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;
49 #ifdef XFS_BUF_LOCK_TRACKING
50 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
51 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
52 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
54 # define XB_SET_OWNER(bp) do { } while (0)
55 # define XB_CLEAR_OWNER(bp) do { } while (0)
56 # define XB_GET_OWNER(bp) do { } while (0)
59 #define xb_to_gfp(flags) \
60 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
61 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
63 #define xb_to_km(flags) \
64 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
72 * Return true if the buffer is vmapped.
74 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
75 * code is clever enough to know it doesn't have to map a single page,
76 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
78 return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
85 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
89 * xfs_buf_lru_add - add a buffer to the LRU.
91 * The LRU takes a new reference to the buffer so that it will only be freed
92 * once the shrinker takes the buffer off the LRU.
98 struct xfs_buftarg *btp = bp->b_target;
100 spin_lock(&btp->bt_lru_lock);
101 if (list_empty(&bp->b_lru)) {
102 atomic_inc(&bp->b_hold);
103 list_add_tail(&bp->b_lru, &btp->bt_lru);
106 spin_unlock(&btp->bt_lru_lock);
110 * xfs_buf_lru_del - remove a buffer from the LRU
112 * The unlocked check is safe here because it only occurs when there are not
113 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
114 * to optimise the shrinker removing the buffer from the LRU and calling
115 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
122 struct xfs_buftarg *btp = bp->b_target;
124 if (list_empty(&bp->b_lru))
127 spin_lock(&btp->bt_lru_lock);
128 if (!list_empty(&bp->b_lru)) {
129 list_del_init(&bp->b_lru);
132 spin_unlock(&btp->bt_lru_lock);
136 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
137 * b_lru_ref count so that the buffer is freed immediately when the buffer
138 * reference count falls to zero. If the buffer is already on the LRU, we need
139 * to remove the reference that LRU holds on the buffer.
141 * This prevents build-up of stale buffers on the LRU.
147 bp->b_flags |= XBF_STALE;
148 xfs_buf_delwri_dequeue(bp);
149 atomic_set(&(bp)->b_lru_ref, 0);
150 if (!list_empty(&bp->b_lru)) {
151 struct xfs_buftarg *btp = bp->b_target;
153 spin_lock(&btp->bt_lru_lock);
154 if (!list_empty(&bp->b_lru)) {
155 list_del_init(&bp->b_lru);
157 atomic_dec(&bp->b_hold);
159 spin_unlock(&btp->bt_lru_lock);
161 ASSERT(atomic_read(&bp->b_hold) >= 1);
166 struct xfs_buftarg *target,
167 xfs_off_t range_base,
169 xfs_buf_flags_t flags)
173 bp = kmem_zone_alloc(xfs_buf_zone, xb_to_km(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);
205 trace_xfs_buf_init(bp, _RET_IP_);
211 * Allocate a page array capable of holding a specified number
212 * of pages, and point the page buf at it.
218 xfs_buf_flags_t flags)
220 /* Make sure that we have a page list */
221 if (bp->b_pages == NULL) {
222 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
223 bp->b_page_count = page_count;
224 if (page_count <= XB_PAGES) {
225 bp->b_pages = bp->b_page_array;
227 bp->b_pages = kmem_alloc(sizeof(struct page *) *
228 page_count, xb_to_km(flags));
229 if (bp->b_pages == NULL)
232 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
238 * Frees b_pages if it was allocated.
244 if (bp->b_pages != bp->b_page_array) {
245 kmem_free(bp->b_pages);
251 * Releases the specified buffer.
253 * The modification state of any associated pages is left unchanged.
254 * The buffer most not be on any hash - use xfs_buf_rele instead for
255 * hashed and refcounted buffers
261 trace_xfs_buf_free(bp, _RET_IP_);
263 ASSERT(list_empty(&bp->b_lru));
265 if (bp->b_flags & _XBF_PAGES) {
268 if (xfs_buf_is_vmapped(bp))
269 vm_unmap_ram(bp->b_addr - bp->b_offset,
272 for (i = 0; i < bp->b_page_count; i++) {
273 struct page *page = bp->b_pages[i];
277 } else if (bp->b_flags & _XBF_KMEM)
278 kmem_free(bp->b_addr);
279 _xfs_buf_free_pages(bp);
280 kmem_zone_free(xfs_buf_zone, bp);
284 * Allocates all the pages for buffer in question and builds it's page list.
287 xfs_buf_allocate_memory(
291 size_t size = bp->b_count_desired;
292 size_t nbytes, offset;
293 gfp_t gfp_mask = xb_to_gfp(flags);
294 unsigned short page_count, i;
299 * for buffers that are contained within a single page, just allocate
300 * the memory from the heap - there's no need for the complexity of
301 * page arrays to keep allocation down to order 0.
303 if (bp->b_buffer_length < PAGE_SIZE) {
304 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
306 /* low memory - use alloc_page loop instead */
310 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
312 ((unsigned long)bp->b_addr & PAGE_MASK)) {
313 /* b_addr spans two pages - use alloc_page instead */
314 kmem_free(bp->b_addr);
318 bp->b_offset = offset_in_page(bp->b_addr);
319 bp->b_pages = bp->b_page_array;
320 bp->b_pages[0] = virt_to_page(bp->b_addr);
321 bp->b_page_count = 1;
322 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
327 end = bp->b_file_offset + bp->b_buffer_length;
328 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
329 error = _xfs_buf_get_pages(bp, page_count, flags);
333 offset = bp->b_offset;
334 bp->b_flags |= _XBF_PAGES;
336 for (i = 0; i < bp->b_page_count; i++) {
340 page = alloc_page(gfp_mask);
341 if (unlikely(page == NULL)) {
342 if (flags & XBF_READ_AHEAD) {
343 bp->b_page_count = i;
349 * This could deadlock.
351 * But until all the XFS lowlevel code is revamped to
352 * handle buffer allocation failures we can't do much.
354 if (!(++retries % 100))
356 "possible memory allocation deadlock in %s (mode:0x%x)",
359 XFS_STATS_INC(xb_page_retries);
360 congestion_wait(BLK_RW_ASYNC, HZ/50);
364 XFS_STATS_INC(xb_page_found);
366 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
368 bp->b_pages[i] = page;
374 for (i = 0; i < bp->b_page_count; i++)
375 __free_page(bp->b_pages[i]);
380 * Map buffer into kernel address-space if necessary.
387 ASSERT(bp->b_flags & _XBF_PAGES);
388 if (bp->b_page_count == 1) {
389 /* A single page buffer is always mappable */
390 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
391 bp->b_flags |= XBF_MAPPED;
392 } else if (flags & XBF_MAPPED) {
396 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
401 } while (retried++ <= 1);
405 bp->b_addr += bp->b_offset;
406 bp->b_flags |= XBF_MAPPED;
413 * Finding and Reading Buffers
417 * Look up, and creates if absent, a lockable buffer for
418 * a given range of an inode. The buffer is returned
419 * locked. No I/O is implied by this call.
423 xfs_buftarg_t *btp, /* block device target */
424 xfs_off_t ioff, /* starting offset of range */
425 size_t isize, /* length of range */
426 xfs_buf_flags_t flags,
429 xfs_off_t range_base;
431 struct xfs_perag *pag;
432 struct rb_node **rbp;
433 struct rb_node *parent;
436 range_base = (ioff << BBSHIFT);
437 range_length = (isize << BBSHIFT);
439 /* Check for IOs smaller than the sector size / not sector aligned */
440 ASSERT(!(range_length < (1 << btp->bt_sshift)));
441 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
444 pag = xfs_perag_get(btp->bt_mount,
445 xfs_daddr_to_agno(btp->bt_mount, ioff));
448 spin_lock(&pag->pag_buf_lock);
449 rbp = &pag->pag_buf_tree.rb_node;
454 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
456 if (range_base < bp->b_file_offset)
457 rbp = &(*rbp)->rb_left;
458 else if (range_base > bp->b_file_offset)
459 rbp = &(*rbp)->rb_right;
462 * found a block offset match. If the range doesn't
463 * match, the only way this is allowed is if the buffer
464 * in the cache is stale and the transaction that made
465 * it stale has not yet committed. i.e. we are
466 * reallocating a busy extent. Skip this buffer and
467 * continue searching to the right for an exact match.
469 if (bp->b_buffer_length != range_length) {
470 ASSERT(bp->b_flags & XBF_STALE);
471 rbp = &(*rbp)->rb_right;
474 atomic_inc(&bp->b_hold);
481 rb_link_node(&new_bp->b_rbnode, parent, rbp);
482 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
483 /* the buffer keeps the perag reference until it is freed */
485 spin_unlock(&pag->pag_buf_lock);
487 XFS_STATS_INC(xb_miss_locked);
488 spin_unlock(&pag->pag_buf_lock);
494 spin_unlock(&pag->pag_buf_lock);
497 if (!xfs_buf_trylock(bp)) {
498 if (flags & XBF_TRYLOCK) {
500 XFS_STATS_INC(xb_busy_locked);
504 XFS_STATS_INC(xb_get_locked_waited);
508 * if the buffer is stale, clear all the external state associated with
509 * it. We need to keep flags such as how we allocated the buffer memory
512 if (bp->b_flags & XBF_STALE) {
513 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
514 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
517 trace_xfs_buf_find(bp, flags, _RET_IP_);
518 XFS_STATS_INC(xb_get_locked);
523 * Assembles a buffer covering the specified range. The code is optimised for
524 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
525 * more hits than misses.
529 xfs_buftarg_t *target,/* target for buffer */
530 xfs_off_t ioff, /* starting offset of range */
531 size_t isize, /* length of range */
532 xfs_buf_flags_t flags)
535 struct xfs_buf *new_bp;
538 bp = _xfs_buf_find(target, ioff, isize, flags, NULL);
542 new_bp = xfs_buf_alloc(target, ioff << BBSHIFT, isize << BBSHIFT,
544 if (unlikely(!new_bp))
547 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
549 kmem_zone_free(xfs_buf_zone, new_bp);
554 error = xfs_buf_allocate_memory(bp, flags);
558 kmem_zone_free(xfs_buf_zone, new_bp);
561 * Now we have a workable buffer, fill in the block number so
562 * that we can do IO on it.
565 bp->b_count_desired = bp->b_buffer_length;
568 if (!(bp->b_flags & XBF_MAPPED)) {
569 error = _xfs_buf_map_pages(bp, flags);
570 if (unlikely(error)) {
571 xfs_warn(target->bt_mount,
572 "%s: failed to map pages\n", __func__);
577 XFS_STATS_INC(xb_get);
578 trace_xfs_buf_get(bp, flags, _RET_IP_);
582 if (flags & (XBF_LOCK | XBF_TRYLOCK))
591 xfs_buf_flags_t flags)
595 ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
596 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
598 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | XBF_READ_AHEAD);
599 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
601 status = xfs_buf_iorequest(bp);
602 if (status || bp->b_error || (flags & XBF_ASYNC))
604 return xfs_buf_iowait(bp);
609 xfs_buftarg_t *target,
612 xfs_buf_flags_t flags)
618 bp = xfs_buf_get(target, ioff, isize, flags);
620 trace_xfs_buf_read(bp, flags, _RET_IP_);
622 if (!XFS_BUF_ISDONE(bp)) {
623 XFS_STATS_INC(xb_get_read);
624 _xfs_buf_read(bp, flags);
625 } else if (flags & XBF_ASYNC) {
627 * Read ahead call which is already satisfied,
632 /* We do not want read in the flags */
633 bp->b_flags &= ~XBF_READ;
640 if (flags & (XBF_LOCK | XBF_TRYLOCK))
647 * If we are not low on memory then do the readahead in a deadlock
652 xfs_buftarg_t *target,
656 if (bdi_read_congested(target->bt_bdi))
659 xfs_buf_read(target, ioff, isize,
660 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
664 * Read an uncached buffer from disk. Allocates and returns a locked
665 * buffer containing the disk contents or nothing.
668 xfs_buf_read_uncached(
669 struct xfs_mount *mp,
670 struct xfs_buftarg *target,
678 bp = xfs_buf_get_uncached(target, length, flags);
682 /* set up the buffer for a read IO */
683 XFS_BUF_SET_ADDR(bp, daddr);
687 error = xfs_buf_iowait(bp);
688 if (error || bp->b_error) {
696 * Return a buffer allocated as an empty buffer and associated to external
697 * memory via xfs_buf_associate_memory() back to it's empty state.
705 _xfs_buf_free_pages(bp);
708 bp->b_page_count = 0;
710 bp->b_file_offset = 0;
711 bp->b_buffer_length = bp->b_count_desired = len;
712 bp->b_bn = XFS_BUF_DADDR_NULL;
713 bp->b_flags &= ~XBF_MAPPED;
716 static inline struct page *
720 if ((!is_vmalloc_addr(addr))) {
721 return virt_to_page(addr);
723 return vmalloc_to_page(addr);
728 xfs_buf_associate_memory(
735 unsigned long pageaddr;
736 unsigned long offset;
740 pageaddr = (unsigned long)mem & PAGE_MASK;
741 offset = (unsigned long)mem - pageaddr;
742 buflen = PAGE_ALIGN(len + offset);
743 page_count = buflen >> PAGE_SHIFT;
745 /* Free any previous set of page pointers */
747 _xfs_buf_free_pages(bp);
752 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
756 bp->b_offset = offset;
758 for (i = 0; i < bp->b_page_count; i++) {
759 bp->b_pages[i] = mem_to_page((void *)pageaddr);
760 pageaddr += PAGE_SIZE;
763 bp->b_count_desired = len;
764 bp->b_buffer_length = buflen;
765 bp->b_flags |= XBF_MAPPED;
771 xfs_buf_get_uncached(
772 struct xfs_buftarg *target,
776 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
780 bp = xfs_buf_alloc(target, 0, len, 0);
781 if (unlikely(bp == NULL))
784 error = _xfs_buf_get_pages(bp, page_count, 0);
788 for (i = 0; i < page_count; i++) {
789 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
793 bp->b_flags |= _XBF_PAGES;
795 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
796 if (unlikely(error)) {
797 xfs_warn(target->bt_mount,
798 "%s: failed to map pages\n", __func__);
802 trace_xfs_buf_get_uncached(bp, _RET_IP_);
807 __free_page(bp->b_pages[i]);
808 _xfs_buf_free_pages(bp);
810 kmem_zone_free(xfs_buf_zone, bp);
816 * Increment reference count on buffer, to hold the buffer concurrently
817 * with another thread which may release (free) the buffer asynchronously.
818 * Must hold the buffer already to call this function.
824 trace_xfs_buf_hold(bp, _RET_IP_);
825 atomic_inc(&bp->b_hold);
829 * Releases a hold on the specified buffer. If the
830 * the hold count is 1, calls xfs_buf_free.
836 struct xfs_perag *pag = bp->b_pag;
838 trace_xfs_buf_rele(bp, _RET_IP_);
841 ASSERT(list_empty(&bp->b_lru));
842 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
843 if (atomic_dec_and_test(&bp->b_hold))
848 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
850 ASSERT(atomic_read(&bp->b_hold) > 0);
851 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
852 if (!(bp->b_flags & XBF_STALE) &&
853 atomic_read(&bp->b_lru_ref)) {
855 spin_unlock(&pag->pag_buf_lock);
858 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
859 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
860 spin_unlock(&pag->pag_buf_lock);
869 * Lock a buffer object, if it is not already locked.
871 * If we come across a stale, pinned, locked buffer, we know that we are
872 * being asked to lock a buffer that has been reallocated. Because it is
873 * pinned, we know that the log has not been pushed to disk and hence it
874 * will still be locked. Rather than continuing to have trylock attempts
875 * fail until someone else pushes the log, push it ourselves before
876 * returning. This means that the xfsaild will not get stuck trying
877 * to push on stale inode buffers.
885 locked = down_trylock(&bp->b_sema) == 0;
888 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
889 xfs_log_force(bp->b_target->bt_mount, 0);
891 trace_xfs_buf_trylock(bp, _RET_IP_);
896 * Lock a buffer object.
898 * If we come across a stale, pinned, locked buffer, we know that we
899 * are being asked to lock a buffer that has been reallocated. Because
900 * it is pinned, we know that the log has not been pushed to disk and
901 * hence it will still be locked. Rather than sleeping until someone
902 * else pushes the log, push it ourselves before trying to get the lock.
908 trace_xfs_buf_lock(bp, _RET_IP_);
910 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
911 xfs_log_force(bp->b_target->bt_mount, 0);
915 trace_xfs_buf_lock_done(bp, _RET_IP_);
919 * Releases the lock on the buffer object.
920 * If the buffer is marked delwri but is not queued, do so before we
921 * unlock the buffer as we need to set flags correctly. We also need to
922 * take a reference for the delwri queue because the unlocker is going to
923 * drop their's and they don't know we just queued it.
932 trace_xfs_buf_unlock(bp, _RET_IP_);
939 DECLARE_WAITQUEUE (wait, current);
941 if (atomic_read(&bp->b_pin_count) == 0)
944 add_wait_queue(&bp->b_waiters, &wait);
946 set_current_state(TASK_UNINTERRUPTIBLE);
947 if (atomic_read(&bp->b_pin_count) == 0)
951 remove_wait_queue(&bp->b_waiters, &wait);
952 set_current_state(TASK_RUNNING);
956 * Buffer Utility Routines
961 struct work_struct *work)
964 container_of(work, xfs_buf_t, b_iodone_work);
967 (*(bp->b_iodone))(bp);
968 else if (bp->b_flags & XBF_ASYNC)
977 trace_xfs_buf_iodone(bp, _RET_IP_);
979 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
980 if (bp->b_error == 0)
981 bp->b_flags |= XBF_DONE;
983 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
985 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
986 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
988 xfs_buf_iodone_work(&bp->b_iodone_work);
991 complete(&bp->b_iowait);
1000 ASSERT(error >= 0 && error <= 0xffff);
1001 bp->b_error = (unsigned short)error;
1002 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1006 xfs_buf_ioerror_alert(
1010 xfs_alert(bp->b_target->bt_mount,
1011 "metadata I/O error: block 0x%llx (\"%s\") error %d buf count %zd",
1012 (__uint64_t)XFS_BUF_ADDR(bp), func,
1013 bp->b_error, XFS_BUF_COUNT(bp));
1022 bp->b_flags |= XBF_WRITE;
1023 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1025 xfs_buf_delwri_dequeue(bp);
1028 error = xfs_buf_iowait(bp);
1030 xfs_force_shutdown(bp->b_target->bt_mount,
1031 SHUTDOWN_META_IO_ERROR);
1037 * Called when we want to stop a buffer from getting written or read.
1038 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1039 * so that the proper iodone callbacks get called.
1045 #ifdef XFSERRORDEBUG
1046 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1050 * No need to wait until the buffer is unpinned, we aren't flushing it.
1052 xfs_buf_ioerror(bp, EIO);
1055 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
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.
1088 bp->b_iodone = NULL;
1089 if (!(fl & XBF_ASYNC)) {
1091 * Mark b_error and B_ERROR _both_.
1092 * Lot's of chunkcache code assumes that.
1093 * There's no reason to mark error for
1096 xfs_buf_ioerror(bp, EIO);
1097 complete(&bp->b_iowait);
1107 * All xfs metadata buffers except log state machine buffers
1108 * get this attached as their b_bdstrat callback function.
1109 * This is so that we can catch a buffer
1110 * after prematurely unpinning it to forcibly shutdown the filesystem.
1116 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1117 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1119 * Metadata write that didn't get logged but
1120 * written delayed anyway. These aren't associated
1121 * with a transaction, and can be ignored.
1123 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1124 return xfs_bioerror_relse(bp);
1126 return xfs_bioerror(bp);
1129 xfs_buf_iorequest(bp);
1134 * Wrapper around bdstrat so that we can stop data from going to disk in case
1135 * we are shutting down the filesystem. Typically user data goes thru this
1136 * path; one of the exceptions is the superblock.
1140 struct xfs_mount *mp,
1143 if (XFS_FORCED_SHUTDOWN(mp)) {
1144 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1145 xfs_bioerror_relse(bp);
1149 xfs_buf_iorequest(bp);
1157 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1158 xfs_buf_ioend(bp, schedule);
1166 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1168 xfs_buf_ioerror(bp, -error);
1170 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1171 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1173 _xfs_buf_ioend(bp, 1);
1181 int rw, map_i, total_nr_pages, nr_pages;
1183 int offset = bp->b_offset;
1184 int size = bp->b_count_desired;
1185 sector_t sector = bp->b_bn;
1187 total_nr_pages = bp->b_page_count;
1190 if (bp->b_flags & XBF_WRITE) {
1191 if (bp->b_flags & XBF_SYNCIO)
1195 if (bp->b_flags & XBF_FUA)
1197 if (bp->b_flags & XBF_FLUSH)
1199 } else if (bp->b_flags & XBF_READ_AHEAD) {
1205 /* we only use the buffer cache for meta-data */
1209 atomic_inc(&bp->b_io_remaining);
1210 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1211 if (nr_pages > total_nr_pages)
1212 nr_pages = total_nr_pages;
1214 bio = bio_alloc(GFP_NOIO, nr_pages);
1215 bio->bi_bdev = bp->b_target->bt_bdev;
1216 bio->bi_sector = sector;
1217 bio->bi_end_io = xfs_buf_bio_end_io;
1218 bio->bi_private = bp;
1221 for (; size && nr_pages; nr_pages--, map_i++) {
1222 int rbytes, nbytes = PAGE_SIZE - offset;
1227 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1228 if (rbytes < nbytes)
1232 sector += nbytes >> BBSHIFT;
1237 if (likely(bio->bi_size)) {
1238 if (xfs_buf_is_vmapped(bp)) {
1239 flush_kernel_vmap_range(bp->b_addr,
1240 xfs_buf_vmap_len(bp));
1242 submit_bio(rw, bio);
1246 xfs_buf_ioerror(bp, EIO);
1255 trace_xfs_buf_iorequest(bp, _RET_IP_);
1257 ASSERT(!(bp->b_flags & XBF_DELWRI));
1259 if (bp->b_flags & XBF_WRITE)
1260 xfs_buf_wait_unpin(bp);
1263 /* Set the count to 1 initially, this will stop an I/O
1264 * completion callout which happens before we have started
1265 * all the I/O from calling xfs_buf_ioend too early.
1267 atomic_set(&bp->b_io_remaining, 1);
1268 _xfs_buf_ioapply(bp);
1269 _xfs_buf_ioend(bp, 0);
1276 * Waits for I/O to complete on the buffer supplied.
1277 * It returns immediately if no I/O is pending.
1278 * It returns the I/O error code, if any, or 0 if there was no error.
1284 trace_xfs_buf_iowait(bp, _RET_IP_);
1286 wait_for_completion(&bp->b_iowait);
1288 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1299 if (bp->b_flags & XBF_MAPPED)
1300 return bp->b_addr + offset;
1302 offset += bp->b_offset;
1303 page = bp->b_pages[offset >> PAGE_SHIFT];
1304 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1308 * Move data into or out of a buffer.
1312 xfs_buf_t *bp, /* buffer to process */
1313 size_t boff, /* starting buffer offset */
1314 size_t bsize, /* length to copy */
1315 void *data, /* data address */
1316 xfs_buf_rw_t mode) /* read/write/zero flag */
1318 size_t bend, cpoff, csize;
1321 bend = boff + bsize;
1322 while (boff < bend) {
1323 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1324 cpoff = xfs_buf_poff(boff + bp->b_offset);
1325 csize = min_t(size_t,
1326 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1328 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1332 memset(page_address(page) + cpoff, 0, csize);
1335 memcpy(data, page_address(page) + cpoff, csize);
1338 memcpy(page_address(page) + cpoff, data, csize);
1347 * Handling of buffer targets (buftargs).
1351 * Wait for any bufs with callbacks that have been submitted but have not yet
1352 * returned. These buffers will have an elevated hold count, so wait on those
1353 * while freeing all the buffers only held by the LRU.
1357 struct xfs_buftarg *btp)
1362 spin_lock(&btp->bt_lru_lock);
1363 while (!list_empty(&btp->bt_lru)) {
1364 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1365 if (atomic_read(&bp->b_hold) > 1) {
1366 spin_unlock(&btp->bt_lru_lock);
1371 * clear the LRU reference count so the buffer doesn't get
1372 * ignored in xfs_buf_rele().
1374 atomic_set(&bp->b_lru_ref, 0);
1375 spin_unlock(&btp->bt_lru_lock);
1377 spin_lock(&btp->bt_lru_lock);
1379 spin_unlock(&btp->bt_lru_lock);
1384 struct shrinker *shrink,
1385 struct shrink_control *sc)
1387 struct xfs_buftarg *btp = container_of(shrink,
1388 struct xfs_buftarg, bt_shrinker);
1390 int nr_to_scan = sc->nr_to_scan;
1394 return btp->bt_lru_nr;
1396 spin_lock(&btp->bt_lru_lock);
1397 while (!list_empty(&btp->bt_lru)) {
1398 if (nr_to_scan-- <= 0)
1401 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1404 * Decrement the b_lru_ref count unless the value is already
1405 * zero. If the value is already zero, we need to reclaim the
1406 * buffer, otherwise it gets another trip through the LRU.
1408 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1409 list_move_tail(&bp->b_lru, &btp->bt_lru);
1414 * remove the buffer from the LRU now to avoid needing another
1415 * lock round trip inside xfs_buf_rele().
1417 list_move(&bp->b_lru, &dispose);
1420 spin_unlock(&btp->bt_lru_lock);
1422 while (!list_empty(&dispose)) {
1423 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1424 list_del_init(&bp->b_lru);
1428 return btp->bt_lru_nr;
1433 struct xfs_mount *mp,
1434 struct xfs_buftarg *btp)
1436 unregister_shrinker(&btp->bt_shrinker);
1438 xfs_flush_buftarg(btp, 1);
1439 if (mp->m_flags & XFS_MOUNT_BARRIER)
1440 xfs_blkdev_issue_flush(btp);
1442 kthread_stop(btp->bt_task);
1447 xfs_setsize_buftarg_flags(
1449 unsigned int blocksize,
1450 unsigned int sectorsize,
1453 btp->bt_bsize = blocksize;
1454 btp->bt_sshift = ffs(sectorsize) - 1;
1455 btp->bt_smask = sectorsize - 1;
1457 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1458 char name[BDEVNAME_SIZE];
1460 bdevname(btp->bt_bdev, name);
1462 xfs_warn(btp->bt_mount,
1463 "Cannot set_blocksize to %u on device %s\n",
1472 * When allocating the initial buffer target we have not yet
1473 * read in the superblock, so don't know what sized sectors
1474 * are being used is at this early stage. Play safe.
1477 xfs_setsize_buftarg_early(
1479 struct block_device *bdev)
1481 return xfs_setsize_buftarg_flags(btp,
1482 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1486 xfs_setsize_buftarg(
1488 unsigned int blocksize,
1489 unsigned int sectorsize)
1491 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1495 xfs_alloc_delwri_queue(
1499 INIT_LIST_HEAD(&btp->bt_delwri_queue);
1500 spin_lock_init(&btp->bt_delwri_lock);
1502 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1503 if (IS_ERR(btp->bt_task))
1504 return PTR_ERR(btp->bt_task);
1510 struct xfs_mount *mp,
1511 struct block_device *bdev,
1517 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1520 btp->bt_dev = bdev->bd_dev;
1521 btp->bt_bdev = bdev;
1522 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1526 INIT_LIST_HEAD(&btp->bt_lru);
1527 spin_lock_init(&btp->bt_lru_lock);
1528 if (xfs_setsize_buftarg_early(btp, bdev))
1530 if (xfs_alloc_delwri_queue(btp, fsname))
1532 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1533 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1534 register_shrinker(&btp->bt_shrinker);
1544 * Delayed write buffer handling
1547 xfs_buf_delwri_queue(
1550 struct xfs_buftarg *btp = bp->b_target;
1552 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1554 ASSERT(!(bp->b_flags & XBF_READ));
1556 spin_lock(&btp->bt_delwri_lock);
1557 if (!list_empty(&bp->b_list)) {
1558 /* if already in the queue, move it to the tail */
1559 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1560 list_move_tail(&bp->b_list, &btp->bt_delwri_queue);
1562 /* start xfsbufd as it is about to have something to do */
1563 if (list_empty(&btp->bt_delwri_queue))
1564 wake_up_process(bp->b_target->bt_task);
1566 atomic_inc(&bp->b_hold);
1567 bp->b_flags |= XBF_DELWRI | _XBF_DELWRI_Q | XBF_ASYNC;
1568 list_add_tail(&bp->b_list, &btp->bt_delwri_queue);
1570 bp->b_queuetime = jiffies;
1571 spin_unlock(&btp->bt_delwri_lock);
1575 xfs_buf_delwri_dequeue(
1580 spin_lock(&bp->b_target->bt_delwri_lock);
1581 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1582 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1583 list_del_init(&bp->b_list);
1586 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1587 spin_unlock(&bp->b_target->bt_delwri_lock);
1592 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1596 * If a delwri buffer needs to be pushed before it has aged out, then promote
1597 * it to the head of the delwri queue so that it will be flushed on the next
1598 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1599 * than the age currently needed to flush the buffer. Hence the next time the
1600 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1603 xfs_buf_delwri_promote(
1606 struct xfs_buftarg *btp = bp->b_target;
1607 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1609 ASSERT(bp->b_flags & XBF_DELWRI);
1610 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1613 * Check the buffer age before locking the delayed write queue as we
1614 * don't need to promote buffers that are already past the flush age.
1616 if (bp->b_queuetime < jiffies - age)
1618 bp->b_queuetime = jiffies - age;
1619 spin_lock(&btp->bt_delwri_lock);
1620 list_move(&bp->b_list, &btp->bt_delwri_queue);
1621 spin_unlock(&btp->bt_delwri_lock);
1625 * Move as many buffers as specified to the supplied list
1626 * idicating if we skipped any buffers to prevent deadlocks.
1629 xfs_buf_delwri_split(
1630 xfs_buftarg_t *target,
1631 struct list_head *list,
1638 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1639 INIT_LIST_HEAD(list);
1640 spin_lock(&target->bt_delwri_lock);
1641 list_for_each_entry_safe(bp, n, &target->bt_delwri_queue, b_list) {
1642 ASSERT(bp->b_flags & XBF_DELWRI);
1644 if (!xfs_buf_ispinned(bp) && xfs_buf_trylock(bp)) {
1646 time_before(jiffies, bp->b_queuetime + age)) {
1651 bp->b_flags &= ~(XBF_DELWRI | _XBF_DELWRI_Q);
1652 bp->b_flags |= XBF_WRITE;
1653 list_move_tail(&bp->b_list, list);
1654 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1659 spin_unlock(&target->bt_delwri_lock);
1664 * Compare function is more complex than it needs to be because
1665 * the return value is only 32 bits and we are doing comparisons
1671 struct list_head *a,
1672 struct list_head *b)
1674 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1675 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1678 diff = ap->b_bn - bp->b_bn;
1690 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1692 current->flags |= PF_MEMALLOC;
1697 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1698 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1699 struct list_head tmp;
1700 struct blk_plug plug;
1702 if (unlikely(freezing(current)))
1705 /* sleep for a long time if there is nothing to do. */
1706 if (list_empty(&target->bt_delwri_queue))
1707 tout = MAX_SCHEDULE_TIMEOUT;
1708 schedule_timeout_interruptible(tout);
1710 xfs_buf_delwri_split(target, &tmp, age);
1711 list_sort(NULL, &tmp, xfs_buf_cmp);
1713 blk_start_plug(&plug);
1714 while (!list_empty(&tmp)) {
1716 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1717 list_del_init(&bp->b_list);
1720 blk_finish_plug(&plug);
1721 } while (!kthread_should_stop());
1727 * Go through all incore buffers, and release buffers if they belong to
1728 * the given device. This is used in filesystem error handling to
1729 * preserve the consistency of its metadata.
1733 xfs_buftarg_t *target,
1738 LIST_HEAD(tmp_list);
1739 LIST_HEAD(wait_list);
1740 struct blk_plug plug;
1742 flush_workqueue(xfslogd_workqueue);
1744 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1745 pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1748 * Dropped the delayed write list lock, now walk the temporary list.
1749 * All I/O is issued async and then if we need to wait for completion
1750 * we do that after issuing all the IO.
1752 list_sort(NULL, &tmp_list, xfs_buf_cmp);
1754 blk_start_plug(&plug);
1755 while (!list_empty(&tmp_list)) {
1756 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1757 ASSERT(target == bp->b_target);
1758 list_del_init(&bp->b_list);
1760 bp->b_flags &= ~XBF_ASYNC;
1761 list_add(&bp->b_list, &wait_list);
1765 blk_finish_plug(&plug);
1768 /* Wait for IO to complete. */
1769 while (!list_empty(&wait_list)) {
1770 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1772 list_del_init(&bp->b_list);
1784 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1785 KM_ZONE_HWALIGN, NULL);
1789 xfslogd_workqueue = alloc_workqueue("xfslogd",
1790 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1791 if (!xfslogd_workqueue)
1792 goto out_free_buf_zone;
1797 kmem_zone_destroy(xfs_buf_zone);
1803 xfs_buf_terminate(void)
1805 destroy_workqueue(xfslogd_workqueue);
1806 kmem_zone_destroy(xfs_buf_zone);