fs/xfs/linux-2.6/xfs_buf.c

   1 /*
   2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include <linux/stddef.h>
  20 #include <linux/errno.h>
  21 #include <linux/slab.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/init.h>
  24 #include <linux/vmalloc.h>
  25 #include <linux/bio.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/proc_fs.h>
  28 #include <linux/workqueue.h>
  29 #include <linux/percpu.h>
  30 #include <linux/blkdev.h>
  31 #include <linux/hash.h>
  32 #include <linux/kthread.h>
  33 #include <linux/migrate.h>
  34 #include <linux/backing-dev.h>
  35 #include <linux/freezer.h>
  36
  37 static kmem_zone_t *xfs_buf_zone;
  38 STATIC int xfsbufd(void *);
  39 STATIC int xfsbufd_wakeup(int, gfp_t);
  40 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
  41 static struct shrinker xfs_buf_shake = {
  42         .shrink = xfsbufd_wakeup,
  43         .seeks = DEFAULT_SEEKS,
  44 };
  45
  46 static struct workqueue_struct *xfslogd_workqueue;
  47 struct workqueue_struct *xfsdatad_workqueue;
  48
  49 #ifdef XFS_BUF_TRACE
  50 void
  51 xfs_buf_trace(
  52         xfs_buf_t       *bp,
  53         char            *id,
  54         void            *data,
  55         void            *ra)
  56 {
  57         ktrace_enter(xfs_buf_trace_buf,
  58                 bp, id,
  59                 (void *)(unsigned long)bp->b_flags,
  60                 (void *)(unsigned long)bp->b_hold.counter,
  61                 (void *)(unsigned long)bp->b_sema.count.counter,
  62                 (void *)current,
  63                 data, ra,
  64                 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
  65                 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
  66                 (void *)(unsigned long)bp->b_buffer_length,
  67                 NULL, NULL, NULL, NULL, NULL);
  68 }
  69 ktrace_t *xfs_buf_trace_buf;
  70 #define XFS_BUF_TRACE_SIZE      4096
  71 #define XB_TRACE(bp, id, data)  \
  72         xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
  73 #else
  74 #define XB_TRACE(bp, id, data)  do { } while (0)
  75 #endif
  76
  77 #ifdef XFS_BUF_LOCK_TRACKING
  78 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  79 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  80 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  81 #else
  82 # define XB_SET_OWNER(bp)       do { } while (0)
  83 # define XB_CLEAR_OWNER(bp)     do { } while (0)
  84 # define XB_GET_OWNER(bp)       do { } while (0)
  85 #endif
  86
  87 #define xb_to_gfp(flags) \
  88         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
  89           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
  90
  91 #define xb_to_km(flags) \
  92          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
  93
  94 #define xfs_buf_allocate(flags) \
  95         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
  96 #define xfs_buf_deallocate(bp) \
  97         kmem_zone_free(xfs_buf_zone, (bp));
  98
  99 /*
 100  *      Page Region interfaces.
 101  *
 102  *      For pages in filesystems where the blocksize is smaller than the
 103  *      pagesize, we use the page->private field (long) to hold a bitmap
 104  *      of uptodate regions within the page.
 105  *
 106  *      Each such region is "bytes per page / bits per long" bytes long.
 107  *
 108  *      NBPPR == number-of-bytes-per-page-region
 109  *      BTOPR == bytes-to-page-region (rounded up)
 110  *      BTOPRT == bytes-to-page-region-truncated (rounded down)
 111  */
 112 #if (BITS_PER_LONG == 32)
 113 #define PRSHIFT         (PAGE_CACHE_SHIFT - 5)  /* (32 == 1<<5) */
 114 #elif (BITS_PER_LONG == 64)
 115 #define PRSHIFT         (PAGE_CACHE_SHIFT - 6)  /* (64 == 1<<6) */
 116 #else
 117 #error BITS_PER_LONG must be 32 or 64
 118 #endif
 119 #define NBPPR           (PAGE_CACHE_SIZE/BITS_PER_LONG)
 120 #define BTOPR(b)        (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
 121 #define BTOPRT(b)       (((unsigned int)(b) >> PRSHIFT))
 122
 123 STATIC unsigned long
 124 page_region_mask(
 125         size_t          offset,
 126         size_t          length)
 127 {
 128         unsigned long   mask;
 129         int             first, final;
 130
 131         first = BTOPR(offset);
 132         final = BTOPRT(offset + length - 1);
 133         first = min(first, final);
 134
 135         mask = ~0UL;
 136         mask <<= BITS_PER_LONG - (final - first);
 137         mask >>= BITS_PER_LONG - (final);
 138
 139         ASSERT(offset + length <= PAGE_CACHE_SIZE);
 140         ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
 141
 142         return mask;
 143 }
 144
 145 STATIC_INLINE void
 146 set_page_region(
 147         struct page     *page,
 148         size_t          offset,
 149         size_t          length)
 150 {
 151         set_page_private(page,
 152                 page_private(page) | page_region_mask(offset, length));
 153         if (page_private(page) == ~0UL)
 154                 SetPageUptodate(page);
 155 }
 156
 157 STATIC_INLINE int
 158 test_page_region(
 159         struct page     *page,
 160         size_t          offset,
 161         size_t          length)
 162 {
 163         unsigned long   mask = page_region_mask(offset, length);
 164
 165         return (mask && (page_private(page) & mask) == mask);
 166 }
 167
 168 /*
 169  *      Mapping of multi-page buffers into contiguous virtual space
 170  */
 171
 172 typedef struct a_list {
 173         void            *vm_addr;
 174         struct a_list   *next;
 175 } a_list_t;
 176
 177 static a_list_t         *as_free_head;
 178 static int              as_list_len;
 179 static DEFINE_SPINLOCK(as_lock);
 180
 181 /*
 182  *      Try to batch vunmaps because they are costly.
 183  */
 184 STATIC void
 185 free_address(
 186         void            *addr)
 187 {
 188         a_list_t        *aentry;
 189
 190 #ifdef CONFIG_XEN
 191         /*
 192          * Xen needs to be able to make sure it can get an exclusive
 193          * RO mapping of pages it wants to turn into a pagetable.  If
 194          * a newly allocated page is also still being vmap()ed by xfs,
 195          * it will cause pagetable construction to fail.  This is a
 196          * quick workaround to always eagerly unmap pages so that Xen
 197          * is happy.
 198          */
 199         vunmap(addr);
 200         return;
 201 #endif
 202
 203         aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
 204         if (likely(aentry)) {
 205                 spin_lock(&as_lock);
 206                 aentry->next = as_free_head;
 207                 aentry->vm_addr = addr;
 208                 as_free_head = aentry;
 209                 as_list_len++;
 210                 spin_unlock(&as_lock);
 211         } else {
 212                 vunmap(addr);
 213         }
 214 }
 215
 216 STATIC void
 217 purge_addresses(void)
 218 {
 219         a_list_t        *aentry, *old;
 220
 221         if (as_free_head == NULL)
 222                 return;
 223
 224         spin_lock(&as_lock);
 225         aentry = as_free_head;
 226         as_free_head = NULL;
 227         as_list_len = 0;
 228         spin_unlock(&as_lock);
 229
 230         while ((old = aentry) != NULL) {
 231                 vunmap(aentry->vm_addr);
 232                 aentry = aentry->next;
 233                 kfree(old);
 234         }
 235 }
 236
 237 /*
 238  *      Internal xfs_buf_t object manipulation
 239  */
 240
 241 STATIC void
 242 _xfs_buf_initialize(
 243         xfs_buf_t               *bp,
 244         xfs_buftarg_t           *target,
 245         xfs_off_t               range_base,
 246         size_t                  range_length,
 247         xfs_buf_flags_t         flags)
 248 {
 249         /*
 250          * We don't want certain flags to appear in b_flags.
 251          */
 252         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
 253
 254         memset(bp, 0, sizeof(xfs_buf_t));
 255         atomic_set(&bp->b_hold, 1);
 256         init_MUTEX_LOCKED(&bp->b_iodonesema);
 257         INIT_LIST_HEAD(&bp->b_list);
 258         INIT_LIST_HEAD(&bp->b_hash_list);
 259         init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
 260         XB_SET_OWNER(bp);
 261         bp->b_target = target;
 262         bp->b_file_offset = range_base;
 263         /*
 264          * Set buffer_length and count_desired to the same value initially.
 265          * I/O routines should use count_desired, which will be the same in
 266          * most cases but may be reset (e.g. XFS recovery).
 267          */
 268         bp->b_buffer_length = bp->b_count_desired = range_length;
 269         bp->b_flags = flags;
 270         bp->b_bn = XFS_BUF_DADDR_NULL;
 271         atomic_set(&bp->b_pin_count, 0);
 272         init_waitqueue_head(&bp->b_waiters);
 273
 274         XFS_STATS_INC(xb_create);
 275         XB_TRACE(bp, "initialize", target);
 276 }
 277
 278 /*
 279  *      Allocate a page array capable of holding a specified number
 280  *      of pages, and point the page buf at it.
 281  */
 282 STATIC int
 283 _xfs_buf_get_pages(
 284         xfs_buf_t               *bp,
 285         int                     page_count,
 286         xfs_buf_flags_t         flags)
 287 {
 288         /* Make sure that we have a page list */
 289         if (bp->b_pages == NULL) {
 290                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
 291                 bp->b_page_count = page_count;
 292                 if (page_count <= XB_PAGES) {
 293                         bp->b_pages = bp->b_page_array;
 294                 } else {
 295                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
 296                                         page_count, xb_to_km(flags));
 297                         if (bp->b_pages == NULL)
 298                                 return -ENOMEM;
 299                 }
 300                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 301         }
 302         return 0;
 303 }
 304
 305 /*
 306  *      Frees b_pages if it was allocated.
 307  */
 308 STATIC void
 309 _xfs_buf_free_pages(
 310         xfs_buf_t       *bp)
 311 {
 312         if (bp->b_pages != bp->b_page_array) {
 313                 kmem_free(bp->b_pages);
 314         }
 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 xfs_buf_t *
 634 xfs_buf_read_flags(
 635         xfs_buftarg_t           *target,
 636         xfs_off_t               ioff,
 637         size_t                  isize,
 638         xfs_buf_flags_t         flags)
 639 {
 640         xfs_buf_t               *bp;
 641
 642         flags |= XBF_READ;
 643
 644         bp = xfs_buf_get_flags(target, ioff, isize, flags);
 645         if (bp) {
 646                 if (!XFS_BUF_ISDONE(bp)) {
 647                         XB_TRACE(bp, "read", (unsigned long)flags);
 648                         XFS_STATS_INC(xb_get_read);
 649                         xfs_buf_iostart(bp, flags);
 650                 } else if (flags & XBF_ASYNC) {
 651                         XB_TRACE(bp, "read_async", (unsigned long)flags);
 652                         /*
 653                          * Read ahead call which is already satisfied,
 654                          * drop the buffer
 655                          */
 656                         goto no_buffer;
 657                 } else {
 658                         XB_TRACE(bp, "read_done", (unsigned long)flags);
 659                         /* We do not want read in the flags */
 660                         bp->b_flags &= ~XBF_READ;
 661                 }
 662         }
 663
 664         return bp;
 665
 666  no_buffer:
 667         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 668                 xfs_buf_unlock(bp);
 669         xfs_buf_rele(bp);
 670         return NULL;
 671 }
 672
 673 /*
 674  *      If we are not low on memory then do the readahead in a deadlock
 675  *      safe manner.
 676  */
 677 void
 678 xfs_buf_readahead(
 679         xfs_buftarg_t           *target,
 680         xfs_off_t               ioff,
 681         size_t                  isize,
 682         xfs_buf_flags_t         flags)
 683 {
 684         struct backing_dev_info *bdi;
 685
 686         bdi = target->bt_mapping->backing_dev_info;
 687         if (bdi_read_congested(bdi))
 688                 return;
 689
 690         flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
 691         xfs_buf_read_flags(target, ioff, isize, flags);
 692 }
 693
 694 xfs_buf_t *
 695 xfs_buf_get_empty(
 696         size_t                  len,
 697         xfs_buftarg_t           *target)
 698 {
 699         xfs_buf_t               *bp;
 700
 701         bp = xfs_buf_allocate(0);
 702         if (bp)
 703                 _xfs_buf_initialize(bp, target, 0, len, 0);
 704         return bp;
 705 }
 706
 707 static inline struct page *
 708 mem_to_page(
 709         void                    *addr)
 710 {
 711         if ((!is_vmalloc_addr(addr))) {
 712                 return virt_to_page(addr);
 713         } else {
 714                 return vmalloc_to_page(addr);
 715         }
 716 }
 717
 718 int
 719 xfs_buf_associate_memory(
 720         xfs_buf_t               *bp,
 721         void                    *mem,
 722         size_t                  len)
 723 {
 724         int                     rval;
 725         int                     i = 0;
 726         unsigned long           pageaddr;
 727         unsigned long           offset;
 728         size_t                  buflen;
 729         int                     page_count;
 730
 731         pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
 732         offset = (unsigned long)mem - pageaddr;
 733         buflen = PAGE_CACHE_ALIGN(len + offset);
 734         page_count = buflen >> PAGE_CACHE_SHIFT;
 735
 736         /* Free any previous set of page pointers */
 737         if (bp->b_pages)
 738                 _xfs_buf_free_pages(bp);
 739
 740         bp->b_pages = NULL;
 741         bp->b_addr = mem;
 742
 743         rval = _xfs_buf_get_pages(bp, page_count, 0);
 744         if (rval)
 745                 return rval;
 746
 747         bp->b_offset = offset;
 748
 749         for (i = 0; i < bp->b_page_count; i++) {
 750                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 751                 pageaddr += PAGE_CACHE_SIZE;
 752         }
 753
 754         bp->b_count_desired = len;
 755         bp->b_buffer_length = buflen;
 756         bp->b_flags |= XBF_MAPPED;
 757         bp->b_flags &= ~_XBF_PAGE_LOCKED;
 758
 759         return 0;
 760 }
 761
 762 xfs_buf_t *
 763 xfs_buf_get_noaddr(
 764         size_t                  len,
 765         xfs_buftarg_t           *target)
 766 {
 767         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 768         int                     error, i;
 769         xfs_buf_t               *bp;
 770
 771         bp = xfs_buf_allocate(0);
 772         if (unlikely(bp == NULL))
 773                 goto fail;
 774         _xfs_buf_initialize(bp, target, 0, len, 0);
 775
 776         error = _xfs_buf_get_pages(bp, page_count, 0);
 777         if (error)
 778                 goto fail_free_buf;
 779
 780         for (i = 0; i < page_count; i++) {
 781                 bp->b_pages[i] = alloc_page(GFP_KERNEL);
 782                 if (!bp->b_pages[i])
 783                         goto fail_free_mem;
 784         }
 785         bp->b_flags |= _XBF_PAGES;
 786
 787         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 788         if (unlikely(error)) {
 789                 printk(KERN_WARNING "%s: failed to map pages\n",
 790                                 __func__);
 791                 goto fail_free_mem;
 792         }
 793
 794         xfs_buf_unlock(bp);
 795
 796         XB_TRACE(bp, "no_daddr", len);
 797         return bp;
 798
 799  fail_free_mem:
 800         while (--i >= 0)
 801                 __free_page(bp->b_pages[i]);
 802         _xfs_buf_free_pages(bp);
 803  fail_free_buf:
 804         xfs_buf_deallocate(bp);
 805  fail:
 806         return NULL;
 807 }
 808
 809 /*
 810  *      Increment reference count on buffer, to hold the buffer concurrently
 811  *      with another thread which may release (free) the buffer asynchronously.
 812  *      Must hold the buffer already to call this function.
 813  */
 814 void
 815 xfs_buf_hold(
 816         xfs_buf_t               *bp)
 817 {
 818         atomic_inc(&bp->b_hold);
 819         XB_TRACE(bp, "hold", 0);
 820 }
 821
 822 /*
 823  *      Releases a hold on the specified buffer.  If the
 824  *      the hold count is 1, calls xfs_buf_free.
 825  */
 826 void
 827 xfs_buf_rele(
 828         xfs_buf_t               *bp)
 829 {
 830         xfs_bufhash_t           *hash = bp->b_hash;
 831
 832         XB_TRACE(bp, "rele", bp->b_relse);
 833
 834         if (unlikely(!hash)) {
 835                 ASSERT(!bp->b_relse);
 836                 if (atomic_dec_and_test(&bp->b_hold))
 837                         xfs_buf_free(bp);
 838                 return;
 839         }
 840
 841         if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
 842                 if (bp->b_relse) {
 843                         atomic_inc(&bp->b_hold);
 844                         spin_unlock(&hash->bh_lock);
 845                         (*(bp->b_relse)) (bp);
 846                 } else if (bp->b_flags & XBF_FS_MANAGED) {
 847                         spin_unlock(&hash->bh_lock);
 848                 } else {
 849                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 850                         list_del_init(&bp->b_hash_list);
 851                         spin_unlock(&hash->bh_lock);
 852                         xfs_buf_free(bp);
 853                 }
 854         } else {
 855                 /*
 856                  * Catch reference count leaks
 857                  */
 858                 ASSERT(atomic_read(&bp->b_hold) >= 0);
 859         }
 860 }
 861
 862
 863 /*
 864  *      Mutual exclusion on buffers.  Locking model:
 865  *
 866  *      Buffers associated with inodes for which buffer locking
 867  *      is not enabled are not protected by semaphores, and are
 868  *      assumed to be exclusively owned by the caller.  There is a
 869  *      spinlock in the buffer, used by the caller when concurrent
 870  *      access is possible.
 871  */
 872
 873 /*
 874  *      Locks a buffer object, if it is not already locked.
 875  *      Note that this in no way locks the underlying pages, so it is only
 876  *      useful for synchronizing concurrent use of buffer objects, not for
 877  *      synchronizing independent access to the underlying pages.
 878  */
 879 int
 880 xfs_buf_cond_lock(
 881         xfs_buf_t               *bp)
 882 {
 883         int                     locked;
 884
 885         locked = down_trylock(&bp->b_sema) == 0;
 886         if (locked) {
 887                 XB_SET_OWNER(bp);
 888         }
 889         XB_TRACE(bp, "cond_lock", (long)locked);
 890         return locked ? 0 : -EBUSY;
 891 }
 892
 893 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
 894 int
 895 xfs_buf_lock_value(
 896         xfs_buf_t               *bp)
 897 {
 898         return bp->b_sema.count;
 899 }
 900 #endif
 901
 902 /*
 903  *      Locks a buffer object.
 904  *      Note that this in no way locks the underlying pages, so it is only
 905  *      useful for synchronizing concurrent use of buffer objects, not for
 906  *      synchronizing independent access to the underlying pages.
 907  */
 908 void
 909 xfs_buf_lock(
 910         xfs_buf_t               *bp)
 911 {
 912         XB_TRACE(bp, "lock", 0);
 913         if (atomic_read(&bp->b_io_remaining))
 914                 blk_run_address_space(bp->b_target->bt_mapping);
 915         down(&bp->b_sema);
 916         XB_SET_OWNER(bp);
 917         XB_TRACE(bp, "locked", 0);
 918 }
 919
 920 /*
 921  *      Releases the lock on the buffer object.
 922  *      If the buffer is marked delwri but is not queued, do so before we
 923  *      unlock the buffer as we need to set flags correctly.  We also need to
 924  *      take a reference for the delwri queue because the unlocker is going to
 925  *      drop their's and they don't know we just queued it.
 926  */
 927 void
 928 xfs_buf_unlock(
 929         xfs_buf_t               *bp)
 930 {
 931         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 932                 atomic_inc(&bp->b_hold);
 933                 bp->b_flags |= XBF_ASYNC;
 934                 xfs_buf_delwri_queue(bp, 0);
 935         }
 936
 937         XB_CLEAR_OWNER(bp);
 938         up(&bp->b_sema);
 939         XB_TRACE(bp, "unlock", 0);
 940 }
 941
 942
 943 /*
 944  *      Pinning Buffer Storage in Memory
 945  *      Ensure that no attempt to force a buffer to disk will succeed.
 946  */
 947 void
 948 xfs_buf_pin(
 949         xfs_buf_t               *bp)
 950 {
 951         atomic_inc(&bp->b_pin_count);
 952         XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
 953 }
 954
 955 void
 956 xfs_buf_unpin(
 957         xfs_buf_t               *bp)
 958 {
 959         if (atomic_dec_and_test(&bp->b_pin_count))
 960                 wake_up_all(&bp->b_waiters);
 961         XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
 962 }
 963
 964 int
 965 xfs_buf_ispin(
 966         xfs_buf_t               *bp)
 967 {
 968         return atomic_read(&bp->b_pin_count);
 969 }
 970
 971 STATIC void
 972 xfs_buf_wait_unpin(
 973         xfs_buf_t               *bp)
 974 {
 975         DECLARE_WAITQUEUE       (wait, current);
 976
 977         if (atomic_read(&bp->b_pin_count) == 0)
 978                 return;
 979
 980         add_wait_queue(&bp->b_waiters, &wait);
 981         for (;;) {
 982                 set_current_state(TASK_UNINTERRUPTIBLE);
 983                 if (atomic_read(&bp->b_pin_count) == 0)
 984                         break;
 985                 if (atomic_read(&bp->b_io_remaining))
 986                         blk_run_address_space(bp->b_target->bt_mapping);
 987                 schedule();
 988         }
 989         remove_wait_queue(&bp->b_waiters, &wait);
 990         set_current_state(TASK_RUNNING);
 991 }
 992
 993 /*
 994  *      Buffer Utility Routines
 995  */
 996
 997 STATIC void
 998 xfs_buf_iodone_work(
 999         struct work_struct      *work)
1000 {
1001         xfs_buf_t               *bp =
1002                 container_of(work, xfs_buf_t, b_iodone_work);
1003
1004         /*
1005          * We can get an EOPNOTSUPP to ordered writes.  Here we clear the
1006          * ordered flag and reissue them.  Because we can't tell the higher
1007          * layers directly that they should not issue ordered I/O anymore, they
1008          * need to check if the ordered flag was cleared during I/O completion.
1009          */
1010         if ((bp->b_error == EOPNOTSUPP) &&
1011             (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1012                 XB_TRACE(bp, "ordered_retry", bp->b_iodone);
1013                 bp->b_flags &= ~XBF_ORDERED;
1014                 xfs_buf_iorequest(bp);
1015         } else if (bp->b_iodone)
1016                 (*(bp->b_iodone))(bp);
1017         else if (bp->b_flags & XBF_ASYNC)
1018                 xfs_buf_relse(bp);
1019 }
1020
1021 void
1022 xfs_buf_ioend(
1023         xfs_buf_t               *bp,
1024         int                     schedule)
1025 {
1026         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1027         if (bp->b_error == 0)
1028                 bp->b_flags |= XBF_DONE;
1029
1030         XB_TRACE(bp, "iodone", bp->b_iodone);
1031
1032         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1033                 if (schedule) {
1034                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1035                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1036                 } else {
1037                         xfs_buf_iodone_work(&bp->b_iodone_work);
1038                 }
1039         } else {
1040                 up(&bp->b_iodonesema);
1041         }
1042 }
1043
1044 void
1045 xfs_buf_ioerror(
1046         xfs_buf_t               *bp,
1047         int                     error)
1048 {
1049         ASSERT(error >= 0 && error <= 0xffff);
1050         bp->b_error = (unsigned short)error;
1051         XB_TRACE(bp, "ioerror", (unsigned long)error);
1052 }
1053
1054 /*
1055  *      Initiate I/O on a buffer, based on the flags supplied.
1056  *      The b_iodone routine in the buffer supplied will only be called
1057  *      when all of the subsidiary I/O requests, if any, have been completed.
1058  */
1059 int
1060 xfs_buf_iostart(
1061         xfs_buf_t               *bp,
1062         xfs_buf_flags_t         flags)
1063 {
1064         int                     status = 0;
1065
1066         XB_TRACE(bp, "iostart", (unsigned long)flags);
1067
1068         if (flags & XBF_DELWRI) {
1069                 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1070                 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1071                 xfs_buf_delwri_queue(bp, 1);
1072                 return 0;
1073         }
1074
1075         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1076                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1077         bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1078                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1079
1080         BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1081
1082         /* For writes allow an alternate strategy routine to precede
1083          * the actual I/O request (which may not be issued at all in
1084          * a shutdown situation, for example).
1085          */
1086         status = (flags & XBF_WRITE) ?
1087                 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1088
1089         /* Wait for I/O if we are not an async request.
1090          * Note: async I/O request completion will release the buffer,
1091          * and that can already be done by this point.  So using the
1092          * buffer pointer from here on, after async I/O, is invalid.
1093          */
1094         if (!status && !(flags & XBF_ASYNC))
1095                 status = xfs_buf_iowait(bp);
1096
1097         return status;
1098 }
1099
1100 STATIC_INLINE void
1101 _xfs_buf_ioend(
1102         xfs_buf_t               *bp,
1103         int                     schedule)
1104 {
1105         if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1106                 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1107                 xfs_buf_ioend(bp, schedule);
1108         }
1109 }
1110
1111 STATIC void
1112 xfs_buf_bio_end_io(
1113         struct bio              *bio,
1114         int                     error)
1115 {
1116         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1117         unsigned int            blocksize = bp->b_target->bt_bsize;
1118         struct bio_vec          *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1119
1120         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1121                 bp->b_error = EIO;
1122
1123         do {
1124                 struct page     *page = bvec->bv_page;
1125
1126                 ASSERT(!PagePrivate(page));
1127                 if (unlikely(bp->b_error)) {
1128                         if (bp->b_flags & XBF_READ)
1129                                 ClearPageUptodate(page);
1130                 } else if (blocksize >= PAGE_CACHE_SIZE) {
1131                         SetPageUptodate(page);
1132                 } else if (!PagePrivate(page) &&
1133                                 (bp->b_flags & _XBF_PAGE_CACHE)) {
1134                         set_page_region(page, bvec->bv_offset, bvec->bv_len);
1135                 }
1136
1137                 if (--bvec >= bio->bi_io_vec)
1138                         prefetchw(&bvec->bv_page->flags);
1139
1140                 if (bp->b_flags & _XBF_PAGE_LOCKED)
1141                         unlock_page(page);
1142         } while (bvec >= bio->bi_io_vec);
1143
1144         _xfs_buf_ioend(bp, 1);
1145         bio_put(bio);
1146 }
1147
1148 STATIC void
1149 _xfs_buf_ioapply(
1150         xfs_buf_t               *bp)
1151 {
1152         int                     rw, map_i, total_nr_pages, nr_pages;
1153         struct bio              *bio;
1154         int                     offset = bp->b_offset;
1155         int                     size = bp->b_count_desired;
1156         sector_t                sector = bp->b_bn;
1157         unsigned int            blocksize = bp->b_target->bt_bsize;
1158
1159         total_nr_pages = bp->b_page_count;
1160         map_i = 0;
1161
1162         if (bp->b_flags & XBF_ORDERED) {
1163                 ASSERT(!(bp->b_flags & XBF_READ));
1164                 rw = WRITE_BARRIER;
1165         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1166                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1167                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1168                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1169         } else {
1170                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1171                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1172         }
1173
1174         /* Special code path for reading a sub page size buffer in --
1175          * we populate up the whole page, and hence the other metadata
1176          * in the same page.  This optimization is only valid when the
1177          * filesystem block size is not smaller than the page size.
1178          */
1179         if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1180             ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1181               (XBF_READ|_XBF_PAGE_LOCKED)) &&
1182             (blocksize >= PAGE_CACHE_SIZE)) {
1183                 bio = bio_alloc(GFP_NOIO, 1);
1184
1185                 bio->bi_bdev = bp->b_target->bt_bdev;
1186                 bio->bi_sector = sector - (offset >> BBSHIFT);
1187                 bio->bi_end_io = xfs_buf_bio_end_io;
1188                 bio->bi_private = bp;
1189
1190                 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1191                 size = 0;
1192
1193                 atomic_inc(&bp->b_io_remaining);
1194
1195                 goto submit_io;
1196         }
1197
1198 next_chunk:
1199         atomic_inc(&bp->b_io_remaining);
1200         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1201         if (nr_pages > total_nr_pages)
1202                 nr_pages = total_nr_pages;
1203
1204         bio = bio_alloc(GFP_NOIO, nr_pages);
1205         bio->bi_bdev = bp->b_target->bt_bdev;
1206         bio->bi_sector = sector;
1207         bio->bi_end_io = xfs_buf_bio_end_io;
1208         bio->bi_private = bp;
1209
1210         for (; size && nr_pages; nr_pages--, map_i++) {
1211                 int     rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1212
1213                 if (nbytes > size)
1214                         nbytes = size;
1215
1216                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1217                 if (rbytes < nbytes)
1218                         break;
1219
1220                 offset = 0;
1221                 sector += nbytes >> BBSHIFT;
1222                 size -= nbytes;
1223                 total_nr_pages--;
1224         }
1225
1226 submit_io:
1227         if (likely(bio->bi_size)) {
1228                 submit_bio(rw, bio);
1229                 if (size)
1230                         goto next_chunk;
1231         } else {
1232                 bio_put(bio);
1233                 xfs_buf_ioerror(bp, EIO);
1234         }
1235 }
1236
1237 int
1238 xfs_buf_iorequest(
1239         xfs_buf_t               *bp)
1240 {
1241         XB_TRACE(bp, "iorequest", 0);
1242
1243         if (bp->b_flags & XBF_DELWRI) {
1244                 xfs_buf_delwri_queue(bp, 1);
1245                 return 0;
1246         }
1247
1248         if (bp->b_flags & XBF_WRITE) {
1249                 xfs_buf_wait_unpin(bp);
1250         }
1251
1252         xfs_buf_hold(bp);
1253
1254         /* Set the count to 1 initially, this will stop an I/O
1255          * completion callout which happens before we have started
1256          * all the I/O from calling xfs_buf_ioend too early.
1257          */
1258         atomic_set(&bp->b_io_remaining, 1);
1259         _xfs_buf_ioapply(bp);
1260         _xfs_buf_ioend(bp, 0);
1261
1262         xfs_buf_rele(bp);
1263         return 0;
1264 }
1265
1266 /*
1267  *      Waits for I/O to complete on the buffer supplied.
1268  *      It returns immediately if no I/O is pending.
1269  *      It returns the I/O error code, if any, or 0 if there was no error.
1270  */
1271 int
1272 xfs_buf_iowait(
1273         xfs_buf_t               *bp)
1274 {
1275         XB_TRACE(bp, "iowait", 0);
1276         if (atomic_read(&bp->b_io_remaining))
1277                 blk_run_address_space(bp->b_target->bt_mapping);
1278         down(&bp->b_iodonesema);
1279         XB_TRACE(bp, "iowaited", (long)bp->b_error);
1280         return bp->b_error;
1281 }
1282
1283 xfs_caddr_t
1284 xfs_buf_offset(
1285         xfs_buf_t               *bp,
1286         size_t                  offset)
1287 {
1288         struct page             *page;
1289
1290         if (bp->b_flags & XBF_MAPPED)
1291                 return XFS_BUF_PTR(bp) + offset;
1292
1293         offset += bp->b_offset;
1294         page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1295         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1296 }
1297
1298 /*
1299  *      Move data into or out of a buffer.
1300  */
1301 void
1302 xfs_buf_iomove(
1303         xfs_buf_t               *bp,    /* buffer to process            */
1304         size_t                  boff,   /* starting buffer offset       */
1305         size_t                  bsize,  /* length to copy               */
1306         caddr_t                 data,   /* data address                 */
1307         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1308 {
1309         size_t                  bend, cpoff, csize;
1310         struct page             *page;
1311
1312         bend = boff + bsize;
1313         while (boff < bend) {
1314                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1315                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1316                 csize = min_t(size_t,
1317                               PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1318
1319                 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1320
1321                 switch (mode) {
1322                 case XBRW_ZERO:
1323                         memset(page_address(page) + cpoff, 0, csize);
1324                         break;
1325                 case XBRW_READ:
1326                         memcpy(data, page_address(page) + cpoff, csize);
1327                         break;
1328                 case XBRW_WRITE:
1329                         memcpy(page_address(page) + cpoff, data, csize);
1330                 }
1331
1332                 boff += csize;
1333                 data += csize;
1334         }
1335 }
1336
1337 /*
1338  *      Handling of buffer targets (buftargs).
1339  */
1340
1341 /*
1342  *      Wait for any bufs with callbacks that have been submitted but
1343  *      have not yet returned... walk the hash list for the target.
1344  */
1345 void
1346 xfs_wait_buftarg(
1347         xfs_buftarg_t   *btp)
1348 {
1349         xfs_buf_t       *bp, *n;
1350         xfs_bufhash_t   *hash;
1351         uint            i;
1352
1353         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1354                 hash = &btp->bt_hash[i];
1355 again:
1356                 spin_lock(&hash->bh_lock);
1357                 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1358                         ASSERT(btp == bp->b_target);
1359                         if (!(bp->b_flags & XBF_FS_MANAGED)) {
1360                                 spin_unlock(&hash->bh_lock);
1361                                 /*
1362                                  * Catch superblock reference count leaks
1363                                  * immediately
1364                                  */
1365                                 BUG_ON(bp->b_bn == 0);
1366                                 delay(100);
1367                                 goto again;
1368                         }
1369                 }
1370                 spin_unlock(&hash->bh_lock);
1371         }
1372 }
1373
1374 /*
1375  *      Allocate buffer hash table for a given target.
1376  *      For devices containing metadata (i.e. not the log/realtime devices)
1377  *      we need to allocate a much larger hash table.
1378  */
1379 STATIC void
1380 xfs_alloc_bufhash(
1381         xfs_buftarg_t           *btp,
1382         int                     external)
1383 {
1384         unsigned int            i;
1385
1386         btp->bt_hashshift = external ? 3 : 8;   /* 8 or 256 buckets */
1387         btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1388         btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1389                                         sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1390         for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1391                 spin_lock_init(&btp->bt_hash[i].bh_lock);
1392                 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1393         }
1394 }
1395
1396 STATIC void
1397 xfs_free_bufhash(
1398         xfs_buftarg_t           *btp)
1399 {
1400         kmem_free(btp->bt_hash);
1401         btp->bt_hash = NULL;
1402 }
1403
1404 /*
1405  *      buftarg list for delwrite queue processing
1406  */
1407 static LIST_HEAD(xfs_buftarg_list);
1408 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1409
1410 STATIC void
1411 xfs_register_buftarg(
1412         xfs_buftarg_t           *btp)
1413 {
1414         spin_lock(&xfs_buftarg_lock);
1415         list_add(&btp->bt_list, &xfs_buftarg_list);
1416         spin_unlock(&xfs_buftarg_lock);
1417 }
1418
1419 STATIC void
1420 xfs_unregister_buftarg(
1421         xfs_buftarg_t           *btp)
1422 {
1423         spin_lock(&xfs_buftarg_lock);
1424         list_del(&btp->bt_list);
1425         spin_unlock(&xfs_buftarg_lock);
1426 }
1427
1428 void
1429 xfs_free_buftarg(
1430         xfs_buftarg_t           *btp)
1431 {
1432         xfs_flush_buftarg(btp, 1);
1433         xfs_blkdev_issue_flush(btp);
1434         xfs_free_bufhash(btp);
1435         iput(btp->bt_mapping->host);
1436
1437         /* Unregister the buftarg first so that we don't get a
1438          * wakeup finding a non-existent task
1439          */
1440         xfs_unregister_buftarg(btp);
1441         kthread_stop(btp->bt_task);
1442
1443         kmem_free(btp);
1444 }
1445
1446 STATIC int
1447 xfs_setsize_buftarg_flags(
1448         xfs_buftarg_t           *btp,
1449         unsigned int            blocksize,
1450         unsigned int            sectorsize,
1451         int                     verbose)
1452 {
1453         btp->bt_bsize = blocksize;
1454         btp->bt_sshift = ffs(sectorsize) - 1;
1455         btp->bt_smask = sectorsize - 1;
1456
1457         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1458                 printk(KERN_WARNING
1459                         "XFS: Cannot set_blocksize to %u on device %s\n",
1460                         sectorsize, XFS_BUFTARG_NAME(btp));
1461                 return EINVAL;
1462         }
1463
1464         if (verbose &&
1465             (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1466                 printk(KERN_WARNING
1467                         "XFS: %u byte sectors in use on device %s.  "
1468                         "This is suboptimal; %u or greater is ideal.\n",
1469                         sectorsize, XFS_BUFTARG_NAME(btp),
1470                         (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1471         }
1472
1473         return 0;
1474 }
1475
1476 /*
1477  *      When allocating the initial buffer target we have not yet
1478  *      read in the superblock, so don't know what sized sectors
1479  *      are being used is at this early stage.  Play safe.
1480  */
1481 STATIC int
1482 xfs_setsize_buftarg_early(
1483         xfs_buftarg_t           *btp,
1484         struct block_device     *bdev)
1485 {
1486         return xfs_setsize_buftarg_flags(btp,
1487                         PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1488 }
1489
1490 int
1491 xfs_setsize_buftarg(
1492         xfs_buftarg_t           *btp,
1493         unsigned int            blocksize,
1494         unsigned int            sectorsize)
1495 {
1496         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1497 }
1498
1499 STATIC int
1500 xfs_mapping_buftarg(
1501         xfs_buftarg_t           *btp,
1502         struct block_device     *bdev)
1503 {
1504         struct backing_dev_info *bdi;
1505         struct inode            *inode;
1506         struct address_space    *mapping;
1507         static const struct address_space_operations mapping_aops = {
1508                 .sync_page = block_sync_page,
1509                 .migratepage = fail_migrate_page,
1510         };
1511
1512         inode = new_inode(bdev->bd_inode->i_sb);
1513         if (!inode) {
1514                 printk(KERN_WARNING
1515                         "XFS: Cannot allocate mapping inode for device %s\n",
1516                         XFS_BUFTARG_NAME(btp));
1517                 return ENOMEM;
1518         }
1519         inode->i_mode = S_IFBLK;
1520         inode->i_bdev = bdev;
1521         inode->i_rdev = bdev->bd_dev;
1522         bdi = blk_get_backing_dev_info(bdev);
1523         if (!bdi)
1524                 bdi = &default_backing_dev_info;
1525         mapping = &inode->i_data;
1526         mapping->a_ops = &mapping_aops;
1527         mapping->backing_dev_info = bdi;
1528         mapping_set_gfp_mask(mapping, GFP_NOFS);
1529         btp->bt_mapping = mapping;
1530         return 0;
1531 }
1532
1533 STATIC int
1534 xfs_alloc_delwrite_queue(
1535         xfs_buftarg_t           *btp)
1536 {
1537         int     error = 0;
1538
1539         INIT_LIST_HEAD(&btp->bt_list);
1540         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1541         spin_lock_init(&btp->bt_delwrite_lock);
1542         btp->bt_flags = 0;
1543         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1544         if (IS_ERR(btp->bt_task)) {
1545                 error = PTR_ERR(btp->bt_task);
1546                 goto out_error;
1547         }
1548         xfs_register_buftarg(btp);
1549 out_error:
1550         return error;
1551 }
1552
1553 xfs_buftarg_t *
1554 xfs_alloc_buftarg(
1555         struct block_device     *bdev,
1556         int                     external)
1557 {
1558         xfs_buftarg_t           *btp;
1559
1560         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1561
1562         btp->bt_dev =  bdev->bd_dev;
1563         btp->bt_bdev = bdev;
1564         if (xfs_setsize_buftarg_early(btp, bdev))
1565                 goto error;
1566         if (xfs_mapping_buftarg(btp, bdev))
1567                 goto error;
1568         if (xfs_alloc_delwrite_queue(btp))
1569                 goto error;
1570         xfs_alloc_bufhash(btp, external);
1571         return btp;
1572
1573 error:
1574         kmem_free(btp);
1575         return NULL;
1576 }
1577
1578
1579 /*
1580  *      Delayed write buffer handling
1581  */
1582 STATIC void
1583 xfs_buf_delwri_queue(
1584         xfs_buf_t               *bp,
1585         int                     unlock)
1586 {
1587         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1588         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1589
1590         XB_TRACE(bp, "delwri_q", (long)unlock);
1591         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1592
1593         spin_lock(dwlk);
1594         /* If already in the queue, dequeue and place at tail */
1595         if (!list_empty(&bp->b_list)) {
1596                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1597                 if (unlock)
1598                         atomic_dec(&bp->b_hold);
1599                 list_del(&bp->b_list);
1600         }
1601
1602         bp->b_flags |= _XBF_DELWRI_Q;
1603         list_add_tail(&bp->b_list, dwq);
1604         bp->b_queuetime = jiffies;
1605         spin_unlock(dwlk);
1606
1607         if (unlock)
1608                 xfs_buf_unlock(bp);
1609 }
1610
1611 void
1612 xfs_buf_delwri_dequeue(
1613         xfs_buf_t               *bp)
1614 {
1615         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1616         int                     dequeued = 0;
1617
1618         spin_lock(dwlk);
1619         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1620                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1621                 list_del_init(&bp->b_list);
1622                 dequeued = 1;
1623         }
1624         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1625         spin_unlock(dwlk);
1626
1627         if (dequeued)
1628                 xfs_buf_rele(bp);
1629
1630         XB_TRACE(bp, "delwri_dq", (long)dequeued);
1631 }
1632
1633 STATIC void
1634 xfs_buf_runall_queues(
1635         struct workqueue_struct *queue)
1636 {
1637         flush_workqueue(queue);
1638 }
1639
1640 STATIC int
1641 xfsbufd_wakeup(
1642         int                     priority,
1643         gfp_t                   mask)
1644 {
1645         xfs_buftarg_t           *btp;
1646
1647         spin_lock(&xfs_buftarg_lock);
1648         list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1649                 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1650                         continue;
1651                 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1652                 wake_up_process(btp->bt_task);
1653         }
1654         spin_unlock(&xfs_buftarg_lock);
1655         return 0;
1656 }
1657
1658 /*
1659  * Move as many buffers as specified to the supplied list
1660  * idicating if we skipped any buffers to prevent deadlocks.
1661  */
1662 STATIC int
1663 xfs_buf_delwri_split(
1664         xfs_buftarg_t   *target,
1665         struct list_head *list,
1666         unsigned long   age)
1667 {
1668         xfs_buf_t       *bp, *n;
1669         struct list_head *dwq = &target->bt_delwrite_queue;
1670         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1671         int             skipped = 0;
1672         int             force;
1673
1674         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1675         INIT_LIST_HEAD(list);
1676         spin_lock(dwlk);
1677         list_for_each_entry_safe(bp, n, dwq, b_list) {
1678                 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1679                 ASSERT(bp->b_flags & XBF_DELWRI);
1680
1681                 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1682                         if (!force &&
1683                             time_before(jiffies, bp->b_queuetime + age)) {
1684                                 xfs_buf_unlock(bp);
1685                                 break;
1686                         }
1687
1688                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1689                                          _XBF_RUN_QUEUES);
1690                         bp->b_flags |= XBF_WRITE;
1691                         list_move_tail(&bp->b_list, list);
1692                 } else
1693                         skipped++;
1694         }
1695         spin_unlock(dwlk);
1696
1697         return skipped;
1698
1699 }
1700
1701 STATIC int
1702 xfsbufd(
1703         void            *data)
1704 {
1705         struct list_head tmp;
1706         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1707         int             count;
1708         xfs_buf_t       *bp;
1709
1710         current->flags |= PF_MEMALLOC;
1711
1712         set_freezable();
1713
1714         do {
1715                 if (unlikely(freezing(current))) {
1716                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1717                         refrigerator();
1718                 } else {
1719                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1720                 }
1721
1722                 schedule_timeout_interruptible(
1723                         xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1724
1725                 xfs_buf_delwri_split(target, &tmp,
1726                                 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1727
1728                 count = 0;
1729                 while (!list_empty(&tmp)) {
1730                         bp = list_entry(tmp.next, xfs_buf_t, b_list);
1731                         ASSERT(target == bp->b_target);
1732
1733                         list_del_init(&bp->b_list);
1734                         xfs_buf_iostrategy(bp);
1735                         count++;
1736                 }
1737
1738                 if (as_list_len > 0)
1739                         purge_addresses();
1740                 if (count)
1741                         blk_run_address_space(target->bt_mapping);
1742
1743         } while (!kthread_should_stop());
1744
1745         return 0;
1746 }
1747
1748 /*
1749  *      Go through all incore buffers, and release buffers if they belong to
1750  *      the given device. This is used in filesystem error handling to
1751  *      preserve the consistency of its metadata.
1752  */
1753 int
1754 xfs_flush_buftarg(
1755         xfs_buftarg_t   *target,
1756         int             wait)
1757 {
1758         struct list_head tmp;
1759         xfs_buf_t       *bp, *n;
1760         int             pincount = 0;
1761
1762         xfs_buf_runall_queues(xfsdatad_workqueue);
1763         xfs_buf_runall_queues(xfslogd_workqueue);
1764
1765         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1766         pincount = xfs_buf_delwri_split(target, &tmp, 0);
1767
1768         /*
1769          * Dropped the delayed write list lock, now walk the temporary list
1770          */
1771         list_for_each_entry_safe(bp, n, &tmp, b_list) {
1772                 ASSERT(target == bp->b_target);
1773                 if (wait)
1774                         bp->b_flags &= ~XBF_ASYNC;
1775                 else
1776                         list_del_init(&bp->b_list);
1777
1778                 xfs_buf_iostrategy(bp);
1779         }
1780
1781         if (wait)
1782                 blk_run_address_space(target->bt_mapping);
1783
1784         /*
1785          * Remaining list items must be flushed before returning
1786          */
1787         while (!list_empty(&tmp)) {
1788                 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1789
1790                 list_del_init(&bp->b_list);
1791                 xfs_iowait(bp);
1792                 xfs_buf_relse(bp);
1793         }
1794
1795         return pincount;
1796 }
1797
1798 int __init
1799 xfs_buf_init(void)
1800 {
1801 #ifdef XFS_BUF_TRACE
1802         xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1803 #endif
1804
1805         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1806                                                 KM_ZONE_HWALIGN, NULL);
1807         if (!xfs_buf_zone)
1808                 goto out_free_trace_buf;
1809
1810         xfslogd_workqueue = create_workqueue("xfslogd");
1811         if (!xfslogd_workqueue)
1812                 goto out_free_buf_zone;
1813
1814         xfsdatad_workqueue = create_workqueue("xfsdatad");
1815         if (!xfsdatad_workqueue)
1816                 goto out_destroy_xfslogd_workqueue;
1817
1818         register_shrinker(&xfs_buf_shake);
1819         return 0;
1820
1821  out_destroy_xfslogd_workqueue:
1822         destroy_workqueue(xfslogd_workqueue);
1823  out_free_buf_zone:
1824         kmem_zone_destroy(xfs_buf_zone);
1825  out_free_trace_buf:
1826 #ifdef XFS_BUF_TRACE
1827         ktrace_free(xfs_buf_trace_buf);
1828 #endif
1829         return -ENOMEM;
1830 }
1831
1832 void
1833 xfs_buf_terminate(void)
1834 {
1835         unregister_shrinker(&xfs_buf_shake);
1836         destroy_workqueue(xfsdatad_workqueue);
1837         destroy_workqueue(xfslogd_workqueue);
1838         kmem_zone_destroy(xfs_buf_zone);
1839 #ifdef XFS_BUF_TRACE
1840         ktrace_free(xfs_buf_trace_buf);
1841 #endif
1842 }
1843
1844 #ifdef CONFIG_KDB_MODULES
1845 struct list_head *
1846 xfs_get_buftarg_list(void)
1847 {
1848         return &xfs_buftarg_list;
1849 }
1850 #endif