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/gfp.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 #include <linux/list_sort.h>
  37
  38 #include "xfs_sb.h"
  39 #include "xfs_inum.h"
  40 #include "xfs_log.h"
  41 #include "xfs_ag.h"
  42 #include "xfs_mount.h"
  43 #include "xfs_trace.h"
  44
  45 static kmem_zone_t *xfs_buf_zone;
  46 STATIC int xfsbufd(void *);
  47 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
  48
  49 static struct workqueue_struct *xfslogd_workqueue;
  50 struct workqueue_struct *xfsdatad_workqueue;
  51 struct workqueue_struct *xfsconvertd_workqueue;
  52
  53 #ifdef XFS_BUF_LOCK_TRACKING
  54 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  55 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  56 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  57 #else
  58 # define XB_SET_OWNER(bp)       do { } while (0)
  59 # define XB_CLEAR_OWNER(bp)     do { } while (0)
  60 # define XB_GET_OWNER(bp)       do { } while (0)
  61 #endif
  62
  63 #define xb_to_gfp(flags) \
  64         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
  65           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
  66
  67 #define xb_to_km(flags) \
  68          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
  69
  70 #define xfs_buf_allocate(flags) \
  71         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
  72 #define xfs_buf_deallocate(bp) \
  73         kmem_zone_free(xfs_buf_zone, (bp));
  74
  75 static inline int
  76 xfs_buf_is_vmapped(
  77         struct xfs_buf  *bp)
  78 {
  79         /*
  80          * Return true if the buffer is vmapped.
  81          *
  82          * The XBF_MAPPED flag is set if the buffer should be mapped, but the
  83          * code is clever enough to know it doesn't have to map a single page,
  84          * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
  85          */
  86         return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
  87 }
  88
  89 static inline int
  90 xfs_buf_vmap_len(
  91         struct xfs_buf  *bp)
  92 {
  93         return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
  94 }
  95
  96 /*
  97  * xfs_buf_lru_add - add a buffer to the LRU.
  98  *
  99  * The LRU takes a new reference to the buffer so that it will only be freed
 100  * once the shrinker takes the buffer off the LRU.
 101  */
 102 STATIC void
 103 xfs_buf_lru_add(
 104         struct xfs_buf  *bp)
 105 {
 106         struct xfs_buftarg *btp = bp->b_target;
 107
 108         spin_lock(&btp->bt_lru_lock);
 109         if (list_empty(&bp->b_lru)) {
 110                 atomic_inc(&bp->b_hold);
 111                 list_add_tail(&bp->b_lru, &btp->bt_lru);
 112                 btp->bt_lru_nr++;
 113         }
 114         spin_unlock(&btp->bt_lru_lock);
 115 }
 116
 117 /*
 118  * xfs_buf_lru_del - remove a buffer from the LRU
 119  *
 120  * The unlocked check is safe here because it only occurs when there are not
 121  * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
 122  * to optimise the shrinker removing the buffer from the LRU and calling
 123  * xfs_buf_free(). i.e. it removes an unneccessary round trip on the
 124  * bt_lru_lock.
 125  */
 126 STATIC void
 127 xfs_buf_lru_del(
 128         struct xfs_buf  *bp)
 129 {
 130         struct xfs_buftarg *btp = bp->b_target;
 131
 132         if (list_empty(&bp->b_lru))
 133                 return;
 134
 135         spin_lock(&btp->bt_lru_lock);
 136         if (!list_empty(&bp->b_lru)) {
 137                 list_del_init(&bp->b_lru);
 138                 btp->bt_lru_nr--;
 139         }
 140         spin_unlock(&btp->bt_lru_lock);
 141 }
 142
 143 /*
 144  * When we mark a buffer stale, we remove the buffer from the LRU and clear the
 145  * b_lru_ref count so that the buffer is freed immediately when the buffer
 146  * reference count falls to zero. If the buffer is already on the LRU, we need
 147  * to remove the reference that LRU holds on the buffer.
 148  *
 149  * This prevents build-up of stale buffers on the LRU.
 150  */
 151 void
 152 xfs_buf_stale(
 153         struct xfs_buf  *bp)
 154 {
 155         bp->b_flags |= XBF_STALE;
 156         atomic_set(&(bp)->b_lru_ref, 0);
 157         if (!list_empty(&bp->b_lru)) {
 158                 struct xfs_buftarg *btp = bp->b_target;
 159
 160                 spin_lock(&btp->bt_lru_lock);
 161                 if (!list_empty(&bp->b_lru)) {
 162                         list_del_init(&bp->b_lru);
 163                         btp->bt_lru_nr--;
 164                         atomic_dec(&bp->b_hold);
 165                 }
 166                 spin_unlock(&btp->bt_lru_lock);
 167         }
 168         ASSERT(atomic_read(&bp->b_hold) >= 1);
 169 }
 170
 171 STATIC void
 172 _xfs_buf_initialize(
 173         xfs_buf_t               *bp,
 174         xfs_buftarg_t           *target,
 175         xfs_off_t               range_base,
 176         size_t                  range_length,
 177         xfs_buf_flags_t         flags)
 178 {
 179         /*
 180          * We don't want certain flags to appear in b_flags.
 181          */
 182         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
 183
 184         memset(bp, 0, sizeof(xfs_buf_t));
 185         atomic_set(&bp->b_hold, 1);
 186         atomic_set(&bp->b_lru_ref, 1);
 187         init_completion(&bp->b_iowait);
 188         INIT_LIST_HEAD(&bp->b_lru);
 189         INIT_LIST_HEAD(&bp->b_list);
 190         RB_CLEAR_NODE(&bp->b_rbnode);
 191         sema_init(&bp->b_sema, 0); /* held, no waiters */
 192         XB_SET_OWNER(bp);
 193         bp->b_target = target;
 194         bp->b_file_offset = range_base;
 195         /*
 196          * Set buffer_length and count_desired to the same value initially.
 197          * I/O routines should use count_desired, which will be the same in
 198          * most cases but may be reset (e.g. XFS recovery).
 199          */
 200         bp->b_buffer_length = bp->b_count_desired = range_length;
 201         bp->b_flags = flags;
 202         bp->b_bn = XFS_BUF_DADDR_NULL;
 203         atomic_set(&bp->b_pin_count, 0);
 204         init_waitqueue_head(&bp->b_waiters);
 205
 206         XFS_STATS_INC(xb_create);
 207
 208         trace_xfs_buf_init(bp, _RET_IP_);
 209 }
 210
 211 /*
 212  *      Allocate a page array capable of holding a specified number
 213  *      of pages, and point the page buf at it.
 214  */
 215 STATIC int
 216 _xfs_buf_get_pages(
 217         xfs_buf_t               *bp,
 218         int                     page_count,
 219         xfs_buf_flags_t         flags)
 220 {
 221         /* Make sure that we have a page list */
 222         if (bp->b_pages == NULL) {
 223                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
 224                 bp->b_page_count = page_count;
 225                 if (page_count <= XB_PAGES) {
 226                         bp->b_pages = bp->b_page_array;
 227                 } else {
 228                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
 229                                         page_count, xb_to_km(flags));
 230                         if (bp->b_pages == NULL)
 231                                 return -ENOMEM;
 232                 }
 233                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 234         }
 235         return 0;
 236 }
 237
 238 /*
 239  *      Frees b_pages if it was allocated.
 240  */
 241 STATIC void
 242 _xfs_buf_free_pages(
 243         xfs_buf_t       *bp)
 244 {
 245         if (bp->b_pages != bp->b_page_array) {
 246                 kmem_free(bp->b_pages);
 247                 bp->b_pages = NULL;
 248         }
 249 }
 250
 251 /*
 252  *      Releases the specified buffer.
 253  *
 254  *      The modification state of any associated pages is left unchanged.
 255  *      The buffer most not be on any hash - use xfs_buf_rele instead for
 256  *      hashed and refcounted buffers
 257  */
 258 void
 259 xfs_buf_free(
 260         xfs_buf_t               *bp)
 261 {
 262         trace_xfs_buf_free(bp, _RET_IP_);
 263
 264         ASSERT(list_empty(&bp->b_lru));
 265
 266         if (bp->b_flags & _XBF_PAGES) {
 267                 uint            i;
 268
 269                 if (xfs_buf_is_vmapped(bp))
 270                         vm_unmap_ram(bp->b_addr - bp->b_offset,
 271                                         bp->b_page_count);
 272
 273                 for (i = 0; i < bp->b_page_count; i++) {
 274                         struct page     *page = bp->b_pages[i];
 275
 276                         __free_page(page);
 277                 }
 278         } else if (bp->b_flags & _XBF_KMEM)
 279                 kmem_free(bp->b_addr);
 280         _xfs_buf_free_pages(bp);
 281         xfs_buf_deallocate(bp);
 282 }
 283
 284 /*
 285  * Allocates all the pages for buffer in question and builds it's page list.
 286  */
 287 STATIC int
 288 xfs_buf_allocate_memory(
 289         xfs_buf_t               *bp,
 290         uint                    flags)
 291 {
 292         size_t                  size = bp->b_count_desired;
 293         size_t                  nbytes, offset;
 294         gfp_t                   gfp_mask = xb_to_gfp(flags);
 295         unsigned short          page_count, i;
 296         pgoff_t                 first;
 297         xfs_off_t               end;
 298         int                     error;
 299
 300         /*
 301          * for buffers that are contained within a single page, just allocate
 302          * the memory from the heap - there's no need for the complexity of
 303          * page arrays to keep allocation down to order 0.
 304          */
 305         if (bp->b_buffer_length < PAGE_SIZE) {
 306                 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
 307                 if (!bp->b_addr) {
 308                         /* low memory - use alloc_page loop instead */
 309                         goto use_alloc_page;
 310                 }
 311
 312                 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
 313                                                                 PAGE_MASK) !=
 314                     ((unsigned long)bp->b_addr & PAGE_MASK)) {
 315                         /* b_addr spans two pages - use alloc_page instead */
 316                         kmem_free(bp->b_addr);
 317                         bp->b_addr = NULL;
 318                         goto use_alloc_page;
 319                 }
 320                 bp->b_offset = offset_in_page(bp->b_addr);
 321                 bp->b_pages = bp->b_page_array;
 322                 bp->b_pages[0] = virt_to_page(bp->b_addr);
 323                 bp->b_page_count = 1;
 324                 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
 325                 return 0;
 326         }
 327
 328 use_alloc_page:
 329         end = bp->b_file_offset + bp->b_buffer_length;
 330         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 331         error = _xfs_buf_get_pages(bp, page_count, flags);
 332         if (unlikely(error))
 333                 return error;
 334
 335         offset = bp->b_offset;
 336         first = bp->b_file_offset >> PAGE_SHIFT;
 337         bp->b_flags |= _XBF_PAGES;
 338
 339         for (i = 0; i < bp->b_page_count; i++) {
 340                 struct page     *page;
 341                 uint            retries = 0;
 342 retry:
 343                 page = alloc_page(gfp_mask);
 344                 if (unlikely(page == NULL)) {
 345                         if (flags & XBF_READ_AHEAD) {
 346                                 bp->b_page_count = i;
 347                                 error = ENOMEM;
 348                                 goto out_free_pages;
 349                         }
 350
 351                         /*
 352                          * This could deadlock.
 353                          *
 354                          * But until all the XFS lowlevel code is revamped to
 355                          * handle buffer allocation failures we can't do much.
 356                          */
 357                         if (!(++retries % 100))
 358                                 xfs_err(NULL,
 359                 "possible memory allocation deadlock in %s (mode:0x%x)",
 360                                         __func__, gfp_mask);
 361
 362                         XFS_STATS_INC(xb_page_retries);
 363                         congestion_wait(BLK_RW_ASYNC, HZ/50);
 364                         goto retry;
 365                 }
 366
 367                 XFS_STATS_INC(xb_page_found);
 368
 369                 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
 370                 size -= nbytes;
 371                 bp->b_pages[i] = page;
 372                 offset = 0;
 373         }
 374         return 0;
 375
 376 out_free_pages:
 377         for (i = 0; i < bp->b_page_count; i++)
 378                 __free_page(bp->b_pages[i]);
 379         return error;
 380 }
 381
 382 /*
 383  *      Map buffer into kernel address-space if nessecary.
 384  */
 385 STATIC int
 386 _xfs_buf_map_pages(
 387         xfs_buf_t               *bp,
 388         uint                    flags)
 389 {
 390         ASSERT(bp->b_flags & _XBF_PAGES);
 391         if (bp->b_page_count == 1) {
 392                 /* A single page buffer is always mappable */
 393                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 394                 bp->b_flags |= XBF_MAPPED;
 395         } else if (flags & XBF_MAPPED) {
 396                 int retried = 0;
 397
 398                 do {
 399                         bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
 400                                                 -1, PAGE_KERNEL);
 401                         if (bp->b_addr)
 402                                 break;
 403                         vm_unmap_aliases();
 404                 } while (retried++ <= 1);
 405
 406                 if (!bp->b_addr)
 407                         return -ENOMEM;
 408                 bp->b_addr += bp->b_offset;
 409                 bp->b_flags |= XBF_MAPPED;
 410         }
 411
 412         return 0;
 413 }
 414
 415 /*
 416  *      Finding and Reading Buffers
 417  */
 418
 419 /*
 420  *      Look up, and creates if absent, a lockable buffer for
 421  *      a given range of an inode.  The buffer is returned
 422  *      locked.  If other overlapping buffers exist, they are
 423  *      released before the new buffer is created and locked,
 424  *      which may imply that this call will block until those buffers
 425  *      are unlocked.  No I/O is implied by this call.
 426  */
 427 xfs_buf_t *
 428 _xfs_buf_find(
 429         xfs_buftarg_t           *btp,   /* block device target          */
 430         xfs_off_t               ioff,   /* starting offset of range     */
 431         size_t                  isize,  /* length of range              */
 432         xfs_buf_flags_t         flags,
 433         xfs_buf_t               *new_bp)
 434 {
 435         xfs_off_t               range_base;
 436         size_t                  range_length;
 437         struct xfs_perag        *pag;
 438         struct rb_node          **rbp;
 439         struct rb_node          *parent;
 440         xfs_buf_t               *bp;
 441
 442         range_base = (ioff << BBSHIFT);
 443         range_length = (isize << BBSHIFT);
 444
 445         /* Check for IOs smaller than the sector size / not sector aligned */
 446         ASSERT(!(range_length < (1 << btp->bt_sshift)));
 447         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
 448
 449         /* get tree root */
 450         pag = xfs_perag_get(btp->bt_mount,
 451                                 xfs_daddr_to_agno(btp->bt_mount, ioff));
 452
 453         /* walk tree */
 454         spin_lock(&pag->pag_buf_lock);
 455         rbp = &pag->pag_buf_tree.rb_node;
 456         parent = NULL;
 457         bp = NULL;
 458         while (*rbp) {
 459                 parent = *rbp;
 460                 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
 461
 462                 if (range_base < bp->b_file_offset)
 463                         rbp = &(*rbp)->rb_left;
 464                 else if (range_base > bp->b_file_offset)
 465                         rbp = &(*rbp)->rb_right;
 466                 else {
 467                         /*
 468                          * found a block offset match. If the range doesn't
 469                          * match, the only way this is allowed is if the buffer
 470                          * in the cache is stale and the transaction that made
 471                          * it stale has not yet committed. i.e. we are
 472                          * reallocating a busy extent. Skip this buffer and
 473                          * continue searching to the right for an exact match.
 474                          */
 475                         if (bp->b_buffer_length != range_length) {
 476                                 ASSERT(bp->b_flags & XBF_STALE);
 477                                 rbp = &(*rbp)->rb_right;
 478                                 continue;
 479                         }
 480                         atomic_inc(&bp->b_hold);
 481                         goto found;
 482                 }
 483         }
 484
 485         /* No match found */
 486         if (new_bp) {
 487                 _xfs_buf_initialize(new_bp, btp, range_base,
 488                                 range_length, flags);
 489                 rb_link_node(&new_bp->b_rbnode, parent, rbp);
 490                 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
 491                 /* the buffer keeps the perag reference until it is freed */
 492                 new_bp->b_pag = pag;
 493                 spin_unlock(&pag->pag_buf_lock);
 494         } else {
 495                 XFS_STATS_INC(xb_miss_locked);
 496                 spin_unlock(&pag->pag_buf_lock);
 497                 xfs_perag_put(pag);
 498         }
 499         return new_bp;
 500
 501 found:
 502         spin_unlock(&pag->pag_buf_lock);
 503         xfs_perag_put(pag);
 504
 505         if (xfs_buf_cond_lock(bp)) {
 506                 /* failed, so wait for the lock if requested. */
 507                 if (!(flags & XBF_TRYLOCK)) {
 508                         xfs_buf_lock(bp);
 509                         XFS_STATS_INC(xb_get_locked_waited);
 510                 } else {
 511                         xfs_buf_rele(bp);
 512                         XFS_STATS_INC(xb_busy_locked);
 513                         return NULL;
 514                 }
 515         }
 516
 517         /*
 518          * if the buffer is stale, clear all the external state associated with
 519          * it. We need to keep flags such as how we allocated the buffer memory
 520          * intact here.
 521          */
 522         if (bp->b_flags & XBF_STALE) {
 523                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 524                 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
 525         }
 526
 527         trace_xfs_buf_find(bp, flags, _RET_IP_);
 528         XFS_STATS_INC(xb_get_locked);
 529         return bp;
 530 }
 531
 532 /*
 533  *      Assembles a buffer covering the specified range.
 534  *      Storage in memory for all portions of the buffer will be allocated,
 535  *      although backing storage may not be.
 536  */
 537 xfs_buf_t *
 538 xfs_buf_get(
 539         xfs_buftarg_t           *target,/* target for buffer            */
 540         xfs_off_t               ioff,   /* starting offset of range     */
 541         size_t                  isize,  /* length of range              */
 542         xfs_buf_flags_t         flags)
 543 {
 544         xfs_buf_t               *bp, *new_bp;
 545         int                     error = 0;
 546
 547         new_bp = xfs_buf_allocate(flags);
 548         if (unlikely(!new_bp))
 549                 return NULL;
 550
 551         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 552         if (bp == new_bp) {
 553                 error = xfs_buf_allocate_memory(bp, flags);
 554                 if (error)
 555                         goto no_buffer;
 556         } else {
 557                 xfs_buf_deallocate(new_bp);
 558                 if (unlikely(bp == NULL))
 559                         return NULL;
 560         }
 561
 562         if (!(bp->b_flags & XBF_MAPPED)) {
 563                 error = _xfs_buf_map_pages(bp, flags);
 564                 if (unlikely(error)) {
 565                         xfs_warn(target->bt_mount,
 566                                 "%s: failed to map pages\n", __func__);
 567                         goto no_buffer;
 568                 }
 569         }
 570
 571         XFS_STATS_INC(xb_get);
 572
 573         /*
 574          * Always fill in the block number now, the mapped cases can do
 575          * their own overlay of this later.
 576          */
 577         bp->b_bn = ioff;
 578         bp->b_count_desired = bp->b_buffer_length;
 579
 580         trace_xfs_buf_get(bp, flags, _RET_IP_);
 581         return bp;
 582
 583  no_buffer:
 584         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 585                 xfs_buf_unlock(bp);
 586         xfs_buf_rele(bp);
 587         return NULL;
 588 }
 589
 590 STATIC int
 591 _xfs_buf_read(
 592         xfs_buf_t               *bp,
 593         xfs_buf_flags_t         flags)
 594 {
 595         int                     status;
 596
 597         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
 598         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
 599
 600         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
 601                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 602         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
 603                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 604
 605         status = xfs_buf_iorequest(bp);
 606         if (status || XFS_BUF_ISERROR(bp) || (flags & XBF_ASYNC))
 607                 return status;
 608         return xfs_buf_iowait(bp);
 609 }
 610
 611 xfs_buf_t *
 612 xfs_buf_read(
 613         xfs_buftarg_t           *target,
 614         xfs_off_t               ioff,
 615         size_t                  isize,
 616         xfs_buf_flags_t         flags)
 617 {
 618         xfs_buf_t               *bp;
 619
 620         flags |= XBF_READ;
 621
 622         bp = xfs_buf_get(target, ioff, isize, flags);
 623         if (bp) {
 624                 trace_xfs_buf_read(bp, flags, _RET_IP_);
 625
 626                 if (!XFS_BUF_ISDONE(bp)) {
 627                         XFS_STATS_INC(xb_get_read);
 628                         _xfs_buf_read(bp, flags);
 629                 } else if (flags & XBF_ASYNC) {
 630                         /*
 631                          * Read ahead call which is already satisfied,
 632                          * drop the buffer
 633                          */
 634                         goto no_buffer;
 635                 } else {
 636                         /* We do not want read in the flags */
 637                         bp->b_flags &= ~XBF_READ;
 638                 }
 639         }
 640
 641         return bp;
 642
 643  no_buffer:
 644         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 645                 xfs_buf_unlock(bp);
 646         xfs_buf_rele(bp);
 647         return NULL;
 648 }
 649
 650 /*
 651  *      If we are not low on memory then do the readahead in a deadlock
 652  *      safe manner.
 653  */
 654 void
 655 xfs_buf_readahead(
 656         xfs_buftarg_t           *target,
 657         xfs_off_t               ioff,
 658         size_t                  isize)
 659 {
 660         struct backing_dev_info *bdi;
 661
 662         if (bdi_read_congested(target->bt_bdi))
 663                 return;
 664
 665         xfs_buf_read(target, ioff, isize,
 666                      XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
 667 }
 668
 669 /*
 670  * Read an uncached buffer from disk. Allocates and returns a locked
 671  * buffer containing the disk contents or nothing.
 672  */
 673 struct xfs_buf *
 674 xfs_buf_read_uncached(
 675         struct xfs_mount        *mp,
 676         struct xfs_buftarg      *target,
 677         xfs_daddr_t             daddr,
 678         size_t                  length,
 679         int                     flags)
 680 {
 681         xfs_buf_t               *bp;
 682         int                     error;
 683
 684         bp = xfs_buf_get_uncached(target, length, flags);
 685         if (!bp)
 686                 return NULL;
 687
 688         /* set up the buffer for a read IO */
 689         xfs_buf_lock(bp);
 690         XFS_BUF_SET_ADDR(bp, daddr);
 691         XFS_BUF_READ(bp);
 692         XFS_BUF_BUSY(bp);
 693
 694         xfsbdstrat(mp, bp);
 695         error = xfs_buf_iowait(bp);
 696         if (error || bp->b_error) {
 697                 xfs_buf_relse(bp);
 698                 return NULL;
 699         }
 700         return bp;
 701 }
 702
 703 xfs_buf_t *
 704 xfs_buf_get_empty(
 705         size_t                  len,
 706         xfs_buftarg_t           *target)
 707 {
 708         xfs_buf_t               *bp;
 709
 710         bp = xfs_buf_allocate(0);
 711         if (bp)
 712                 _xfs_buf_initialize(bp, target, 0, len, 0);
 713         return bp;
 714 }
 715
 716 static inline struct page *
 717 mem_to_page(
 718         void                    *addr)
 719 {
 720         if ((!is_vmalloc_addr(addr))) {
 721                 return virt_to_page(addr);
 722         } else {
 723                 return vmalloc_to_page(addr);
 724         }
 725 }
 726
 727 int
 728 xfs_buf_associate_memory(
 729         xfs_buf_t               *bp,
 730         void                    *mem,
 731         size_t                  len)
 732 {
 733         int                     rval;
 734         int                     i = 0;
 735         unsigned long           pageaddr;
 736         unsigned long           offset;
 737         size_t                  buflen;
 738         int                     page_count;
 739
 740         pageaddr = (unsigned long)mem & PAGE_MASK;
 741         offset = (unsigned long)mem - pageaddr;
 742         buflen = PAGE_ALIGN(len + offset);
 743         page_count = buflen >> PAGE_SHIFT;
 744
 745         /* Free any previous set of page pointers */
 746         if (bp->b_pages)
 747                 _xfs_buf_free_pages(bp);
 748
 749         bp->b_pages = NULL;
 750         bp->b_addr = mem;
 751
 752         rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
 753         if (rval)
 754                 return rval;
 755
 756         bp->b_offset = offset;
 757
 758         for (i = 0; i < bp->b_page_count; i++) {
 759                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 760                 pageaddr += PAGE_SIZE;
 761         }
 762
 763         bp->b_count_desired = len;
 764         bp->b_buffer_length = buflen;
 765         bp->b_flags |= XBF_MAPPED;
 766
 767         return 0;
 768 }
 769
 770 xfs_buf_t *
 771 xfs_buf_get_uncached(
 772         struct xfs_buftarg      *target,
 773         size_t                  len,
 774         int                     flags)
 775 {
 776         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 777         int                     error, i;
 778         xfs_buf_t               *bp;
 779
 780         bp = xfs_buf_allocate(0);
 781         if (unlikely(bp == NULL))
 782                 goto fail;
 783         _xfs_buf_initialize(bp, target, 0, len, 0);
 784
 785         error = _xfs_buf_get_pages(bp, page_count, 0);
 786         if (error)
 787                 goto fail_free_buf;
 788
 789         for (i = 0; i < page_count; i++) {
 790                 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
 791                 if (!bp->b_pages[i])
 792                         goto fail_free_mem;
 793         }
 794         bp->b_flags |= _XBF_PAGES;
 795
 796         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 797         if (unlikely(error)) {
 798                 xfs_warn(target->bt_mount,
 799                         "%s: failed to map pages\n", __func__);
 800                 goto fail_free_mem;
 801         }
 802
 803         xfs_buf_unlock(bp);
 804
 805         trace_xfs_buf_get_uncached(bp, _RET_IP_);
 806         return bp;
 807
 808  fail_free_mem:
 809         while (--i >= 0)
 810                 __free_page(bp->b_pages[i]);
 811         _xfs_buf_free_pages(bp);
 812  fail_free_buf:
 813         xfs_buf_deallocate(bp);
 814  fail:
 815         return NULL;
 816 }
 817
 818 /*
 819  *      Increment reference count on buffer, to hold the buffer concurrently
 820  *      with another thread which may release (free) the buffer asynchronously.
 821  *      Must hold the buffer already to call this function.
 822  */
 823 void
 824 xfs_buf_hold(
 825         xfs_buf_t               *bp)
 826 {
 827         trace_xfs_buf_hold(bp, _RET_IP_);
 828         atomic_inc(&bp->b_hold);
 829 }
 830
 831 /*
 832  *      Releases a hold on the specified buffer.  If the
 833  *      the hold count is 1, calls xfs_buf_free.
 834  */
 835 void
 836 xfs_buf_rele(
 837         xfs_buf_t               *bp)
 838 {
 839         struct xfs_perag        *pag = bp->b_pag;
 840
 841         trace_xfs_buf_rele(bp, _RET_IP_);
 842
 843         if (!pag) {
 844                 ASSERT(list_empty(&bp->b_lru));
 845                 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
 846                 if (atomic_dec_and_test(&bp->b_hold))
 847                         xfs_buf_free(bp);
 848                 return;
 849         }
 850
 851         ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
 852
 853         ASSERT(atomic_read(&bp->b_hold) > 0);
 854         if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
 855                 if (!(bp->b_flags & XBF_STALE) &&
 856                            atomic_read(&bp->b_lru_ref)) {
 857                         xfs_buf_lru_add(bp);
 858                         spin_unlock(&pag->pag_buf_lock);
 859                 } else {
 860                         xfs_buf_lru_del(bp);
 861                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 862                         rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
 863                         spin_unlock(&pag->pag_buf_lock);
 864                         xfs_perag_put(pag);
 865                         xfs_buf_free(bp);
 866                 }
 867         }
 868 }
 869
 870
 871 /*
 872  *      Lock a buffer object, if it is not already locked.
 873  *
 874  *      If we come across a stale, pinned, locked buffer, we know that we are
 875  *      being asked to lock a buffer that has been reallocated. Because it is
 876  *      pinned, we know that the log has not been pushed to disk and hence it
 877  *      will still be locked.  Rather than continuing to have trylock attempts
 878  *      fail until someone else pushes the log, push it ourselves before
 879  *      returning.  This means that the xfsaild will not get stuck trying
 880  *      to push on stale inode buffers.
 881  */
 882 int
 883 xfs_buf_cond_lock(
 884         xfs_buf_t               *bp)
 885 {
 886         int                     locked;
 887
 888         locked = down_trylock(&bp->b_sema) == 0;
 889         if (locked)
 890                 XB_SET_OWNER(bp);
 891         else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 892                 xfs_log_force(bp->b_target->bt_mount, 0);
 893
 894         trace_xfs_buf_cond_lock(bp, _RET_IP_);
 895         return locked ? 0 : -EBUSY;
 896 }
 897
 898 int
 899 xfs_buf_lock_value(
 900         xfs_buf_t               *bp)
 901 {
 902         return bp->b_sema.count;
 903 }
 904
 905 /*
 906  *      Lock a buffer object.
 907  *
 908  *      If we come across a stale, pinned, locked buffer, we know that we
 909  *      are being asked to lock a buffer that has been reallocated. Because
 910  *      it is pinned, we know that the log has not been pushed to disk and
 911  *      hence it will still be locked. Rather than sleeping until someone
 912  *      else pushes the log, push it ourselves before trying to get the lock.
 913  */
 914 void
 915 xfs_buf_lock(
 916         xfs_buf_t               *bp)
 917 {
 918         trace_xfs_buf_lock(bp, _RET_IP_);
 919
 920         if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 921                 xfs_log_force(bp->b_target->bt_mount, 0);
 922         if (atomic_read(&bp->b_io_remaining))
 923                 blk_run_backing_dev(bp->b_target->bt_bdi, NULL);
 924         down(&bp->b_sema);
 925         XB_SET_OWNER(bp);
 926
 927         trace_xfs_buf_lock_done(bp, _RET_IP_);
 928 }
 929
 930 /*
 931  *      Releases the lock on the buffer object.
 932  *      If the buffer is marked delwri but is not queued, do so before we
 933  *      unlock the buffer as we need to set flags correctly.  We also need to
 934  *      take a reference for the delwri queue because the unlocker is going to
 935  *      drop their's and they don't know we just queued it.
 936  */
 937 void
 938 xfs_buf_unlock(
 939         xfs_buf_t               *bp)
 940 {
 941         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 942                 atomic_inc(&bp->b_hold);
 943                 bp->b_flags |= XBF_ASYNC;
 944                 xfs_buf_delwri_queue(bp, 0);
 945         }
 946
 947         XB_CLEAR_OWNER(bp);
 948         up(&bp->b_sema);
 949
 950         trace_xfs_buf_unlock(bp, _RET_IP_);
 951 }
 952
 953 STATIC void
 954 xfs_buf_wait_unpin(
 955         xfs_buf_t               *bp)
 956 {
 957         DECLARE_WAITQUEUE       (wait, current);
 958
 959         if (atomic_read(&bp->b_pin_count) == 0)
 960                 return;
 961
 962         add_wait_queue(&bp->b_waiters, &wait);
 963         for (;;) {
 964                 set_current_state(TASK_UNINTERRUPTIBLE);
 965                 if (atomic_read(&bp->b_pin_count) == 0)
 966                         break;
 967                 if (atomic_read(&bp->b_io_remaining))
 968                         blk_run_backing_dev(bp->b_target->bt_bdi, NULL);
 969                 schedule();
 970         }
 971         remove_wait_queue(&bp->b_waiters, &wait);
 972         set_current_state(TASK_RUNNING);
 973 }
 974
 975 /*
 976  *      Buffer Utility Routines
 977  */
 978
 979 STATIC void
 980 xfs_buf_iodone_work(
 981         struct work_struct      *work)
 982 {
 983         xfs_buf_t               *bp =
 984                 container_of(work, xfs_buf_t, b_iodone_work);
 985
 986         if (bp->b_iodone)
 987                 (*(bp->b_iodone))(bp);
 988         else if (bp->b_flags & XBF_ASYNC)
 989                 xfs_buf_relse(bp);
 990 }
 991
 992 void
 993 xfs_buf_ioend(
 994         xfs_buf_t               *bp,
 995         int                     schedule)
 996 {
 997         trace_xfs_buf_iodone(bp, _RET_IP_);
 998
 999         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1000         if (bp->b_error == 0)
1001                 bp->b_flags |= XBF_DONE;
1002
1003         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1004                 if (schedule) {
1005                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1006                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1007                 } else {
1008                         xfs_buf_iodone_work(&bp->b_iodone_work);
1009                 }
1010         } else {
1011                 complete(&bp->b_iowait);
1012         }
1013 }
1014
1015 void
1016 xfs_buf_ioerror(
1017         xfs_buf_t               *bp,
1018         int                     error)
1019 {
1020         ASSERT(error >= 0 && error <= 0xffff);
1021         bp->b_error = (unsigned short)error;
1022         trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1023 }
1024
1025 int
1026 xfs_bwrite(
1027         struct xfs_mount        *mp,
1028         struct xfs_buf          *bp)
1029 {
1030         int                     error;
1031
1032         bp->b_flags |= XBF_WRITE;
1033         bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1034
1035         xfs_buf_delwri_dequeue(bp);
1036         xfs_bdstrat_cb(bp);
1037
1038         error = xfs_buf_iowait(bp);
1039         if (error)
1040                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1041         xfs_buf_relse(bp);
1042         return error;
1043 }
1044
1045 void
1046 xfs_bdwrite(
1047         void                    *mp,
1048         struct xfs_buf          *bp)
1049 {
1050         trace_xfs_buf_bdwrite(bp, _RET_IP_);
1051
1052         bp->b_flags &= ~XBF_READ;
1053         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1054
1055         xfs_buf_delwri_queue(bp, 1);
1056 }
1057
1058 /*
1059  * Called when we want to stop a buffer from getting written or read.
1060  * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1061  * so that the proper iodone callbacks get called.
1062  */
1063 STATIC int
1064 xfs_bioerror(
1065         xfs_buf_t *bp)
1066 {
1067 #ifdef XFSERRORDEBUG
1068         ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1069 #endif
1070
1071         /*
1072          * No need to wait until the buffer is unpinned, we aren't flushing it.
1073          */
1074         XFS_BUF_ERROR(bp, EIO);
1075
1076         /*
1077          * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1078          */
1079         XFS_BUF_UNREAD(bp);
1080         XFS_BUF_UNDELAYWRITE(bp);
1081         XFS_BUF_UNDONE(bp);
1082         XFS_BUF_STALE(bp);
1083
1084         xfs_buf_ioend(bp, 0);
1085
1086         return EIO;
1087 }
1088
1089 /*
1090  * Same as xfs_bioerror, except that we are releasing the buffer
1091  * here ourselves, and avoiding the xfs_buf_ioend call.
1092  * This is meant for userdata errors; metadata bufs come with
1093  * iodone functions attached, so that we can track down errors.
1094  */
1095 STATIC int
1096 xfs_bioerror_relse(
1097         struct xfs_buf  *bp)
1098 {
1099         int64_t         fl = XFS_BUF_BFLAGS(bp);
1100         /*
1101          * No need to wait until the buffer is unpinned.
1102          * We aren't flushing it.
1103          *
1104          * chunkhold expects B_DONE to be set, whether
1105          * we actually finish the I/O or not. We don't want to
1106          * change that interface.
1107          */
1108         XFS_BUF_UNREAD(bp);
1109         XFS_BUF_UNDELAYWRITE(bp);
1110         XFS_BUF_DONE(bp);
1111         XFS_BUF_STALE(bp);
1112         XFS_BUF_CLR_IODONE_FUNC(bp);
1113         if (!(fl & XBF_ASYNC)) {
1114                 /*
1115                  * Mark b_error and B_ERROR _both_.
1116                  * Lot's of chunkcache code assumes that.
1117                  * There's no reason to mark error for
1118                  * ASYNC buffers.
1119                  */
1120                 XFS_BUF_ERROR(bp, EIO);
1121                 XFS_BUF_FINISH_IOWAIT(bp);
1122         } else {
1123                 xfs_buf_relse(bp);
1124         }
1125
1126         return EIO;
1127 }
1128
1129
1130 /*
1131  * All xfs metadata buffers except log state machine buffers
1132  * get this attached as their b_bdstrat callback function.
1133  * This is so that we can catch a buffer
1134  * after prematurely unpinning it to forcibly shutdown the filesystem.
1135  */
1136 int
1137 xfs_bdstrat_cb(
1138         struct xfs_buf  *bp)
1139 {
1140         if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1141                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1142                 /*
1143                  * Metadata write that didn't get logged but
1144                  * written delayed anyway. These aren't associated
1145                  * with a transaction, and can be ignored.
1146                  */
1147                 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1148                         return xfs_bioerror_relse(bp);
1149                 else
1150                         return xfs_bioerror(bp);
1151         }
1152
1153         xfs_buf_iorequest(bp);
1154         return 0;
1155 }
1156
1157 /*
1158  * Wrapper around bdstrat so that we can stop data from going to disk in case
1159  * we are shutting down the filesystem.  Typically user data goes thru this
1160  * path; one of the exceptions is the superblock.
1161  */
1162 void
1163 xfsbdstrat(
1164         struct xfs_mount        *mp,
1165         struct xfs_buf          *bp)
1166 {
1167         if (XFS_FORCED_SHUTDOWN(mp)) {
1168                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1169                 xfs_bioerror_relse(bp);
1170                 return;
1171         }
1172
1173         xfs_buf_iorequest(bp);
1174 }
1175
1176 STATIC void
1177 _xfs_buf_ioend(
1178         xfs_buf_t               *bp,
1179         int                     schedule)
1180 {
1181         if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1182                 xfs_buf_ioend(bp, schedule);
1183 }
1184
1185 STATIC void
1186 xfs_buf_bio_end_io(
1187         struct bio              *bio,
1188         int                     error)
1189 {
1190         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1191
1192         xfs_buf_ioerror(bp, -error);
1193
1194         if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1195                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1196
1197         _xfs_buf_ioend(bp, 1);
1198         bio_put(bio);
1199 }
1200
1201 STATIC void
1202 _xfs_buf_ioapply(
1203         xfs_buf_t               *bp)
1204 {
1205         int                     rw, map_i, total_nr_pages, nr_pages;
1206         struct bio              *bio;
1207         int                     offset = bp->b_offset;
1208         int                     size = bp->b_count_desired;
1209         sector_t                sector = bp->b_bn;
1210
1211         total_nr_pages = bp->b_page_count;
1212         map_i = 0;
1213
1214         if (bp->b_flags & XBF_ORDERED) {
1215                 ASSERT(!(bp->b_flags & XBF_READ));
1216                 rw = WRITE_FLUSH_FUA;
1217         } else if (bp->b_flags & XBF_LOG_BUFFER) {
1218                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1219                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1220                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1221         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1222                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1223                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1224                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1225         } else {
1226                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1227                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1228         }
1229
1230
1231 next_chunk:
1232         atomic_inc(&bp->b_io_remaining);
1233         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1234         if (nr_pages > total_nr_pages)
1235                 nr_pages = total_nr_pages;
1236
1237         bio = bio_alloc(GFP_NOIO, nr_pages);
1238         bio->bi_bdev = bp->b_target->bt_bdev;
1239         bio->bi_sector = sector;
1240         bio->bi_end_io = xfs_buf_bio_end_io;
1241         bio->bi_private = bp;
1242
1243
1244         for (; size && nr_pages; nr_pages--, map_i++) {
1245                 int     rbytes, nbytes = PAGE_SIZE - offset;
1246
1247                 if (nbytes > size)
1248                         nbytes = size;
1249
1250                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1251                 if (rbytes < nbytes)
1252                         break;
1253
1254                 offset = 0;
1255                 sector += nbytes >> BBSHIFT;
1256                 size -= nbytes;
1257                 total_nr_pages--;
1258         }
1259
1260         if (likely(bio->bi_size)) {
1261                 if (xfs_buf_is_vmapped(bp)) {
1262                         flush_kernel_vmap_range(bp->b_addr,
1263                                                 xfs_buf_vmap_len(bp));
1264                 }
1265                 submit_bio(rw, bio);
1266                 if (size)
1267                         goto next_chunk;
1268         } else {
1269                 xfs_buf_ioerror(bp, EIO);
1270                 bio_put(bio);
1271         }
1272 }
1273
1274 int
1275 xfs_buf_iorequest(
1276         xfs_buf_t               *bp)
1277 {
1278         trace_xfs_buf_iorequest(bp, _RET_IP_);
1279
1280         if (bp->b_flags & XBF_DELWRI) {
1281                 xfs_buf_delwri_queue(bp, 1);
1282                 return 0;
1283         }
1284
1285         if (bp->b_flags & XBF_WRITE) {
1286                 xfs_buf_wait_unpin(bp);
1287         }
1288
1289         xfs_buf_hold(bp);
1290
1291         /* Set the count to 1 initially, this will stop an I/O
1292          * completion callout which happens before we have started
1293          * all the I/O from calling xfs_buf_ioend too early.
1294          */
1295         atomic_set(&bp->b_io_remaining, 1);
1296         _xfs_buf_ioapply(bp);
1297         _xfs_buf_ioend(bp, 0);
1298
1299         xfs_buf_rele(bp);
1300         return 0;
1301 }
1302
1303 /*
1304  *      Waits for I/O to complete on the buffer supplied.
1305  *      It returns immediately if no I/O is pending.
1306  *      It returns the I/O error code, if any, or 0 if there was no error.
1307  */
1308 int
1309 xfs_buf_iowait(
1310         xfs_buf_t               *bp)
1311 {
1312         trace_xfs_buf_iowait(bp, _RET_IP_);
1313
1314         if (atomic_read(&bp->b_io_remaining))
1315                 blk_run_backing_dev(bp->b_target->bt_bdi, NULL);
1316         wait_for_completion(&bp->b_iowait);
1317
1318         trace_xfs_buf_iowait_done(bp, _RET_IP_);
1319         return bp->b_error;
1320 }
1321
1322 xfs_caddr_t
1323 xfs_buf_offset(
1324         xfs_buf_t               *bp,
1325         size_t                  offset)
1326 {
1327         struct page             *page;
1328
1329         if (bp->b_flags & XBF_MAPPED)
1330                 return XFS_BUF_PTR(bp) + offset;
1331
1332         offset += bp->b_offset;
1333         page = bp->b_pages[offset >> PAGE_SHIFT];
1334         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1335 }
1336
1337 /*
1338  *      Move data into or out of a buffer.
1339  */
1340 void
1341 xfs_buf_iomove(
1342         xfs_buf_t               *bp,    /* buffer to process            */
1343         size_t                  boff,   /* starting buffer offset       */
1344         size_t                  bsize,  /* length to copy               */
1345         void                    *data,  /* data address                 */
1346         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1347 {
1348         size_t                  bend, cpoff, csize;
1349         struct page             *page;
1350
1351         bend = boff + bsize;
1352         while (boff < bend) {
1353                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1354                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1355                 csize = min_t(size_t,
1356                               PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1357
1358                 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1359
1360                 switch (mode) {
1361                 case XBRW_ZERO:
1362                         memset(page_address(page) + cpoff, 0, csize);
1363                         break;
1364                 case XBRW_READ:
1365                         memcpy(data, page_address(page) + cpoff, csize);
1366                         break;
1367                 case XBRW_WRITE:
1368                         memcpy(page_address(page) + cpoff, data, csize);
1369                 }
1370
1371                 boff += csize;
1372                 data += csize;
1373         }
1374 }
1375
1376 /*
1377  *      Handling of buffer targets (buftargs).
1378  */
1379
1380 /*
1381  * Wait for any bufs with callbacks that have been submitted but have not yet
1382  * returned. These buffers will have an elevated hold count, so wait on those
1383  * while freeing all the buffers only held by the LRU.
1384  */
1385 void
1386 xfs_wait_buftarg(
1387         struct xfs_buftarg      *btp)
1388 {
1389         struct xfs_buf          *bp;
1390
1391 restart:
1392         spin_lock(&btp->bt_lru_lock);
1393         while (!list_empty(&btp->bt_lru)) {
1394                 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1395                 if (atomic_read(&bp->b_hold) > 1) {
1396                         spin_unlock(&btp->bt_lru_lock);
1397                         delay(100);
1398                         goto restart;
1399                 }
1400                 /*
1401                  * clear the LRU reference count so the bufer doesn't get
1402                  * ignored in xfs_buf_rele().
1403                  */
1404                 atomic_set(&bp->b_lru_ref, 0);
1405                 spin_unlock(&btp->bt_lru_lock);
1406                 xfs_buf_rele(bp);
1407                 spin_lock(&btp->bt_lru_lock);
1408         }
1409         spin_unlock(&btp->bt_lru_lock);
1410 }
1411
1412 int
1413 xfs_buftarg_shrink(
1414         struct shrinker         *shrink,
1415         int                     nr_to_scan,
1416         gfp_t                   mask)
1417 {
1418         struct xfs_buftarg      *btp = container_of(shrink,
1419                                         struct xfs_buftarg, bt_shrinker);
1420         struct xfs_buf          *bp;
1421         LIST_HEAD(dispose);
1422
1423         if (!nr_to_scan)
1424                 return btp->bt_lru_nr;
1425
1426         spin_lock(&btp->bt_lru_lock);
1427         while (!list_empty(&btp->bt_lru)) {
1428                 if (nr_to_scan-- <= 0)
1429                         break;
1430
1431                 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1432
1433                 /*
1434                  * Decrement the b_lru_ref count unless the value is already
1435                  * zero. If the value is already zero, we need to reclaim the
1436                  * buffer, otherwise it gets another trip through the LRU.
1437                  */
1438                 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1439                         list_move_tail(&bp->b_lru, &btp->bt_lru);
1440                         continue;
1441                 }
1442
1443                 /*
1444                  * remove the buffer from the LRU now to avoid needing another
1445                  * lock round trip inside xfs_buf_rele().
1446                  */
1447                 list_move(&bp->b_lru, &dispose);
1448                 btp->bt_lru_nr--;
1449         }
1450         spin_unlock(&btp->bt_lru_lock);
1451
1452         while (!list_empty(&dispose)) {
1453                 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1454                 list_del_init(&bp->b_lru);
1455                 xfs_buf_rele(bp);
1456         }
1457
1458         return btp->bt_lru_nr;
1459 }
1460
1461 void
1462 xfs_free_buftarg(
1463         struct xfs_mount        *mp,
1464         struct xfs_buftarg      *btp)
1465 {
1466         unregister_shrinker(&btp->bt_shrinker);
1467
1468         xfs_flush_buftarg(btp, 1);
1469         if (mp->m_flags & XFS_MOUNT_BARRIER)
1470                 xfs_blkdev_issue_flush(btp);
1471
1472         kthread_stop(btp->bt_task);
1473         kmem_free(btp);
1474 }
1475
1476 STATIC int
1477 xfs_setsize_buftarg_flags(
1478         xfs_buftarg_t           *btp,
1479         unsigned int            blocksize,
1480         unsigned int            sectorsize,
1481         int                     verbose)
1482 {
1483         btp->bt_bsize = blocksize;
1484         btp->bt_sshift = ffs(sectorsize) - 1;
1485         btp->bt_smask = sectorsize - 1;
1486
1487         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1488                 xfs_warn(btp->bt_mount,
1489                         "Cannot set_blocksize to %u on device %s\n",
1490                         sectorsize, XFS_BUFTARG_NAME(btp));
1491                 return EINVAL;
1492         }
1493
1494         return 0;
1495 }
1496
1497 /*
1498  *      When allocating the initial buffer target we have not yet
1499  *      read in the superblock, so don't know what sized sectors
1500  *      are being used is at this early stage.  Play safe.
1501  */
1502 STATIC int
1503 xfs_setsize_buftarg_early(
1504         xfs_buftarg_t           *btp,
1505         struct block_device     *bdev)
1506 {
1507         return xfs_setsize_buftarg_flags(btp,
1508                         PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1509 }
1510
1511 int
1512 xfs_setsize_buftarg(
1513         xfs_buftarg_t           *btp,
1514         unsigned int            blocksize,
1515         unsigned int            sectorsize)
1516 {
1517         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1518 }
1519
1520 STATIC int
1521 xfs_alloc_delwrite_queue(
1522         xfs_buftarg_t           *btp,
1523         const char              *fsname)
1524 {
1525         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1526         spin_lock_init(&btp->bt_delwrite_lock);
1527         btp->bt_flags = 0;
1528         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1529         if (IS_ERR(btp->bt_task))
1530                 return PTR_ERR(btp->bt_task);
1531         return 0;
1532 }
1533
1534 xfs_buftarg_t *
1535 xfs_alloc_buftarg(
1536         struct xfs_mount        *mp,
1537         struct block_device     *bdev,
1538         int                     external,
1539         const char              *fsname)
1540 {
1541         xfs_buftarg_t           *btp;
1542
1543         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1544
1545         btp->bt_mount = mp;
1546         btp->bt_dev =  bdev->bd_dev;
1547         btp->bt_bdev = bdev;
1548         btp->bt_bdi = blk_get_backing_dev_info(bdev);
1549         if (!btp->bt_bdi)
1550                 goto error;
1551
1552         INIT_LIST_HEAD(&btp->bt_lru);
1553         spin_lock_init(&btp->bt_lru_lock);
1554         if (xfs_setsize_buftarg_early(btp, bdev))
1555                 goto error;
1556         if (xfs_alloc_delwrite_queue(btp, fsname))
1557                 goto error;
1558         btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1559         btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1560         register_shrinker(&btp->bt_shrinker);
1561         return btp;
1562
1563 error:
1564         kmem_free(btp);
1565         return NULL;
1566 }
1567
1568
1569 /*
1570  *      Delayed write buffer handling
1571  */
1572 STATIC void
1573 xfs_buf_delwri_queue(
1574         xfs_buf_t               *bp,
1575         int                     unlock)
1576 {
1577         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1578         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1579
1580         trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1581
1582         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1583
1584         spin_lock(dwlk);
1585         /* If already in the queue, dequeue and place at tail */
1586         if (!list_empty(&bp->b_list)) {
1587                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1588                 if (unlock)
1589                         atomic_dec(&bp->b_hold);
1590                 list_del(&bp->b_list);
1591         }
1592
1593         if (list_empty(dwq)) {
1594                 /* start xfsbufd as it is about to have something to do */
1595                 wake_up_process(bp->b_target->bt_task);
1596         }
1597
1598         bp->b_flags |= _XBF_DELWRI_Q;
1599         list_add_tail(&bp->b_list, dwq);
1600         bp->b_queuetime = jiffies;
1601         spin_unlock(dwlk);
1602
1603         if (unlock)
1604                 xfs_buf_unlock(bp);
1605 }
1606
1607 void
1608 xfs_buf_delwri_dequeue(
1609         xfs_buf_t               *bp)
1610 {
1611         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1612         int                     dequeued = 0;
1613
1614         spin_lock(dwlk);
1615         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1616                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1617                 list_del_init(&bp->b_list);
1618                 dequeued = 1;
1619         }
1620         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1621         spin_unlock(dwlk);
1622
1623         if (dequeued)
1624                 xfs_buf_rele(bp);
1625
1626         trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1627 }
1628
1629 /*
1630  * If a delwri buffer needs to be pushed before it has aged out, then promote
1631  * it to the head of the delwri queue so that it will be flushed on the next
1632  * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1633  * than the age currently needed to flush the buffer. Hence the next time the
1634  * xfsbufd sees it is guaranteed to be considered old enough to flush.
1635  */
1636 void
1637 xfs_buf_delwri_promote(
1638         struct xfs_buf  *bp)
1639 {
1640         struct xfs_buftarg *btp = bp->b_target;
1641         long            age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1642
1643         ASSERT(bp->b_flags & XBF_DELWRI);
1644         ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1645
1646         /*
1647          * Check the buffer age before locking the delayed write queue as we
1648          * don't need to promote buffers that are already past the flush age.
1649          */
1650         if (bp->b_queuetime < jiffies - age)
1651                 return;
1652         bp->b_queuetime = jiffies - age;
1653         spin_lock(&btp->bt_delwrite_lock);
1654         list_move(&bp->b_list, &btp->bt_delwrite_queue);
1655         spin_unlock(&btp->bt_delwrite_lock);
1656 }
1657
1658 STATIC void
1659 xfs_buf_runall_queues(
1660         struct workqueue_struct *queue)
1661 {
1662         flush_workqueue(queue);
1663 }
1664
1665 /*
1666  * Move as many buffers as specified to the supplied list
1667  * idicating if we skipped any buffers to prevent deadlocks.
1668  */
1669 STATIC int
1670 xfs_buf_delwri_split(
1671         xfs_buftarg_t   *target,
1672         struct list_head *list,
1673         unsigned long   age)
1674 {
1675         xfs_buf_t       *bp, *n;
1676         struct list_head *dwq = &target->bt_delwrite_queue;
1677         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1678         int             skipped = 0;
1679         int             force;
1680
1681         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1682         INIT_LIST_HEAD(list);
1683         spin_lock(dwlk);
1684         list_for_each_entry_safe(bp, n, dwq, b_list) {
1685                 ASSERT(bp->b_flags & XBF_DELWRI);
1686
1687                 if (!XFS_BUF_ISPINNED(bp) && !xfs_buf_cond_lock(bp)) {
1688                         if (!force &&
1689                             time_before(jiffies, bp->b_queuetime + age)) {
1690                                 xfs_buf_unlock(bp);
1691                                 break;
1692                         }
1693
1694                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1695                                          _XBF_RUN_QUEUES);
1696                         bp->b_flags |= XBF_WRITE;
1697                         list_move_tail(&bp->b_list, list);
1698                         trace_xfs_buf_delwri_split(bp, _RET_IP_);
1699                 } else
1700                         skipped++;
1701         }
1702         spin_unlock(dwlk);
1703
1704         return skipped;
1705
1706 }
1707
1708 /*
1709  * Compare function is more complex than it needs to be because
1710  * the return value is only 32 bits and we are doing comparisons
1711  * on 64 bit values
1712  */
1713 static int
1714 xfs_buf_cmp(
1715         void            *priv,
1716         struct list_head *a,
1717         struct list_head *b)
1718 {
1719         struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1720         struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1721         xfs_daddr_t             diff;
1722
1723         diff = ap->b_bn - bp->b_bn;
1724         if (diff < 0)
1725                 return -1;
1726         if (diff > 0)
1727                 return 1;
1728         return 0;
1729 }
1730
1731 void
1732 xfs_buf_delwri_sort(
1733         xfs_buftarg_t   *target,
1734         struct list_head *list)
1735 {
1736         list_sort(NULL, list, xfs_buf_cmp);
1737 }
1738
1739 STATIC int
1740 xfsbufd(
1741         void            *data)
1742 {
1743         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1744
1745         current->flags |= PF_MEMALLOC;
1746
1747         set_freezable();
1748
1749         do {
1750                 long    age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1751                 long    tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1752                 int     count = 0;
1753                 struct list_head tmp;
1754
1755                 if (unlikely(freezing(current))) {
1756                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1757                         refrigerator();
1758                 } else {
1759                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1760                 }
1761
1762                 /* sleep for a long time if there is nothing to do. */
1763                 if (list_empty(&target->bt_delwrite_queue))
1764                         tout = MAX_SCHEDULE_TIMEOUT;
1765                 schedule_timeout_interruptible(tout);
1766
1767                 xfs_buf_delwri_split(target, &tmp, age);
1768                 list_sort(NULL, &tmp, xfs_buf_cmp);
1769                 while (!list_empty(&tmp)) {
1770                         struct xfs_buf *bp;
1771                         bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1772                         list_del_init(&bp->b_list);
1773                         xfs_bdstrat_cb(bp);
1774                         count++;
1775                 }
1776                 if (count)
1777                         blk_run_backing_dev(target->bt_bdi, NULL);
1778
1779         } while (!kthread_should_stop());
1780
1781         return 0;
1782 }
1783
1784 /*
1785  *      Go through all incore buffers, and release buffers if they belong to
1786  *      the given device. This is used in filesystem error handling to
1787  *      preserve the consistency of its metadata.
1788  */
1789 int
1790 xfs_flush_buftarg(
1791         xfs_buftarg_t   *target,
1792         int             wait)
1793 {
1794         xfs_buf_t       *bp;
1795         int             pincount = 0;
1796         LIST_HEAD(tmp_list);
1797         LIST_HEAD(wait_list);
1798
1799         xfs_buf_runall_queues(xfsconvertd_workqueue);
1800         xfs_buf_runall_queues(xfsdatad_workqueue);
1801         xfs_buf_runall_queues(xfslogd_workqueue);
1802
1803         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1804         pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1805
1806         /*
1807          * Dropped the delayed write list lock, now walk the temporary list.
1808          * All I/O is issued async and then if we need to wait for completion
1809          * we do that after issuing all the IO.
1810          */
1811         list_sort(NULL, &tmp_list, xfs_buf_cmp);
1812         while (!list_empty(&tmp_list)) {
1813                 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1814                 ASSERT(target == bp->b_target);
1815                 list_del_init(&bp->b_list);
1816                 if (wait) {
1817                         bp->b_flags &= ~XBF_ASYNC;
1818                         list_add(&bp->b_list, &wait_list);
1819                 }
1820                 xfs_bdstrat_cb(bp);
1821         }
1822
1823         if (wait) {
1824                 /* Expedite and wait for IO to complete. */
1825                 blk_run_backing_dev(target->bt_bdi, NULL);
1826                 while (!list_empty(&wait_list)) {
1827                         bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1828
1829                         list_del_init(&bp->b_list);
1830                         xfs_buf_iowait(bp);
1831                         xfs_buf_relse(bp);
1832                 }
1833         }
1834
1835         return pincount;
1836 }
1837
1838 int __init
1839 xfs_buf_init(void)
1840 {
1841         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1842                                                 KM_ZONE_HWALIGN, NULL);
1843         if (!xfs_buf_zone)
1844                 goto out;
1845
1846         xfslogd_workqueue = alloc_workqueue("xfslogd",
1847                                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1848         if (!xfslogd_workqueue)
1849                 goto out_free_buf_zone;
1850
1851         xfsdatad_workqueue = create_workqueue("xfsdatad");
1852         if (!xfsdatad_workqueue)
1853                 goto out_destroy_xfslogd_workqueue;
1854
1855         xfsconvertd_workqueue = create_workqueue("xfsconvertd");
1856         if (!xfsconvertd_workqueue)
1857                 goto out_destroy_xfsdatad_workqueue;
1858
1859         return 0;
1860
1861  out_destroy_xfsdatad_workqueue:
1862         destroy_workqueue(xfsdatad_workqueue);
1863  out_destroy_xfslogd_workqueue:
1864         destroy_workqueue(xfslogd_workqueue);
1865  out_free_buf_zone:
1866         kmem_zone_destroy(xfs_buf_zone);
1867  out:
1868         return -ENOMEM;
1869 }
1870
1871 void
1872 xfs_buf_terminate(void)
1873 {
1874         destroy_workqueue(xfsconvertd_workqueue);
1875         destroy_workqueue(xfsdatad_workqueue);
1876         destroy_workqueue(xfslogd_workqueue);
1877         kmem_zone_destroy(xfs_buf_zone);
1878 }
1879
1880 #ifdef CONFIG_KDB_MODULES
1881 struct list_head *
1882 xfs_get_buftarg_list(void)
1883 {
1884         return &xfs_buftarg_list;
1885 }
1886 #endif