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 #include "xfs_sb.h"
  38 #include "xfs_inum.h"
  39 #include "xfs_ag.h"
  40 #include "xfs_dmapi.h"
  41 #include "xfs_mount.h"
  42 #include "xfs_trace.h"
  43
  44 static kmem_zone_t *xfs_buf_zone;
  45 STATIC int xfsbufd(void *);
  46 STATIC int xfsbufd_wakeup(int, gfp_t);
  47 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
  48 static struct shrinker xfs_buf_shake = {
  49         .shrink = xfsbufd_wakeup,
  50         .seeks = DEFAULT_SEEKS,
  51 };
  52
  53 static struct workqueue_struct *xfslogd_workqueue;
  54 struct workqueue_struct *xfsdatad_workqueue;
  55 struct workqueue_struct *xfsconvertd_workqueue;
  56
  57 #ifdef XFS_BUF_LOCK_TRACKING
  58 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  59 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  60 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  61 #else
  62 # define XB_SET_OWNER(bp)       do { } while (0)
  63 # define XB_CLEAR_OWNER(bp)     do { } while (0)
  64 # define XB_GET_OWNER(bp)       do { } while (0)
  65 #endif
  66
  67 #define xb_to_gfp(flags) \
  68         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
  69           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
  70
  71 #define xb_to_km(flags) \
  72          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
  73
  74 #define xfs_buf_allocate(flags) \
  75         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
  76 #define xfs_buf_deallocate(bp) \
  77         kmem_zone_free(xfs_buf_zone, (bp));
  78
  79 static inline int
  80 xfs_buf_is_vmapped(
  81         struct xfs_buf  *bp)
  82 {
  83         /*
  84          * Return true if the buffer is vmapped.
  85          *
  86          * The XBF_MAPPED flag is set if the buffer should be mapped, but the
  87          * code is clever enough to know it doesn't have to map a single page,
  88          * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
  89          */
  90         return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
  91 }
  92
  93 static inline int
  94 xfs_buf_vmap_len(
  95         struct xfs_buf  *bp)
  96 {
  97         return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
  98 }
  99
 100 /*
 101  *      Page Region interfaces.
 102  *
 103  *      For pages in filesystems where the blocksize is smaller than the
 104  *      pagesize, we use the page->private field (long) to hold a bitmap
 105  *      of uptodate regions within the page.
 106  *
 107  *      Each such region is "bytes per page / bits per long" bytes long.
 108  *
 109  *      NBPPR == number-of-bytes-per-page-region
 110  *      BTOPR == bytes-to-page-region (rounded up)
 111  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
 112  */
 113 #if (BITS_PER_LONG == 32)
 114 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
 115 #elif (BITS_PER_LONG == 64)
 116 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
 117 #else
 118 #error BITS_PER_LONG must be 32 or 64
 119 #endif
 120 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
 121 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
 122 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
 123
 124 STATIC unsigned long
 125 page_region_mask(
 126         size_t          offset,
 127         size_t          length)
 128 {
 129         unsigned long   mask;
 130         int             first, final;
 131
 132         first = BTOPR(offset);
 133         final = BTOPRT(offset + length - 1);
 134         first = min(first, final);
 135
 136         mask = ~0UL;
 137         mask <<= BITS_PER_LONG - (final - first);
 138         mask >>= BITS_PER_LONG - (final);
 139
 140         ASSERT(offset + length <= PAGE_CACHE_SIZE);
 141         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
 142
 143         return mask;
 144 }
 145
 146 STATIC void
 147 set_page_region(
 148         struct page     *page,
 149         size_t          offset,
 150         size_t          length)
 151 {
 152         set_page_private(page,
 153                 page_private(page) | page_region_mask(offset, length));
 154         if (page_private(page) == ~0UL)
 155                 SetPageUptodate(page);
 156 }
 157
 158 STATIC int
 159 test_page_region(
 160         struct page     *page,
 161         size_t          offset,
 162         size_t          length)
 163 {
 164         unsigned long   mask = page_region_mask(offset, length);
 165
 166         return (mask && (page_private(page) & mask) == mask);
 167 }
 168
 169 /*
 170  *      Mapping of multi-page buffers into contiguous virtual space
 171  */
 172
 173 typedef struct a_list {
 174         void            *vm_addr;
 175         struct a_list   *next;
 176 } a_list_t;
 177
 178 static a_list_t         *as_free_head;
 179 static int              as_list_len;
 180 static DEFINE_SPINLOCK(as_lock);
 181
 182 /*
 183  *      Try to batch vunmaps because they are costly.
 184  */
 185 STATIC void
 186 free_address(
 187         void            *addr)
 188 {
 189         a_list_t        *aentry;
 190
 191 #ifdef CONFIG_XEN
 192         /*
 193          * Xen needs to be able to make sure it can get an exclusive
 194          * RO mapping of pages it wants to turn into a pagetable.  If
 195          * a newly allocated page is also still being vmap()ed by xfs,
 196          * it will cause pagetable construction to fail.  This is a
 197          * quick workaround to always eagerly unmap pages so that Xen
 198          * is happy.
 199          */
 200         vunmap(addr);
 201         return;
 202 #endif
 203
 204         aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
 205         if (likely(aentry)) {
 206                 spin_lock(&as_lock);
 207                 aentry->next = as_free_head;
 208                 aentry->vm_addr = addr;
 209                 as_free_head = aentry;
 210                 as_list_len++;
 211                 spin_unlock(&as_lock);
 212         } else {
 213                 vunmap(addr);
 214         }
 215 }
 216
 217 STATIC void
 218 purge_addresses(void)
 219 {
 220         a_list_t        *aentry, *old;
 221
 222         if (as_free_head == NULL)
 223                 return;
 224
 225         spin_lock(&as_lock);
 226         aentry = as_free_head;
 227         as_free_head = NULL;
 228         as_list_len = 0;
 229         spin_unlock(&as_lock);
 230
 231         while ((old = aentry) != NULL) {
 232                 vunmap(aentry->vm_addr);
 233                 aentry = aentry->next;
 234                 kfree(old);
 235         }
 236 }
 237
 238 /*
 239  *      Internal xfs_buf_t object manipulation
 240  */
 241
 242 STATIC void
 243 _xfs_buf_initialize(
 244         xfs_buf_t               *bp,
 245         xfs_buftarg_t           *target,
 246         xfs_off_t               range_base,
 247         size_t                  range_length,
 248         xfs_buf_flags_t         flags)
 249 {
 250         /*
 251          * We don't want certain flags to appear in b_flags.
 252          */
 253         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
 254
 255         memset(bp, 0, sizeof(xfs_buf_t));
 256         atomic_set(&bp->b_hold, 1);
 257         init_completion(&bp->b_iowait);
 258         INIT_LIST_HEAD(&bp->b_list);
 259         INIT_LIST_HEAD(&bp->b_hash_list);
 260         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
 261         XB_SET_OWNER(bp);
 262         bp->b_target = target;
 263         bp->b_file_offset = range_base;
 264         /*
 265          * Set buffer_length and count_desired to the same value initially.
 266          * I/O routines should use count_desired, which will be the same in
 267          * most cases but may be reset (e.g. XFS recovery).
 268          */
 269         bp->b_buffer_length = bp->b_count_desired = range_length;
 270         bp->b_flags = flags;
 271         bp->b_bn = XFS_BUF_DADDR_NULL;
 272         atomic_set(&bp->b_pin_count, 0);
 273         init_waitqueue_head(&bp->b_waiters);
 274
 275         XFS_STATS_INC(xb_create);
 276
 277         trace_xfs_buf_init(bp, _RET_IP_);
 278 }
 279
 280 /*
 281  *      Allocate a page array capable of holding a specified number
 282  *      of pages, and point the page buf at it.
 283  */
 284 STATIC int
 285 _xfs_buf_get_pages(
 286         xfs_buf_t               *bp,
 287         int                     page_count,
 288         xfs_buf_flags_t         flags)
 289 {
 290         /* Make sure that we have a page list */
 291         if (bp->b_pages == NULL) {
 292                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
 293                 bp->b_page_count = page_count;
 294                 if (page_count <= XB_PAGES) {
 295                         bp->b_pages = bp->b_page_array;
 296                 } else {
 297                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
 298                                         page_count, xb_to_km(flags));
 299                         if (bp->b_pages == NULL)
 300                                 return -ENOMEM;
 301                 }
 302                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 303         }
 304         return 0;
 305 }
 306
 307 /*
 308  *      Frees b_pages if it was allocated.
 309  */
 310 STATIC void
 311 _xfs_buf_free_pages(
 312         xfs_buf_t       *bp)
 313 {
 314         if (bp->b_pages != bp->b_page_array) {
 315                 kmem_free(bp->b_pages);
 316                 bp->b_pages = NULL;
 317         }
 318 }
 319
 320 /*
 321  *      Releases the specified buffer.
 322  *
 323  *      The modification state of any associated pages is left unchanged.
 324  *      The buffer most not be on any hash - use xfs_buf_rele instead for
 325  *      hashed and refcounted buffers
 326  */
 327 void
 328 xfs_buf_free(
 329         xfs_buf_t               *bp)
 330 {
 331         trace_xfs_buf_free(bp, _RET_IP_);
 332
 333         ASSERT(list_empty(&bp->b_hash_list));
 334
 335         if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
 336                 uint            i;
 337
 338                 if (xfs_buf_is_vmapped(bp))
 339                         free_address(bp->b_addr - bp->b_offset);
 340
 341                 for (i = 0; i < bp->b_page_count; i++) {
 342                         struct page     *page = bp->b_pages[i];
 343
 344                         if (bp->b_flags & _XBF_PAGE_CACHE)
 345                                 ASSERT(!PagePrivate(page));
 346                         page_cache_release(page);
 347                 }
 348         }
 349         _xfs_buf_free_pages(bp);
 350         xfs_buf_deallocate(bp);
 351 }
 352
 353 /*
 354  *      Finds all pages for buffer in question and builds it's page list.
 355  */
 356 STATIC int
 357 _xfs_buf_lookup_pages(
 358         xfs_buf_t               *bp,
 359         uint                    flags)
 360 {
 361         struct address_space    *mapping = bp->b_target->bt_mapping;
 362         size_t                  blocksize = bp->b_target->bt_bsize;
 363         size_t                  size = bp->b_count_desired;
 364         size_t                  nbytes, offset;
 365         gfp_t                   gfp_mask = xb_to_gfp(flags);
 366         unsigned short          page_count, i;
 367         pgoff_t                 first;
 368         xfs_off_t               end;
 369         int                     error;
 370
 371         end = bp->b_file_offset + bp->b_buffer_length;
 372         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 373
 374         error = _xfs_buf_get_pages(bp, page_count, flags);
 375         if (unlikely(error))
 376                 return error;
 377         bp->b_flags |= _XBF_PAGE_CACHE;
 378
 379         offset = bp->b_offset;
 380         first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
 381
 382         for (i = 0; i < bp->b_page_count; i++) {
 383                 struct page     *page;
 384                 uint            retries = 0;
 385
 386               retry:
 387                 page = find_or_create_page(mapping, first + i, gfp_mask);
 388                 if (unlikely(page == NULL)) {
 389                         if (flags & XBF_READ_AHEAD) {
 390                                 bp->b_page_count = i;
 391                                 for (i = 0; i < bp->b_page_count; i++)
 392                                         unlock_page(bp->b_pages[i]);
 393                                 return -ENOMEM;
 394                         }
 395
 396                         /*
 397                          * This could deadlock.
 398                          *
 399                          * But until all the XFS lowlevel code is revamped to
 400                          * handle buffer allocation failures we can't do much.
 401                          */
 402                         if (!(++retries % 100))
 403                                 printk(KERN_ERR
 404                                         "XFS: possible memory allocation "
 405                                         "deadlock in %s (mode:0x%x)\n",
 406                                         __func__, gfp_mask);
 407
 408                         XFS_STATS_INC(xb_page_retries);
 409                         xfsbufd_wakeup(0, gfp_mask);
 410                         congestion_wait(BLK_RW_ASYNC, HZ/50);
 411                         goto retry;
 412                 }
 413
 414                 XFS_STATS_INC(xb_page_found);
 415
 416                 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
 417                 size -= nbytes;
 418
 419                 ASSERT(!PagePrivate(page));
 420                 if (!PageUptodate(page)) {
 421                         page_count--;
 422                         if (blocksize >= PAGE_CACHE_SIZE) {
 423                                 if (flags & XBF_READ)
 424                                         bp->b_flags |= _XBF_PAGE_LOCKED;
 425                         } else if (!PagePrivate(page)) {
 426                                 if (test_page_region(page, offset, nbytes))
 427                                         page_count++;
 428                         }
 429                 }
 430
 431                 bp->b_pages[i] = page;
 432                 offset = 0;
 433         }
 434
 435         if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
 436                 for (i = 0; i < bp->b_page_count; i++)
 437                         unlock_page(bp->b_pages[i]);
 438         }
 439
 440         if (page_count == bp->b_page_count)
 441                 bp->b_flags |= XBF_DONE;
 442
 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                         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
 568         trace_xfs_buf_find(bp, flags, _RET_IP_);
 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(
 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                                         __func__);
 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         trace_xfs_buf_get(bp, flags, _RET_IP_);
 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 STATIC int
 635 _xfs_buf_read(
 636         xfs_buf_t               *bp,
 637         xfs_buf_flags_t         flags)
 638 {
 639         int                     status;
 640
 641         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
 642         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
 643
 644         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
 645                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 646         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
 647                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 648
 649         status = xfs_buf_iorequest(bp);
 650         if (!status && !(flags & XBF_ASYNC))
 651                 status = xfs_buf_iowait(bp);
 652         return status;
 653 }
 654
 655 xfs_buf_t *
 656 xfs_buf_read(
 657         xfs_buftarg_t           *target,
 658         xfs_off_t               ioff,
 659         size_t                  isize,
 660         xfs_buf_flags_t         flags)
 661 {
 662         xfs_buf_t               *bp;
 663
 664         flags |= XBF_READ;
 665
 666         bp = xfs_buf_get(target, ioff, isize, flags);
 667         if (bp) {
 668                 trace_xfs_buf_read(bp, flags, _RET_IP_);
 669
 670                 if (!XFS_BUF_ISDONE(bp)) {
 671                         XFS_STATS_INC(xb_get_read);
 672                         _xfs_buf_read(bp, flags);
 673                 } else if (flags & XBF_ASYNC) {
 674                         /*
 675                          * Read ahead call which is already satisfied,
 676                          * drop the buffer
 677                          */
 678                         goto no_buffer;
 679                 } else {
 680                         /* We do not want read in the flags */
 681                         bp->b_flags &= ~XBF_READ;
 682                 }
 683         }
 684
 685         return bp;
 686
 687  no_buffer:
 688         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 689                 xfs_buf_unlock(bp);
 690         xfs_buf_rele(bp);
 691         return NULL;
 692 }
 693
 694 /*
 695  *      If we are not low on memory then do the readahead in a deadlock
 696  *      safe manner.
 697  */
 698 void
 699 xfs_buf_readahead(
 700         xfs_buftarg_t           *target,
 701         xfs_off_t               ioff,
 702         size_t                  isize,
 703         xfs_buf_flags_t         flags)
 704 {
 705         struct backing_dev_info *bdi;
 706
 707         bdi = target->bt_mapping->backing_dev_info;
 708         if (bdi_read_congested(bdi))
 709                 return;
 710
 711         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 712         xfs_buf_read(target, ioff, isize, flags);
 713 }
 714
 715 xfs_buf_t *
 716 xfs_buf_get_empty(
 717         size_t                  len,
 718         xfs_buftarg_t           *target)
 719 {
 720         xfs_buf_t               *bp;
 721
 722         bp = xfs_buf_allocate(0);
 723         if (bp)
 724                 _xfs_buf_initialize(bp, target, 0, len, 0);
 725         return bp;
 726 }
 727
 728 static inline struct page *
 729 mem_to_page(
 730         void                    *addr)
 731 {
 732         if ((!is_vmalloc_addr(addr))) {
 733                 return virt_to_page(addr);
 734         } else {
 735                 return vmalloc_to_page(addr);
 736         }
 737 }
 738
 739 int
 740 xfs_buf_associate_memory(
 741         xfs_buf_t               *bp,
 742         void                    *mem,
 743         size_t                  len)
 744 {
 745         int                     rval;
 746         int                     i = 0;
 747         unsigned long           pageaddr;
 748         unsigned long           offset;
 749         size_t                  buflen;
 750         int                     page_count;
 751
 752         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
 753         offset = (unsigned long)mem - pageaddr;
 754         buflen = PAGE_CACHE_ALIGN(len + offset);
 755         page_count = buflen >> PAGE_CACHE_SHIFT;
 756
 757         /* Free any previous set of page pointers */
 758         if (bp->b_pages)
 759                 _xfs_buf_free_pages(bp);
 760
 761         bp->b_pages = NULL;
 762         bp->b_addr = mem;
 763
 764         rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
 765         if (rval)
 766                 return rval;
 767
 768         bp->b_offset = offset;
 769
 770         for (i = 0; i < bp->b_page_count; i++) {
 771                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 772                 pageaddr += PAGE_CACHE_SIZE;
 773         }
 774
 775         bp->b_count_desired = len;
 776         bp->b_buffer_length = buflen;
 777         bp->b_flags |= XBF_MAPPED;
 778         bp->b_flags &= ~_XBF_PAGE_LOCKED;
 779
 780         return 0;
 781 }
 782
 783 xfs_buf_t *
 784 xfs_buf_get_noaddr(
 785         size_t                  len,
 786         xfs_buftarg_t           *target)
 787 {
 788         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 789         int                     error, i;
 790         xfs_buf_t               *bp;
 791
 792         bp = xfs_buf_allocate(0);
 793         if (unlikely(bp == NULL))
 794                 goto fail;
 795         _xfs_buf_initialize(bp, target, 0, len, 0);
 796
 797         error = _xfs_buf_get_pages(bp, page_count, 0);
 798         if (error)
 799                 goto fail_free_buf;
 800
 801         for (i = 0; i < page_count; i++) {
 802                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
 803                 if (!bp->b_pages[i])
 804                         goto fail_free_mem;
 805         }
 806         bp->b_flags |= _XBF_PAGES;
 807
 808         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 809         if (unlikely(error)) {
 810                 printk(KERN_WARNING "%s: failed to map pages\n",
 811                                 __func__);
 812                 goto fail_free_mem;
 813         }
 814
 815         xfs_buf_unlock(bp);
 816
 817         trace_xfs_buf_get_noaddr(bp, _RET_IP_);
 818         return bp;
 819
 820  fail_free_mem:
 821         while (--i >= 0)
 822                 __free_page(bp->b_pages[i]);
 823         _xfs_buf_free_pages(bp);
 824  fail_free_buf:
 825         xfs_buf_deallocate(bp);
 826  fail:
 827         return NULL;
 828 }
 829
 830 /*
 831  *      Increment reference count on buffer, to hold the buffer concurrently
 832  *      with another thread which may release (free) the buffer asynchronously.
 833  *      Must hold the buffer already to call this function.
 834  */
 835 void
 836 xfs_buf_hold(
 837         xfs_buf_t               *bp)
 838 {
 839         trace_xfs_buf_hold(bp, _RET_IP_);
 840         atomic_inc(&bp->b_hold);
 841 }
 842
 843 /*
 844  *      Releases a hold on the specified buffer.  If the
 845  *      the hold count is 1, calls xfs_buf_free.
 846  */
 847 void
 848 xfs_buf_rele(
 849         xfs_buf_t               *bp)
 850 {
 851         xfs_bufhash_t           *hash = bp->b_hash;
 852
 853         trace_xfs_buf_rele(bp, _RET_IP_);
 854
 855         if (unlikely(!hash)) {
 856                 ASSERT(!bp->b_relse);
 857                 if (atomic_dec_and_test(&bp->b_hold))
 858                         xfs_buf_free(bp);
 859                 return;
 860         }
 861
 862         ASSERT(atomic_read(&bp->b_hold) > 0);
 863         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
 864                 if (bp->b_relse) {
 865                         atomic_inc(&bp->b_hold);
 866                         spin_unlock(&hash->bh_lock);
 867                         (*(bp->b_relse)) (bp);
 868                 } else if (bp->b_flags & XBF_FS_MANAGED) {
 869                         spin_unlock(&hash->bh_lock);
 870                 } else {
 871                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 872                         list_del_init(&bp->b_hash_list);
 873                         spin_unlock(&hash->bh_lock);
 874                         xfs_buf_free(bp);
 875                 }
 876         }
 877 }
 878
 879
 880 /*
 881  *      Mutual exclusion on buffers.  Locking model:
 882  *
 883  *      Buffers associated with inodes for which buffer locking
 884  *      is not enabled are not protected by semaphores, and are
 885  *      assumed to be exclusively owned by the caller.  There is a
 886  *      spinlock in the buffer, used by the caller when concurrent
 887  *      access is possible.
 888  */
 889
 890 /*
 891  *      Locks a buffer object, if it is not already locked.
 892  *      Note that this in no way locks the underlying pages, so it is only
 893  *      useful for synchronizing concurrent use of buffer objects, not for
 894  *      synchronizing independent access to the underlying pages.
 895  */
 896 int
 897 xfs_buf_cond_lock(
 898         xfs_buf_t               *bp)
 899 {
 900         int                     locked;
 901
 902         locked = down_trylock(&bp->b_sema) == 0;
 903         if (locked)
 904                 XB_SET_OWNER(bp);
 905
 906         trace_xfs_buf_cond_lock(bp, _RET_IP_);
 907         return locked ? 0 : -EBUSY;
 908 }
 909
 910 int
 911 xfs_buf_lock_value(
 912         xfs_buf_t               *bp)
 913 {
 914         return bp->b_sema.count;
 915 }
 916
 917 /*
 918  *      Locks a buffer object.
 919  *      Note that this in no way locks the underlying pages, so it is only
 920  *      useful for synchronizing concurrent use of buffer objects, not for
 921  *      synchronizing independent access to the underlying pages.
 922  */
 923 void
 924 xfs_buf_lock(
 925         xfs_buf_t               *bp)
 926 {
 927         trace_xfs_buf_lock(bp, _RET_IP_);
 928
 929         if (atomic_read(&bp->b_io_remaining))
 930                 blk_run_address_space(bp->b_target->bt_mapping);
 931         down(&bp->b_sema);
 932         XB_SET_OWNER(bp);
 933
 934         trace_xfs_buf_lock_done(bp, _RET_IP_);
 935 }
 936
 937 /*
 938  *      Releases the lock on the buffer object.
 939  *      If the buffer is marked delwri but is not queued, do so before we
 940  *      unlock the buffer as we need to set flags correctly.  We also need to
 941  *      take a reference for the delwri queue because the unlocker is going to
 942  *      drop their's and they don't know we just queued it.
 943  */
 944 void
 945 xfs_buf_unlock(
 946         xfs_buf_t               *bp)
 947 {
 948         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 949                 atomic_inc(&bp->b_hold);
 950                 bp->b_flags |= XBF_ASYNC;
 951                 xfs_buf_delwri_queue(bp, 0);
 952         }
 953
 954         XB_CLEAR_OWNER(bp);
 955         up(&bp->b_sema);
 956
 957         trace_xfs_buf_unlock(bp, _RET_IP_);
 958 }
 959
 960
 961 /*
 962  *      Pinning Buffer Storage in Memory
 963  *      Ensure that no attempt to force a buffer to disk will succeed.
 964  */
 965 void
 966 xfs_buf_pin(
 967         xfs_buf_t               *bp)
 968 {
 969         trace_xfs_buf_pin(bp, _RET_IP_);
 970         atomic_inc(&bp->b_pin_count);
 971 }
 972
 973 void
 974 xfs_buf_unpin(
 975         xfs_buf_t               *bp)
 976 {
 977         trace_xfs_buf_unpin(bp, _RET_IP_);
 978
 979         if (atomic_dec_and_test(&bp->b_pin_count))
 980                 wake_up_all(&bp->b_waiters);
 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                 trace_xfs_buf_ordered_retry(bp, _RET_IP_);
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         trace_xfs_buf_iodone(bp, _RET_IP_);
1047
1048         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1049         if (bp->b_error == 0)
1050                 bp->b_flags |= XBF_DONE;
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         trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1072 }
1073
1074 int
1075 xfs_bawrite(
1076         void                    *mp,
1077         struct xfs_buf          *bp)
1078 {
1079         trace_xfs_buf_bawrite(bp, _RET_IP_);
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         trace_xfs_buf_bdwrite(bp, _RET_IP_);
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 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         if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1132                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
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 (bp->b_flags & _XBF_PAGE_LOCKED)
1152                         unlock_page(page);
1153         } while (bvec >= bio->bi_io_vec);
1154
1155         _xfs_buf_ioend(bp, 1);
1156         bio_put(bio);
1157 }
1158
1159 STATIC void
1160 _xfs_buf_ioapply(
1161         xfs_buf_t               *bp)
1162 {
1163         int                     rw, map_i, total_nr_pages, nr_pages;
1164         struct bio              *bio;
1165         int                     offset = bp->b_offset;
1166         int                     size = bp->b_count_desired;
1167         sector_t                sector = bp->b_bn;
1168         unsigned int            blocksize = bp->b_target->bt_bsize;
1169
1170         total_nr_pages = bp->b_page_count;
1171         map_i = 0;
1172
1173         if (bp->b_flags & XBF_ORDERED) {
1174                 ASSERT(!(bp->b_flags & XBF_READ));
1175                 rw = WRITE_BARRIER;
1176         } else if (bp->b_flags & XBF_LOG_BUFFER) {
1177                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1178                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1179                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1180         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1181                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1182                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1183                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1184         } else {
1185                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1186                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1187         }
1188
1189         /* Special code path for reading a sub page size buffer in --
1190          * we populate up the whole page, and hence the other metadata
1191          * in the same page.  This optimization is only valid when the
1192          * filesystem block size is not smaller than the page size.
1193          */
1194         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1195             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1196               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1197             (blocksize >= PAGE_CACHE_SIZE)) {
1198                 bio = bio_alloc(GFP_NOIO, 1);
1199
1200                 bio->bi_bdev = bp->b_target->bt_bdev;
1201                 bio->bi_sector = sector - (offset >> BBSHIFT);
1202                 bio->bi_end_io = xfs_buf_bio_end_io;
1203                 bio->bi_private = bp;
1204
1205                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1206                 size = 0;
1207
1208                 atomic_inc(&bp->b_io_remaining);
1209
1210                 goto submit_io;
1211         }
1212
1213 next_chunk:
1214         atomic_inc(&bp->b_io_remaining);
1215         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1216         if (nr_pages > total_nr_pages)
1217                 nr_pages = total_nr_pages;
1218
1219         bio = bio_alloc(GFP_NOIO, nr_pages);
1220         bio->bi_bdev = bp->b_target->bt_bdev;
1221         bio->bi_sector = sector;
1222         bio->bi_end_io = xfs_buf_bio_end_io;
1223         bio->bi_private = bp;
1224
1225         for (; size && nr_pages; nr_pages--, map_i++) {
1226                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1227
1228                 if (nbytes > size)
1229                         nbytes = size;
1230
1231                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1232                 if (rbytes < nbytes)
1233                         break;
1234
1235                 offset = 0;
1236                 sector += nbytes >> BBSHIFT;
1237                 size -= nbytes;
1238                 total_nr_pages--;
1239         }
1240
1241 submit_io:
1242         if (likely(bio->bi_size)) {
1243                 if (xfs_buf_is_vmapped(bp)) {
1244                         flush_kernel_vmap_range(bp->b_addr,
1245                                                 xfs_buf_vmap_len(bp));
1246                 }
1247                 submit_bio(rw, bio);
1248                 if (size)
1249                         goto next_chunk;
1250         } else {
1251                 bio_put(bio);
1252                 xfs_buf_ioerror(bp, EIO);
1253         }
1254 }
1255
1256 int
1257 xfs_buf_iorequest(
1258         xfs_buf_t               *bp)
1259 {
1260         trace_xfs_buf_iorequest(bp, _RET_IP_);
1261
1262         if (bp->b_flags & XBF_DELWRI) {
1263                 xfs_buf_delwri_queue(bp, 1);
1264                 return 0;
1265         }
1266
1267         if (bp->b_flags & XBF_WRITE) {
1268                 xfs_buf_wait_unpin(bp);
1269         }
1270
1271         xfs_buf_hold(bp);
1272
1273         /* Set the count to 1 initially, this will stop an I/O
1274          * completion callout which happens before we have started
1275          * all the I/O from calling xfs_buf_ioend too early.
1276          */
1277         atomic_set(&bp->b_io_remaining, 1);
1278         _xfs_buf_ioapply(bp);
1279         _xfs_buf_ioend(bp, 0);
1280
1281         xfs_buf_rele(bp);
1282         return 0;
1283 }
1284
1285 /*
1286  *      Waits for I/O to complete on the buffer supplied.
1287  *      It returns immediately if no I/O is pending.
1288  *      It returns the I/O error code, if any, or 0 if there was no error.
1289  */
1290 int
1291 xfs_buf_iowait(
1292         xfs_buf_t               *bp)
1293 {
1294         trace_xfs_buf_iowait(bp, _RET_IP_);
1295
1296         if (atomic_read(&bp->b_io_remaining))
1297                 blk_run_address_space(bp->b_target->bt_mapping);
1298         wait_for_completion(&bp->b_iowait);
1299
1300         trace_xfs_buf_iowait_done(bp, _RET_IP_);
1301         return bp->b_error;
1302 }
1303
1304 xfs_caddr_t
1305 xfs_buf_offset(
1306         xfs_buf_t               *bp,
1307         size_t                  offset)
1308 {
1309         struct page             *page;
1310
1311         if (bp->b_flags & XBF_MAPPED)
1312                 return XFS_BUF_PTR(bp) + offset;
1313
1314         offset += bp->b_offset;
1315         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1316         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1317 }
1318
1319 /*
1320  *      Move data into or out of a buffer.
1321  */
1322 void
1323 xfs_buf_iomove(
1324         xfs_buf_t               *bp,    /* buffer to process            */
1325         size_t                  boff,   /* starting buffer offset       */
1326         size_t                  bsize,  /* length to copy               */
1327         caddr_t                 data,   /* data address                 */
1328         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1329 {
1330         size_t                  bend, cpoff, csize;
1331         struct page             *page;
1332
1333         bend = boff + bsize;
1334         while (boff < bend) {
1335                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1336                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1337                 csize = min_t(size_t,
1338                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1339
1340                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1341
1342                 switch (mode) {
1343                 case XBRW_ZERO:
1344                         memset(page_address(page) + cpoff, 0, csize);
1345                         break;
1346                 case XBRW_READ:
1347                         memcpy(data, page_address(page) + cpoff, csize);
1348                         break;
1349                 case XBRW_WRITE:
1350                         memcpy(page_address(page) + cpoff, data, csize);
1351                 }
1352
1353                 boff += csize;
1354                 data += csize;
1355         }
1356 }
1357
1358 /*
1359  *      Handling of buffer targets (buftargs).
1360  */
1361
1362 /*
1363  *      Wait for any bufs with callbacks that have been submitted but
1364  *      have not yet returned... walk the hash list for the target.
1365  */
1366 void
1367 xfs_wait_buftarg(
1368         xfs_buftarg_t   *btp)
1369 {
1370         xfs_buf_t       *bp, *n;
1371         xfs_bufhash_t   *hash;
1372         uint            i;
1373
1374         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1375                 hash = &btp->bt_hash[i];
1376 again:
1377                 spin_lock(&hash->bh_lock);
1378                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1379                         ASSERT(btp == bp->b_target);
1380                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1381                                 spin_unlock(&hash->bh_lock);
1382                                 /*
1383                                  * Catch superblock reference count leaks
1384                                  * immediately
1385                                  */
1386                                 BUG_ON(bp->b_bn == 0);
1387                                 delay(100);
1388                                 goto again;
1389                         }
1390                 }
1391                 spin_unlock(&hash->bh_lock);
1392         }
1393 }
1394
1395 /*
1396  *      Allocate buffer hash table for a given target.
1397  *      For devices containing metadata (i.e. not the log/realtime devices)
1398  *      we need to allocate a much larger hash table.
1399  */
1400 STATIC void
1401 xfs_alloc_bufhash(
1402         xfs_buftarg_t           *btp,
1403         int                     external)
1404 {
1405         unsigned int            i;
1406
1407         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1408         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1409         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1410                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1411         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1412                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1413                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1414         }
1415 }
1416
1417 STATIC void
1418 xfs_free_bufhash(
1419         xfs_buftarg_t           *btp)
1420 {
1421         kmem_free(btp->bt_hash);
1422         btp->bt_hash = NULL;
1423 }
1424
1425 /*
1426  *      buftarg list for delwrite queue processing
1427  */
1428 static LIST_HEAD(xfs_buftarg_list);
1429 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1430
1431 STATIC void
1432 xfs_register_buftarg(
1433         xfs_buftarg_t           *btp)
1434 {
1435         spin_lock(&xfs_buftarg_lock);
1436         list_add(&btp->bt_list, &xfs_buftarg_list);
1437         spin_unlock(&xfs_buftarg_lock);
1438 }
1439
1440 STATIC void
1441 xfs_unregister_buftarg(
1442         xfs_buftarg_t           *btp)
1443 {
1444         spin_lock(&xfs_buftarg_lock);
1445         list_del(&btp->bt_list);
1446         spin_unlock(&xfs_buftarg_lock);
1447 }
1448
1449 void
1450 xfs_free_buftarg(
1451         struct xfs_mount        *mp,
1452         struct xfs_buftarg      *btp)
1453 {
1454         xfs_flush_buftarg(btp, 1);
1455         if (mp->m_flags & XFS_MOUNT_BARRIER)
1456                 xfs_blkdev_issue_flush(btp);
1457         xfs_free_bufhash(btp);
1458         iput(btp->bt_mapping->host);
1459
1460         /* Unregister the buftarg first so that we don't get a
1461          * wakeup finding a non-existent task
1462          */
1463         xfs_unregister_buftarg(btp);
1464         kthread_stop(btp->bt_task);
1465
1466         kmem_free(btp);
1467 }
1468
1469 STATIC int
1470 xfs_setsize_buftarg_flags(
1471         xfs_buftarg_t           *btp,
1472         unsigned int            blocksize,
1473         unsigned int            sectorsize,
1474         int                     verbose)
1475 {
1476         btp->bt_bsize = blocksize;
1477         btp->bt_sshift = ffs(sectorsize) - 1;
1478         btp->bt_smask = sectorsize - 1;
1479
1480         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1481                 printk(KERN_WARNING
1482                         "XFS: Cannot set_blocksize to %u on device %s\n",
1483                         sectorsize, XFS_BUFTARG_NAME(btp));
1484                 return EINVAL;
1485         }
1486
1487         if (verbose &&
1488             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1489                 printk(KERN_WARNING
1490                         "XFS: %u byte sectors in use on device %s.  "
1491                         "This is suboptimal; %u or greater is ideal.\n",
1492                         sectorsize, XFS_BUFTARG_NAME(btp),
1493                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1494         }
1495
1496         return 0;
1497 }
1498
1499 /*
1500  *      When allocating the initial buffer target we have not yet
1501  *      read in the superblock, so don't know what sized sectors
1502  *      are being used is at this early stage.  Play safe.
1503  */
1504 STATIC int
1505 xfs_setsize_buftarg_early(
1506         xfs_buftarg_t           *btp,
1507         struct block_device     *bdev)
1508 {
1509         return xfs_setsize_buftarg_flags(btp,
1510                         PAGE_CACHE_SIZE, bdev_logical_block_size(bdev), 0);
1511 }
1512
1513 int
1514 xfs_setsize_buftarg(
1515         xfs_buftarg_t           *btp,
1516         unsigned int            blocksize,
1517         unsigned int            sectorsize)
1518 {
1519         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1520 }
1521
1522 STATIC int
1523 xfs_mapping_buftarg(
1524         xfs_buftarg_t           *btp,
1525         struct block_device     *bdev)
1526 {
1527         struct backing_dev_info *bdi;
1528         struct inode            *inode;
1529         struct address_space    *mapping;
1530         static const struct address_space_operations mapping_aops = {
1531                 .sync_page = block_sync_page,
1532                 .migratepage = fail_migrate_page,
1533         };
1534
1535         inode = new_inode(bdev->bd_inode->i_sb);
1536         if (!inode) {
1537                 printk(KERN_WARNING
1538                         "XFS: Cannot allocate mapping inode for device %s\n",
1539                         XFS_BUFTARG_NAME(btp));
1540                 return ENOMEM;
1541         }
1542         inode->i_mode = S_IFBLK;
1543         inode->i_bdev = bdev;
1544         inode->i_rdev = bdev->bd_dev;
1545         bdi = blk_get_backing_dev_info(bdev);
1546         if (!bdi)
1547                 bdi = &default_backing_dev_info;
1548         mapping = &inode->i_data;
1549         mapping->a_ops = &mapping_aops;
1550         mapping->backing_dev_info = bdi;
1551         mapping_set_gfp_mask(mapping, GFP_NOFS);
1552         btp->bt_mapping = mapping;
1553         return 0;
1554 }
1555
1556 STATIC int
1557 xfs_alloc_delwrite_queue(
1558         xfs_buftarg_t           *btp)
1559 {
1560         int     error = 0;
1561
1562         INIT_LIST_HEAD(&btp->bt_list);
1563         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1564         spin_lock_init(&btp->bt_delwrite_lock);
1565         btp->bt_flags = 0;
1566         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1567         if (IS_ERR(btp->bt_task)) {
1568                 error = PTR_ERR(btp->bt_task);
1569                 goto out_error;
1570         }
1571         xfs_register_buftarg(btp);
1572 out_error:
1573         return error;
1574 }
1575
1576 xfs_buftarg_t *
1577 xfs_alloc_buftarg(
1578         struct block_device     *bdev,
1579         int                     external)
1580 {
1581         xfs_buftarg_t           *btp;
1582
1583         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1584
1585         btp->bt_dev =  bdev->bd_dev;
1586         btp->bt_bdev = bdev;
1587         if (xfs_setsize_buftarg_early(btp, bdev))
1588                 goto error;
1589         if (xfs_mapping_buftarg(btp, bdev))
1590                 goto error;
1591         if (xfs_alloc_delwrite_queue(btp))
1592                 goto error;
1593         xfs_alloc_bufhash(btp, external);
1594         return btp;
1595
1596 error:
1597         kmem_free(btp);
1598         return NULL;
1599 }
1600
1601
1602 /*
1603  *      Delayed write buffer handling
1604  */
1605 STATIC void
1606 xfs_buf_delwri_queue(
1607         xfs_buf_t               *bp,
1608         int                     unlock)
1609 {
1610         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1611         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1612
1613         trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1614
1615         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1616
1617         spin_lock(dwlk);
1618         /* If already in the queue, dequeue and place at tail */
1619         if (!list_empty(&bp->b_list)) {
1620                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1621                 if (unlock)
1622                         atomic_dec(&bp->b_hold);
1623                 list_del(&bp->b_list);
1624         }
1625
1626         bp->b_flags |= _XBF_DELWRI_Q;
1627         list_add_tail(&bp->b_list, dwq);
1628         bp->b_queuetime = jiffies;
1629         spin_unlock(dwlk);
1630
1631         if (unlock)
1632                 xfs_buf_unlock(bp);
1633 }
1634
1635 void
1636 xfs_buf_delwri_dequeue(
1637         xfs_buf_t               *bp)
1638 {
1639         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1640         int                     dequeued = 0;
1641
1642         spin_lock(dwlk);
1643         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1644                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1645                 list_del_init(&bp->b_list);
1646                 dequeued = 1;
1647         }
1648         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1649         spin_unlock(dwlk);
1650
1651         if (dequeued)
1652                 xfs_buf_rele(bp);
1653
1654         trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1655 }
1656
1657 STATIC void
1658 xfs_buf_runall_queues(
1659         struct workqueue_struct *queue)
1660 {
1661         flush_workqueue(queue);
1662 }
1663
1664 STATIC int
1665 xfsbufd_wakeup(
1666         int                     priority,
1667         gfp_t                   mask)
1668 {
1669         xfs_buftarg_t           *btp;
1670
1671         spin_lock(&xfs_buftarg_lock);
1672         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1673                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1674                         continue;
1675                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1676                 wake_up_process(btp->bt_task);
1677         }
1678         spin_unlock(&xfs_buftarg_lock);
1679         return 0;
1680 }
1681
1682 /*
1683  * Move as many buffers as specified to the supplied list
1684  * idicating if we skipped any buffers to prevent deadlocks.
1685  */
1686 STATIC int
1687 xfs_buf_delwri_split(
1688         xfs_buftarg_t   *target,
1689         struct list_head *list,
1690         unsigned long   age)
1691 {
1692         xfs_buf_t       *bp, *n;
1693         struct list_head *dwq = &target->bt_delwrite_queue;
1694         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1695         int             skipped = 0;
1696         int             force;
1697
1698         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1699         INIT_LIST_HEAD(list);
1700         spin_lock(dwlk);
1701         list_for_each_entry_safe(bp, n, dwq, b_list) {
1702                 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1703                 ASSERT(bp->b_flags & XBF_DELWRI);
1704
1705                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1706                         if (!force &&
1707                             time_before(jiffies, bp->b_queuetime + age)) {
1708                                 xfs_buf_unlock(bp);
1709                                 break;
1710                         }
1711
1712                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1713                                          _XBF_RUN_QUEUES);
1714                         bp->b_flags |= XBF_WRITE;
1715                         list_move_tail(&bp->b_list, list);
1716                 } else
1717                         skipped++;
1718         }
1719         spin_unlock(dwlk);
1720
1721         return skipped;
1722
1723 }
1724
1725 STATIC int
1726 xfsbufd(
1727         void            *data)
1728 {
1729         struct list_head tmp;
1730         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1731         int             count;
1732         xfs_buf_t       *bp;
1733
1734         current->flags |= PF_MEMALLOC;
1735
1736         set_freezable();
1737
1738         do {
1739                 if (unlikely(freezing(current))) {
1740                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1741                         refrigerator();
1742                 } else {
1743                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1744                 }
1745
1746                 schedule_timeout_interruptible(
1747                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1748
1749                 xfs_buf_delwri_split(target, &tmp,
1750                                 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1751
1752                 count = 0;
1753                 while (!list_empty(&tmp)) {
1754                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1755                         ASSERT(target == bp->b_target);
1756
1757                         list_del_init(&bp->b_list);
1758                         xfs_buf_iostrategy(bp);
1759                         count++;
1760                 }
1761
1762                 if (as_list_len > 0)
1763                         purge_addresses();
1764                 if (count)
1765                         blk_run_address_space(target->bt_mapping);
1766
1767         } while (!kthread_should_stop());
1768
1769         return 0;
1770 }
1771
1772 /*
1773  *      Go through all incore buffers, and release buffers if they belong to
1774  *      the given device. This is used in filesystem error handling to
1775  *      preserve the consistency of its metadata.
1776  */
1777 int
1778 xfs_flush_buftarg(
1779         xfs_buftarg_t   *target,
1780         int             wait)
1781 {
1782         struct list_head tmp;
1783         xfs_buf_t       *bp, *n;
1784         int             pincount = 0;
1785
1786         xfs_buf_runall_queues(xfsconvertd_workqueue);
1787         xfs_buf_runall_queues(xfsdatad_workqueue);
1788         xfs_buf_runall_queues(xfslogd_workqueue);
1789
1790         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1791         pincount = xfs_buf_delwri_split(target, &tmp, 0);
1792
1793         /*
1794          * Dropped the delayed write list lock, now walk the temporary list
1795          */
1796         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1797                 ASSERT(target == bp->b_target);
1798                 if (wait)
1799                         bp->b_flags &= ~XBF_ASYNC;
1800                 else
1801                         list_del_init(&bp->b_list);
1802
1803                 xfs_buf_iostrategy(bp);
1804         }
1805
1806         if (wait)
1807                 blk_run_address_space(target->bt_mapping);
1808
1809         /*
1810          * Remaining list items must be flushed before returning
1811          */
1812         while (!list_empty(&tmp)) {
1813                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1814
1815                 list_del_init(&bp->b_list);
1816                 xfs_iowait(bp);
1817                 xfs_buf_relse(bp);
1818         }
1819
1820         return pincount;
1821 }
1822
1823 int __init
1824 xfs_buf_init(void)
1825 {
1826         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1827                                                 KM_ZONE_HWALIGN, NULL);
1828         if (!xfs_buf_zone)
1829                 goto out;
1830
1831         xfslogd_workqueue = create_workqueue("xfslogd");
1832         if (!xfslogd_workqueue)
1833                 goto out_free_buf_zone;
1834
1835         xfsdatad_workqueue = create_workqueue("xfsdatad");
1836         if (!xfsdatad_workqueue)
1837                 goto out_destroy_xfslogd_workqueue;
1838
1839         xfsconvertd_workqueue = create_workqueue("xfsconvertd");
1840         if (!xfsconvertd_workqueue)
1841                 goto out_destroy_xfsdatad_workqueue;
1842
1843         register_shrinker(&xfs_buf_shake);
1844         return 0;
1845
1846  out_destroy_xfsdatad_workqueue:
1847         destroy_workqueue(xfsdatad_workqueue);
1848  out_destroy_xfslogd_workqueue:
1849         destroy_workqueue(xfslogd_workqueue);
1850  out_free_buf_zone:
1851         kmem_zone_destroy(xfs_buf_zone);
1852  out:
1853         return -ENOMEM;
1854 }
1855
1856 void
1857 xfs_buf_terminate(void)
1858 {
1859         unregister_shrinker(&xfs_buf_shake);
1860         destroy_workqueue(xfsconvertd_workqueue);
1861         destroy_workqueue(xfsdatad_workqueue);
1862         destroy_workqueue(xfslogd_workqueue);
1863         kmem_zone_destroy(xfs_buf_zone);
1864 }
1865
1866 #ifdef CONFIG_KDB_MODULES
1867 struct list_head *
1868 xfs_get_buftarg_list(void)
1869 {
1870         return &xfs_buftarg_list;
1871 }
1872 #endif