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>
37 static kmem_zone_t *xfs_buf_zone;
38 static kmem_shaker_t xfs_buf_shake;
39 STATIC int xfsbufd(void *);
40 STATIC int xfsbufd_wakeup(int, gfp_t);
41 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
43 static struct workqueue_struct *xfslogd_workqueue;
44 struct workqueue_struct *xfsdatad_workqueue;
54 ktrace_enter(xfs_buf_trace_buf,
56 (void *)(unsigned long)bp->b_flags,
57 (void *)(unsigned long)bp->b_hold.counter,
58 (void *)(unsigned long)bp->b_sema.count.counter,
61 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
62 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
63 (void *)(unsigned long)bp->b_buffer_length,
64 NULL, NULL, NULL, NULL, NULL);
66 ktrace_t *xfs_buf_trace_buf;
67 #define XFS_BUF_TRACE_SIZE 4096
68 #define XB_TRACE(bp, id, data) \
69 xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
71 #define XB_TRACE(bp, id, data) do { } while (0)
74 #ifdef XFS_BUF_LOCK_TRACKING
75 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
76 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
77 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
79 # define XB_SET_OWNER(bp) do { } while (0)
80 # define XB_CLEAR_OWNER(bp) do { } while (0)
81 # define XB_GET_OWNER(bp) do { } while (0)
84 #define xb_to_gfp(flags) \
85 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
86 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
88 #define xb_to_km(flags) \
89 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
91 #define xfs_buf_allocate(flags) \
92 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
93 #define xfs_buf_deallocate(bp) \
94 kmem_zone_free(xfs_buf_zone, (bp));
97 * Page Region interfaces.
99 * For pages in filesystems where the blocksize is smaller than the
100 * pagesize, we use the page->private field (long) to hold a bitmap
101 * of uptodate regions within the page.
103 * Each such region is "bytes per page / bits per long" bytes long.
105 * NBPPR == number-of-bytes-per-page-region
106 * BTOPR == bytes-to-page-region (rounded up)
107 * BTOPRT == bytes-to-page-region-truncated (rounded down)
109 #if (BITS_PER_LONG == 32)
110 #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */
111 #elif (BITS_PER_LONG == 64)
112 #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */
114 #error BITS_PER_LONG must be 32 or 64
116 #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG)
117 #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
118 #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT))
128 first = BTOPR(offset);
129 final = BTOPRT(offset + length - 1);
130 first = min(first, final);
133 mask <<= BITS_PER_LONG - (final - first);
134 mask >>= BITS_PER_LONG - (final);
136 ASSERT(offset + length <= PAGE_CACHE_SIZE);
137 ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
148 set_page_private(page,
149 page_private(page) | page_region_mask(offset, length));
150 if (page_private(page) == ~0UL)
151 SetPageUptodate(page);
160 unsigned long mask = page_region_mask(offset, length);
162 return (mask && (page_private(page) & mask) == mask);
166 * Mapping of multi-page buffers into contiguous virtual space
169 typedef struct a_list {
174 static a_list_t *as_free_head;
175 static int as_list_len;
176 static DEFINE_SPINLOCK(as_lock);
179 * Try to batch vunmaps because they are costly.
187 aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
188 if (likely(aentry)) {
190 aentry->next = as_free_head;
191 aentry->vm_addr = addr;
192 as_free_head = aentry;
194 spin_unlock(&as_lock);
201 purge_addresses(void)
203 a_list_t *aentry, *old;
205 if (as_free_head == NULL)
209 aentry = as_free_head;
212 spin_unlock(&as_lock);
214 while ((old = aentry) != NULL) {
215 vunmap(aentry->vm_addr);
216 aentry = aentry->next;
222 * Internal xfs_buf_t object manipulation
228 xfs_buftarg_t *target,
229 xfs_off_t range_base,
231 xfs_buf_flags_t flags)
234 * We don't want certain flags to appear in b_flags.
236 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
238 memset(bp, 0, sizeof(xfs_buf_t));
239 atomic_set(&bp->b_hold, 1);
240 init_MUTEX_LOCKED(&bp->b_iodonesema);
241 INIT_LIST_HEAD(&bp->b_list);
242 INIT_LIST_HEAD(&bp->b_hash_list);
243 init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
245 bp->b_target = target;
246 bp->b_file_offset = range_base;
248 * Set buffer_length and count_desired to the same value initially.
249 * I/O routines should use count_desired, which will be the same in
250 * most cases but may be reset (e.g. XFS recovery).
252 bp->b_buffer_length = bp->b_count_desired = range_length;
254 bp->b_bn = XFS_BUF_DADDR_NULL;
255 atomic_set(&bp->b_pin_count, 0);
256 init_waitqueue_head(&bp->b_waiters);
258 XFS_STATS_INC(xb_create);
259 XB_TRACE(bp, "initialize", target);
263 * Allocate a page array capable of holding a specified number
264 * of pages, and point the page buf at it.
270 xfs_buf_flags_t flags)
272 /* Make sure that we have a page list */
273 if (bp->b_pages == NULL) {
274 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
275 bp->b_page_count = page_count;
276 if (page_count <= XB_PAGES) {
277 bp->b_pages = bp->b_page_array;
279 bp->b_pages = kmem_alloc(sizeof(struct page *) *
280 page_count, xb_to_km(flags));
281 if (bp->b_pages == NULL)
284 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
290 * Frees b_pages if it was allocated.
296 if (bp->b_pages != bp->b_page_array) {
297 kmem_free(bp->b_pages,
298 bp->b_page_count * sizeof(struct page *));
303 * Releases the specified buffer.
305 * The modification state of any associated pages is left unchanged.
306 * The buffer most not be on any hash - use xfs_buf_rele instead for
307 * hashed and refcounted buffers
313 XB_TRACE(bp, "free", 0);
315 ASSERT(list_empty(&bp->b_hash_list));
317 if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
320 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
321 free_address(bp->b_addr - bp->b_offset);
323 for (i = 0; i < bp->b_page_count; i++) {
324 struct page *page = bp->b_pages[i];
326 if (bp->b_flags & _XBF_PAGE_CACHE)
327 ASSERT(!PagePrivate(page));
328 page_cache_release(page);
330 _xfs_buf_free_pages(bp);
333 xfs_buf_deallocate(bp);
337 * Finds all pages for buffer in question and builds it's page list.
340 _xfs_buf_lookup_pages(
344 struct address_space *mapping = bp->b_target->bt_mapping;
345 size_t blocksize = bp->b_target->bt_bsize;
346 size_t size = bp->b_count_desired;
347 size_t nbytes, offset;
348 gfp_t gfp_mask = xb_to_gfp(flags);
349 unsigned short page_count, i;
354 end = bp->b_file_offset + bp->b_buffer_length;
355 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
357 error = _xfs_buf_get_pages(bp, page_count, flags);
360 bp->b_flags |= _XBF_PAGE_CACHE;
362 offset = bp->b_offset;
363 first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
365 for (i = 0; i < bp->b_page_count; i++) {
370 page = find_or_create_page(mapping, first + i, gfp_mask);
371 if (unlikely(page == NULL)) {
372 if (flags & XBF_READ_AHEAD) {
373 bp->b_page_count = i;
374 for (i = 0; i < bp->b_page_count; i++)
375 unlock_page(bp->b_pages[i]);
380 * This could deadlock.
382 * But until all the XFS lowlevel code is revamped to
383 * handle buffer allocation failures we can't do much.
385 if (!(++retries % 100))
387 "XFS: possible memory allocation "
388 "deadlock in %s (mode:0x%x)\n",
389 __FUNCTION__, gfp_mask);
391 XFS_STATS_INC(xb_page_retries);
392 xfsbufd_wakeup(0, gfp_mask);
393 congestion_wait(WRITE, HZ/50);
397 XFS_STATS_INC(xb_page_found);
399 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
402 ASSERT(!PagePrivate(page));
403 if (!PageUptodate(page)) {
405 if (blocksize >= PAGE_CACHE_SIZE) {
406 if (flags & XBF_READ)
408 } else if (!PagePrivate(page)) {
409 if (test_page_region(page, offset, nbytes))
414 bp->b_pages[i] = page;
419 for (i = 0; i < bp->b_page_count; i++)
420 unlock_page(bp->b_pages[i]);
423 if (page_count == bp->b_page_count)
424 bp->b_flags |= XBF_DONE;
426 XB_TRACE(bp, "lookup_pages", (long)page_count);
431 * Map buffer into kernel address-space if nessecary.
438 /* A single page buffer is always mappable */
439 if (bp->b_page_count == 1) {
440 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
441 bp->b_flags |= XBF_MAPPED;
442 } else if (flags & XBF_MAPPED) {
443 if (as_list_len > 64)
445 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
446 VM_MAP, PAGE_KERNEL);
447 if (unlikely(bp->b_addr == NULL))
449 bp->b_addr += bp->b_offset;
450 bp->b_flags |= XBF_MAPPED;
457 * Finding and Reading Buffers
461 * Look up, and creates if absent, a lockable buffer for
462 * a given range of an inode. The buffer is returned
463 * locked. If other overlapping buffers exist, they are
464 * released before the new buffer is created and locked,
465 * which may imply that this call will block until those buffers
466 * are unlocked. No I/O is implied by this call.
470 xfs_buftarg_t *btp, /* block device target */
471 xfs_off_t ioff, /* starting offset of range */
472 size_t isize, /* length of range */
473 xfs_buf_flags_t flags,
476 xfs_off_t range_base;
481 range_base = (ioff << BBSHIFT);
482 range_length = (isize << BBSHIFT);
484 /* Check for IOs smaller than the sector size / not sector aligned */
485 ASSERT(!(range_length < (1 << btp->bt_sshift)));
486 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
488 hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
490 spin_lock(&hash->bh_lock);
492 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
493 ASSERT(btp == bp->b_target);
494 if (bp->b_file_offset == range_base &&
495 bp->b_buffer_length == range_length) {
497 * If we look at something, bring it to the
498 * front of the list for next time.
500 atomic_inc(&bp->b_hold);
501 list_move(&bp->b_hash_list, &hash->bh_list);
508 _xfs_buf_initialize(new_bp, btp, range_base,
509 range_length, flags);
510 new_bp->b_hash = hash;
511 list_add(&new_bp->b_hash_list, &hash->bh_list);
513 XFS_STATS_INC(xb_miss_locked);
516 spin_unlock(&hash->bh_lock);
520 spin_unlock(&hash->bh_lock);
522 /* Attempt to get the semaphore without sleeping,
523 * if this does not work then we need to drop the
524 * spinlock and do a hard attempt on the semaphore.
526 if (down_trylock(&bp->b_sema)) {
527 if (!(flags & XBF_TRYLOCK)) {
528 /* wait for buffer ownership */
529 XB_TRACE(bp, "get_lock", 0);
531 XFS_STATS_INC(xb_get_locked_waited);
533 /* We asked for a trylock and failed, no need
534 * to look at file offset and length here, we
535 * know that this buffer at least overlaps our
536 * buffer and is locked, therefore our buffer
537 * either does not exist, or is this buffer.
540 XFS_STATS_INC(xb_busy_locked);
548 if (bp->b_flags & XBF_STALE) {
549 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
550 bp->b_flags &= XBF_MAPPED;
552 XB_TRACE(bp, "got_lock", 0);
553 XFS_STATS_INC(xb_get_locked);
558 * Assembles a buffer covering the specified range.
559 * Storage in memory for all portions of the buffer will be allocated,
560 * although backing storage may not be.
564 xfs_buftarg_t *target,/* target for buffer */
565 xfs_off_t ioff, /* starting offset of range */
566 size_t isize, /* length of range */
567 xfs_buf_flags_t flags)
569 xfs_buf_t *bp, *new_bp;
572 new_bp = xfs_buf_allocate(flags);
573 if (unlikely(!new_bp))
576 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
578 error = _xfs_buf_lookup_pages(bp, flags);
582 xfs_buf_deallocate(new_bp);
583 if (unlikely(bp == NULL))
587 for (i = 0; i < bp->b_page_count; i++)
588 mark_page_accessed(bp->b_pages[i]);
590 if (!(bp->b_flags & XBF_MAPPED)) {
591 error = _xfs_buf_map_pages(bp, flags);
592 if (unlikely(error)) {
593 printk(KERN_WARNING "%s: failed to map pages\n",
599 XFS_STATS_INC(xb_get);
602 * Always fill in the block number now, the mapped cases can do
603 * their own overlay of this later.
606 bp->b_count_desired = bp->b_buffer_length;
608 XB_TRACE(bp, "get", (unsigned long)flags);
612 if (flags & (XBF_LOCK | XBF_TRYLOCK))
620 xfs_buftarg_t *target,
623 xfs_buf_flags_t flags)
629 bp = xfs_buf_get_flags(target, ioff, isize, flags);
631 if (!XFS_BUF_ISDONE(bp)) {
632 XB_TRACE(bp, "read", (unsigned long)flags);
633 XFS_STATS_INC(xb_get_read);
634 xfs_buf_iostart(bp, flags);
635 } else if (flags & XBF_ASYNC) {
636 XB_TRACE(bp, "read_async", (unsigned long)flags);
638 * Read ahead call which is already satisfied,
643 XB_TRACE(bp, "read_done", (unsigned long)flags);
644 /* We do not want read in the flags */
645 bp->b_flags &= ~XBF_READ;
652 if (flags & (XBF_LOCK | XBF_TRYLOCK))
659 * If we are not low on memory then do the readahead in a deadlock
664 xfs_buftarg_t *target,
667 xfs_buf_flags_t flags)
669 struct backing_dev_info *bdi;
671 bdi = target->bt_mapping->backing_dev_info;
672 if (bdi_read_congested(bdi))
675 flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
676 xfs_buf_read_flags(target, ioff, isize, flags);
682 xfs_buftarg_t *target)
686 bp = xfs_buf_allocate(0);
688 _xfs_buf_initialize(bp, target, 0, len, 0);
692 static inline struct page *
696 if (((unsigned long)addr < VMALLOC_START) ||
697 ((unsigned long)addr >= VMALLOC_END)) {
698 return virt_to_page(addr);
700 return vmalloc_to_page(addr);
705 xfs_buf_associate_memory(
717 page_count = PAGE_CACHE_ALIGN(len) >> PAGE_CACHE_SHIFT;
718 offset = (off_t) mem - ((off_t)mem & PAGE_CACHE_MASK);
719 if (offset && (len > PAGE_CACHE_SIZE))
722 /* Free any previous set of page pointers */
724 _xfs_buf_free_pages(bp);
729 rval = _xfs_buf_get_pages(bp, page_count, 0);
733 bp->b_offset = offset;
734 ptr = (size_t) mem & PAGE_CACHE_MASK;
735 end = PAGE_CACHE_ALIGN((size_t) mem + len);
737 /* set up first page */
738 bp->b_pages[0] = mem_to_page(mem);
740 ptr += PAGE_CACHE_SIZE;
741 bp->b_page_count = ++i;
743 bp->b_pages[i] = mem_to_page((void *)ptr);
744 bp->b_page_count = ++i;
745 ptr += PAGE_CACHE_SIZE;
749 bp->b_count_desired = bp->b_buffer_length = len;
750 bp->b_flags |= XBF_MAPPED;
758 xfs_buftarg_t *target)
760 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
764 bp = xfs_buf_allocate(0);
765 if (unlikely(bp == NULL))
767 _xfs_buf_initialize(bp, target, 0, len, 0);
769 error = _xfs_buf_get_pages(bp, page_count, 0);
773 for (i = 0; i < page_count; i++) {
774 bp->b_pages[i] = alloc_page(GFP_KERNEL);
778 bp->b_flags |= _XBF_PAGES;
780 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
781 if (unlikely(error)) {
782 printk(KERN_WARNING "%s: failed to map pages\n",
789 XB_TRACE(bp, "no_daddr", len);
794 __free_page(bp->b_pages[i]);
795 _xfs_buf_free_pages(bp);
797 xfs_buf_deallocate(bp);
803 * Increment reference count on buffer, to hold the buffer concurrently
804 * with another thread which may release (free) the buffer asynchronously.
805 * Must hold the buffer already to call this function.
811 atomic_inc(&bp->b_hold);
812 XB_TRACE(bp, "hold", 0);
816 * Releases a hold on the specified buffer. If the
817 * the hold count is 1, calls xfs_buf_free.
823 xfs_bufhash_t *hash = bp->b_hash;
825 XB_TRACE(bp, "rele", bp->b_relse);
827 if (unlikely(!hash)) {
828 ASSERT(!bp->b_relse);
829 if (atomic_dec_and_test(&bp->b_hold))
834 if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
836 atomic_inc(&bp->b_hold);
837 spin_unlock(&hash->bh_lock);
838 (*(bp->b_relse)) (bp);
839 } else if (bp->b_flags & XBF_FS_MANAGED) {
840 spin_unlock(&hash->bh_lock);
842 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
843 list_del_init(&bp->b_hash_list);
844 spin_unlock(&hash->bh_lock);
849 * Catch reference count leaks
851 ASSERT(atomic_read(&bp->b_hold) >= 0);
857 * Mutual exclusion on buffers. Locking model:
859 * Buffers associated with inodes for which buffer locking
860 * is not enabled are not protected by semaphores, and are
861 * assumed to be exclusively owned by the caller. There is a
862 * spinlock in the buffer, used by the caller when concurrent
863 * access is possible.
867 * Locks a buffer object, if it is not already locked.
868 * Note that this in no way locks the underlying pages, so it is only
869 * useful for synchronizing concurrent use of buffer objects, not for
870 * synchronizing independent access to the underlying pages.
878 locked = down_trylock(&bp->b_sema) == 0;
882 XB_TRACE(bp, "cond_lock", (long)locked);
883 return locked ? 0 : -EBUSY;
886 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
891 return atomic_read(&bp->b_sema.count);
896 * Locks a buffer object.
897 * Note that this in no way locks the underlying pages, so it is only
898 * useful for synchronizing concurrent use of buffer objects, not for
899 * synchronizing independent access to the underlying pages.
905 XB_TRACE(bp, "lock", 0);
906 if (atomic_read(&bp->b_io_remaining))
907 blk_run_address_space(bp->b_target->bt_mapping);
910 XB_TRACE(bp, "locked", 0);
914 * Releases the lock on the buffer object.
915 * If the buffer is marked delwri but is not queued, do so before we
916 * unlock the buffer as we need to set flags correctly. We also need to
917 * take a reference for the delwri queue because the unlocker is going to
918 * drop their's and they don't know we just queued it.
924 if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
925 atomic_inc(&bp->b_hold);
926 bp->b_flags |= XBF_ASYNC;
927 xfs_buf_delwri_queue(bp, 0);
932 XB_TRACE(bp, "unlock", 0);
937 * Pinning Buffer Storage in Memory
938 * Ensure that no attempt to force a buffer to disk will succeed.
944 atomic_inc(&bp->b_pin_count);
945 XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
952 if (atomic_dec_and_test(&bp->b_pin_count))
953 wake_up_all(&bp->b_waiters);
954 XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
961 return atomic_read(&bp->b_pin_count);
968 DECLARE_WAITQUEUE (wait, current);
970 if (atomic_read(&bp->b_pin_count) == 0)
973 add_wait_queue(&bp->b_waiters, &wait);
975 set_current_state(TASK_UNINTERRUPTIBLE);
976 if (atomic_read(&bp->b_pin_count) == 0)
978 if (atomic_read(&bp->b_io_remaining))
979 blk_run_address_space(bp->b_target->bt_mapping);
982 remove_wait_queue(&bp->b_waiters, &wait);
983 set_current_state(TASK_RUNNING);
987 * Buffer Utility Routines
992 struct work_struct *work)
995 container_of(work, xfs_buf_t, b_iodone_work);
998 (*(bp->b_iodone))(bp);
999 else if (bp->b_flags & XBF_ASYNC)
1008 bp->b_flags &= ~(XBF_READ | XBF_WRITE);
1009 if (bp->b_error == 0)
1010 bp->b_flags |= XBF_DONE;
1012 XB_TRACE(bp, "iodone", bp->b_iodone);
1014 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1016 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1017 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1019 xfs_buf_iodone_work(&bp->b_iodone_work);
1022 up(&bp->b_iodonesema);
1031 ASSERT(error >= 0 && error <= 0xffff);
1032 bp->b_error = (unsigned short)error;
1033 XB_TRACE(bp, "ioerror", (unsigned long)error);
1037 * Initiate I/O on a buffer, based on the flags supplied.
1038 * The b_iodone routine in the buffer supplied will only be called
1039 * when all of the subsidiary I/O requests, if any, have been completed.
1044 xfs_buf_flags_t flags)
1048 XB_TRACE(bp, "iostart", (unsigned long)flags);
1050 if (flags & XBF_DELWRI) {
1051 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1052 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1053 xfs_buf_delwri_queue(bp, 1);
1057 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1058 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1059 bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1060 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1062 BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1064 /* For writes allow an alternate strategy routine to precede
1065 * the actual I/O request (which may not be issued at all in
1066 * a shutdown situation, for example).
1068 status = (flags & XBF_WRITE) ?
1069 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1071 /* Wait for I/O if we are not an async request.
1072 * Note: async I/O request completion will release the buffer,
1073 * and that can already be done by this point. So using the
1074 * buffer pointer from here on, after async I/O, is invalid.
1076 if (!status && !(flags & XBF_ASYNC))
1077 status = xfs_buf_iowait(bp);
1086 ASSERT(bp->b_flags & (XBF_READ | XBF_WRITE));
1087 if (bp->b_flags & XBF_READ)
1088 return bp->b_locked;
1097 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1099 xfs_buf_ioend(bp, schedule);
1106 unsigned int bytes_done,
1109 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1110 unsigned int blocksize = bp->b_target->bt_bsize;
1111 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1116 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1120 struct page *page = bvec->bv_page;
1122 ASSERT(!PagePrivate(page));
1123 if (unlikely(bp->b_error)) {
1124 if (bp->b_flags & XBF_READ)
1125 ClearPageUptodate(page);
1126 } else if (blocksize >= PAGE_CACHE_SIZE) {
1127 SetPageUptodate(page);
1128 } else if (!PagePrivate(page) &&
1129 (bp->b_flags & _XBF_PAGE_CACHE)) {
1130 set_page_region(page, bvec->bv_offset, bvec->bv_len);
1133 if (--bvec >= bio->bi_io_vec)
1134 prefetchw(&bvec->bv_page->flags);
1136 if (_xfs_buf_iolocked(bp)) {
1139 } while (bvec >= bio->bi_io_vec);
1141 _xfs_buf_ioend(bp, 1);
1150 int i, rw, map_i, total_nr_pages, nr_pages;
1152 int offset = bp->b_offset;
1153 int size = bp->b_count_desired;
1154 sector_t sector = bp->b_bn;
1155 unsigned int blocksize = bp->b_target->bt_bsize;
1156 int locking = _xfs_buf_iolocked(bp);
1158 total_nr_pages = bp->b_page_count;
1161 if (bp->b_flags & XBF_ORDERED) {
1162 ASSERT(!(bp->b_flags & XBF_READ));
1164 } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1165 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1166 bp->b_flags &= ~_XBF_RUN_QUEUES;
1167 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1169 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1170 (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1173 /* Special code path for reading a sub page size buffer in --
1174 * we populate up the whole page, and hence the other metadata
1175 * in the same page. This optimization is only valid when the
1176 * filesystem block size is not smaller than the page size.
1178 if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1179 (bp->b_flags & XBF_READ) && locking &&
1180 (blocksize >= PAGE_CACHE_SIZE)) {
1181 bio = bio_alloc(GFP_NOIO, 1);
1183 bio->bi_bdev = bp->b_target->bt_bdev;
1184 bio->bi_sector = sector - (offset >> BBSHIFT);
1185 bio->bi_end_io = xfs_buf_bio_end_io;
1186 bio->bi_private = bp;
1188 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1191 atomic_inc(&bp->b_io_remaining);
1196 /* Lock down the pages which we need to for the request */
1197 if (locking && (bp->b_flags & XBF_WRITE) && (bp->b_locked == 0)) {
1198 for (i = 0; size; i++) {
1199 int nbytes = PAGE_CACHE_SIZE - offset;
1200 struct page *page = bp->b_pages[i];
1210 offset = bp->b_offset;
1211 size = bp->b_count_desired;
1215 atomic_inc(&bp->b_io_remaining);
1216 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1217 if (nr_pages > total_nr_pages)
1218 nr_pages = total_nr_pages;
1220 bio = bio_alloc(GFP_NOIO, nr_pages);
1221 bio->bi_bdev = bp->b_target->bt_bdev;
1222 bio->bi_sector = sector;
1223 bio->bi_end_io = xfs_buf_bio_end_io;
1224 bio->bi_private = bp;
1226 for (; size && nr_pages; nr_pages--, map_i++) {
1227 int rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1232 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1233 if (rbytes < nbytes)
1237 sector += nbytes >> BBSHIFT;
1243 if (likely(bio->bi_size)) {
1244 submit_bio(rw, bio);
1249 xfs_buf_ioerror(bp, EIO);
1257 XB_TRACE(bp, "iorequest", 0);
1259 if (bp->b_flags & XBF_DELWRI) {
1260 xfs_buf_delwri_queue(bp, 1);
1264 if (bp->b_flags & XBF_WRITE) {
1265 xfs_buf_wait_unpin(bp);
1270 /* Set the count to 1 initially, this will stop an I/O
1271 * completion callout which happens before we have started
1272 * all the I/O from calling xfs_buf_ioend too early.
1274 atomic_set(&bp->b_io_remaining, 1);
1275 _xfs_buf_ioapply(bp);
1276 _xfs_buf_ioend(bp, 0);
1283 * Waits for I/O to complete on the buffer supplied.
1284 * It returns immediately if no I/O is pending.
1285 * It returns the I/O error code, if any, or 0 if there was no error.
1291 XB_TRACE(bp, "iowait", 0);
1292 if (atomic_read(&bp->b_io_remaining))
1293 blk_run_address_space(bp->b_target->bt_mapping);
1294 down(&bp->b_iodonesema);
1295 XB_TRACE(bp, "iowaited", (long)bp->b_error);
1306 if (bp->b_flags & XBF_MAPPED)
1307 return XFS_BUF_PTR(bp) + offset;
1309 offset += bp->b_offset;
1310 page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1311 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1315 * Move data into or out of a buffer.
1319 xfs_buf_t *bp, /* buffer to process */
1320 size_t boff, /* starting buffer offset */
1321 size_t bsize, /* length to copy */
1322 caddr_t data, /* data address */
1323 xfs_buf_rw_t mode) /* read/write/zero flag */
1325 size_t bend, cpoff, csize;
1328 bend = boff + bsize;
1329 while (boff < bend) {
1330 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1331 cpoff = xfs_buf_poff(boff + bp->b_offset);
1332 csize = min_t(size_t,
1333 PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1335 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1339 memset(page_address(page) + cpoff, 0, csize);
1342 memcpy(data, page_address(page) + cpoff, csize);
1345 memcpy(page_address(page) + cpoff, data, csize);
1354 * Handling of buffer targets (buftargs).
1358 * Wait for any bufs with callbacks that have been submitted but
1359 * have not yet returned... walk the hash list for the target.
1366 xfs_bufhash_t *hash;
1369 for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1370 hash = &btp->bt_hash[i];
1372 spin_lock(&hash->bh_lock);
1373 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1374 ASSERT(btp == bp->b_target);
1375 if (!(bp->b_flags & XBF_FS_MANAGED)) {
1376 spin_unlock(&hash->bh_lock);
1378 * Catch superblock reference count leaks
1381 BUG_ON(bp->b_bn == 0);
1386 spin_unlock(&hash->bh_lock);
1391 * Allocate buffer hash table for a given target.
1392 * For devices containing metadata (i.e. not the log/realtime devices)
1393 * we need to allocate a much larger hash table.
1402 btp->bt_hashshift = external ? 3 : 8; /* 8 or 256 buckets */
1403 btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1404 btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1405 sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1406 for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1407 spin_lock_init(&btp->bt_hash[i].bh_lock);
1408 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1416 kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1417 btp->bt_hash = NULL;
1421 * buftarg list for delwrite queue processing
1423 static LIST_HEAD(xfs_buftarg_list);
1424 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1427 xfs_register_buftarg(
1430 spin_lock(&xfs_buftarg_lock);
1431 list_add(&btp->bt_list, &xfs_buftarg_list);
1432 spin_unlock(&xfs_buftarg_lock);
1436 xfs_unregister_buftarg(
1439 spin_lock(&xfs_buftarg_lock);
1440 list_del(&btp->bt_list);
1441 spin_unlock(&xfs_buftarg_lock);
1449 xfs_flush_buftarg(btp, 1);
1451 xfs_blkdev_put(btp->bt_bdev);
1452 xfs_free_bufhash(btp);
1453 iput(btp->bt_mapping->host);
1455 /* Unregister the buftarg first so that we don't get a
1456 * wakeup finding a non-existent task
1458 xfs_unregister_buftarg(btp);
1459 kthread_stop(btp->bt_task);
1461 kmem_free(btp, sizeof(*btp));
1465 xfs_setsize_buftarg_flags(
1467 unsigned int blocksize,
1468 unsigned int sectorsize,
1471 btp->bt_bsize = blocksize;
1472 btp->bt_sshift = ffs(sectorsize) - 1;
1473 btp->bt_smask = sectorsize - 1;
1475 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1477 "XFS: Cannot set_blocksize to %u on device %s\n",
1478 sectorsize, XFS_BUFTARG_NAME(btp));
1483 (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1485 "XFS: %u byte sectors in use on device %s. "
1486 "This is suboptimal; %u or greater is ideal.\n",
1487 sectorsize, XFS_BUFTARG_NAME(btp),
1488 (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1495 * When allocating the initial buffer target we have not yet
1496 * read in the superblock, so don't know what sized sectors
1497 * are being used is at this early stage. Play safe.
1500 xfs_setsize_buftarg_early(
1502 struct block_device *bdev)
1504 return xfs_setsize_buftarg_flags(btp,
1505 PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1509 xfs_setsize_buftarg(
1511 unsigned int blocksize,
1512 unsigned int sectorsize)
1514 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1518 xfs_mapping_buftarg(
1520 struct block_device *bdev)
1522 struct backing_dev_info *bdi;
1523 struct inode *inode;
1524 struct address_space *mapping;
1525 static const struct address_space_operations mapping_aops = {
1526 .sync_page = block_sync_page,
1527 .migratepage = fail_migrate_page,
1530 inode = new_inode(bdev->bd_inode->i_sb);
1533 "XFS: Cannot allocate mapping inode for device %s\n",
1534 XFS_BUFTARG_NAME(btp));
1537 inode->i_mode = S_IFBLK;
1538 inode->i_bdev = bdev;
1539 inode->i_rdev = bdev->bd_dev;
1540 bdi = blk_get_backing_dev_info(bdev);
1542 bdi = &default_backing_dev_info;
1543 mapping = &inode->i_data;
1544 mapping->a_ops = &mapping_aops;
1545 mapping->backing_dev_info = bdi;
1546 mapping_set_gfp_mask(mapping, GFP_NOFS);
1547 btp->bt_mapping = mapping;
1552 xfs_alloc_delwrite_queue(
1557 INIT_LIST_HEAD(&btp->bt_list);
1558 INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1559 spinlock_init(&btp->bt_delwrite_lock, "delwri_lock");
1561 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1562 if (IS_ERR(btp->bt_task)) {
1563 error = PTR_ERR(btp->bt_task);
1566 xfs_register_buftarg(btp);
1573 struct block_device *bdev,
1578 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1580 btp->bt_dev = bdev->bd_dev;
1581 btp->bt_bdev = bdev;
1582 if (xfs_setsize_buftarg_early(btp, bdev))
1584 if (xfs_mapping_buftarg(btp, bdev))
1586 if (xfs_alloc_delwrite_queue(btp))
1588 xfs_alloc_bufhash(btp, external);
1592 kmem_free(btp, sizeof(*btp));
1598 * Delayed write buffer handling
1601 xfs_buf_delwri_queue(
1605 struct list_head *dwq = &bp->b_target->bt_delwrite_queue;
1606 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1608 XB_TRACE(bp, "delwri_q", (long)unlock);
1609 ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1612 /* If already in the queue, dequeue and place at tail */
1613 if (!list_empty(&bp->b_list)) {
1614 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1616 atomic_dec(&bp->b_hold);
1617 list_del(&bp->b_list);
1620 bp->b_flags |= _XBF_DELWRI_Q;
1621 list_add_tail(&bp->b_list, dwq);
1622 bp->b_queuetime = jiffies;
1630 xfs_buf_delwri_dequeue(
1633 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1637 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1638 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1639 list_del_init(&bp->b_list);
1642 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1648 XB_TRACE(bp, "delwri_dq", (long)dequeued);
1652 xfs_buf_runall_queues(
1653 struct workqueue_struct *queue)
1655 flush_workqueue(queue);
1665 spin_lock(&xfs_buftarg_lock);
1666 list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1667 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1669 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1670 wake_up_process(btp->bt_task);
1672 spin_unlock(&xfs_buftarg_lock);
1677 * Move as many buffers as specified to the supplied list
1678 * idicating if we skipped any buffers to prevent deadlocks.
1681 xfs_buf_delwri_split(
1682 xfs_buftarg_t *target,
1683 struct list_head *list,
1687 struct list_head *dwq = &target->bt_delwrite_queue;
1688 spinlock_t *dwlk = &target->bt_delwrite_lock;
1692 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1693 INIT_LIST_HEAD(list);
1695 list_for_each_entry_safe(bp, n, dwq, b_list) {
1696 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1697 ASSERT(bp->b_flags & XBF_DELWRI);
1699 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1701 time_before(jiffies, bp->b_queuetime + age)) {
1706 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1708 bp->b_flags |= XBF_WRITE;
1709 list_move_tail(&bp->b_list, list);
1723 struct list_head tmp;
1724 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1728 current->flags |= PF_MEMALLOC;
1731 if (unlikely(freezing(current))) {
1732 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1735 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1738 schedule_timeout_interruptible(
1739 xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1741 xfs_buf_delwri_split(target, &tmp,
1742 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1745 while (!list_empty(&tmp)) {
1746 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1747 ASSERT(target == bp->b_target);
1749 list_del_init(&bp->b_list);
1750 xfs_buf_iostrategy(bp);
1754 if (as_list_len > 0)
1757 blk_run_address_space(target->bt_mapping);
1759 } while (!kthread_should_stop());
1765 * Go through all incore buffers, and release buffers if they belong to
1766 * the given device. This is used in filesystem error handling to
1767 * preserve the consistency of its metadata.
1771 xfs_buftarg_t *target,
1774 struct list_head tmp;
1778 xfs_buf_runall_queues(xfsdatad_workqueue);
1779 xfs_buf_runall_queues(xfslogd_workqueue);
1781 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1782 pincount = xfs_buf_delwri_split(target, &tmp, 0);
1785 * Dropped the delayed write list lock, now walk the temporary list
1787 list_for_each_entry_safe(bp, n, &tmp, b_list) {
1788 ASSERT(target == bp->b_target);
1790 bp->b_flags &= ~XBF_ASYNC;
1792 list_del_init(&bp->b_list);
1794 xfs_buf_iostrategy(bp);
1798 blk_run_address_space(target->bt_mapping);
1801 * Remaining list items must be flushed before returning
1803 while (!list_empty(&tmp)) {
1804 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1806 list_del_init(&bp->b_list);
1817 #ifdef XFS_BUF_TRACE
1818 xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1821 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1822 KM_ZONE_HWALIGN, NULL);
1824 goto out_free_trace_buf;
1826 xfslogd_workqueue = create_workqueue("xfslogd");
1827 if (!xfslogd_workqueue)
1828 goto out_free_buf_zone;
1830 xfsdatad_workqueue = create_workqueue("xfsdatad");
1831 if (!xfsdatad_workqueue)
1832 goto out_destroy_xfslogd_workqueue;
1834 xfs_buf_shake = kmem_shake_register(xfsbufd_wakeup);
1836 goto out_destroy_xfsdatad_workqueue;
1840 out_destroy_xfsdatad_workqueue:
1841 destroy_workqueue(xfsdatad_workqueue);
1842 out_destroy_xfslogd_workqueue:
1843 destroy_workqueue(xfslogd_workqueue);
1845 kmem_zone_destroy(xfs_buf_zone);
1847 #ifdef XFS_BUF_TRACE
1848 ktrace_free(xfs_buf_trace_buf);
1854 xfs_buf_terminate(void)
1856 kmem_shake_deregister(xfs_buf_shake);
1857 destroy_workqueue(xfsdatad_workqueue);
1858 destroy_workqueue(xfslogd_workqueue);
1859 kmem_zone_destroy(xfs_buf_zone);
1860 #ifdef XFS_BUF_TRACE
1861 ktrace_free(xfs_buf_trace_buf);
1865 #ifdef CONFIG_KDB_MODULES
1867 xfs_get_buftarg_list(void)
1869 return &xfs_buftarg_list;