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
  37 #include "xfs_sb.h"
  38 #include "xfs_inum.h"
  39 #include "xfs_log.h"
  40 #include "xfs_ag.h"
  41 #include "xfs_mount.h"
  42 #include "xfs_trace.h"
  43
  44 static kmem_zone_t *xfs_buf_zone;
  45 STATIC int xfsbufd(void *);
  46 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
  47
  48 static struct workqueue_struct *xfslogd_workqueue;
  49 struct workqueue_struct *xfsdatad_workqueue;
  50 struct workqueue_struct *xfsconvertd_workqueue;
  51
  52 #ifdef XFS_BUF_LOCK_TRACKING
  53 # define XB_SET_OWNER(bp)       ((bp)->b_last_holder = current->pid)
  54 # define XB_CLEAR_OWNER(bp)     ((bp)->b_last_holder = -1)
  55 # define XB_GET_OWNER(bp)       ((bp)->b_last_holder)
  56 #else
  57 # define XB_SET_OWNER(bp)       do { } while (0)
  58 # define XB_CLEAR_OWNER(bp)     do { } while (0)
  59 # define XB_GET_OWNER(bp)       do { } while (0)
  60 #endif
  61
  62 #define xb_to_gfp(flags) \
  63         ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
  64           ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
  65
  66 #define xb_to_km(flags) \
  67          (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
  68
  69 #define xfs_buf_allocate(flags) \
  70         kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
  71 #define xfs_buf_deallocate(bp) \
  72         kmem_zone_free(xfs_buf_zone, (bp));
  73
  74 static inline int
  75 xfs_buf_is_vmapped(
  76         struct xfs_buf  *bp)
  77 {
  78         /*
  79          * Return true if the buffer is vmapped.
  80          *
  81          * The XBF_MAPPED flag is set if the buffer should be mapped, but the
  82          * code is clever enough to know it doesn't have to map a single page,
  83          * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
  84          */
  85         return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
  86 }
  87
  88 static inline int
  89 xfs_buf_vmap_len(
  90         struct xfs_buf  *bp)
  91 {
  92         return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
  93 }
  94
  95 /*
  96  * xfs_buf_lru_add - add a buffer to the LRU.
  97  *
  98  * The LRU takes a new reference to the buffer so that it will only be freed
  99  * once the shrinker takes the buffer off the LRU.
 100  */
 101 STATIC void
 102 xfs_buf_lru_add(
 103         struct xfs_buf  *bp)
 104 {
 105         struct xfs_buftarg *btp = bp->b_target;
 106
 107         spin_lock(&btp->bt_lru_lock);
 108         if (list_empty(&bp->b_lru)) {
 109                 atomic_inc(&bp->b_hold);
 110                 list_add_tail(&bp->b_lru, &btp->bt_lru);
 111                 btp->bt_lru_nr++;
 112         }
 113         spin_unlock(&btp->bt_lru_lock);
 114 }
 115
 116 /*
 117  * xfs_buf_lru_del - remove a buffer from the LRU
 118  *
 119  * The unlocked check is safe here because it only occurs when there are not
 120  * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
 121  * to optimise the shrinker removing the buffer from the LRU and calling
 122  * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
 123  * bt_lru_lock.
 124  */
 125 STATIC void
 126 xfs_buf_lru_del(
 127         struct xfs_buf  *bp)
 128 {
 129         struct xfs_buftarg *btp = bp->b_target;
 130
 131         if (list_empty(&bp->b_lru))
 132                 return;
 133
 134         spin_lock(&btp->bt_lru_lock);
 135         if (!list_empty(&bp->b_lru)) {
 136                 list_del_init(&bp->b_lru);
 137                 btp->bt_lru_nr--;
 138         }
 139         spin_unlock(&btp->bt_lru_lock);
 140 }
 141
 142 /*
 143  * When we mark a buffer stale, we remove the buffer from the LRU and clear the
 144  * b_lru_ref count so that the buffer is freed immediately when the buffer
 145  * reference count falls to zero. If the buffer is already on the LRU, we need
 146  * to remove the reference that LRU holds on the buffer.
 147  *
 148  * This prevents build-up of stale buffers on the LRU.
 149  */
 150 void
 151 xfs_buf_stale(
 152         struct xfs_buf  *bp)
 153 {
 154         bp->b_flags |= XBF_STALE;
 155         atomic_set(&(bp)->b_lru_ref, 0);
 156         if (!list_empty(&bp->b_lru)) {
 157                 struct xfs_buftarg *btp = bp->b_target;
 158
 159                 spin_lock(&btp->bt_lru_lock);
 160                 if (!list_empty(&bp->b_lru)) {
 161                         list_del_init(&bp->b_lru);
 162                         btp->bt_lru_nr--;
 163                         atomic_dec(&bp->b_hold);
 164                 }
 165                 spin_unlock(&btp->bt_lru_lock);
 166         }
 167         ASSERT(atomic_read(&bp->b_hold) >= 1);
 168 }
 169
 170 STATIC void
 171 _xfs_buf_initialize(
 172         xfs_buf_t               *bp,
 173         xfs_buftarg_t           *target,
 174         xfs_off_t               range_base,
 175         size_t                  range_length,
 176         xfs_buf_flags_t         flags)
 177 {
 178         /*
 179          * We don't want certain flags to appear in b_flags.
 180          */
 181         flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
 182
 183         memset(bp, 0, sizeof(xfs_buf_t));
 184         atomic_set(&bp->b_hold, 1);
 185         atomic_set(&bp->b_lru_ref, 1);
 186         init_completion(&bp->b_iowait);
 187         INIT_LIST_HEAD(&bp->b_lru);
 188         INIT_LIST_HEAD(&bp->b_list);
 189         RB_CLEAR_NODE(&bp->b_rbnode);
 190         sema_init(&bp->b_sema, 0); /* held, no waiters */
 191         XB_SET_OWNER(bp);
 192         bp->b_target = target;
 193         bp->b_file_offset = range_base;
 194         /*
 195          * Set buffer_length and count_desired to the same value initially.
 196          * I/O routines should use count_desired, which will be the same in
 197          * most cases but may be reset (e.g. XFS recovery).
 198          */
 199         bp->b_buffer_length = bp->b_count_desired = range_length;
 200         bp->b_flags = flags;
 201         bp->b_bn = XFS_BUF_DADDR_NULL;
 202         atomic_set(&bp->b_pin_count, 0);
 203         init_waitqueue_head(&bp->b_waiters);
 204
 205         XFS_STATS_INC(xb_create);
 206
 207         trace_xfs_buf_init(bp, _RET_IP_);
 208 }
 209
 210 /*
 211  *      Allocate a page array capable of holding a specified number
 212  *      of pages, and point the page buf at it.
 213  */
 214 STATIC int
 215 _xfs_buf_get_pages(
 216         xfs_buf_t               *bp,
 217         int                     page_count,
 218         xfs_buf_flags_t         flags)
 219 {
 220         /* Make sure that we have a page list */
 221         if (bp->b_pages == NULL) {
 222                 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
 223                 bp->b_page_count = page_count;
 224                 if (page_count <= XB_PAGES) {
 225                         bp->b_pages = bp->b_page_array;
 226                 } else {
 227                         bp->b_pages = kmem_alloc(sizeof(struct page *) *
 228                                         page_count, xb_to_km(flags));
 229                         if (bp->b_pages == NULL)
 230                                 return -ENOMEM;
 231                 }
 232                 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
 233         }
 234         return 0;
 235 }
 236
 237 /*
 238  *      Frees b_pages if it was allocated.
 239  */
 240 STATIC void
 241 _xfs_buf_free_pages(
 242         xfs_buf_t       *bp)
 243 {
 244         if (bp->b_pages != bp->b_page_array) {
 245                 kmem_free(bp->b_pages);
 246                 bp->b_pages = NULL;
 247         }
 248 }
 249
 250 /*
 251  *      Releases the specified buffer.
 252  *
 253  *      The modification state of any associated pages is left unchanged.
 254  *      The buffer most not be on any hash - use xfs_buf_rele instead for
 255  *      hashed and refcounted buffers
 256  */
 257 void
 258 xfs_buf_free(
 259         xfs_buf_t               *bp)
 260 {
 261         trace_xfs_buf_free(bp, _RET_IP_);
 262
 263         ASSERT(list_empty(&bp->b_lru));
 264
 265         if (bp->b_flags & _XBF_PAGES) {
 266                 uint            i;
 267
 268                 if (xfs_buf_is_vmapped(bp))
 269                         vm_unmap_ram(bp->b_addr - bp->b_offset,
 270                                         bp->b_page_count);
 271
 272                 for (i = 0; i < bp->b_page_count; i++) {
 273                         struct page     *page = bp->b_pages[i];
 274
 275                         __free_page(page);
 276                 }
 277         } else if (bp->b_flags & _XBF_KMEM)
 278                 kmem_free(bp->b_addr);
 279         _xfs_buf_free_pages(bp);
 280         xfs_buf_deallocate(bp);
 281 }
 282
 283 /*
 284  * Allocates all the pages for buffer in question and builds it's page list.
 285  */
 286 STATIC int
 287 xfs_buf_allocate_memory(
 288         xfs_buf_t               *bp,
 289         uint                    flags)
 290 {
 291         size_t                  size = bp->b_count_desired;
 292         size_t                  nbytes, offset;
 293         gfp_t                   gfp_mask = xb_to_gfp(flags);
 294         unsigned short          page_count, i;
 295         xfs_off_t               end;
 296         int                     error;
 297
 298         /*
 299          * for buffers that are contained within a single page, just allocate
 300          * the memory from the heap - there's no need for the complexity of
 301          * page arrays to keep allocation down to order 0.
 302          */
 303         if (bp->b_buffer_length < PAGE_SIZE) {
 304                 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
 305                 if (!bp->b_addr) {
 306                         /* low memory - use alloc_page loop instead */
 307                         goto use_alloc_page;
 308                 }
 309
 310                 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
 311                                                                 PAGE_MASK) !=
 312                     ((unsigned long)bp->b_addr & PAGE_MASK)) {
 313                         /* b_addr spans two pages - use alloc_page instead */
 314                         kmem_free(bp->b_addr);
 315                         bp->b_addr = NULL;
 316                         goto use_alloc_page;
 317                 }
 318                 bp->b_offset = offset_in_page(bp->b_addr);
 319                 bp->b_pages = bp->b_page_array;
 320                 bp->b_pages[0] = virt_to_page(bp->b_addr);
 321                 bp->b_page_count = 1;
 322                 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
 323                 return 0;
 324         }
 325
 326 use_alloc_page:
 327         end = bp->b_file_offset + bp->b_buffer_length;
 328         page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
 329         error = _xfs_buf_get_pages(bp, page_count, flags);
 330         if (unlikely(error))
 331                 return error;
 332
 333         offset = bp->b_offset;
 334         bp->b_flags |= _XBF_PAGES;
 335
 336         for (i = 0; i < bp->b_page_count; i++) {
 337                 struct page     *page;
 338                 uint            retries = 0;
 339 retry:
 340                 page = alloc_page(gfp_mask);
 341                 if (unlikely(page == NULL)) {
 342                         if (flags & XBF_READ_AHEAD) {
 343                                 bp->b_page_count = i;
 344                                 error = ENOMEM;
 345                                 goto out_free_pages;
 346                         }
 347
 348                         /*
 349                          * This could deadlock.
 350                          *
 351                          * But until all the XFS lowlevel code is revamped to
 352                          * handle buffer allocation failures we can't do much.
 353                          */
 354                         if (!(++retries % 100))
 355                                 xfs_err(NULL,
 356                 "possible memory allocation deadlock in %s (mode:0x%x)",
 357                                         __func__, gfp_mask);
 358
 359                         XFS_STATS_INC(xb_page_retries);
 360                         congestion_wait(BLK_RW_ASYNC, HZ/50);
 361                         goto retry;
 362                 }
 363
 364                 XFS_STATS_INC(xb_page_found);
 365
 366                 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
 367                 size -= nbytes;
 368                 bp->b_pages[i] = page;
 369                 offset = 0;
 370         }
 371         return 0;
 372
 373 out_free_pages:
 374         for (i = 0; i < bp->b_page_count; i++)
 375                 __free_page(bp->b_pages[i]);
 376         return error;
 377 }
 378
 379 /*
 380  *      Map buffer into kernel address-space if necessary.
 381  */
 382 STATIC int
 383 _xfs_buf_map_pages(
 384         xfs_buf_t               *bp,
 385         uint                    flags)
 386 {
 387         ASSERT(bp->b_flags & _XBF_PAGES);
 388         if (bp->b_page_count == 1) {
 389                 /* A single page buffer is always mappable */
 390                 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
 391                 bp->b_flags |= XBF_MAPPED;
 392         } else if (flags & XBF_MAPPED) {
 393                 int retried = 0;
 394
 395                 do {
 396                         bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
 397                                                 -1, PAGE_KERNEL);
 398                         if (bp->b_addr)
 399                                 break;
 400                         vm_unmap_aliases();
 401                 } while (retried++ <= 1);
 402
 403                 if (!bp->b_addr)
 404                         return -ENOMEM;
 405                 bp->b_addr += bp->b_offset;
 406                 bp->b_flags |= XBF_MAPPED;
 407         }
 408
 409         return 0;
 410 }
 411
 412 /*
 413  *      Finding and Reading Buffers
 414  */
 415
 416 /*
 417  *      Look up, and creates if absent, a lockable buffer for
 418  *      a given range of an inode.  The buffer is returned
 419  *      locked.  If other overlapping buffers exist, they are
 420  *      released before the new buffer is created and locked,
 421  *      which may imply that this call will block until those buffers
 422  *      are unlocked.  No I/O is implied by this call.
 423  */
 424 xfs_buf_t *
 425 _xfs_buf_find(
 426         xfs_buftarg_t           *btp,   /* block device target          */
 427         xfs_off_t               ioff,   /* starting offset of range     */
 428         size_t                  isize,  /* length of range              */
 429         xfs_buf_flags_t         flags,
 430         xfs_buf_t               *new_bp)
 431 {
 432         xfs_off_t               range_base;
 433         size_t                  range_length;
 434         struct xfs_perag        *pag;
 435         struct rb_node          **rbp;
 436         struct rb_node          *parent;
 437         xfs_buf_t               *bp;
 438
 439         range_base = (ioff << BBSHIFT);
 440         range_length = (isize << BBSHIFT);
 441
 442         /* Check for IOs smaller than the sector size / not sector aligned */
 443         ASSERT(!(range_length < (1 << btp->bt_sshift)));
 444         ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
 445
 446         /* get tree root */
 447         pag = xfs_perag_get(btp->bt_mount,
 448                                 xfs_daddr_to_agno(btp->bt_mount, ioff));
 449
 450         /* walk tree */
 451         spin_lock(&pag->pag_buf_lock);
 452         rbp = &pag->pag_buf_tree.rb_node;
 453         parent = NULL;
 454         bp = NULL;
 455         while (*rbp) {
 456                 parent = *rbp;
 457                 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
 458
 459                 if (range_base < bp->b_file_offset)
 460                         rbp = &(*rbp)->rb_left;
 461                 else if (range_base > bp->b_file_offset)
 462                         rbp = &(*rbp)->rb_right;
 463                 else {
 464                         /*
 465                          * found a block offset match. If the range doesn't
 466                          * match, the only way this is allowed is if the buffer
 467                          * in the cache is stale and the transaction that made
 468                          * it stale has not yet committed. i.e. we are
 469                          * reallocating a busy extent. Skip this buffer and
 470                          * continue searching to the right for an exact match.
 471                          */
 472                         if (bp->b_buffer_length != range_length) {
 473                                 ASSERT(bp->b_flags & XBF_STALE);
 474                                 rbp = &(*rbp)->rb_right;
 475                                 continue;
 476                         }
 477                         atomic_inc(&bp->b_hold);
 478                         goto found;
 479                 }
 480         }
 481
 482         /* No match found */
 483         if (new_bp) {
 484                 _xfs_buf_initialize(new_bp, btp, range_base,
 485                                 range_length, flags);
 486                 rb_link_node(&new_bp->b_rbnode, parent, rbp);
 487                 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
 488                 /* the buffer keeps the perag reference until it is freed */
 489                 new_bp->b_pag = pag;
 490                 spin_unlock(&pag->pag_buf_lock);
 491         } else {
 492                 XFS_STATS_INC(xb_miss_locked);
 493                 spin_unlock(&pag->pag_buf_lock);
 494                 xfs_perag_put(pag);
 495         }
 496         return new_bp;
 497
 498 found:
 499         spin_unlock(&pag->pag_buf_lock);
 500         xfs_perag_put(pag);
 501
 502         if (!xfs_buf_trylock(bp)) {
 503                 if (flags & XBF_TRYLOCK) {
 504                         xfs_buf_rele(bp);
 505                         XFS_STATS_INC(xb_busy_locked);
 506                         return NULL;
 507                 }
 508                 xfs_buf_lock(bp);
 509                 XFS_STATS_INC(xb_get_locked_waited);
 510         }
 511
 512         /*
 513          * if the buffer is stale, clear all the external state associated with
 514          * it. We need to keep flags such as how we allocated the buffer memory
 515          * intact here.
 516          */
 517         if (bp->b_flags & XBF_STALE) {
 518                 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
 519                 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
 520         }
 521
 522         trace_xfs_buf_find(bp, flags, _RET_IP_);
 523         XFS_STATS_INC(xb_get_locked);
 524         return bp;
 525 }
 526
 527 /*
 528  *      Assembles a buffer covering the specified range.
 529  *      Storage in memory for all portions of the buffer will be allocated,
 530  *      although backing storage may not be.
 531  */
 532 xfs_buf_t *
 533 xfs_buf_get(
 534         xfs_buftarg_t           *target,/* target for buffer            */
 535         xfs_off_t               ioff,   /* starting offset of range     */
 536         size_t                  isize,  /* length of range              */
 537         xfs_buf_flags_t         flags)
 538 {
 539         xfs_buf_t               *bp, *new_bp;
 540         int                     error = 0;
 541
 542         new_bp = xfs_buf_allocate(flags);
 543         if (unlikely(!new_bp))
 544                 return NULL;
 545
 546         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
 547         if (bp == new_bp) {
 548                 error = xfs_buf_allocate_memory(bp, flags);
 549                 if (error)
 550                         goto no_buffer;
 551         } else {
 552                 xfs_buf_deallocate(new_bp);
 553                 if (unlikely(bp == NULL))
 554                         return NULL;
 555         }
 556
 557         if (!(bp->b_flags & XBF_MAPPED)) {
 558                 error = _xfs_buf_map_pages(bp, flags);
 559                 if (unlikely(error)) {
 560                         xfs_warn(target->bt_mount,
 561                                 "%s: failed to map pages\n", __func__);
 562                         goto no_buffer;
 563                 }
 564         }
 565
 566         XFS_STATS_INC(xb_get);
 567
 568         /*
 569          * Always fill in the block number now, the mapped cases can do
 570          * their own overlay of this later.
 571          */
 572         bp->b_bn = ioff;
 573         bp->b_count_desired = bp->b_buffer_length;
 574
 575         trace_xfs_buf_get(bp, flags, _RET_IP_);
 576         return bp;
 577
 578  no_buffer:
 579         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 580                 xfs_buf_unlock(bp);
 581         xfs_buf_rele(bp);
 582         return NULL;
 583 }
 584
 585 STATIC int
 586 _xfs_buf_read(
 587         xfs_buf_t               *bp,
 588         xfs_buf_flags_t         flags)
 589 {
 590         int                     status;
 591
 592         ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
 593         ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
 594
 595         bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
 596                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 597         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
 598                         XBF_READ_AHEAD | _XBF_RUN_QUEUES);
 599
 600         status = xfs_buf_iorequest(bp);
 601         if (status || XFS_BUF_ISERROR(bp) || (flags & XBF_ASYNC))
 602                 return status;
 603         return xfs_buf_iowait(bp);
 604 }
 605
 606 xfs_buf_t *
 607 xfs_buf_read(
 608         xfs_buftarg_t           *target,
 609         xfs_off_t               ioff,
 610         size_t                  isize,
 611         xfs_buf_flags_t         flags)
 612 {
 613         xfs_buf_t               *bp;
 614
 615         flags |= XBF_READ;
 616
 617         bp = xfs_buf_get(target, ioff, isize, flags);
 618         if (bp) {
 619                 trace_xfs_buf_read(bp, flags, _RET_IP_);
 620
 621                 if (!XFS_BUF_ISDONE(bp)) {
 622                         XFS_STATS_INC(xb_get_read);
 623                         _xfs_buf_read(bp, flags);
 624                 } else if (flags & XBF_ASYNC) {
 625                         /*
 626                          * Read ahead call which is already satisfied,
 627                          * drop the buffer
 628                          */
 629                         goto no_buffer;
 630                 } else {
 631                         /* We do not want read in the flags */
 632                         bp->b_flags &= ~XBF_READ;
 633                 }
 634         }
 635
 636         return bp;
 637
 638  no_buffer:
 639         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 640                 xfs_buf_unlock(bp);
 641         xfs_buf_rele(bp);
 642         return NULL;
 643 }
 644
 645 /*
 646  *      If we are not low on memory then do the readahead in a deadlock
 647  *      safe manner.
 648  */
 649 void
 650 xfs_buf_readahead(
 651         xfs_buftarg_t           *target,
 652         xfs_off_t               ioff,
 653         size_t                  isize)
 654 {
 655         if (bdi_read_congested(target->bt_bdi))
 656                 return;
 657
 658         xfs_buf_read(target, ioff, isize,
 659                      XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
 660 }
 661
 662 /*
 663  * Read an uncached buffer from disk. Allocates and returns a locked
 664  * buffer containing the disk contents or nothing.
 665  */
 666 struct xfs_buf *
 667 xfs_buf_read_uncached(
 668         struct xfs_mount        *mp,
 669         struct xfs_buftarg      *target,
 670         xfs_daddr_t             daddr,
 671         size_t                  length,
 672         int                     flags)
 673 {
 674         xfs_buf_t               *bp;
 675         int                     error;
 676
 677         bp = xfs_buf_get_uncached(target, length, flags);
 678         if (!bp)
 679                 return NULL;
 680
 681         /* set up the buffer for a read IO */
 682         xfs_buf_lock(bp);
 683         XFS_BUF_SET_ADDR(bp, daddr);
 684         XFS_BUF_READ(bp);
 685         XFS_BUF_BUSY(bp);
 686
 687         xfsbdstrat(mp, bp);
 688         error = xfs_buf_iowait(bp);
 689         if (error || bp->b_error) {
 690                 xfs_buf_relse(bp);
 691                 return NULL;
 692         }
 693         return bp;
 694 }
 695
 696 xfs_buf_t *
 697 xfs_buf_get_empty(
 698         size_t                  len,
 699         xfs_buftarg_t           *target)
 700 {
 701         xfs_buf_t               *bp;
 702
 703         bp = xfs_buf_allocate(0);
 704         if (bp)
 705                 _xfs_buf_initialize(bp, target, 0, len, 0);
 706         return bp;
 707 }
 708
 709 /*
 710  * Return a buffer allocated as an empty buffer and associated to external
 711  * memory via xfs_buf_associate_memory() back to it's empty state.
 712  */
 713 void
 714 xfs_buf_set_empty(
 715         struct xfs_buf          *bp,
 716         size_t                  len)
 717 {
 718         if (bp->b_pages)
 719                 _xfs_buf_free_pages(bp);
 720
 721         bp->b_pages = NULL;
 722         bp->b_page_count = 0;
 723         bp->b_addr = NULL;
 724         bp->b_file_offset = 0;
 725         bp->b_buffer_length = bp->b_count_desired = len;
 726         bp->b_bn = XFS_BUF_DADDR_NULL;
 727         bp->b_flags &= ~XBF_MAPPED;
 728 }
 729
 730 static inline struct page *
 731 mem_to_page(
 732         void                    *addr)
 733 {
 734         if ((!is_vmalloc_addr(addr))) {
 735                 return virt_to_page(addr);
 736         } else {
 737                 return vmalloc_to_page(addr);
 738         }
 739 }
 740
 741 int
 742 xfs_buf_associate_memory(
 743         xfs_buf_t               *bp,
 744         void                    *mem,
 745         size_t                  len)
 746 {
 747         int                     rval;
 748         int                     i = 0;
 749         unsigned long           pageaddr;
 750         unsigned long           offset;
 751         size_t                  buflen;
 752         int                     page_count;
 753
 754         pageaddr = (unsigned long)mem & PAGE_MASK;
 755         offset = (unsigned long)mem - pageaddr;
 756         buflen = PAGE_ALIGN(len + offset);
 757         page_count = buflen >> PAGE_SHIFT;
 758
 759         /* Free any previous set of page pointers */
 760         if (bp->b_pages)
 761                 _xfs_buf_free_pages(bp);
 762
 763         bp->b_pages = NULL;
 764         bp->b_addr = mem;
 765
 766         rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
 767         if (rval)
 768                 return rval;
 769
 770         bp->b_offset = offset;
 771
 772         for (i = 0; i < bp->b_page_count; i++) {
 773                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 774                 pageaddr += PAGE_SIZE;
 775         }
 776
 777         bp->b_count_desired = len;
 778         bp->b_buffer_length = buflen;
 779         bp->b_flags |= XBF_MAPPED;
 780
 781         return 0;
 782 }
 783
 784 xfs_buf_t *
 785 xfs_buf_get_uncached(
 786         struct xfs_buftarg      *target,
 787         size_t                  len,
 788         int                     flags)
 789 {
 790         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 791         int                     error, i;
 792         xfs_buf_t               *bp;
 793
 794         bp = xfs_buf_allocate(0);
 795         if (unlikely(bp == NULL))
 796                 goto fail;
 797         _xfs_buf_initialize(bp, target, 0, len, 0);
 798
 799         error = _xfs_buf_get_pages(bp, page_count, 0);
 800         if (error)
 801                 goto fail_free_buf;
 802
 803         for (i = 0; i < page_count; i++) {
 804                 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
 805                 if (!bp->b_pages[i])
 806                         goto fail_free_mem;
 807         }
 808         bp->b_flags |= _XBF_PAGES;
 809
 810         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 811         if (unlikely(error)) {
 812                 xfs_warn(target->bt_mount,
 813                         "%s: failed to map pages\n", __func__);
 814                 goto fail_free_mem;
 815         }
 816
 817         xfs_buf_unlock(bp);
 818
 819         trace_xfs_buf_get_uncached(bp, _RET_IP_);
 820         return bp;
 821
 822  fail_free_mem:
 823         while (--i >= 0)
 824                 __free_page(bp->b_pages[i]);
 825         _xfs_buf_free_pages(bp);
 826  fail_free_buf:
 827         xfs_buf_deallocate(bp);
 828  fail:
 829         return NULL;
 830 }
 831
 832 /*
 833  *      Increment reference count on buffer, to hold the buffer concurrently
 834  *      with another thread which may release (free) the buffer asynchronously.
 835  *      Must hold the buffer already to call this function.
 836  */
 837 void
 838 xfs_buf_hold(
 839         xfs_buf_t               *bp)
 840 {
 841         trace_xfs_buf_hold(bp, _RET_IP_);
 842         atomic_inc(&bp->b_hold);
 843 }
 844
 845 /*
 846  *      Releases a hold on the specified buffer.  If the
 847  *      the hold count is 1, calls xfs_buf_free.
 848  */
 849 void
 850 xfs_buf_rele(
 851         xfs_buf_t               *bp)
 852 {
 853         struct xfs_perag        *pag = bp->b_pag;
 854
 855         trace_xfs_buf_rele(bp, _RET_IP_);
 856
 857         if (!pag) {
 858                 ASSERT(list_empty(&bp->b_lru));
 859                 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
 860                 if (atomic_dec_and_test(&bp->b_hold))
 861                         xfs_buf_free(bp);
 862                 return;
 863         }
 864
 865         ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
 866
 867         ASSERT(atomic_read(&bp->b_hold) > 0);
 868         if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
 869                 if (!(bp->b_flags & XBF_STALE) &&
 870                            atomic_read(&bp->b_lru_ref)) {
 871                         xfs_buf_lru_add(bp);
 872                         spin_unlock(&pag->pag_buf_lock);
 873                 } else {
 874                         xfs_buf_lru_del(bp);
 875                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 876                         rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
 877                         spin_unlock(&pag->pag_buf_lock);
 878                         xfs_perag_put(pag);
 879                         xfs_buf_free(bp);
 880                 }
 881         }
 882 }
 883
 884
 885 /*
 886  *      Lock a buffer object, if it is not already locked.
 887  *
 888  *      If we come across a stale, pinned, locked buffer, we know that we are
 889  *      being asked to lock a buffer that has been reallocated. Because it is
 890  *      pinned, we know that the log has not been pushed to disk and hence it
 891  *      will still be locked.  Rather than continuing to have trylock attempts
 892  *      fail until someone else pushes the log, push it ourselves before
 893  *      returning.  This means that the xfsaild will not get stuck trying
 894  *      to push on stale inode buffers.
 895  */
 896 int
 897 xfs_buf_trylock(
 898         struct xfs_buf          *bp)
 899 {
 900         int                     locked;
 901
 902         locked = down_trylock(&bp->b_sema) == 0;
 903         if (locked)
 904                 XB_SET_OWNER(bp);
 905         else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 906                 xfs_log_force(bp->b_target->bt_mount, 0);
 907
 908         trace_xfs_buf_trylock(bp, _RET_IP_);
 909         return locked;
 910 }
 911
 912 /*
 913  *      Lock a buffer object.
 914  *
 915  *      If we come across a stale, pinned, locked buffer, we know that we
 916  *      are being asked to lock a buffer that has been reallocated. Because
 917  *      it is pinned, we know that the log has not been pushed to disk and
 918  *      hence it will still be locked. Rather than sleeping until someone
 919  *      else pushes the log, push it ourselves before trying to get the lock.
 920  */
 921 void
 922 xfs_buf_lock(
 923         struct xfs_buf          *bp)
 924 {
 925         trace_xfs_buf_lock(bp, _RET_IP_);
 926
 927         if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 928                 xfs_log_force(bp->b_target->bt_mount, 0);
 929         down(&bp->b_sema);
 930         XB_SET_OWNER(bp);
 931
 932         trace_xfs_buf_lock_done(bp, _RET_IP_);
 933 }
 934
 935 /*
 936  *      Releases the lock on the buffer object.
 937  *      If the buffer is marked delwri but is not queued, do so before we
 938  *      unlock the buffer as we need to set flags correctly.  We also need to
 939  *      take a reference for the delwri queue because the unlocker is going to
 940  *      drop their's and they don't know we just queued it.
 941  */
 942 void
 943 xfs_buf_unlock(
 944         struct xfs_buf          *bp)
 945 {
 946         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 947                 atomic_inc(&bp->b_hold);
 948                 bp->b_flags |= XBF_ASYNC;
 949                 xfs_buf_delwri_queue(bp, 0);
 950         }
 951
 952         XB_CLEAR_OWNER(bp);
 953         up(&bp->b_sema);
 954
 955         trace_xfs_buf_unlock(bp, _RET_IP_);
 956 }
 957
 958 STATIC void
 959 xfs_buf_wait_unpin(
 960         xfs_buf_t               *bp)
 961 {
 962         DECLARE_WAITQUEUE       (wait, current);
 963
 964         if (atomic_read(&bp->b_pin_count) == 0)
 965                 return;
 966
 967         add_wait_queue(&bp->b_waiters, &wait);
 968         for (;;) {
 969                 set_current_state(TASK_UNINTERRUPTIBLE);
 970                 if (atomic_read(&bp->b_pin_count) == 0)
 971                         break;
 972                 io_schedule();
 973         }
 974         remove_wait_queue(&bp->b_waiters, &wait);
 975         set_current_state(TASK_RUNNING);
 976 }
 977
 978 /*
 979  *      Buffer Utility Routines
 980  */
 981
 982 STATIC void
 983 xfs_buf_iodone_work(
 984         struct work_struct      *work)
 985 {
 986         xfs_buf_t               *bp =
 987                 container_of(work, xfs_buf_t, b_iodone_work);
 988
 989         if (bp->b_iodone)
 990                 (*(bp->b_iodone))(bp);
 991         else if (bp->b_flags & XBF_ASYNC)
 992                 xfs_buf_relse(bp);
 993 }
 994
 995 void
 996 xfs_buf_ioend(
 997         xfs_buf_t               *bp,
 998         int                     schedule)
 999 {
1000         trace_xfs_buf_iodone(bp, _RET_IP_);
1001
1002         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1003         if (bp->b_error == 0)
1004                 bp->b_flags |= XBF_DONE;
1005
1006         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1007                 if (schedule) {
1008                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1009                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1010                 } else {
1011                         xfs_buf_iodone_work(&bp->b_iodone_work);
1012                 }
1013         } else {
1014                 complete(&bp->b_iowait);
1015         }
1016 }
1017
1018 void
1019 xfs_buf_ioerror(
1020         xfs_buf_t               *bp,
1021         int                     error)
1022 {
1023         ASSERT(error >= 0 && error <= 0xffff);
1024         bp->b_error = (unsigned short)error;
1025         trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1026 }
1027
1028 int
1029 xfs_bwrite(
1030         struct xfs_mount        *mp,
1031         struct xfs_buf          *bp)
1032 {
1033         int                     error;
1034
1035         bp->b_flags |= XBF_WRITE;
1036         bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1037
1038         xfs_buf_delwri_dequeue(bp);
1039         xfs_bdstrat_cb(bp);
1040
1041         error = xfs_buf_iowait(bp);
1042         if (error)
1043                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1044         xfs_buf_relse(bp);
1045         return error;
1046 }
1047
1048 void
1049 xfs_bdwrite(
1050         void                    *mp,
1051         struct xfs_buf          *bp)
1052 {
1053         trace_xfs_buf_bdwrite(bp, _RET_IP_);
1054
1055         bp->b_flags &= ~XBF_READ;
1056         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1057
1058         xfs_buf_delwri_queue(bp, 1);
1059 }
1060
1061 /*
1062  * Called when we want to stop a buffer from getting written or read.
1063  * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1064  * so that the proper iodone callbacks get called.
1065  */
1066 STATIC int
1067 xfs_bioerror(
1068         xfs_buf_t *bp)
1069 {
1070 #ifdef XFSERRORDEBUG
1071         ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1072 #endif
1073
1074         /*
1075          * No need to wait until the buffer is unpinned, we aren't flushing it.
1076          */
1077         XFS_BUF_ERROR(bp, EIO);
1078
1079         /*
1080          * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1081          */
1082         XFS_BUF_UNREAD(bp);
1083         XFS_BUF_UNDELAYWRITE(bp);
1084         XFS_BUF_UNDONE(bp);
1085         XFS_BUF_STALE(bp);
1086
1087         xfs_buf_ioend(bp, 0);
1088
1089         return EIO;
1090 }
1091
1092 /*
1093  * Same as xfs_bioerror, except that we are releasing the buffer
1094  * here ourselves, and avoiding the xfs_buf_ioend call.
1095  * This is meant for userdata errors; metadata bufs come with
1096  * iodone functions attached, so that we can track down errors.
1097  */
1098 STATIC int
1099 xfs_bioerror_relse(
1100         struct xfs_buf  *bp)
1101 {
1102         int64_t         fl = XFS_BUF_BFLAGS(bp);
1103         /*
1104          * No need to wait until the buffer is unpinned.
1105          * We aren't flushing it.
1106          *
1107          * chunkhold expects B_DONE to be set, whether
1108          * we actually finish the I/O or not. We don't want to
1109          * change that interface.
1110          */
1111         XFS_BUF_UNREAD(bp);
1112         XFS_BUF_UNDELAYWRITE(bp);
1113         XFS_BUF_DONE(bp);
1114         XFS_BUF_STALE(bp);
1115         XFS_BUF_CLR_IODONE_FUNC(bp);
1116         if (!(fl & XBF_ASYNC)) {
1117                 /*
1118                  * Mark b_error and B_ERROR _both_.
1119                  * Lot's of chunkcache code assumes that.
1120                  * There's no reason to mark error for
1121                  * ASYNC buffers.
1122                  */
1123                 XFS_BUF_ERROR(bp, EIO);
1124                 XFS_BUF_FINISH_IOWAIT(bp);
1125         } else {
1126                 xfs_buf_relse(bp);
1127         }
1128
1129         return EIO;
1130 }
1131
1132
1133 /*
1134  * All xfs metadata buffers except log state machine buffers
1135  * get this attached as their b_bdstrat callback function.
1136  * This is so that we can catch a buffer
1137  * after prematurely unpinning it to forcibly shutdown the filesystem.
1138  */
1139 int
1140 xfs_bdstrat_cb(
1141         struct xfs_buf  *bp)
1142 {
1143         if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1144                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1145                 /*
1146                  * Metadata write that didn't get logged but
1147                  * written delayed anyway. These aren't associated
1148                  * with a transaction, and can be ignored.
1149                  */
1150                 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1151                         return xfs_bioerror_relse(bp);
1152                 else
1153                         return xfs_bioerror(bp);
1154         }
1155
1156         xfs_buf_iorequest(bp);
1157         return 0;
1158 }
1159
1160 /*
1161  * Wrapper around bdstrat so that we can stop data from going to disk in case
1162  * we are shutting down the filesystem.  Typically user data goes thru this
1163  * path; one of the exceptions is the superblock.
1164  */
1165 void
1166 xfsbdstrat(
1167         struct xfs_mount        *mp,
1168         struct xfs_buf          *bp)
1169 {
1170         if (XFS_FORCED_SHUTDOWN(mp)) {
1171                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1172                 xfs_bioerror_relse(bp);
1173                 return;
1174         }
1175
1176         xfs_buf_iorequest(bp);
1177 }
1178
1179 STATIC void
1180 _xfs_buf_ioend(
1181         xfs_buf_t               *bp,
1182         int                     schedule)
1183 {
1184         if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1185                 xfs_buf_ioend(bp, schedule);
1186 }
1187
1188 STATIC void
1189 xfs_buf_bio_end_io(
1190         struct bio              *bio,
1191         int                     error)
1192 {
1193         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1194
1195         xfs_buf_ioerror(bp, -error);
1196
1197         if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1198                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1199
1200         _xfs_buf_ioend(bp, 1);
1201         bio_put(bio);
1202 }
1203
1204 STATIC void
1205 _xfs_buf_ioapply(
1206         xfs_buf_t               *bp)
1207 {
1208         int                     rw, map_i, total_nr_pages, nr_pages;
1209         struct bio              *bio;
1210         int                     offset = bp->b_offset;
1211         int                     size = bp->b_count_desired;
1212         sector_t                sector = bp->b_bn;
1213
1214         total_nr_pages = bp->b_page_count;
1215         map_i = 0;
1216
1217         if (bp->b_flags & XBF_ORDERED) {
1218                 ASSERT(!(bp->b_flags & XBF_READ));
1219                 rw = WRITE_FLUSH_FUA;
1220         } else if (bp->b_flags & XBF_LOG_BUFFER) {
1221                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1222                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1223                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1224         } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1225                 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1226                 bp->b_flags &= ~_XBF_RUN_QUEUES;
1227                 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1228         } else {
1229                 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1230                      (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1231         }
1232
1233
1234 next_chunk:
1235         atomic_inc(&bp->b_io_remaining);
1236         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1237         if (nr_pages > total_nr_pages)
1238                 nr_pages = total_nr_pages;
1239
1240         bio = bio_alloc(GFP_NOIO, nr_pages);
1241         bio->bi_bdev = bp->b_target->bt_bdev;
1242         bio->bi_sector = sector;
1243         bio->bi_end_io = xfs_buf_bio_end_io;
1244         bio->bi_private = bp;
1245
1246
1247         for (; size && nr_pages; nr_pages--, map_i++) {
1248                 int     rbytes, nbytes = PAGE_SIZE - offset;
1249
1250                 if (nbytes > size)
1251                         nbytes = size;
1252
1253                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1254                 if (rbytes < nbytes)
1255                         break;
1256
1257                 offset = 0;
1258                 sector += nbytes >> BBSHIFT;
1259                 size -= nbytes;
1260                 total_nr_pages--;
1261         }
1262
1263         if (likely(bio->bi_size)) {
1264                 if (xfs_buf_is_vmapped(bp)) {
1265                         flush_kernel_vmap_range(bp->b_addr,
1266                                                 xfs_buf_vmap_len(bp));
1267                 }
1268                 submit_bio(rw, bio);
1269                 if (size)
1270                         goto next_chunk;
1271         } else {
1272                 xfs_buf_ioerror(bp, EIO);
1273                 bio_put(bio);
1274         }
1275 }
1276
1277 int
1278 xfs_buf_iorequest(
1279         xfs_buf_t               *bp)
1280 {
1281         trace_xfs_buf_iorequest(bp, _RET_IP_);
1282
1283         if (bp->b_flags & XBF_DELWRI) {
1284                 xfs_buf_delwri_queue(bp, 1);
1285                 return 0;
1286         }
1287
1288         if (bp->b_flags & XBF_WRITE) {
1289                 xfs_buf_wait_unpin(bp);
1290         }
1291
1292         xfs_buf_hold(bp);
1293
1294         /* Set the count to 1 initially, this will stop an I/O
1295          * completion callout which happens before we have started
1296          * all the I/O from calling xfs_buf_ioend too early.
1297          */
1298         atomic_set(&bp->b_io_remaining, 1);
1299         _xfs_buf_ioapply(bp);
1300         _xfs_buf_ioend(bp, 0);
1301
1302         xfs_buf_rele(bp);
1303         return 0;
1304 }
1305
1306 /*
1307  *      Waits for I/O to complete on the buffer supplied.
1308  *      It returns immediately if no I/O is pending.
1309  *      It returns the I/O error code, if any, or 0 if there was no error.
1310  */
1311 int
1312 xfs_buf_iowait(
1313         xfs_buf_t               *bp)
1314 {
1315         trace_xfs_buf_iowait(bp, _RET_IP_);
1316
1317         wait_for_completion(&bp->b_iowait);
1318
1319         trace_xfs_buf_iowait_done(bp, _RET_IP_);
1320         return bp->b_error;
1321 }
1322
1323 xfs_caddr_t
1324 xfs_buf_offset(
1325         xfs_buf_t               *bp,
1326         size_t                  offset)
1327 {
1328         struct page             *page;
1329
1330         if (bp->b_flags & XBF_MAPPED)
1331                 return XFS_BUF_PTR(bp) + offset;
1332
1333         offset += bp->b_offset;
1334         page = bp->b_pages[offset >> PAGE_SHIFT];
1335         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1336 }
1337
1338 /*
1339  *      Move data into or out of a buffer.
1340  */
1341 void
1342 xfs_buf_iomove(
1343         xfs_buf_t               *bp,    /* buffer to process            */
1344         size_t                  boff,   /* starting buffer offset       */
1345         size_t                  bsize,  /* length to copy               */
1346         void                    *data,  /* data address                 */
1347         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1348 {
1349         size_t                  bend, cpoff, csize;
1350         struct page             *page;
1351
1352         bend = boff + bsize;
1353         while (boff < bend) {
1354                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1355                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1356                 csize = min_t(size_t,
1357                               PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1358
1359                 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1360
1361                 switch (mode) {
1362                 case XBRW_ZERO:
1363                         memset(page_address(page) + cpoff, 0, csize);
1364                         break;
1365                 case XBRW_READ:
1366                         memcpy(data, page_address(page) + cpoff, csize);
1367                         break;
1368                 case XBRW_WRITE:
1369                         memcpy(page_address(page) + cpoff, data, csize);
1370                 }
1371
1372                 boff += csize;
1373                 data += csize;
1374         }
1375 }
1376
1377 /*
1378  *      Handling of buffer targets (buftargs).
1379  */
1380
1381 /*
1382  * Wait for any bufs with callbacks that have been submitted but have not yet
1383  * returned. These buffers will have an elevated hold count, so wait on those
1384  * while freeing all the buffers only held by the LRU.
1385  */
1386 void
1387 xfs_wait_buftarg(
1388         struct xfs_buftarg      *btp)
1389 {
1390         struct xfs_buf          *bp;
1391
1392 restart:
1393         spin_lock(&btp->bt_lru_lock);
1394         while (!list_empty(&btp->bt_lru)) {
1395                 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1396                 if (atomic_read(&bp->b_hold) > 1) {
1397                         spin_unlock(&btp->bt_lru_lock);
1398                         delay(100);
1399                         goto restart;
1400                 }
1401                 /*
1402                  * clear the LRU reference count so the bufer doesn't get
1403                  * ignored in xfs_buf_rele().
1404                  */
1405                 atomic_set(&bp->b_lru_ref, 0);
1406                 spin_unlock(&btp->bt_lru_lock);
1407                 xfs_buf_rele(bp);
1408                 spin_lock(&btp->bt_lru_lock);
1409         }
1410         spin_unlock(&btp->bt_lru_lock);
1411 }
1412
1413 int
1414 xfs_buftarg_shrink(
1415         struct shrinker         *shrink,
1416         struct shrink_control   *sc)
1417 {
1418         struct xfs_buftarg      *btp = container_of(shrink,
1419                                         struct xfs_buftarg, bt_shrinker);
1420         struct xfs_buf          *bp;
1421         int nr_to_scan = sc->nr_to_scan;
1422         LIST_HEAD(dispose);
1423
1424         if (!nr_to_scan)
1425                 return btp->bt_lru_nr;
1426
1427         spin_lock(&btp->bt_lru_lock);
1428         while (!list_empty(&btp->bt_lru)) {
1429                 if (nr_to_scan-- <= 0)
1430                         break;
1431
1432                 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1433
1434                 /*
1435                  * Decrement the b_lru_ref count unless the value is already
1436                  * zero. If the value is already zero, we need to reclaim the
1437                  * buffer, otherwise it gets another trip through the LRU.
1438                  */
1439                 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1440                         list_move_tail(&bp->b_lru, &btp->bt_lru);
1441                         continue;
1442                 }
1443
1444                 /*
1445                  * remove the buffer from the LRU now to avoid needing another
1446                  * lock round trip inside xfs_buf_rele().
1447                  */
1448                 list_move(&bp->b_lru, &dispose);
1449                 btp->bt_lru_nr--;
1450         }
1451         spin_unlock(&btp->bt_lru_lock);
1452
1453         while (!list_empty(&dispose)) {
1454                 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1455                 list_del_init(&bp->b_lru);
1456                 xfs_buf_rele(bp);
1457         }
1458
1459         return btp->bt_lru_nr;
1460 }
1461
1462 void
1463 xfs_free_buftarg(
1464         struct xfs_mount        *mp,
1465         struct xfs_buftarg      *btp)
1466 {
1467         unregister_shrinker(&btp->bt_shrinker);
1468
1469         xfs_flush_buftarg(btp, 1);
1470         if (mp->m_flags & XFS_MOUNT_BARRIER)
1471                 xfs_blkdev_issue_flush(btp);
1472
1473         kthread_stop(btp->bt_task);
1474         kmem_free(btp);
1475 }
1476
1477 STATIC int
1478 xfs_setsize_buftarg_flags(
1479         xfs_buftarg_t           *btp,
1480         unsigned int            blocksize,
1481         unsigned int            sectorsize,
1482         int                     verbose)
1483 {
1484         btp->bt_bsize = blocksize;
1485         btp->bt_sshift = ffs(sectorsize) - 1;
1486         btp->bt_smask = sectorsize - 1;
1487
1488         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1489                 xfs_warn(btp->bt_mount,
1490                         "Cannot set_blocksize to %u on device %s\n",
1491                         sectorsize, XFS_BUFTARG_NAME(btp));
1492                 return EINVAL;
1493         }
1494
1495         return 0;
1496 }
1497
1498 /*
1499  *      When allocating the initial buffer target we have not yet
1500  *      read in the superblock, so don't know what sized sectors
1501  *      are being used is at this early stage.  Play safe.
1502  */
1503 STATIC int
1504 xfs_setsize_buftarg_early(
1505         xfs_buftarg_t           *btp,
1506         struct block_device     *bdev)
1507 {
1508         return xfs_setsize_buftarg_flags(btp,
1509                         PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1510 }
1511
1512 int
1513 xfs_setsize_buftarg(
1514         xfs_buftarg_t           *btp,
1515         unsigned int            blocksize,
1516         unsigned int            sectorsize)
1517 {
1518         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1519 }
1520
1521 STATIC int
1522 xfs_alloc_delwrite_queue(
1523         xfs_buftarg_t           *btp,
1524         const char              *fsname)
1525 {
1526         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1527         spin_lock_init(&btp->bt_delwrite_lock);
1528         btp->bt_flags = 0;
1529         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1530         if (IS_ERR(btp->bt_task))
1531                 return PTR_ERR(btp->bt_task);
1532         return 0;
1533 }
1534
1535 xfs_buftarg_t *
1536 xfs_alloc_buftarg(
1537         struct xfs_mount        *mp,
1538         struct block_device     *bdev,
1539         int                     external,
1540         const char              *fsname)
1541 {
1542         xfs_buftarg_t           *btp;
1543
1544         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1545
1546         btp->bt_mount = mp;
1547         btp->bt_dev =  bdev->bd_dev;
1548         btp->bt_bdev = bdev;
1549         btp->bt_bdi = blk_get_backing_dev_info(bdev);
1550         if (!btp->bt_bdi)
1551                 goto error;
1552
1553         INIT_LIST_HEAD(&btp->bt_lru);
1554         spin_lock_init(&btp->bt_lru_lock);
1555         if (xfs_setsize_buftarg_early(btp, bdev))
1556                 goto error;
1557         if (xfs_alloc_delwrite_queue(btp, fsname))
1558                 goto error;
1559         btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1560         btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1561         register_shrinker(&btp->bt_shrinker);
1562         return btp;
1563
1564 error:
1565         kmem_free(btp);
1566         return NULL;
1567 }
1568
1569
1570 /*
1571  *      Delayed write buffer handling
1572  */
1573 STATIC void
1574 xfs_buf_delwri_queue(
1575         xfs_buf_t               *bp,
1576         int                     unlock)
1577 {
1578         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1579         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1580
1581         trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1582
1583         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1584
1585         spin_lock(dwlk);
1586         /* If already in the queue, dequeue and place at tail */
1587         if (!list_empty(&bp->b_list)) {
1588                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1589                 if (unlock)
1590                         atomic_dec(&bp->b_hold);
1591                 list_del(&bp->b_list);
1592         }
1593
1594         if (list_empty(dwq)) {
1595                 /* start xfsbufd as it is about to have something to do */
1596                 wake_up_process(bp->b_target->bt_task);
1597         }
1598
1599         bp->b_flags |= _XBF_DELWRI_Q;
1600         list_add_tail(&bp->b_list, dwq);
1601         bp->b_queuetime = jiffies;
1602         spin_unlock(dwlk);
1603
1604         if (unlock)
1605                 xfs_buf_unlock(bp);
1606 }
1607
1608 void
1609 xfs_buf_delwri_dequeue(
1610         xfs_buf_t               *bp)
1611 {
1612         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1613         int                     dequeued = 0;
1614
1615         spin_lock(dwlk);
1616         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1617                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1618                 list_del_init(&bp->b_list);
1619                 dequeued = 1;
1620         }
1621         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1622         spin_unlock(dwlk);
1623
1624         if (dequeued)
1625                 xfs_buf_rele(bp);
1626
1627         trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1628 }
1629
1630 /*
1631  * If a delwri buffer needs to be pushed before it has aged out, then promote
1632  * it to the head of the delwri queue so that it will be flushed on the next
1633  * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1634  * than the age currently needed to flush the buffer. Hence the next time the
1635  * xfsbufd sees it is guaranteed to be considered old enough to flush.
1636  */
1637 void
1638 xfs_buf_delwri_promote(
1639         struct xfs_buf  *bp)
1640 {
1641         struct xfs_buftarg *btp = bp->b_target;
1642         long            age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1643
1644         ASSERT(bp->b_flags & XBF_DELWRI);
1645         ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1646
1647         /*
1648          * Check the buffer age before locking the delayed write queue as we
1649          * don't need to promote buffers that are already past the flush age.
1650          */
1651         if (bp->b_queuetime < jiffies - age)
1652                 return;
1653         bp->b_queuetime = jiffies - age;
1654         spin_lock(&btp->bt_delwrite_lock);
1655         list_move(&bp->b_list, &btp->bt_delwrite_queue);
1656         spin_unlock(&btp->bt_delwrite_lock);
1657 }
1658
1659 STATIC void
1660 xfs_buf_runall_queues(
1661         struct workqueue_struct *queue)
1662 {
1663         flush_workqueue(queue);
1664 }
1665
1666 /*
1667  * Move as many buffers as specified to the supplied list
1668  * idicating if we skipped any buffers to prevent deadlocks.
1669  */
1670 STATIC int
1671 xfs_buf_delwri_split(
1672         xfs_buftarg_t   *target,
1673         struct list_head *list,
1674         unsigned long   age)
1675 {
1676         xfs_buf_t       *bp, *n;
1677         struct list_head *dwq = &target->bt_delwrite_queue;
1678         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1679         int             skipped = 0;
1680         int             force;
1681
1682         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1683         INIT_LIST_HEAD(list);
1684         spin_lock(dwlk);
1685         list_for_each_entry_safe(bp, n, dwq, b_list) {
1686                 ASSERT(bp->b_flags & XBF_DELWRI);
1687
1688                 if (!XFS_BUF_ISPINNED(bp) && xfs_buf_trylock(bp)) {
1689                         if (!force &&
1690                             time_before(jiffies, bp->b_queuetime + age)) {
1691                                 xfs_buf_unlock(bp);
1692                                 break;
1693                         }
1694
1695                         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1696                                          _XBF_RUN_QUEUES);
1697                         bp->b_flags |= XBF_WRITE;
1698                         list_move_tail(&bp->b_list, list);
1699                         trace_xfs_buf_delwri_split(bp, _RET_IP_);
1700                 } else
1701                         skipped++;
1702         }
1703         spin_unlock(dwlk);
1704
1705         return skipped;
1706
1707 }
1708
1709 /*
1710  * Compare function is more complex than it needs to be because
1711  * the return value is only 32 bits and we are doing comparisons
1712  * on 64 bit values
1713  */
1714 static int
1715 xfs_buf_cmp(
1716         void            *priv,
1717         struct list_head *a,
1718         struct list_head *b)
1719 {
1720         struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1721         struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1722         xfs_daddr_t             diff;
1723
1724         diff = ap->b_bn - bp->b_bn;
1725         if (diff < 0)
1726                 return -1;
1727         if (diff > 0)
1728                 return 1;
1729         return 0;
1730 }
1731
1732 void
1733 xfs_buf_delwri_sort(
1734         xfs_buftarg_t   *target,
1735         struct list_head *list)
1736 {
1737         list_sort(NULL, list, xfs_buf_cmp);
1738 }
1739
1740 STATIC int
1741 xfsbufd(
1742         void            *data)
1743 {
1744         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1745
1746         current->flags |= PF_MEMALLOC;
1747
1748         set_freezable();
1749
1750         do {
1751                 long    age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1752                 long    tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1753                 struct list_head tmp;
1754                 struct blk_plug plug;
1755
1756                 if (unlikely(freezing(current))) {
1757                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1758                         refrigerator();
1759                 } else {
1760                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1761                 }
1762
1763                 /* sleep for a long time if there is nothing to do. */
1764                 if (list_empty(&target->bt_delwrite_queue))
1765                         tout = MAX_SCHEDULE_TIMEOUT;
1766                 schedule_timeout_interruptible(tout);
1767
1768                 xfs_buf_delwri_split(target, &tmp, age);
1769                 list_sort(NULL, &tmp, xfs_buf_cmp);
1770
1771                 blk_start_plug(&plug);
1772                 while (!list_empty(&tmp)) {
1773                         struct xfs_buf *bp;
1774                         bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1775                         list_del_init(&bp->b_list);
1776                         xfs_bdstrat_cb(bp);
1777                 }
1778                 blk_finish_plug(&plug);
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         struct blk_plug plug;
1799
1800         xfs_buf_runall_queues(xfsconvertd_workqueue);
1801         xfs_buf_runall_queues(xfsdatad_workqueue);
1802         xfs_buf_runall_queues(xfslogd_workqueue);
1803
1804         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1805         pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1806
1807         /*
1808          * Dropped the delayed write list lock, now walk the temporary list.
1809          * All I/O is issued async and then if we need to wait for completion
1810          * we do that after issuing all the IO.
1811          */
1812         list_sort(NULL, &tmp_list, xfs_buf_cmp);
1813
1814         blk_start_plug(&plug);
1815         while (!list_empty(&tmp_list)) {
1816                 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1817                 ASSERT(target == bp->b_target);
1818                 list_del_init(&bp->b_list);
1819                 if (wait) {
1820                         bp->b_flags &= ~XBF_ASYNC;
1821                         list_add(&bp->b_list, &wait_list);
1822                 }
1823                 xfs_bdstrat_cb(bp);
1824         }
1825         blk_finish_plug(&plug);
1826
1827         if (wait) {
1828                 /* Wait for IO to complete. */
1829                 while (!list_empty(&wait_list)) {
1830                         bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1831
1832                         list_del_init(&bp->b_list);
1833                         xfs_buf_iowait(bp);
1834                         xfs_buf_relse(bp);
1835                 }
1836         }
1837
1838         return pincount;
1839 }
1840
1841 int __init
1842 xfs_buf_init(void)
1843 {
1844         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1845                                                 KM_ZONE_HWALIGN, NULL);
1846         if (!xfs_buf_zone)
1847                 goto out;
1848
1849         xfslogd_workqueue = alloc_workqueue("xfslogd",
1850                                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1851         if (!xfslogd_workqueue)
1852                 goto out_free_buf_zone;
1853
1854         xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1855         if (!xfsdatad_workqueue)
1856                 goto out_destroy_xfslogd_workqueue;
1857
1858         xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1859                                                 WQ_MEM_RECLAIM, 1);
1860         if (!xfsconvertd_workqueue)
1861                 goto out_destroy_xfsdatad_workqueue;
1862
1863         return 0;
1864
1865  out_destroy_xfsdatad_workqueue:
1866         destroy_workqueue(xfsdatad_workqueue);
1867  out_destroy_xfslogd_workqueue:
1868         destroy_workqueue(xfslogd_workqueue);
1869  out_free_buf_zone:
1870         kmem_zone_destroy(xfs_buf_zone);
1871  out:
1872         return -ENOMEM;
1873 }
1874
1875 void
1876 xfs_buf_terminate(void)
1877 {
1878         destroy_workqueue(xfsconvertd_workqueue);
1879         destroy_workqueue(xfsdatad_workqueue);
1880         destroy_workqueue(xfslogd_workqueue);
1881         kmem_zone_destroy(xfs_buf_zone);
1882 }
1883
1884 #ifdef CONFIG_KDB_MODULES
1885 struct list_head *
1886 xfs_get_buftarg_list(void)
1887 {
1888         return &xfs_buftarg_list;
1889 }
1890 #endif