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)
74 * Return true if the buffer is vmapped.
76 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
77 * code is clever enough to know it doesn't have to map a single page,
78 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
80 return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
87 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
91 * xfs_buf_lru_add - add a buffer to the LRU.
93 * The LRU takes a new reference to the buffer so that it will only be freed
94 * once the shrinker takes the buffer off the LRU.
100 struct xfs_buftarg *btp = bp->b_target;
102 spin_lock(&btp->bt_lru_lock);
103 if (list_empty(&bp->b_lru)) {
104 atomic_inc(&bp->b_hold);
105 list_add_tail(&bp->b_lru, &btp->bt_lru);
108 spin_unlock(&btp->bt_lru_lock);
112 * xfs_buf_lru_del - remove a buffer from the LRU
114 * The unlocked check is safe here because it only occurs when there are not
115 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
116 * to optimise the shrinker removing the buffer from the LRU and calling
117 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
124 struct xfs_buftarg *btp = bp->b_target;
126 if (list_empty(&bp->b_lru))
129 spin_lock(&btp->bt_lru_lock);
130 if (!list_empty(&bp->b_lru)) {
131 list_del_init(&bp->b_lru);
134 spin_unlock(&btp->bt_lru_lock);
138 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
139 * b_lru_ref count so that the buffer is freed immediately when the buffer
140 * reference count falls to zero. If the buffer is already on the LRU, we need
141 * to remove the reference that LRU holds on the buffer.
143 * This prevents build-up of stale buffers on the LRU.
149 bp->b_flags |= XBF_STALE;
150 xfs_buf_delwri_dequeue(bp);
151 atomic_set(&(bp)->b_lru_ref, 0);
152 if (!list_empty(&bp->b_lru)) {
153 struct xfs_buftarg *btp = bp->b_target;
155 spin_lock(&btp->bt_lru_lock);
156 if (!list_empty(&bp->b_lru)) {
157 list_del_init(&bp->b_lru);
159 atomic_dec(&bp->b_hold);
161 spin_unlock(&btp->bt_lru_lock);
163 ASSERT(atomic_read(&bp->b_hold) >= 1);
168 struct xfs_buftarg *target,
169 xfs_off_t range_base,
171 xfs_buf_flags_t flags)
175 bp = kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags));
180 * We don't want certain flags to appear in b_flags.
182 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
184 memset(bp, 0, sizeof(xfs_buf_t));
185 atomic_set(&bp->b_hold, 1);
186 atomic_set(&bp->b_lru_ref, 1);
187 init_completion(&bp->b_iowait);
188 INIT_LIST_HEAD(&bp->b_lru);
189 INIT_LIST_HEAD(&bp->b_list);
190 RB_CLEAR_NODE(&bp->b_rbnode);
191 sema_init(&bp->b_sema, 0); /* held, no waiters */
193 bp->b_target = target;
194 bp->b_file_offset = range_base;
196 * Set buffer_length and count_desired to the same value initially.
197 * I/O routines should use count_desired, which will be the same in
198 * most cases but may be reset (e.g. XFS recovery).
200 bp->b_buffer_length = bp->b_count_desired = range_length;
202 bp->b_bn = XFS_BUF_DADDR_NULL;
203 atomic_set(&bp->b_pin_count, 0);
204 init_waitqueue_head(&bp->b_waiters);
206 XFS_STATS_INC(xb_create);
207 trace_xfs_buf_init(bp, _RET_IP_);
213 * Allocate a page array capable of holding a specified number
214 * of pages, and point the page buf at it.
220 xfs_buf_flags_t flags)
222 /* Make sure that we have a page list */
223 if (bp->b_pages == NULL) {
224 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
225 bp->b_page_count = page_count;
226 if (page_count <= XB_PAGES) {
227 bp->b_pages = bp->b_page_array;
229 bp->b_pages = kmem_alloc(sizeof(struct page *) *
230 page_count, xb_to_km(flags));
231 if (bp->b_pages == NULL)
234 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
240 * Frees b_pages if it was allocated.
246 if (bp->b_pages != bp->b_page_array) {
247 kmem_free(bp->b_pages);
253 * Releases the specified buffer.
255 * The modification state of any associated pages is left unchanged.
256 * The buffer most not be on any hash - use xfs_buf_rele instead for
257 * hashed and refcounted buffers
263 trace_xfs_buf_free(bp, _RET_IP_);
265 ASSERT(list_empty(&bp->b_lru));
267 if (bp->b_flags & _XBF_PAGES) {
270 if (xfs_buf_is_vmapped(bp))
271 vm_unmap_ram(bp->b_addr - bp->b_offset,
274 for (i = 0; i < bp->b_page_count; i++) {
275 struct page *page = bp->b_pages[i];
279 } else if (bp->b_flags & _XBF_KMEM)
280 kmem_free(bp->b_addr);
281 _xfs_buf_free_pages(bp);
282 kmem_zone_free(xfs_buf_zone, bp);
286 * Allocates all the pages for buffer in question and builds it's page list.
289 xfs_buf_allocate_memory(
293 size_t size = bp->b_count_desired;
294 size_t nbytes, offset;
295 gfp_t gfp_mask = xb_to_gfp(flags);
296 unsigned short page_count, i;
301 * for buffers that are contained within a single page, just allocate
302 * the memory from the heap - there's no need for the complexity of
303 * page arrays to keep allocation down to order 0.
305 if (bp->b_buffer_length < PAGE_SIZE) {
306 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
308 /* low memory - use alloc_page loop instead */
312 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
314 ((unsigned long)bp->b_addr & PAGE_MASK)) {
315 /* b_addr spans two pages - use alloc_page instead */
316 kmem_free(bp->b_addr);
320 bp->b_offset = offset_in_page(bp->b_addr);
321 bp->b_pages = bp->b_page_array;
322 bp->b_pages[0] = virt_to_page(bp->b_addr);
323 bp->b_page_count = 1;
324 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
329 end = bp->b_file_offset + bp->b_buffer_length;
330 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
331 error = _xfs_buf_get_pages(bp, page_count, flags);
335 offset = bp->b_offset;
336 bp->b_flags |= _XBF_PAGES;
338 for (i = 0; i < bp->b_page_count; i++) {
342 page = alloc_page(gfp_mask);
343 if (unlikely(page == NULL)) {
344 if (flags & XBF_READ_AHEAD) {
345 bp->b_page_count = i;
351 * This could deadlock.
353 * But until all the XFS lowlevel code is revamped to
354 * handle buffer allocation failures we can't do much.
356 if (!(++retries % 100))
358 "possible memory allocation deadlock in %s (mode:0x%x)",
361 XFS_STATS_INC(xb_page_retries);
362 congestion_wait(BLK_RW_ASYNC, HZ/50);
366 XFS_STATS_INC(xb_page_found);
368 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
370 bp->b_pages[i] = page;
376 for (i = 0; i < bp->b_page_count; i++)
377 __free_page(bp->b_pages[i]);
382 * Map buffer into kernel address-space if necessary.
389 ASSERT(bp->b_flags & _XBF_PAGES);
390 if (bp->b_page_count == 1) {
391 /* A single page buffer is always mappable */
392 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
393 bp->b_flags |= XBF_MAPPED;
394 } else if (flags & XBF_MAPPED) {
398 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
403 } while (retried++ <= 1);
407 bp->b_addr += bp->b_offset;
408 bp->b_flags |= XBF_MAPPED;
415 * Finding and Reading Buffers
419 * Look up, and creates if absent, a lockable buffer for
420 * a given range of an inode. The buffer is returned
421 * locked. 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 rb_link_node(&new_bp->b_rbnode, parent, rbp);
484 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
485 /* the buffer keeps the perag reference until it is freed */
487 spin_unlock(&pag->pag_buf_lock);
489 XFS_STATS_INC(xb_miss_locked);
490 spin_unlock(&pag->pag_buf_lock);
496 spin_unlock(&pag->pag_buf_lock);
499 if (!xfs_buf_trylock(bp)) {
500 if (flags & XBF_TRYLOCK) {
502 XFS_STATS_INC(xb_busy_locked);
506 XFS_STATS_INC(xb_get_locked_waited);
510 * if the buffer is stale, clear all the external state associated with
511 * it. We need to keep flags such as how we allocated the buffer memory
514 if (bp->b_flags & XBF_STALE) {
515 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
516 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
519 trace_xfs_buf_find(bp, flags, _RET_IP_);
520 XFS_STATS_INC(xb_get_locked);
525 * Assembles a buffer covering the specified range. The code is optimised for
526 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
527 * more hits than misses.
531 xfs_buftarg_t *target,/* target for buffer */
532 xfs_off_t ioff, /* starting offset of range */
533 size_t isize, /* length of range */
534 xfs_buf_flags_t flags)
537 struct xfs_buf *new_bp;
540 bp = _xfs_buf_find(target, ioff, isize, flags, NULL);
544 new_bp = xfs_buf_alloc(target, ioff << BBSHIFT, isize << BBSHIFT,
546 if (unlikely(!new_bp))
549 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
551 kmem_zone_free(xfs_buf_zone, new_bp);
556 error = xfs_buf_allocate_memory(bp, flags);
560 kmem_zone_free(xfs_buf_zone, new_bp);
563 * Now we have a workable buffer, fill in the block number so
564 * that we can do IO on it.
567 bp->b_count_desired = bp->b_buffer_length;
570 if (!(bp->b_flags & XBF_MAPPED)) {
571 error = _xfs_buf_map_pages(bp, flags);
572 if (unlikely(error)) {
573 xfs_warn(target->bt_mount,
574 "%s: failed to map pages\n", __func__);
579 XFS_STATS_INC(xb_get);
580 trace_xfs_buf_get(bp, flags, _RET_IP_);
584 if (flags & (XBF_LOCK | XBF_TRYLOCK))
593 xfs_buf_flags_t flags)
597 ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
598 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
600 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | XBF_READ_AHEAD);
601 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
603 status = xfs_buf_iorequest(bp);
604 if (status || bp->b_error || (flags & XBF_ASYNC))
606 return xfs_buf_iowait(bp);
611 xfs_buftarg_t *target,
614 xfs_buf_flags_t flags)
620 bp = xfs_buf_get(target, ioff, isize, flags);
622 trace_xfs_buf_read(bp, flags, _RET_IP_);
624 if (!XFS_BUF_ISDONE(bp)) {
625 XFS_STATS_INC(xb_get_read);
626 _xfs_buf_read(bp, flags);
627 } else if (flags & XBF_ASYNC) {
629 * Read ahead call which is already satisfied,
634 /* We do not want read in the flags */
635 bp->b_flags &= ~XBF_READ;
642 if (flags & (XBF_LOCK | XBF_TRYLOCK))
649 * If we are not low on memory then do the readahead in a deadlock
654 xfs_buftarg_t *target,
658 if (bdi_read_congested(target->bt_bdi))
661 xfs_buf_read(target, ioff, isize,
662 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
666 * Read an uncached buffer from disk. Allocates and returns a locked
667 * buffer containing the disk contents or nothing.
670 xfs_buf_read_uncached(
671 struct xfs_mount *mp,
672 struct xfs_buftarg *target,
680 bp = xfs_buf_get_uncached(target, length, flags);
684 /* set up the buffer for a read IO */
685 XFS_BUF_SET_ADDR(bp, daddr);
689 error = xfs_buf_iowait(bp);
690 if (error || bp->b_error) {
698 * Return a buffer allocated as an empty buffer and associated to external
699 * memory via xfs_buf_associate_memory() back to it's empty state.
707 _xfs_buf_free_pages(bp);
710 bp->b_page_count = 0;
712 bp->b_file_offset = 0;
713 bp->b_buffer_length = bp->b_count_desired = len;
714 bp->b_bn = XFS_BUF_DADDR_NULL;
715 bp->b_flags &= ~XBF_MAPPED;
718 static inline struct page *
722 if ((!is_vmalloc_addr(addr))) {
723 return virt_to_page(addr);
725 return vmalloc_to_page(addr);
730 xfs_buf_associate_memory(
737 unsigned long pageaddr;
738 unsigned long offset;
742 pageaddr = (unsigned long)mem & PAGE_MASK;
743 offset = (unsigned long)mem - pageaddr;
744 buflen = PAGE_ALIGN(len + offset);
745 page_count = buflen >> PAGE_SHIFT;
747 /* Free any previous set of page pointers */
749 _xfs_buf_free_pages(bp);
754 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
758 bp->b_offset = offset;
760 for (i = 0; i < bp->b_page_count; i++) {
761 bp->b_pages[i] = mem_to_page((void *)pageaddr);
762 pageaddr += PAGE_SIZE;
765 bp->b_count_desired = len;
766 bp->b_buffer_length = buflen;
767 bp->b_flags |= XBF_MAPPED;
773 xfs_buf_get_uncached(
774 struct xfs_buftarg *target,
778 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
782 bp = xfs_buf_alloc(target, 0, len, 0);
783 if (unlikely(bp == NULL))
786 error = _xfs_buf_get_pages(bp, page_count, 0);
790 for (i = 0; i < page_count; i++) {
791 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
795 bp->b_flags |= _XBF_PAGES;
797 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
798 if (unlikely(error)) {
799 xfs_warn(target->bt_mount,
800 "%s: failed to map pages\n", __func__);
804 trace_xfs_buf_get_uncached(bp, _RET_IP_);
809 __free_page(bp->b_pages[i]);
810 _xfs_buf_free_pages(bp);
812 kmem_zone_free(xfs_buf_zone, bp);
818 * Increment reference count on buffer, to hold the buffer concurrently
819 * with another thread which may release (free) the buffer asynchronously.
820 * Must hold the buffer already to call this function.
826 trace_xfs_buf_hold(bp, _RET_IP_);
827 atomic_inc(&bp->b_hold);
831 * Releases a hold on the specified buffer. If the
832 * the hold count is 1, calls xfs_buf_free.
838 struct xfs_perag *pag = bp->b_pag;
840 trace_xfs_buf_rele(bp, _RET_IP_);
843 ASSERT(list_empty(&bp->b_lru));
844 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
845 if (atomic_dec_and_test(&bp->b_hold))
850 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
852 ASSERT(atomic_read(&bp->b_hold) > 0);
853 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
854 if (!(bp->b_flags & XBF_STALE) &&
855 atomic_read(&bp->b_lru_ref)) {
857 spin_unlock(&pag->pag_buf_lock);
860 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
861 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
862 spin_unlock(&pag->pag_buf_lock);
871 * Lock a buffer object, if it is not already locked.
873 * If we come across a stale, pinned, locked buffer, we know that we are
874 * being asked to lock a buffer that has been reallocated. Because it is
875 * pinned, we know that the log has not been pushed to disk and hence it
876 * will still be locked. Rather than continuing to have trylock attempts
877 * fail until someone else pushes the log, push it ourselves before
878 * returning. This means that the xfsaild will not get stuck trying
879 * to push on stale inode buffers.
887 locked = down_trylock(&bp->b_sema) == 0;
890 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
891 xfs_log_force(bp->b_target->bt_mount, 0);
893 trace_xfs_buf_trylock(bp, _RET_IP_);
898 * Lock a buffer object.
900 * If we come across a stale, pinned, locked buffer, we know that we
901 * are being asked to lock a buffer that has been reallocated. Because
902 * it is pinned, we know that the log has not been pushed to disk and
903 * hence it will still be locked. Rather than sleeping until someone
904 * else pushes the log, push it ourselves before trying to get the lock.
910 trace_xfs_buf_lock(bp, _RET_IP_);
912 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
913 xfs_log_force(bp->b_target->bt_mount, 0);
917 trace_xfs_buf_lock_done(bp, _RET_IP_);
921 * Releases the lock on the buffer object.
922 * If the buffer is marked delwri but is not queued, do so before we
923 * unlock the buffer as we need to set flags correctly. We also need to
924 * take a reference for the delwri queue because the unlocker is going to
925 * drop their's and they don't know we just queued it.
934 trace_xfs_buf_unlock(bp, _RET_IP_);
941 DECLARE_WAITQUEUE (wait, current);
943 if (atomic_read(&bp->b_pin_count) == 0)
946 add_wait_queue(&bp->b_waiters, &wait);
948 set_current_state(TASK_UNINTERRUPTIBLE);
949 if (atomic_read(&bp->b_pin_count) == 0)
953 remove_wait_queue(&bp->b_waiters, &wait);
954 set_current_state(TASK_RUNNING);
958 * Buffer Utility Routines
963 struct work_struct *work)
966 container_of(work, xfs_buf_t, b_iodone_work);
969 (*(bp->b_iodone))(bp);
970 else if (bp->b_flags & XBF_ASYNC)
979 trace_xfs_buf_iodone(bp, _RET_IP_);
981 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
982 if (bp->b_error == 0)
983 bp->b_flags |= XBF_DONE;
985 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
987 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
988 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
990 xfs_buf_iodone_work(&bp->b_iodone_work);
993 complete(&bp->b_iowait);
1002 ASSERT(error >= 0 && error <= 0xffff);
1003 bp->b_error = (unsigned short)error;
1004 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1008 xfs_buf_ioerror_alert(
1012 xfs_alert(bp->b_target->bt_mount,
1013 "metadata I/O error: block 0x%llx (\"%s\") error %d buf count %zd",
1014 (__uint64_t)XFS_BUF_ADDR(bp), func,
1015 bp->b_error, XFS_BUF_COUNT(bp));
1024 bp->b_flags |= XBF_WRITE;
1025 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1027 xfs_buf_delwri_dequeue(bp);
1030 error = xfs_buf_iowait(bp);
1032 xfs_force_shutdown(bp->b_target->bt_mount,
1033 SHUTDOWN_META_IO_ERROR);
1039 * Called when we want to stop a buffer from getting written or read.
1040 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1041 * so that the proper iodone callbacks get called.
1047 #ifdef XFSERRORDEBUG
1048 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1052 * No need to wait until the buffer is unpinned, we aren't flushing it.
1054 xfs_buf_ioerror(bp, EIO);
1057 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1063 xfs_buf_ioend(bp, 0);
1069 * Same as xfs_bioerror, except that we are releasing the buffer
1070 * here ourselves, and avoiding the xfs_buf_ioend call.
1071 * This is meant for userdata errors; metadata bufs come with
1072 * iodone functions attached, so that we can track down errors.
1078 int64_t fl = bp->b_flags;
1080 * No need to wait until the buffer is unpinned.
1081 * We aren't flushing it.
1083 * chunkhold expects B_DONE to be set, whether
1084 * we actually finish the I/O or not. We don't want to
1085 * change that interface.
1090 bp->b_iodone = NULL;
1091 if (!(fl & XBF_ASYNC)) {
1093 * Mark b_error and B_ERROR _both_.
1094 * Lot's of chunkcache code assumes that.
1095 * There's no reason to mark error for
1098 xfs_buf_ioerror(bp, EIO);
1099 complete(&bp->b_iowait);
1109 * All xfs metadata buffers except log state machine buffers
1110 * get this attached as their b_bdstrat callback function.
1111 * This is so that we can catch a buffer
1112 * after prematurely unpinning it to forcibly shutdown the filesystem.
1118 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1119 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1121 * Metadata write that didn't get logged but
1122 * written delayed anyway. These aren't associated
1123 * with a transaction, and can be ignored.
1125 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1126 return xfs_bioerror_relse(bp);
1128 return xfs_bioerror(bp);
1131 xfs_buf_iorequest(bp);
1136 * Wrapper around bdstrat so that we can stop data from going to disk in case
1137 * we are shutting down the filesystem. Typically user data goes thru this
1138 * path; one of the exceptions is the superblock.
1142 struct xfs_mount *mp,
1145 if (XFS_FORCED_SHUTDOWN(mp)) {
1146 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1147 xfs_bioerror_relse(bp);
1151 xfs_buf_iorequest(bp);
1159 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1160 xfs_buf_ioend(bp, schedule);
1168 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1170 xfs_buf_ioerror(bp, -error);
1172 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1173 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1175 _xfs_buf_ioend(bp, 1);
1183 int rw, map_i, total_nr_pages, nr_pages;
1185 int offset = bp->b_offset;
1186 int size = bp->b_count_desired;
1187 sector_t sector = bp->b_bn;
1189 total_nr_pages = bp->b_page_count;
1192 if (bp->b_flags & XBF_WRITE) {
1193 if (bp->b_flags & XBF_SYNCIO)
1197 if (bp->b_flags & XBF_FUA)
1199 if (bp->b_flags & XBF_FLUSH)
1201 } else if (bp->b_flags & XBF_READ_AHEAD) {
1207 /* we only use the buffer cache for meta-data */
1211 atomic_inc(&bp->b_io_remaining);
1212 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1213 if (nr_pages > total_nr_pages)
1214 nr_pages = total_nr_pages;
1216 bio = bio_alloc(GFP_NOIO, nr_pages);
1217 bio->bi_bdev = bp->b_target->bt_bdev;
1218 bio->bi_sector = sector;
1219 bio->bi_end_io = xfs_buf_bio_end_io;
1220 bio->bi_private = bp;
1223 for (; size && nr_pages; nr_pages--, map_i++) {
1224 int rbytes, nbytes = PAGE_SIZE - offset;
1229 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1230 if (rbytes < nbytes)
1234 sector += nbytes >> BBSHIFT;
1239 if (likely(bio->bi_size)) {
1240 if (xfs_buf_is_vmapped(bp)) {
1241 flush_kernel_vmap_range(bp->b_addr,
1242 xfs_buf_vmap_len(bp));
1244 submit_bio(rw, bio);
1248 xfs_buf_ioerror(bp, EIO);
1257 trace_xfs_buf_iorequest(bp, _RET_IP_);
1259 ASSERT(!(bp->b_flags & XBF_DELWRI));
1261 if (bp->b_flags & XBF_WRITE)
1262 xfs_buf_wait_unpin(bp);
1265 /* Set the count to 1 initially, this will stop an I/O
1266 * completion callout which happens before we have started
1267 * all the I/O from calling xfs_buf_ioend too early.
1269 atomic_set(&bp->b_io_remaining, 1);
1270 _xfs_buf_ioapply(bp);
1271 _xfs_buf_ioend(bp, 0);
1278 * Waits for I/O to complete on the buffer supplied.
1279 * It returns immediately if no I/O is pending.
1280 * It returns the I/O error code, if any, or 0 if there was no error.
1286 trace_xfs_buf_iowait(bp, _RET_IP_);
1288 wait_for_completion(&bp->b_iowait);
1290 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1301 if (bp->b_flags & XBF_MAPPED)
1302 return bp->b_addr + offset;
1304 offset += bp->b_offset;
1305 page = bp->b_pages[offset >> PAGE_SHIFT];
1306 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1310 * Move data into or out of a buffer.
1314 xfs_buf_t *bp, /* buffer to process */
1315 size_t boff, /* starting buffer offset */
1316 size_t bsize, /* length to copy */
1317 void *data, /* data address */
1318 xfs_buf_rw_t mode) /* read/write/zero flag */
1320 size_t bend, cpoff, csize;
1323 bend = boff + bsize;
1324 while (boff < bend) {
1325 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1326 cpoff = xfs_buf_poff(boff + bp->b_offset);
1327 csize = min_t(size_t,
1328 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1330 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1334 memset(page_address(page) + cpoff, 0, csize);
1337 memcpy(data, page_address(page) + cpoff, csize);
1340 memcpy(page_address(page) + cpoff, data, csize);
1349 * Handling of buffer targets (buftargs).
1353 * Wait for any bufs with callbacks that have been submitted but have not yet
1354 * returned. These buffers will have an elevated hold count, so wait on those
1355 * while freeing all the buffers only held by the LRU.
1359 struct xfs_buftarg *btp)
1364 spin_lock(&btp->bt_lru_lock);
1365 while (!list_empty(&btp->bt_lru)) {
1366 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1367 if (atomic_read(&bp->b_hold) > 1) {
1368 spin_unlock(&btp->bt_lru_lock);
1373 * clear the LRU reference count so the bufer doesn't get
1374 * ignored in xfs_buf_rele().
1376 atomic_set(&bp->b_lru_ref, 0);
1377 spin_unlock(&btp->bt_lru_lock);
1379 spin_lock(&btp->bt_lru_lock);
1381 spin_unlock(&btp->bt_lru_lock);
1386 struct shrinker *shrink,
1387 struct shrink_control *sc)
1389 struct xfs_buftarg *btp = container_of(shrink,
1390 struct xfs_buftarg, bt_shrinker);
1392 int nr_to_scan = sc->nr_to_scan;
1396 return btp->bt_lru_nr;
1398 spin_lock(&btp->bt_lru_lock);
1399 while (!list_empty(&btp->bt_lru)) {
1400 if (nr_to_scan-- <= 0)
1403 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1406 * Decrement the b_lru_ref count unless the value is already
1407 * zero. If the value is already zero, we need to reclaim the
1408 * buffer, otherwise it gets another trip through the LRU.
1410 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1411 list_move_tail(&bp->b_lru, &btp->bt_lru);
1416 * remove the buffer from the LRU now to avoid needing another
1417 * lock round trip inside xfs_buf_rele().
1419 list_move(&bp->b_lru, &dispose);
1422 spin_unlock(&btp->bt_lru_lock);
1424 while (!list_empty(&dispose)) {
1425 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1426 list_del_init(&bp->b_lru);
1430 return btp->bt_lru_nr;
1435 struct xfs_mount *mp,
1436 struct xfs_buftarg *btp)
1438 unregister_shrinker(&btp->bt_shrinker);
1440 xfs_flush_buftarg(btp, 1);
1441 if (mp->m_flags & XFS_MOUNT_BARRIER)
1442 xfs_blkdev_issue_flush(btp);
1444 kthread_stop(btp->bt_task);
1449 xfs_setsize_buftarg_flags(
1451 unsigned int blocksize,
1452 unsigned int sectorsize,
1455 btp->bt_bsize = blocksize;
1456 btp->bt_sshift = ffs(sectorsize) - 1;
1457 btp->bt_smask = sectorsize - 1;
1459 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1460 char name[BDEVNAME_SIZE];
1462 bdevname(btp->bt_bdev, name);
1464 xfs_warn(btp->bt_mount,
1465 "Cannot set_blocksize to %u on device %s\n",
1474 * When allocating the initial buffer target we have not yet
1475 * read in the superblock, so don't know what sized sectors
1476 * are being used is at this early stage. Play safe.
1479 xfs_setsize_buftarg_early(
1481 struct block_device *bdev)
1483 return xfs_setsize_buftarg_flags(btp,
1484 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1488 xfs_setsize_buftarg(
1490 unsigned int blocksize,
1491 unsigned int sectorsize)
1493 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1497 xfs_alloc_delwri_queue(
1501 INIT_LIST_HEAD(&btp->bt_delwri_queue);
1502 spin_lock_init(&btp->bt_delwri_lock);
1504 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1505 if (IS_ERR(btp->bt_task))
1506 return PTR_ERR(btp->bt_task);
1512 struct xfs_mount *mp,
1513 struct block_device *bdev,
1519 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1522 btp->bt_dev = bdev->bd_dev;
1523 btp->bt_bdev = bdev;
1524 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1528 INIT_LIST_HEAD(&btp->bt_lru);
1529 spin_lock_init(&btp->bt_lru_lock);
1530 if (xfs_setsize_buftarg_early(btp, bdev))
1532 if (xfs_alloc_delwri_queue(btp, fsname))
1534 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1535 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1536 register_shrinker(&btp->bt_shrinker);
1546 * Delayed write buffer handling
1549 xfs_buf_delwri_queue(
1552 struct xfs_buftarg *btp = bp->b_target;
1554 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1556 ASSERT(!(bp->b_flags & XBF_READ));
1558 spin_lock(&btp->bt_delwri_lock);
1559 if (!list_empty(&bp->b_list)) {
1560 /* if already in the queue, move it to the tail */
1561 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1562 list_move_tail(&bp->b_list, &btp->bt_delwri_queue);
1564 /* start xfsbufd as it is about to have something to do */
1565 if (list_empty(&btp->bt_delwri_queue))
1566 wake_up_process(bp->b_target->bt_task);
1568 atomic_inc(&bp->b_hold);
1569 bp->b_flags |= XBF_DELWRI | _XBF_DELWRI_Q | XBF_ASYNC;
1570 list_add_tail(&bp->b_list, &btp->bt_delwri_queue);
1572 bp->b_queuetime = jiffies;
1573 spin_unlock(&btp->bt_delwri_lock);
1577 xfs_buf_delwri_dequeue(
1582 spin_lock(&bp->b_target->bt_delwri_lock);
1583 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1584 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1585 list_del_init(&bp->b_list);
1588 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1589 spin_unlock(&bp->b_target->bt_delwri_lock);
1594 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1598 * If a delwri buffer needs to be pushed before it has aged out, then promote
1599 * it to the head of the delwri queue so that it will be flushed on the next
1600 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1601 * than the age currently needed to flush the buffer. Hence the next time the
1602 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1605 xfs_buf_delwri_promote(
1608 struct xfs_buftarg *btp = bp->b_target;
1609 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1611 ASSERT(bp->b_flags & XBF_DELWRI);
1612 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1615 * Check the buffer age before locking the delayed write queue as we
1616 * don't need to promote buffers that are already past the flush age.
1618 if (bp->b_queuetime < jiffies - age)
1620 bp->b_queuetime = jiffies - age;
1621 spin_lock(&btp->bt_delwri_lock);
1622 list_move(&bp->b_list, &btp->bt_delwri_queue);
1623 spin_unlock(&btp->bt_delwri_lock);
1627 * Move as many buffers as specified to the supplied list
1628 * idicating if we skipped any buffers to prevent deadlocks.
1631 xfs_buf_delwri_split(
1632 xfs_buftarg_t *target,
1633 struct list_head *list,
1640 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1641 INIT_LIST_HEAD(list);
1642 spin_lock(&target->bt_delwri_lock);
1643 list_for_each_entry_safe(bp, n, &target->bt_delwri_queue, b_list) {
1644 ASSERT(bp->b_flags & XBF_DELWRI);
1646 if (!xfs_buf_ispinned(bp) && xfs_buf_trylock(bp)) {
1648 time_before(jiffies, bp->b_queuetime + age)) {
1653 bp->b_flags &= ~(XBF_DELWRI | _XBF_DELWRI_Q);
1654 bp->b_flags |= XBF_WRITE;
1655 list_move_tail(&bp->b_list, list);
1656 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1661 spin_unlock(&target->bt_delwri_lock);
1666 * Compare function is more complex than it needs to be because
1667 * the return value is only 32 bits and we are doing comparisons
1673 struct list_head *a,
1674 struct list_head *b)
1676 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1677 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1680 diff = ap->b_bn - bp->b_bn;
1692 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1694 current->flags |= PF_MEMALLOC;
1699 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1700 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1701 struct list_head tmp;
1702 struct blk_plug plug;
1704 if (unlikely(freezing(current))) {
1705 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1708 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1711 /* sleep for a long time if there is nothing to do. */
1712 if (list_empty(&target->bt_delwri_queue))
1713 tout = MAX_SCHEDULE_TIMEOUT;
1714 schedule_timeout_interruptible(tout);
1716 xfs_buf_delwri_split(target, &tmp, age);
1717 list_sort(NULL, &tmp, xfs_buf_cmp);
1719 blk_start_plug(&plug);
1720 while (!list_empty(&tmp)) {
1722 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1723 list_del_init(&bp->b_list);
1726 blk_finish_plug(&plug);
1727 } while (!kthread_should_stop());
1733 * Go through all incore buffers, and release buffers if they belong to
1734 * the given device. This is used in filesystem error handling to
1735 * preserve the consistency of its metadata.
1739 xfs_buftarg_t *target,
1744 LIST_HEAD(tmp_list);
1745 LIST_HEAD(wait_list);
1746 struct blk_plug plug;
1748 flush_workqueue(xfslogd_workqueue);
1750 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1751 pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1754 * Dropped the delayed write list lock, now walk the temporary list.
1755 * All I/O is issued async and then if we need to wait for completion
1756 * we do that after issuing all the IO.
1758 list_sort(NULL, &tmp_list, xfs_buf_cmp);
1760 blk_start_plug(&plug);
1761 while (!list_empty(&tmp_list)) {
1762 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1763 ASSERT(target == bp->b_target);
1764 list_del_init(&bp->b_list);
1766 bp->b_flags &= ~XBF_ASYNC;
1767 list_add(&bp->b_list, &wait_list);
1771 blk_finish_plug(&plug);
1774 /* Wait for IO to complete. */
1775 while (!list_empty(&wait_list)) {
1776 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1778 list_del_init(&bp->b_list);
1790 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1791 KM_ZONE_HWALIGN, NULL);
1795 xfslogd_workqueue = alloc_workqueue("xfslogd",
1796 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1797 if (!xfslogd_workqueue)
1798 goto out_free_buf_zone;
1800 xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1801 if (!xfsdatad_workqueue)
1802 goto out_destroy_xfslogd_workqueue;
1804 xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1806 if (!xfsconvertd_workqueue)
1807 goto out_destroy_xfsdatad_workqueue;
1811 out_destroy_xfsdatad_workqueue:
1812 destroy_workqueue(xfsdatad_workqueue);
1813 out_destroy_xfslogd_workqueue:
1814 destroy_workqueue(xfslogd_workqueue);
1816 kmem_zone_destroy(xfs_buf_zone);
1822 xfs_buf_terminate(void)
1824 destroy_workqueue(xfsconvertd_workqueue);
1825 destroy_workqueue(xfsdatad_workqueue);
1826 destroy_workqueue(xfslogd_workqueue);
1827 kmem_zone_destroy(xfs_buf_zone);