fs/xfs/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 | XBF_READ_AHEAD);
 596         bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
 597
 598         status = xfs_buf_iorequest(bp);
 599         if (status || bp->b_error || (flags & XBF_ASYNC))
 600                 return status;
 601         return xfs_buf_iowait(bp);
 602 }
 603
 604 xfs_buf_t *
 605 xfs_buf_read(
 606         xfs_buftarg_t           *target,
 607         xfs_off_t               ioff,
 608         size_t                  isize,
 609         xfs_buf_flags_t         flags)
 610 {
 611         xfs_buf_t               *bp;
 612
 613         flags |= XBF_READ;
 614
 615         bp = xfs_buf_get(target, ioff, isize, flags);
 616         if (bp) {
 617                 trace_xfs_buf_read(bp, flags, _RET_IP_);
 618
 619                 if (!XFS_BUF_ISDONE(bp)) {
 620                         XFS_STATS_INC(xb_get_read);
 621                         _xfs_buf_read(bp, flags);
 622                 } else if (flags & XBF_ASYNC) {
 623                         /*
 624                          * Read ahead call which is already satisfied,
 625                          * drop the buffer
 626                          */
 627                         goto no_buffer;
 628                 } else {
 629                         /* We do not want read in the flags */
 630                         bp->b_flags &= ~XBF_READ;
 631                 }
 632         }
 633
 634         return bp;
 635
 636  no_buffer:
 637         if (flags & (XBF_LOCK | XBF_TRYLOCK))
 638                 xfs_buf_unlock(bp);
 639         xfs_buf_rele(bp);
 640         return NULL;
 641 }
 642
 643 /*
 644  *      If we are not low on memory then do the readahead in a deadlock
 645  *      safe manner.
 646  */
 647 void
 648 xfs_buf_readahead(
 649         xfs_buftarg_t           *target,
 650         xfs_off_t               ioff,
 651         size_t                  isize)
 652 {
 653         if (bdi_read_congested(target->bt_bdi))
 654                 return;
 655
 656         xfs_buf_read(target, ioff, isize,
 657                      XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
 658 }
 659
 660 /*
 661  * Read an uncached buffer from disk. Allocates and returns a locked
 662  * buffer containing the disk contents or nothing.
 663  */
 664 struct xfs_buf *
 665 xfs_buf_read_uncached(
 666         struct xfs_mount        *mp,
 667         struct xfs_buftarg      *target,
 668         xfs_daddr_t             daddr,
 669         size_t                  length,
 670         int                     flags)
 671 {
 672         xfs_buf_t               *bp;
 673         int                     error;
 674
 675         bp = xfs_buf_get_uncached(target, length, flags);
 676         if (!bp)
 677                 return NULL;
 678
 679         /* set up the buffer for a read IO */
 680         XFS_BUF_SET_ADDR(bp, daddr);
 681         XFS_BUF_READ(bp);
 682
 683         xfsbdstrat(mp, bp);
 684         error = xfs_buf_iowait(bp);
 685         if (error || bp->b_error) {
 686                 xfs_buf_relse(bp);
 687                 return NULL;
 688         }
 689         return bp;
 690 }
 691
 692 xfs_buf_t *
 693 xfs_buf_get_empty(
 694         size_t                  len,
 695         xfs_buftarg_t           *target)
 696 {
 697         xfs_buf_t               *bp;
 698
 699         bp = xfs_buf_allocate(0);
 700         if (bp)
 701                 _xfs_buf_initialize(bp, target, 0, len, 0);
 702         return bp;
 703 }
 704
 705 /*
 706  * Return a buffer allocated as an empty buffer and associated to external
 707  * memory via xfs_buf_associate_memory() back to it's empty state.
 708  */
 709 void
 710 xfs_buf_set_empty(
 711         struct xfs_buf          *bp,
 712         size_t                  len)
 713 {
 714         if (bp->b_pages)
 715                 _xfs_buf_free_pages(bp);
 716
 717         bp->b_pages = NULL;
 718         bp->b_page_count = 0;
 719         bp->b_addr = NULL;
 720         bp->b_file_offset = 0;
 721         bp->b_buffer_length = bp->b_count_desired = len;
 722         bp->b_bn = XFS_BUF_DADDR_NULL;
 723         bp->b_flags &= ~XBF_MAPPED;
 724 }
 725
 726 static inline struct page *
 727 mem_to_page(
 728         void                    *addr)
 729 {
 730         if ((!is_vmalloc_addr(addr))) {
 731                 return virt_to_page(addr);
 732         } else {
 733                 return vmalloc_to_page(addr);
 734         }
 735 }
 736
 737 int
 738 xfs_buf_associate_memory(
 739         xfs_buf_t               *bp,
 740         void                    *mem,
 741         size_t                  len)
 742 {
 743         int                     rval;
 744         int                     i = 0;
 745         unsigned long           pageaddr;
 746         unsigned long           offset;
 747         size_t                  buflen;
 748         int                     page_count;
 749
 750         pageaddr = (unsigned long)mem & PAGE_MASK;
 751         offset = (unsigned long)mem - pageaddr;
 752         buflen = PAGE_ALIGN(len + offset);
 753         page_count = buflen >> PAGE_SHIFT;
 754
 755         /* Free any previous set of page pointers */
 756         if (bp->b_pages)
 757                 _xfs_buf_free_pages(bp);
 758
 759         bp->b_pages = NULL;
 760         bp->b_addr = mem;
 761
 762         rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
 763         if (rval)
 764                 return rval;
 765
 766         bp->b_offset = offset;
 767
 768         for (i = 0; i < bp->b_page_count; i++) {
 769                 bp->b_pages[i] = mem_to_page((void *)pageaddr);
 770                 pageaddr += PAGE_SIZE;
 771         }
 772
 773         bp->b_count_desired = len;
 774         bp->b_buffer_length = buflen;
 775         bp->b_flags |= XBF_MAPPED;
 776
 777         return 0;
 778 }
 779
 780 xfs_buf_t *
 781 xfs_buf_get_uncached(
 782         struct xfs_buftarg      *target,
 783         size_t                  len,
 784         int                     flags)
 785 {
 786         unsigned long           page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
 787         int                     error, i;
 788         xfs_buf_t               *bp;
 789
 790         bp = xfs_buf_allocate(0);
 791         if (unlikely(bp == NULL))
 792                 goto fail;
 793         _xfs_buf_initialize(bp, target, 0, len, 0);
 794
 795         error = _xfs_buf_get_pages(bp, page_count, 0);
 796         if (error)
 797                 goto fail_free_buf;
 798
 799         for (i = 0; i < page_count; i++) {
 800                 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
 801                 if (!bp->b_pages[i])
 802                         goto fail_free_mem;
 803         }
 804         bp->b_flags |= _XBF_PAGES;
 805
 806         error = _xfs_buf_map_pages(bp, XBF_MAPPED);
 807         if (unlikely(error)) {
 808                 xfs_warn(target->bt_mount,
 809                         "%s: failed to map pages\n", __func__);
 810                 goto fail_free_mem;
 811         }
 812
 813         trace_xfs_buf_get_uncached(bp, _RET_IP_);
 814         return bp;
 815
 816  fail_free_mem:
 817         while (--i >= 0)
 818                 __free_page(bp->b_pages[i]);
 819         _xfs_buf_free_pages(bp);
 820  fail_free_buf:
 821         xfs_buf_deallocate(bp);
 822  fail:
 823         return NULL;
 824 }
 825
 826 /*
 827  *      Increment reference count on buffer, to hold the buffer concurrently
 828  *      with another thread which may release (free) the buffer asynchronously.
 829  *      Must hold the buffer already to call this function.
 830  */
 831 void
 832 xfs_buf_hold(
 833         xfs_buf_t               *bp)
 834 {
 835         trace_xfs_buf_hold(bp, _RET_IP_);
 836         atomic_inc(&bp->b_hold);
 837 }
 838
 839 /*
 840  *      Releases a hold on the specified buffer.  If the
 841  *      the hold count is 1, calls xfs_buf_free.
 842  */
 843 void
 844 xfs_buf_rele(
 845         xfs_buf_t               *bp)
 846 {
 847         struct xfs_perag        *pag = bp->b_pag;
 848
 849         trace_xfs_buf_rele(bp, _RET_IP_);
 850
 851         if (!pag) {
 852                 ASSERT(list_empty(&bp->b_lru));
 853                 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
 854                 if (atomic_dec_and_test(&bp->b_hold))
 855                         xfs_buf_free(bp);
 856                 return;
 857         }
 858
 859         ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
 860
 861         ASSERT(atomic_read(&bp->b_hold) > 0);
 862         if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
 863                 if (!(bp->b_flags & XBF_STALE) &&
 864                            atomic_read(&bp->b_lru_ref)) {
 865                         xfs_buf_lru_add(bp);
 866                         spin_unlock(&pag->pag_buf_lock);
 867                 } else {
 868                         xfs_buf_lru_del(bp);
 869                         ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
 870                         rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
 871                         spin_unlock(&pag->pag_buf_lock);
 872                         xfs_perag_put(pag);
 873                         xfs_buf_free(bp);
 874                 }
 875         }
 876 }
 877
 878
 879 /*
 880  *      Lock a buffer object, if it is not already locked.
 881  *
 882  *      If we come across a stale, pinned, locked buffer, we know that we are
 883  *      being asked to lock a buffer that has been reallocated. Because it is
 884  *      pinned, we know that the log has not been pushed to disk and hence it
 885  *      will still be locked.  Rather than continuing to have trylock attempts
 886  *      fail until someone else pushes the log, push it ourselves before
 887  *      returning.  This means that the xfsaild will not get stuck trying
 888  *      to push on stale inode buffers.
 889  */
 890 int
 891 xfs_buf_trylock(
 892         struct xfs_buf          *bp)
 893 {
 894         int                     locked;
 895
 896         locked = down_trylock(&bp->b_sema) == 0;
 897         if (locked)
 898                 XB_SET_OWNER(bp);
 899         else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 900                 xfs_log_force(bp->b_target->bt_mount, 0);
 901
 902         trace_xfs_buf_trylock(bp, _RET_IP_);
 903         return locked;
 904 }
 905
 906 /*
 907  *      Lock a buffer object.
 908  *
 909  *      If we come across a stale, pinned, locked buffer, we know that we
 910  *      are being asked to lock a buffer that has been reallocated. Because
 911  *      it is pinned, we know that the log has not been pushed to disk and
 912  *      hence it will still be locked. Rather than sleeping until someone
 913  *      else pushes the log, push it ourselves before trying to get the lock.
 914  */
 915 void
 916 xfs_buf_lock(
 917         struct xfs_buf          *bp)
 918 {
 919         trace_xfs_buf_lock(bp, _RET_IP_);
 920
 921         if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
 922                 xfs_log_force(bp->b_target->bt_mount, 0);
 923         down(&bp->b_sema);
 924         XB_SET_OWNER(bp);
 925
 926         trace_xfs_buf_lock_done(bp, _RET_IP_);
 927 }
 928
 929 /*
 930  *      Releases the lock on the buffer object.
 931  *      If the buffer is marked delwri but is not queued, do so before we
 932  *      unlock the buffer as we need to set flags correctly.  We also need to
 933  *      take a reference for the delwri queue because the unlocker is going to
 934  *      drop their's and they don't know we just queued it.
 935  */
 936 void
 937 xfs_buf_unlock(
 938         struct xfs_buf          *bp)
 939 {
 940         if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
 941                 atomic_inc(&bp->b_hold);
 942                 bp->b_flags |= XBF_ASYNC;
 943                 xfs_buf_delwri_queue(bp, 0);
 944         }
 945
 946         XB_CLEAR_OWNER(bp);
 947         up(&bp->b_sema);
 948
 949         trace_xfs_buf_unlock(bp, _RET_IP_);
 950 }
 951
 952 STATIC void
 953 xfs_buf_wait_unpin(
 954         xfs_buf_t               *bp)
 955 {
 956         DECLARE_WAITQUEUE       (wait, current);
 957
 958         if (atomic_read(&bp->b_pin_count) == 0)
 959                 return;
 960
 961         add_wait_queue(&bp->b_waiters, &wait);
 962         for (;;) {
 963                 set_current_state(TASK_UNINTERRUPTIBLE);
 964                 if (atomic_read(&bp->b_pin_count) == 0)
 965                         break;
 966                 io_schedule();
 967         }
 968         remove_wait_queue(&bp->b_waiters, &wait);
 969         set_current_state(TASK_RUNNING);
 970 }
 971
 972 /*
 973  *      Buffer Utility Routines
 974  */
 975
 976 STATIC void
 977 xfs_buf_iodone_work(
 978         struct work_struct      *work)
 979 {
 980         xfs_buf_t               *bp =
 981                 container_of(work, xfs_buf_t, b_iodone_work);
 982
 983         if (bp->b_iodone)
 984                 (*(bp->b_iodone))(bp);
 985         else if (bp->b_flags & XBF_ASYNC)
 986                 xfs_buf_relse(bp);
 987 }
 988
 989 void
 990 xfs_buf_ioend(
 991         xfs_buf_t               *bp,
 992         int                     schedule)
 993 {
 994         trace_xfs_buf_iodone(bp, _RET_IP_);
 995
 996         bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
 997         if (bp->b_error == 0)
 998                 bp->b_flags |= XBF_DONE;
 999
1000         if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1001                 if (schedule) {
1002                         INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1003                         queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1004                 } else {
1005                         xfs_buf_iodone_work(&bp->b_iodone_work);
1006                 }
1007         } else {
1008                 complete(&bp->b_iowait);
1009         }
1010 }
1011
1012 void
1013 xfs_buf_ioerror(
1014         xfs_buf_t               *bp,
1015         int                     error)
1016 {
1017         ASSERT(error >= 0 && error <= 0xffff);
1018         bp->b_error = (unsigned short)error;
1019         trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1020 }
1021
1022 int
1023 xfs_bwrite(
1024         struct xfs_mount        *mp,
1025         struct xfs_buf          *bp)
1026 {
1027         int                     error;
1028
1029         bp->b_flags |= XBF_WRITE;
1030         bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1031
1032         xfs_buf_delwri_dequeue(bp);
1033         xfs_bdstrat_cb(bp);
1034
1035         error = xfs_buf_iowait(bp);
1036         if (error)
1037                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1038         xfs_buf_relse(bp);
1039         return error;
1040 }
1041
1042 void
1043 xfs_bdwrite(
1044         void                    *mp,
1045         struct xfs_buf          *bp)
1046 {
1047         trace_xfs_buf_bdwrite(bp, _RET_IP_);
1048
1049         bp->b_flags &= ~XBF_READ;
1050         bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1051
1052         xfs_buf_delwri_queue(bp, 1);
1053 }
1054
1055 /*
1056  * Called when we want to stop a buffer from getting written or read.
1057  * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1058  * so that the proper iodone callbacks get called.
1059  */
1060 STATIC int
1061 xfs_bioerror(
1062         xfs_buf_t *bp)
1063 {
1064 #ifdef XFSERRORDEBUG
1065         ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1066 #endif
1067
1068         /*
1069          * No need to wait until the buffer is unpinned, we aren't flushing it.
1070          */
1071         xfs_buf_ioerror(bp, EIO);
1072
1073         /*
1074          * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1075          */
1076         XFS_BUF_UNREAD(bp);
1077         XFS_BUF_UNDELAYWRITE(bp);
1078         XFS_BUF_UNDONE(bp);
1079         XFS_BUF_STALE(bp);
1080
1081         xfs_buf_ioend(bp, 0);
1082
1083         return EIO;
1084 }
1085
1086 /*
1087  * Same as xfs_bioerror, except that we are releasing the buffer
1088  * here ourselves, and avoiding the xfs_buf_ioend call.
1089  * This is meant for userdata errors; metadata bufs come with
1090  * iodone functions attached, so that we can track down errors.
1091  */
1092 STATIC int
1093 xfs_bioerror_relse(
1094         struct xfs_buf  *bp)
1095 {
1096         int64_t         fl = bp->b_flags;
1097         /*
1098          * No need to wait until the buffer is unpinned.
1099          * We aren't flushing it.
1100          *
1101          * chunkhold expects B_DONE to be set, whether
1102          * we actually finish the I/O or not. We don't want to
1103          * change that interface.
1104          */
1105         XFS_BUF_UNREAD(bp);
1106         XFS_BUF_UNDELAYWRITE(bp);
1107         XFS_BUF_DONE(bp);
1108         XFS_BUF_STALE(bp);
1109         bp->b_iodone = NULL;
1110         if (!(fl & XBF_ASYNC)) {
1111                 /*
1112                  * Mark b_error and B_ERROR _both_.
1113                  * Lot's of chunkcache code assumes that.
1114                  * There's no reason to mark error for
1115                  * ASYNC buffers.
1116                  */
1117                 xfs_buf_ioerror(bp, EIO);
1118                 XFS_BUF_FINISH_IOWAIT(bp);
1119         } else {
1120                 xfs_buf_relse(bp);
1121         }
1122
1123         return EIO;
1124 }
1125
1126
1127 /*
1128  * All xfs metadata buffers except log state machine buffers
1129  * get this attached as their b_bdstrat callback function.
1130  * This is so that we can catch a buffer
1131  * after prematurely unpinning it to forcibly shutdown the filesystem.
1132  */
1133 int
1134 xfs_bdstrat_cb(
1135         struct xfs_buf  *bp)
1136 {
1137         if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1138                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1139                 /*
1140                  * Metadata write that didn't get logged but
1141                  * written delayed anyway. These aren't associated
1142                  * with a transaction, and can be ignored.
1143                  */
1144                 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1145                         return xfs_bioerror_relse(bp);
1146                 else
1147                         return xfs_bioerror(bp);
1148         }
1149
1150         xfs_buf_iorequest(bp);
1151         return 0;
1152 }
1153
1154 /*
1155  * Wrapper around bdstrat so that we can stop data from going to disk in case
1156  * we are shutting down the filesystem.  Typically user data goes thru this
1157  * path; one of the exceptions is the superblock.
1158  */
1159 void
1160 xfsbdstrat(
1161         struct xfs_mount        *mp,
1162         struct xfs_buf          *bp)
1163 {
1164         if (XFS_FORCED_SHUTDOWN(mp)) {
1165                 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1166                 xfs_bioerror_relse(bp);
1167                 return;
1168         }
1169
1170         xfs_buf_iorequest(bp);
1171 }
1172
1173 STATIC void
1174 _xfs_buf_ioend(
1175         xfs_buf_t               *bp,
1176         int                     schedule)
1177 {
1178         if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1179                 xfs_buf_ioend(bp, schedule);
1180 }
1181
1182 STATIC void
1183 xfs_buf_bio_end_io(
1184         struct bio              *bio,
1185         int                     error)
1186 {
1187         xfs_buf_t               *bp = (xfs_buf_t *)bio->bi_private;
1188
1189         xfs_buf_ioerror(bp, -error);
1190
1191         if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1192                 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1193
1194         _xfs_buf_ioend(bp, 1);
1195         bio_put(bio);
1196 }
1197
1198 STATIC void
1199 _xfs_buf_ioapply(
1200         xfs_buf_t               *bp)
1201 {
1202         int                     rw, map_i, total_nr_pages, nr_pages;
1203         struct bio              *bio;
1204         int                     offset = bp->b_offset;
1205         int                     size = bp->b_count_desired;
1206         sector_t                sector = bp->b_bn;
1207
1208         total_nr_pages = bp->b_page_count;
1209         map_i = 0;
1210
1211         if (bp->b_flags & XBF_WRITE) {
1212                 if (bp->b_flags & XBF_SYNCIO)
1213                         rw = WRITE_SYNC;
1214                 else
1215                         rw = WRITE;
1216                 if (bp->b_flags & XBF_FUA)
1217                         rw |= REQ_FUA;
1218                 if (bp->b_flags & XBF_FLUSH)
1219                         rw |= REQ_FLUSH;
1220         } else if (bp->b_flags & XBF_READ_AHEAD) {
1221                 rw = READA;
1222         } else {
1223                 rw = READ;
1224         }
1225
1226         /* we only use the buffer cache for meta-data */
1227         rw |= REQ_META;
1228
1229 next_chunk:
1230         atomic_inc(&bp->b_io_remaining);
1231         nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1232         if (nr_pages > total_nr_pages)
1233                 nr_pages = total_nr_pages;
1234
1235         bio = bio_alloc(GFP_NOIO, nr_pages);
1236         bio->bi_bdev = bp->b_target->bt_bdev;
1237         bio->bi_sector = sector;
1238         bio->bi_end_io = xfs_buf_bio_end_io;
1239         bio->bi_private = bp;
1240
1241
1242         for (; size && nr_pages; nr_pages--, map_i++) {
1243                 int     rbytes, nbytes = PAGE_SIZE - offset;
1244
1245                 if (nbytes > size)
1246                         nbytes = size;
1247
1248                 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1249                 if (rbytes < nbytes)
1250                         break;
1251
1252                 offset = 0;
1253                 sector += nbytes >> BBSHIFT;
1254                 size -= nbytes;
1255                 total_nr_pages--;
1256         }
1257
1258         if (likely(bio->bi_size)) {
1259                 if (xfs_buf_is_vmapped(bp)) {
1260                         flush_kernel_vmap_range(bp->b_addr,
1261                                                 xfs_buf_vmap_len(bp));
1262                 }
1263                 submit_bio(rw, bio);
1264                 if (size)
1265                         goto next_chunk;
1266         } else {
1267                 xfs_buf_ioerror(bp, EIO);
1268                 bio_put(bio);
1269         }
1270 }
1271
1272 int
1273 xfs_buf_iorequest(
1274         xfs_buf_t               *bp)
1275 {
1276         trace_xfs_buf_iorequest(bp, _RET_IP_);
1277
1278         if (bp->b_flags & XBF_DELWRI) {
1279                 xfs_buf_delwri_queue(bp, 1);
1280                 return 0;
1281         }
1282
1283         if (bp->b_flags & XBF_WRITE) {
1284                 xfs_buf_wait_unpin(bp);
1285         }
1286
1287         xfs_buf_hold(bp);
1288
1289         /* Set the count to 1 initially, this will stop an I/O
1290          * completion callout which happens before we have started
1291          * all the I/O from calling xfs_buf_ioend too early.
1292          */
1293         atomic_set(&bp->b_io_remaining, 1);
1294         _xfs_buf_ioapply(bp);
1295         _xfs_buf_ioend(bp, 0);
1296
1297         xfs_buf_rele(bp);
1298         return 0;
1299 }
1300
1301 /*
1302  *      Waits for I/O to complete on the buffer supplied.
1303  *      It returns immediately if no I/O is pending.
1304  *      It returns the I/O error code, if any, or 0 if there was no error.
1305  */
1306 int
1307 xfs_buf_iowait(
1308         xfs_buf_t               *bp)
1309 {
1310         trace_xfs_buf_iowait(bp, _RET_IP_);
1311
1312         wait_for_completion(&bp->b_iowait);
1313
1314         trace_xfs_buf_iowait_done(bp, _RET_IP_);
1315         return bp->b_error;
1316 }
1317
1318 xfs_caddr_t
1319 xfs_buf_offset(
1320         xfs_buf_t               *bp,
1321         size_t                  offset)
1322 {
1323         struct page             *page;
1324
1325         if (bp->b_flags & XBF_MAPPED)
1326                 return bp->b_addr + offset;
1327
1328         offset += bp->b_offset;
1329         page = bp->b_pages[offset >> PAGE_SHIFT];
1330         return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1331 }
1332
1333 /*
1334  *      Move data into or out of a buffer.
1335  */
1336 void
1337 xfs_buf_iomove(
1338         xfs_buf_t               *bp,    /* buffer to process            */
1339         size_t                  boff,   /* starting buffer offset       */
1340         size_t                  bsize,  /* length to copy               */
1341         void                    *data,  /* data address                 */
1342         xfs_buf_rw_t            mode)   /* read/write/zero flag         */
1343 {
1344         size_t                  bend, cpoff, csize;
1345         struct page             *page;
1346
1347         bend = boff + bsize;
1348         while (boff < bend) {
1349                 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1350                 cpoff = xfs_buf_poff(boff + bp->b_offset);
1351                 csize = min_t(size_t,
1352                               PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1353
1354                 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1355
1356                 switch (mode) {
1357                 case XBRW_ZERO:
1358                         memset(page_address(page) + cpoff, 0, csize);
1359                         break;
1360                 case XBRW_READ:
1361                         memcpy(data, page_address(page) + cpoff, csize);
1362                         break;
1363                 case XBRW_WRITE:
1364                         memcpy(page_address(page) + cpoff, data, csize);
1365                 }
1366
1367                 boff += csize;
1368                 data += csize;
1369         }
1370 }
1371
1372 /*
1373  *      Handling of buffer targets (buftargs).
1374  */
1375
1376 /*
1377  * Wait for any bufs with callbacks that have been submitted but have not yet
1378  * returned. These buffers will have an elevated hold count, so wait on those
1379  * while freeing all the buffers only held by the LRU.
1380  */
1381 void
1382 xfs_wait_buftarg(
1383         struct xfs_buftarg      *btp)
1384 {
1385         struct xfs_buf          *bp;
1386
1387 restart:
1388         spin_lock(&btp->bt_lru_lock);
1389         while (!list_empty(&btp->bt_lru)) {
1390                 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1391                 if (atomic_read(&bp->b_hold) > 1) {
1392                         spin_unlock(&btp->bt_lru_lock);
1393                         delay(100);
1394                         goto restart;
1395                 }
1396                 /*
1397                  * clear the LRU reference count so the bufer doesn't get
1398                  * ignored in xfs_buf_rele().
1399                  */
1400                 atomic_set(&bp->b_lru_ref, 0);
1401                 spin_unlock(&btp->bt_lru_lock);
1402                 xfs_buf_rele(bp);
1403                 spin_lock(&btp->bt_lru_lock);
1404         }
1405         spin_unlock(&btp->bt_lru_lock);
1406 }
1407
1408 int
1409 xfs_buftarg_shrink(
1410         struct shrinker         *shrink,
1411         struct shrink_control   *sc)
1412 {
1413         struct xfs_buftarg      *btp = container_of(shrink,
1414                                         struct xfs_buftarg, bt_shrinker);
1415         struct xfs_buf          *bp;
1416         int nr_to_scan = sc->nr_to_scan;
1417         LIST_HEAD(dispose);
1418
1419         if (!nr_to_scan)
1420                 return btp->bt_lru_nr;
1421
1422         spin_lock(&btp->bt_lru_lock);
1423         while (!list_empty(&btp->bt_lru)) {
1424                 if (nr_to_scan-- <= 0)
1425                         break;
1426
1427                 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1428
1429                 /*
1430                  * Decrement the b_lru_ref count unless the value is already
1431                  * zero. If the value is already zero, we need to reclaim the
1432                  * buffer, otherwise it gets another trip through the LRU.
1433                  */
1434                 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1435                         list_move_tail(&bp->b_lru, &btp->bt_lru);
1436                         continue;
1437                 }
1438
1439                 /*
1440                  * remove the buffer from the LRU now to avoid needing another
1441                  * lock round trip inside xfs_buf_rele().
1442                  */
1443                 list_move(&bp->b_lru, &dispose);
1444                 btp->bt_lru_nr--;
1445         }
1446         spin_unlock(&btp->bt_lru_lock);
1447
1448         while (!list_empty(&dispose)) {
1449                 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1450                 list_del_init(&bp->b_lru);
1451                 xfs_buf_rele(bp);
1452         }
1453
1454         return btp->bt_lru_nr;
1455 }
1456
1457 void
1458 xfs_free_buftarg(
1459         struct xfs_mount        *mp,
1460         struct xfs_buftarg      *btp)
1461 {
1462         unregister_shrinker(&btp->bt_shrinker);
1463
1464         xfs_flush_buftarg(btp, 1);
1465         if (mp->m_flags & XFS_MOUNT_BARRIER)
1466                 xfs_blkdev_issue_flush(btp);
1467
1468         kthread_stop(btp->bt_task);
1469         kmem_free(btp);
1470 }
1471
1472 STATIC int
1473 xfs_setsize_buftarg_flags(
1474         xfs_buftarg_t           *btp,
1475         unsigned int            blocksize,
1476         unsigned int            sectorsize,
1477         int                     verbose)
1478 {
1479         btp->bt_bsize = blocksize;
1480         btp->bt_sshift = ffs(sectorsize) - 1;
1481         btp->bt_smask = sectorsize - 1;
1482
1483         if (set_blocksize(btp->bt_bdev, sectorsize)) {
1484                 xfs_warn(btp->bt_mount,
1485                         "Cannot set_blocksize to %u on device %s\n",
1486                         sectorsize, xfs_buf_target_name(btp));
1487                 return EINVAL;
1488         }
1489
1490         return 0;
1491 }
1492
1493 /*
1494  *      When allocating the initial buffer target we have not yet
1495  *      read in the superblock, so don't know what sized sectors
1496  *      are being used is at this early stage.  Play safe.
1497  */
1498 STATIC int
1499 xfs_setsize_buftarg_early(
1500         xfs_buftarg_t           *btp,
1501         struct block_device     *bdev)
1502 {
1503         return xfs_setsize_buftarg_flags(btp,
1504                         PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1505 }
1506
1507 int
1508 xfs_setsize_buftarg(
1509         xfs_buftarg_t           *btp,
1510         unsigned int            blocksize,
1511         unsigned int            sectorsize)
1512 {
1513         return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1514 }
1515
1516 STATIC int
1517 xfs_alloc_delwrite_queue(
1518         xfs_buftarg_t           *btp,
1519         const char              *fsname)
1520 {
1521         INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1522         spin_lock_init(&btp->bt_delwrite_lock);
1523         btp->bt_flags = 0;
1524         btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1525         if (IS_ERR(btp->bt_task))
1526                 return PTR_ERR(btp->bt_task);
1527         return 0;
1528 }
1529
1530 xfs_buftarg_t *
1531 xfs_alloc_buftarg(
1532         struct xfs_mount        *mp,
1533         struct block_device     *bdev,
1534         int                     external,
1535         const char              *fsname)
1536 {
1537         xfs_buftarg_t           *btp;
1538
1539         btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1540
1541         btp->bt_mount = mp;
1542         btp->bt_dev =  bdev->bd_dev;
1543         btp->bt_bdev = bdev;
1544         btp->bt_bdi = blk_get_backing_dev_info(bdev);
1545         if (!btp->bt_bdi)
1546                 goto error;
1547
1548         INIT_LIST_HEAD(&btp->bt_lru);
1549         spin_lock_init(&btp->bt_lru_lock);
1550         if (xfs_setsize_buftarg_early(btp, bdev))
1551                 goto error;
1552         if (xfs_alloc_delwrite_queue(btp, fsname))
1553                 goto error;
1554         btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1555         btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1556         register_shrinker(&btp->bt_shrinker);
1557         return btp;
1558
1559 error:
1560         kmem_free(btp);
1561         return NULL;
1562 }
1563
1564
1565 /*
1566  *      Delayed write buffer handling
1567  */
1568 STATIC void
1569 xfs_buf_delwri_queue(
1570         xfs_buf_t               *bp,
1571         int                     unlock)
1572 {
1573         struct list_head        *dwq = &bp->b_target->bt_delwrite_queue;
1574         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1575
1576         trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1577
1578         ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1579
1580         spin_lock(dwlk);
1581         /* If already in the queue, dequeue and place at tail */
1582         if (!list_empty(&bp->b_list)) {
1583                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1584                 if (unlock)
1585                         atomic_dec(&bp->b_hold);
1586                 list_del(&bp->b_list);
1587         }
1588
1589         if (list_empty(dwq)) {
1590                 /* start xfsbufd as it is about to have something to do */
1591                 wake_up_process(bp->b_target->bt_task);
1592         }
1593
1594         bp->b_flags |= _XBF_DELWRI_Q;
1595         list_add_tail(&bp->b_list, dwq);
1596         bp->b_queuetime = jiffies;
1597         spin_unlock(dwlk);
1598
1599         if (unlock)
1600                 xfs_buf_unlock(bp);
1601 }
1602
1603 void
1604 xfs_buf_delwri_dequeue(
1605         xfs_buf_t               *bp)
1606 {
1607         spinlock_t              *dwlk = &bp->b_target->bt_delwrite_lock;
1608         int                     dequeued = 0;
1609
1610         spin_lock(dwlk);
1611         if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1612                 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1613                 list_del_init(&bp->b_list);
1614                 dequeued = 1;
1615         }
1616         bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1617         spin_unlock(dwlk);
1618
1619         if (dequeued)
1620                 xfs_buf_rele(bp);
1621
1622         trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1623 }
1624
1625 /*
1626  * If a delwri buffer needs to be pushed before it has aged out, then promote
1627  * it to the head of the delwri queue so that it will be flushed on the next
1628  * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1629  * than the age currently needed to flush the buffer. Hence the next time the
1630  * xfsbufd sees it is guaranteed to be considered old enough to flush.
1631  */
1632 void
1633 xfs_buf_delwri_promote(
1634         struct xfs_buf  *bp)
1635 {
1636         struct xfs_buftarg *btp = bp->b_target;
1637         long            age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1638
1639         ASSERT(bp->b_flags & XBF_DELWRI);
1640         ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1641
1642         /*
1643          * Check the buffer age before locking the delayed write queue as we
1644          * don't need to promote buffers that are already past the flush age.
1645          */
1646         if (bp->b_queuetime < jiffies - age)
1647                 return;
1648         bp->b_queuetime = jiffies - age;
1649         spin_lock(&btp->bt_delwrite_lock);
1650         list_move(&bp->b_list, &btp->bt_delwrite_queue);
1651         spin_unlock(&btp->bt_delwrite_lock);
1652 }
1653
1654 STATIC void
1655 xfs_buf_runall_queues(
1656         struct workqueue_struct *queue)
1657 {
1658         flush_workqueue(queue);
1659 }
1660
1661 /*
1662  * Move as many buffers as specified to the supplied list
1663  * idicating if we skipped any buffers to prevent deadlocks.
1664  */
1665 STATIC int
1666 xfs_buf_delwri_split(
1667         xfs_buftarg_t   *target,
1668         struct list_head *list,
1669         unsigned long   age)
1670 {
1671         xfs_buf_t       *bp, *n;
1672         struct list_head *dwq = &target->bt_delwrite_queue;
1673         spinlock_t      *dwlk = &target->bt_delwrite_lock;
1674         int             skipped = 0;
1675         int             force;
1676
1677         force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1678         INIT_LIST_HEAD(list);
1679         spin_lock(dwlk);
1680         list_for_each_entry_safe(bp, n, dwq, b_list) {
1681                 ASSERT(bp->b_flags & XBF_DELWRI);
1682
1683                 if (!xfs_buf_ispinned(bp) && xfs_buf_trylock(bp)) {
1684                         if (!force &&
1685                             time_before(jiffies, bp->b_queuetime + age)) {
1686                                 xfs_buf_unlock(bp);
1687                                 break;
1688                         }
1689
1690                         bp->b_flags &= ~(XBF_DELWRI | _XBF_DELWRI_Q);
1691                         bp->b_flags |= XBF_WRITE;
1692                         list_move_tail(&bp->b_list, list);
1693                         trace_xfs_buf_delwri_split(bp, _RET_IP_);
1694                 } else
1695                         skipped++;
1696         }
1697         spin_unlock(dwlk);
1698
1699         return skipped;
1700
1701 }
1702
1703 /*
1704  * Compare function is more complex than it needs to be because
1705  * the return value is only 32 bits and we are doing comparisons
1706  * on 64 bit values
1707  */
1708 static int
1709 xfs_buf_cmp(
1710         void            *priv,
1711         struct list_head *a,
1712         struct list_head *b)
1713 {
1714         struct xfs_buf  *ap = container_of(a, struct xfs_buf, b_list);
1715         struct xfs_buf  *bp = container_of(b, struct xfs_buf, b_list);
1716         xfs_daddr_t             diff;
1717
1718         diff = ap->b_bn - bp->b_bn;
1719         if (diff < 0)
1720                 return -1;
1721         if (diff > 0)
1722                 return 1;
1723         return 0;
1724 }
1725
1726 STATIC int
1727 xfsbufd(
1728         void            *data)
1729 {
1730         xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
1731
1732         current->flags |= PF_MEMALLOC;
1733
1734         set_freezable();
1735
1736         do {
1737                 long    age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1738                 long    tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1739                 struct list_head tmp;
1740                 struct blk_plug plug;
1741
1742                 if (unlikely(freezing(current))) {
1743                         set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1744                         refrigerator();
1745                 } else {
1746                         clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1747                 }
1748
1749                 /* sleep for a long time if there is nothing to do. */
1750                 if (list_empty(&target->bt_delwrite_queue))
1751                         tout = MAX_SCHEDULE_TIMEOUT;
1752                 schedule_timeout_interruptible(tout);
1753
1754                 xfs_buf_delwri_split(target, &tmp, age);
1755                 list_sort(NULL, &tmp, xfs_buf_cmp);
1756
1757                 blk_start_plug(&plug);
1758                 while (!list_empty(&tmp)) {
1759                         struct xfs_buf *bp;
1760                         bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1761                         list_del_init(&bp->b_list);
1762                         xfs_bdstrat_cb(bp);
1763                 }
1764                 blk_finish_plug(&plug);
1765         } while (!kthread_should_stop());
1766
1767         return 0;
1768 }
1769
1770 /*
1771  *      Go through all incore buffers, and release buffers if they belong to
1772  *      the given device. This is used in filesystem error handling to
1773  *      preserve the consistency of its metadata.
1774  */
1775 int
1776 xfs_flush_buftarg(
1777         xfs_buftarg_t   *target,
1778         int             wait)
1779 {
1780         xfs_buf_t       *bp;
1781         int             pincount = 0;
1782         LIST_HEAD(tmp_list);
1783         LIST_HEAD(wait_list);
1784         struct blk_plug plug;
1785
1786         xfs_buf_runall_queues(xfsconvertd_workqueue);
1787         xfs_buf_runall_queues(xfsdatad_workqueue);
1788         xfs_buf_runall_queues(xfslogd_workqueue);
1789
1790         set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1791         pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1792
1793         /*
1794          * Dropped the delayed write list lock, now walk the temporary list.
1795          * All I/O is issued async and then if we need to wait for completion
1796          * we do that after issuing all the IO.
1797          */
1798         list_sort(NULL, &tmp_list, xfs_buf_cmp);
1799
1800         blk_start_plug(&plug);
1801         while (!list_empty(&tmp_list)) {
1802                 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1803                 ASSERT(target == bp->b_target);
1804                 list_del_init(&bp->b_list);
1805                 if (wait) {
1806                         bp->b_flags &= ~XBF_ASYNC;
1807                         list_add(&bp->b_list, &wait_list);
1808                 }
1809                 xfs_bdstrat_cb(bp);
1810         }
1811         blk_finish_plug(&plug);
1812
1813         if (wait) {
1814                 /* Wait for IO to complete. */
1815                 while (!list_empty(&wait_list)) {
1816                         bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1817
1818                         list_del_init(&bp->b_list);
1819                         xfs_buf_iowait(bp);
1820                         xfs_buf_relse(bp);
1821                 }
1822         }
1823
1824         return pincount;
1825 }
1826
1827 int __init
1828 xfs_buf_init(void)
1829 {
1830         xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1831                                                 KM_ZONE_HWALIGN, NULL);
1832         if (!xfs_buf_zone)
1833                 goto out;
1834
1835         xfslogd_workqueue = alloc_workqueue("xfslogd",
1836                                         WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1837         if (!xfslogd_workqueue)
1838                 goto out_free_buf_zone;
1839
1840         xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1841         if (!xfsdatad_workqueue)
1842                 goto out_destroy_xfslogd_workqueue;
1843
1844         xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1845                                                 WQ_MEM_RECLAIM, 1);
1846         if (!xfsconvertd_workqueue)
1847                 goto out_destroy_xfsdatad_workqueue;
1848
1849         return 0;
1850
1851  out_destroy_xfsdatad_workqueue:
1852         destroy_workqueue(xfsdatad_workqueue);
1853  out_destroy_xfslogd_workqueue:
1854         destroy_workqueue(xfslogd_workqueue);
1855  out_free_buf_zone:
1856         kmem_zone_destroy(xfs_buf_zone);
1857  out:
1858         return -ENOMEM;
1859 }
1860
1861 void
1862 xfs_buf_terminate(void)
1863 {
1864         destroy_workqueue(xfsconvertd_workqueue);
1865         destroy_workqueue(xfsdatad_workqueue);
1866         destroy_workqueue(xfslogd_workqueue);
1867         kmem_zone_destroy(xfs_buf_zone);
1868 }
1869
1870 #ifdef CONFIG_KDB_MODULES
1871 struct list_head *
1872 xfs_get_buftarg_list(void)
1873 {
1874         return &xfs_buftarg_list;
1875 }
1876 #endif