2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
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.
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.
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
19 #include <linux/stddef.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/vmalloc.h>
25 #include <linux/bio.h>
26 #include <linux/sysctl.h>
27 #include <linux/proc_fs.h>
28 #include <linux/workqueue.h>
29 #include <linux/percpu.h>
30 #include <linux/blkdev.h>
31 #include <linux/hash.h>
32 #include <linux/kthread.h>
33 #include <linux/migrate.h>
34 #include <linux/backing-dev.h>
35 #include <linux/freezer.h>
40 #include "xfs_dmapi.h"
41 #include "xfs_mount.h"
42 #include "xfs_trace.h"
44 static kmem_zone_t *xfs_buf_zone;
45 STATIC int xfsbufd(void *);
46 STATIC int xfsbufd_wakeup(int, gfp_t);
47 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
48 static struct shrinker xfs_buf_shake = {
49 .shrink = xfsbufd_wakeup,
50 .seeks = DEFAULT_SEEKS,
53 static struct workqueue_struct *xfslogd_workqueue;
54 struct workqueue_struct *xfsdatad_workqueue;
55 struct workqueue_struct *xfsconvertd_workqueue;
57 #ifdef XFS_BUF_LOCK_TRACKING
58 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
59 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
60 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
62 # define XB_SET_OWNER(bp) do { } while (0)
63 # define XB_CLEAR_OWNER(bp) do { } while (0)
64 # define XB_GET_OWNER(bp) do { } while (0)
67 #define xb_to_gfp(flags) \
68 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
69 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
71 #define xb_to_km(flags) \
72 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
74 #define xfs_buf_allocate(flags) \
75 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
76 #define xfs_buf_deallocate(bp) \
77 kmem_zone_free(xfs_buf_zone, (bp));
80 * Page Region interfaces.
82 * For pages in filesystems where the blocksize is smaller than the
83 * pagesize, we use the page->private field (long) to hold a bitmap
84 * of uptodate regions within the page.
86 * Each such region is "bytes per page / bits per long" bytes long.
88 * NBPPR == number-of-bytes-per-page-region
89 * BTOPR == bytes-to-page-region (rounded up)
90 * BTOPRT == bytes-to-page-region-truncated (rounded down)
92 #if (BITS_PER_LONG == 32)
93 #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */
94 #elif (BITS_PER_LONG == 64)
95 #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */
97 #error BITS_PER_LONG must be 32 or 64
99 #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG)
100 #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
101 #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT))
111 first = BTOPR(offset);
112 final = BTOPRT(offset + length - 1);
113 first = min(first, final);
116 mask <<= BITS_PER_LONG - (final - first);
117 mask >>= BITS_PER_LONG - (final);
119 ASSERT(offset + length <= PAGE_CACHE_SIZE);
120 ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
131 set_page_private(page,
132 page_private(page) | page_region_mask(offset, length));
133 if (page_private(page) == ~0UL)
134 SetPageUptodate(page);
143 unsigned long mask = page_region_mask(offset, length);
145 return (mask && (page_private(page) & mask) == mask);
149 * Mapping of multi-page buffers into contiguous virtual space
152 typedef struct a_list {
157 static a_list_t *as_free_head;
158 static int as_list_len;
159 static DEFINE_SPINLOCK(as_lock);
162 * Try to batch vunmaps because they are costly.
172 * Xen needs to be able to make sure it can get an exclusive
173 * RO mapping of pages it wants to turn into a pagetable. If
174 * a newly allocated page is also still being vmap()ed by xfs,
175 * it will cause pagetable construction to fail. This is a
176 * quick workaround to always eagerly unmap pages so that Xen
183 aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
184 if (likely(aentry)) {
186 aentry->next = as_free_head;
187 aentry->vm_addr = addr;
188 as_free_head = aentry;
190 spin_unlock(&as_lock);
197 purge_addresses(void)
199 a_list_t *aentry, *old;
201 if (as_free_head == NULL)
205 aentry = as_free_head;
208 spin_unlock(&as_lock);
210 while ((old = aentry) != NULL) {
211 vunmap(aentry->vm_addr);
212 aentry = aentry->next;
218 * Internal xfs_buf_t object manipulation
224 xfs_buftarg_t *target,
225 xfs_off_t range_base,
227 xfs_buf_flags_t flags)
230 * We don't want certain flags to appear in b_flags.
232 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
234 memset(bp, 0, sizeof(xfs_buf_t));
235 atomic_set(&bp->b_hold, 1);
236 init_completion(&bp->b_iowait);
237 INIT_LIST_HEAD(&bp->b_list);
238 INIT_LIST_HEAD(&bp->b_hash_list);
239 init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
241 bp->b_target = target;
242 bp->b_file_offset = range_base;
244 * Set buffer_length and count_desired to the same value initially.
245 * I/O routines should use count_desired, which will be the same in
246 * most cases but may be reset (e.g. XFS recovery).
248 bp->b_buffer_length = bp->b_count_desired = range_length;
250 bp->b_bn = XFS_BUF_DADDR_NULL;
251 atomic_set(&bp->b_pin_count, 0);
252 init_waitqueue_head(&bp->b_waiters);
254 XFS_STATS_INC(xb_create);
256 trace_xfs_buf_init(bp, _RET_IP_);
260 * Allocate a page array capable of holding a specified number
261 * of pages, and point the page buf at it.
267 xfs_buf_flags_t flags)
269 /* Make sure that we have a page list */
270 if (bp->b_pages == NULL) {
271 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
272 bp->b_page_count = page_count;
273 if (page_count <= XB_PAGES) {
274 bp->b_pages = bp->b_page_array;
276 bp->b_pages = kmem_alloc(sizeof(struct page *) *
277 page_count, xb_to_km(flags));
278 if (bp->b_pages == NULL)
281 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
287 * Frees b_pages if it was allocated.
293 if (bp->b_pages != bp->b_page_array) {
294 kmem_free(bp->b_pages);
300 * Releases the specified buffer.
302 * The modification state of any associated pages is left unchanged.
303 * The buffer most not be on any hash - use xfs_buf_rele instead for
304 * hashed and refcounted buffers
310 trace_xfs_buf_free(bp, _RET_IP_);
312 ASSERT(list_empty(&bp->b_hash_list));
314 if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
317 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
318 free_address(bp->b_addr - bp->b_offset);
320 for (i = 0; i < bp->b_page_count; i++) {
321 struct page *page = bp->b_pages[i];
323 if (bp->b_flags & _XBF_PAGE_CACHE)
324 ASSERT(!PagePrivate(page));
325 page_cache_release(page);
328 _xfs_buf_free_pages(bp);
329 xfs_buf_deallocate(bp);
333 * Finds all pages for buffer in question and builds it's page list.
336 _xfs_buf_lookup_pages(
340 struct address_space *mapping = bp->b_target->bt_mapping;
341 size_t blocksize = bp->b_target->bt_bsize;
342 size_t size = bp->b_count_desired;
343 size_t nbytes, offset;
344 gfp_t gfp_mask = xb_to_gfp(flags);
345 unsigned short page_count, i;
350 end = bp->b_file_offset + bp->b_buffer_length;
351 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
353 error = _xfs_buf_get_pages(bp, page_count, flags);
356 bp->b_flags |= _XBF_PAGE_CACHE;
358 offset = bp->b_offset;
359 first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
361 for (i = 0; i < bp->b_page_count; i++) {
366 page = find_or_create_page(mapping, first + i, gfp_mask);
367 if (unlikely(page == NULL)) {
368 if (flags & XBF_READ_AHEAD) {
369 bp->b_page_count = i;
370 for (i = 0; i < bp->b_page_count; i++)
371 unlock_page(bp->b_pages[i]);
376 * This could deadlock.
378 * But until all the XFS lowlevel code is revamped to
379 * handle buffer allocation failures we can't do much.
381 if (!(++retries % 100))
383 "XFS: possible memory allocation "
384 "deadlock in %s (mode:0x%x)\n",
387 XFS_STATS_INC(xb_page_retries);
388 xfsbufd_wakeup(0, gfp_mask);
389 congestion_wait(BLK_RW_ASYNC, HZ/50);
393 XFS_STATS_INC(xb_page_found);
395 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
398 ASSERT(!PagePrivate(page));
399 if (!PageUptodate(page)) {
401 if (blocksize >= PAGE_CACHE_SIZE) {
402 if (flags & XBF_READ)
403 bp->b_flags |= _XBF_PAGE_LOCKED;
404 } else if (!PagePrivate(page)) {
405 if (test_page_region(page, offset, nbytes))
410 bp->b_pages[i] = page;
414 if (!(bp->b_flags & _XBF_PAGE_LOCKED)) {
415 for (i = 0; i < bp->b_page_count; i++)
416 unlock_page(bp->b_pages[i]);
419 if (page_count == bp->b_page_count)
420 bp->b_flags |= XBF_DONE;
426 * Map buffer into kernel address-space if nessecary.
433 /* A single page buffer is always mappable */
434 if (bp->b_page_count == 1) {
435 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
436 bp->b_flags |= XBF_MAPPED;
437 } else if (flags & XBF_MAPPED) {
438 if (as_list_len > 64)
440 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
441 VM_MAP, PAGE_KERNEL);
442 if (unlikely(bp->b_addr == NULL))
444 bp->b_addr += bp->b_offset;
445 bp->b_flags |= XBF_MAPPED;
452 * Finding and Reading Buffers
456 * Look up, and creates if absent, a lockable buffer for
457 * a given range of an inode. The buffer is returned
458 * locked. If other overlapping buffers exist, they are
459 * released before the new buffer is created and locked,
460 * which may imply that this call will block until those buffers
461 * are unlocked. No I/O is implied by this call.
465 xfs_buftarg_t *btp, /* block device target */
466 xfs_off_t ioff, /* starting offset of range */
467 size_t isize, /* length of range */
468 xfs_buf_flags_t flags,
471 xfs_off_t range_base;
476 range_base = (ioff << BBSHIFT);
477 range_length = (isize << BBSHIFT);
479 /* Check for IOs smaller than the sector size / not sector aligned */
480 ASSERT(!(range_length < (1 << btp->bt_sshift)));
481 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
483 hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
485 spin_lock(&hash->bh_lock);
487 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
488 ASSERT(btp == bp->b_target);
489 if (bp->b_file_offset == range_base &&
490 bp->b_buffer_length == range_length) {
492 * If we look at something, bring it to the
493 * front of the list for next time.
495 atomic_inc(&bp->b_hold);
496 list_move(&bp->b_hash_list, &hash->bh_list);
503 _xfs_buf_initialize(new_bp, btp, range_base,
504 range_length, flags);
505 new_bp->b_hash = hash;
506 list_add(&new_bp->b_hash_list, &hash->bh_list);
508 XFS_STATS_INC(xb_miss_locked);
511 spin_unlock(&hash->bh_lock);
515 spin_unlock(&hash->bh_lock);
517 /* Attempt to get the semaphore without sleeping,
518 * if this does not work then we need to drop the
519 * spinlock and do a hard attempt on the semaphore.
521 if (down_trylock(&bp->b_sema)) {
522 if (!(flags & XBF_TRYLOCK)) {
523 /* wait for buffer ownership */
525 XFS_STATS_INC(xb_get_locked_waited);
527 /* We asked for a trylock and failed, no need
528 * to look at file offset and length here, we
529 * know that this buffer at least overlaps our
530 * buffer and is locked, therefore our buffer
531 * either does not exist, or is this buffer.
534 XFS_STATS_INC(xb_busy_locked);
542 if (bp->b_flags & XBF_STALE) {
543 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
544 bp->b_flags &= XBF_MAPPED;
547 trace_xfs_buf_find(bp, flags, _RET_IP_);
548 XFS_STATS_INC(xb_get_locked);
553 * Assembles a buffer covering the specified range.
554 * Storage in memory for all portions of the buffer will be allocated,
555 * although backing storage may not be.
559 xfs_buftarg_t *target,/* target for buffer */
560 xfs_off_t ioff, /* starting offset of range */
561 size_t isize, /* length of range */
562 xfs_buf_flags_t flags)
564 xfs_buf_t *bp, *new_bp;
567 new_bp = xfs_buf_allocate(flags);
568 if (unlikely(!new_bp))
571 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
573 error = _xfs_buf_lookup_pages(bp, flags);
577 xfs_buf_deallocate(new_bp);
578 if (unlikely(bp == NULL))
582 for (i = 0; i < bp->b_page_count; i++)
583 mark_page_accessed(bp->b_pages[i]);
585 if (!(bp->b_flags & XBF_MAPPED)) {
586 error = _xfs_buf_map_pages(bp, flags);
587 if (unlikely(error)) {
588 printk(KERN_WARNING "%s: failed to map pages\n",
594 XFS_STATS_INC(xb_get);
597 * Always fill in the block number now, the mapped cases can do
598 * their own overlay of this later.
601 bp->b_count_desired = bp->b_buffer_length;
603 trace_xfs_buf_get(bp, flags, _RET_IP_);
607 if (flags & (XBF_LOCK | XBF_TRYLOCK))
616 xfs_buf_flags_t flags)
620 ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
621 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
623 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
624 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
625 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
626 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
628 status = xfs_buf_iorequest(bp);
629 if (!status && !(flags & XBF_ASYNC))
630 status = xfs_buf_iowait(bp);
636 xfs_buftarg_t *target,
639 xfs_buf_flags_t flags)
645 bp = xfs_buf_get(target, ioff, isize, flags);
647 trace_xfs_buf_read(bp, flags, _RET_IP_);
649 if (!XFS_BUF_ISDONE(bp)) {
650 XFS_STATS_INC(xb_get_read);
651 _xfs_buf_read(bp, flags);
652 } else if (flags & XBF_ASYNC) {
654 * Read ahead call which is already satisfied,
659 /* We do not want read in the flags */
660 bp->b_flags &= ~XBF_READ;
667 if (flags & (XBF_LOCK | XBF_TRYLOCK))
674 * If we are not low on memory then do the readahead in a deadlock
679 xfs_buftarg_t *target,
682 xfs_buf_flags_t flags)
684 struct backing_dev_info *bdi;
686 bdi = target->bt_mapping->backing_dev_info;
687 if (bdi_read_congested(bdi))
690 flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
691 xfs_buf_read(target, ioff, isize, flags);
697 xfs_buftarg_t *target)
701 bp = xfs_buf_allocate(0);
703 _xfs_buf_initialize(bp, target, 0, len, 0);
707 static inline struct page *
711 if ((!is_vmalloc_addr(addr))) {
712 return virt_to_page(addr);
714 return vmalloc_to_page(addr);
719 xfs_buf_associate_memory(
726 unsigned long pageaddr;
727 unsigned long offset;
731 pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
732 offset = (unsigned long)mem - pageaddr;
733 buflen = PAGE_CACHE_ALIGN(len + offset);
734 page_count = buflen >> PAGE_CACHE_SHIFT;
736 /* Free any previous set of page pointers */
738 _xfs_buf_free_pages(bp);
743 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
747 bp->b_offset = offset;
749 for (i = 0; i < bp->b_page_count; i++) {
750 bp->b_pages[i] = mem_to_page((void *)pageaddr);
751 pageaddr += PAGE_CACHE_SIZE;
754 bp->b_count_desired = len;
755 bp->b_buffer_length = buflen;
756 bp->b_flags |= XBF_MAPPED;
757 bp->b_flags &= ~_XBF_PAGE_LOCKED;
765 xfs_buftarg_t *target)
767 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
771 bp = xfs_buf_allocate(0);
772 if (unlikely(bp == NULL))
774 _xfs_buf_initialize(bp, target, 0, len, 0);
776 error = _xfs_buf_get_pages(bp, page_count, 0);
780 for (i = 0; i < page_count; i++) {
781 bp->b_pages[i] = alloc_page(GFP_KERNEL);
785 bp->b_flags |= _XBF_PAGES;
787 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
788 if (unlikely(error)) {
789 printk(KERN_WARNING "%s: failed to map pages\n",
796 trace_xfs_buf_get_noaddr(bp, _RET_IP_);
801 __free_page(bp->b_pages[i]);
802 _xfs_buf_free_pages(bp);
804 xfs_buf_deallocate(bp);
810 * Increment reference count on buffer, to hold the buffer concurrently
811 * with another thread which may release (free) the buffer asynchronously.
812 * Must hold the buffer already to call this function.
818 trace_xfs_buf_hold(bp, _RET_IP_);
819 atomic_inc(&bp->b_hold);
823 * Releases a hold on the specified buffer. If the
824 * the hold count is 1, calls xfs_buf_free.
830 xfs_bufhash_t *hash = bp->b_hash;
832 trace_xfs_buf_rele(bp, _RET_IP_);
834 if (unlikely(!hash)) {
835 ASSERT(!bp->b_relse);
836 if (atomic_dec_and_test(&bp->b_hold))
841 ASSERT(atomic_read(&bp->b_hold) > 0);
842 if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
844 atomic_inc(&bp->b_hold);
845 spin_unlock(&hash->bh_lock);
846 (*(bp->b_relse)) (bp);
847 } else if (bp->b_flags & XBF_FS_MANAGED) {
848 spin_unlock(&hash->bh_lock);
850 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
851 list_del_init(&bp->b_hash_list);
852 spin_unlock(&hash->bh_lock);
860 * Mutual exclusion on buffers. Locking model:
862 * Buffers associated with inodes for which buffer locking
863 * is not enabled are not protected by semaphores, and are
864 * assumed to be exclusively owned by the caller. There is a
865 * spinlock in the buffer, used by the caller when concurrent
866 * access is possible.
870 * Locks a buffer object, if it is not already locked.
871 * Note that this in no way locks the underlying pages, so it is only
872 * useful for synchronizing concurrent use of buffer objects, not for
873 * synchronizing independent access to the underlying pages.
881 locked = down_trylock(&bp->b_sema) == 0;
885 trace_xfs_buf_cond_lock(bp, _RET_IP_);
886 return locked ? 0 : -EBUSY;
893 return bp->b_sema.count;
897 * Locks a buffer object.
898 * Note that this in no way locks the underlying pages, so it is only
899 * useful for synchronizing concurrent use of buffer objects, not for
900 * synchronizing independent access to the underlying pages.
906 trace_xfs_buf_lock(bp, _RET_IP_);
908 if (atomic_read(&bp->b_io_remaining))
909 blk_run_address_space(bp->b_target->bt_mapping);
913 trace_xfs_buf_lock_done(bp, _RET_IP_);
917 * Releases the lock on the buffer object.
918 * If the buffer is marked delwri but is not queued, do so before we
919 * unlock the buffer as we need to set flags correctly. We also need to
920 * take a reference for the delwri queue because the unlocker is going to
921 * drop their's and they don't know we just queued it.
927 if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
928 atomic_inc(&bp->b_hold);
929 bp->b_flags |= XBF_ASYNC;
930 xfs_buf_delwri_queue(bp, 0);
936 trace_xfs_buf_unlock(bp, _RET_IP_);
941 * Pinning Buffer Storage in Memory
942 * Ensure that no attempt to force a buffer to disk will succeed.
948 trace_xfs_buf_pin(bp, _RET_IP_);
949 atomic_inc(&bp->b_pin_count);
956 trace_xfs_buf_unpin(bp, _RET_IP_);
958 if (atomic_dec_and_test(&bp->b_pin_count))
959 wake_up_all(&bp->b_waiters);
966 return atomic_read(&bp->b_pin_count);
973 DECLARE_WAITQUEUE (wait, current);
975 if (atomic_read(&bp->b_pin_count) == 0)
978 add_wait_queue(&bp->b_waiters, &wait);
980 set_current_state(TASK_UNINTERRUPTIBLE);
981 if (atomic_read(&bp->b_pin_count) == 0)
983 if (atomic_read(&bp->b_io_remaining))
984 blk_run_address_space(bp->b_target->bt_mapping);
987 remove_wait_queue(&bp->b_waiters, &wait);
988 set_current_state(TASK_RUNNING);
992 * Buffer Utility Routines
997 struct work_struct *work)
1000 container_of(work, xfs_buf_t, b_iodone_work);
1003 * We can get an EOPNOTSUPP to ordered writes. Here we clear the
1004 * ordered flag and reissue them. Because we can't tell the higher
1005 * layers directly that they should not issue ordered I/O anymore, they
1006 * need to check if the _XFS_BARRIER_FAILED flag was set during I/O completion.
1008 if ((bp->b_error == EOPNOTSUPP) &&
1009 (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1010 trace_xfs_buf_ordered_retry(bp, _RET_IP_);
1011 bp->b_flags &= ~XBF_ORDERED;
1012 bp->b_flags |= _XFS_BARRIER_FAILED;
1013 xfs_buf_iorequest(bp);
1014 } else if (bp->b_iodone)
1015 (*(bp->b_iodone))(bp);
1016 else if (bp->b_flags & XBF_ASYNC)
1025 trace_xfs_buf_iodone(bp, _RET_IP_);
1027 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1028 if (bp->b_error == 0)
1029 bp->b_flags |= XBF_DONE;
1031 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1033 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1034 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1036 xfs_buf_iodone_work(&bp->b_iodone_work);
1039 complete(&bp->b_iowait);
1048 ASSERT(error >= 0 && error <= 0xffff);
1049 bp->b_error = (unsigned short)error;
1050 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1058 trace_xfs_buf_bawrite(bp, _RET_IP_);
1060 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
1062 xfs_buf_delwri_dequeue(bp);
1064 bp->b_flags &= ~(XBF_READ | XBF_DELWRI | XBF_READ_AHEAD);
1065 bp->b_flags |= (XBF_WRITE | XBF_ASYNC | _XBF_RUN_QUEUES);
1068 bp->b_strat = xfs_bdstrat_cb;
1069 return xfs_bdstrat_cb(bp);
1077 trace_xfs_buf_bdwrite(bp, _RET_IP_);
1079 bp->b_strat = xfs_bdstrat_cb;
1082 bp->b_flags &= ~XBF_READ;
1083 bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1085 xfs_buf_delwri_queue(bp, 1);
1093 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1094 bp->b_flags &= ~_XBF_PAGE_LOCKED;
1095 xfs_buf_ioend(bp, schedule);
1104 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1105 unsigned int blocksize = bp->b_target->bt_bsize;
1106 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1108 xfs_buf_ioerror(bp, -error);
1111 struct page *page = bvec->bv_page;
1113 ASSERT(!PagePrivate(page));
1114 if (unlikely(bp->b_error)) {
1115 if (bp->b_flags & XBF_READ)
1116 ClearPageUptodate(page);
1117 } else if (blocksize >= PAGE_CACHE_SIZE) {
1118 SetPageUptodate(page);
1119 } else if (!PagePrivate(page) &&
1120 (bp->b_flags & _XBF_PAGE_CACHE)) {
1121 set_page_region(page, bvec->bv_offset, bvec->bv_len);
1124 if (--bvec >= bio->bi_io_vec)
1125 prefetchw(&bvec->bv_page->flags);
1127 if (bp->b_flags & _XBF_PAGE_LOCKED)
1129 } while (bvec >= bio->bi_io_vec);
1131 _xfs_buf_ioend(bp, 1);
1139 int rw, map_i, total_nr_pages, nr_pages;
1141 int offset = bp->b_offset;
1142 int size = bp->b_count_desired;
1143 sector_t sector = bp->b_bn;
1144 unsigned int blocksize = bp->b_target->bt_bsize;
1146 total_nr_pages = bp->b_page_count;
1149 if (bp->b_flags & XBF_ORDERED) {
1150 ASSERT(!(bp->b_flags & XBF_READ));
1152 } else if (bp->b_flags & XBF_LOG_BUFFER) {
1153 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1154 bp->b_flags &= ~_XBF_RUN_QUEUES;
1155 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1156 } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1157 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1158 bp->b_flags &= ~_XBF_RUN_QUEUES;
1159 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1161 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1162 (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1165 /* Special code path for reading a sub page size buffer in --
1166 * we populate up the whole page, and hence the other metadata
1167 * in the same page. This optimization is only valid when the
1168 * filesystem block size is not smaller than the page size.
1170 if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1171 ((bp->b_flags & (XBF_READ|_XBF_PAGE_LOCKED)) ==
1172 (XBF_READ|_XBF_PAGE_LOCKED)) &&
1173 (blocksize >= PAGE_CACHE_SIZE)) {
1174 bio = bio_alloc(GFP_NOIO, 1);
1176 bio->bi_bdev = bp->b_target->bt_bdev;
1177 bio->bi_sector = sector - (offset >> BBSHIFT);
1178 bio->bi_end_io = xfs_buf_bio_end_io;
1179 bio->bi_private = bp;
1181 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1184 atomic_inc(&bp->b_io_remaining);
1190 atomic_inc(&bp->b_io_remaining);
1191 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1192 if (nr_pages > total_nr_pages)
1193 nr_pages = total_nr_pages;
1195 bio = bio_alloc(GFP_NOIO, nr_pages);
1196 bio->bi_bdev = bp->b_target->bt_bdev;
1197 bio->bi_sector = sector;
1198 bio->bi_end_io = xfs_buf_bio_end_io;
1199 bio->bi_private = bp;
1201 for (; size && nr_pages; nr_pages--, map_i++) {
1202 int rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1207 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1208 if (rbytes < nbytes)
1212 sector += nbytes >> BBSHIFT;
1218 if (likely(bio->bi_size)) {
1219 submit_bio(rw, bio);
1224 xfs_buf_ioerror(bp, EIO);
1232 trace_xfs_buf_iorequest(bp, _RET_IP_);
1234 if (bp->b_flags & XBF_DELWRI) {
1235 xfs_buf_delwri_queue(bp, 1);
1239 if (bp->b_flags & XBF_WRITE) {
1240 xfs_buf_wait_unpin(bp);
1245 /* Set the count to 1 initially, this will stop an I/O
1246 * completion callout which happens before we have started
1247 * all the I/O from calling xfs_buf_ioend too early.
1249 atomic_set(&bp->b_io_remaining, 1);
1250 _xfs_buf_ioapply(bp);
1251 _xfs_buf_ioend(bp, 0);
1258 * Waits for I/O to complete on the buffer supplied.
1259 * It returns immediately if no I/O is pending.
1260 * It returns the I/O error code, if any, or 0 if there was no error.
1266 trace_xfs_buf_iowait(bp, _RET_IP_);
1268 if (atomic_read(&bp->b_io_remaining))
1269 blk_run_address_space(bp->b_target->bt_mapping);
1270 wait_for_completion(&bp->b_iowait);
1272 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1283 if (bp->b_flags & XBF_MAPPED)
1284 return XFS_BUF_PTR(bp) + offset;
1286 offset += bp->b_offset;
1287 page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1288 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1292 * Move data into or out of a buffer.
1296 xfs_buf_t *bp, /* buffer to process */
1297 size_t boff, /* starting buffer offset */
1298 size_t bsize, /* length to copy */
1299 caddr_t data, /* data address */
1300 xfs_buf_rw_t mode) /* read/write/zero flag */
1302 size_t bend, cpoff, csize;
1305 bend = boff + bsize;
1306 while (boff < bend) {
1307 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1308 cpoff = xfs_buf_poff(boff + bp->b_offset);
1309 csize = min_t(size_t,
1310 PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1312 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1316 memset(page_address(page) + cpoff, 0, csize);
1319 memcpy(data, page_address(page) + cpoff, csize);
1322 memcpy(page_address(page) + cpoff, data, csize);
1331 * Handling of buffer targets (buftargs).
1335 * Wait for any bufs with callbacks that have been submitted but
1336 * have not yet returned... walk the hash list for the target.
1343 xfs_bufhash_t *hash;
1346 for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1347 hash = &btp->bt_hash[i];
1349 spin_lock(&hash->bh_lock);
1350 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1351 ASSERT(btp == bp->b_target);
1352 if (!(bp->b_flags & XBF_FS_MANAGED)) {
1353 spin_unlock(&hash->bh_lock);
1355 * Catch superblock reference count leaks
1358 BUG_ON(bp->b_bn == 0);
1363 spin_unlock(&hash->bh_lock);
1368 * Allocate buffer hash table for a given target.
1369 * For devices containing metadata (i.e. not the log/realtime devices)
1370 * we need to allocate a much larger hash table.
1379 btp->bt_hashshift = external ? 3 : 8; /* 8 or 256 buckets */
1380 btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1381 btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1382 sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1383 for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1384 spin_lock_init(&btp->bt_hash[i].bh_lock);
1385 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1393 kmem_free(btp->bt_hash);
1394 btp->bt_hash = NULL;
1398 * buftarg list for delwrite queue processing
1400 static LIST_HEAD(xfs_buftarg_list);
1401 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1404 xfs_register_buftarg(
1407 spin_lock(&xfs_buftarg_lock);
1408 list_add(&btp->bt_list, &xfs_buftarg_list);
1409 spin_unlock(&xfs_buftarg_lock);
1413 xfs_unregister_buftarg(
1416 spin_lock(&xfs_buftarg_lock);
1417 list_del(&btp->bt_list);
1418 spin_unlock(&xfs_buftarg_lock);
1423 struct xfs_mount *mp,
1424 struct xfs_buftarg *btp)
1426 xfs_flush_buftarg(btp, 1);
1427 if (mp->m_flags & XFS_MOUNT_BARRIER)
1428 xfs_blkdev_issue_flush(btp);
1429 xfs_free_bufhash(btp);
1430 iput(btp->bt_mapping->host);
1432 /* Unregister the buftarg first so that we don't get a
1433 * wakeup finding a non-existent task
1435 xfs_unregister_buftarg(btp);
1436 kthread_stop(btp->bt_task);
1442 xfs_setsize_buftarg_flags(
1444 unsigned int blocksize,
1445 unsigned int sectorsize,
1448 btp->bt_bsize = blocksize;
1449 btp->bt_sshift = ffs(sectorsize) - 1;
1450 btp->bt_smask = sectorsize - 1;
1452 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1454 "XFS: Cannot set_blocksize to %u on device %s\n",
1455 sectorsize, XFS_BUFTARG_NAME(btp));
1460 (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1462 "XFS: %u byte sectors in use on device %s. "
1463 "This is suboptimal; %u or greater is ideal.\n",
1464 sectorsize, XFS_BUFTARG_NAME(btp),
1465 (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1472 * When allocating the initial buffer target we have not yet
1473 * read in the superblock, so don't know what sized sectors
1474 * are being used is at this early stage. Play safe.
1477 xfs_setsize_buftarg_early(
1479 struct block_device *bdev)
1481 return xfs_setsize_buftarg_flags(btp,
1482 PAGE_CACHE_SIZE, bdev_logical_block_size(bdev), 0);
1486 xfs_setsize_buftarg(
1488 unsigned int blocksize,
1489 unsigned int sectorsize)
1491 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1495 xfs_mapping_buftarg(
1497 struct block_device *bdev)
1499 struct backing_dev_info *bdi;
1500 struct inode *inode;
1501 struct address_space *mapping;
1502 static const struct address_space_operations mapping_aops = {
1503 .sync_page = block_sync_page,
1504 .migratepage = fail_migrate_page,
1507 inode = new_inode(bdev->bd_inode->i_sb);
1510 "XFS: Cannot allocate mapping inode for device %s\n",
1511 XFS_BUFTARG_NAME(btp));
1514 inode->i_mode = S_IFBLK;
1515 inode->i_bdev = bdev;
1516 inode->i_rdev = bdev->bd_dev;
1517 bdi = blk_get_backing_dev_info(bdev);
1519 bdi = &default_backing_dev_info;
1520 mapping = &inode->i_data;
1521 mapping->a_ops = &mapping_aops;
1522 mapping->backing_dev_info = bdi;
1523 mapping_set_gfp_mask(mapping, GFP_NOFS);
1524 btp->bt_mapping = mapping;
1529 xfs_alloc_delwrite_queue(
1534 INIT_LIST_HEAD(&btp->bt_list);
1535 INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1536 spin_lock_init(&btp->bt_delwrite_lock);
1538 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1539 if (IS_ERR(btp->bt_task)) {
1540 error = PTR_ERR(btp->bt_task);
1543 xfs_register_buftarg(btp);
1550 struct block_device *bdev,
1555 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1557 btp->bt_dev = bdev->bd_dev;
1558 btp->bt_bdev = bdev;
1559 if (xfs_setsize_buftarg_early(btp, bdev))
1561 if (xfs_mapping_buftarg(btp, bdev))
1563 if (xfs_alloc_delwrite_queue(btp))
1565 xfs_alloc_bufhash(btp, external);
1575 * Delayed write buffer handling
1578 xfs_buf_delwri_queue(
1582 struct list_head *dwq = &bp->b_target->bt_delwrite_queue;
1583 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1585 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1587 ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1590 /* If already in the queue, dequeue and place at tail */
1591 if (!list_empty(&bp->b_list)) {
1592 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1594 atomic_dec(&bp->b_hold);
1595 list_del(&bp->b_list);
1598 bp->b_flags |= _XBF_DELWRI_Q;
1599 list_add_tail(&bp->b_list, dwq);
1600 bp->b_queuetime = jiffies;
1608 xfs_buf_delwri_dequeue(
1611 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1615 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1616 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1617 list_del_init(&bp->b_list);
1620 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1626 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1630 xfs_buf_runall_queues(
1631 struct workqueue_struct *queue)
1633 flush_workqueue(queue);
1643 spin_lock(&xfs_buftarg_lock);
1644 list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1645 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1647 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1648 wake_up_process(btp->bt_task);
1650 spin_unlock(&xfs_buftarg_lock);
1655 * Move as many buffers as specified to the supplied list
1656 * idicating if we skipped any buffers to prevent deadlocks.
1659 xfs_buf_delwri_split(
1660 xfs_buftarg_t *target,
1661 struct list_head *list,
1665 struct list_head *dwq = &target->bt_delwrite_queue;
1666 spinlock_t *dwlk = &target->bt_delwrite_lock;
1670 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1671 INIT_LIST_HEAD(list);
1673 list_for_each_entry_safe(bp, n, dwq, b_list) {
1674 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1675 ASSERT(bp->b_flags & XBF_DELWRI);
1677 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1679 time_before(jiffies, bp->b_queuetime + age)) {
1684 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1686 bp->b_flags |= XBF_WRITE;
1687 list_move_tail(&bp->b_list, list);
1701 struct list_head tmp;
1702 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1706 current->flags |= PF_MEMALLOC;
1711 if (unlikely(freezing(current))) {
1712 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1715 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1718 schedule_timeout_interruptible(
1719 xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1721 xfs_buf_delwri_split(target, &tmp,
1722 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1725 while (!list_empty(&tmp)) {
1726 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1727 ASSERT(target == bp->b_target);
1729 list_del_init(&bp->b_list);
1730 xfs_buf_iostrategy(bp);
1734 if (as_list_len > 0)
1737 blk_run_address_space(target->bt_mapping);
1739 } while (!kthread_should_stop());
1745 * Go through all incore buffers, and release buffers if they belong to
1746 * the given device. This is used in filesystem error handling to
1747 * preserve the consistency of its metadata.
1751 xfs_buftarg_t *target,
1754 struct list_head tmp;
1758 xfs_buf_runall_queues(xfsconvertd_workqueue);
1759 xfs_buf_runall_queues(xfsdatad_workqueue);
1760 xfs_buf_runall_queues(xfslogd_workqueue);
1762 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1763 pincount = xfs_buf_delwri_split(target, &tmp, 0);
1766 * Dropped the delayed write list lock, now walk the temporary list
1768 list_for_each_entry_safe(bp, n, &tmp, b_list) {
1769 ASSERT(target == bp->b_target);
1771 bp->b_flags &= ~XBF_ASYNC;
1773 list_del_init(&bp->b_list);
1775 xfs_buf_iostrategy(bp);
1779 blk_run_address_space(target->bt_mapping);
1782 * Remaining list items must be flushed before returning
1784 while (!list_empty(&tmp)) {
1785 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1787 list_del_init(&bp->b_list);
1798 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1799 KM_ZONE_HWALIGN, NULL);
1803 xfslogd_workqueue = create_workqueue("xfslogd");
1804 if (!xfslogd_workqueue)
1805 goto out_free_buf_zone;
1807 xfsdatad_workqueue = create_workqueue("xfsdatad");
1808 if (!xfsdatad_workqueue)
1809 goto out_destroy_xfslogd_workqueue;
1811 xfsconvertd_workqueue = create_workqueue("xfsconvertd");
1812 if (!xfsconvertd_workqueue)
1813 goto out_destroy_xfsdatad_workqueue;
1815 register_shrinker(&xfs_buf_shake);
1818 out_destroy_xfsdatad_workqueue:
1819 destroy_workqueue(xfsdatad_workqueue);
1820 out_destroy_xfslogd_workqueue:
1821 destroy_workqueue(xfslogd_workqueue);
1823 kmem_zone_destroy(xfs_buf_zone);
1829 xfs_buf_terminate(void)
1831 unregister_shrinker(&xfs_buf_shake);
1832 destroy_workqueue(xfsconvertd_workqueue);
1833 destroy_workqueue(xfsdatad_workqueue);
1834 destroy_workqueue(xfslogd_workqueue);
1835 kmem_zone_destroy(xfs_buf_zone);
1838 #ifdef CONFIG_KDB_MODULES
1840 xfs_get_buftarg_list(void)
1842 return &xfs_buftarg_list;