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 int xfsbufd(void *);
39 STATIC int xfsbufd_wakeup(int, gfp_t);
40 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
41 static struct shrinker xfs_buf_shake = {
42 .shrink = xfsbufd_wakeup,
43 .seeks = DEFAULT_SEEKS,
46 static struct workqueue_struct *xfslogd_workqueue;
47 struct workqueue_struct *xfsdatad_workqueue;
57 ktrace_enter(xfs_buf_trace_buf,
59 (void *)(unsigned long)bp->b_flags,
60 (void *)(unsigned long)bp->b_hold.counter,
61 (void *)(unsigned long)bp->b_sema.count.counter,
64 (void *)(unsigned long)((bp->b_file_offset>>32) & 0xffffffff),
65 (void *)(unsigned long)(bp->b_file_offset & 0xffffffff),
66 (void *)(unsigned long)bp->b_buffer_length,
67 NULL, NULL, NULL, NULL, NULL);
69 ktrace_t *xfs_buf_trace_buf;
70 #define XFS_BUF_TRACE_SIZE 4096
71 #define XB_TRACE(bp, id, data) \
72 xfs_buf_trace(bp, id, (void *)data, (void *)__builtin_return_address(0))
74 #define XB_TRACE(bp, id, data) do { } while (0)
77 #ifdef XFS_BUF_LOCK_TRACKING
78 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
79 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
80 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
82 # define XB_SET_OWNER(bp) do { } while (0)
83 # define XB_CLEAR_OWNER(bp) do { } while (0)
84 # define XB_GET_OWNER(bp) do { } while (0)
87 #define xb_to_gfp(flags) \
88 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
89 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
91 #define xb_to_km(flags) \
92 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
94 #define xfs_buf_allocate(flags) \
95 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
96 #define xfs_buf_deallocate(bp) \
97 kmem_zone_free(xfs_buf_zone, (bp));
100 * Page Region interfaces.
102 * For pages in filesystems where the blocksize is smaller than the
103 * pagesize, we use the page->private field (long) to hold a bitmap
104 * of uptodate regions within the page.
106 * Each such region is "bytes per page / bits per long" bytes long.
108 * NBPPR == number-of-bytes-per-page-region
109 * BTOPR == bytes-to-page-region (rounded up)
110 * BTOPRT == bytes-to-page-region-truncated (rounded down)
112 #if (BITS_PER_LONG == 32)
113 #define PRSHIFT (PAGE_CACHE_SHIFT - 5) /* (32 == 1<<5) */
114 #elif (BITS_PER_LONG == 64)
115 #define PRSHIFT (PAGE_CACHE_SHIFT - 6) /* (64 == 1<<6) */
117 #error BITS_PER_LONG must be 32 or 64
119 #define NBPPR (PAGE_CACHE_SIZE/BITS_PER_LONG)
120 #define BTOPR(b) (((unsigned int)(b) + (NBPPR - 1)) >> PRSHIFT)
121 #define BTOPRT(b) (((unsigned int)(b) >> PRSHIFT))
131 first = BTOPR(offset);
132 final = BTOPRT(offset + length - 1);
133 first = min(first, final);
136 mask <<= BITS_PER_LONG - (final - first);
137 mask >>= BITS_PER_LONG - (final);
139 ASSERT(offset + length <= PAGE_CACHE_SIZE);
140 ASSERT((final - first) < BITS_PER_LONG && (final - first) >= 0);
151 set_page_private(page,
152 page_private(page) | page_region_mask(offset, length));
153 if (page_private(page) == ~0UL)
154 SetPageUptodate(page);
163 unsigned long mask = page_region_mask(offset, length);
165 return (mask && (page_private(page) & mask) == mask);
169 * Mapping of multi-page buffers into contiguous virtual space
172 typedef struct a_list {
177 static a_list_t *as_free_head;
178 static int as_list_len;
179 static DEFINE_SPINLOCK(as_lock);
182 * Try to batch vunmaps because they are costly.
192 * Xen needs to be able to make sure it can get an exclusive
193 * RO mapping of pages it wants to turn into a pagetable. If
194 * a newly allocated page is also still being vmap()ed by xfs,
195 * it will cause pagetable construction to fail. This is a
196 * quick workaround to always eagerly unmap pages so that Xen
203 aentry = kmalloc(sizeof(a_list_t), GFP_NOWAIT);
204 if (likely(aentry)) {
206 aentry->next = as_free_head;
207 aentry->vm_addr = addr;
208 as_free_head = aentry;
210 spin_unlock(&as_lock);
217 purge_addresses(void)
219 a_list_t *aentry, *old;
221 if (as_free_head == NULL)
225 aentry = as_free_head;
228 spin_unlock(&as_lock);
230 while ((old = aentry) != NULL) {
231 vunmap(aentry->vm_addr);
232 aentry = aentry->next;
238 * Internal xfs_buf_t object manipulation
244 xfs_buftarg_t *target,
245 xfs_off_t range_base,
247 xfs_buf_flags_t flags)
250 * We don't want certain flags to appear in b_flags.
252 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
254 memset(bp, 0, sizeof(xfs_buf_t));
255 atomic_set(&bp->b_hold, 1);
256 init_MUTEX_LOCKED(&bp->b_iodonesema);
257 INIT_LIST_HEAD(&bp->b_list);
258 INIT_LIST_HEAD(&bp->b_hash_list);
259 init_MUTEX_LOCKED(&bp->b_sema); /* held, no waiters */
261 bp->b_target = target;
262 bp->b_file_offset = range_base;
264 * Set buffer_length and count_desired to the same value initially.
265 * I/O routines should use count_desired, which will be the same in
266 * most cases but may be reset (e.g. XFS recovery).
268 bp->b_buffer_length = bp->b_count_desired = range_length;
270 bp->b_bn = XFS_BUF_DADDR_NULL;
271 atomic_set(&bp->b_pin_count, 0);
272 init_waitqueue_head(&bp->b_waiters);
274 XFS_STATS_INC(xb_create);
275 XB_TRACE(bp, "initialize", target);
279 * Allocate a page array capable of holding a specified number
280 * of pages, and point the page buf at it.
286 xfs_buf_flags_t flags)
288 /* Make sure that we have a page list */
289 if (bp->b_pages == NULL) {
290 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
291 bp->b_page_count = page_count;
292 if (page_count <= XB_PAGES) {
293 bp->b_pages = bp->b_page_array;
295 bp->b_pages = kmem_alloc(sizeof(struct page *) *
296 page_count, xb_to_km(flags));
297 if (bp->b_pages == NULL)
300 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
306 * Frees b_pages if it was allocated.
312 if (bp->b_pages != bp->b_page_array) {
313 kmem_free(bp->b_pages,
314 bp->b_page_count * sizeof(struct page *));
319 * Releases the specified buffer.
321 * The modification state of any associated pages is left unchanged.
322 * The buffer most not be on any hash - use xfs_buf_rele instead for
323 * hashed and refcounted buffers
329 XB_TRACE(bp, "free", 0);
331 ASSERT(list_empty(&bp->b_hash_list));
333 if (bp->b_flags & (_XBF_PAGE_CACHE|_XBF_PAGES)) {
336 if ((bp->b_flags & XBF_MAPPED) && (bp->b_page_count > 1))
337 free_address(bp->b_addr - bp->b_offset);
339 for (i = 0; i < bp->b_page_count; i++) {
340 struct page *page = bp->b_pages[i];
342 if (bp->b_flags & _XBF_PAGE_CACHE)
343 ASSERT(!PagePrivate(page));
344 page_cache_release(page);
346 _xfs_buf_free_pages(bp);
349 xfs_buf_deallocate(bp);
353 * Finds all pages for buffer in question and builds it's page list.
356 _xfs_buf_lookup_pages(
360 struct address_space *mapping = bp->b_target->bt_mapping;
361 size_t blocksize = bp->b_target->bt_bsize;
362 size_t size = bp->b_count_desired;
363 size_t nbytes, offset;
364 gfp_t gfp_mask = xb_to_gfp(flags);
365 unsigned short page_count, i;
370 end = bp->b_file_offset + bp->b_buffer_length;
371 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
373 error = _xfs_buf_get_pages(bp, page_count, flags);
376 bp->b_flags |= _XBF_PAGE_CACHE;
378 offset = bp->b_offset;
379 first = bp->b_file_offset >> PAGE_CACHE_SHIFT;
381 for (i = 0; i < bp->b_page_count; i++) {
386 page = find_or_create_page(mapping, first + i, gfp_mask);
387 if (unlikely(page == NULL)) {
388 if (flags & XBF_READ_AHEAD) {
389 bp->b_page_count = i;
390 for (i = 0; i < bp->b_page_count; i++)
391 unlock_page(bp->b_pages[i]);
396 * This could deadlock.
398 * But until all the XFS lowlevel code is revamped to
399 * handle buffer allocation failures we can't do much.
401 if (!(++retries % 100))
403 "XFS: possible memory allocation "
404 "deadlock in %s (mode:0x%x)\n",
405 __FUNCTION__, gfp_mask);
407 XFS_STATS_INC(xb_page_retries);
408 xfsbufd_wakeup(0, gfp_mask);
409 congestion_wait(WRITE, HZ/50);
414 XFS_STATS_INC(xb_page_found);
416 nbytes = min_t(size_t, size, PAGE_CACHE_SIZE - offset);
419 ASSERT(!PagePrivate(page));
420 if (!PageUptodate(page)) {
422 if (blocksize < PAGE_CACHE_SIZE && !PagePrivate(page)) {
423 if (test_page_region(page, offset, nbytes))
428 bp->b_pages[i] = page;
432 if (page_count == bp->b_page_count)
433 bp->b_flags |= XBF_DONE;
435 XB_TRACE(bp, "lookup_pages", (long)page_count);
440 * Map buffer into kernel address-space if nessecary.
447 /* A single page buffer is always mappable */
448 if (bp->b_page_count == 1) {
449 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
450 bp->b_flags |= XBF_MAPPED;
451 } else if (flags & XBF_MAPPED) {
452 if (as_list_len > 64)
454 bp->b_addr = vmap(bp->b_pages, bp->b_page_count,
455 VM_MAP, PAGE_KERNEL);
456 if (unlikely(bp->b_addr == NULL))
458 bp->b_addr += bp->b_offset;
459 bp->b_flags |= XBF_MAPPED;
466 * Finding and Reading Buffers
470 * Look up, and creates if absent, a lockable buffer for
471 * a given range of an inode. The buffer is returned
472 * locked. If other overlapping buffers exist, they are
473 * released before the new buffer is created and locked,
474 * which may imply that this call will block until those buffers
475 * are unlocked. No I/O is implied by this call.
479 xfs_buftarg_t *btp, /* block device target */
480 xfs_off_t ioff, /* starting offset of range */
481 size_t isize, /* length of range */
482 xfs_buf_flags_t flags,
485 xfs_off_t range_base;
490 range_base = (ioff << BBSHIFT);
491 range_length = (isize << BBSHIFT);
493 /* Check for IOs smaller than the sector size / not sector aligned */
494 ASSERT(!(range_length < (1 << btp->bt_sshift)));
495 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
497 hash = &btp->bt_hash[hash_long((unsigned long)ioff, btp->bt_hashshift)];
499 spin_lock(&hash->bh_lock);
501 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
502 ASSERT(btp == bp->b_target);
503 if (bp->b_file_offset == range_base &&
504 bp->b_buffer_length == range_length) {
506 * If we look at something, bring it to the
507 * front of the list for next time.
509 atomic_inc(&bp->b_hold);
510 list_move(&bp->b_hash_list, &hash->bh_list);
517 _xfs_buf_initialize(new_bp, btp, range_base,
518 range_length, flags);
519 new_bp->b_hash = hash;
520 list_add(&new_bp->b_hash_list, &hash->bh_list);
522 XFS_STATS_INC(xb_miss_locked);
525 spin_unlock(&hash->bh_lock);
529 spin_unlock(&hash->bh_lock);
531 /* Attempt to get the semaphore without sleeping,
532 * if this does not work then we need to drop the
533 * spinlock and do a hard attempt on the semaphore.
535 if (down_trylock(&bp->b_sema)) {
536 if (!(flags & XBF_TRYLOCK)) {
537 /* wait for buffer ownership */
538 XB_TRACE(bp, "get_lock", 0);
540 XFS_STATS_INC(xb_get_locked_waited);
542 /* We asked for a trylock and failed, no need
543 * to look at file offset and length here, we
544 * know that this buffer at least overlaps our
545 * buffer and is locked, therefore our buffer
546 * either does not exist, or is this buffer.
549 XFS_STATS_INC(xb_busy_locked);
557 if (bp->b_flags & XBF_STALE) {
558 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
559 bp->b_flags &= XBF_MAPPED;
561 XB_TRACE(bp, "got_lock", 0);
562 XFS_STATS_INC(xb_get_locked);
567 * Assembles a buffer covering the specified range.
568 * Storage in memory for all portions of the buffer will be allocated,
569 * although backing storage may not be.
573 xfs_buftarg_t *target,/* target for buffer */
574 xfs_off_t ioff, /* starting offset of range */
575 size_t isize, /* length of range */
576 xfs_buf_flags_t flags)
578 xfs_buf_t *bp, *new_bp;
581 new_bp = xfs_buf_allocate(flags);
582 if (unlikely(!new_bp))
585 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
587 error = _xfs_buf_lookup_pages(bp, flags);
591 xfs_buf_deallocate(new_bp);
592 if (unlikely(bp == NULL))
596 for (i = 0; i < bp->b_page_count; i++)
597 mark_page_accessed(bp->b_pages[i]);
599 if (!(bp->b_flags & XBF_MAPPED)) {
600 error = _xfs_buf_map_pages(bp, flags);
601 if (unlikely(error)) {
602 printk(KERN_WARNING "%s: failed to map pages\n",
608 XFS_STATS_INC(xb_get);
611 * Always fill in the block number now, the mapped cases can do
612 * their own overlay of this later.
615 bp->b_count_desired = bp->b_buffer_length;
617 XB_TRACE(bp, "get", (unsigned long)flags);
621 if (flags & (XBF_LOCK | XBF_TRYLOCK))
629 xfs_buftarg_t *target,
632 xfs_buf_flags_t flags)
638 bp = xfs_buf_get_flags(target, ioff, isize, flags);
640 if (!XFS_BUF_ISDONE(bp)) {
641 XB_TRACE(bp, "read", (unsigned long)flags);
642 XFS_STATS_INC(xb_get_read);
643 xfs_buf_iostart(bp, flags);
644 } else if (flags & XBF_ASYNC) {
645 XB_TRACE(bp, "read_async", (unsigned long)flags);
647 * Read ahead call which is already satisfied,
652 XB_TRACE(bp, "read_done", (unsigned long)flags);
653 /* We do not want read in the flags */
654 bp->b_flags &= ~XBF_READ;
661 if (flags & (XBF_LOCK | XBF_TRYLOCK))
668 * If we are not low on memory then do the readahead in a deadlock
673 xfs_buftarg_t *target,
676 xfs_buf_flags_t flags)
678 struct backing_dev_info *bdi;
680 bdi = target->bt_mapping->backing_dev_info;
681 if (bdi_read_congested(bdi))
684 flags |= (XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD);
685 xfs_buf_read_flags(target, ioff, isize, flags);
691 xfs_buftarg_t *target)
695 bp = xfs_buf_allocate(0);
697 _xfs_buf_initialize(bp, target, 0, len, 0);
701 static inline struct page *
705 if (((unsigned long)addr < VMALLOC_START) ||
706 ((unsigned long)addr >= VMALLOC_END)) {
707 return virt_to_page(addr);
709 return vmalloc_to_page(addr);
714 xfs_buf_associate_memory(
721 unsigned long pageaddr;
722 unsigned long offset;
726 pageaddr = (unsigned long)mem & PAGE_CACHE_MASK;
727 offset = (unsigned long)mem - pageaddr;
728 buflen = PAGE_CACHE_ALIGN(len + offset);
729 page_count = buflen >> PAGE_CACHE_SHIFT;
731 /* Free any previous set of page pointers */
733 _xfs_buf_free_pages(bp);
738 rval = _xfs_buf_get_pages(bp, page_count, 0);
742 bp->b_offset = offset;
744 for (i = 0; i < bp->b_page_count; i++) {
745 bp->b_pages[i] = mem_to_page((void *)pageaddr);
746 pageaddr += PAGE_CACHE_SIZE;
749 bp->b_count_desired = len;
750 bp->b_buffer_length = buflen;
751 bp->b_flags |= XBF_MAPPED;
759 xfs_buftarg_t *target)
761 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
765 bp = xfs_buf_allocate(0);
766 if (unlikely(bp == NULL))
768 _xfs_buf_initialize(bp, target, 0, len, 0);
770 error = _xfs_buf_get_pages(bp, page_count, 0);
774 for (i = 0; i < page_count; i++) {
775 bp->b_pages[i] = alloc_page(GFP_KERNEL);
779 bp->b_flags |= _XBF_PAGES;
781 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
782 if (unlikely(error)) {
783 printk(KERN_WARNING "%s: failed to map pages\n",
790 XB_TRACE(bp, "no_daddr", len);
795 __free_page(bp->b_pages[i]);
796 _xfs_buf_free_pages(bp);
798 xfs_buf_deallocate(bp);
804 * Increment reference count on buffer, to hold the buffer concurrently
805 * with another thread which may release (free) the buffer asynchronously.
806 * Must hold the buffer already to call this function.
812 atomic_inc(&bp->b_hold);
813 XB_TRACE(bp, "hold", 0);
817 * Releases a hold on the specified buffer. If the
818 * the hold count is 1, calls xfs_buf_free.
824 xfs_bufhash_t *hash = bp->b_hash;
826 XB_TRACE(bp, "rele", bp->b_relse);
828 if (unlikely(!hash)) {
829 ASSERT(!bp->b_relse);
830 if (atomic_dec_and_test(&bp->b_hold))
835 if (atomic_dec_and_lock(&bp->b_hold, &hash->bh_lock)) {
837 atomic_inc(&bp->b_hold);
838 spin_unlock(&hash->bh_lock);
839 (*(bp->b_relse)) (bp);
840 } else if (bp->b_flags & XBF_FS_MANAGED) {
841 spin_unlock(&hash->bh_lock);
843 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
844 list_del_init(&bp->b_hash_list);
845 spin_unlock(&hash->bh_lock);
850 * Catch reference count leaks
852 ASSERT(atomic_read(&bp->b_hold) >= 0);
858 * Mutual exclusion on buffers. Locking model:
860 * Buffers associated with inodes for which buffer locking
861 * is not enabled are not protected by semaphores, and are
862 * assumed to be exclusively owned by the caller. There is a
863 * spinlock in the buffer, used by the caller when concurrent
864 * access is possible.
868 * Locks a buffer object, if it is not already locked.
869 * Note that this in no way locks the underlying pages, so it is only
870 * useful for synchronizing concurrent use of buffer objects, not for
871 * synchronizing independent access to the underlying pages.
879 locked = down_trylock(&bp->b_sema) == 0;
883 XB_TRACE(bp, "cond_lock", (long)locked);
884 return locked ? 0 : -EBUSY;
887 #if defined(DEBUG) || defined(XFS_BLI_TRACE)
892 return atomic_read(&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 XB_TRACE(bp, "lock", 0);
907 if (atomic_read(&bp->b_io_remaining))
908 blk_run_address_space(bp->b_target->bt_mapping);
911 XB_TRACE(bp, "locked", 0);
915 * Releases the lock on the buffer object.
916 * If the buffer is marked delwri but is not queued, do so before we
917 * unlock the buffer as we need to set flags correctly. We also need to
918 * take a reference for the delwri queue because the unlocker is going to
919 * drop their's and they don't know we just queued it.
925 if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
926 atomic_inc(&bp->b_hold);
927 bp->b_flags |= XBF_ASYNC;
928 xfs_buf_delwri_queue(bp, 0);
933 XB_TRACE(bp, "unlock", 0);
938 * Pinning Buffer Storage in Memory
939 * Ensure that no attempt to force a buffer to disk will succeed.
945 atomic_inc(&bp->b_pin_count);
946 XB_TRACE(bp, "pin", (long)bp->b_pin_count.counter);
953 if (atomic_dec_and_test(&bp->b_pin_count))
954 wake_up_all(&bp->b_waiters);
955 XB_TRACE(bp, "unpin", (long)bp->b_pin_count.counter);
962 return atomic_read(&bp->b_pin_count);
969 DECLARE_WAITQUEUE (wait, current);
971 if (atomic_read(&bp->b_pin_count) == 0)
974 add_wait_queue(&bp->b_waiters, &wait);
976 set_current_state(TASK_UNINTERRUPTIBLE);
977 if (atomic_read(&bp->b_pin_count) == 0)
979 if (atomic_read(&bp->b_io_remaining))
980 blk_run_address_space(bp->b_target->bt_mapping);
983 remove_wait_queue(&bp->b_waiters, &wait);
984 set_current_state(TASK_RUNNING);
988 * Buffer Utility Routines
993 struct work_struct *work)
996 container_of(work, xfs_buf_t, b_iodone_work);
999 * We can get an EOPNOTSUPP to ordered writes. Here we clear the
1000 * ordered flag and reissue them. Because we can't tell the higher
1001 * layers directly that they should not issue ordered I/O anymore, they
1002 * need to check if the ordered flag was cleared during I/O completion.
1004 if ((bp->b_error == EOPNOTSUPP) &&
1005 (bp->b_flags & (XBF_ORDERED|XBF_ASYNC)) == (XBF_ORDERED|XBF_ASYNC)) {
1006 XB_TRACE(bp, "ordered_retry", bp->b_iodone);
1007 bp->b_flags &= ~XBF_ORDERED;
1008 xfs_buf_iorequest(bp);
1009 } else if (bp->b_iodone)
1010 (*(bp->b_iodone))(bp);
1011 else if (bp->b_flags & XBF_ASYNC)
1020 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1021 if (bp->b_error == 0)
1022 bp->b_flags |= XBF_DONE;
1024 XB_TRACE(bp, "iodone", bp->b_iodone);
1026 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1028 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1029 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1031 xfs_buf_iodone_work(&bp->b_iodone_work);
1034 up(&bp->b_iodonesema);
1043 ASSERT(error >= 0 && error <= 0xffff);
1044 bp->b_error = (unsigned short)error;
1045 XB_TRACE(bp, "ioerror", (unsigned long)error);
1049 * Initiate I/O on a buffer, based on the flags supplied.
1050 * The b_iodone routine in the buffer supplied will only be called
1051 * when all of the subsidiary I/O requests, if any, have been completed.
1056 xfs_buf_flags_t flags)
1060 XB_TRACE(bp, "iostart", (unsigned long)flags);
1062 if (flags & XBF_DELWRI) {
1063 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC);
1064 bp->b_flags |= flags & (XBF_DELWRI | XBF_ASYNC);
1065 xfs_buf_delwri_queue(bp, 1);
1069 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
1070 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1071 bp->b_flags |= flags & (XBF_READ | XBF_WRITE | XBF_ASYNC | \
1072 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
1074 BUG_ON(bp->b_bn == XFS_BUF_DADDR_NULL);
1076 /* For writes allow an alternate strategy routine to precede
1077 * the actual I/O request (which may not be issued at all in
1078 * a shutdown situation, for example).
1080 status = (flags & XBF_WRITE) ?
1081 xfs_buf_iostrategy(bp) : xfs_buf_iorequest(bp);
1083 /* Wait for I/O if we are not an async request.
1084 * Note: async I/O request completion will release the buffer,
1085 * and that can already be done by this point. So using the
1086 * buffer pointer from here on, after async I/O, is invalid.
1088 if (!status && !(flags & XBF_ASYNC))
1089 status = xfs_buf_iowait(bp);
1099 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1100 xfs_buf_ioend(bp, schedule);
1108 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1109 unsigned int blocksize = bp->b_target->bt_bsize;
1110 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1112 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1116 struct page *page = bvec->bv_page;
1118 ASSERT(!PagePrivate(page));
1119 if (unlikely(bp->b_error)) {
1120 if (bp->b_flags & XBF_READ)
1121 ClearPageUptodate(page);
1122 } else if (blocksize >= PAGE_CACHE_SIZE) {
1123 SetPageUptodate(page);
1124 } else if (!PagePrivate(page) &&
1125 (bp->b_flags & _XBF_PAGE_CACHE)) {
1126 set_page_region(page, bvec->bv_offset, bvec->bv_len);
1129 if (--bvec >= bio->bi_io_vec)
1130 prefetchw(&bvec->bv_page->flags);
1131 } while (bvec >= bio->bi_io_vec);
1133 _xfs_buf_ioend(bp, 1);
1141 int rw, map_i, total_nr_pages, nr_pages;
1143 int offset = bp->b_offset;
1144 int size = bp->b_count_desired;
1145 sector_t sector = bp->b_bn;
1146 unsigned int blocksize = bp->b_target->bt_bsize;
1148 total_nr_pages = bp->b_page_count;
1151 if (bp->b_flags & XBF_ORDERED) {
1152 ASSERT(!(bp->b_flags & XBF_READ));
1154 } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1155 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1156 bp->b_flags &= ~_XBF_RUN_QUEUES;
1157 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1159 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1160 (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1163 /* Special code path for reading a sub page size buffer in --
1164 * we populate up the whole page, and hence the other metadata
1165 * in the same page. This optimization is only valid when the
1166 * filesystem block size is not smaller than the page size.
1168 if ((bp->b_buffer_length < PAGE_CACHE_SIZE) &&
1169 (bp->b_flags & XBF_READ) &&
1170 (blocksize >= PAGE_CACHE_SIZE)) {
1171 bio = bio_alloc(GFP_NOIO, 1);
1173 bio->bi_bdev = bp->b_target->bt_bdev;
1174 bio->bi_sector = sector - (offset >> BBSHIFT);
1175 bio->bi_end_io = xfs_buf_bio_end_io;
1176 bio->bi_private = bp;
1178 bio_add_page(bio, bp->b_pages[0], PAGE_CACHE_SIZE, 0);
1181 atomic_inc(&bp->b_io_remaining);
1187 atomic_inc(&bp->b_io_remaining);
1188 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1189 if (nr_pages > total_nr_pages)
1190 nr_pages = total_nr_pages;
1192 bio = bio_alloc(GFP_NOIO, nr_pages);
1193 bio->bi_bdev = bp->b_target->bt_bdev;
1194 bio->bi_sector = sector;
1195 bio->bi_end_io = xfs_buf_bio_end_io;
1196 bio->bi_private = bp;
1198 for (; size && nr_pages; nr_pages--, map_i++) {
1199 int rbytes, nbytes = PAGE_CACHE_SIZE - offset;
1204 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1205 if (rbytes < nbytes)
1209 sector += nbytes >> BBSHIFT;
1215 if (likely(bio->bi_size)) {
1216 submit_bio(rw, bio);
1221 xfs_buf_ioerror(bp, EIO);
1229 XB_TRACE(bp, "iorequest", 0);
1231 if (bp->b_flags & XBF_DELWRI) {
1232 xfs_buf_delwri_queue(bp, 1);
1236 if (bp->b_flags & XBF_WRITE) {
1237 xfs_buf_wait_unpin(bp);
1242 /* Set the count to 1 initially, this will stop an I/O
1243 * completion callout which happens before we have started
1244 * all the I/O from calling xfs_buf_ioend too early.
1246 atomic_set(&bp->b_io_remaining, 1);
1247 _xfs_buf_ioapply(bp);
1248 _xfs_buf_ioend(bp, 0);
1255 * Waits for I/O to complete on the buffer supplied.
1256 * It returns immediately if no I/O is pending.
1257 * It returns the I/O error code, if any, or 0 if there was no error.
1263 XB_TRACE(bp, "iowait", 0);
1264 if (atomic_read(&bp->b_io_remaining))
1265 blk_run_address_space(bp->b_target->bt_mapping);
1266 down(&bp->b_iodonesema);
1267 XB_TRACE(bp, "iowaited", (long)bp->b_error);
1278 if (bp->b_flags & XBF_MAPPED)
1279 return XFS_BUF_PTR(bp) + offset;
1281 offset += bp->b_offset;
1282 page = bp->b_pages[offset >> PAGE_CACHE_SHIFT];
1283 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_CACHE_SIZE-1));
1287 * Move data into or out of a buffer.
1291 xfs_buf_t *bp, /* buffer to process */
1292 size_t boff, /* starting buffer offset */
1293 size_t bsize, /* length to copy */
1294 caddr_t data, /* data address */
1295 xfs_buf_rw_t mode) /* read/write/zero flag */
1297 size_t bend, cpoff, csize;
1300 bend = boff + bsize;
1301 while (boff < bend) {
1302 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1303 cpoff = xfs_buf_poff(boff + bp->b_offset);
1304 csize = min_t(size_t,
1305 PAGE_CACHE_SIZE-cpoff, bp->b_count_desired-boff);
1307 ASSERT(((csize + cpoff) <= PAGE_CACHE_SIZE));
1311 memset(page_address(page) + cpoff, 0, csize);
1314 memcpy(data, page_address(page) + cpoff, csize);
1317 memcpy(page_address(page) + cpoff, data, csize);
1326 * Handling of buffer targets (buftargs).
1330 * Wait for any bufs with callbacks that have been submitted but
1331 * have not yet returned... walk the hash list for the target.
1338 xfs_bufhash_t *hash;
1341 for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1342 hash = &btp->bt_hash[i];
1344 spin_lock(&hash->bh_lock);
1345 list_for_each_entry_safe(bp, n, &hash->bh_list, b_hash_list) {
1346 ASSERT(btp == bp->b_target);
1347 if (!(bp->b_flags & XBF_FS_MANAGED)) {
1348 spin_unlock(&hash->bh_lock);
1350 * Catch superblock reference count leaks
1353 BUG_ON(bp->b_bn == 0);
1358 spin_unlock(&hash->bh_lock);
1363 * Allocate buffer hash table for a given target.
1364 * For devices containing metadata (i.e. not the log/realtime devices)
1365 * we need to allocate a much larger hash table.
1374 btp->bt_hashshift = external ? 3 : 8; /* 8 or 256 buckets */
1375 btp->bt_hashmask = (1 << btp->bt_hashshift) - 1;
1376 btp->bt_hash = kmem_zalloc((1 << btp->bt_hashshift) *
1377 sizeof(xfs_bufhash_t), KM_SLEEP | KM_LARGE);
1378 for (i = 0; i < (1 << btp->bt_hashshift); i++) {
1379 spin_lock_init(&btp->bt_hash[i].bh_lock);
1380 INIT_LIST_HEAD(&btp->bt_hash[i].bh_list);
1388 kmem_free(btp->bt_hash, (1<<btp->bt_hashshift) * sizeof(xfs_bufhash_t));
1389 btp->bt_hash = NULL;
1393 * buftarg list for delwrite queue processing
1395 static LIST_HEAD(xfs_buftarg_list);
1396 static DEFINE_SPINLOCK(xfs_buftarg_lock);
1399 xfs_register_buftarg(
1402 spin_lock(&xfs_buftarg_lock);
1403 list_add(&btp->bt_list, &xfs_buftarg_list);
1404 spin_unlock(&xfs_buftarg_lock);
1408 xfs_unregister_buftarg(
1411 spin_lock(&xfs_buftarg_lock);
1412 list_del(&btp->bt_list);
1413 spin_unlock(&xfs_buftarg_lock);
1421 xfs_flush_buftarg(btp, 1);
1422 xfs_blkdev_issue_flush(btp);
1424 xfs_blkdev_put(btp->bt_bdev);
1425 xfs_free_bufhash(btp);
1426 iput(btp->bt_mapping->host);
1428 /* Unregister the buftarg first so that we don't get a
1429 * wakeup finding a non-existent task
1431 xfs_unregister_buftarg(btp);
1432 kthread_stop(btp->bt_task);
1434 kmem_free(btp, sizeof(*btp));
1438 xfs_setsize_buftarg_flags(
1440 unsigned int blocksize,
1441 unsigned int sectorsize,
1444 btp->bt_bsize = blocksize;
1445 btp->bt_sshift = ffs(sectorsize) - 1;
1446 btp->bt_smask = sectorsize - 1;
1448 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1450 "XFS: Cannot set_blocksize to %u on device %s\n",
1451 sectorsize, XFS_BUFTARG_NAME(btp));
1456 (PAGE_CACHE_SIZE / BITS_PER_LONG) > sectorsize) {
1458 "XFS: %u byte sectors in use on device %s. "
1459 "This is suboptimal; %u or greater is ideal.\n",
1460 sectorsize, XFS_BUFTARG_NAME(btp),
1461 (unsigned int)PAGE_CACHE_SIZE / BITS_PER_LONG);
1468 * When allocating the initial buffer target we have not yet
1469 * read in the superblock, so don't know what sized sectors
1470 * are being used is at this early stage. Play safe.
1473 xfs_setsize_buftarg_early(
1475 struct block_device *bdev)
1477 return xfs_setsize_buftarg_flags(btp,
1478 PAGE_CACHE_SIZE, bdev_hardsect_size(bdev), 0);
1482 xfs_setsize_buftarg(
1484 unsigned int blocksize,
1485 unsigned int sectorsize)
1487 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1491 xfs_mapping_buftarg(
1493 struct block_device *bdev)
1495 struct backing_dev_info *bdi;
1496 struct inode *inode;
1497 struct address_space *mapping;
1498 static const struct address_space_operations mapping_aops = {
1499 .sync_page = block_sync_page,
1500 .migratepage = fail_migrate_page,
1503 inode = new_inode(bdev->bd_inode->i_sb);
1506 "XFS: Cannot allocate mapping inode for device %s\n",
1507 XFS_BUFTARG_NAME(btp));
1510 inode->i_mode = S_IFBLK;
1511 inode->i_bdev = bdev;
1512 inode->i_rdev = bdev->bd_dev;
1513 bdi = blk_get_backing_dev_info(bdev);
1515 bdi = &default_backing_dev_info;
1516 mapping = &inode->i_data;
1517 mapping->a_ops = &mapping_aops;
1518 mapping->backing_dev_info = bdi;
1519 mapping_set_gfp_mask(mapping, GFP_NOFS);
1520 btp->bt_mapping = mapping;
1525 xfs_alloc_delwrite_queue(
1530 INIT_LIST_HEAD(&btp->bt_list);
1531 INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1532 spin_lock_init(&btp->bt_delwrite_lock);
1534 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd");
1535 if (IS_ERR(btp->bt_task)) {
1536 error = PTR_ERR(btp->bt_task);
1539 xfs_register_buftarg(btp);
1546 struct block_device *bdev,
1551 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1553 btp->bt_dev = bdev->bd_dev;
1554 btp->bt_bdev = bdev;
1555 if (xfs_setsize_buftarg_early(btp, bdev))
1557 if (xfs_mapping_buftarg(btp, bdev))
1559 if (xfs_alloc_delwrite_queue(btp))
1561 xfs_alloc_bufhash(btp, external);
1565 kmem_free(btp, sizeof(*btp));
1571 * Delayed write buffer handling
1574 xfs_buf_delwri_queue(
1578 struct list_head *dwq = &bp->b_target->bt_delwrite_queue;
1579 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1581 XB_TRACE(bp, "delwri_q", (long)unlock);
1582 ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1585 /* If already in the queue, dequeue and place at tail */
1586 if (!list_empty(&bp->b_list)) {
1587 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1589 atomic_dec(&bp->b_hold);
1590 list_del(&bp->b_list);
1593 bp->b_flags |= _XBF_DELWRI_Q;
1594 list_add_tail(&bp->b_list, dwq);
1595 bp->b_queuetime = jiffies;
1603 xfs_buf_delwri_dequeue(
1606 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1610 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1611 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1612 list_del_init(&bp->b_list);
1615 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1621 XB_TRACE(bp, "delwri_dq", (long)dequeued);
1625 xfs_buf_runall_queues(
1626 struct workqueue_struct *queue)
1628 flush_workqueue(queue);
1638 spin_lock(&xfs_buftarg_lock);
1639 list_for_each_entry(btp, &xfs_buftarg_list, bt_list) {
1640 if (test_bit(XBT_FORCE_SLEEP, &btp->bt_flags))
1642 set_bit(XBT_FORCE_FLUSH, &btp->bt_flags);
1643 wake_up_process(btp->bt_task);
1645 spin_unlock(&xfs_buftarg_lock);
1650 * Move as many buffers as specified to the supplied list
1651 * idicating if we skipped any buffers to prevent deadlocks.
1654 xfs_buf_delwri_split(
1655 xfs_buftarg_t *target,
1656 struct list_head *list,
1660 struct list_head *dwq = &target->bt_delwrite_queue;
1661 spinlock_t *dwlk = &target->bt_delwrite_lock;
1665 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1666 INIT_LIST_HEAD(list);
1668 list_for_each_entry_safe(bp, n, dwq, b_list) {
1669 XB_TRACE(bp, "walkq1", (long)xfs_buf_ispin(bp));
1670 ASSERT(bp->b_flags & XBF_DELWRI);
1672 if (!xfs_buf_ispin(bp) && !xfs_buf_cond_lock(bp)) {
1674 time_before(jiffies, bp->b_queuetime + age)) {
1679 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1681 bp->b_flags |= XBF_WRITE;
1682 list_move_tail(&bp->b_list, list);
1696 struct list_head tmp;
1697 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1701 current->flags |= PF_MEMALLOC;
1706 if (unlikely(freezing(current))) {
1707 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1710 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1713 schedule_timeout_interruptible(
1714 xfs_buf_timer_centisecs * msecs_to_jiffies(10));
1716 xfs_buf_delwri_split(target, &tmp,
1717 xfs_buf_age_centisecs * msecs_to_jiffies(10));
1720 while (!list_empty(&tmp)) {
1721 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1722 ASSERT(target == bp->b_target);
1724 list_del_init(&bp->b_list);
1725 xfs_buf_iostrategy(bp);
1729 if (as_list_len > 0)
1732 blk_run_address_space(target->bt_mapping);
1734 } while (!kthread_should_stop());
1740 * Go through all incore buffers, and release buffers if they belong to
1741 * the given device. This is used in filesystem error handling to
1742 * preserve the consistency of its metadata.
1746 xfs_buftarg_t *target,
1749 struct list_head tmp;
1753 xfs_buf_runall_queues(xfsdatad_workqueue);
1754 xfs_buf_runall_queues(xfslogd_workqueue);
1756 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1757 pincount = xfs_buf_delwri_split(target, &tmp, 0);
1760 * Dropped the delayed write list lock, now walk the temporary list
1762 list_for_each_entry_safe(bp, n, &tmp, b_list) {
1763 ASSERT(target == bp->b_target);
1765 bp->b_flags &= ~XBF_ASYNC;
1767 list_del_init(&bp->b_list);
1769 xfs_buf_iostrategy(bp);
1773 blk_run_address_space(target->bt_mapping);
1776 * Remaining list items must be flushed before returning
1778 while (!list_empty(&tmp)) {
1779 bp = list_entry(tmp.next, xfs_buf_t, b_list);
1781 list_del_init(&bp->b_list);
1792 #ifdef XFS_BUF_TRACE
1793 xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
1796 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1797 KM_ZONE_HWALIGN, NULL);
1799 goto out_free_trace_buf;
1801 xfslogd_workqueue = create_workqueue("xfslogd");
1802 if (!xfslogd_workqueue)
1803 goto out_free_buf_zone;
1805 xfsdatad_workqueue = create_workqueue("xfsdatad");
1806 if (!xfsdatad_workqueue)
1807 goto out_destroy_xfslogd_workqueue;
1809 register_shrinker(&xfs_buf_shake);
1812 out_destroy_xfslogd_workqueue:
1813 destroy_workqueue(xfslogd_workqueue);
1815 kmem_zone_destroy(xfs_buf_zone);
1817 #ifdef XFS_BUF_TRACE
1818 ktrace_free(xfs_buf_trace_buf);
1824 xfs_buf_terminate(void)
1826 unregister_shrinker(&xfs_buf_shake);
1827 destroy_workqueue(xfsdatad_workqueue);
1828 destroy_workqueue(xfslogd_workqueue);
1829 kmem_zone_destroy(xfs_buf_zone);
1830 #ifdef XFS_BUF_TRACE
1831 ktrace_free(xfs_buf_trace_buf);
1835 #ifdef CONFIG_KDB_MODULES
1837 xfs_get_buftarg_list(void)
1839 return &xfs_buftarg_list;