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/gfp.h>
22 #include <linux/pagemap.h>
23 #include <linux/init.h>
24 #include <linux/vmalloc.h>
25 #include <linux/bio.h>
26 #include <linux/sysctl.h>
27 #include <linux/proc_fs.h>
28 #include <linux/workqueue.h>
29 #include <linux/percpu.h>
30 #include <linux/blkdev.h>
31 #include <linux/hash.h>
32 #include <linux/kthread.h>
33 #include <linux/migrate.h>
34 #include <linux/backing-dev.h>
35 #include <linux/freezer.h>
41 #include "xfs_mount.h"
42 #include "xfs_trace.h"
44 static kmem_zone_t *xfs_buf_zone;
45 STATIC int xfsbufd(void *);
47 static struct workqueue_struct *xfslogd_workqueue;
48 struct workqueue_struct *xfsdatad_workqueue;
49 struct workqueue_struct *xfsconvertd_workqueue;
51 #ifdef XFS_BUF_LOCK_TRACKING
52 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
53 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
54 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
56 # define XB_SET_OWNER(bp) do { } while (0)
57 # define XB_CLEAR_OWNER(bp) do { } while (0)
58 # define XB_GET_OWNER(bp) do { } while (0)
61 #define xb_to_gfp(flags) \
62 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
63 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
65 #define xb_to_km(flags) \
66 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
68 #define xfs_buf_allocate(flags) \
69 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
70 #define xfs_buf_deallocate(bp) \
71 kmem_zone_free(xfs_buf_zone, (bp));
78 * Return true if the buffer is vmapped.
80 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
81 * code is clever enough to know it doesn't have to map a single page,
82 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
84 return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
91 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
95 * xfs_buf_lru_add - add a buffer to the LRU.
97 * The LRU takes a new reference to the buffer so that it will only be freed
98 * once the shrinker takes the buffer off the LRU.
104 struct xfs_buftarg *btp = bp->b_target;
106 spin_lock(&btp->bt_lru_lock);
107 if (list_empty(&bp->b_lru)) {
108 atomic_inc(&bp->b_hold);
109 list_add_tail(&bp->b_lru, &btp->bt_lru);
112 spin_unlock(&btp->bt_lru_lock);
116 * xfs_buf_lru_del - remove a buffer from the LRU
118 * The unlocked check is safe here because it only occurs when there are not
119 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
120 * to optimise the shrinker removing the buffer from the LRU and calling
121 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
128 struct xfs_buftarg *btp = bp->b_target;
130 if (list_empty(&bp->b_lru))
133 spin_lock(&btp->bt_lru_lock);
134 if (!list_empty(&bp->b_lru)) {
135 list_del_init(&bp->b_lru);
138 spin_unlock(&btp->bt_lru_lock);
142 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
143 * b_lru_ref count so that the buffer is freed immediately when the buffer
144 * reference count falls to zero. If the buffer is already on the LRU, we need
145 * to remove the reference that LRU holds on the buffer.
147 * This prevents build-up of stale buffers on the LRU.
153 bp->b_flags |= XBF_STALE;
154 atomic_set(&(bp)->b_lru_ref, 0);
155 if (!list_empty(&bp->b_lru)) {
156 struct xfs_buftarg *btp = bp->b_target;
158 spin_lock(&btp->bt_lru_lock);
159 if (!list_empty(&bp->b_lru)) {
160 list_del_init(&bp->b_lru);
162 atomic_dec(&bp->b_hold);
164 spin_unlock(&btp->bt_lru_lock);
166 ASSERT(atomic_read(&bp->b_hold) >= 1);
172 xfs_buftarg_t *target,
173 xfs_off_t range_base,
175 xfs_buf_flags_t flags)
178 * We don't want certain flags to appear in b_flags.
180 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
182 memset(bp, 0, sizeof(xfs_buf_t));
183 atomic_set(&bp->b_hold, 1);
184 atomic_set(&bp->b_lru_ref, 1);
185 init_completion(&bp->b_iowait);
186 INIT_LIST_HEAD(&bp->b_lru);
187 INIT_LIST_HEAD(&bp->b_list);
188 RB_CLEAR_NODE(&bp->b_rbnode);
189 sema_init(&bp->b_sema, 0); /* held, no waiters */
191 bp->b_target = target;
192 bp->b_file_offset = range_base;
194 * Set buffer_length and count_desired to the same value initially.
195 * I/O routines should use count_desired, which will be the same in
196 * most cases but may be reset (e.g. XFS recovery).
198 bp->b_buffer_length = bp->b_count_desired = range_length;
200 bp->b_bn = XFS_BUF_DADDR_NULL;
201 atomic_set(&bp->b_pin_count, 0);
202 init_waitqueue_head(&bp->b_waiters);
204 XFS_STATS_INC(xb_create);
206 trace_xfs_buf_init(bp, _RET_IP_);
210 * Allocate a page array capable of holding a specified number
211 * of pages, and point the page buf at it.
217 xfs_buf_flags_t flags)
219 /* Make sure that we have a page list */
220 if (bp->b_pages == NULL) {
221 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
222 bp->b_page_count = page_count;
223 if (page_count <= XB_PAGES) {
224 bp->b_pages = bp->b_page_array;
226 bp->b_pages = kmem_alloc(sizeof(struct page *) *
227 page_count, xb_to_km(flags));
228 if (bp->b_pages == NULL)
231 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
237 * Frees b_pages if it was allocated.
243 if (bp->b_pages != bp->b_page_array) {
244 kmem_free(bp->b_pages);
250 * Releases the specified buffer.
252 * The modification state of any associated pages is left unchanged.
253 * The buffer most not be on any hash - use xfs_buf_rele instead for
254 * hashed and refcounted buffers
260 trace_xfs_buf_free(bp, _RET_IP_);
262 ASSERT(list_empty(&bp->b_lru));
264 if (bp->b_flags & _XBF_PAGES) {
267 if (xfs_buf_is_vmapped(bp))
268 vm_unmap_ram(bp->b_addr - bp->b_offset,
271 for (i = 0; i < bp->b_page_count; i++) {
272 struct page *page = bp->b_pages[i];
276 } else if (bp->b_flags & _XBF_KMEM)
277 kmem_free(bp->b_addr);
278 _xfs_buf_free_pages(bp);
279 xfs_buf_deallocate(bp);
283 * Allocates all the pages for buffer in question and builds it's page list.
286 xfs_buf_allocate_memory(
290 size_t size = bp->b_count_desired;
291 size_t nbytes, offset;
292 gfp_t gfp_mask = xb_to_gfp(flags);
293 unsigned short page_count, i;
298 * for buffers that are contained within a single page, just allocate
299 * the memory from the heap - there's no need for the complexity of
300 * page arrays to keep allocation down to order 0.
302 if (bp->b_buffer_length < PAGE_SIZE) {
303 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
305 /* low memory - use alloc_page loop instead */
309 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
311 ((unsigned long)bp->b_addr & PAGE_MASK)) {
312 /* b_addr spans two pages - use alloc_page instead */
313 kmem_free(bp->b_addr);
317 bp->b_offset = offset_in_page(bp->b_addr);
318 bp->b_pages = bp->b_page_array;
319 bp->b_pages[0] = virt_to_page(bp->b_addr);
320 bp->b_page_count = 1;
321 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
326 end = bp->b_file_offset + bp->b_buffer_length;
327 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
328 error = _xfs_buf_get_pages(bp, page_count, flags);
332 offset = bp->b_offset;
333 bp->b_flags |= _XBF_PAGES;
335 for (i = 0; i < bp->b_page_count; i++) {
339 page = alloc_page(gfp_mask);
340 if (unlikely(page == NULL)) {
341 if (flags & XBF_READ_AHEAD) {
342 bp->b_page_count = i;
348 * This could deadlock.
350 * But until all the XFS lowlevel code is revamped to
351 * handle buffer allocation failures we can't do much.
353 if (!(++retries % 100))
355 "possible memory allocation deadlock in %s (mode:0x%x)",
358 XFS_STATS_INC(xb_page_retries);
359 congestion_wait(BLK_RW_ASYNC, HZ/50);
363 XFS_STATS_INC(xb_page_found);
365 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
367 bp->b_pages[i] = page;
373 for (i = 0; i < bp->b_page_count; i++)
374 __free_page(bp->b_pages[i]);
379 * Map buffer into kernel address-space if necessary.
386 ASSERT(bp->b_flags & _XBF_PAGES);
387 if (bp->b_page_count == 1) {
388 /* A single page buffer is always mappable */
389 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
390 bp->b_flags |= XBF_MAPPED;
391 } else if (flags & XBF_MAPPED) {
395 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
400 } while (retried++ <= 1);
404 bp->b_addr += bp->b_offset;
405 bp->b_flags |= XBF_MAPPED;
412 * Finding and Reading Buffers
416 * Look up, and creates if absent, a lockable buffer for
417 * a given range of an inode. The buffer is returned
418 * locked. If other overlapping buffers exist, they are
419 * released before the new buffer is created and locked,
420 * which may imply that this call will block until those buffers
421 * are unlocked. No I/O is implied by this call.
425 xfs_buftarg_t *btp, /* block device target */
426 xfs_off_t ioff, /* starting offset of range */
427 size_t isize, /* length of range */
428 xfs_buf_flags_t flags,
431 xfs_off_t range_base;
433 struct xfs_perag *pag;
434 struct rb_node **rbp;
435 struct rb_node *parent;
438 range_base = (ioff << BBSHIFT);
439 range_length = (isize << BBSHIFT);
441 /* Check for IOs smaller than the sector size / not sector aligned */
442 ASSERT(!(range_length < (1 << btp->bt_sshift)));
443 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
446 pag = xfs_perag_get(btp->bt_mount,
447 xfs_daddr_to_agno(btp->bt_mount, ioff));
450 spin_lock(&pag->pag_buf_lock);
451 rbp = &pag->pag_buf_tree.rb_node;
456 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
458 if (range_base < bp->b_file_offset)
459 rbp = &(*rbp)->rb_left;
460 else if (range_base > bp->b_file_offset)
461 rbp = &(*rbp)->rb_right;
464 * found a block offset match. If the range doesn't
465 * match, the only way this is allowed is if the buffer
466 * in the cache is stale and the transaction that made
467 * it stale has not yet committed. i.e. we are
468 * reallocating a busy extent. Skip this buffer and
469 * continue searching to the right for an exact match.
471 if (bp->b_buffer_length != range_length) {
472 ASSERT(bp->b_flags & XBF_STALE);
473 rbp = &(*rbp)->rb_right;
476 atomic_inc(&bp->b_hold);
483 _xfs_buf_initialize(new_bp, btp, range_base,
484 range_length, flags);
485 rb_link_node(&new_bp->b_rbnode, parent, rbp);
486 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
487 /* the buffer keeps the perag reference until it is freed */
489 spin_unlock(&pag->pag_buf_lock);
491 XFS_STATS_INC(xb_miss_locked);
492 spin_unlock(&pag->pag_buf_lock);
498 spin_unlock(&pag->pag_buf_lock);
501 if (!xfs_buf_trylock(bp)) {
502 if (flags & XBF_TRYLOCK) {
504 XFS_STATS_INC(xb_busy_locked);
508 XFS_STATS_INC(xb_get_locked_waited);
512 * if the buffer is stale, clear all the external state associated with
513 * it. We need to keep flags such as how we allocated the buffer memory
516 if (bp->b_flags & XBF_STALE) {
517 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
518 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
521 trace_xfs_buf_find(bp, flags, _RET_IP_);
522 XFS_STATS_INC(xb_get_locked);
527 * Assembles a buffer covering the specified range.
528 * Storage in memory for all portions of the buffer will be allocated,
529 * although backing storage may not be.
533 xfs_buftarg_t *target,/* target for buffer */
534 xfs_off_t ioff, /* starting offset of range */
535 size_t isize, /* length of range */
536 xfs_buf_flags_t flags)
538 xfs_buf_t *bp, *new_bp;
541 new_bp = xfs_buf_allocate(flags);
542 if (unlikely(!new_bp))
545 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
547 error = xfs_buf_allocate_memory(bp, flags);
551 xfs_buf_deallocate(new_bp);
552 if (unlikely(bp == NULL))
556 if (!(bp->b_flags & XBF_MAPPED)) {
557 error = _xfs_buf_map_pages(bp, flags);
558 if (unlikely(error)) {
559 xfs_warn(target->bt_mount,
560 "%s: failed to map pages\n", __func__);
565 XFS_STATS_INC(xb_get);
568 * Always fill in the block number now, the mapped cases can do
569 * their own overlay of this later.
572 bp->b_count_desired = bp->b_buffer_length;
574 trace_xfs_buf_get(bp, flags, _RET_IP_);
578 if (flags & (XBF_LOCK | XBF_TRYLOCK))
587 xfs_buf_flags_t flags)
591 ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
592 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
594 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | XBF_READ_AHEAD);
595 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
597 status = xfs_buf_iorequest(bp);
598 if (status || bp->b_error || (flags & XBF_ASYNC))
600 return xfs_buf_iowait(bp);
605 xfs_buftarg_t *target,
608 xfs_buf_flags_t flags)
614 bp = xfs_buf_get(target, ioff, isize, flags);
616 trace_xfs_buf_read(bp, flags, _RET_IP_);
618 if (!XFS_BUF_ISDONE(bp)) {
619 XFS_STATS_INC(xb_get_read);
620 _xfs_buf_read(bp, flags);
621 } else if (flags & XBF_ASYNC) {
623 * Read ahead call which is already satisfied,
628 /* We do not want read in the flags */
629 bp->b_flags &= ~XBF_READ;
636 if (flags & (XBF_LOCK | XBF_TRYLOCK))
643 * If we are not low on memory then do the readahead in a deadlock
648 xfs_buftarg_t *target,
652 if (bdi_read_congested(target->bt_bdi))
655 xfs_buf_read(target, ioff, isize,
656 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
660 * Read an uncached buffer from disk. Allocates and returns a locked
661 * buffer containing the disk contents or nothing.
664 xfs_buf_read_uncached(
665 struct xfs_mount *mp,
666 struct xfs_buftarg *target,
674 bp = xfs_buf_get_uncached(target, length, flags);
678 /* set up the buffer for a read IO */
679 XFS_BUF_SET_ADDR(bp, daddr);
683 error = xfs_buf_iowait(bp);
684 if (error || bp->b_error) {
694 xfs_buftarg_t *target)
698 bp = xfs_buf_allocate(0);
700 _xfs_buf_initialize(bp, target, 0, len, 0);
705 * Return a buffer allocated as an empty buffer and associated to external
706 * memory via xfs_buf_associate_memory() back to it's empty state.
714 _xfs_buf_free_pages(bp);
717 bp->b_page_count = 0;
719 bp->b_file_offset = 0;
720 bp->b_buffer_length = bp->b_count_desired = len;
721 bp->b_bn = XFS_BUF_DADDR_NULL;
722 bp->b_flags &= ~XBF_MAPPED;
725 static inline struct page *
729 if ((!is_vmalloc_addr(addr))) {
730 return virt_to_page(addr);
732 return vmalloc_to_page(addr);
737 xfs_buf_associate_memory(
744 unsigned long pageaddr;
745 unsigned long offset;
749 pageaddr = (unsigned long)mem & PAGE_MASK;
750 offset = (unsigned long)mem - pageaddr;
751 buflen = PAGE_ALIGN(len + offset);
752 page_count = buflen >> PAGE_SHIFT;
754 /* Free any previous set of page pointers */
756 _xfs_buf_free_pages(bp);
761 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
765 bp->b_offset = offset;
767 for (i = 0; i < bp->b_page_count; i++) {
768 bp->b_pages[i] = mem_to_page((void *)pageaddr);
769 pageaddr += PAGE_SIZE;
772 bp->b_count_desired = len;
773 bp->b_buffer_length = buflen;
774 bp->b_flags |= XBF_MAPPED;
780 xfs_buf_get_uncached(
781 struct xfs_buftarg *target,
785 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
789 bp = xfs_buf_allocate(0);
790 if (unlikely(bp == NULL))
792 _xfs_buf_initialize(bp, target, 0, len, 0);
794 error = _xfs_buf_get_pages(bp, page_count, 0);
798 for (i = 0; i < page_count; i++) {
799 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
803 bp->b_flags |= _XBF_PAGES;
805 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
806 if (unlikely(error)) {
807 xfs_warn(target->bt_mount,
808 "%s: failed to map pages\n", __func__);
812 trace_xfs_buf_get_uncached(bp, _RET_IP_);
817 __free_page(bp->b_pages[i]);
818 _xfs_buf_free_pages(bp);
820 xfs_buf_deallocate(bp);
826 * Increment reference count on buffer, to hold the buffer concurrently
827 * with another thread which may release (free) the buffer asynchronously.
828 * Must hold the buffer already to call this function.
834 trace_xfs_buf_hold(bp, _RET_IP_);
835 atomic_inc(&bp->b_hold);
839 * Releases a hold on the specified buffer. If the
840 * the hold count is 1, calls xfs_buf_free.
846 struct xfs_perag *pag = bp->b_pag;
848 trace_xfs_buf_rele(bp, _RET_IP_);
851 ASSERT(list_empty(&bp->b_lru));
852 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
853 if (atomic_dec_and_test(&bp->b_hold))
858 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
860 ASSERT(atomic_read(&bp->b_hold) > 0);
861 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
862 if (!(bp->b_flags & XBF_STALE) &&
863 atomic_read(&bp->b_lru_ref)) {
865 spin_unlock(&pag->pag_buf_lock);
868 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
869 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
870 spin_unlock(&pag->pag_buf_lock);
879 * Lock a buffer object, if it is not already locked.
881 * If we come across a stale, pinned, locked buffer, we know that we are
882 * being asked to lock a buffer that has been reallocated. Because it is
883 * pinned, we know that the log has not been pushed to disk and hence it
884 * will still be locked. Rather than continuing to have trylock attempts
885 * fail until someone else pushes the log, push it ourselves before
886 * returning. This means that the xfsaild will not get stuck trying
887 * to push on stale inode buffers.
895 locked = down_trylock(&bp->b_sema) == 0;
898 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
899 xfs_log_force(bp->b_target->bt_mount, 0);
901 trace_xfs_buf_trylock(bp, _RET_IP_);
906 * Lock a buffer object.
908 * If we come across a stale, pinned, locked buffer, we know that we
909 * are being asked to lock a buffer that has been reallocated. Because
910 * it is pinned, we know that the log has not been pushed to disk and
911 * hence it will still be locked. Rather than sleeping until someone
912 * else pushes the log, push it ourselves before trying to get the lock.
918 trace_xfs_buf_lock(bp, _RET_IP_);
920 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
921 xfs_log_force(bp->b_target->bt_mount, 0);
925 trace_xfs_buf_lock_done(bp, _RET_IP_);
929 * Releases the lock on the buffer object.
930 * If the buffer is marked delwri but is not queued, do so before we
931 * unlock the buffer as we need to set flags correctly. We also need to
932 * take a reference for the delwri queue because the unlocker is going to
933 * drop their's and they don't know we just queued it.
942 trace_xfs_buf_unlock(bp, _RET_IP_);
949 DECLARE_WAITQUEUE (wait, current);
951 if (atomic_read(&bp->b_pin_count) == 0)
954 add_wait_queue(&bp->b_waiters, &wait);
956 set_current_state(TASK_UNINTERRUPTIBLE);
957 if (atomic_read(&bp->b_pin_count) == 0)
961 remove_wait_queue(&bp->b_waiters, &wait);
962 set_current_state(TASK_RUNNING);
966 * Buffer Utility Routines
971 struct work_struct *work)
974 container_of(work, xfs_buf_t, b_iodone_work);
977 (*(bp->b_iodone))(bp);
978 else if (bp->b_flags & XBF_ASYNC)
987 trace_xfs_buf_iodone(bp, _RET_IP_);
989 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
990 if (bp->b_error == 0)
991 bp->b_flags |= XBF_DONE;
993 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
995 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
996 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
998 xfs_buf_iodone_work(&bp->b_iodone_work);
1001 complete(&bp->b_iowait);
1010 ASSERT(error >= 0 && error <= 0xffff);
1011 bp->b_error = (unsigned short)error;
1012 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1021 bp->b_flags |= XBF_WRITE;
1022 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1024 xfs_buf_delwri_dequeue(bp);
1027 error = xfs_buf_iowait(bp);
1029 xfs_force_shutdown(bp->b_target->bt_mount,
1030 SHUTDOWN_META_IO_ERROR);
1036 * Called when we want to stop a buffer from getting written or read.
1037 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1038 * so that the proper iodone callbacks get called.
1044 #ifdef XFSERRORDEBUG
1045 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1049 * No need to wait until the buffer is unpinned, we aren't flushing it.
1051 xfs_buf_ioerror(bp, EIO);
1054 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1057 xfs_buf_delwri_dequeue(bp);
1061 xfs_buf_ioend(bp, 0);
1067 * Same as xfs_bioerror, except that we are releasing the buffer
1068 * here ourselves, and avoiding the xfs_buf_ioend call.
1069 * This is meant for userdata errors; metadata bufs come with
1070 * iodone functions attached, so that we can track down errors.
1076 int64_t fl = bp->b_flags;
1078 * No need to wait until the buffer is unpinned.
1079 * We aren't flushing it.
1081 * chunkhold expects B_DONE to be set, whether
1082 * we actually finish the I/O or not. We don't want to
1083 * change that interface.
1086 xfs_buf_delwri_dequeue(bp);
1089 bp->b_iodone = NULL;
1090 if (!(fl & XBF_ASYNC)) {
1092 * Mark b_error and B_ERROR _both_.
1093 * Lot's of chunkcache code assumes that.
1094 * There's no reason to mark error for
1097 xfs_buf_ioerror(bp, EIO);
1098 XFS_BUF_FINISH_IOWAIT(bp);
1108 * All xfs metadata buffers except log state machine buffers
1109 * get this attached as their b_bdstrat callback function.
1110 * This is so that we can catch a buffer
1111 * after prematurely unpinning it to forcibly shutdown the filesystem.
1117 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1118 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1120 * Metadata write that didn't get logged but
1121 * written delayed anyway. These aren't associated
1122 * with a transaction, and can be ignored.
1124 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1125 return xfs_bioerror_relse(bp);
1127 return xfs_bioerror(bp);
1130 xfs_buf_iorequest(bp);
1135 * Wrapper around bdstrat so that we can stop data from going to disk in case
1136 * we are shutting down the filesystem. Typically user data goes thru this
1137 * path; one of the exceptions is the superblock.
1141 struct xfs_mount *mp,
1144 if (XFS_FORCED_SHUTDOWN(mp)) {
1145 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1146 xfs_bioerror_relse(bp);
1150 xfs_buf_iorequest(bp);
1158 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1159 xfs_buf_ioend(bp, schedule);
1167 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1169 xfs_buf_ioerror(bp, -error);
1171 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1172 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1174 _xfs_buf_ioend(bp, 1);
1182 int rw, map_i, total_nr_pages, nr_pages;
1184 int offset = bp->b_offset;
1185 int size = bp->b_count_desired;
1186 sector_t sector = bp->b_bn;
1188 total_nr_pages = bp->b_page_count;
1191 if (bp->b_flags & XBF_WRITE) {
1192 if (bp->b_flags & XBF_SYNCIO)
1196 if (bp->b_flags & XBF_FUA)
1198 if (bp->b_flags & XBF_FLUSH)
1200 } else if (bp->b_flags & XBF_READ_AHEAD) {
1206 /* we only use the buffer cache for meta-data */
1210 atomic_inc(&bp->b_io_remaining);
1211 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1212 if (nr_pages > total_nr_pages)
1213 nr_pages = total_nr_pages;
1215 bio = bio_alloc(GFP_NOIO, nr_pages);
1216 bio->bi_bdev = bp->b_target->bt_bdev;
1217 bio->bi_sector = sector;
1218 bio->bi_end_io = xfs_buf_bio_end_io;
1219 bio->bi_private = bp;
1222 for (; size && nr_pages; nr_pages--, map_i++) {
1223 int rbytes, nbytes = PAGE_SIZE - offset;
1228 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1229 if (rbytes < nbytes)
1233 sector += nbytes >> BBSHIFT;
1238 if (likely(bio->bi_size)) {
1239 if (xfs_buf_is_vmapped(bp)) {
1240 flush_kernel_vmap_range(bp->b_addr,
1241 xfs_buf_vmap_len(bp));
1243 submit_bio(rw, bio);
1247 xfs_buf_ioerror(bp, EIO);
1256 trace_xfs_buf_iorequest(bp, _RET_IP_);
1258 ASSERT(!(bp->b_flags & XBF_DELWRI));
1260 if (bp->b_flags & XBF_WRITE)
1261 xfs_buf_wait_unpin(bp);
1264 /* Set the count to 1 initially, this will stop an I/O
1265 * completion callout which happens before we have started
1266 * all the I/O from calling xfs_buf_ioend too early.
1268 atomic_set(&bp->b_io_remaining, 1);
1269 _xfs_buf_ioapply(bp);
1270 _xfs_buf_ioend(bp, 0);
1277 * Waits for I/O to complete on the buffer supplied.
1278 * It returns immediately if no I/O is pending.
1279 * It returns the I/O error code, if any, or 0 if there was no error.
1285 trace_xfs_buf_iowait(bp, _RET_IP_);
1287 wait_for_completion(&bp->b_iowait);
1289 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1300 if (bp->b_flags & XBF_MAPPED)
1301 return bp->b_addr + offset;
1303 offset += bp->b_offset;
1304 page = bp->b_pages[offset >> PAGE_SHIFT];
1305 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1309 * Move data into or out of a buffer.
1313 xfs_buf_t *bp, /* buffer to process */
1314 size_t boff, /* starting buffer offset */
1315 size_t bsize, /* length to copy */
1316 void *data, /* data address */
1317 xfs_buf_rw_t mode) /* read/write/zero flag */
1319 size_t bend, cpoff, csize;
1322 bend = boff + bsize;
1323 while (boff < bend) {
1324 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1325 cpoff = xfs_buf_poff(boff + bp->b_offset);
1326 csize = min_t(size_t,
1327 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1329 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1333 memset(page_address(page) + cpoff, 0, csize);
1336 memcpy(data, page_address(page) + cpoff, csize);
1339 memcpy(page_address(page) + cpoff, data, csize);
1348 * Handling of buffer targets (buftargs).
1352 * Wait for any bufs with callbacks that have been submitted but have not yet
1353 * returned. These buffers will have an elevated hold count, so wait on those
1354 * while freeing all the buffers only held by the LRU.
1358 struct xfs_buftarg *btp)
1363 spin_lock(&btp->bt_lru_lock);
1364 while (!list_empty(&btp->bt_lru)) {
1365 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1366 if (atomic_read(&bp->b_hold) > 1) {
1367 spin_unlock(&btp->bt_lru_lock);
1372 * clear the LRU reference count so the bufer doesn't get
1373 * ignored in xfs_buf_rele().
1375 atomic_set(&bp->b_lru_ref, 0);
1376 spin_unlock(&btp->bt_lru_lock);
1378 spin_lock(&btp->bt_lru_lock);
1380 spin_unlock(&btp->bt_lru_lock);
1385 struct shrinker *shrink,
1386 struct shrink_control *sc)
1388 struct xfs_buftarg *btp = container_of(shrink,
1389 struct xfs_buftarg, bt_shrinker);
1391 int nr_to_scan = sc->nr_to_scan;
1395 return btp->bt_lru_nr;
1397 spin_lock(&btp->bt_lru_lock);
1398 while (!list_empty(&btp->bt_lru)) {
1399 if (nr_to_scan-- <= 0)
1402 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1405 * Decrement the b_lru_ref count unless the value is already
1406 * zero. If the value is already zero, we need to reclaim the
1407 * buffer, otherwise it gets another trip through the LRU.
1409 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1410 list_move_tail(&bp->b_lru, &btp->bt_lru);
1415 * remove the buffer from the LRU now to avoid needing another
1416 * lock round trip inside xfs_buf_rele().
1418 list_move(&bp->b_lru, &dispose);
1421 spin_unlock(&btp->bt_lru_lock);
1423 while (!list_empty(&dispose)) {
1424 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1425 list_del_init(&bp->b_lru);
1429 return btp->bt_lru_nr;
1434 struct xfs_mount *mp,
1435 struct xfs_buftarg *btp)
1437 unregister_shrinker(&btp->bt_shrinker);
1439 xfs_flush_buftarg(btp, 1);
1440 if (mp->m_flags & XFS_MOUNT_BARRIER)
1441 xfs_blkdev_issue_flush(btp);
1443 kthread_stop(btp->bt_task);
1448 xfs_setsize_buftarg_flags(
1450 unsigned int blocksize,
1451 unsigned int sectorsize,
1454 btp->bt_bsize = blocksize;
1455 btp->bt_sshift = ffs(sectorsize) - 1;
1456 btp->bt_smask = sectorsize - 1;
1458 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1459 xfs_warn(btp->bt_mount,
1460 "Cannot set_blocksize to %u on device %s\n",
1461 sectorsize, xfs_buf_target_name(btp));
1469 * When allocating the initial buffer target we have not yet
1470 * read in the superblock, so don't know what sized sectors
1471 * are being used is at this early stage. Play safe.
1474 xfs_setsize_buftarg_early(
1476 struct block_device *bdev)
1478 return xfs_setsize_buftarg_flags(btp,
1479 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1483 xfs_setsize_buftarg(
1485 unsigned int blocksize,
1486 unsigned int sectorsize)
1488 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1492 xfs_alloc_delwri_queue(
1496 INIT_LIST_HEAD(&btp->bt_delwri_queue);
1497 spin_lock_init(&btp->bt_delwri_lock);
1499 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1500 if (IS_ERR(btp->bt_task))
1501 return PTR_ERR(btp->bt_task);
1507 struct xfs_mount *mp,
1508 struct block_device *bdev,
1514 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1517 btp->bt_dev = bdev->bd_dev;
1518 btp->bt_bdev = bdev;
1519 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1523 INIT_LIST_HEAD(&btp->bt_lru);
1524 spin_lock_init(&btp->bt_lru_lock);
1525 if (xfs_setsize_buftarg_early(btp, bdev))
1527 if (xfs_alloc_delwri_queue(btp, fsname))
1529 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1530 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1531 register_shrinker(&btp->bt_shrinker);
1541 * Delayed write buffer handling
1544 xfs_buf_delwri_queue(
1547 struct xfs_buftarg *btp = bp->b_target;
1549 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1551 ASSERT(!(bp->b_flags & XBF_READ));
1553 spin_lock(&btp->bt_delwri_lock);
1554 if (!list_empty(&bp->b_list)) {
1555 /* if already in the queue, move it to the tail */
1556 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1557 list_move_tail(&bp->b_list, &btp->bt_delwri_queue);
1559 /* start xfsbufd as it is about to have something to do */
1560 if (list_empty(&btp->bt_delwri_queue))
1561 wake_up_process(bp->b_target->bt_task);
1563 atomic_inc(&bp->b_hold);
1564 bp->b_flags |= XBF_DELWRI | _XBF_DELWRI_Q | XBF_ASYNC;
1565 list_add_tail(&bp->b_list, &btp->bt_delwri_queue);
1567 bp->b_queuetime = jiffies;
1568 spin_unlock(&btp->bt_delwri_lock);
1572 xfs_buf_delwri_dequeue(
1577 spin_lock(&bp->b_target->bt_delwri_lock);
1578 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1579 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1580 list_del_init(&bp->b_list);
1583 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1584 spin_unlock(&bp->b_target->bt_delwri_lock);
1589 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1593 * If a delwri buffer needs to be pushed before it has aged out, then promote
1594 * it to the head of the delwri queue so that it will be flushed on the next
1595 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1596 * than the age currently needed to flush the buffer. Hence the next time the
1597 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1600 xfs_buf_delwri_promote(
1603 struct xfs_buftarg *btp = bp->b_target;
1604 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1606 ASSERT(bp->b_flags & XBF_DELWRI);
1607 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1610 * Check the buffer age before locking the delayed write queue as we
1611 * don't need to promote buffers that are already past the flush age.
1613 if (bp->b_queuetime < jiffies - age)
1615 bp->b_queuetime = jiffies - age;
1616 spin_lock(&btp->bt_delwri_lock);
1617 list_move(&bp->b_list, &btp->bt_delwri_queue);
1618 spin_unlock(&btp->bt_delwri_lock);
1622 xfs_buf_runall_queues(
1623 struct workqueue_struct *queue)
1625 flush_workqueue(queue);
1629 * Move as many buffers as specified to the supplied list
1630 * idicating if we skipped any buffers to prevent deadlocks.
1633 xfs_buf_delwri_split(
1634 xfs_buftarg_t *target,
1635 struct list_head *list,
1642 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1643 INIT_LIST_HEAD(list);
1644 spin_lock(&target->bt_delwri_lock);
1645 list_for_each_entry_safe(bp, n, &target->bt_delwri_queue, b_list) {
1646 ASSERT(bp->b_flags & XBF_DELWRI);
1648 if (!xfs_buf_ispinned(bp) && xfs_buf_trylock(bp)) {
1650 time_before(jiffies, bp->b_queuetime + age)) {
1655 bp->b_flags &= ~(XBF_DELWRI | _XBF_DELWRI_Q);
1656 bp->b_flags |= XBF_WRITE;
1657 list_move_tail(&bp->b_list, list);
1658 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1663 spin_unlock(&target->bt_delwri_lock);
1668 * Compare function is more complex than it needs to be because
1669 * the return value is only 32 bits and we are doing comparisons
1675 struct list_head *a,
1676 struct list_head *b)
1678 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1679 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1682 diff = ap->b_bn - bp->b_bn;
1694 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1696 current->flags |= PF_MEMALLOC;
1701 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1702 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1703 struct list_head tmp;
1704 struct blk_plug plug;
1706 if (unlikely(freezing(current))) {
1707 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1710 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1713 /* sleep for a long time if there is nothing to do. */
1714 if (list_empty(&target->bt_delwri_queue))
1715 tout = MAX_SCHEDULE_TIMEOUT;
1716 schedule_timeout_interruptible(tout);
1718 xfs_buf_delwri_split(target, &tmp, age);
1719 list_sort(NULL, &tmp, xfs_buf_cmp);
1721 blk_start_plug(&plug);
1722 while (!list_empty(&tmp)) {
1724 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1725 list_del_init(&bp->b_list);
1728 blk_finish_plug(&plug);
1729 } while (!kthread_should_stop());
1735 * Go through all incore buffers, and release buffers if they belong to
1736 * the given device. This is used in filesystem error handling to
1737 * preserve the consistency of its metadata.
1741 xfs_buftarg_t *target,
1746 LIST_HEAD(tmp_list);
1747 LIST_HEAD(wait_list);
1748 struct blk_plug plug;
1750 xfs_buf_runall_queues(xfsconvertd_workqueue);
1751 xfs_buf_runall_queues(xfsdatad_workqueue);
1752 xfs_buf_runall_queues(xfslogd_workqueue);
1754 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1755 pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1758 * Dropped the delayed write list lock, now walk the temporary list.
1759 * All I/O is issued async and then if we need to wait for completion
1760 * we do that after issuing all the IO.
1762 list_sort(NULL, &tmp_list, xfs_buf_cmp);
1764 blk_start_plug(&plug);
1765 while (!list_empty(&tmp_list)) {
1766 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1767 ASSERT(target == bp->b_target);
1768 list_del_init(&bp->b_list);
1770 bp->b_flags &= ~XBF_ASYNC;
1771 list_add(&bp->b_list, &wait_list);
1775 blk_finish_plug(&plug);
1778 /* Wait for IO to complete. */
1779 while (!list_empty(&wait_list)) {
1780 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1782 list_del_init(&bp->b_list);
1794 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1795 KM_ZONE_HWALIGN, NULL);
1799 xfslogd_workqueue = alloc_workqueue("xfslogd",
1800 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1801 if (!xfslogd_workqueue)
1802 goto out_free_buf_zone;
1804 xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1805 if (!xfsdatad_workqueue)
1806 goto out_destroy_xfslogd_workqueue;
1808 xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1810 if (!xfsconvertd_workqueue)
1811 goto out_destroy_xfsdatad_workqueue;
1815 out_destroy_xfsdatad_workqueue:
1816 destroy_workqueue(xfsdatad_workqueue);
1817 out_destroy_xfslogd_workqueue:
1818 destroy_workqueue(xfslogd_workqueue);
1820 kmem_zone_destroy(xfs_buf_zone);
1826 xfs_buf_terminate(void)
1828 destroy_workqueue(xfsconvertd_workqueue);
1829 destroy_workqueue(xfsdatad_workqueue);
1830 destroy_workqueue(xfslogd_workqueue);
1831 kmem_zone_destroy(xfs_buf_zone);
1834 #ifdef CONFIG_KDB_MODULES
1836 xfs_get_buftarg_list(void)
1838 return &xfs_buftarg_list;