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 int xfsbufd(void *);
  39 STATIC int xfsbufd_wakeup(int, gfp_t);
  40 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
  41 static struct shrinker xfs_buf_shake = {
  42         .shrink = xfsbufd_wakeup,
  43         .seeks = DEFAULT_SEEKS,
  44 };
  45
  46 static struct workqueue_struct *xfslogd_workqueue;
  47 struct workqueue_struct *xfsdatad_workqueue;
  48
  49 #ifdef XFS_BUF_TRACE
  50 void
  51 xfs_buf_trace(
  52         xfs_buf_t       *bp,
  53         char            *id,
  54         void            *data,
  55         void            *ra)
  56 {
  57         ktrace_enter(xfs_buf_trace_buf,
  58                 bp, id,
  59                 (void *)(unsigned long)bp->b_flags,
  60                 (void *)(unsigned long)bp->b_hold.counter,
  61                 (void *)(unsigned long)bp->b_sema.count.counter,
  62                 (void *)current,
  63                 data, ra,
  64                 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
  65                 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
  66                 (void *)(unsigned long)bp->b_buffer_length,
  67                 NULL, NULL, NULL, NULL, NULL);
  68 }
  69 ktrace_t *xfs_buf_trace_buf;
  70 #define XFS_BUF_TRACE_SIZE      4096
  71 #define XB_TRACE(bp, id, data)  \
  72         xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
  73 #else
  74 #define XB_TRACE(bp, id, data)  do { } while (0)
  75 #endif
  76
  77 #ifdef XFS_BUF_LOCK_TRACKING
  78 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  79 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  80 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  81 #else
  82 # define XB_SET_OWNER(bp)       do { } while (0)
  83 # define XB_CLEAR_OWNER(bp)     do { } while (0)
  84 # define XB_GET_OWNER(bp)       do { } while (0)
  85 #endif
  86
  87 #define xb_to_gfp(flags) \
  88         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
  89           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
  90
  91 #define xb_to_km(flags) \
  92          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
  93
  94 #define xfs_buf_allocate(flags) \
  95         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
  96 #define xfs_buf_deallocate(bp) \
  97         kmem_zone_free(xfs_buf_zone, (bp));
  98
  99 /*
 100  *      Page Region interfaces.
 101  *
 102  *      For pages in filesystems where the blocksize is smaller than the
 103  *      pagesize, we use the page->private field (long) to hold a bitmap
 104  *      of uptodate regions within the page.
 105  *
 106  *      Each such region is "bytes per page / bits per long" bytes long.
 107  *
 108  *      NBPPR == number-of-bytes-per-page-region
 109  *      BTOPR == bytes-to-page-region (rounded up)
 110  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
 111  */
 112 #if (BITS_PER_LONG == 32)
 113 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
 114 #elif (BITS_PER_LONG == 64)
 115 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
 116 #else
 117 #error BITS_PER_LONG must be 32 or 64
 118 #endif
 119 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
 120 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
 121 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
 122
 123 STATIC unsigned long
 124 page_region_mask(
 125         size_t          offset,
 126         size_t          length)
 127 {
 128         unsigned long   mask;
 129         int             first, final;
 130
 131         first = BTOPR(offset);
 132         final = BTOPRT(offset + length - 1);
 133         first = min(first, final);
 134
 135         mask = ~0UL;
 136         mask <<= BITS_PER_LONG - (final - first);
 137         mask >>= BITS_PER_LONG - (final);
 138
 139         ASSERT(offset + length <= PAGE_CACHE_SIZE);
 140         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
 141
 142         return mask;
 143 }
 144
 145 STATIC_INLINE void
 146 set_page_region(
 147         struct page     *page,
 148         size_t          offset,
 149         size_t          length)
 150 {
 151         set_page_private(page,
 152                 page_private(page) | page_region_mask(offset, length));
 153         if (page_private(page) == ~0UL)
 154                 SetPageUptodate(page);
 155 }
 156
 157 STATIC_INLINE int
 158 test_page_region(
 159         struct page     *page,
 160         size_t          offset,
 161         size_t          length)
 162 {
 163         unsigned long   mask = page_region_mask(offset, length);
 164
 165         return (mask && (page_private(page) & mask) == mask);
 166 }
 167
 168 /*
 169  *      Mapping of multi-page buffers into contiguous virtual space
 170  */
 171
 172 typedef struct a_list {
 173         void            *vm_addr;
 174         struct a_list   *next;
 175 } a_list_t;
 176
 177 static a_list_t         *as_free_head;
 178 static int              as_list_len;
 179 static DEFINE_SPINLOCK(as_lock);
 180
 181 /*
 182  *      Try to batch vunmaps because they are costly.
 183  */
 184 STATIC void
 185 free_address(
 186         void            *addr)
 187 {
 188         a_list_t        *aentry;
 189
 190 #ifdef CONFIG_XEN
 191         /*
 192          * Xen needs to be able to make sure it can get an exclusive
 193          * RO mapping of pages it wants to turn into a pagetable.  If
 194          * a newly allocated page is also still being vmap()ed by xfs,
 195          * it will cause pagetable construction to fail.  This is a
 196          * quick workaround to always eagerly unmap pages so that Xen
 197          * is happy.
 198          */
 199         vunmap(addr);
 200         return;
 201 #endif
 202
 203         aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
 204         if (likely(aentry)) {
 205                 spin_lock(&as_lock);
 206                 aentry->next = as_free_head;
 207                 aentry->vm_addr = addr;
 208                 as_free_head = aentry;
 209                 as_list_len++;
 210                 spin_unlock(&as_lock);
 211         } else {
 212                 vunmap(addr);
 213         }
 214 }
 215
 216 STATIC void
 217 purge_addresses(void)
 218 {
 219         a_list_t        *aentry, *old;
 220
 221         if (as_free_head == NULL)
 222                 return;
 223
 224         spin_lock(&as_lock);
 225         aentry = as_free_head;
 226         as_free_head = NULL;
 227         as_list_len = 0;
 228         spin_unlock(&as_lock);
 229
 230         while ((old = aentry) != NULL) {
 231                 vunmap(aentry->vm_addr);
 232                 aentry = aentry->next;
 233                 kfree(old);
 234         }
 235 }
 236
 237 /*
 238  *      Internal xfs_buf_t object manipulation
 239  */
 240
 241 STATIC void
 242 _xfs_buf_initialize(
 243         xfs_buf_t               *bp,
 244         xfs_buftarg_t           *target,
 245         xfs_off_t               range_base,
 246         size_t                  range_length,
 247         xfs_buf_flags_t         flags)
 248 {
 249         /*
 250          * We don't want certain flags to appear in b_flags.
 251          */
 252         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
 253
 254         memset(bp, 0, sizeof(xfs_buf_t));
 255         atomic_set(&bp->b_hold, 1);
 256         init_MUTEX_LOCKED(&bp->b_iodonesema);
 257         INIT_LIST_HEAD(&bp->b_list);
 258         INIT_LIST_HEAD(&bp->b_hash_list);
 259         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
 260         XB_SET_OWNER(bp);
 261         bp->b_target = target;
 262         bp->b_file_offset = range_base;
 263         /*
 264          * Set buffer_length and count_desired to the same value initially.
 265          * I/O routines should use count_desired, which will be the same in
 266          * most cases but may be reset (e.g. XFS recovery).
 267          */
 268         bp->b_buffer_length = bp->b_count_desired = range_length;
 269         bp->b_flags = flags;
 270         bp->b_bn = XFS_BUF_DADDR_NULL;
 271         atomic_set(&bp->b_pin_count, 0);
 272         init_waitqueue_head(&bp->b_waiters);
 273
 274         XFS_STATS_INC(xb_create);
 275         XB_TRACE(bp, "initialize", target);
 276 }
 277
 278 /*
 279  *      Allocate a page array capable of holding a specified number
 280  *      of pages, and point the page buf at it.
 281  */
 282 STATIC int
 283 _xfs_buf_get_pages(
 284         xfs_buf_t               *bp,
 285         int                     page_count,
 286         xfs_buf_flags_t         flags)
 287 {
 288         /* Make sure that we have a page list */
 289         if (bp->b_pages == NULL) {
 290                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
 291                 bp->b_page_count = page_count;
 292                 if (page_count <= XB_PAGES) {
 293                         bp->b_pages = bp->b_page_array;
 294                 } else {
 295                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
 296                                         page_count, xb_to_km(flags));
 297                         if (bp->b_pages == NULL)
 298                                 return -ENOMEM;
 299                 }
 300                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 301         }
 302         return 0;
 303 }
 304
 305 /*
 306  *      Frees b_pages if it was allocated.
 307  */
 308 STATIC void
 309 _xfs_buf_free_pages(
 310         xfs_buf_t       *bp)
 311 {
 312         if (bp->b_pages != bp->b_page_array) {
 313                 kmem_free(bp->b_pages,
 314                           bp->b_page_count * sizeof(struct page *));
 315         }
 316 }
 317
 318 /*
 319  *      Releases the specified buffer.
 320  *
 321  *      The modification state of any associated pages is left unchanged.
 322  *      The buffer most not be on any hash - use xfs_buf_rele instead for
 323  *      hashed and refcounted buffers
 324  */
 325 void
 326 xfs_buf_free(
 327         xfs_buf_t               *bp)
 328 {
 329         XB_TRACE(bp, "free", 0);
 330
 331         ASSERT(list_empty(&bp->b_hash_list));
 332
 333         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
 334                 uint            i;
 335
 336                 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
 337                         free_address(bp->b_addr - bp->b_offset);
 338
 339                 for (i = 0; i < bp->b_page_count; i++) {
 340                         struct page     *page = bp->b_pages[i];
 341
 342                         if (bp->b_flags & _XBF_PAGE_CACHE)
 343                                 ASSERT(!PagePrivate(page));
 344                         page_cache_release(page);
 345                 }
 346                 _xfs_buf_free_pages(bp);
 347         }
 348
 349         xfs_buf_deallocate(bp);
 350 }
 351
 352 /*
 353  *      Finds all pages for buffer in question and builds it's page list.
 354  */
 355 STATIC int
 356 _xfs_buf_lookup_pages(
 357         xfs_buf_t               *bp,
 358         uint                    flags)
 359 {
 360         struct address_space    *mapping = bp->b_target->bt_mapping;
 361         size_t                  blocksize = bp->b_target->bt_bsize;
 362         size_t                  size = bp->b_count_desired;
 363         size_t                  nbytes, offset;
 364         gfp_t                   gfp_mask = xb_to_gfp(flags);
 365         unsigned short          page_count, i;
 366         pgoff_t                 first;
 367         xfs_off_t               end;
 368         int                     error;
 369
 370         end = bp->b_file_offset + bp->b_buffer_length;
 371         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 372
 373         error = _xfs_buf_get_pages(bp, page_count, flags);
 374         if (unlikely(error))
 375                 return error;
 376         bp->b_flags |= _XBF_PAGE_CACHE;
 377
 378         offset = bp->b_offset;
 379         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
 380
 381         for (i = 0; i < bp->b_page_count; i++) {
 382                 struct page     *page;
 383                 uint            retries = 0;
 384
 385               retry:
 386                 page = find_or_create_page(mapping, first + i, gfp_mask);
 387                 if (unlikely(page == NULL)) {
 388                         if (flags & XBF_READ_AHEAD) {
 389                                 bp->b_page_count = i;
 390                                 for (i = 0; i < bp->b_page_count; i++)
 391                                         unlock_page(bp->b_pages[i]);
 392                                 return -ENOMEM;
 393                         }
 394
 395                         /*
 396                          * This could deadlock.
 397                          *
 398                          * But until all the XFS lowlevel code is revamped to
 399                          * handle buffer allocation failures we can't do much.
 400                          */
 401                         if (!(++retries % 100))
 402                                 printk(KERN_ERR
 403                                         "XFS: possible memory allocation "
 404                                         "deadlock in %s (mode:0x%x)\n",
 405                                         __FUNCTION__, gfp_mask);
 406
 407                         XFS_STATS_INC(xb_page_retries);
 408                         xfsbufd_wakeup(0, gfp_mask);
 409                         congestion_wait(WRITE, HZ/50);
 410                         goto retry;
 411                 }
 412
 413                 XFS_STATS_INC(xb_page_found);
 414
 415                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
 416                 size -= nbytes;
 417
 418                 ASSERT(!PagePrivate(page));
 419                 if (!PageUptodate(page)) {
 420                         page_count--;
 421                         if (blocksize >= PAGE_CACHE_SIZE) {
 422                                 if (flags & XBF_READ)
 423                                         bp->b_locked = 1;
 424                         } else if (!PagePrivate(page)) {
 425                                 if (test_page_region(page, offset, nbytes))
 426                                         page_count++;
 427                         }
 428                 }
 429
 430                 bp->b_pages[i] = page;
 431                 offset = 0;
 432         }
 433
 434         if (!bp->b_locked) {
 435                 for (i = 0; i < bp->b_page_count; i++)
 436                         unlock_page(bp->b_pages[i]);
 437         }
 438
 439         if (page_count == bp->b_page_count)
 440                 bp->b_flags |= XBF_DONE;
 441
 442         XB_TRACE(bp, "lookup_pages", (long)page_count);
 443         return error;
 444 }
 445
 446 /*
 447  *      Map buffer into kernel address-space if nessecary.
 448  */
 449 STATIC int
 450 _xfs_buf_map_pages(
 451         xfs_buf_t               *bp,
 452         uint                    flags)
 453 {
 454         /* A single page buffer is always mappable */
 455         if (bp->b_page_count == 1) {
 456                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 457                 bp->b_flags |= XBF_MAPPED;
 458         } else if (flags & XBF_MAPPED) {
 459                 if (as_list_len > 64)
 460                         purge_addresses();
 461                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
 462                                         VM_MAP, PAGE_KERNEL);
 463                 if (unlikely(bp->b_addr == NULL))
 464                         return -ENOMEM;
 465                 bp->b_addr += bp->b_offset;
 466                 bp->b_flags |= XBF_MAPPED;
 467         }
 468
 469         return 0;
 470 }
 471
 472 /*
 473  *      Finding and Reading Buffers
 474  */
 475
 476 /*
 477  *      Look up, and creates if absent, a lockable buffer for
 478  *      a given range of an inode.  The buffer is returned
 479  *      locked.  If other overlapping buffers exist, they are
 480  *      released before the new buffer is created and locked,
 481  *      which may imply that this call will block until those buffers
 482  *      are unlocked.  No I/O is implied by this call.
 483  */
 484 xfs_buf_t *
 485 _xfs_buf_find(
 486         xfs_buftarg_t           *btp,   /* block device target          */
 487         xfs_off_t               ioff,   /* starting offset of range     */
 488         size_t                  isize,  /* length of range              */
 489         xfs_buf_flags_t         flags,
 490         xfs_buf_t               *new_bp)
 491 {
 492         xfs_off_t               range_base;
 493         size_t                  range_length;
 494         xfs_bufhash_t           *hash;
 495         xfs_buf_t               *bp, *n;
 496
 497         range_base = (ioff << BBSHIFT);
 498         range_length = (isize << BBSHIFT);
 499
 500         /* Check for IOs smaller than the sector size / not sector aligned */
 501         ASSERT(!(range_length < (1 << btp->bt_sshift)));
 502         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
 503
 504         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
 505
 506         spin_lock(&hash->bh_lock);
 507
 508         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
 509                 ASSERT(btp == bp->b_target);
 510                 if (bp->b_file_offset == range_base &&
 511                     bp->b_buffer_length == range_length) {
 512                         /*
 513                          * If we look at something, bring it to the
 514                          * front of the list for next time.
 515                          */
 516                         atomic_inc(&bp->b_hold);
 517                         list_move(&bp->b_hash_list, &hash->bh_list);
 518                         goto found;
 519                 }
 520         }
 521
 522         /* No match found */
 523         if (new_bp) {
 524                 _xfs_buf_initialize(new_bp, btp, range_base,
 525                                 range_length, flags);
 526                 new_bp->b_hash = hash;
 527                 list_add(&new_bp->b_hash_list, &hash->bh_list);
 528         } else {
 529                 XFS_STATS_INC(xb_miss_locked);
 530         }
 531
 532         spin_unlock(&hash->bh_lock);
 533         return new_bp;
 534
 535 found:
 536         spin_unlock(&hash->bh_lock);
 537
 538         /* Attempt to get the semaphore without sleeping,
 539          * if this does not work then we need to drop the
 540          * spinlock and do a hard attempt on the semaphore.
 541          */
 542         if (down_trylock(&bp->b_sema)) {
 543                 if (!(flags & XBF_TRYLOCK)) {
 544                         /* wait for buffer ownership */
 545                         XB_TRACE(bp, "get_lock", 0);
 546                         xfs_buf_lock(bp);
 547                         XFS_STATS_INC(xb_get_locked_waited);
 548                 } else {
 549                         /* We asked for a trylock and failed, no need
 550                          * to look at file offset and length here, we
 551                          * know that this buffer at least overlaps our
 552                          * buffer and is locked, therefore our buffer
 553                          * either does not exist, or is this buffer.
 554                          */
 555                         xfs_buf_rele(bp);
 556                         XFS_STATS_INC(xb_busy_locked);
 557                         return NULL;
 558                 }
 559         } else {
 560                 /* trylock worked */
 561                 XB_SET_OWNER(bp);
 562         }
 563
 564         if (bp->b_flags & XBF_STALE) {
 565                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 566                 bp->b_flags &= XBF_MAPPED;
 567         }
 568         XB_TRACE(bp, "got_lock", 0);
 569         XFS_STATS_INC(xb_get_locked);
 570         return bp;
 571 }
 572
 573 /*
 574  *      Assembles a buffer covering the specified range.
 575  *      Storage in memory for all portions of the buffer will be allocated,
 576  *      although backing storage may not be.
 577  */
 578 xfs_buf_t *
 579 xfs_buf_get_flags(
 580         xfs_buftarg_t           *target,/* target for buffer            */
 581         xfs_off_t               ioff,   /* starting offset of range     */
 582         size_t                  isize,  /* length of range              */
 583         xfs_buf_flags_t         flags)
 584 {
 585         xfs_buf_t               *bp, *new_bp;
 586         int                     error = 0, i;
 587
 588         new_bp = xfs_buf_allocate(flags);
 589         if (unlikely(!new_bp))
 590                 return NULL;
 591
 592         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 593         if (bp == new_bp) {
 594                 error = _xfs_buf_lookup_pages(bp, flags);
 595                 if (error)
 596                         goto no_buffer;
 597         } else {
 598                 xfs_buf_deallocate(new_bp);
 599                 if (unlikely(bp == NULL))
 600                         return NULL;
 601         }
 602
 603         for (i = 0; i < bp->b_page_count; i++)
 604                 mark_page_accessed(bp->b_pages[i]);
 605
 606         if (!(bp->b_flags & XBF_MAPPED)) {
 607                 error = _xfs_buf_map_pages(bp, flags);
 608                 if (unlikely(error)) {
 609                         printk(KERN_WARNING "%s: failed to map pages\n",
 610                                         __FUNCTION__);
 611                         goto no_buffer;
 612                 }
 613         }
 614
 615         XFS_STATS_INC(xb_get);
 616
 617         /*
 618          * Always fill in the block number now, the mapped cases can do
 619          * their own overlay of this later.
 620          */
 621         bp->b_bn = ioff;
 622         bp->b_count_desired = bp->b_buffer_length;
 623
 624         XB_TRACE(bp, "get", (unsigned long)flags);
 625         return bp;
 626
 627  no_buffer:
 628         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 629                 xfs_buf_unlock(bp);
 630         xfs_buf_rele(bp);
 631         return NULL;
 632 }
 633
 634 xfs_buf_t *
 635 xfs_buf_read_flags(
 636         xfs_buftarg_t           *target,
 637         xfs_off_t               ioff,
 638         size_t                  isize,
 639         xfs_buf_flags_t         flags)
 640 {
 641         xfs_buf_t               *bp;
 642
 643         flags |= XBF_READ;
 644
 645         bp = xfs_buf_get_flags(target, ioff, isize, flags);
 646         if (bp) {
 647                 if (!XFS_BUF_ISDONE(bp)) {
 648                         XB_TRACE(bp, "read", (unsigned long)flags);
 649                         XFS_STATS_INC(xb_get_read);
 650                         xfs_buf_iostart(bp, flags);
 651                 } else if (flags & XBF_ASYNC) {
 652                         XB_TRACE(bp, "read_async", (unsigned long)flags);
 653                         /*
 654                          * Read ahead call which is already satisfied,
 655                          * drop the buffer
 656                          */
 657                         goto no_buffer;
 658                 } else {
 659                         XB_TRACE(bp, "read_done", (unsigned long)flags);
 660                         /* We do not want read in the flags */
 661                         bp->b_flags &= ~XBF_READ;
 662                 }
 663         }
 664
 665         return bp;
 666
 667  no_buffer:
 668         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 669                 xfs_buf_unlock(bp);
 670         xfs_buf_rele(bp);
 671         return NULL;
 672 }
 673
 674 /*
 675  *      If we are not low on memory then do the readahead in a deadlock
 676  *      safe manner.
 677  */
 678 void
 679 xfs_buf_readahead(
 680         xfs_buftarg_t           *target,
 681         xfs_off_t               ioff,
 682         size_t                  isize,
 683         xfs_buf_flags_t         flags)
 684 {
 685         struct backing_dev_info *bdi;
 686
 687         bdi = target->bt_mapping->backing_dev_info;
 688         if (bdi_read_congested(bdi))
 689                 return;
 690
 691         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 692         xfs_buf_read_flags(target, ioff, isize, flags);
 693 }
 694
 695 xfs_buf_t *
 696 xfs_buf_get_empty(
 697         size_t                  len,
 698         xfs_buftarg_t           *target)
 699 {
 700         xfs_buf_t               *bp;
 701
 702         bp = xfs_buf_allocate(0);
 703         if (bp)
 704                 _xfs_buf_initialize(bp, target, 0, len, 0);
 705         return bp;
 706 }
 707
 708 static inline struct page *
 709 mem_to_page(
 710         void                    *addr)
 711 {
 712         if ((!is_vmalloc_addr(addr))) {
 713                 return virt_to_page(addr);
 714         } else {
 715                 return vmalloc_to_page(addr);
 716         }
 717 }
 718
 719 int
 720 xfs_buf_associate_memory(
 721         xfs_buf_t               *bp,
 722         void                    *mem,
 723         size_t                  len)
 724 {
 725         int                     rval;
 726         int                     i = 0;
 727         unsigned long           pageaddr;
 728         unsigned long           offset;
 729         size_t                  buflen;
 730         int                     page_count;
 731
 732         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
 733         offset = (unsigned long)mem - pageaddr;
 734         buflen = PAGE_CACHE_ALIGN(len + offset);
 735         page_count = buflen >> PAGE_CACHE_SHIFT;
 736
 737         /* Free any previous set of page pointers */
 738         if (bp->b_pages)
 739                 _xfs_buf_free_pages(bp);
 740
 741         bp->b_pages = NULL;
 742         bp->b_addr = mem;
 743
 744         rval = _xfs_buf_get_pages(bp, page_count, 0);
 745         if (rval)
 746                 return rval;
 747
 748         bp->b_offset = offset;
 749
 750         for (i = 0; i < bp->b_page_count; i++) {
 751                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 752                 pageaddr += PAGE_CACHE_SIZE;
 753         }
 754         bp->b_locked = 0;
 755
 756         bp->b_count_desired = len;
 757         bp->b_buffer_length = buflen;
 758         bp->b_flags |= XBF_MAPPED;
 759
 760         return 0;
 761 }
 762
 763 xfs_buf_t *
 764 xfs_buf_get_noaddr(
 765         size_t                  len,
 766         xfs_buftarg_t           *target)
 767 {
 768         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 769         int                     error, i;
 770         xfs_buf_t               *bp;
 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         error = _xfs_buf_get_pages(bp, page_count, 0);
 778         if (error)
 779                 goto fail_free_buf;
 780
 781         for (i = 0; i < page_count; i++) {
 782                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
 783                 if (!bp->b_pages[i])
 784                         goto fail_free_mem;
 785         }
 786         bp->b_flags |= _XBF_PAGES;
 787
 788         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 789         if (unlikely(error)) {
 790                 printk(KERN_WARNING "%s: failed to map pages\n",
 791                                 __FUNCTION__);
 792                 goto fail_free_mem;
 793         }
 794
 795         xfs_buf_unlock(bp);
 796
 797         XB_TRACE(bp, "no_daddr", len);
 798         return bp;
 799
 800  fail_free_mem:
 801         while (--i >= 0)
 802                 __free_page(bp->b_pages[i]);
 803         _xfs_buf_free_pages(bp);
 804  fail_free_buf:
 805         xfs_buf_deallocate(bp);
 806  fail:
 807         return NULL;
 808 }
 809
 810 /*
 811  *      Increment reference count on buffer, to hold the buffer concurrently
 812  *      with another thread which may release (free) the buffer asynchronously.
 813  *      Must hold the buffer already to call this function.
 814  */
 815 void
 816 xfs_buf_hold(
 817         xfs_buf_t               *bp)
 818 {
 819         atomic_inc(&bp->b_hold);
 820         XB_TRACE(bp, "hold", 0);
 821 }
 822
 823 /*
 824  *      Releases a hold on the specified buffer.  If the
 825  *      the hold count is 1, calls xfs_buf_free.
 826  */
 827 void
 828 xfs_buf_rele(
 829         xfs_buf_t               *bp)
 830 {
 831         xfs_bufhash_t           *hash = bp->b_hash;
 832
 833         XB_TRACE(bp, "rele", bp->b_relse);
 834
 835         if (unlikely(!hash)) {
 836                 ASSERT(!bp->b_relse);
 837                 if (atomic_dec_and_test(&bp->b_hold))
 838                         xfs_buf_free(bp);
 839                 return;
 840         }
 841
 842         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
 843                 if (bp->b_relse) {
 844                         atomic_inc(&bp->b_hold);
 845                         spin_unlock(&hash->bh_lock);
 846                         (*(bp->b_relse)) (bp);
 847                 } else if (bp->b_flags & XBF_FS_MANAGED) {
 848                         spin_unlock(&hash->bh_lock);
 849                 } else {
 850                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 851                         list_del_init(&bp->b_hash_list);
 852                         spin_unlock(&hash->bh_lock);
 853                         xfs_buf_free(bp);
 854                 }
 855         } else {
 856                 /*
 857                  * Catch reference count leaks
 858                  */
 859                 ASSERT(atomic_read(&bp->b_hold) >= 0);
 860         }
 861 }
 862
 863
 864 /*
 865  *      Mutual exclusion on buffers.  Locking model:
 866  *
 867  *      Buffers associated with inodes for which buffer locking
 868  *      is not enabled are not protected by semaphores, and are
 869  *      assumed to be exclusively owned by the caller.  There is a
 870  *      spinlock in the buffer, used by the caller when concurrent
 871  *      access is possible.
 872  */
 873
 874 /*
 875  *      Locks a buffer object, if it is not already locked.
 876  *      Note that this in no way locks the underlying pages, so it is only
 877  *      useful for synchronizing concurrent use of buffer objects, not for
 878  *      synchronizing independent access to the underlying pages.
 879  */
 880 int
 881 xfs_buf_cond_lock(
 882         xfs_buf_t               *bp)
 883 {
 884         int                     locked;
 885
 886         locked = down_trylock(&bp->b_sema) == 0;
 887         if (locked) {
 888                 XB_SET_OWNER(bp);
 889         }
 890         XB_TRACE(bp, "cond_lock", (long)locked);
 891         return locked ? 0 : -EBUSY;
 892 }
 893
 894 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
 895 int
 896 xfs_buf_lock_value(
 897         xfs_buf_t               *bp)
 898 {
 899         return atomic_read(&bp->b_sema.count);
 900 }
 901 #endif
 902
 903 /*
 904  *      Locks a buffer object.
 905  *      Note that this in no way locks the underlying pages, so it is only
 906  *      useful for synchronizing concurrent use of buffer objects, not for
 907  *      synchronizing independent access to the underlying pages.
 908  */
 909 void
 910 xfs_buf_lock(
 911         xfs_buf_t               *bp)
 912 {
 913         XB_TRACE(bp, "lock", 0);
 914         if (atomic_read(&bp->b_io_remaining))
 915                 blk_run_address_space(bp->b_target->bt_mapping);
 916         down(&bp->b_sema);
 917         XB_SET_OWNER(bp);
 918         XB_TRACE(bp, "locked", 0);
 919 }
 920
 921 /*
 922  *      Releases the lock on the buffer object.
 923  *      If the buffer is marked delwri but is not queued, do so before we
 924  *      unlock the buffer as we need to set flags correctly.  We also need to
 925  *      take a reference for the delwri queue because the unlocker is going to
 926  *      drop their's and they don't know we just queued it.
 927  */
 928 void
 929 xfs_buf_unlock(
 930         xfs_buf_t               *bp)
 931 {
 932         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 933                 atomic_inc(&bp->b_hold);
 934                 bp->b_flags |= XBF_ASYNC;
 935                 xfs_buf_delwri_queue(bp, 0);
 936         }
 937
 938         XB_CLEAR_OWNER(bp);
 939         up(&bp->b_sema);
 940         XB_TRACE(bp, "unlock", 0);
 941 }
 942
 943
 944 /*
 945  *      Pinning Buffer Storage in Memory
 946  *      Ensure that no attempt to force a buffer to disk will succeed.
 947  */
 948 void
 949 xfs_buf_pin(
 950         xfs_buf_t               *bp)
 951 {
 952         atomic_inc(&bp->b_pin_count);
 953         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
 954 }
 955
 956 void
 957 xfs_buf_unpin(
 958         xfs_buf_t               *bp)
 959 {
 960         if (atomic_dec_and_test(&bp->b_pin_count))
 961                 wake_up_all(&bp->b_waiters);
 962         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
 963 }
 964
 965 int
 966 xfs_buf_ispin(
 967         xfs_buf_t               *bp)
 968 {
 969         return atomic_read(&bp->b_pin_count);
 970 }
 971
 972 STATIC void
 973 xfs_buf_wait_unpin(
 974         xfs_buf_t               *bp)
 975 {
 976         DECLARE_WAITQUEUE       (wait, current);
 977
 978         if (atomic_read(&bp->b_pin_count) == 0)
 979                 return;
 980
 981         add_wait_queue(&bp->b_waiters, &wait);
 982         for (;;) {
 983                 set_current_state(TASK_UNINTERRUPTIBLE);
 984                 if (atomic_read(&bp->b_pin_count) == 0)
 985                         break;
 986                 if (atomic_read(&bp->b_io_remaining))
 987                         blk_run_address_space(bp->b_target->bt_mapping);
 988                 schedule();
 989         }
 990         remove_wait_queue(&bp->b_waiters, &wait);
 991         set_current_state(TASK_RUNNING);
 992 }
 993
 994 /*
 995  *      Buffer Utility Routines
 996  */
 997
 998 STATIC void
 999 xfs_buf_iodone_work(
1000         struct work_struct      *work)
1001 {
1002         xfs_buf_t               *bp =
1003                 container_of(work, xfs_buf_t, b_iodone_work);
1004
1005         /*
1006          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
1007          * ordered flag and reissue them.  Because we can't tell the higher
1008          * layers directly that they should not issue ordered I/O anymore, they
1009          * need to check if the ordered flag was cleared during I/O completion.
1010          */
1011         if ((bp->b_error == EOPNOTSUPP) &&
1012             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1013                 XB_TRACE(bp, "ordered_retry", bp->b_iodone);
1014                 bp->b_flags &= ~XBF_ORDERED;
1015                 xfs_buf_iorequest(bp);
1016         } else if (bp->b_iodone)
1017                 (*(bp->b_iodone))(bp);
1018         else if (bp->b_flags & XBF_ASYNC)
1019                 xfs_buf_relse(bp);
1020 }
1021
1022 void
1023 xfs_buf_ioend(
1024         xfs_buf_t               *bp,
1025         int                     schedule)
1026 {
1027         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1028         if (bp->b_error == 0)
1029                 bp->b_flags |= XBF_DONE;
1030
1031         XB_TRACE(bp, "iodone", bp->b_iodone);
1032
1033         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1034                 if (schedule) {
1035                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1036                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1037                 } else {
1038                         xfs_buf_iodone_work(&bp->b_iodone_work);
1039                 }
1040         } else {
1041                 up(&bp->b_iodonesema);
1042         }
1043 }
1044
1045 void
1046 xfs_buf_ioerror(
1047         xfs_buf_t               *bp,
1048         int                     error)
1049 {
1050         ASSERT(error >= 0 && error <= 0xffff);
1051         bp->b_error = (unsigned short)error;
1052         XB_TRACE(bp, "ioerror", (unsigned long)error);
1053 }
1054
1055 /*
1056  *      Initiate I/O on a buffer, based on the flags supplied.
1057  *      The b_iodone routine in the buffer supplied will only be called
1058  *      when all of the subsidiary I/O requests, if any, have been completed.
1059  */
1060 int
1061 xfs_buf_iostart(
1062         xfs_buf_t               *bp,
1063         xfs_buf_flags_t         flags)
1064 {
1065         int                     status = 0;
1066
1067         XB_TRACE(bp, "iostart", (unsigned long)flags);
1068
1069         if (flags & XBF_DELWRI) {
1070                 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1071                 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1072                 xfs_buf_delwri_queue(bp, 1);
1073                 return status;
1074         }
1075
1076         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1077                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1078         bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1079                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1080
1081         BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1082
1083         /* For writes allow an alternate strategy routine to precede
1084          * the actual I/O request (which may not be issued at all in
1085          * a shutdown situation, for example).
1086          */
1087         status = (flags & XBF_WRITE) ?
1088                 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1089
1090         /* Wait for I/O if we are not an async request.
1091          * Note: async I/O request completion will release the buffer,
1092          * and that can already be done by this point.  So using the
1093          * buffer pointer from here on, after async I/O, is invalid.
1094          */
1095         if (!status && !(flags & XBF_ASYNC))
1096                 status = xfs_buf_iowait(bp);
1097
1098         return status;
1099 }
1100
1101 STATIC_INLINE int
1102 _xfs_buf_iolocked(
1103         xfs_buf_t               *bp)
1104 {
1105         ASSERT(bp->b_flags & (XBF_READ | XBF_WRITE));
1106         if (bp->b_flags & XBF_READ)
1107                 return bp->b_locked;
1108         return 0;
1109 }
1110
1111 STATIC_INLINE void
1112 _xfs_buf_ioend(
1113         xfs_buf_t               *bp,
1114         int                     schedule)
1115 {
1116         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1117                 bp->b_locked = 0;
1118                 xfs_buf_ioend(bp, schedule);
1119         }
1120 }
1121
1122 STATIC void
1123 xfs_buf_bio_end_io(
1124         struct bio              *bio,
1125         int                     error)
1126 {
1127         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1128         unsigned int            blocksize = bp->b_target->bt_bsize;
1129         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1130
1131         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1132                 bp->b_error = EIO;
1133
1134         do {
1135                 struct page     *page = bvec->bv_page;
1136
1137                 ASSERT(!PagePrivate(page));
1138                 if (unlikely(bp->b_error)) {
1139                         if (bp->b_flags & XBF_READ)
1140                                 ClearPageUptodate(page);
1141                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1142                         SetPageUptodate(page);
1143                 } else if (!PagePrivate(page) &&
1144                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1145                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1146                 }
1147
1148                 if (--bvec >= bio->bi_io_vec)
1149                         prefetchw(&bvec->bv_page->flags);
1150
1151                 if (_xfs_buf_iolocked(bp)) {
1152                         unlock_page(page);
1153                 }
1154         } while (bvec >= bio->bi_io_vec);
1155
1156         _xfs_buf_ioend(bp, 1);
1157         bio_put(bio);
1158 }
1159
1160 STATIC void
1161 _xfs_buf_ioapply(
1162         xfs_buf_t               *bp)
1163 {
1164         int                     i, rw, map_i, total_nr_pages, nr_pages;
1165         struct bio              *bio;
1166         int                     offset = bp->b_offset;
1167         int                     size = bp->b_count_desired;
1168         sector_t                sector = bp->b_bn;
1169         unsigned int            blocksize = bp->b_target->bt_bsize;
1170         int                     locking = _xfs_buf_iolocked(bp);
1171
1172         total_nr_pages = bp->b_page_count;
1173         map_i = 0;
1174
1175         if (bp->b_flags & XBF_ORDERED) {
1176                 ASSERT(!(bp->b_flags & XBF_READ));
1177                 rw = WRITE_BARRIER;
1178         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1179                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1180                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1181                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1182         } else {
1183                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1184                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1185         }
1186
1187         /* Special code path for reading a sub page size buffer in --
1188          * we populate up the whole page, and hence the other metadata
1189          * in the same page.  This optimization is only valid when the
1190          * filesystem block size is not smaller than the page size.
1191          */
1192         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1193             (bp->b_flags & XBF_READ) && locking &&
1194             (blocksize >= PAGE_CACHE_SIZE)) {
1195                 bio = bio_alloc(GFP_NOIO, 1);
1196
1197                 bio->bi_bdev = bp->b_target->bt_bdev;
1198                 bio->bi_sector = sector - (offset >> BBSHIFT);
1199                 bio->bi_end_io = xfs_buf_bio_end_io;
1200                 bio->bi_private = bp;
1201
1202                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1203                 size = 0;
1204
1205                 atomic_inc(&bp->b_io_remaining);
1206
1207                 goto submit_io;
1208         }
1209
1210         /* Lock down the pages which we need to for the request */
1211         if (locking && (bp->b_flags & XBF_WRITE) && (bp->b_locked == 0)) {
1212                 for (i = 0; size; i++) {
1213                         int             nbytes = PAGE_CACHE_SIZE - offset;
1214                         struct page     *page = bp->b_pages[i];
1215
1216                         if (nbytes > size)
1217                                 nbytes = size;
1218
1219                         lock_page(page);
1220
1221                         size -= nbytes;
1222                         offset = 0;
1223                 }
1224                 offset = bp->b_offset;
1225                 size = bp->b_count_desired;
1226         }
1227
1228 next_chunk:
1229         atomic_inc(&bp->b_io_remaining);
1230         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1231         if (nr_pages > total_nr_pages)
1232                 nr_pages = total_nr_pages;
1233
1234         bio = bio_alloc(GFP_NOIO, nr_pages);
1235         bio->bi_bdev = bp->b_target->bt_bdev;
1236         bio->bi_sector = sector;
1237         bio->bi_end_io = xfs_buf_bio_end_io;
1238         bio->bi_private = bp;
1239
1240         for (; size && nr_pages; nr_pages--, map_i++) {
1241                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1242
1243                 if (nbytes > size)
1244                         nbytes = size;
1245
1246                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1247                 if (rbytes < nbytes)
1248                         break;
1249
1250                 offset = 0;
1251                 sector += nbytes >> BBSHIFT;
1252                 size -= nbytes;
1253                 total_nr_pages--;
1254         }
1255
1256 submit_io:
1257         if (likely(bio->bi_size)) {
1258                 submit_bio(rw, bio);
1259                 if (size)
1260                         goto next_chunk;
1261         } else {
1262                 bio_put(bio);
1263                 xfs_buf_ioerror(bp, EIO);
1264         }
1265 }
1266
1267 int
1268 xfs_buf_iorequest(
1269         xfs_buf_t               *bp)
1270 {
1271         XB_TRACE(bp, "iorequest", 0);
1272
1273         if (bp->b_flags & XBF_DELWRI) {
1274                 xfs_buf_delwri_queue(bp, 1);
1275                 return 0;
1276         }
1277
1278         if (bp->b_flags & XBF_WRITE) {
1279                 xfs_buf_wait_unpin(bp);
1280         }
1281
1282         xfs_buf_hold(bp);
1283
1284         /* Set the count to 1 initially, this will stop an I/O
1285          * completion callout which happens before we have started
1286          * all the I/O from calling xfs_buf_ioend too early.
1287          */
1288         atomic_set(&bp->b_io_remaining, 1);
1289         _xfs_buf_ioapply(bp);
1290         _xfs_buf_ioend(bp, 0);
1291
1292         xfs_buf_rele(bp);
1293         return 0;
1294 }
1295
1296 /*
1297  *      Waits for I/O to complete on the buffer supplied.
1298  *      It returns immediately if no I/O is pending.
1299  *      It returns the I/O error code, if any, or 0 if there was no error.
1300  */
1301 int
1302 xfs_buf_iowait(
1303         xfs_buf_t               *bp)
1304 {
1305         XB_TRACE(bp, "iowait", 0);
1306         if (atomic_read(&bp->b_io_remaining))
1307                 blk_run_address_space(bp->b_target->bt_mapping);
1308         down(&bp->b_iodonesema);
1309         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1310         return bp->b_error;
1311 }
1312
1313 xfs_caddr_t
1314 xfs_buf_offset(
1315         xfs_buf_t               *bp,
1316         size_t                  offset)
1317 {
1318         struct page             *page;
1319
1320         if (bp->b_flags & XBF_MAPPED)
1321                 return XFS_BUF_PTR(bp) + offset;
1322
1323         offset += bp->b_offset;
1324         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1325         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1326 }
1327
1328 /*
1329  *      Move data into or out of a buffer.
1330  */
1331 void
1332 xfs_buf_iomove(
1333         xfs_buf_t               *bp,    /* buffer to process            */
1334         size_t                  boff,   /* starting buffer offset       */
1335         size_t                  bsize,  /* length to copy               */
1336         caddr_t                 data,   /* data address                 */
1337         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1338 {
1339         size_t                  bend, cpoff, csize;
1340         struct page             *page;
1341
1342         bend = boff + bsize;
1343         while (boff < bend) {
1344                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1345                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1346                 csize = min_t(size_t,
1347                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1348
1349                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1350
1351                 switch (mode) {
1352                 case XBRW_ZERO:
1353                         memset(page_address(page) + cpoff, 0, csize);
1354                         break;
1355                 case XBRW_READ:
1356                         memcpy(data, page_address(page) + cpoff, csize);
1357                         break;
1358                 case XBRW_WRITE:
1359                         memcpy(page_address(page) + cpoff, data, csize);
1360                 }
1361
1362                 boff += csize;
1363                 data += csize;
1364         }
1365 }
1366
1367 /*
1368  *      Handling of buffer targets (buftargs).
1369  */
1370
1371 /*
1372  *      Wait for any bufs with callbacks that have been submitted but
1373  *      have not yet returned... walk the hash list for the target.
1374  */
1375 void
1376 xfs_wait_buftarg(
1377         xfs_buftarg_t   *btp)
1378 {
1379         xfs_buf_t       *bp, *n;
1380         xfs_bufhash_t   *hash;
1381         uint            i;
1382
1383         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1384                 hash = &btp->bt_hash[i];
1385 again:
1386                 spin_lock(&hash->bh_lock);
1387                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1388                         ASSERT(btp == bp->b_target);
1389                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1390                                 spin_unlock(&hash->bh_lock);
1391                                 /*
1392                                  * Catch superblock reference count leaks
1393                                  * immediately
1394                                  */
1395                                 BUG_ON(bp->b_bn == 0);
1396                                 delay(100);
1397                                 goto again;
1398                         }
1399                 }
1400                 spin_unlock(&hash->bh_lock);
1401         }
1402 }
1403
1404 /*
1405  *      Allocate buffer hash table for a given target.
1406  *      For devices containing metadata (i.e. not the log/realtime devices)
1407  *      we need to allocate a much larger hash table.
1408  */
1409 STATIC void
1410 xfs_alloc_bufhash(
1411         xfs_buftarg_t           *btp,
1412         int                     external)
1413 {
1414         unsigned int            i;
1415
1416         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1417         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1418         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1419                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1420         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1421                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1422                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1423         }
1424 }
1425
1426 STATIC void
1427 xfs_free_bufhash(
1428         xfs_buftarg_t           *btp)
1429 {
1430         kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1431         btp->bt_hash = NULL;
1432 }
1433
1434 /*
1435  *      buftarg list for delwrite queue processing
1436  */
1437 static LIST_HEAD(xfs_buftarg_list);
1438 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1439
1440 STATIC void
1441 xfs_register_buftarg(
1442         xfs_buftarg_t           *btp)
1443 {
1444         spin_lock(&xfs_buftarg_lock);
1445         list_add(&btp->bt_list, &xfs_buftarg_list);
1446         spin_unlock(&xfs_buftarg_lock);
1447 }
1448
1449 STATIC void
1450 xfs_unregister_buftarg(
1451         xfs_buftarg_t           *btp)
1452 {
1453         spin_lock(&xfs_buftarg_lock);
1454         list_del(&btp->bt_list);
1455         spin_unlock(&xfs_buftarg_lock);
1456 }
1457
1458 void
1459 xfs_free_buftarg(
1460         xfs_buftarg_t           *btp,
1461         int                     external)
1462 {
1463         xfs_flush_buftarg(btp, 1);
1464         xfs_blkdev_issue_flush(btp);
1465         if (external)
1466                 xfs_blkdev_put(btp->bt_bdev);
1467         xfs_free_bufhash(btp);
1468         iput(btp->bt_mapping->host);
1469
1470         /* Unregister the buftarg first so that we don't get a
1471          * wakeup finding a non-existent task
1472          */
1473         xfs_unregister_buftarg(btp);
1474         kthread_stop(btp->bt_task);
1475
1476         kmem_free(btp, sizeof(*btp));
1477 }
1478
1479 STATIC int
1480 xfs_setsize_buftarg_flags(
1481         xfs_buftarg_t           *btp,
1482         unsigned int            blocksize,
1483         unsigned int            sectorsize,
1484         int                     verbose)
1485 {
1486         btp->bt_bsize = blocksize;
1487         btp->bt_sshift = ffs(sectorsize) - 1;
1488         btp->bt_smask = sectorsize - 1;
1489
1490         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1491                 printk(KERN_WARNING
1492                         "XFS: Cannot set_blocksize to %u on device %s\n",
1493                         sectorsize, XFS_BUFTARG_NAME(btp));
1494                 return EINVAL;
1495         }
1496
1497         if (verbose &&
1498             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1499                 printk(KERN_WARNING
1500                         "XFS: %u byte sectors in use on device %s.  "
1501                         "This is suboptimal; %u or greater is ideal.\n",
1502                         sectorsize, XFS_BUFTARG_NAME(btp),
1503                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1504         }
1505
1506         return 0;
1507 }
1508
1509 /*
1510  *      When allocating the initial buffer target we have not yet
1511  *      read in the superblock, so don't know what sized sectors
1512  *      are being used is at this early stage.  Play safe.
1513  */
1514 STATIC int
1515 xfs_setsize_buftarg_early(
1516         xfs_buftarg_t           *btp,
1517         struct block_device     *bdev)
1518 {
1519         return xfs_setsize_buftarg_flags(btp,
1520                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1521 }
1522
1523 int
1524 xfs_setsize_buftarg(
1525         xfs_buftarg_t           *btp,
1526         unsigned int            blocksize,
1527         unsigned int            sectorsize)
1528 {
1529         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1530 }
1531
1532 STATIC int
1533 xfs_mapping_buftarg(
1534         xfs_buftarg_t           *btp,
1535         struct block_device     *bdev)
1536 {
1537         struct backing_dev_info *bdi;
1538         struct inode            *inode;
1539         struct address_space    *mapping;
1540         static const struct address_space_operations mapping_aops = {
1541                 .sync_page = block_sync_page,
1542                 .migratepage = fail_migrate_page,
1543         };
1544
1545         inode = new_inode(bdev->bd_inode->i_sb);
1546         if (!inode) {
1547                 printk(KERN_WARNING
1548                         "XFS: Cannot allocate mapping inode for device %s\n",
1549                         XFS_BUFTARG_NAME(btp));
1550                 return ENOMEM;
1551         }
1552         inode->i_mode = S_IFBLK;
1553         inode->i_bdev = bdev;
1554         inode->i_rdev = bdev->bd_dev;
1555         bdi = blk_get_backing_dev_info(bdev);
1556         if (!bdi)
1557                 bdi = &default_backing_dev_info;
1558         mapping = &inode->i_data;
1559         mapping->a_ops = &mapping_aops;
1560         mapping->backing_dev_info = bdi;
1561         mapping_set_gfp_mask(mapping, GFP_NOFS);
1562         btp->bt_mapping = mapping;
1563         return 0;
1564 }
1565
1566 STATIC int
1567 xfs_alloc_delwrite_queue(
1568         xfs_buftarg_t           *btp)
1569 {
1570         int     error = 0;
1571
1572         INIT_LIST_HEAD(&btp->bt_list);
1573         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1574         spinlock_init(&btp->bt_delwrite_lock, "delwri_lock");
1575         btp->bt_flags = 0;
1576         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1577         if (IS_ERR(btp->bt_task)) {
1578                 error = PTR_ERR(btp->bt_task);
1579                 goto out_error;
1580         }
1581         xfs_register_buftarg(btp);
1582 out_error:
1583         return error;
1584 }
1585
1586 xfs_buftarg_t *
1587 xfs_alloc_buftarg(
1588         struct block_device     *bdev,
1589         int                     external)
1590 {
1591         xfs_buftarg_t           *btp;
1592
1593         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1594
1595         btp->bt_dev =  bdev->bd_dev;
1596         btp->bt_bdev = bdev;
1597         if (xfs_setsize_buftarg_early(btp, bdev))
1598                 goto error;
1599         if (xfs_mapping_buftarg(btp, bdev))
1600                 goto error;
1601         if (xfs_alloc_delwrite_queue(btp))
1602                 goto error;
1603         xfs_alloc_bufhash(btp, external);
1604         return btp;
1605
1606 error:
1607         kmem_free(btp, sizeof(*btp));
1608         return NULL;
1609 }
1610
1611
1612 /*
1613  *      Delayed write buffer handling
1614  */
1615 STATIC void
1616 xfs_buf_delwri_queue(
1617         xfs_buf_t               *bp,
1618         int                     unlock)
1619 {
1620         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1621         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1622
1623         XB_TRACE(bp, "delwri_q", (long)unlock);
1624         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1625
1626         spin_lock(dwlk);
1627         /* If already in the queue, dequeue and place at tail */
1628         if (!list_empty(&bp->b_list)) {
1629                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1630                 if (unlock)
1631                         atomic_dec(&bp->b_hold);
1632                 list_del(&bp->b_list);
1633         }
1634
1635         bp->b_flags |= _XBF_DELWRI_Q;
1636         list_add_tail(&bp->b_list, dwq);
1637         bp->b_queuetime = jiffies;
1638         spin_unlock(dwlk);
1639
1640         if (unlock)
1641                 xfs_buf_unlock(bp);
1642 }
1643
1644 void
1645 xfs_buf_delwri_dequeue(
1646         xfs_buf_t               *bp)
1647 {
1648         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1649         int                     dequeued = 0;
1650
1651         spin_lock(dwlk);
1652         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1653                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1654                 list_del_init(&bp->b_list);
1655                 dequeued = 1;
1656         }
1657         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1658         spin_unlock(dwlk);
1659
1660         if (dequeued)
1661                 xfs_buf_rele(bp);
1662
1663         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1664 }
1665
1666 STATIC void
1667 xfs_buf_runall_queues(
1668         struct workqueue_struct *queue)
1669 {
1670         flush_workqueue(queue);
1671 }
1672
1673 STATIC int
1674 xfsbufd_wakeup(
1675         int                     priority,
1676         gfp_t                   mask)
1677 {
1678         xfs_buftarg_t           *btp;
1679
1680         spin_lock(&xfs_buftarg_lock);
1681         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1682                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1683                         continue;
1684                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1685                 wake_up_process(btp->bt_task);
1686         }
1687         spin_unlock(&xfs_buftarg_lock);
1688         return 0;
1689 }
1690
1691 /*
1692  * Move as many buffers as specified to the supplied list
1693  * idicating if we skipped any buffers to prevent deadlocks.
1694  */
1695 STATIC int
1696 xfs_buf_delwri_split(
1697         xfs_buftarg_t   *target,
1698         struct list_head *list,
1699         unsigned long   age)
1700 {
1701         xfs_buf_t       *bp, *n;
1702         struct list_head *dwq = &target->bt_delwrite_queue;
1703         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1704         int             skipped = 0;
1705         int             force;
1706
1707         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1708         INIT_LIST_HEAD(list);
1709         spin_lock(dwlk);
1710         list_for_each_entry_safe(bp, n, dwq, b_list) {
1711                 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1712                 ASSERT(bp->b_flags & XBF_DELWRI);
1713
1714                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1715                         if (!force &&
1716                             time_before(jiffies, bp->b_queuetime + age)) {
1717                                 xfs_buf_unlock(bp);
1718                                 break;
1719                         }
1720
1721                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1722                                          _XBF_RUN_QUEUES);
1723                         bp->b_flags |= XBF_WRITE;
1724                         list_move_tail(&bp->b_list, list);
1725                 } else
1726                         skipped++;
1727         }
1728         spin_unlock(dwlk);
1729
1730         return skipped;
1731
1732 }
1733
1734 STATIC int
1735 xfsbufd(
1736         void            *data)
1737 {
1738         struct list_head tmp;
1739         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1740         int             count;
1741         xfs_buf_t       *bp;
1742
1743         current->flags |= PF_MEMALLOC;
1744
1745         set_freezable();
1746
1747         do {
1748                 if (unlikely(freezing(current))) {
1749                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1750                         refrigerator();
1751                 } else {
1752                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1753                 }
1754
1755                 schedule_timeout_interruptible(
1756                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1757
1758                 xfs_buf_delwri_split(target, &tmp,
1759                                 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1760
1761                 count = 0;
1762                 while (!list_empty(&tmp)) {
1763                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1764                         ASSERT(target == bp->b_target);
1765
1766                         list_del_init(&bp->b_list);
1767                         xfs_buf_iostrategy(bp);
1768                         count++;
1769                 }
1770
1771                 if (as_list_len > 0)
1772                         purge_addresses();
1773                 if (count)
1774                         blk_run_address_space(target->bt_mapping);
1775
1776         } while (!kthread_should_stop());
1777
1778         return 0;
1779 }
1780
1781 /*
1782  *      Go through all incore buffers, and release buffers if they belong to
1783  *      the given device. This is used in filesystem error handling to
1784  *      preserve the consistency of its metadata.
1785  */
1786 int
1787 xfs_flush_buftarg(
1788         xfs_buftarg_t   *target,
1789         int             wait)
1790 {
1791         struct list_head tmp;
1792         xfs_buf_t       *bp, *n;
1793         int             pincount = 0;
1794
1795         xfs_buf_runall_queues(xfsdatad_workqueue);
1796         xfs_buf_runall_queues(xfslogd_workqueue);
1797
1798         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1799         pincount = xfs_buf_delwri_split(target, &tmp, 0);
1800
1801         /*
1802          * Dropped the delayed write list lock, now walk the temporary list
1803          */
1804         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1805                 ASSERT(target == bp->b_target);
1806                 if (wait)
1807                         bp->b_flags &= ~XBF_ASYNC;
1808                 else
1809                         list_del_init(&bp->b_list);
1810
1811                 xfs_buf_iostrategy(bp);
1812         }
1813
1814         if (wait)
1815                 blk_run_address_space(target->bt_mapping);
1816
1817         /*
1818          * Remaining list items must be flushed before returning
1819          */
1820         while (!list_empty(&tmp)) {
1821                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1822
1823                 list_del_init(&bp->b_list);
1824                 xfs_iowait(bp);
1825                 xfs_buf_relse(bp);
1826         }
1827
1828         return pincount;
1829 }
1830
1831 int __init
1832 xfs_buf_init(void)
1833 {
1834 #ifdef XFS_BUF_TRACE
1835         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1836 #endif
1837
1838         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1839                                                 KM_ZONE_HWALIGN, NULL);
1840         if (!xfs_buf_zone)
1841                 goto out_free_trace_buf;
1842
1843         xfslogd_workqueue = create_workqueue("xfslogd");
1844         if (!xfslogd_workqueue)
1845                 goto out_free_buf_zone;
1846
1847         xfsdatad_workqueue = create_workqueue("xfsdatad");
1848         if (!xfsdatad_workqueue)
1849                 goto out_destroy_xfslogd_workqueue;
1850
1851         register_shrinker(&xfs_buf_shake);
1852         return 0;
1853
1854  out_destroy_xfslogd_workqueue:
1855         destroy_workqueue(xfslogd_workqueue);
1856  out_free_buf_zone:
1857         kmem_zone_destroy(xfs_buf_zone);
1858  out_free_trace_buf:
1859 #ifdef XFS_BUF_TRACE
1860         ktrace_free(xfs_buf_trace_buf);
1861 #endif
1862         return -ENOMEM;
1863 }
1864
1865 void
1866 xfs_buf_terminate(void)
1867 {
1868         unregister_shrinker(&xfs_buf_shake);
1869         destroy_workqueue(xfsdatad_workqueue);
1870         destroy_workqueue(xfslogd_workqueue);
1871         kmem_zone_destroy(xfs_buf_zone);
1872 #ifdef XFS_BUF_TRACE
1873         ktrace_free(xfs_buf_trace_buf);
1874 #endif
1875 }
1876
1877 #ifdef CONFIG_KDB_MODULES
1878 struct list_head *
1879 xfs_get_buftarg_list(void)
1880 {
1881         return &xfs_buftarg_list;
1882 }
1883 #endif