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,
  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_completion(&bp->b_iowait);
 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         }
 315 }
 316
 317 /*
 318  *      Releases the specified buffer.
 319  *
 320  *      The modification state of any associated pages is left unchanged.
 321  *      The buffer most not be on any hash - use xfs_buf_rele instead for
 322  *      hashed and refcounted buffers
 323  */
 324 void
 325 xfs_buf_free(
 326         xfs_buf_t               *bp)
 327 {
 328         XB_TRACE(bp, "free", 0);
 329
 330         ASSERT(list_empty(&bp->b_hash_list));
 331
 332         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
 333                 uint            i;
 334
 335                 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
 336                         free_address(bp->b_addr - bp->b_offset);
 337
 338                 for (i = 0; i < bp->b_page_count; i++) {
 339                         struct page     *page = bp->b_pages[i];
 340
 341                         if (bp->b_flags & _XBF_PAGE_CACHE)
 342                                 ASSERT(!PagePrivate(page));
 343                         page_cache_release(page);
 344                 }
 345                 _xfs_buf_free_pages(bp);
 346         }
 347
 348         xfs_buf_deallocate(bp);
 349 }
 350
 351 /*
 352  *      Finds all pages for buffer in question and builds it's page list.
 353  */
 354 STATIC int
 355 _xfs_buf_lookup_pages(
 356         xfs_buf_t               *bp,
 357         uint                    flags)
 358 {
 359         struct address_space    *mapping = bp->b_target->bt_mapping;
 360         size_t                  blocksize = bp->b_target->bt_bsize;
 361         size_t                  size = bp->b_count_desired;
 362         size_t                  nbytes, offset;
 363         gfp_t                   gfp_mask = xb_to_gfp(flags);
 364         unsigned short          page_count, i;
 365         pgoff_t                 first;
 366         xfs_off_t               end;
 367         int                     error;
 368
 369         end = bp->b_file_offset + bp->b_buffer_length;
 370         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 371
 372         error = _xfs_buf_get_pages(bp, page_count, flags);
 373         if (unlikely(error))
 374                 return error;
 375         bp->b_flags |= _XBF_PAGE_CACHE;
 376
 377         offset = bp->b_offset;
 378         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
 379
 380         for (i = 0; i < bp->b_page_count; i++) {
 381                 struct page     *page;
 382                 uint            retries = 0;
 383
 384               retry:
 385                 page = find_or_create_page(mapping, first + i, gfp_mask);
 386                 if (unlikely(page == NULL)) {
 387                         if (flags & XBF_READ_AHEAD) {
 388                                 bp->b_page_count = i;
 389                                 for (i = 0; i < bp->b_page_count; i++)
 390                                         unlock_page(bp->b_pages[i]);
 391                                 return -ENOMEM;
 392                         }
 393
 394                         /*
 395                          * This could deadlock.
 396                          *
 397                          * But until all the XFS lowlevel code is revamped to
 398                          * handle buffer allocation failures we can't do much.
 399                          */
 400                         if (!(++retries % 100))
 401                                 printk(KERN_ERR
 402                                         "XFS: possible memory allocation "
 403                                         "deadlock in %s (mode:0x%x)\n",
 404                                         __func__, gfp_mask);
 405
 406                         XFS_STATS_INC(xb_page_retries);
 407                         xfsbufd_wakeup(0, gfp_mask);
 408                         congestion_wait(WRITE, HZ/50);
 409                         goto retry;
 410                 }
 411
 412                 XFS_STATS_INC(xb_page_found);
 413
 414                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
 415                 size -= nbytes;
 416
 417                 ASSERT(!PagePrivate(page));
 418                 if (!PageUptodate(page)) {
 419                         page_count--;
 420                         if (blocksize >= PAGE_CACHE_SIZE) {
 421                                 if (flags & XBF_READ)
 422                                         bp->b_flags |= _XBF_PAGE_LOCKED;
 423                         } else if (!PagePrivate(page)) {
 424                                 if (test_page_region(page, offset, nbytes))
 425                                         page_count++;
 426                         }
 427                 }
 428
 429                 bp->b_pages[i] = page;
 430                 offset = 0;
 431         }
 432
 433         if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
 434                 for (i = 0; i < bp->b_page_count; i++)
 435                         unlock_page(bp->b_pages[i]);
 436         }
 437
 438         if (page_count == bp->b_page_count)
 439                 bp->b_flags |= XBF_DONE;
 440
 441         XB_TRACE(bp, "lookup_pages", (long)page_count);
 442         return error;
 443 }
 444
 445 /*
 446  *      Map buffer into kernel address-space if nessecary.
 447  */
 448 STATIC int
 449 _xfs_buf_map_pages(
 450         xfs_buf_t               *bp,
 451         uint                    flags)
 452 {
 453         /* A single page buffer is always mappable */
 454         if (bp->b_page_count == 1) {
 455                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 456                 bp->b_flags |= XBF_MAPPED;
 457         } else if (flags & XBF_MAPPED) {
 458                 if (as_list_len > 64)
 459                         purge_addresses();
 460                 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
 461                                         VM_MAP, PAGE_KERNEL);
 462                 if (unlikely(bp->b_addr == NULL))
 463                         return -ENOMEM;
 464                 bp->b_addr += bp->b_offset;
 465                 bp->b_flags |= XBF_MAPPED;
 466         }
 467
 468         return 0;
 469 }
 470
 471 /*
 472  *      Finding and Reading Buffers
 473  */
 474
 475 /*
 476  *      Look up, and creates if absent, a lockable buffer for
 477  *      a given range of an inode.  The buffer is returned
 478  *      locked.  If other overlapping buffers exist, they are
 479  *      released before the new buffer is created and locked,
 480  *      which may imply that this call will block until those buffers
 481  *      are unlocked.  No I/O is implied by this call.
 482  */
 483 xfs_buf_t *
 484 _xfs_buf_find(
 485         xfs_buftarg_t           *btp,   /* block device target          */
 486         xfs_off_t               ioff,   /* starting offset of range     */
 487         size_t                  isize,  /* length of range              */
 488         xfs_buf_flags_t         flags,
 489         xfs_buf_t               *new_bp)
 490 {
 491         xfs_off_t               range_base;
 492         size_t                  range_length;
 493         xfs_bufhash_t           *hash;
 494         xfs_buf_t               *bp, *n;
 495
 496         range_base = (ioff << BBSHIFT);
 497         range_length = (isize << BBSHIFT);
 498
 499         /* Check for IOs smaller than the sector size / not sector aligned */
 500         ASSERT(!(range_length < (1 << btp->bt_sshift)));
 501         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
 502
 503         hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
 504
 505         spin_lock(&hash->bh_lock);
 506
 507         list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
 508                 ASSERT(btp == bp->b_target);
 509                 if (bp->b_file_offset == range_base &&
 510                     bp->b_buffer_length == range_length) {
 511                         /*
 512                          * If we look at something, bring it to the
 513                          * front of the list for next time.
 514                          */
 515                         atomic_inc(&bp->b_hold);
 516                         list_move(&bp->b_hash_list, &hash->bh_list);
 517                         goto found;
 518                 }
 519         }
 520
 521         /* No match found */
 522         if (new_bp) {
 523                 _xfs_buf_initialize(new_bp, btp, range_base,
 524                                 range_length, flags);
 525                 new_bp->b_hash = hash;
 526                 list_add(&new_bp->b_hash_list, &hash->bh_list);
 527         } else {
 528                 XFS_STATS_INC(xb_miss_locked);
 529         }
 530
 531         spin_unlock(&hash->bh_lock);
 532         return new_bp;
 533
 534 found:
 535         spin_unlock(&hash->bh_lock);
 536
 537         /* Attempt to get the semaphore without sleeping,
 538          * if this does not work then we need to drop the
 539          * spinlock and do a hard attempt on the semaphore.
 540          */
 541         if (down_trylock(&bp->b_sema)) {
 542                 if (!(flags & XBF_TRYLOCK)) {
 543                         /* wait for buffer ownership */
 544                         XB_TRACE(bp, "get_lock", 0);
 545                         xfs_buf_lock(bp);
 546                         XFS_STATS_INC(xb_get_locked_waited);
 547                 } else {
 548                         /* We asked for a trylock and failed, no need
 549                          * to look at file offset and length here, we
 550                          * know that this buffer at least overlaps our
 551                          * buffer and is locked, therefore our buffer
 552                          * either does not exist, or is this buffer.
 553                          */
 554                         xfs_buf_rele(bp);
 555                         XFS_STATS_INC(xb_busy_locked);
 556                         return NULL;
 557                 }
 558         } else {
 559                 /* trylock worked */
 560                 XB_SET_OWNER(bp);
 561         }
 562
 563         if (bp->b_flags & XBF_STALE) {
 564                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 565                 bp->b_flags &= XBF_MAPPED;
 566         }
 567         XB_TRACE(bp, "got_lock", 0);
 568         XFS_STATS_INC(xb_get_locked);
 569         return bp;
 570 }
 571
 572 /*
 573  *      Assembles a buffer covering the specified range.
 574  *      Storage in memory for all portions of the buffer will be allocated,
 575  *      although backing storage may not be.
 576  */
 577 xfs_buf_t *
 578 xfs_buf_get_flags(
 579         xfs_buftarg_t           *target,/* target for buffer            */
 580         xfs_off_t               ioff,   /* starting offset of range     */
 581         size_t                  isize,  /* length of range              */
 582         xfs_buf_flags_t         flags)
 583 {
 584         xfs_buf_t               *bp, *new_bp;
 585         int                     error = 0, i;
 586
 587         new_bp = xfs_buf_allocate(flags);
 588         if (unlikely(!new_bp))
 589                 return NULL;
 590
 591         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 592         if (bp == new_bp) {
 593                 error = _xfs_buf_lookup_pages(bp, flags);
 594                 if (error)
 595                         goto no_buffer;
 596         } else {
 597                 xfs_buf_deallocate(new_bp);
 598                 if (unlikely(bp == NULL))
 599                         return NULL;
 600         }
 601
 602         for (i = 0; i < bp->b_page_count; i++)
 603                 mark_page_accessed(bp->b_pages[i]);
 604
 605         if (!(bp->b_flags & XBF_MAPPED)) {
 606                 error = _xfs_buf_map_pages(bp, flags);
 607                 if (unlikely(error)) {
 608                         printk(KERN_WARNING "%s: failed to map pages\n",
 609                                         __func__);
 610                         goto no_buffer;
 611                 }
 612         }
 613
 614         XFS_STATS_INC(xb_get);
 615
 616         /*
 617          * Always fill in the block number now, the mapped cases can do
 618          * their own overlay of this later.
 619          */
 620         bp->b_bn = ioff;
 621         bp->b_count_desired = bp->b_buffer_length;
 622
 623         XB_TRACE(bp, "get", (unsigned long)flags);
 624         return bp;
 625
 626  no_buffer:
 627         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 628                 xfs_buf_unlock(bp);
 629         xfs_buf_rele(bp);
 630         return NULL;
 631 }
 632
 633 STATIC int
 634 _xfs_buf_read(
 635         xfs_buf_t               *bp,
 636         xfs_buf_flags_t         flags)
 637 {
 638         int                     status;
 639
 640         XB_TRACE(bp, "_xfs_buf_read", (unsigned long)flags);
 641
 642         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
 643         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
 644
 645         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
 646                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 647         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
 648                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 649
 650         status = xfs_buf_iorequest(bp);
 651         if (!status && !(flags & XBF_ASYNC))
 652                 status = xfs_buf_iowait(bp);
 653         return status;
 654 }
 655
 656 xfs_buf_t *
 657 xfs_buf_read_flags(
 658         xfs_buftarg_t           *target,
 659         xfs_off_t               ioff,
 660         size_t                  isize,
 661         xfs_buf_flags_t         flags)
 662 {
 663         xfs_buf_t               *bp;
 664
 665         flags |= XBF_READ;
 666
 667         bp = xfs_buf_get_flags(target, ioff, isize, flags);
 668         if (bp) {
 669                 if (!XFS_BUF_ISDONE(bp)) {
 670                         XB_TRACE(bp, "read", (unsigned long)flags);
 671                         XFS_STATS_INC(xb_get_read);
 672                         _xfs_buf_read(bp, flags);
 673                 } else if (flags & XBF_ASYNC) {
 674                         XB_TRACE(bp, "read_async", (unsigned long)flags);
 675                         /*
 676                          * Read ahead call which is already satisfied,
 677                          * drop the buffer
 678                          */
 679                         goto no_buffer;
 680                 } else {
 681                         XB_TRACE(bp, "read_done", (unsigned long)flags);
 682                         /* We do not want read in the flags */
 683                         bp->b_flags &= ~XBF_READ;
 684                 }
 685         }
 686
 687         return bp;
 688
 689  no_buffer:
 690         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 691                 xfs_buf_unlock(bp);
 692         xfs_buf_rele(bp);
 693         return NULL;
 694 }
 695
 696 /*
 697  *      If we are not low on memory then do the readahead in a deadlock
 698  *      safe manner.
 699  */
 700 void
 701 xfs_buf_readahead(
 702         xfs_buftarg_t           *target,
 703         xfs_off_t               ioff,
 704         size_t                  isize,
 705         xfs_buf_flags_t         flags)
 706 {
 707         struct backing_dev_info *bdi;
 708
 709         bdi = target->bt_mapping->backing_dev_info;
 710         if (bdi_read_congested(bdi))
 711                 return;
 712
 713         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 714         xfs_buf_read_flags(target, ioff, isize, flags);
 715 }
 716
 717 xfs_buf_t *
 718 xfs_buf_get_empty(
 719         size_t                  len,
 720         xfs_buftarg_t           *target)
 721 {
 722         xfs_buf_t               *bp;
 723
 724         bp = xfs_buf_allocate(0);
 725         if (bp)
 726                 _xfs_buf_initialize(bp, target, 0, len, 0);
 727         return bp;
 728 }
 729
 730 static inline struct page *
 731 mem_to_page(
 732         void                    *addr)
 733 {
 734         if ((!is_vmalloc_addr(addr))) {
 735                 return virt_to_page(addr);
 736         } else {
 737                 return vmalloc_to_page(addr);
 738         }
 739 }
 740
 741 int
 742 xfs_buf_associate_memory(
 743         xfs_buf_t               *bp,
 744         void                    *mem,
 745         size_t                  len)
 746 {
 747         int                     rval;
 748         int                     i = 0;
 749         unsigned long           pageaddr;
 750         unsigned long           offset;
 751         size_t                  buflen;
 752         int                     page_count;
 753
 754         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
 755         offset = (unsigned long)mem - pageaddr;
 756         buflen = PAGE_CACHE_ALIGN(len + offset);
 757         page_count = buflen >> PAGE_CACHE_SHIFT;
 758
 759         /* Free any previous set of page pointers */
 760         if (bp->b_pages)
 761                 _xfs_buf_free_pages(bp);
 762
 763         bp->b_pages = NULL;
 764         bp->b_addr = mem;
 765
 766         rval = _xfs_buf_get_pages(bp, page_count, 0);
 767         if (rval)
 768                 return rval;
 769
 770         bp->b_offset = offset;
 771
 772         for (i = 0; i < bp->b_page_count; i++) {
 773                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 774                 pageaddr += PAGE_CACHE_SIZE;
 775         }
 776
 777         bp->b_count_desired = len;
 778         bp->b_buffer_length = buflen;
 779         bp->b_flags |= XBF_MAPPED;
 780         bp->b_flags &= ~_XBF_PAGE_LOCKED;
 781
 782         return 0;
 783 }
 784
 785 xfs_buf_t *
 786 xfs_buf_get_noaddr(
 787         size_t                  len,
 788         xfs_buftarg_t           *target)
 789 {
 790         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 791         int                     error, i;
 792         xfs_buf_t               *bp;
 793
 794         bp = xfs_buf_allocate(0);
 795         if (unlikely(bp == NULL))
 796                 goto fail;
 797         _xfs_buf_initialize(bp, target, 0, len, 0);
 798
 799         error = _xfs_buf_get_pages(bp, page_count, 0);
 800         if (error)
 801                 goto fail_free_buf;
 802
 803         for (i = 0; i < page_count; i++) {
 804                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
 805                 if (!bp->b_pages[i])
 806                         goto fail_free_mem;
 807         }
 808         bp->b_flags |= _XBF_PAGES;
 809
 810         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 811         if (unlikely(error)) {
 812                 printk(KERN_WARNING "%s: failed to map pages\n",
 813                                 __func__);
 814                 goto fail_free_mem;
 815         }
 816
 817         xfs_buf_unlock(bp);
 818
 819         XB_TRACE(bp, "no_daddr", len);
 820         return bp;
 821
 822  fail_free_mem:
 823         while (--i >= 0)
 824                 __free_page(bp->b_pages[i]);
 825         _xfs_buf_free_pages(bp);
 826  fail_free_buf:
 827         xfs_buf_deallocate(bp);
 828  fail:
 829         return NULL;
 830 }
 831
 832 /*
 833  *      Increment reference count on buffer, to hold the buffer concurrently
 834  *      with another thread which may release (free) the buffer asynchronously.
 835  *      Must hold the buffer already to call this function.
 836  */
 837 void
 838 xfs_buf_hold(
 839         xfs_buf_t               *bp)
 840 {
 841         atomic_inc(&bp->b_hold);
 842         XB_TRACE(bp, "hold", 0);
 843 }
 844
 845 /*
 846  *      Releases a hold on the specified buffer.  If the
 847  *      the hold count is 1, calls xfs_buf_free.
 848  */
 849 void
 850 xfs_buf_rele(
 851         xfs_buf_t               *bp)
 852 {
 853         xfs_bufhash_t           *hash = bp->b_hash;
 854
 855         XB_TRACE(bp, "rele", bp->b_relse);
 856
 857         if (unlikely(!hash)) {
 858                 ASSERT(!bp->b_relse);
 859                 if (atomic_dec_and_test(&bp->b_hold))
 860                         xfs_buf_free(bp);
 861                 return;
 862         }
 863
 864         ASSERT(atomic_read(&bp->b_hold) > 0);
 865         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
 866                 if (bp->b_relse) {
 867                         atomic_inc(&bp->b_hold);
 868                         spin_unlock(&hash->bh_lock);
 869                         (*(bp->b_relse)) (bp);
 870                 } else if (bp->b_flags & XBF_FS_MANAGED) {
 871                         spin_unlock(&hash->bh_lock);
 872                 } else {
 873                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 874                         list_del_init(&bp->b_hash_list);
 875                         spin_unlock(&hash->bh_lock);
 876                         xfs_buf_free(bp);
 877                 }
 878         }
 879 }
 880
 881
 882 /*
 883  *      Mutual exclusion on buffers.  Locking model:
 884  *
 885  *      Buffers associated with inodes for which buffer locking
 886  *      is not enabled are not protected by semaphores, and are
 887  *      assumed to be exclusively owned by the caller.  There is a
 888  *      spinlock in the buffer, used by the caller when concurrent
 889  *      access is possible.
 890  */
 891
 892 /*
 893  *      Locks a buffer object, if it is not already locked.
 894  *      Note that this in no way locks the underlying pages, so it is only
 895  *      useful for synchronizing concurrent use of buffer objects, not for
 896  *      synchronizing independent access to the underlying pages.
 897  */
 898 int
 899 xfs_buf_cond_lock(
 900         xfs_buf_t               *bp)
 901 {
 902         int                     locked;
 903
 904         locked = down_trylock(&bp->b_sema) == 0;
 905         if (locked) {
 906                 XB_SET_OWNER(bp);
 907         }
 908         XB_TRACE(bp, "cond_lock", (long)locked);
 909         return locked ? 0 : -EBUSY;
 910 }
 911
 912 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
 913 int
 914 xfs_buf_lock_value(
 915         xfs_buf_t               *bp)
 916 {
 917         return bp->b_sema.count;
 918 }
 919 #endif
 920
 921 /*
 922  *      Locks a buffer object.
 923  *      Note that this in no way locks the underlying pages, so it is only
 924  *      useful for synchronizing concurrent use of buffer objects, not for
 925  *      synchronizing independent access to the underlying pages.
 926  */
 927 void
 928 xfs_buf_lock(
 929         xfs_buf_t               *bp)
 930 {
 931         XB_TRACE(bp, "lock", 0);
 932         if (atomic_read(&bp->b_io_remaining))
 933                 blk_run_address_space(bp->b_target->bt_mapping);
 934         down(&bp->b_sema);
 935         XB_SET_OWNER(bp);
 936         XB_TRACE(bp, "locked", 0);
 937 }
 938
 939 /*
 940  *      Releases the lock on the buffer object.
 941  *      If the buffer is marked delwri but is not queued, do so before we
 942  *      unlock the buffer as we need to set flags correctly.  We also need to
 943  *      take a reference for the delwri queue because the unlocker is going to
 944  *      drop their's and they don't know we just queued it.
 945  */
 946 void
 947 xfs_buf_unlock(
 948         xfs_buf_t               *bp)
 949 {
 950         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 951                 atomic_inc(&bp->b_hold);
 952                 bp->b_flags |= XBF_ASYNC;
 953                 xfs_buf_delwri_queue(bp, 0);
 954         }
 955
 956         XB_CLEAR_OWNER(bp);
 957         up(&bp->b_sema);
 958         XB_TRACE(bp, "unlock", 0);
 959 }
 960
 961
 962 /*
 963  *      Pinning Buffer Storage in Memory
 964  *      Ensure that no attempt to force a buffer to disk will succeed.
 965  */
 966 void
 967 xfs_buf_pin(
 968         xfs_buf_t               *bp)
 969 {
 970         atomic_inc(&bp->b_pin_count);
 971         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
 972 }
 973
 974 void
 975 xfs_buf_unpin(
 976         xfs_buf_t               *bp)
 977 {
 978         if (atomic_dec_and_test(&bp->b_pin_count))
 979                 wake_up_all(&bp->b_waiters);
 980         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
 981 }
 982
 983 int
 984 xfs_buf_ispin(
 985         xfs_buf_t               *bp)
 986 {
 987         return atomic_read(&bp->b_pin_count);
 988 }
 989
 990 STATIC void
 991 xfs_buf_wait_unpin(
 992         xfs_buf_t               *bp)
 993 {
 994         DECLARE_WAITQUEUE       (wait, current);
 995
 996         if (atomic_read(&bp->b_pin_count) == 0)
 997                 return;
 998
 999         add_wait_queue(&bp->b_waiters, &wait);
1000         for (;;) {
1001                 set_current_state(TASK_UNINTERRUPTIBLE);
1002                 if (atomic_read(&bp->b_pin_count) == 0)
1003                         break;
1004                 if (atomic_read(&bp->b_io_remaining))
1005                         blk_run_address_space(bp->b_target->bt_mapping);
1006                 schedule();
1007         }
1008         remove_wait_queue(&bp->b_waiters, &wait);
1009         set_current_state(TASK_RUNNING);
1010 }
1011
1012 /*
1013  *      Buffer Utility Routines
1014  */
1015
1016 STATIC void
1017 xfs_buf_iodone_work(
1018         struct work_struct      *work)
1019 {
1020         xfs_buf_t               *bp =
1021                 container_of(work, xfs_buf_t, b_iodone_work);
1022
1023         /*
1024          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
1025          * ordered flag and reissue them.  Because we can't tell the higher
1026          * layers directly that they should not issue ordered I/O anymore, they
1027          * need to check if the _XFS_BARRIER_FAILED flag was set during I/O completion.
1028          */
1029         if ((bp->b_error == EOPNOTSUPP) &&
1030             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1031                 XB_TRACE(bp, "ordered_retry", bp->b_iodone);
1032                 bp->b_flags &= ~XBF_ORDERED;
1033                 bp->b_flags |= _XFS_BARRIER_FAILED;
1034                 xfs_buf_iorequest(bp);
1035         } else if (bp->b_iodone)
1036                 (*(bp->b_iodone))(bp);
1037         else if (bp->b_flags & XBF_ASYNC)
1038                 xfs_buf_relse(bp);
1039 }
1040
1041 void
1042 xfs_buf_ioend(
1043         xfs_buf_t               *bp,
1044         int                     schedule)
1045 {
1046         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1047         if (bp->b_error == 0)
1048                 bp->b_flags |= XBF_DONE;
1049
1050         XB_TRACE(bp, "iodone", bp->b_iodone);
1051
1052         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1053                 if (schedule) {
1054                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1055                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1056                 } else {
1057                         xfs_buf_iodone_work(&bp->b_iodone_work);
1058                 }
1059         } else {
1060                 complete(&bp->b_iowait);
1061         }
1062 }
1063
1064 void
1065 xfs_buf_ioerror(
1066         xfs_buf_t               *bp,
1067         int                     error)
1068 {
1069         ASSERT(error >= 0 && error <= 0xffff);
1070         bp->b_error = (unsigned short)error;
1071         XB_TRACE(bp, "ioerror", (unsigned long)error);
1072 }
1073
1074 int
1075 xfs_bawrite(
1076         void                    *mp,
1077         struct xfs_buf          *bp)
1078 {
1079         XB_TRACE(bp, "bawrite", 0);
1080
1081         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
1082
1083         xfs_buf_delwri_dequeue(bp);
1084
1085         bp->b_flags &= ~(XBF_READ | XBF_DELWRI | XBF_READ_AHEAD);
1086         bp->b_flags |= (XBF_WRITE | XBF_ASYNC | _XBF_RUN_QUEUES);
1087
1088         bp->b_mount = mp;
1089         bp->b_strat = xfs_bdstrat_cb;
1090         return xfs_bdstrat_cb(bp);
1091 }
1092
1093 void
1094 xfs_bdwrite(
1095         void                    *mp,
1096         struct xfs_buf          *bp)
1097 {
1098         XB_TRACE(bp, "bdwrite", 0);
1099
1100         bp->b_strat = xfs_bdstrat_cb;
1101         bp->b_mount = mp;
1102
1103         bp->b_flags &= ~XBF_READ;
1104         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1105
1106         xfs_buf_delwri_queue(bp, 1);
1107 }
1108
1109 STATIC_INLINE void
1110 _xfs_buf_ioend(
1111         xfs_buf_t               *bp,
1112         int                     schedule)
1113 {
1114         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1115                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1116                 xfs_buf_ioend(bp, schedule);
1117         }
1118 }
1119
1120 STATIC void
1121 xfs_buf_bio_end_io(
1122         struct bio              *bio,
1123         int                     error)
1124 {
1125         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1126         unsigned int            blocksize = bp->b_target->bt_bsize;
1127         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1128
1129         xfs_buf_ioerror(bp, -error);
1130
1131         do {
1132                 struct page     *page = bvec->bv_page;
1133
1134                 ASSERT(!PagePrivate(page));
1135                 if (unlikely(bp->b_error)) {
1136                         if (bp->b_flags & XBF_READ)
1137                                 ClearPageUptodate(page);
1138                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1139                         SetPageUptodate(page);
1140                 } else if (!PagePrivate(page) &&
1141                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1142                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1143                 }
1144
1145                 if (--bvec >= bio->bi_io_vec)
1146                         prefetchw(&bvec->bv_page->flags);
1147
1148                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1149                         unlock_page(page);
1150         } while (bvec >= bio->bi_io_vec);
1151
1152         _xfs_buf_ioend(bp, 1);
1153         bio_put(bio);
1154 }
1155
1156 STATIC void
1157 _xfs_buf_ioapply(
1158         xfs_buf_t               *bp)
1159 {
1160         int                     rw, map_i, total_nr_pages, nr_pages;
1161         struct bio              *bio;
1162         int                     offset = bp->b_offset;
1163         int                     size = bp->b_count_desired;
1164         sector_t                sector = bp->b_bn;
1165         unsigned int            blocksize = bp->b_target->bt_bsize;
1166
1167         total_nr_pages = bp->b_page_count;
1168         map_i = 0;
1169
1170         if (bp->b_flags & XBF_ORDERED) {
1171                 ASSERT(!(bp->b_flags & XBF_READ));
1172                 rw = WRITE_BARRIER;
1173         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1174                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1175                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1176                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1177         } else {
1178                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1179                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1180         }
1181
1182         /* Special code path for reading a sub page size buffer in --
1183          * we populate up the whole page, and hence the other metadata
1184          * in the same page.  This optimization is only valid when the
1185          * filesystem block size is not smaller than the page size.
1186          */
1187         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1188             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1189               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1190             (blocksize >= PAGE_CACHE_SIZE)) {
1191                 bio = bio_alloc(GFP_NOIO, 1);
1192
1193                 bio->bi_bdev = bp->b_target->bt_bdev;
1194                 bio->bi_sector = sector - (offset >> BBSHIFT);
1195                 bio->bi_end_io = xfs_buf_bio_end_io;
1196                 bio->bi_private = bp;
1197
1198                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1199                 size = 0;
1200
1201                 atomic_inc(&bp->b_io_remaining);
1202
1203                 goto submit_io;
1204         }
1205
1206 next_chunk:
1207         atomic_inc(&bp->b_io_remaining);
1208         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1209         if (nr_pages > total_nr_pages)
1210                 nr_pages = total_nr_pages;
1211
1212         bio = bio_alloc(GFP_NOIO, nr_pages);
1213         bio->bi_bdev = bp->b_target->bt_bdev;
1214         bio->bi_sector = sector;
1215         bio->bi_end_io = xfs_buf_bio_end_io;
1216         bio->bi_private = bp;
1217
1218         for (; size && nr_pages; nr_pages--, map_i++) {
1219                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1220
1221                 if (nbytes > size)
1222                         nbytes = size;
1223
1224                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1225                 if (rbytes < nbytes)
1226                         break;
1227
1228                 offset = 0;
1229                 sector += nbytes >> BBSHIFT;
1230                 size -= nbytes;
1231                 total_nr_pages--;
1232         }
1233
1234 submit_io:
1235         if (likely(bio->bi_size)) {
1236                 submit_bio(rw, bio);
1237                 if (size)
1238                         goto next_chunk;
1239         } else {
1240                 bio_put(bio);
1241                 xfs_buf_ioerror(bp, EIO);
1242         }
1243 }
1244
1245 int
1246 xfs_buf_iorequest(
1247         xfs_buf_t               *bp)
1248 {
1249         XB_TRACE(bp, "iorequest", 0);
1250
1251         if (bp->b_flags & XBF_DELWRI) {
1252                 xfs_buf_delwri_queue(bp, 1);
1253                 return 0;
1254         }
1255
1256         if (bp->b_flags & XBF_WRITE) {
1257                 xfs_buf_wait_unpin(bp);
1258         }
1259
1260         xfs_buf_hold(bp);
1261
1262         /* Set the count to 1 initially, this will stop an I/O
1263          * completion callout which happens before we have started
1264          * all the I/O from calling xfs_buf_ioend too early.
1265          */
1266         atomic_set(&bp->b_io_remaining, 1);
1267         _xfs_buf_ioapply(bp);
1268         _xfs_buf_ioend(bp, 0);
1269
1270         xfs_buf_rele(bp);
1271         return 0;
1272 }
1273
1274 /*
1275  *      Waits for I/O to complete on the buffer supplied.
1276  *      It returns immediately if no I/O is pending.
1277  *      It returns the I/O error code, if any, or 0 if there was no error.
1278  */
1279 int
1280 xfs_buf_iowait(
1281         xfs_buf_t               *bp)
1282 {
1283         XB_TRACE(bp, "iowait", 0);
1284         if (atomic_read(&bp->b_io_remaining))
1285                 blk_run_address_space(bp->b_target->bt_mapping);
1286         wait_for_completion(&bp->b_iowait);
1287         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1288         return bp->b_error;
1289 }
1290
1291 xfs_caddr_t
1292 xfs_buf_offset(
1293         xfs_buf_t               *bp,
1294         size_t                  offset)
1295 {
1296         struct page             *page;
1297
1298         if (bp->b_flags & XBF_MAPPED)
1299                 return XFS_BUF_PTR(bp) + offset;
1300
1301         offset += bp->b_offset;
1302         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1303         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1304 }
1305
1306 /*
1307  *      Move data into or out of a buffer.
1308  */
1309 void
1310 xfs_buf_iomove(
1311         xfs_buf_t               *bp,    /* buffer to process            */
1312         size_t                  boff,   /* starting buffer offset       */
1313         size_t                  bsize,  /* length to copy               */
1314         caddr_t                 data,   /* data address                 */
1315         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1316 {
1317         size_t                  bend, cpoff, csize;
1318         struct page             *page;
1319
1320         bend = boff + bsize;
1321         while (boff < bend) {
1322                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1323                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1324                 csize = min_t(size_t,
1325                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1326
1327                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1328
1329                 switch (mode) {
1330                 case XBRW_ZERO:
1331                         memset(page_address(page) + cpoff, 0, csize);
1332                         break;
1333                 case XBRW_READ:
1334                         memcpy(data, page_address(page) + cpoff, csize);
1335                         break;
1336                 case XBRW_WRITE:
1337                         memcpy(page_address(page) + cpoff, data, csize);
1338                 }
1339
1340                 boff += csize;
1341                 data += csize;
1342         }
1343 }
1344
1345 /*
1346  *      Handling of buffer targets (buftargs).
1347  */
1348
1349 /*
1350  *      Wait for any bufs with callbacks that have been submitted but
1351  *      have not yet returned... walk the hash list for the target.
1352  */
1353 void
1354 xfs_wait_buftarg(
1355         xfs_buftarg_t   *btp)
1356 {
1357         xfs_buf_t       *bp, *n;
1358         xfs_bufhash_t   *hash;
1359         uint            i;
1360
1361         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1362                 hash = &btp->bt_hash[i];
1363 again:
1364                 spin_lock(&hash->bh_lock);
1365                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1366                         ASSERT(btp == bp->b_target);
1367                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1368                                 spin_unlock(&hash->bh_lock);
1369                                 /*
1370                                  * Catch superblock reference count leaks
1371                                  * immediately
1372                                  */
1373                                 BUG_ON(bp->b_bn == 0);
1374                                 delay(100);
1375                                 goto again;
1376                         }
1377                 }
1378                 spin_unlock(&hash->bh_lock);
1379         }
1380 }
1381
1382 /*
1383  *      Allocate buffer hash table for a given target.
1384  *      For devices containing metadata (i.e. not the log/realtime devices)
1385  *      we need to allocate a much larger hash table.
1386  */
1387 STATIC void
1388 xfs_alloc_bufhash(
1389         xfs_buftarg_t           *btp,
1390         int                     external)
1391 {
1392         unsigned int            i;
1393
1394         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1395         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1396         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1397                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1398         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1399                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1400                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1401         }
1402 }
1403
1404 STATIC void
1405 xfs_free_bufhash(
1406         xfs_buftarg_t           *btp)
1407 {
1408         kmem_free(btp->bt_hash);
1409         btp->bt_hash = NULL;
1410 }
1411
1412 /*
1413  *      buftarg list for delwrite queue processing
1414  */
1415 static LIST_HEAD(xfs_buftarg_list);
1416 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1417
1418 STATIC void
1419 xfs_register_buftarg(
1420         xfs_buftarg_t           *btp)
1421 {
1422         spin_lock(&xfs_buftarg_lock);
1423         list_add(&btp->bt_list, &xfs_buftarg_list);
1424         spin_unlock(&xfs_buftarg_lock);
1425 }
1426
1427 STATIC void
1428 xfs_unregister_buftarg(
1429         xfs_buftarg_t           *btp)
1430 {
1431         spin_lock(&xfs_buftarg_lock);
1432         list_del(&btp->bt_list);
1433         spin_unlock(&xfs_buftarg_lock);
1434 }
1435
1436 void
1437 xfs_free_buftarg(
1438         xfs_buftarg_t           *btp)
1439 {
1440         xfs_flush_buftarg(btp, 1);
1441         xfs_blkdev_issue_flush(btp);
1442         xfs_free_bufhash(btp);
1443         iput(btp->bt_mapping->host);
1444
1445         /* Unregister the buftarg first so that we don't get a
1446          * wakeup finding a non-existent task
1447          */
1448         xfs_unregister_buftarg(btp);
1449         kthread_stop(btp->bt_task);
1450
1451         kmem_free(btp);
1452 }
1453
1454 STATIC int
1455 xfs_setsize_buftarg_flags(
1456         xfs_buftarg_t           *btp,
1457         unsigned int            blocksize,
1458         unsigned int            sectorsize,
1459         int                     verbose)
1460 {
1461         btp->bt_bsize = blocksize;
1462         btp->bt_sshift = ffs(sectorsize) - 1;
1463         btp->bt_smask = sectorsize - 1;
1464
1465         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1466                 printk(KERN_WARNING
1467                         "XFS: Cannot set_blocksize to %u on device %s\n",
1468                         sectorsize, XFS_BUFTARG_NAME(btp));
1469                 return EINVAL;
1470         }
1471
1472         if (verbose &&
1473             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1474                 printk(KERN_WARNING
1475                         "XFS: %u byte sectors in use on device %s.  "
1476                         "This is suboptimal; %u or greater is ideal.\n",
1477                         sectorsize, XFS_BUFTARG_NAME(btp),
1478                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1479         }
1480
1481         return 0;
1482 }
1483
1484 /*
1485  *      When allocating the initial buffer target we have not yet
1486  *      read in the superblock, so don't know what sized sectors
1487  *      are being used is at this early stage.  Play safe.
1488  */
1489 STATIC int
1490 xfs_setsize_buftarg_early(
1491         xfs_buftarg_t           *btp,
1492         struct block_device     *bdev)
1493 {
1494         return xfs_setsize_buftarg_flags(btp,
1495                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1496 }
1497
1498 int
1499 xfs_setsize_buftarg(
1500         xfs_buftarg_t           *btp,
1501         unsigned int            blocksize,
1502         unsigned int            sectorsize)
1503 {
1504         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1505 }
1506
1507 STATIC int
1508 xfs_mapping_buftarg(
1509         xfs_buftarg_t           *btp,
1510         struct block_device     *bdev)
1511 {
1512         struct backing_dev_info *bdi;
1513         struct inode            *inode;
1514         struct address_space    *mapping;
1515         static const struct address_space_operations mapping_aops = {
1516                 .sync_page = block_sync_page,
1517                 .migratepage = fail_migrate_page,
1518         };
1519
1520         inode = new_inode(bdev->bd_inode->i_sb);
1521         if (!inode) {
1522                 printk(KERN_WARNING
1523                         "XFS: Cannot allocate mapping inode for device %s\n",
1524                         XFS_BUFTARG_NAME(btp));
1525                 return ENOMEM;
1526         }
1527         inode->i_mode = S_IFBLK;
1528         inode->i_bdev = bdev;
1529         inode->i_rdev = bdev->bd_dev;
1530         bdi = blk_get_backing_dev_info(bdev);
1531         if (!bdi)
1532                 bdi = &default_backing_dev_info;
1533         mapping = &inode->i_data;
1534         mapping->a_ops = &mapping_aops;
1535         mapping->backing_dev_info = bdi;
1536         mapping_set_gfp_mask(mapping, GFP_NOFS);
1537         btp->bt_mapping = mapping;
1538         return 0;
1539 }
1540
1541 STATIC int
1542 xfs_alloc_delwrite_queue(
1543         xfs_buftarg_t           *btp)
1544 {
1545         int     error = 0;
1546
1547         INIT_LIST_HEAD(&btp->bt_list);
1548         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1549         spin_lock_init(&btp->bt_delwrite_lock);
1550         btp->bt_flags = 0;
1551         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1552         if (IS_ERR(btp->bt_task)) {
1553                 error = PTR_ERR(btp->bt_task);
1554                 goto out_error;
1555         }
1556         xfs_register_buftarg(btp);
1557 out_error:
1558         return error;
1559 }
1560
1561 xfs_buftarg_t *
1562 xfs_alloc_buftarg(
1563         struct block_device     *bdev,
1564         int                     external)
1565 {
1566         xfs_buftarg_t           *btp;
1567
1568         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1569
1570         btp->bt_dev =  bdev->bd_dev;
1571         btp->bt_bdev = bdev;
1572         if (xfs_setsize_buftarg_early(btp, bdev))
1573                 goto error;
1574         if (xfs_mapping_buftarg(btp, bdev))
1575                 goto error;
1576         if (xfs_alloc_delwrite_queue(btp))
1577                 goto error;
1578         xfs_alloc_bufhash(btp, external);
1579         return btp;
1580
1581 error:
1582         kmem_free(btp);
1583         return NULL;
1584 }
1585
1586
1587 /*
1588  *      Delayed write buffer handling
1589  */
1590 STATIC void
1591 xfs_buf_delwri_queue(
1592         xfs_buf_t               *bp,
1593         int                     unlock)
1594 {
1595         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1596         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1597
1598         XB_TRACE(bp, "delwri_q", (long)unlock);
1599         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1600
1601         spin_lock(dwlk);
1602         /* If already in the queue, dequeue and place at tail */
1603         if (!list_empty(&bp->b_list)) {
1604                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1605                 if (unlock)
1606                         atomic_dec(&bp->b_hold);
1607                 list_del(&bp->b_list);
1608         }
1609
1610         bp->b_flags |= _XBF_DELWRI_Q;
1611         list_add_tail(&bp->b_list, dwq);
1612         bp->b_queuetime = jiffies;
1613         spin_unlock(dwlk);
1614
1615         if (unlock)
1616                 xfs_buf_unlock(bp);
1617 }
1618
1619 void
1620 xfs_buf_delwri_dequeue(
1621         xfs_buf_t               *bp)
1622 {
1623         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1624         int                     dequeued = 0;
1625
1626         spin_lock(dwlk);
1627         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1628                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1629                 list_del_init(&bp->b_list);
1630                 dequeued = 1;
1631         }
1632         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1633         spin_unlock(dwlk);
1634
1635         if (dequeued)
1636                 xfs_buf_rele(bp);
1637
1638         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1639 }
1640
1641 STATIC void
1642 xfs_buf_runall_queues(
1643         struct workqueue_struct *queue)
1644 {
1645         flush_workqueue(queue);
1646 }
1647
1648 STATIC int
1649 xfsbufd_wakeup(
1650         int                     priority,
1651         gfp_t                   mask)
1652 {
1653         xfs_buftarg_t           *btp;
1654
1655         spin_lock(&xfs_buftarg_lock);
1656         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1657                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1658                         continue;
1659                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1660                 wake_up_process(btp->bt_task);
1661         }
1662         spin_unlock(&xfs_buftarg_lock);
1663         return 0;
1664 }
1665
1666 /*
1667  * Move as many buffers as specified to the supplied list
1668  * idicating if we skipped any buffers to prevent deadlocks.
1669  */
1670 STATIC int
1671 xfs_buf_delwri_split(
1672         xfs_buftarg_t   *target,
1673         struct list_head *list,
1674         unsigned long   age)
1675 {
1676         xfs_buf_t       *bp, *n;
1677         struct list_head *dwq = &target->bt_delwrite_queue;
1678         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1679         int             skipped = 0;
1680         int             force;
1681
1682         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1683         INIT_LIST_HEAD(list);
1684         spin_lock(dwlk);
1685         list_for_each_entry_safe(bp, n, dwq, b_list) {
1686                 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1687                 ASSERT(bp->b_flags & XBF_DELWRI);
1688
1689                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1690                         if (!force &&
1691                             time_before(jiffies, bp->b_queuetime + age)) {
1692                                 xfs_buf_unlock(bp);
1693                                 break;
1694                         }
1695
1696                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1697                                          _XBF_RUN_QUEUES);
1698                         bp->b_flags |= XBF_WRITE;
1699                         list_move_tail(&bp->b_list, list);
1700                 } else
1701                         skipped++;
1702         }
1703         spin_unlock(dwlk);
1704
1705         return skipped;
1706
1707 }
1708
1709 STATIC int
1710 xfsbufd(
1711         void            *data)
1712 {
1713         struct list_head tmp;
1714         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1715         int             count;
1716         xfs_buf_t       *bp;
1717
1718         current->flags |= PF_MEMALLOC;
1719
1720         set_freezable();
1721
1722         do {
1723                 if (unlikely(freezing(current))) {
1724                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1725                         refrigerator();
1726                 } else {
1727                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1728                 }
1729
1730                 schedule_timeout_interruptible(
1731                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1732
1733                 xfs_buf_delwri_split(target, &tmp,
1734                                 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1735
1736                 count = 0;
1737                 while (!list_empty(&tmp)) {
1738                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1739                         ASSERT(target == bp->b_target);
1740
1741                         list_del_init(&bp->b_list);
1742                         xfs_buf_iostrategy(bp);
1743                         count++;
1744                 }
1745
1746                 if (as_list_len > 0)
1747                         purge_addresses();
1748                 if (count)
1749                         blk_run_address_space(target->bt_mapping);
1750
1751         } while (!kthread_should_stop());
1752
1753         return 0;
1754 }
1755
1756 /*
1757  *      Go through all incore buffers, and release buffers if they belong to
1758  *      the given device. This is used in filesystem error handling to
1759  *      preserve the consistency of its metadata.
1760  */
1761 int
1762 xfs_flush_buftarg(
1763         xfs_buftarg_t   *target,
1764         int             wait)
1765 {
1766         struct list_head tmp;
1767         xfs_buf_t       *bp, *n;
1768         int             pincount = 0;
1769
1770         xfs_buf_runall_queues(xfsdatad_workqueue);
1771         xfs_buf_runall_queues(xfslogd_workqueue);
1772
1773         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1774         pincount = xfs_buf_delwri_split(target, &tmp, 0);
1775
1776         /*
1777          * Dropped the delayed write list lock, now walk the temporary list
1778          */
1779         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1780                 ASSERT(target == bp->b_target);
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         if (wait)
1790                 blk_run_address_space(target->bt_mapping);
1791
1792         /*
1793          * Remaining list items must be flushed before returning
1794          */
1795         while (!list_empty(&tmp)) {
1796                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1797
1798                 list_del_init(&bp->b_list);
1799                 xfs_iowait(bp);
1800                 xfs_buf_relse(bp);
1801         }
1802
1803         return pincount;
1804 }
1805
1806 int __init
1807 xfs_buf_init(void)
1808 {
1809 #ifdef XFS_BUF_TRACE
1810         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_NOFS);
1811 #endif
1812
1813         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1814                                                 KM_ZONE_HWALIGN, NULL);
1815         if (!xfs_buf_zone)
1816                 goto out_free_trace_buf;
1817
1818         xfslogd_workqueue = create_workqueue("xfslogd");
1819         if (!xfslogd_workqueue)
1820                 goto out_free_buf_zone;
1821
1822         xfsdatad_workqueue = create_workqueue("xfsdatad");
1823         if (!xfsdatad_workqueue)
1824                 goto out_destroy_xfslogd_workqueue;
1825
1826         register_shrinker(&xfs_buf_shake);
1827         return 0;
1828
1829  out_destroy_xfslogd_workqueue:
1830         destroy_workqueue(xfslogd_workqueue);
1831  out_free_buf_zone:
1832         kmem_zone_destroy(xfs_buf_zone);
1833  out_free_trace_buf:
1834 #ifdef XFS_BUF_TRACE
1835         ktrace_free(xfs_buf_trace_buf);
1836 #endif
1837         return -ENOMEM;
1838 }
1839
1840 void
1841 xfs_buf_terminate(void)
1842 {
1843         unregister_shrinker(&xfs_buf_shake);
1844         destroy_workqueue(xfsdatad_workqueue);
1845         destroy_workqueue(xfslogd_workqueue);
1846         kmem_zone_destroy(xfs_buf_zone);
1847 #ifdef XFS_BUF_TRACE
1848         ktrace_free(xfs_buf_trace_buf);
1849 #endif
1850 }
1851
1852 #ifdef CONFIG_KDB_MODULES
1853 struct list_head *
1854 xfs_get_buftarg_list(void)
1855 {
1856         return &xfs_buftarg_list;
1857 }
1858 #endif