fs/xfs/linux-2.6/xfs_buf.c

   1 /*
   2  * Copyright (c) 2000-2005 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 <linux/stddef.h>
  19 #include <linux/errno.h>
  20 #include <linux/slab.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/init.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/bio.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/percpu.h>
  29 #include <linux/blkdev.h>
  30 #include <linux/hash.h>
  31 #include <linux/kthread.h>
  32 #include <linux/migrate.h>
  33 #include "xfs_linux.h"
  34
  35 STATIC kmem_zone_t *xfs_buf_zone;
  36 STATIC kmem_shaker_t xfs_buf_shake;
  37 STATIC int xfsbufd(void *);
  38 STATIC int xfsbufd_wakeup(int, gfp_t);
  39 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
  40
  41 STATIC struct workqueue_struct *xfslogd_workqueue;
  42 struct workqueue_struct *xfsdatad_workqueue;
  43
  44 #ifdef XFS_BUF_TRACE
  45 void
  46 xfs_buf_trace(
  47         xfs_buf_t       *bp,
  48         char            *id,
  49         void            *data,
  50         void            *ra)
  51 {
  52         ktrace_enter(xfs_buf_trace_buf,
  53                 bp, id,
  54                 (void *)(unsigned long)bp->b_flags,
  55                 (void *)(unsigned long)bp->b_hold.counter,
  56                 (void *)(unsigned long)bp->b_sema.count.counter,
  57                 (void *)current,
  58                 data, ra,
  59                 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
  60                 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
  61                 (void *)(unsigned long)bp->b_buffer_length,
  62                 NULL, NULL, NULL, NULL, NULL);
  63 }
  64 ktrace_t *xfs_buf_trace_buf;
  65 #define XFS_BUF_TRACE_SIZE      4096
  66 #define XB_TRACE(bp, id, data)  \
  67         xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
  68 #else
  69 #define XB_TRACE(bp, id, data)  do { } while (0)
  70 #endif
  71
  72 #ifdef XFS_BUF_LOCK_TRACKING
  73 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  74 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  75 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  76 #else
  77 # define XB_SET_OWNER(bp)       do { } while (0)
  78 # define XB_CLEAR_OWNER(bp)     do { } while (0)
  79 # define XB_GET_OWNER(bp)       do { } while (0)
  80 #endif
  81
  82 #define xb_to_gfp(flags) \
  83         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
  84           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
  85
  86 #define xb_to_km(flags) \
  87          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
  88
  89 #define xfs_buf_allocate(flags) \
  90         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
  91 #define xfs_buf_deallocate(bp) \
  92         kmem_zone_free(xfs_buf_zone, (bp));
  93
  94 /*
  95  *      Page Region interfaces.
  96  *
  97  *      For pages in filesystems where the blocksize is smaller than the
  98  *      pagesize, we use the page->private field (long) to hold a bitmap
  99  *      of uptodate regions within the page.
 100  *
 101  *      Each such region is "bytes per page / bits per long" bytes long.
 102  *
 103  *      NBPPR == number-of-bytes-per-page-region
 104  *      BTOPR == bytes-to-page-region (rounded up)
 105  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
 106  */
 107 #if (BITS_PER_LONG == 32)
 108 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
 109 #elif (BITS_PER_LONG == 64)
 110 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
 111 #else
 112 #error BITS_PER_LONG must be 32 or 64
 113 #endif
 114 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
 115 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
 116 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
 117
 118 STATIC unsigned long
 119 page_region_mask(
 120         size_t          offset,
 121         size_t          length)
 122 {
 123         unsigned long   mask;
 124         int             first, final;
 125
 126         first = BTOPR(offset);
 127         final = BTOPRT(offset + length - 1);
 128         first = min(first, final);
 129
 130         mask = ~0UL;
 131         mask <<= BITS_PER_LONG - (final - first);
 132         mask >>= BITS_PER_LONG - (final);
 133
 134         ASSERT(offset + length <= PAGE_CACHE_SIZE);
 135         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
 136
 137         return mask;
 138 }
 139
 140 STATIC inline void
 141 set_page_region(
 142         struct page     *page,
 143         size_t          offset,
 144         size_t          length)
 145 {
 146         set_page_private(page,
 147                 page_private(page) | page_region_mask(offset, length));
 148         if (page_private(page) == ~0UL)
 149                 SetPageUptodate(page);
 150 }
 151
 152 STATIC inline int
 153 test_page_region(
 154         struct page     *page,
 155         size_t          offset,
 156         size_t          length)
 157 {
 158         unsigned long   mask = page_region_mask(offset, length);
 159
 160         return (mask && (page_private(page) & mask) == mask);
 161 }
 162
 163 /*
 164  *      Mapping of multi-page buffers into contiguous virtual space
 165  */
 166
 167 typedef struct a_list {
 168         void            *vm_addr;
 169         struct a_list   *next;
 170 } a_list_t;
 171
 172 STATIC a_list_t         *as_free_head;
 173 STATIC int              as_list_len;
 174 STATIC DEFINE_SPINLOCK(as_lock);
 175
 176 /*
 177  *      Try to batch vunmaps because they are costly.
 178  */
 179 STATIC void
 180 free_address(
 181         void            *addr)
 182 {
 183         a_list_t        *aentry;
 184
 185         aentry = kmalloc(sizeof(a_list_t), GFP_ATOMIC & ~__GFP_HIGH);
 186         if (likely(aentry)) {
 187                 spin_lock(&as_lock);
 188                 aentry->next = as_free_head;
 189                 aentry->vm_addr = addr;
 190                 as_free_head = aentry;
 191                 as_list_len++;
 192                 spin_unlock(&as_lock);
 193         } else {
 194                 vunmap(addr);
 195         }
 196 }
 197
 198 STATIC void
 199 purge_addresses(void)
 200 {
 201         a_list_t        *aentry, *old;
 202
 203         if (as_free_head == NULL)
 204                 return;
 205
 206         spin_lock(&as_lock);
 207         aentry = as_free_head;
 208         as_free_head = NULL;
 209         as_list_len = 0;
 210         spin_unlock(&as_lock);
 211
 212         while ((old = aentry) != NULL) {
 213                 vunmap(aentry->vm_addr);
 214                 aentry = aentry->next;
 215                 kfree(old);
 216         }
 217 }
 218
 219 /*
 220  *      Internal xfs_buf_t object manipulation
 221  */
 222
 223 STATIC void
 224 _xfs_buf_initialize(
 225         xfs_buf_t               *bp,
 226         xfs_buftarg_t           *target,
 227         xfs_off_t               range_base,
 228         size_t                  range_length,
 229         xfs_buf_flags_t         flags)
 230 {
 231         /*
 232          * We don't want certain flags to appear in b_flags.
 233          */
 234         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
 235
 236         memset(bp, 0, sizeof(xfs_buf_t));
 237         atomic_set(&bp->b_hold, 1);
 238         init_MUTEX_LOCKED(&bp->b_iodonesema);
 239         INIT_LIST_HEAD(&bp->b_list);
 240         INIT_LIST_HEAD(&bp->b_hash_list);
 241         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
 242         XB_SET_OWNER(bp);
 243         bp->b_target = target;
 244         bp->b_file_offset = range_base;
 245         /*
 246          * Set buffer_length and count_desired to the same value initially.
 247          * I/O routines should use count_desired, which will be the same in
 248          * most cases but may be reset (e.g. XFS recovery).
 249          */
 250         bp->b_buffer_length = bp->b_count_desired = range_length;
 251         bp->b_flags = flags;
 252         bp->b_bn = XFS_BUF_DADDR_NULL;
 253         atomic_set(&bp->b_pin_count, 0);
 254         init_waitqueue_head(&bp->b_waiters);
 255
 256         XFS_STATS_INC(xb_create);
 257         XB_TRACE(bp, "initialize", target);
 258 }
 259
 260 /*
 261  *      Allocate a page array capable of holding a specified number
 262  *      of pages, and point the page buf at it.
 263  */
 264 STATIC int
 265 _xfs_buf_get_pages(
 266         xfs_buf_t               *bp,
 267         int                     page_count,
 268         xfs_buf_flags_t         flags)
 269 {
 270         /* Make sure that we have a page list */
 271         if (bp->b_pages == NULL) {
 272                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
 273                 bp->b_page_count = page_count;
 274                 if (page_count <= XB_PAGES) {
 275                         bp->b_pages = bp->b_page_array;
 276                 } else {
 277                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
 278                                         page_count, xb_to_km(flags));
 279                         if (bp->b_pages == NULL)
 280                                 return -ENOMEM;
 281                 }
 282                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 283         }
 284         return 0;
 285 }
 286
 287 /*
 288  *      Frees b_pages if it was allocated.
 289  */
 290 STATIC void
 291 _xfs_buf_free_pages(
 292         xfs_buf_t       *bp)
 293 {
 294         if (bp->b_pages != bp->b_page_array) {
 295                 kmem_free(bp->b_pages,
 296                           bp->b_page_count * sizeof(struct page *));
 297         }
 298 }
 299
 300 /*
 301  *      Releases the specified buffer.
 302  *
 303  *      The modification state of any associated pages is left unchanged.
 304  *      The buffer most not be on any hash - use xfs_buf_rele instead for
 305  *      hashed and refcounted buffers
 306  */
 307 void
 308 xfs_buf_free(
 309         xfs_buf_t               *bp)
 310 {
 311         XB_TRACE(bp, "free", 0);
 312
 313         ASSERT(list_empty(&bp->b_hash_list));
 314
 315         if (bp->b_flags & _XBF_PAGE_CACHE) {
 316                 uint            i;
 317
 318                 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
 319                         free_address(bp->b_addr - bp->b_offset);
 320
 321                 for (i = 0; i < bp->b_page_count; i++)
 322                         page_cache_release(bp->b_pages[i]);
 323                 _xfs_buf_free_pages(bp);
 324         } else if (bp->b_flags & _XBF_KMEM_ALLOC) {
 325                  /*
 326                   * XXX(hch): bp->b_count_desired might be incorrect (see
 327                   * xfs_buf_associate_memory for details), but fortunately
 328                   * the Linux version of kmem_free ignores the len argument..
 329                   */
 330                 kmem_free(bp->b_addr, bp->b_count_desired);
 331                 _xfs_buf_free_pages(bp);
 332         }
 333
 334         xfs_buf_deallocate(bp);
 335 }
 336
 337 /*
 338  *      Finds all pages for buffer in question and builds it's page list.
 339  */
 340 STATIC int
 341 _xfs_buf_lookup_pages(
 342         xfs_buf_t               *bp,
 343         uint                    flags)
 344 {
 345         struct address_space    *mapping = bp->b_target->bt_mapping;
 346         size_t                  blocksize = bp->b_target->bt_bsize;
 347         size_t                  size = bp->b_count_desired;
 348         size_t                  nbytes, offset;
 349         gfp_t                   gfp_mask = xb_to_gfp(flags);
 350         unsigned short          page_count, i;
 351         pgoff_t                 first;
 352         xfs_off_t               end;
 353         int                     error;
 354
 355         end = bp->b_file_offset + bp->b_buffer_length;
 356         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 357
 358         error = _xfs_buf_get_pages(bp, page_count, flags);
 359         if (unlikely(error))
 360                 return error;
 361         bp->b_flags |= _XBF_PAGE_CACHE;
 362
 363         offset = bp->b_offset;
 364         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
 365
 366         for (i = 0; i < bp->b_page_count; i++) {
 367                 struct page     *page;
 368                 uint            retries = 0;
 369
 370               retry:
 371                 page = find_or_create_page(mapping, first + i, gfp_mask);
 372                 if (unlikely(page == NULL)) {
 373                         if (flags & XBF_READ_AHEAD) {
 374                                 bp->b_page_count = i;
 375                                 for (i = 0; i < bp->b_page_count; i++)
 376                                         unlock_page(bp->b_pages[i]);
 377                                 return -ENOMEM;
 378                         }
 379
 380                         /*
 381                          * This could deadlock.
 382                          *
 383                          * But until all the XFS lowlevel code is revamped to
 384                          * handle buffer allocation failures we can't do much.
 385                          */
 386                         if (!(++retries % 100))
 387                                 printk(KERN_ERR
 388                                         "XFS: possible memory allocation "
 389                                         "deadlock in %s (mode:0x%x)\n",
 390                                         __FUNCTION__, gfp_mask);
 391
 392                         XFS_STATS_INC(xb_page_retries);
 393                         xfsbufd_wakeup(0, gfp_mask);
 394                         blk_congestion_wait(WRITE, HZ/50);
 395                         goto retry;
 396                 }
 397
 398                 XFS_STATS_INC(xb_page_found);
 399
 400                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
 401                 size -= nbytes;
 402
 403                 if (!PageUptodate(page)) {
 404                         page_count--;
 405                         if (blocksize >= PAGE_CACHE_SIZE) {
 406                                 if (flags & XBF_READ)
 407                                         bp->b_locked = 1;
 408                         } else if (!PagePrivate(page)) {
 409                                 if (test_page_region(page, offset, nbytes))
 410                                         page_count++;
 411                         }
 412                 }
 413
 414                 bp->b_pages[i] = page;
 415                 offset = 0;
 416         }
 417
 418         if (!bp->b_locked) {
 419                 for (i = 0; i < bp->b_page_count; i++)
 420                         unlock_page(bp->b_pages[i]);
 421         }
 422
 423         if (page_count == bp->b_page_count)
 424                 bp->b_flags |= XBF_DONE;
 425
 426         XB_TRACE(bp, "lookup_pages", (long)page_count);
 427         return error;
 428 }
 429
 430 /*
 431  *      Map buffer into kernel address-space if nessecary.
 432  */
 433 STATIC int
 434 _xfs_buf_map_pages(
 435         xfs_buf_t               *bp,
 436         uint                    flags)
 437 {
 438         /* A single page buffer is always mappable */
 439         if (bp->b_page_count == 1) {
 440                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 441                 bp->b_flags |= XBF_MAPPED;
 442         } else if (flags & XBF_MAPPED) {
 443                 if (as_list_len > 64)
 444                         purge_addresses();
 445                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
 446                                         VM_MAP, PAGE_KERNEL);
 447                 if (unlikely(bp->b_addr == NULL))
 448                         return -ENOMEM;
 449                 bp->b_addr += bp->b_offset;
 450                 bp->b_flags |= XBF_MAPPED;
 451         }
 452
 453         return 0;
 454 }
 455
 456 /*
 457  *      Finding and Reading Buffers
 458  */
 459
 460 /*
 461  *      Look up, and creates if absent, a lockable buffer for
 462  *      a given range of an inode.  The buffer is returned
 463  *      locked.  If other overlapping buffers exist, they are
 464  *      released before the new buffer is created and locked,
 465  *      which may imply that this call will block until those buffers
 466  *      are unlocked.  No I/O is implied by this call.
 467  */
 468 xfs_buf_t *
 469 _xfs_buf_find(
 470         xfs_buftarg_t           *btp,   /* block device target          */
 471         xfs_off_t               ioff,   /* starting offset of range     */
 472         size_t                  isize,  /* length of range              */
 473         xfs_buf_flags_t         flags,
 474         xfs_buf_t               *new_bp)
 475 {
 476         xfs_off_t               range_base;
 477         size_t                  range_length;
 478         xfs_bufhash_t           *hash;
 479         xfs_buf_t               *bp, *n;
 480
 481         range_base = (ioff << BBSHIFT);
 482         range_length = (isize << BBSHIFT);
 483
 484         /* Check for IOs smaller than the sector size / not sector aligned */
 485         ASSERT(!(range_length < (1 << btp->bt_sshift)));
 486         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
 487
 488         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
 489
 490         spin_lock(&hash->bh_lock);
 491
 492         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
 493                 ASSERT(btp == bp->b_target);
 494                 if (bp->b_file_offset == range_base &&
 495                     bp->b_buffer_length == range_length) {
 496                         /*
 497                          * If we look at something, bring it to the
 498                          * front of the list for next time.
 499                          */
 500                         atomic_inc(&bp->b_hold);
 501                         list_move(&bp->b_hash_list, &hash->bh_list);
 502                         goto found;
 503                 }
 504         }
 505
 506         /* No match found */
 507         if (new_bp) {
 508                 _xfs_buf_initialize(new_bp, btp, range_base,
 509                                 range_length, flags);
 510                 new_bp->b_hash = hash;
 511                 list_add(&new_bp->b_hash_list, &hash->bh_list);
 512         } else {
 513                 XFS_STATS_INC(xb_miss_locked);
 514         }
 515
 516         spin_unlock(&hash->bh_lock);
 517         return new_bp;
 518
 519 found:
 520         spin_unlock(&hash->bh_lock);
 521
 522         /* Attempt to get the semaphore without sleeping,
 523          * if this does not work then we need to drop the
 524          * spinlock and do a hard attempt on the semaphore.
 525          */
 526         if (down_trylock(&bp->b_sema)) {
 527                 if (!(flags & XBF_TRYLOCK)) {
 528                         /* wait for buffer ownership */
 529                         XB_TRACE(bp, "get_lock", 0);
 530                         xfs_buf_lock(bp);
 531                         XFS_STATS_INC(xb_get_locked_waited);
 532                 } else {
 533                         /* We asked for a trylock and failed, no need
 534                          * to look at file offset and length here, we
 535                          * know that this buffer at least overlaps our
 536                          * buffer and is locked, therefore our buffer
 537                          * either does not exist, or is this buffer.
 538                          */
 539                         xfs_buf_rele(bp);
 540                         XFS_STATS_INC(xb_busy_locked);
 541                         return NULL;
 542                 }
 543         } else {
 544                 /* trylock worked */
 545                 XB_SET_OWNER(bp);
 546         }
 547
 548         if (bp->b_flags & XBF_STALE) {
 549                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 550                 bp->b_flags &= XBF_MAPPED;
 551         }
 552         XB_TRACE(bp, "got_lock", 0);
 553         XFS_STATS_INC(xb_get_locked);
 554         return bp;
 555 }
 556
 557 /*
 558  *      Assembles a buffer covering the specified range.
 559  *      Storage in memory for all portions of the buffer will be allocated,
 560  *      although backing storage may not be.
 561  */
 562 xfs_buf_t *
 563 xfs_buf_get_flags(
 564         xfs_buftarg_t           *target,/* target for buffer            */
 565         xfs_off_t               ioff,   /* starting offset of range     */
 566         size_t                  isize,  /* length of range              */
 567         xfs_buf_flags_t         flags)
 568 {
 569         xfs_buf_t               *bp, *new_bp;
 570         int                     error = 0, i;
 571
 572         new_bp = xfs_buf_allocate(flags);
 573         if (unlikely(!new_bp))
 574                 return NULL;
 575
 576         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 577         if (bp == new_bp) {
 578                 error = _xfs_buf_lookup_pages(bp, flags);
 579                 if (error)
 580                         goto no_buffer;
 581         } else {
 582                 xfs_buf_deallocate(new_bp);
 583                 if (unlikely(bp == NULL))
 584                         return NULL;
 585         }
 586
 587         for (i = 0; i < bp->b_page_count; i++)
 588                 mark_page_accessed(bp->b_pages[i]);
 589
 590         if (!(bp->b_flags & XBF_MAPPED)) {
 591                 error = _xfs_buf_map_pages(bp, flags);
 592                 if (unlikely(error)) {
 593                         printk(KERN_WARNING "%s: failed to map pages\n",
 594                                         __FUNCTION__);
 595                         goto no_buffer;
 596                 }
 597         }
 598
 599         XFS_STATS_INC(xb_get);
 600
 601         /*
 602          * Always fill in the block number now, the mapped cases can do
 603          * their own overlay of this later.
 604          */
 605         bp->b_bn = ioff;
 606         bp->b_count_desired = bp->b_buffer_length;
 607
 608         XB_TRACE(bp, "get", (unsigned long)flags);
 609         return bp;
 610
 611  no_buffer:
 612         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 613                 xfs_buf_unlock(bp);
 614         xfs_buf_rele(bp);
 615         return NULL;
 616 }
 617
 618 xfs_buf_t *
 619 xfs_buf_read_flags(
 620         xfs_buftarg_t           *target,
 621         xfs_off_t               ioff,
 622         size_t                  isize,
 623         xfs_buf_flags_t         flags)
 624 {
 625         xfs_buf_t               *bp;
 626
 627         flags |= XBF_READ;
 628
 629         bp = xfs_buf_get_flags(target, ioff, isize, flags);
 630         if (bp) {
 631                 if (!XFS_BUF_ISDONE(bp)) {
 632                         XB_TRACE(bp, "read", (unsigned long)flags);
 633                         XFS_STATS_INC(xb_get_read);
 634                         xfs_buf_iostart(bp, flags);
 635                 } else if (flags & XBF_ASYNC) {
 636                         XB_TRACE(bp, "read_async", (unsigned long)flags);
 637                         /*
 638                          * Read ahead call which is already satisfied,
 639                          * drop the buffer
 640                          */
 641                         goto no_buffer;
 642                 } else {
 643                         XB_TRACE(bp, "read_done", (unsigned long)flags);
 644                         /* We do not want read in the flags */
 645                         bp->b_flags &= ~XBF_READ;
 646                 }
 647         }
 648
 649         return bp;
 650
 651  no_buffer:
 652         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 653                 xfs_buf_unlock(bp);
 654         xfs_buf_rele(bp);
 655         return NULL;
 656 }
 657
 658 /*
 659  *      If we are not low on memory then do the readahead in a deadlock
 660  *      safe manner.
 661  */
 662 void
 663 xfs_buf_readahead(
 664         xfs_buftarg_t           *target,
 665         xfs_off_t               ioff,
 666         size_t                  isize,
 667         xfs_buf_flags_t         flags)
 668 {
 669         struct backing_dev_info *bdi;
 670
 671         bdi = target->bt_mapping->backing_dev_info;
 672         if (bdi_read_congested(bdi))
 673                 return;
 674
 675         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 676         xfs_buf_read_flags(target, ioff, isize, flags);
 677 }
 678
 679 xfs_buf_t *
 680 xfs_buf_get_empty(
 681         size_t                  len,
 682         xfs_buftarg_t           *target)
 683 {
 684         xfs_buf_t               *bp;
 685
 686         bp = xfs_buf_allocate(0);
 687         if (bp)
 688                 _xfs_buf_initialize(bp, target, 0, len, 0);
 689         return bp;
 690 }
 691
 692 static inline struct page *
 693 mem_to_page(
 694         void                    *addr)
 695 {
 696         if (((unsigned long)addr < VMALLOC_START) ||
 697             ((unsigned long)addr >= VMALLOC_END)) {
 698                 return virt_to_page(addr);
 699         } else {
 700                 return vmalloc_to_page(addr);
 701         }
 702 }
 703
 704 int
 705 xfs_buf_associate_memory(
 706         xfs_buf_t               *bp,
 707         void                    *mem,
 708         size_t                  len)
 709 {
 710         int                     rval;
 711         int                     i = 0;
 712         size_t                  ptr;
 713         size_t                  end, end_cur;
 714         off_t                   offset;
 715         int                     page_count;
 716
 717         page_count = PAGE_CACHE_ALIGN(len) >> PAGE_CACHE_SHIFT;
 718         offset = (off_t) mem - ((off_t)mem & PAGE_CACHE_MASK);
 719         if (offset && (len > PAGE_CACHE_SIZE))
 720                 page_count++;
 721
 722         /* Free any previous set of page pointers */
 723         if (bp->b_pages)
 724                 _xfs_buf_free_pages(bp);
 725
 726         bp->b_pages = NULL;
 727         bp->b_addr = mem;
 728
 729         rval = _xfs_buf_get_pages(bp, page_count, 0);
 730         if (rval)
 731                 return rval;
 732
 733         bp->b_offset = offset;
 734         ptr = (size_t) mem & PAGE_CACHE_MASK;
 735         end = PAGE_CACHE_ALIGN((size_t) mem + len);
 736         end_cur = end;
 737         /* set up first page */
 738         bp->b_pages[0] = mem_to_page(mem);
 739
 740         ptr += PAGE_CACHE_SIZE;
 741         bp->b_page_count = ++i;
 742         while (ptr < end) {
 743                 bp->b_pages[i] = mem_to_page((void *)ptr);
 744                 bp->b_page_count = ++i;
 745                 ptr += PAGE_CACHE_SIZE;
 746         }
 747         bp->b_locked = 0;
 748
 749         bp->b_count_desired = bp->b_buffer_length = len;
 750         bp->b_flags |= XBF_MAPPED;
 751
 752         return 0;
 753 }
 754
 755 xfs_buf_t *
 756 xfs_buf_get_noaddr(
 757         size_t                  len,
 758         xfs_buftarg_t           *target)
 759 {
 760         size_t                  malloc_len = len;
 761         xfs_buf_t               *bp;
 762         void                    *data;
 763         int                     error;
 764
 765         bp = xfs_buf_allocate(0);
 766         if (unlikely(bp == NULL))
 767                 goto fail;
 768         _xfs_buf_initialize(bp, target, 0, len, 0);
 769
 770  try_again:
 771         data = kmem_alloc(malloc_len, KM_SLEEP | KM_MAYFAIL);
 772         if (unlikely(data == NULL))
 773                 goto fail_free_buf;
 774
 775         /* check whether alignment matches.. */
 776         if ((__psunsigned_t)data !=
 777             ((__psunsigned_t)data & ~target->bt_smask)) {
 778                 /* .. else double the size and try again */
 779                 kmem_free(data, malloc_len);
 780                 malloc_len <<= 1;
 781                 goto try_again;
 782         }
 783
 784         error = xfs_buf_associate_memory(bp, data, len);
 785         if (error)
 786                 goto fail_free_mem;
 787         bp->b_flags |= _XBF_KMEM_ALLOC;
 788
 789         xfs_buf_unlock(bp);
 790
 791         XB_TRACE(bp, "no_daddr", data);
 792         return bp;
 793  fail_free_mem:
 794         kmem_free(data, malloc_len);
 795  fail_free_buf:
 796         xfs_buf_free(bp);
 797  fail:
 798         return NULL;
 799 }
 800
 801 /*
 802  *      Increment reference count on buffer, to hold the buffer concurrently
 803  *      with another thread which may release (free) the buffer asynchronously.
 804  *      Must hold the buffer already to call this function.
 805  */
 806 void
 807 xfs_buf_hold(
 808         xfs_buf_t               *bp)
 809 {
 810         atomic_inc(&bp->b_hold);
 811         XB_TRACE(bp, "hold", 0);
 812 }
 813
 814 /*
 815  *      Releases a hold on the specified buffer.  If the
 816  *      the hold count is 1, calls xfs_buf_free.
 817  */
 818 void
 819 xfs_buf_rele(
 820         xfs_buf_t               *bp)
 821 {
 822         xfs_bufhash_t           *hash = bp->b_hash;
 823
 824         XB_TRACE(bp, "rele", bp->b_relse);
 825
 826         if (unlikely(!hash)) {
 827                 ASSERT(!bp->b_relse);
 828                 if (atomic_dec_and_test(&bp->b_hold))
 829                         xfs_buf_free(bp);
 830                 return;
 831         }
 832
 833         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
 834                 if (bp->b_relse) {
 835                         atomic_inc(&bp->b_hold);
 836                         spin_unlock(&hash->bh_lock);
 837                         (*(bp->b_relse)) (bp);
 838                 } else if (bp->b_flags & XBF_FS_MANAGED) {
 839                         spin_unlock(&hash->bh_lock);
 840                 } else {
 841                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 842                         list_del_init(&bp->b_hash_list);
 843                         spin_unlock(&hash->bh_lock);
 844                         xfs_buf_free(bp);
 845                 }
 846         } else {
 847                 /*
 848                  * Catch reference count leaks
 849                  */
 850                 ASSERT(atomic_read(&bp->b_hold) >= 0);
 851         }
 852 }
 853
 854
 855 /*
 856  *      Mutual exclusion on buffers.  Locking model:
 857  *
 858  *      Buffers associated with inodes for which buffer locking
 859  *      is not enabled are not protected by semaphores, and are
 860  *      assumed to be exclusively owned by the caller.  There is a
 861  *      spinlock in the buffer, used by the caller when concurrent
 862  *      access is possible.
 863  */
 864
 865 /*
 866  *      Locks a buffer object, if it is not already locked.
 867  *      Note that this in no way locks the underlying pages, so it is only
 868  *      useful for synchronizing concurrent use of buffer objects, not for
 869  *      synchronizing independent access to the underlying pages.
 870  */
 871 int
 872 xfs_buf_cond_lock(
 873         xfs_buf_t               *bp)
 874 {
 875         int                     locked;
 876
 877         locked = down_trylock(&bp->b_sema) == 0;
 878         if (locked) {
 879                 XB_SET_OWNER(bp);
 880         }
 881         XB_TRACE(bp, "cond_lock", (long)locked);
 882         return locked ? 0 : -EBUSY;
 883 }
 884
 885 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
 886 int
 887 xfs_buf_lock_value(
 888         xfs_buf_t               *bp)
 889 {
 890         return atomic_read(&bp->b_sema.count);
 891 }
 892 #endif
 893
 894 /*
 895  *      Locks a buffer object.
 896  *      Note that this in no way locks the underlying pages, so it is only
 897  *      useful for synchronizing concurrent use of buffer objects, not for
 898  *      synchronizing independent access to the underlying pages.
 899  */
 900 void
 901 xfs_buf_lock(
 902         xfs_buf_t               *bp)
 903 {
 904         XB_TRACE(bp, "lock", 0);
 905         if (atomic_read(&bp->b_io_remaining))
 906                 blk_run_address_space(bp->b_target->bt_mapping);
 907         down(&bp->b_sema);
 908         XB_SET_OWNER(bp);
 909         XB_TRACE(bp, "locked", 0);
 910 }
 911
 912 /*
 913  *      Releases the lock on the buffer object.
 914  *      If the buffer is marked delwri but is not queued, do so before we
 915  *      unlock the buffer as we need to set flags correctly.  We also need to
 916  *      take a reference for the delwri queue because the unlocker is going to
 917  *      drop their's and they don't know we just queued it.
 918  */
 919 void
 920 xfs_buf_unlock(
 921         xfs_buf_t               *bp)
 922 {
 923         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 924                 atomic_inc(&bp->b_hold);
 925                 bp->b_flags |= XBF_ASYNC;
 926                 xfs_buf_delwri_queue(bp, 0);
 927         }
 928
 929         XB_CLEAR_OWNER(bp);
 930         up(&bp->b_sema);
 931         XB_TRACE(bp, "unlock", 0);
 932 }
 933
 934
 935 /*
 936  *      Pinning Buffer Storage in Memory
 937  *      Ensure that no attempt to force a buffer to disk will succeed.
 938  */
 939 void
 940 xfs_buf_pin(
 941         xfs_buf_t               *bp)
 942 {
 943         atomic_inc(&bp->b_pin_count);
 944         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
 945 }
 946
 947 void
 948 xfs_buf_unpin(
 949         xfs_buf_t               *bp)
 950 {
 951         if (atomic_dec_and_test(&bp->b_pin_count))
 952                 wake_up_all(&bp->b_waiters);
 953         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
 954 }
 955
 956 int
 957 xfs_buf_ispin(
 958         xfs_buf_t               *bp)
 959 {
 960         return atomic_read(&bp->b_pin_count);
 961 }
 962
 963 STATIC void
 964 xfs_buf_wait_unpin(
 965         xfs_buf_t               *bp)
 966 {
 967         DECLARE_WAITQUEUE       (wait, current);
 968
 969         if (atomic_read(&bp->b_pin_count) == 0)
 970                 return;
 971
 972         add_wait_queue(&bp->b_waiters, &wait);
 973         for (;;) {
 974                 set_current_state(TASK_UNINTERRUPTIBLE);
 975                 if (atomic_read(&bp->b_pin_count) == 0)
 976                         break;
 977                 if (atomic_read(&bp->b_io_remaining))
 978                         blk_run_address_space(bp->b_target->bt_mapping);
 979                 schedule();
 980         }
 981         remove_wait_queue(&bp->b_waiters, &wait);
 982         set_current_state(TASK_RUNNING);
 983 }
 984
 985 /*
 986  *      Buffer Utility Routines
 987  */
 988
 989 STATIC void
 990 xfs_buf_iodone_work(
 991         void                    *v)
 992 {
 993         xfs_buf_t               *bp = (xfs_buf_t *)v;
 994
 995         if (bp->b_iodone)
 996                 (*(bp->b_iodone))(bp);
 997         else if (bp->b_flags & XBF_ASYNC)
 998                 xfs_buf_relse(bp);
 999 }
1000
1001 void
1002 xfs_buf_ioend(
1003         xfs_buf_t               *bp,
1004         int                     schedule)
1005 {
1006         bp->b_flags &= ~(XBF_READ | XBF_WRITE);
1007         if (bp->b_error == 0)
1008                 bp->b_flags |= XBF_DONE;
1009
1010         XB_TRACE(bp, "iodone", bp->b_iodone);
1011
1012         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1013                 if (schedule) {
1014                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work, bp);
1015                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1016                 } else {
1017                         xfs_buf_iodone_work(bp);
1018                 }
1019         } else {
1020                 up(&bp->b_iodonesema);
1021         }
1022 }
1023
1024 void
1025 xfs_buf_ioerror(
1026         xfs_buf_t               *bp,
1027         int                     error)
1028 {
1029         ASSERT(error >= 0 && error <= 0xffff);
1030         bp->b_error = (unsigned short)error;
1031         XB_TRACE(bp, "ioerror", (unsigned long)error);
1032 }
1033
1034 /*
1035  *      Initiate I/O on a buffer, based on the flags supplied.
1036  *      The b_iodone routine in the buffer supplied will only be called
1037  *      when all of the subsidiary I/O requests, if any, have been completed.
1038  */
1039 int
1040 xfs_buf_iostart(
1041         xfs_buf_t               *bp,
1042         xfs_buf_flags_t         flags)
1043 {
1044         int                     status = 0;
1045
1046         XB_TRACE(bp, "iostart", (unsigned long)flags);
1047
1048         if (flags & XBF_DELWRI) {
1049                 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1050                 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1051                 xfs_buf_delwri_queue(bp, 1);
1052                 return status;
1053         }
1054
1055         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1056                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1057         bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1058                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1059
1060         BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1061
1062         /* For writes allow an alternate strategy routine to precede
1063          * the actual I/O request (which may not be issued at all in
1064          * a shutdown situation, for example).
1065          */
1066         status = (flags & XBF_WRITE) ?
1067                 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1068
1069         /* Wait for I/O if we are not an async request.
1070          * Note: async I/O request completion will release the buffer,
1071          * and that can already be done by this point.  So using the
1072          * buffer pointer from here on, after async I/O, is invalid.
1073          */
1074         if (!status && !(flags & XBF_ASYNC))
1075                 status = xfs_buf_iowait(bp);
1076
1077         return status;
1078 }
1079
1080 STATIC __inline__ int
1081 _xfs_buf_iolocked(
1082         xfs_buf_t               *bp)
1083 {
1084         ASSERT(bp->b_flags & (XBF_READ | XBF_WRITE));
1085         if (bp->b_flags & XBF_READ)
1086                 return bp->b_locked;
1087         return 0;
1088 }
1089
1090 STATIC __inline__ void
1091 _xfs_buf_ioend(
1092         xfs_buf_t               *bp,
1093         int                     schedule)
1094 {
1095         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1096                 bp->b_locked = 0;
1097                 xfs_buf_ioend(bp, schedule);
1098         }
1099 }
1100
1101 STATIC int
1102 xfs_buf_bio_end_io(
1103         struct bio              *bio,
1104         unsigned int            bytes_done,
1105         int                     error)
1106 {
1107         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1108         unsigned int            blocksize = bp->b_target->bt_bsize;
1109         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1110
1111         if (bio->bi_size)
1112                 return 1;
1113
1114         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1115                 bp->b_error = EIO;
1116
1117         do {
1118                 struct page     *page = bvec->bv_page;
1119
1120                 if (unlikely(bp->b_error)) {
1121                         if (bp->b_flags & XBF_READ)
1122                                 ClearPageUptodate(page);
1123                         SetPageError(page);
1124                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1125                         SetPageUptodate(page);
1126                 } else if (!PagePrivate(page) &&
1127                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1128                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1129                 }
1130
1131                 if (--bvec >= bio->bi_io_vec)
1132                         prefetchw(&bvec->bv_page->flags);
1133
1134                 if (_xfs_buf_iolocked(bp)) {
1135                         unlock_page(page);
1136                 }
1137         } while (bvec >= bio->bi_io_vec);
1138
1139         _xfs_buf_ioend(bp, 1);
1140         bio_put(bio);
1141         return 0;
1142 }
1143
1144 STATIC void
1145 _xfs_buf_ioapply(
1146         xfs_buf_t               *bp)
1147 {
1148         int                     i, rw, map_i, total_nr_pages, nr_pages;
1149         struct bio              *bio;
1150         int                     offset = bp->b_offset;
1151         int                     size = bp->b_count_desired;
1152         sector_t                sector = bp->b_bn;
1153         unsigned int            blocksize = bp->b_target->bt_bsize;
1154         int                     locking = _xfs_buf_iolocked(bp);
1155
1156         total_nr_pages = bp->b_page_count;
1157         map_i = 0;
1158
1159         if (bp->b_flags & _XBF_RUN_QUEUES) {
1160                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1161                 rw = (bp->b_flags & XBF_READ) ? READ_SYNC : WRITE_SYNC;
1162         } else {
1163                 rw = (bp->b_flags & XBF_READ) ? READ : WRITE;
1164         }
1165
1166         if (bp->b_flags & XBF_ORDERED) {
1167                 ASSERT(!(bp->b_flags & XBF_READ));
1168                 rw = WRITE_BARRIER;
1169         }
1170
1171         /* Special code path for reading a sub page size buffer in --
1172          * we populate up the whole page, and hence the other metadata
1173          * in the same page.  This optimization is only valid when the
1174          * filesystem block size is not smaller than the page size.
1175          */
1176         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1177             (bp->b_flags & XBF_READ) && locking &&
1178             (blocksize >= PAGE_CACHE_SIZE)) {
1179                 bio = bio_alloc(GFP_NOIO, 1);
1180
1181                 bio->bi_bdev = bp->b_target->bt_bdev;
1182                 bio->bi_sector = sector - (offset >> BBSHIFT);
1183                 bio->bi_end_io = xfs_buf_bio_end_io;
1184                 bio->bi_private = bp;
1185
1186                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1187                 size = 0;
1188
1189                 atomic_inc(&bp->b_io_remaining);
1190
1191                 goto submit_io;
1192         }
1193
1194         /* Lock down the pages which we need to for the request */
1195         if (locking && (bp->b_flags & XBF_WRITE) && (bp->b_locked == 0)) {
1196                 for (i = 0; size; i++) {
1197                         int             nbytes = PAGE_CACHE_SIZE - offset;
1198                         struct page     *page = bp->b_pages[i];
1199
1200                         if (nbytes > size)
1201                                 nbytes = size;
1202
1203                         lock_page(page);
1204
1205                         size -= nbytes;
1206                         offset = 0;
1207                 }
1208                 offset = bp->b_offset;
1209                 size = bp->b_count_desired;
1210         }
1211
1212 next_chunk:
1213         atomic_inc(&bp->b_io_remaining);
1214         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1215         if (nr_pages > total_nr_pages)
1216                 nr_pages = total_nr_pages;
1217
1218         bio = bio_alloc(GFP_NOIO, nr_pages);
1219         bio->bi_bdev = bp->b_target->bt_bdev;
1220         bio->bi_sector = sector;
1221         bio->bi_end_io = xfs_buf_bio_end_io;
1222         bio->bi_private = bp;
1223
1224         for (; size && nr_pages; nr_pages--, map_i++) {
1225                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1226
1227                 if (nbytes > size)
1228                         nbytes = size;
1229
1230                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1231                 if (rbytes < nbytes)
1232                         break;
1233
1234                 offset = 0;
1235                 sector += nbytes >> BBSHIFT;
1236                 size -= nbytes;
1237                 total_nr_pages--;
1238         }
1239
1240 submit_io:
1241         if (likely(bio->bi_size)) {
1242                 submit_bio(rw, bio);
1243                 if (size)
1244                         goto next_chunk;
1245         } else {
1246                 bio_put(bio);
1247                 xfs_buf_ioerror(bp, EIO);
1248         }
1249 }
1250
1251 int
1252 xfs_buf_iorequest(
1253         xfs_buf_t               *bp)
1254 {
1255         XB_TRACE(bp, "iorequest", 0);
1256
1257         if (bp->b_flags & XBF_DELWRI) {
1258                 xfs_buf_delwri_queue(bp, 1);
1259                 return 0;
1260         }
1261
1262         if (bp->b_flags & XBF_WRITE) {
1263                 xfs_buf_wait_unpin(bp);
1264         }
1265
1266         xfs_buf_hold(bp);
1267
1268         /* Set the count to 1 initially, this will stop an I/O
1269          * completion callout which happens before we have started
1270          * all the I/O from calling xfs_buf_ioend too early.
1271          */
1272         atomic_set(&bp->b_io_remaining, 1);
1273         _xfs_buf_ioapply(bp);
1274         _xfs_buf_ioend(bp, 0);
1275
1276         xfs_buf_rele(bp);
1277         return 0;
1278 }
1279
1280 /*
1281  *      Waits for I/O to complete on the buffer supplied.
1282  *      It returns immediately if no I/O is pending.
1283  *      It returns the I/O error code, if any, or 0 if there was no error.
1284  */
1285 int
1286 xfs_buf_iowait(
1287         xfs_buf_t               *bp)
1288 {
1289         XB_TRACE(bp, "iowait", 0);
1290         if (atomic_read(&bp->b_io_remaining))
1291                 blk_run_address_space(bp->b_target->bt_mapping);
1292         down(&bp->b_iodonesema);
1293         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1294         return bp->b_error;
1295 }
1296
1297 xfs_caddr_t
1298 xfs_buf_offset(
1299         xfs_buf_t               *bp,
1300         size_t                  offset)
1301 {
1302         struct page             *page;
1303
1304         if (bp->b_flags & XBF_MAPPED)
1305                 return XFS_BUF_PTR(bp) + offset;
1306
1307         offset += bp->b_offset;
1308         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1309         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1310 }
1311
1312 /*
1313  *      Move data into or out of a buffer.
1314  */
1315 void
1316 xfs_buf_iomove(
1317         xfs_buf_t               *bp,    /* buffer to process            */
1318         size_t                  boff,   /* starting buffer offset       */
1319         size_t                  bsize,  /* length to copy               */
1320         caddr_t                 data,   /* data address                 */
1321         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1322 {
1323         size_t                  bend, cpoff, csize;
1324         struct page             *page;
1325
1326         bend = boff + bsize;
1327         while (boff < bend) {
1328                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1329                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1330                 csize = min_t(size_t,
1331                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1332
1333                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1334
1335                 switch (mode) {
1336                 case XBRW_ZERO:
1337                         memset(page_address(page) + cpoff, 0, csize);
1338                         break;
1339                 case XBRW_READ:
1340                         memcpy(data, page_address(page) + cpoff, csize);
1341                         break;
1342                 case XBRW_WRITE:
1343                         memcpy(page_address(page) + cpoff, data, csize);
1344                 }
1345
1346                 boff += csize;
1347                 data += csize;
1348         }
1349 }
1350
1351 /*
1352  *      Handling of buffer targets (buftargs).
1353  */
1354
1355 /*
1356  *      Wait for any bufs with callbacks that have been submitted but
1357  *      have not yet returned... walk the hash list for the target.
1358  */
1359 void
1360 xfs_wait_buftarg(
1361         xfs_buftarg_t   *btp)
1362 {
1363         xfs_buf_t       *bp, *n;
1364         xfs_bufhash_t   *hash;
1365         uint            i;
1366
1367         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1368                 hash = &btp->bt_hash[i];
1369 again:
1370                 spin_lock(&hash->bh_lock);
1371                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1372                         ASSERT(btp == bp->b_target);
1373                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1374                                 spin_unlock(&hash->bh_lock);
1375                                 /*
1376                                  * Catch superblock reference count leaks
1377                                  * immediately
1378                                  */
1379                                 BUG_ON(bp->b_bn == 0);
1380                                 delay(100);
1381                                 goto again;
1382                         }
1383                 }
1384                 spin_unlock(&hash->bh_lock);
1385         }
1386 }
1387
1388 /*
1389  *      Allocate buffer hash table for a given target.
1390  *      For devices containing metadata (i.e. not the log/realtime devices)
1391  *      we need to allocate a much larger hash table.
1392  */
1393 STATIC void
1394 xfs_alloc_bufhash(
1395         xfs_buftarg_t           *btp,
1396         int                     external)
1397 {
1398         unsigned int            i;
1399
1400         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1401         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1402         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1403                                         sizeof(xfs_bufhash_t), KM_SLEEP);
1404         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1405                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1406                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1407         }
1408 }
1409
1410 STATIC void
1411 xfs_free_bufhash(
1412         xfs_buftarg_t           *btp)
1413 {
1414         kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1415         btp->bt_hash = NULL;
1416 }
1417
1418 /*
1419  *      buftarg list for delwrite queue processing
1420  */
1421 STATIC LIST_HEAD(xfs_buftarg_list);
1422 STATIC DEFINE_SPINLOCK(xfs_buftarg_lock);
1423
1424 STATIC void
1425 xfs_register_buftarg(
1426         xfs_buftarg_t           *btp)
1427 {
1428         spin_lock(&xfs_buftarg_lock);
1429         list_add(&btp->bt_list, &xfs_buftarg_list);
1430         spin_unlock(&xfs_buftarg_lock);
1431 }
1432
1433 STATIC void
1434 xfs_unregister_buftarg(
1435         xfs_buftarg_t           *btp)
1436 {
1437         spin_lock(&xfs_buftarg_lock);
1438         list_del(&btp->bt_list);
1439         spin_unlock(&xfs_buftarg_lock);
1440 }
1441
1442 void
1443 xfs_free_buftarg(
1444         xfs_buftarg_t           *btp,
1445         int                     external)
1446 {
1447         xfs_flush_buftarg(btp, 1);
1448         if (external)
1449                 xfs_blkdev_put(btp->bt_bdev);
1450         xfs_free_bufhash(btp);
1451         iput(btp->bt_mapping->host);
1452
1453         /* Unregister the buftarg first so that we don't get a
1454          * wakeup finding a non-existent task
1455          */
1456         xfs_unregister_buftarg(btp);
1457         kthread_stop(btp->bt_task);
1458
1459         kmem_free(btp, sizeof(*btp));
1460 }
1461
1462 STATIC int
1463 xfs_setsize_buftarg_flags(
1464         xfs_buftarg_t           *btp,
1465         unsigned int            blocksize,
1466         unsigned int            sectorsize,
1467         int                     verbose)
1468 {
1469         btp->bt_bsize = blocksize;
1470         btp->bt_sshift = ffs(sectorsize) - 1;
1471         btp->bt_smask = sectorsize - 1;
1472
1473         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1474                 printk(KERN_WARNING
1475                         "XFS: Cannot set_blocksize to %u on device %s\n",
1476                         sectorsize, XFS_BUFTARG_NAME(btp));
1477                 return EINVAL;
1478         }
1479
1480         if (verbose &&
1481             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1482                 printk(KERN_WARNING
1483                         "XFS: %u byte sectors in use on device %s.  "
1484                         "This is suboptimal; %u or greater is ideal.\n",
1485                         sectorsize, XFS_BUFTARG_NAME(btp),
1486                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1487         }
1488
1489         return 0;
1490 }
1491
1492 /*
1493  *      When allocating the initial buffer target we have not yet
1494  *      read in the superblock, so don't know what sized sectors
1495  *      are being used is at this early stage.  Play safe.
1496  */
1497 STATIC int
1498 xfs_setsize_buftarg_early(
1499         xfs_buftarg_t           *btp,
1500         struct block_device     *bdev)
1501 {
1502         return xfs_setsize_buftarg_flags(btp,
1503                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1504 }
1505
1506 int
1507 xfs_setsize_buftarg(
1508         xfs_buftarg_t           *btp,
1509         unsigned int            blocksize,
1510         unsigned int            sectorsize)
1511 {
1512         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1513 }
1514
1515 STATIC int
1516 xfs_mapping_buftarg(
1517         xfs_buftarg_t           *btp,
1518         struct block_device     *bdev)
1519 {
1520         struct backing_dev_info *bdi;
1521         struct inode            *inode;
1522         struct address_space    *mapping;
1523         static struct address_space_operations mapping_aops = {
1524                 .sync_page = block_sync_page,
1525                 .migratepage = fail_migrate_page,
1526         };
1527
1528         inode = new_inode(bdev->bd_inode->i_sb);
1529         if (!inode) {
1530                 printk(KERN_WARNING
1531                         "XFS: Cannot allocate mapping inode for device %s\n",
1532                         XFS_BUFTARG_NAME(btp));
1533                 return ENOMEM;
1534         }
1535         inode->i_mode = S_IFBLK;
1536         inode->i_bdev = bdev;
1537         inode->i_rdev = bdev->bd_dev;
1538         bdi = blk_get_backing_dev_info(bdev);
1539         if (!bdi)
1540                 bdi = &default_backing_dev_info;
1541         mapping = &inode->i_data;
1542         mapping->a_ops = &mapping_aops;
1543         mapping->backing_dev_info = bdi;
1544         mapping_set_gfp_mask(mapping, GFP_NOFS);
1545         btp->bt_mapping = mapping;
1546         return 0;
1547 }
1548
1549 STATIC int
1550 xfs_alloc_delwrite_queue(
1551         xfs_buftarg_t           *btp)
1552 {
1553         int     error = 0;
1554
1555         INIT_LIST_HEAD(&btp->bt_list);
1556         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1557         spinlock_init(&btp->bt_delwrite_lock, "delwri_lock");
1558         btp->bt_flags = 0;
1559         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1560         if (IS_ERR(btp->bt_task)) {
1561                 error = PTR_ERR(btp->bt_task);
1562                 goto out_error;
1563         }
1564         xfs_register_buftarg(btp);
1565 out_error:
1566         return error;
1567 }
1568
1569 xfs_buftarg_t *
1570 xfs_alloc_buftarg(
1571         struct block_device     *bdev,
1572         int                     external)
1573 {
1574         xfs_buftarg_t           *btp;
1575
1576         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1577
1578         btp->bt_dev =  bdev->bd_dev;
1579         btp->bt_bdev = bdev;
1580         if (xfs_setsize_buftarg_early(btp, bdev))
1581                 goto error;
1582         if (xfs_mapping_buftarg(btp, bdev))
1583                 goto error;
1584         if (xfs_alloc_delwrite_queue(btp))
1585                 goto error;
1586         xfs_alloc_bufhash(btp, external);
1587         return btp;
1588
1589 error:
1590         kmem_free(btp, sizeof(*btp));
1591         return NULL;
1592 }
1593
1594
1595 /*
1596  *      Delayed write buffer handling
1597  */
1598 STATIC void
1599 xfs_buf_delwri_queue(
1600         xfs_buf_t               *bp,
1601         int                     unlock)
1602 {
1603         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1604         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1605
1606         XB_TRACE(bp, "delwri_q", (long)unlock);
1607         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1608
1609         spin_lock(dwlk);
1610         /* If already in the queue, dequeue and place at tail */
1611         if (!list_empty(&bp->b_list)) {
1612                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1613                 if (unlock)
1614                         atomic_dec(&bp->b_hold);
1615                 list_del(&bp->b_list);
1616         }
1617
1618         bp->b_flags |= _XBF_DELWRI_Q;
1619         list_add_tail(&bp->b_list, dwq);
1620         bp->b_queuetime = jiffies;
1621         spin_unlock(dwlk);
1622
1623         if (unlock)
1624                 xfs_buf_unlock(bp);
1625 }
1626
1627 void
1628 xfs_buf_delwri_dequeue(
1629         xfs_buf_t               *bp)
1630 {
1631         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1632         int                     dequeued = 0;
1633
1634         spin_lock(dwlk);
1635         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1636                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1637                 list_del_init(&bp->b_list);
1638                 dequeued = 1;
1639         }
1640         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1641         spin_unlock(dwlk);
1642
1643         if (dequeued)
1644                 xfs_buf_rele(bp);
1645
1646         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1647 }
1648
1649 STATIC void
1650 xfs_buf_runall_queues(
1651         struct workqueue_struct *queue)
1652 {
1653         flush_workqueue(queue);
1654 }
1655
1656 STATIC int
1657 xfsbufd_wakeup(
1658         int                     priority,
1659         gfp_t                   mask)
1660 {
1661         xfs_buftarg_t           *btp;
1662
1663         spin_lock(&xfs_buftarg_lock);
1664         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1665                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1666                         continue;
1667                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1668                 wake_up_process(btp->bt_task);
1669         }
1670         spin_unlock(&xfs_buftarg_lock);
1671         return 0;
1672 }
1673
1674 STATIC int
1675 xfsbufd(
1676         void                    *data)
1677 {
1678         struct list_head        tmp;
1679         unsigned long           age;
1680         xfs_buftarg_t           *target = (xfs_buftarg_t *)data;
1681         xfs_buf_t               *bp, *n;
1682         struct list_head        *dwq = &target->bt_delwrite_queue;
1683         spinlock_t              *dwlk = &target->bt_delwrite_lock;
1684
1685         current->flags |= PF_MEMALLOC;
1686
1687         INIT_LIST_HEAD(&tmp);
1688         do {
1689                 if (unlikely(freezing(current))) {
1690                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1691                         refrigerator();
1692                 } else {
1693                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1694                 }
1695
1696                 schedule_timeout_interruptible(
1697                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1698
1699                 age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1700                 spin_lock(dwlk);
1701                 list_for_each_entry_safe(bp, n, dwq, b_list) {
1702                         XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1703                         ASSERT(bp->b_flags & XBF_DELWRI);
1704
1705                         if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1706                                 if (!test_bit(XBT_FORCE_FLUSH,
1707                                                 &target->bt_flags) &&
1708                                     time_before(jiffies,
1709                                                 bp->b_queuetime + age)) {
1710                                         xfs_buf_unlock(bp);
1711                                         break;
1712                                 }
1713
1714                                 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1715                                 bp->b_flags |= XBF_WRITE;
1716                                 list_move(&bp->b_list, &tmp);
1717                         }
1718                 }
1719                 spin_unlock(dwlk);
1720
1721                 while (!list_empty(&tmp)) {
1722                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1723                         ASSERT(target == bp->b_target);
1724
1725                         list_del_init(&bp->b_list);
1726                         xfs_buf_iostrategy(bp);
1727
1728                         blk_run_address_space(target->bt_mapping);
1729                 }
1730
1731                 if (as_list_len > 0)
1732                         purge_addresses();
1733
1734                 clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1735         } while (!kthread_should_stop());
1736
1737         return 0;
1738 }
1739
1740 /*
1741  *      Go through all incore buffers, and release buffers if they belong to
1742  *      the given device. This is used in filesystem error handling to
1743  *      preserve the consistency of its metadata.
1744  */
1745 int
1746 xfs_flush_buftarg(
1747         xfs_buftarg_t           *target,
1748         int                     wait)
1749 {
1750         struct list_head        tmp;
1751         xfs_buf_t               *bp, *n;
1752         int                     pincount = 0;
1753         struct list_head        *dwq = &target->bt_delwrite_queue;
1754         spinlock_t              *dwlk = &target->bt_delwrite_lock;
1755
1756         xfs_buf_runall_queues(xfsdatad_workqueue);
1757         xfs_buf_runall_queues(xfslogd_workqueue);
1758
1759         INIT_LIST_HEAD(&tmp);
1760         spin_lock(dwlk);
1761         list_for_each_entry_safe(bp, n, dwq, b_list) {
1762                 ASSERT(bp->b_target == target);
1763                 ASSERT(bp->b_flags & (XBF_DELWRI | _XBF_DELWRI_Q));
1764                 XB_TRACE(bp, "walkq2", (long)xfs_buf_ispin(bp));
1765                 if (xfs_buf_ispin(bp)) {
1766                         pincount++;
1767                         continue;
1768                 }
1769
1770                 list_move(&bp->b_list, &tmp);
1771         }
1772         spin_unlock(dwlk);
1773
1774         /*
1775          * Dropped the delayed write list lock, now walk the temporary list
1776          */
1777         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1778                 xfs_buf_lock(bp);
1779                 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1780                 bp->b_flags |= XBF_WRITE;
1781                 if (wait)
1782                         bp->b_flags &= ~XBF_ASYNC;
1783                 else
1784                         list_del_init(&bp->b_list);
1785
1786                 xfs_buf_iostrategy(bp);
1787         }
1788
1789         /*
1790          * Remaining list items must be flushed before returning
1791          */
1792         while (!list_empty(&tmp)) {
1793                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1794
1795                 list_del_init(&bp->b_list);
1796                 xfs_iowait(bp);
1797                 xfs_buf_relse(bp);
1798         }
1799
1800         if (wait)
1801                 blk_run_address_space(target->bt_mapping);
1802
1803         return pincount;
1804 }
1805
1806 int __init
1807 xfs_buf_init(void)
1808 {
1809         int             error = -ENOMEM;
1810
1811 #ifdef XFS_BUF_TRACE
1812         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1813 #endif
1814
1815         xfs_buf_zone = kmem_zone_init(sizeof(xfs_buf_t), "xfs_buf");
1816         if (!xfs_buf_zone)
1817                 goto out_free_trace_buf;
1818
1819         xfslogd_workqueue = create_workqueue("xfslogd");
1820         if (!xfslogd_workqueue)
1821                 goto out_free_buf_zone;
1822
1823         xfsdatad_workqueue = create_workqueue("xfsdatad");
1824         if (!xfsdatad_workqueue)
1825                 goto out_destroy_xfslogd_workqueue;
1826
1827         xfs_buf_shake = kmem_shake_register(xfsbufd_wakeup);
1828         if (!xfs_buf_shake)
1829                 goto out_destroy_xfsdatad_workqueue;
1830
1831         return 0;
1832
1833  out_destroy_xfsdatad_workqueue:
1834         destroy_workqueue(xfsdatad_workqueue);
1835  out_destroy_xfslogd_workqueue:
1836         destroy_workqueue(xfslogd_workqueue);
1837  out_free_buf_zone:
1838         kmem_zone_destroy(xfs_buf_zone);
1839  out_free_trace_buf:
1840 #ifdef XFS_BUF_TRACE
1841         ktrace_free(xfs_buf_trace_buf);
1842 #endif
1843         return error;
1844 }
1845
1846 void
1847 xfs_buf_terminate(void)
1848 {
1849         kmem_shake_deregister(xfs_buf_shake);
1850         destroy_workqueue(xfsdatad_workqueue);
1851         destroy_workqueue(xfslogd_workqueue);
1852         kmem_zone_destroy(xfs_buf_zone);
1853 #ifdef XFS_BUF_TRACE
1854         ktrace_free(xfs_buf_trace_buf);
1855 #endif
1856 }