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>
36 #include <linux/list_sort.h>
42 #include "xfs_mount.h"
43 #include "xfs_trace.h"
45 static kmem_zone_t *xfs_buf_zone;
46 STATIC int xfsbufd(void *);
47 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
49 static struct workqueue_struct *xfslogd_workqueue;
50 struct workqueue_struct *xfsdatad_workqueue;
51 struct workqueue_struct *xfsconvertd_workqueue;
53 #ifdef XFS_BUF_LOCK_TRACKING
54 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
55 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
56 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
58 # define XB_SET_OWNER(bp) do { } while (0)
59 # define XB_CLEAR_OWNER(bp) do { } while (0)
60 # define XB_GET_OWNER(bp) do { } while (0)
63 #define xb_to_gfp(flags) \
64 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
65 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
67 #define xb_to_km(flags) \
68 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
70 #define xfs_buf_allocate(flags) \
71 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
72 #define xfs_buf_deallocate(bp) \
73 kmem_zone_free(xfs_buf_zone, (bp));
80 * Return true if the buffer is vmapped.
82 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
83 * code is clever enough to know it doesn't have to map a single page,
84 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
86 return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
93 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
97 * xfs_buf_lru_add - add a buffer to the LRU.
99 * The LRU takes a new reference to the buffer so that it will only be freed
100 * once the shrinker takes the buffer off the LRU.
106 struct xfs_buftarg *btp = bp->b_target;
108 spin_lock(&btp->bt_lru_lock);
109 if (list_empty(&bp->b_lru)) {
110 atomic_inc(&bp->b_hold);
111 list_add_tail(&bp->b_lru, &btp->bt_lru);
114 spin_unlock(&btp->bt_lru_lock);
118 * xfs_buf_lru_del - remove a buffer from the LRU
120 * The unlocked check is safe here because it only occurs when there are not
121 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
122 * to optimise the shrinker removing the buffer from the LRU and calling
123 * xfs_buf_free(). i.e. it removes an unneccessary round trip on the
130 struct xfs_buftarg *btp = bp->b_target;
132 if (list_empty(&bp->b_lru))
135 spin_lock(&btp->bt_lru_lock);
136 if (!list_empty(&bp->b_lru)) {
137 list_del_init(&bp->b_lru);
140 spin_unlock(&btp->bt_lru_lock);
144 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
145 * b_lru_ref count so that the buffer is freed immediately when the buffer
146 * reference count falls to zero. If the buffer is already on the LRU, we need
147 * to remove the reference that LRU holds on the buffer.
149 * This prevents build-up of stale buffers on the LRU.
155 bp->b_flags |= XBF_STALE;
156 atomic_set(&(bp)->b_lru_ref, 0);
157 if (!list_empty(&bp->b_lru)) {
158 struct xfs_buftarg *btp = bp->b_target;
160 spin_lock(&btp->bt_lru_lock);
161 if (!list_empty(&bp->b_lru)) {
162 list_del_init(&bp->b_lru);
164 atomic_dec(&bp->b_hold);
166 spin_unlock(&btp->bt_lru_lock);
168 ASSERT(atomic_read(&bp->b_hold) >= 1);
174 xfs_buftarg_t *target,
175 xfs_off_t range_base,
177 xfs_buf_flags_t 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);
208 trace_xfs_buf_init(bp, _RET_IP_);
212 * Allocate a page array capable of holding a specified number
213 * of pages, and point the page buf at it.
219 xfs_buf_flags_t flags)
221 /* Make sure that we have a page list */
222 if (bp->b_pages == NULL) {
223 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
224 bp->b_page_count = page_count;
225 if (page_count <= XB_PAGES) {
226 bp->b_pages = bp->b_page_array;
228 bp->b_pages = kmem_alloc(sizeof(struct page *) *
229 page_count, xb_to_km(flags));
230 if (bp->b_pages == NULL)
233 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
239 * Frees b_pages if it was allocated.
245 if (bp->b_pages != bp->b_page_array) {
246 kmem_free(bp->b_pages);
252 * Releases the specified buffer.
254 * The modification state of any associated pages is left unchanged.
255 * The buffer most not be on any hash - use xfs_buf_rele instead for
256 * hashed and refcounted buffers
262 trace_xfs_buf_free(bp, _RET_IP_);
264 ASSERT(list_empty(&bp->b_lru));
266 if (bp->b_flags & _XBF_PAGES) {
269 if (xfs_buf_is_vmapped(bp))
270 vm_unmap_ram(bp->b_addr - bp->b_offset,
273 for (i = 0; i < bp->b_page_count; i++) {
274 struct page *page = bp->b_pages[i];
278 } else if (bp->b_flags & _XBF_KMEM)
279 kmem_free(bp->b_addr);
280 _xfs_buf_free_pages(bp);
281 xfs_buf_deallocate(bp);
285 * Allocates all the pages for buffer in question and builds it's page list.
288 xfs_buf_allocate_memory(
292 size_t size = bp->b_count_desired;
293 size_t nbytes, offset;
294 gfp_t gfp_mask = xb_to_gfp(flags);
295 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 first = bp->b_file_offset >> PAGE_SHIFT;
337 bp->b_flags |= _XBF_PAGES;
339 for (i = 0; i < bp->b_page_count; i++) {
343 page = alloc_page(gfp_mask);
344 if (unlikely(page == NULL)) {
345 if (flags & XBF_READ_AHEAD) {
346 bp->b_page_count = i;
352 * This could deadlock.
354 * But until all the XFS lowlevel code is revamped to
355 * handle buffer allocation failures we can't do much.
357 if (!(++retries % 100))
359 "possible memory allocation deadlock in %s (mode:0x%x)",
362 XFS_STATS_INC(xb_page_retries);
363 congestion_wait(BLK_RW_ASYNC, HZ/50);
367 XFS_STATS_INC(xb_page_found);
369 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
371 bp->b_pages[i] = page;
377 for (i = 0; i < bp->b_page_count; i++)
378 __free_page(bp->b_pages[i]);
383 * Map buffer into kernel address-space if nessecary.
390 ASSERT(bp->b_flags & _XBF_PAGES);
391 if (bp->b_page_count == 1) {
392 /* A single page buffer is always mappable */
393 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
394 bp->b_flags |= XBF_MAPPED;
395 } else if (flags & XBF_MAPPED) {
399 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
404 } while (retried++ <= 1);
408 bp->b_addr += bp->b_offset;
409 bp->b_flags |= XBF_MAPPED;
416 * Finding and Reading Buffers
420 * Look up, and creates if absent, a lockable buffer for
421 * a given range of an inode. The buffer is returned
422 * locked. If other overlapping buffers exist, they are
423 * released before the new buffer is created and locked,
424 * which may imply that this call will block until those buffers
425 * are unlocked. No I/O is implied by this call.
429 xfs_buftarg_t *btp, /* block device target */
430 xfs_off_t ioff, /* starting offset of range */
431 size_t isize, /* length of range */
432 xfs_buf_flags_t flags,
435 xfs_off_t range_base;
437 struct xfs_perag *pag;
438 struct rb_node **rbp;
439 struct rb_node *parent;
442 range_base = (ioff << BBSHIFT);
443 range_length = (isize << BBSHIFT);
445 /* Check for IOs smaller than the sector size / not sector aligned */
446 ASSERT(!(range_length < (1 << btp->bt_sshift)));
447 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
450 pag = xfs_perag_get(btp->bt_mount,
451 xfs_daddr_to_agno(btp->bt_mount, ioff));
454 spin_lock(&pag->pag_buf_lock);
455 rbp = &pag->pag_buf_tree.rb_node;
460 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
462 if (range_base < bp->b_file_offset)
463 rbp = &(*rbp)->rb_left;
464 else if (range_base > bp->b_file_offset)
465 rbp = &(*rbp)->rb_right;
468 * found a block offset match. If the range doesn't
469 * match, the only way this is allowed is if the buffer
470 * in the cache is stale and the transaction that made
471 * it stale has not yet committed. i.e. we are
472 * reallocating a busy extent. Skip this buffer and
473 * continue searching to the right for an exact match.
475 if (bp->b_buffer_length != range_length) {
476 ASSERT(bp->b_flags & XBF_STALE);
477 rbp = &(*rbp)->rb_right;
480 atomic_inc(&bp->b_hold);
487 _xfs_buf_initialize(new_bp, btp, range_base,
488 range_length, flags);
489 rb_link_node(&new_bp->b_rbnode, parent, rbp);
490 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
491 /* the buffer keeps the perag reference until it is freed */
493 spin_unlock(&pag->pag_buf_lock);
495 XFS_STATS_INC(xb_miss_locked);
496 spin_unlock(&pag->pag_buf_lock);
502 spin_unlock(&pag->pag_buf_lock);
505 if (xfs_buf_cond_lock(bp)) {
506 /* failed, so wait for the lock if requested. */
507 if (!(flags & XBF_TRYLOCK)) {
509 XFS_STATS_INC(xb_get_locked_waited);
512 XFS_STATS_INC(xb_busy_locked);
518 * if the buffer is stale, clear all the external state associated with
519 * it. We need to keep flags such as how we allocated the buffer memory
522 if (bp->b_flags & XBF_STALE) {
523 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
524 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
527 trace_xfs_buf_find(bp, flags, _RET_IP_);
528 XFS_STATS_INC(xb_get_locked);
533 * Assembles a buffer covering the specified range.
534 * Storage in memory for all portions of the buffer will be allocated,
535 * although backing storage may not be.
539 xfs_buftarg_t *target,/* target for buffer */
540 xfs_off_t ioff, /* starting offset of range */
541 size_t isize, /* length of range */
542 xfs_buf_flags_t flags)
544 xfs_buf_t *bp, *new_bp;
547 new_bp = xfs_buf_allocate(flags);
548 if (unlikely(!new_bp))
551 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
553 error = xfs_buf_allocate_memory(bp, flags);
557 xfs_buf_deallocate(new_bp);
558 if (unlikely(bp == NULL))
562 if (!(bp->b_flags & XBF_MAPPED)) {
563 error = _xfs_buf_map_pages(bp, flags);
564 if (unlikely(error)) {
565 xfs_warn(target->bt_mount,
566 "%s: failed to map pages\n", __func__);
571 XFS_STATS_INC(xb_get);
574 * Always fill in the block number now, the mapped cases can do
575 * their own overlay of this later.
578 bp->b_count_desired = bp->b_buffer_length;
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 | \
601 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
602 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
603 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
605 status = xfs_buf_iorequest(bp);
606 if (status || XFS_BUF_ISERROR(bp) || (flags & XBF_ASYNC))
608 return xfs_buf_iowait(bp);
613 xfs_buftarg_t *target,
616 xfs_buf_flags_t flags)
622 bp = xfs_buf_get(target, ioff, isize, flags);
624 trace_xfs_buf_read(bp, flags, _RET_IP_);
626 if (!XFS_BUF_ISDONE(bp)) {
627 XFS_STATS_INC(xb_get_read);
628 _xfs_buf_read(bp, flags);
629 } else if (flags & XBF_ASYNC) {
631 * Read ahead call which is already satisfied,
636 /* We do not want read in the flags */
637 bp->b_flags &= ~XBF_READ;
644 if (flags & (XBF_LOCK | XBF_TRYLOCK))
651 * If we are not low on memory then do the readahead in a deadlock
656 xfs_buftarg_t *target,
660 struct backing_dev_info *bdi;
662 if (bdi_read_congested(target->bt_bdi))
665 xfs_buf_read(target, ioff, isize,
666 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
670 * Read an uncached buffer from disk. Allocates and returns a locked
671 * buffer containing the disk contents or nothing.
674 xfs_buf_read_uncached(
675 struct xfs_mount *mp,
676 struct xfs_buftarg *target,
684 bp = xfs_buf_get_uncached(target, length, flags);
688 /* set up the buffer for a read IO */
690 XFS_BUF_SET_ADDR(bp, daddr);
695 error = xfs_buf_iowait(bp);
696 if (error || bp->b_error) {
706 xfs_buftarg_t *target)
710 bp = xfs_buf_allocate(0);
712 _xfs_buf_initialize(bp, target, 0, len, 0);
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_allocate(0);
781 if (unlikely(bp == NULL))
783 _xfs_buf_initialize(bp, target, 0, len, 0);
785 error = _xfs_buf_get_pages(bp, page_count, 0);
789 for (i = 0; i < page_count; i++) {
790 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
794 bp->b_flags |= _XBF_PAGES;
796 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
797 if (unlikely(error)) {
798 xfs_warn(target->bt_mount,
799 "%s: failed to map pages\n", __func__);
805 trace_xfs_buf_get_uncached(bp, _RET_IP_);
810 __free_page(bp->b_pages[i]);
811 _xfs_buf_free_pages(bp);
813 xfs_buf_deallocate(bp);
819 * Increment reference count on buffer, to hold the buffer concurrently
820 * with another thread which may release (free) the buffer asynchronously.
821 * Must hold the buffer already to call this function.
827 trace_xfs_buf_hold(bp, _RET_IP_);
828 atomic_inc(&bp->b_hold);
832 * Releases a hold on the specified buffer. If the
833 * the hold count is 1, calls xfs_buf_free.
839 struct xfs_perag *pag = bp->b_pag;
841 trace_xfs_buf_rele(bp, _RET_IP_);
844 ASSERT(list_empty(&bp->b_lru));
845 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
846 if (atomic_dec_and_test(&bp->b_hold))
851 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
853 ASSERT(atomic_read(&bp->b_hold) > 0);
854 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
855 if (!(bp->b_flags & XBF_STALE) &&
856 atomic_read(&bp->b_lru_ref)) {
858 spin_unlock(&pag->pag_buf_lock);
861 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
862 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
863 spin_unlock(&pag->pag_buf_lock);
872 * Lock a buffer object, if it is not already locked.
874 * If we come across a stale, pinned, locked buffer, we know that we are
875 * being asked to lock a buffer that has been reallocated. Because it is
876 * pinned, we know that the log has not been pushed to disk and hence it
877 * will still be locked. Rather than continuing to have trylock attempts
878 * fail until someone else pushes the log, push it ourselves before
879 * returning. This means that the xfsaild will not get stuck trying
880 * to push on stale inode buffers.
888 locked = down_trylock(&bp->b_sema) == 0;
891 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
892 xfs_log_force(bp->b_target->bt_mount, 0);
894 trace_xfs_buf_cond_lock(bp, _RET_IP_);
895 return locked ? 0 : -EBUSY;
902 return bp->b_sema.count;
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);
922 if (atomic_read(&bp->b_io_remaining))
923 blk_flush_plug(current);
927 trace_xfs_buf_lock_done(bp, _RET_IP_);
931 * Releases the lock on the buffer object.
932 * If the buffer is marked delwri but is not queued, do so before we
933 * unlock the buffer as we need to set flags correctly. We also need to
934 * take a reference for the delwri queue because the unlocker is going to
935 * drop their's and they don't know we just queued it.
941 if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
942 atomic_inc(&bp->b_hold);
943 bp->b_flags |= XBF_ASYNC;
944 xfs_buf_delwri_queue(bp, 0);
950 trace_xfs_buf_unlock(bp, _RET_IP_);
957 DECLARE_WAITQUEUE (wait, current);
959 if (atomic_read(&bp->b_pin_count) == 0)
962 add_wait_queue(&bp->b_waiters, &wait);
964 set_current_state(TASK_UNINTERRUPTIBLE);
965 if (atomic_read(&bp->b_pin_count) == 0)
969 remove_wait_queue(&bp->b_waiters, &wait);
970 set_current_state(TASK_RUNNING);
974 * Buffer Utility Routines
979 struct work_struct *work)
982 container_of(work, xfs_buf_t, b_iodone_work);
985 (*(bp->b_iodone))(bp);
986 else if (bp->b_flags & XBF_ASYNC)
995 trace_xfs_buf_iodone(bp, _RET_IP_);
997 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
998 if (bp->b_error == 0)
999 bp->b_flags |= XBF_DONE;
1001 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1003 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1004 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1006 xfs_buf_iodone_work(&bp->b_iodone_work);
1009 complete(&bp->b_iowait);
1018 ASSERT(error >= 0 && error <= 0xffff);
1019 bp->b_error = (unsigned short)error;
1020 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1025 struct xfs_mount *mp,
1030 bp->b_flags |= XBF_WRITE;
1031 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1033 xfs_buf_delwri_dequeue(bp);
1036 error = xfs_buf_iowait(bp);
1038 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1048 trace_xfs_buf_bdwrite(bp, _RET_IP_);
1050 bp->b_flags &= ~XBF_READ;
1051 bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1053 xfs_buf_delwri_queue(bp, 1);
1057 * Called when we want to stop a buffer from getting written or read.
1058 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1059 * so that the proper iodone callbacks get called.
1065 #ifdef XFSERRORDEBUG
1066 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1070 * No need to wait until the buffer is unpinned, we aren't flushing it.
1072 XFS_BUF_ERROR(bp, EIO);
1075 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1078 XFS_BUF_UNDELAYWRITE(bp);
1082 xfs_buf_ioend(bp, 0);
1088 * Same as xfs_bioerror, except that we are releasing the buffer
1089 * here ourselves, and avoiding the xfs_buf_ioend call.
1090 * This is meant for userdata errors; metadata bufs come with
1091 * iodone functions attached, so that we can track down errors.
1097 int64_t fl = XFS_BUF_BFLAGS(bp);
1099 * No need to wait until the buffer is unpinned.
1100 * We aren't flushing it.
1102 * chunkhold expects B_DONE to be set, whether
1103 * we actually finish the I/O or not. We don't want to
1104 * change that interface.
1107 XFS_BUF_UNDELAYWRITE(bp);
1110 XFS_BUF_CLR_IODONE_FUNC(bp);
1111 if (!(fl & XBF_ASYNC)) {
1113 * Mark b_error and B_ERROR _both_.
1114 * Lot's of chunkcache code assumes that.
1115 * There's no reason to mark error for
1118 XFS_BUF_ERROR(bp, EIO);
1119 XFS_BUF_FINISH_IOWAIT(bp);
1129 * All xfs metadata buffers except log state machine buffers
1130 * get this attached as their b_bdstrat callback function.
1131 * This is so that we can catch a buffer
1132 * after prematurely unpinning it to forcibly shutdown the filesystem.
1138 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1139 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1141 * Metadata write that didn't get logged but
1142 * written delayed anyway. These aren't associated
1143 * with a transaction, and can be ignored.
1145 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1146 return xfs_bioerror_relse(bp);
1148 return xfs_bioerror(bp);
1151 xfs_buf_iorequest(bp);
1156 * Wrapper around bdstrat so that we can stop data from going to disk in case
1157 * we are shutting down the filesystem. Typically user data goes thru this
1158 * path; one of the exceptions is the superblock.
1162 struct xfs_mount *mp,
1165 if (XFS_FORCED_SHUTDOWN(mp)) {
1166 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1167 xfs_bioerror_relse(bp);
1171 xfs_buf_iorequest(bp);
1179 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1180 xfs_buf_ioend(bp, schedule);
1188 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1190 xfs_buf_ioerror(bp, -error);
1192 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1193 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1195 _xfs_buf_ioend(bp, 1);
1203 int rw, map_i, total_nr_pages, nr_pages;
1205 int offset = bp->b_offset;
1206 int size = bp->b_count_desired;
1207 sector_t sector = bp->b_bn;
1209 total_nr_pages = bp->b_page_count;
1212 if (bp->b_flags & XBF_ORDERED) {
1213 ASSERT(!(bp->b_flags & XBF_READ));
1214 rw = WRITE_FLUSH_FUA;
1215 } else if (bp->b_flags & XBF_LOG_BUFFER) {
1216 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1217 bp->b_flags &= ~_XBF_RUN_QUEUES;
1218 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1219 } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1220 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1221 bp->b_flags &= ~_XBF_RUN_QUEUES;
1222 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1224 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1225 (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1230 atomic_inc(&bp->b_io_remaining);
1231 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1232 if (nr_pages > total_nr_pages)
1233 nr_pages = total_nr_pages;
1235 bio = bio_alloc(GFP_NOIO, nr_pages);
1236 bio->bi_bdev = bp->b_target->bt_bdev;
1237 bio->bi_sector = sector;
1238 bio->bi_end_io = xfs_buf_bio_end_io;
1239 bio->bi_private = bp;
1242 for (; size && nr_pages; nr_pages--, map_i++) {
1243 int rbytes, nbytes = PAGE_SIZE - offset;
1248 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1249 if (rbytes < nbytes)
1253 sector += nbytes >> BBSHIFT;
1258 if (likely(bio->bi_size)) {
1259 if (xfs_buf_is_vmapped(bp)) {
1260 flush_kernel_vmap_range(bp->b_addr,
1261 xfs_buf_vmap_len(bp));
1263 submit_bio(rw, bio);
1267 xfs_buf_ioerror(bp, EIO);
1276 trace_xfs_buf_iorequest(bp, _RET_IP_);
1278 if (bp->b_flags & XBF_DELWRI) {
1279 xfs_buf_delwri_queue(bp, 1);
1283 if (bp->b_flags & XBF_WRITE) {
1284 xfs_buf_wait_unpin(bp);
1289 /* Set the count to 1 initially, this will stop an I/O
1290 * completion callout which happens before we have started
1291 * all the I/O from calling xfs_buf_ioend too early.
1293 atomic_set(&bp->b_io_remaining, 1);
1294 _xfs_buf_ioapply(bp);
1295 _xfs_buf_ioend(bp, 0);
1302 * Waits for I/O to complete on the buffer supplied.
1303 * It returns immediately if no I/O is pending.
1304 * It returns the I/O error code, if any, or 0 if there was no error.
1310 trace_xfs_buf_iowait(bp, _RET_IP_);
1312 if (atomic_read(&bp->b_io_remaining))
1313 blk_flush_plug(current);
1314 wait_for_completion(&bp->b_iowait);
1316 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1327 if (bp->b_flags & XBF_MAPPED)
1328 return XFS_BUF_PTR(bp) + offset;
1330 offset += bp->b_offset;
1331 page = bp->b_pages[offset >> PAGE_SHIFT];
1332 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1336 * Move data into or out of a buffer.
1340 xfs_buf_t *bp, /* buffer to process */
1341 size_t boff, /* starting buffer offset */
1342 size_t bsize, /* length to copy */
1343 void *data, /* data address */
1344 xfs_buf_rw_t mode) /* read/write/zero flag */
1346 size_t bend, cpoff, csize;
1349 bend = boff + bsize;
1350 while (boff < bend) {
1351 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1352 cpoff = xfs_buf_poff(boff + bp->b_offset);
1353 csize = min_t(size_t,
1354 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1356 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1360 memset(page_address(page) + cpoff, 0, csize);
1363 memcpy(data, page_address(page) + cpoff, csize);
1366 memcpy(page_address(page) + cpoff, data, csize);
1375 * Handling of buffer targets (buftargs).
1379 * Wait for any bufs with callbacks that have been submitted but have not yet
1380 * returned. These buffers will have an elevated hold count, so wait on those
1381 * while freeing all the buffers only held by the LRU.
1385 struct xfs_buftarg *btp)
1390 spin_lock(&btp->bt_lru_lock);
1391 while (!list_empty(&btp->bt_lru)) {
1392 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1393 if (atomic_read(&bp->b_hold) > 1) {
1394 spin_unlock(&btp->bt_lru_lock);
1399 * clear the LRU reference count so the bufer doesn't get
1400 * ignored in xfs_buf_rele().
1402 atomic_set(&bp->b_lru_ref, 0);
1403 spin_unlock(&btp->bt_lru_lock);
1405 spin_lock(&btp->bt_lru_lock);
1407 spin_unlock(&btp->bt_lru_lock);
1412 struct shrinker *shrink,
1416 struct xfs_buftarg *btp = container_of(shrink,
1417 struct xfs_buftarg, bt_shrinker);
1422 return btp->bt_lru_nr;
1424 spin_lock(&btp->bt_lru_lock);
1425 while (!list_empty(&btp->bt_lru)) {
1426 if (nr_to_scan-- <= 0)
1429 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1432 * Decrement the b_lru_ref count unless the value is already
1433 * zero. If the value is already zero, we need to reclaim the
1434 * buffer, otherwise it gets another trip through the LRU.
1436 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1437 list_move_tail(&bp->b_lru, &btp->bt_lru);
1442 * remove the buffer from the LRU now to avoid needing another
1443 * lock round trip inside xfs_buf_rele().
1445 list_move(&bp->b_lru, &dispose);
1448 spin_unlock(&btp->bt_lru_lock);
1450 while (!list_empty(&dispose)) {
1451 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1452 list_del_init(&bp->b_lru);
1456 return btp->bt_lru_nr;
1461 struct xfs_mount *mp,
1462 struct xfs_buftarg *btp)
1464 unregister_shrinker(&btp->bt_shrinker);
1466 xfs_flush_buftarg(btp, 1);
1467 if (mp->m_flags & XFS_MOUNT_BARRIER)
1468 xfs_blkdev_issue_flush(btp);
1470 kthread_stop(btp->bt_task);
1475 xfs_setsize_buftarg_flags(
1477 unsigned int blocksize,
1478 unsigned int sectorsize,
1481 btp->bt_bsize = blocksize;
1482 btp->bt_sshift = ffs(sectorsize) - 1;
1483 btp->bt_smask = sectorsize - 1;
1485 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1486 xfs_warn(btp->bt_mount,
1487 "Cannot set_blocksize to %u on device %s\n",
1488 sectorsize, XFS_BUFTARG_NAME(btp));
1496 * When allocating the initial buffer target we have not yet
1497 * read in the superblock, so don't know what sized sectors
1498 * are being used is at this early stage. Play safe.
1501 xfs_setsize_buftarg_early(
1503 struct block_device *bdev)
1505 return xfs_setsize_buftarg_flags(btp,
1506 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1510 xfs_setsize_buftarg(
1512 unsigned int blocksize,
1513 unsigned int sectorsize)
1515 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1519 xfs_alloc_delwrite_queue(
1523 INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1524 spin_lock_init(&btp->bt_delwrite_lock);
1526 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1527 if (IS_ERR(btp->bt_task))
1528 return PTR_ERR(btp->bt_task);
1534 struct xfs_mount *mp,
1535 struct block_device *bdev,
1541 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1544 btp->bt_dev = bdev->bd_dev;
1545 btp->bt_bdev = bdev;
1546 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1550 INIT_LIST_HEAD(&btp->bt_lru);
1551 spin_lock_init(&btp->bt_lru_lock);
1552 if (xfs_setsize_buftarg_early(btp, bdev))
1554 if (xfs_alloc_delwrite_queue(btp, fsname))
1556 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1557 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1558 register_shrinker(&btp->bt_shrinker);
1568 * Delayed write buffer handling
1571 xfs_buf_delwri_queue(
1575 struct list_head *dwq = &bp->b_target->bt_delwrite_queue;
1576 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1578 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1580 ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1583 /* If already in the queue, dequeue and place at tail */
1584 if (!list_empty(&bp->b_list)) {
1585 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1587 atomic_dec(&bp->b_hold);
1588 list_del(&bp->b_list);
1591 if (list_empty(dwq)) {
1592 /* start xfsbufd as it is about to have something to do */
1593 wake_up_process(bp->b_target->bt_task);
1596 bp->b_flags |= _XBF_DELWRI_Q;
1597 list_add_tail(&bp->b_list, dwq);
1598 bp->b_queuetime = jiffies;
1606 xfs_buf_delwri_dequeue(
1609 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1613 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1614 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1615 list_del_init(&bp->b_list);
1618 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1624 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1628 * If a delwri buffer needs to be pushed before it has aged out, then promote
1629 * it to the head of the delwri queue so that it will be flushed on the next
1630 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1631 * than the age currently needed to flush the buffer. Hence the next time the
1632 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1635 xfs_buf_delwri_promote(
1638 struct xfs_buftarg *btp = bp->b_target;
1639 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1641 ASSERT(bp->b_flags & XBF_DELWRI);
1642 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1645 * Check the buffer age before locking the delayed write queue as we
1646 * don't need to promote buffers that are already past the flush age.
1648 if (bp->b_queuetime < jiffies - age)
1650 bp->b_queuetime = jiffies - age;
1651 spin_lock(&btp->bt_delwrite_lock);
1652 list_move(&bp->b_list, &btp->bt_delwrite_queue);
1653 spin_unlock(&btp->bt_delwrite_lock);
1657 xfs_buf_runall_queues(
1658 struct workqueue_struct *queue)
1660 flush_workqueue(queue);
1664 * Move as many buffers as specified to the supplied list
1665 * idicating if we skipped any buffers to prevent deadlocks.
1668 xfs_buf_delwri_split(
1669 xfs_buftarg_t *target,
1670 struct list_head *list,
1674 struct list_head *dwq = &target->bt_delwrite_queue;
1675 spinlock_t *dwlk = &target->bt_delwrite_lock;
1679 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1680 INIT_LIST_HEAD(list);
1682 list_for_each_entry_safe(bp, n, dwq, b_list) {
1683 ASSERT(bp->b_flags & XBF_DELWRI);
1685 if (!XFS_BUF_ISPINNED(bp) && !xfs_buf_cond_lock(bp)) {
1687 time_before(jiffies, bp->b_queuetime + age)) {
1692 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1694 bp->b_flags |= XBF_WRITE;
1695 list_move_tail(&bp->b_list, list);
1696 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1707 * Compare function is more complex than it needs to be because
1708 * the return value is only 32 bits and we are doing comparisons
1714 struct list_head *a,
1715 struct list_head *b)
1717 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1718 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1721 diff = ap->b_bn - bp->b_bn;
1730 xfs_buf_delwri_sort(
1731 xfs_buftarg_t *target,
1732 struct list_head *list)
1734 list_sort(NULL, list, xfs_buf_cmp);
1741 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1743 current->flags |= PF_MEMALLOC;
1748 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1749 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1751 struct list_head tmp;
1753 if (unlikely(freezing(current))) {
1754 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1757 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1760 /* sleep for a long time if there is nothing to do. */
1761 if (list_empty(&target->bt_delwrite_queue))
1762 tout = MAX_SCHEDULE_TIMEOUT;
1763 schedule_timeout_interruptible(tout);
1765 xfs_buf_delwri_split(target, &tmp, age);
1766 list_sort(NULL, &tmp, xfs_buf_cmp);
1767 while (!list_empty(&tmp)) {
1769 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1770 list_del_init(&bp->b_list);
1775 blk_flush_plug(current);
1777 } while (!kthread_should_stop());
1783 * Go through all incore buffers, and release buffers if they belong to
1784 * the given device. This is used in filesystem error handling to
1785 * preserve the consistency of its metadata.
1789 xfs_buftarg_t *target,
1794 LIST_HEAD(tmp_list);
1795 LIST_HEAD(wait_list);
1797 xfs_buf_runall_queues(xfsconvertd_workqueue);
1798 xfs_buf_runall_queues(xfsdatad_workqueue);
1799 xfs_buf_runall_queues(xfslogd_workqueue);
1801 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1802 pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1805 * Dropped the delayed write list lock, now walk the temporary list.
1806 * All I/O is issued async and then if we need to wait for completion
1807 * we do that after issuing all the IO.
1809 list_sort(NULL, &tmp_list, xfs_buf_cmp);
1810 while (!list_empty(&tmp_list)) {
1811 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1812 ASSERT(target == bp->b_target);
1813 list_del_init(&bp->b_list);
1815 bp->b_flags &= ~XBF_ASYNC;
1816 list_add(&bp->b_list, &wait_list);
1822 /* Expedite and wait for IO to complete. */
1823 blk_flush_plug(current);
1824 while (!list_empty(&wait_list)) {
1825 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1827 list_del_init(&bp->b_list);
1839 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1840 KM_ZONE_HWALIGN, NULL);
1844 xfslogd_workqueue = alloc_workqueue("xfslogd",
1845 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1846 if (!xfslogd_workqueue)
1847 goto out_free_buf_zone;
1849 xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1850 if (!xfsdatad_workqueue)
1851 goto out_destroy_xfslogd_workqueue;
1853 xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1855 if (!xfsconvertd_workqueue)
1856 goto out_destroy_xfsdatad_workqueue;
1860 out_destroy_xfsdatad_workqueue:
1861 destroy_workqueue(xfsdatad_workqueue);
1862 out_destroy_xfslogd_workqueue:
1863 destroy_workqueue(xfslogd_workqueue);
1865 kmem_zone_destroy(xfs_buf_zone);
1871 xfs_buf_terminate(void)
1873 destroy_workqueue(xfsconvertd_workqueue);
1874 destroy_workqueue(xfsdatad_workqueue);
1875 destroy_workqueue(xfslogd_workqueue);
1876 kmem_zone_destroy(xfs_buf_zone);
1879 #ifdef CONFIG_KDB_MODULES
1881 xfs_get_buftarg_list(void)
1883 return &xfs_buftarg_list;