(btp->bt_task)) {
1569                 error = PTR_ERR(btp->bt_task);
1570                 goto out_error;
1571         }
1572         xfs_register_buftarg(btp);
1573 out_error:
1574         return error;
1575 }
1576
1577 xfs_buftarg_t *
1578 xfs_alloc_buftarg(
1579         struct block_device     *bdev,
1580         int                     external)
1581 {
1582         xfs_buftarg_t           *btp;
1583
1584         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1585
1586         btp->bt_dev =  bdev->bd_dev;
1587         btp->bt_bdev = bdev;
1588         if (xfs_setsize_buftarg_early(btp, bdev))
1589                 goto error;
1590         if (xfs_mapping_buftarg(btp, bdev))
1591                 goto error;
1592         if (xfs_alloc_delwrite_queue(btp))
1593                 goto error;
1594         xfs_alloc_bufhash(btp, external);
1595         return btp;
1596
1597 error:
1598         kmem_free(btp, sizeof(*btp));
1599         return NULL;
1600 }
1601
1602
1603 /*
1604  *      Delayed write buffer handling
1605  */
1606 STATIC void
1607 xfs_buf_delwri_queue(
1608         xfs_buf_t               *bp,
1609         int                     unlock)
1610 {
1611         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1612         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1613
1614         XB_TRACE(bp, "delwri_q", (long)unlock);
1615         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1616
1617         spin_lock(dwlk);
1618         /* If already in the queue, dequeue and place at tail */
1619         if (!list_empty(&bp->b_list)) {
1620                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1621                 if (unlock)
1622                         atomic_dec(&bp->b_hold);
1623                 list_del(&bp->b_list);
1624         }
1625
1626         bp->b_flags |= _XBF_DELWRI_Q;
1627         list_add_tail(&bp->b_list, dwq);
1628         bp->b_queuetime = jiffies;
1629         spin_unlock(dwlk);
1630
1631         if (unlock)
1632                 xfs_buf_unlock(bp);
1633 }
1634
1635 void
1636 xfs_buf_delwri_dequeue(
1637         xfs_buf_t               *bp)
1638 {
1639         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1640         int                     dequeued = 0;
1641
1642         spin_lock(dwlk);
1643         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1644                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1645                 list_del_init(&bp->b_list);
1646                 dequeued = 1;
1647         }
1648         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1649         spin_unlock(dwlk);
1650
1651         if (dequeued)
1652                 xfs_buf_rele(bp);
1653
1654         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1655 }
1656
1657 STATIC void
1658 xfs_buf_runall_queues(
1659         struct workqueue_struct *queue)
1660 {
1661         flush_workqueue(queue);
1662 }
1663
1664 STATIC int
1665 xfsbufd_wakeup(
1666         int                     priority,
1667         gfp_t                   mask)
1668 {
1669         xfs_buftarg_t           *btp;
1670
1671         spin_lock(&xfs_buftarg_lock);
1672         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1673                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1674                         continue;
1675                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1676                 wake_up_process(btp->bt_task);
1677         }
1678         spin_unlock(&xfs_buftarg_lock);
1679         return 0;
1680 }
1681
1682 STATIC int
1683 xfsbufd(
1684         void                    *data)
1685 {
1686         struct list_head        tmp;
1687         unsigned long           age;
1688         xfs_buftarg_t           *target = (xfs_buftarg_t *)data;
1689         xfs_buf_t               *bp, *n;
1690         struct list_head        *dwq = &target->bt_delwrite_queue;
1691         spinlock_t              *dwlk = &target->bt_delwrite_lock;
1692         int                     count;
1693
1694         current->flags |= PF_MEMALLOC;
1695
1696         INIT_LIST_HEAD(&tmp);
1697         do {
1698                 if (unlikely(freezing(current))) {
1699                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1700                         refrigerator();
1701                 } else {
1702                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1703                 }
1704
1705                 schedule_timeout_interruptible(
1706                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1707
1708                 count = 0;
1709                 age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1710                 spin_lock(dwlk);
1711                 list_for_each_entry_safe(bp, n, dwq, b_list) {
1712                         XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1713                         ASSERT(bp->b_flags & XBF_DELWRI);
1714
1715                         if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1716                                 if (!test_bit(XBT_FORCE_FLUSH,
1717                                                 &target->bt_flags) &&
1718                                     time_before(jiffies,
1719                                                 bp->b_queuetime + age)) {
1720                                         xfs_buf_unlock(bp);
1721                                         break;
1722                                 }
1723
1724                                 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1725                                                  _XBF_RUN_QUEUES);
1726                                 bp->b_flags |= XBF_WRITE;
1727                                 list_move_tail(&bp->b_list, &tmp);
1728                                 count++;
1729                         }
1730                 }
1731                 spin_unlock(dwlk);
1732
1733                 while (!list_empty(&tmp)) {
1734                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1735                         ASSERT(target == bp->b_target);
1736
1737                         list_del_init(&bp->b_list);
1738                         xfs_buf_iostrategy(bp);
1739                 }
1740
1741                 if (as_list_len > 0)
1742                         purge_addresses();
1743                 if (count)
1744                         blk_run_address_space(target->bt_mapping);
1745
1746                 clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1747         } while (!kthread_should_stop());
1748
1749         return 0;
1750 }
1751
1752 /*
1753  *      Go through all incore buffers, and release buffers if they belong to
1754  *      the given device. This is used in filesystem error handling to
1755  *      preserve the consistency of its metadata.
1756  */
1757 int
1758 xfs_flush_buftarg(
1759         xfs_buftarg_t           *target,
1760         int                     wait)
1761 {
1762         struct list_head        tmp;
1763         xfs_buf_t               *bp, *n;
1764         int                     pincount = 0;
1765         struct list_head        *dwq = &target->bt_delwrite_queue;
1766         spinlock_t              *dwlk = &target->bt_delwrite_lock;
1767
1768         xfs_buf_runall_queues(xfsdatad_workqueue);
1769         xfs_buf_runall_queues(xfslogd_workqueue);
1770
1771         INIT_LIST_HEAD(&tmp);
1772         spin_lock(dwlk);
1773         list_for_each_entry_safe(bp, n, dwq, b_list) {
1774                 ASSERT(bp->b_target == target);
1775                 ASSERT(bp->b_flags & (XBF_DELWRI | _XBF_DELWRI_Q));
1776                 XB_TRACE(bp, "walkq2", (long)xfs_buf_ispin(bp));
1777                 if (xfs_buf_ispin(bp)) {
1778                         pincount++;
1779                         continue;
1780                 }
1781
1782                 list_move_tail(&bp->b_list, &tmp);
1783         }
1784         spin_unlock(dwlk);
1785
1786         /*
1787          * Dropped the delayed write list lock, now walk the temporary list
1788          */
1789         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1790                 xfs_buf_lock(bp);
1791                 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|_XBF_RUN_QUEUES);
1792                 bp->b_flags |= XBF_WRITE;
1793                 if (wait)
1794                         bp->b_flags &= ~XBF_ASYNC;
1795                 else
1796                         list_del_init(&bp->b_list);
1797
1798                 xfs_buf_iostrategy(bp);
1799         }
1800
1801         if (wait)
1802                 blk_run_address_space(target->bt_mapping);
1803
1804         /*
1805          * Remaining list items must be flushed before returning
1806          */
1807         while (!list_empty(&tmp)) {
1808                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1809
1810                 list_del_init(&bp->b_list);
1811                 xfs_iowait(bp);
1812                 xfs_buf_relse(bp);
1813         }
1814
1815         return pincount;
1816 }
1817
1818 int __init
1819 xfs_buf_init(void)
1820 {
1821 #ifdef XFS_BUF_TRACE
1822         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1823 #endif
1824
1825         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1826                                                 KM_ZONE_HWALIGN, NULL);
1827         if (!xfs_buf_zone)
1828                 goto out_free_trace_buf;
1829
1830         xfslogd_workqueue = create_freezeable_workqueue("xfslogd");
1831         if (!xfslogd_workqueue)
1832                 goto out_free_buf_zone;
1833
1834         xfsdatad_workqueue = create_freezeable_workqueue("xfsdatad");
1835         if (!xfsdatad_workqueue)
1836                 goto out_destroy_xfslogd_workqueue;
1837
1838         xfs_buf_shake = kmem_shake_register(xfsbufd_wakeup);
1839         if (!xfs_buf_shake)
1840                 goto out_destroy_xfsdatad_workqueue;
1841
1842         return 0;
1843
1844  out_destroy_xfsdatad_workqueue:
1845         destroy_workqueue(xfsdatad_workqueue);
1846  out_destroy_xfslogd_workqueue:
1847         destroy_workqueue(xfslogd_workqueue);
1848  out_free_buf_zone:
1849         kmem_zone_destroy(xfs_buf_zone);
1850  out_free_trace_buf:
1851 #ifdef XFS_BUF_TRACE
1852         ktrace_free(xfs_buf_trace_buf);
1853 #endif
1854         return -ENOMEM;
1855 }
1856
1857 void
1858 xfs_buf_terminate(void)
1859 {
1860         kmem_shake_deregister(xfs_buf_shake);
1861         destroy_workqueue(xfsdatad_workqueue);
1862         destroy_workqueue(xfslogd_workqueue);
1863         kmem_zone_destroy(xfs_buf_zone);
1864 #ifdef XFS_BUF_TRACE
1865         ktrace_free(xfs_buf_trace_buf);
1866 #endif
1867 }