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 *);
46 STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
48 static struct workqueue_struct *xfslogd_workqueue;
49 struct workqueue_struct *xfsdatad_workqueue;
50 struct workqueue_struct *xfsconvertd_workqueue;
52 #ifdef XFS_BUF_LOCK_TRACKING
53 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
54 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
55 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
57 # define XB_SET_OWNER(bp) do { } while (0)
58 # define XB_CLEAR_OWNER(bp) do { } while (0)
59 # define XB_GET_OWNER(bp) do { } while (0)
62 #define xb_to_gfp(flags) \
63 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
64 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
66 #define xb_to_km(flags) \
67 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
69 #define xfs_buf_allocate(flags) \
70 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
71 #define xfs_buf_deallocate(bp) \
72 kmem_zone_free(xfs_buf_zone, (bp));
79 * Return true if the buffer is vmapped.
81 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
82 * code is clever enough to know it doesn't have to map a single page,
83 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
85 return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
92 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
96 * xfs_buf_lru_add - add a buffer to the LRU.
98 * The LRU takes a new reference to the buffer so that it will only be freed
99 * once the shrinker takes the buffer off the LRU.
105 struct xfs_buftarg *btp = bp->b_target;
107 spin_lock(&btp->bt_lru_lock);
108 if (list_empty(&bp->b_lru)) {
109 atomic_inc(&bp->b_hold);
110 list_add_tail(&bp->b_lru, &btp->bt_lru);
113 spin_unlock(&btp->bt_lru_lock);
117 * xfs_buf_lru_del - remove a buffer from the LRU
119 * The unlocked check is safe here because it only occurs when there are not
120 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
121 * to optimise the shrinker removing the buffer from the LRU and calling
122 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
129 struct xfs_buftarg *btp = bp->b_target;
131 if (list_empty(&bp->b_lru))
134 spin_lock(&btp->bt_lru_lock);
135 if (!list_empty(&bp->b_lru)) {
136 list_del_init(&bp->b_lru);
139 spin_unlock(&btp->bt_lru_lock);
143 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
144 * b_lru_ref count so that the buffer is freed immediately when the buffer
145 * reference count falls to zero. If the buffer is already on the LRU, we need
146 * to remove the reference that LRU holds on the buffer.
148 * This prevents build-up of stale buffers on the LRU.
154 bp->b_flags |= XBF_STALE;
155 atomic_set(&(bp)->b_lru_ref, 0);
156 if (!list_empty(&bp->b_lru)) {
157 struct xfs_buftarg *btp = bp->b_target;
159 spin_lock(&btp->bt_lru_lock);
160 if (!list_empty(&bp->b_lru)) {
161 list_del_init(&bp->b_lru);
163 atomic_dec(&bp->b_hold);
165 spin_unlock(&btp->bt_lru_lock);
167 ASSERT(atomic_read(&bp->b_hold) >= 1);
173 xfs_buftarg_t *target,
174 xfs_off_t range_base,
176 xfs_buf_flags_t flags)
179 * We don't want certain flags to appear in b_flags.
181 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
183 memset(bp, 0, sizeof(xfs_buf_t));
184 atomic_set(&bp->b_hold, 1);
185 atomic_set(&bp->b_lru_ref, 1);
186 init_completion(&bp->b_iowait);
187 INIT_LIST_HEAD(&bp->b_lru);
188 INIT_LIST_HEAD(&bp->b_list);
189 RB_CLEAR_NODE(&bp->b_rbnode);
190 sema_init(&bp->b_sema, 0); /* held, no waiters */
192 bp->b_target = target;
193 bp->b_file_offset = range_base;
195 * Set buffer_length and count_desired to the same value initially.
196 * I/O routines should use count_desired, which will be the same in
197 * most cases but may be reset (e.g. XFS recovery).
199 bp->b_buffer_length = bp->b_count_desired = range_length;
201 bp->b_bn = XFS_BUF_DADDR_NULL;
202 atomic_set(&bp->b_pin_count, 0);
203 init_waitqueue_head(&bp->b_waiters);
205 XFS_STATS_INC(xb_create);
207 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 xfs_buf_deallocate(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. If other overlapping buffers exist, they are
420 * released before the new buffer is created and locked,
421 * which may imply that this call will block until those buffers
422 * are unlocked. No I/O is implied by this call.
426 xfs_buftarg_t *btp, /* block device target */
427 xfs_off_t ioff, /* starting offset of range */
428 size_t isize, /* length of range */
429 xfs_buf_flags_t flags,
432 xfs_off_t range_base;
434 struct xfs_perag *pag;
435 struct rb_node **rbp;
436 struct rb_node *parent;
439 range_base = (ioff << BBSHIFT);
440 range_length = (isize << BBSHIFT);
442 /* Check for IOs smaller than the sector size / not sector aligned */
443 ASSERT(!(range_length < (1 << btp->bt_sshift)));
444 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
447 pag = xfs_perag_get(btp->bt_mount,
448 xfs_daddr_to_agno(btp->bt_mount, ioff));
451 spin_lock(&pag->pag_buf_lock);
452 rbp = &pag->pag_buf_tree.rb_node;
457 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
459 if (range_base < bp->b_file_offset)
460 rbp = &(*rbp)->rb_left;
461 else if (range_base > bp->b_file_offset)
462 rbp = &(*rbp)->rb_right;
465 * found a block offset match. If the range doesn't
466 * match, the only way this is allowed is if the buffer
467 * in the cache is stale and the transaction that made
468 * it stale has not yet committed. i.e. we are
469 * reallocating a busy extent. Skip this buffer and
470 * continue searching to the right for an exact match.
472 if (bp->b_buffer_length != range_length) {
473 ASSERT(bp->b_flags & XBF_STALE);
474 rbp = &(*rbp)->rb_right;
477 atomic_inc(&bp->b_hold);
484 _xfs_buf_initialize(new_bp, btp, range_base,
485 range_length, flags);
486 rb_link_node(&new_bp->b_rbnode, parent, rbp);
487 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
488 /* the buffer keeps the perag reference until it is freed */
490 spin_unlock(&pag->pag_buf_lock);
492 XFS_STATS_INC(xb_miss_locked);
493 spin_unlock(&pag->pag_buf_lock);
499 spin_unlock(&pag->pag_buf_lock);
502 if (xfs_buf_cond_lock(bp)) {
503 /* failed, so wait for the lock if requested. */
504 if (!(flags & XBF_TRYLOCK)) {
506 XFS_STATS_INC(xb_get_locked_waited);
509 XFS_STATS_INC(xb_busy_locked);
515 * if the buffer is stale, clear all the external state associated with
516 * it. We need to keep flags such as how we allocated the buffer memory
519 if (bp->b_flags & XBF_STALE) {
520 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
521 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
524 trace_xfs_buf_find(bp, flags, _RET_IP_);
525 XFS_STATS_INC(xb_get_locked);
530 * Assembles a buffer covering the specified range.
531 * Storage in memory for all portions of the buffer will be allocated,
532 * although backing storage may not be.
536 xfs_buftarg_t *target,/* target for buffer */
537 xfs_off_t ioff, /* starting offset of range */
538 size_t isize, /* length of range */
539 xfs_buf_flags_t flags)
541 xfs_buf_t *bp, *new_bp;
544 new_bp = xfs_buf_allocate(flags);
545 if (unlikely(!new_bp))
548 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
550 error = xfs_buf_allocate_memory(bp, flags);
554 xfs_buf_deallocate(new_bp);
555 if (unlikely(bp == NULL))
559 if (!(bp->b_flags & XBF_MAPPED)) {
560 error = _xfs_buf_map_pages(bp, flags);
561 if (unlikely(error)) {
562 xfs_warn(target->bt_mount,
563 "%s: failed to map pages\n", __func__);
568 XFS_STATS_INC(xb_get);
571 * Always fill in the block number now, the mapped cases can do
572 * their own overlay of this later.
575 bp->b_count_desired = bp->b_buffer_length;
577 trace_xfs_buf_get(bp, flags, _RET_IP_);
581 if (flags & (XBF_LOCK | XBF_TRYLOCK))
590 xfs_buf_flags_t flags)
594 ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
595 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
597 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | \
598 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
599 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | \
600 XBF_READ_AHEAD | _XBF_RUN_QUEUES);
602 status = xfs_buf_iorequest(bp);
603 if (status || XFS_BUF_ISERROR(bp) || (flags & XBF_ASYNC))
605 return xfs_buf_iowait(bp);
610 xfs_buftarg_t *target,
613 xfs_buf_flags_t flags)
619 bp = xfs_buf_get(target, ioff, isize, flags);
621 trace_xfs_buf_read(bp, flags, _RET_IP_);
623 if (!XFS_BUF_ISDONE(bp)) {
624 XFS_STATS_INC(xb_get_read);
625 _xfs_buf_read(bp, flags);
626 } else if (flags & XBF_ASYNC) {
628 * Read ahead call which is already satisfied,
633 /* We do not want read in the flags */
634 bp->b_flags &= ~XBF_READ;
641 if (flags & (XBF_LOCK | XBF_TRYLOCK))
648 * If we are not low on memory then do the readahead in a deadlock
653 xfs_buftarg_t *target,
657 if (bdi_read_congested(target->bt_bdi))
660 xfs_buf_read(target, ioff, isize,
661 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
665 * Read an uncached buffer from disk. Allocates and returns a locked
666 * buffer containing the disk contents or nothing.
669 xfs_buf_read_uncached(
670 struct xfs_mount *mp,
671 struct xfs_buftarg *target,
679 bp = xfs_buf_get_uncached(target, length, flags);
683 /* set up the buffer for a read IO */
685 XFS_BUF_SET_ADDR(bp, daddr);
690 error = xfs_buf_iowait(bp);
691 if (error || bp->b_error) {
701 xfs_buftarg_t *target)
705 bp = xfs_buf_allocate(0);
707 _xfs_buf_initialize(bp, target, 0, len, 0);
712 * Return a buffer allocated as an empty buffer and associated to external
713 * memory via xfs_buf_associate_memory() back to it's empty state.
721 _xfs_buf_free_pages(bp);
724 bp->b_page_count = 0;
726 bp->b_file_offset = 0;
727 bp->b_buffer_length = bp->b_count_desired = len;
728 bp->b_bn = XFS_BUF_DADDR_NULL;
729 bp->b_flags &= ~XBF_MAPPED;
732 static inline struct page *
736 if ((!is_vmalloc_addr(addr))) {
737 return virt_to_page(addr);
739 return vmalloc_to_page(addr);
744 xfs_buf_associate_memory(
751 unsigned long pageaddr;
752 unsigned long offset;
756 pageaddr = (unsigned long)mem & PAGE_MASK;
757 offset = (unsigned long)mem - pageaddr;
758 buflen = PAGE_ALIGN(len + offset);
759 page_count = buflen >> PAGE_SHIFT;
761 /* Free any previous set of page pointers */
763 _xfs_buf_free_pages(bp);
768 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
772 bp->b_offset = offset;
774 for (i = 0; i < bp->b_page_count; i++) {
775 bp->b_pages[i] = mem_to_page((void *)pageaddr);
776 pageaddr += PAGE_SIZE;
779 bp->b_count_desired = len;
780 bp->b_buffer_length = buflen;
781 bp->b_flags |= XBF_MAPPED;
787 xfs_buf_get_uncached(
788 struct xfs_buftarg *target,
792 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
796 bp = xfs_buf_allocate(0);
797 if (unlikely(bp == NULL))
799 _xfs_buf_initialize(bp, target, 0, len, 0);
801 error = _xfs_buf_get_pages(bp, page_count, 0);
805 for (i = 0; i < page_count; i++) {
806 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
810 bp->b_flags |= _XBF_PAGES;
812 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
813 if (unlikely(error)) {
814 xfs_warn(target->bt_mount,
815 "%s: failed to map pages\n", __func__);
821 trace_xfs_buf_get_uncached(bp, _RET_IP_);
826 __free_page(bp->b_pages[i]);
827 _xfs_buf_free_pages(bp);
829 xfs_buf_deallocate(bp);
835 * Increment reference count on buffer, to hold the buffer concurrently
836 * with another thread which may release (free) the buffer asynchronously.
837 * Must hold the buffer already to call this function.
843 trace_xfs_buf_hold(bp, _RET_IP_);
844 atomic_inc(&bp->b_hold);
848 * Releases a hold on the specified buffer. If the
849 * the hold count is 1, calls xfs_buf_free.
855 struct xfs_perag *pag = bp->b_pag;
857 trace_xfs_buf_rele(bp, _RET_IP_);
860 ASSERT(list_empty(&bp->b_lru));
861 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
862 if (atomic_dec_and_test(&bp->b_hold))
867 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
869 ASSERT(atomic_read(&bp->b_hold) > 0);
870 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
871 if (!(bp->b_flags & XBF_STALE) &&
872 atomic_read(&bp->b_lru_ref)) {
874 spin_unlock(&pag->pag_buf_lock);
877 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
878 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
879 spin_unlock(&pag->pag_buf_lock);
888 * Lock a buffer object, if it is not already locked.
890 * If we come across a stale, pinned, locked buffer, we know that we are
891 * being asked to lock a buffer that has been reallocated. Because it is
892 * pinned, we know that the log has not been pushed to disk and hence it
893 * will still be locked. Rather than continuing to have trylock attempts
894 * fail until someone else pushes the log, push it ourselves before
895 * returning. This means that the xfsaild will not get stuck trying
896 * to push on stale inode buffers.
904 locked = down_trylock(&bp->b_sema) == 0;
907 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
908 xfs_log_force(bp->b_target->bt_mount, 0);
910 trace_xfs_buf_cond_lock(bp, _RET_IP_);
911 return locked ? 0 : -EBUSY;
918 return bp->b_sema.count;
922 * Lock a buffer object.
924 * If we come across a stale, pinned, locked buffer, we know that we
925 * are being asked to lock a buffer that has been reallocated. Because
926 * it is pinned, we know that the log has not been pushed to disk and
927 * hence it will still be locked. Rather than sleeping until someone
928 * else pushes the log, push it ourselves before trying to get the lock.
934 trace_xfs_buf_lock(bp, _RET_IP_);
936 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
937 xfs_log_force(bp->b_target->bt_mount, 0);
941 trace_xfs_buf_lock_done(bp, _RET_IP_);
945 * Releases the lock on the buffer object.
946 * If the buffer is marked delwri but is not queued, do so before we
947 * unlock the buffer as we need to set flags correctly. We also need to
948 * take a reference for the delwri queue because the unlocker is going to
949 * drop their's and they don't know we just queued it.
955 if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
956 atomic_inc(&bp->b_hold);
957 bp->b_flags |= XBF_ASYNC;
958 xfs_buf_delwri_queue(bp, 0);
964 trace_xfs_buf_unlock(bp, _RET_IP_);
971 DECLARE_WAITQUEUE (wait, current);
973 if (atomic_read(&bp->b_pin_count) == 0)
976 add_wait_queue(&bp->b_waiters, &wait);
978 set_current_state(TASK_UNINTERRUPTIBLE);
979 if (atomic_read(&bp->b_pin_count) == 0)
983 remove_wait_queue(&bp->b_waiters, &wait);
984 set_current_state(TASK_RUNNING);
988 * Buffer Utility Routines
993 struct work_struct *work)
996 container_of(work, xfs_buf_t, b_iodone_work);
999 (*(bp->b_iodone))(bp);
1000 else if (bp->b_flags & XBF_ASYNC)
1009 trace_xfs_buf_iodone(bp, _RET_IP_);
1011 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1012 if (bp->b_error == 0)
1013 bp->b_flags |= XBF_DONE;
1015 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1017 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1018 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1020 xfs_buf_iodone_work(&bp->b_iodone_work);
1023 complete(&bp->b_iowait);
1032 ASSERT(error >= 0 && error <= 0xffff);
1033 bp->b_error = (unsigned short)error;
1034 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1039 struct xfs_mount *mp,
1044 bp->b_flags |= XBF_WRITE;
1045 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1047 xfs_buf_delwri_dequeue(bp);
1050 error = xfs_buf_iowait(bp);
1052 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1062 trace_xfs_buf_bdwrite(bp, _RET_IP_);
1064 bp->b_flags &= ~XBF_READ;
1065 bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1067 xfs_buf_delwri_queue(bp, 1);
1071 * Called when we want to stop a buffer from getting written or read.
1072 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1073 * so that the proper iodone callbacks get called.
1079 #ifdef XFSERRORDEBUG
1080 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1084 * No need to wait until the buffer is unpinned, we aren't flushing it.
1086 XFS_BUF_ERROR(bp, EIO);
1089 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1092 XFS_BUF_UNDELAYWRITE(bp);
1096 xfs_buf_ioend(bp, 0);
1102 * Same as xfs_bioerror, except that we are releasing the buffer
1103 * here ourselves, and avoiding the xfs_buf_ioend call.
1104 * This is meant for userdata errors; metadata bufs come with
1105 * iodone functions attached, so that we can track down errors.
1111 int64_t fl = XFS_BUF_BFLAGS(bp);
1113 * No need to wait until the buffer is unpinned.
1114 * We aren't flushing it.
1116 * chunkhold expects B_DONE to be set, whether
1117 * we actually finish the I/O or not. We don't want to
1118 * change that interface.
1121 XFS_BUF_UNDELAYWRITE(bp);
1124 XFS_BUF_CLR_IODONE_FUNC(bp);
1125 if (!(fl & XBF_ASYNC)) {
1127 * Mark b_error and B_ERROR _both_.
1128 * Lot's of chunkcache code assumes that.
1129 * There's no reason to mark error for
1132 XFS_BUF_ERROR(bp, EIO);
1133 XFS_BUF_FINISH_IOWAIT(bp);
1143 * All xfs metadata buffers except log state machine buffers
1144 * get this attached as their b_bdstrat callback function.
1145 * This is so that we can catch a buffer
1146 * after prematurely unpinning it to forcibly shutdown the filesystem.
1152 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1153 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1155 * Metadata write that didn't get logged but
1156 * written delayed anyway. These aren't associated
1157 * with a transaction, and can be ignored.
1159 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1160 return xfs_bioerror_relse(bp);
1162 return xfs_bioerror(bp);
1165 xfs_buf_iorequest(bp);
1170 * Wrapper around bdstrat so that we can stop data from going to disk in case
1171 * we are shutting down the filesystem. Typically user data goes thru this
1172 * path; one of the exceptions is the superblock.
1176 struct xfs_mount *mp,
1179 if (XFS_FORCED_SHUTDOWN(mp)) {
1180 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1181 xfs_bioerror_relse(bp);
1185 xfs_buf_iorequest(bp);
1193 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1194 xfs_buf_ioend(bp, schedule);
1202 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1204 xfs_buf_ioerror(bp, -error);
1206 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1207 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1209 _xfs_buf_ioend(bp, 1);
1217 int rw, map_i, total_nr_pages, nr_pages;
1219 int offset = bp->b_offset;
1220 int size = bp->b_count_desired;
1221 sector_t sector = bp->b_bn;
1223 total_nr_pages = bp->b_page_count;
1226 if (bp->b_flags & XBF_ORDERED) {
1227 ASSERT(!(bp->b_flags & XBF_READ));
1228 rw = WRITE_FLUSH_FUA;
1229 } else if (bp->b_flags & XBF_LOG_BUFFER) {
1230 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1231 bp->b_flags &= ~_XBF_RUN_QUEUES;
1232 rw = (bp->b_flags & XBF_WRITE) ? WRITE_SYNC : READ_SYNC;
1233 } else if (bp->b_flags & _XBF_RUN_QUEUES) {
1234 ASSERT(!(bp->b_flags & XBF_READ_AHEAD));
1235 bp->b_flags &= ~_XBF_RUN_QUEUES;
1236 rw = (bp->b_flags & XBF_WRITE) ? WRITE_META : READ_META;
1238 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1239 (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1244 atomic_inc(&bp->b_io_remaining);
1245 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1246 if (nr_pages > total_nr_pages)
1247 nr_pages = total_nr_pages;
1249 bio = bio_alloc(GFP_NOIO, nr_pages);
1250 bio->bi_bdev = bp->b_target->bt_bdev;
1251 bio->bi_sector = sector;
1252 bio->bi_end_io = xfs_buf_bio_end_io;
1253 bio->bi_private = bp;
1256 for (; size && nr_pages; nr_pages--, map_i++) {
1257 int rbytes, nbytes = PAGE_SIZE - offset;
1262 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1263 if (rbytes < nbytes)
1267 sector += nbytes >> BBSHIFT;
1272 if (likely(bio->bi_size)) {
1273 if (xfs_buf_is_vmapped(bp)) {
1274 flush_kernel_vmap_range(bp->b_addr,
1275 xfs_buf_vmap_len(bp));
1277 submit_bio(rw, bio);
1281 xfs_buf_ioerror(bp, EIO);
1290 trace_xfs_buf_iorequest(bp, _RET_IP_);
1292 if (bp->b_flags & XBF_DELWRI) {
1293 xfs_buf_delwri_queue(bp, 1);
1297 if (bp->b_flags & XBF_WRITE) {
1298 xfs_buf_wait_unpin(bp);
1303 /* Set the count to 1 initially, this will stop an I/O
1304 * completion callout which happens before we have started
1305 * all the I/O from calling xfs_buf_ioend too early.
1307 atomic_set(&bp->b_io_remaining, 1);
1308 _xfs_buf_ioapply(bp);
1309 _xfs_buf_ioend(bp, 0);
1316 * Waits for I/O to complete on the buffer supplied.
1317 * It returns immediately if no I/O is pending.
1318 * It returns the I/O error code, if any, or 0 if there was no error.
1324 trace_xfs_buf_iowait(bp, _RET_IP_);
1326 wait_for_completion(&bp->b_iowait);
1328 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1339 if (bp->b_flags & XBF_MAPPED)
1340 return XFS_BUF_PTR(bp) + offset;
1342 offset += bp->b_offset;
1343 page = bp->b_pages[offset >> PAGE_SHIFT];
1344 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1348 * Move data into or out of a buffer.
1352 xfs_buf_t *bp, /* buffer to process */
1353 size_t boff, /* starting buffer offset */
1354 size_t bsize, /* length to copy */
1355 void *data, /* data address */
1356 xfs_buf_rw_t mode) /* read/write/zero flag */
1358 size_t bend, cpoff, csize;
1361 bend = boff + bsize;
1362 while (boff < bend) {
1363 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1364 cpoff = xfs_buf_poff(boff + bp->b_offset);
1365 csize = min_t(size_t,
1366 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1368 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1372 memset(page_address(page) + cpoff, 0, csize);
1375 memcpy(data, page_address(page) + cpoff, csize);
1378 memcpy(page_address(page) + cpoff, data, csize);
1387 * Handling of buffer targets (buftargs).
1391 * Wait for any bufs with callbacks that have been submitted but have not yet
1392 * returned. These buffers will have an elevated hold count, so wait on those
1393 * while freeing all the buffers only held by the LRU.
1397 struct xfs_buftarg *btp)
1402 spin_lock(&btp->bt_lru_lock);
1403 while (!list_empty(&btp->bt_lru)) {
1404 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1405 if (atomic_read(&bp->b_hold) > 1) {
1406 spin_unlock(&btp->bt_lru_lock);
1411 * clear the LRU reference count so the bufer doesn't get
1412 * ignored in xfs_buf_rele().
1414 atomic_set(&bp->b_lru_ref, 0);
1415 spin_unlock(&btp->bt_lru_lock);
1417 spin_lock(&btp->bt_lru_lock);
1419 spin_unlock(&btp->bt_lru_lock);
1424 struct shrinker *shrink,
1425 struct shrink_control *sc)
1427 struct xfs_buftarg *btp = container_of(shrink,
1428 struct xfs_buftarg, bt_shrinker);
1430 int nr_to_scan = sc->nr_to_scan;
1434 return btp->bt_lru_nr;
1436 spin_lock(&btp->bt_lru_lock);
1437 while (!list_empty(&btp->bt_lru)) {
1438 if (nr_to_scan-- <= 0)
1441 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1444 * Decrement the b_lru_ref count unless the value is already
1445 * zero. If the value is already zero, we need to reclaim the
1446 * buffer, otherwise it gets another trip through the LRU.
1448 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1449 list_move_tail(&bp->b_lru, &btp->bt_lru);
1454 * remove the buffer from the LRU now to avoid needing another
1455 * lock round trip inside xfs_buf_rele().
1457 list_move(&bp->b_lru, &dispose);
1460 spin_unlock(&btp->bt_lru_lock);
1462 while (!list_empty(&dispose)) {
1463 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1464 list_del_init(&bp->b_lru);
1468 return btp->bt_lru_nr;
1473 struct xfs_mount *mp,
1474 struct xfs_buftarg *btp)
1476 unregister_shrinker(&btp->bt_shrinker);
1478 xfs_flush_buftarg(btp, 1);
1479 if (mp->m_flags & XFS_MOUNT_BARRIER)
1480 xfs_blkdev_issue_flush(btp);
1482 kthread_stop(btp->bt_task);
1487 xfs_setsize_buftarg_flags(
1489 unsigned int blocksize,
1490 unsigned int sectorsize,
1493 btp->bt_bsize = blocksize;
1494 btp->bt_sshift = ffs(sectorsize) - 1;
1495 btp->bt_smask = sectorsize - 1;
1497 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1498 xfs_warn(btp->bt_mount,
1499 "Cannot set_blocksize to %u on device %s\n",
1500 sectorsize, XFS_BUFTARG_NAME(btp));
1508 * When allocating the initial buffer target we have not yet
1509 * read in the superblock, so don't know what sized sectors
1510 * are being used is at this early stage. Play safe.
1513 xfs_setsize_buftarg_early(
1515 struct block_device *bdev)
1517 return xfs_setsize_buftarg_flags(btp,
1518 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1522 xfs_setsize_buftarg(
1524 unsigned int blocksize,
1525 unsigned int sectorsize)
1527 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1531 xfs_alloc_delwrite_queue(
1535 INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1536 spin_lock_init(&btp->bt_delwrite_lock);
1538 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1539 if (IS_ERR(btp->bt_task))
1540 return PTR_ERR(btp->bt_task);
1546 struct xfs_mount *mp,
1547 struct block_device *bdev,
1553 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1556 btp->bt_dev = bdev->bd_dev;
1557 btp->bt_bdev = bdev;
1558 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1562 INIT_LIST_HEAD(&btp->bt_lru);
1563 spin_lock_init(&btp->bt_lru_lock);
1564 if (xfs_setsize_buftarg_early(btp, bdev))
1566 if (xfs_alloc_delwrite_queue(btp, fsname))
1568 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1569 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1570 register_shrinker(&btp->bt_shrinker);
1580 * Delayed write buffer handling
1583 xfs_buf_delwri_queue(
1587 struct list_head *dwq = &bp->b_target->bt_delwrite_queue;
1588 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1590 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1592 ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1595 /* If already in the queue, dequeue and place at tail */
1596 if (!list_empty(&bp->b_list)) {
1597 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1599 atomic_dec(&bp->b_hold);
1600 list_del(&bp->b_list);
1603 if (list_empty(dwq)) {
1604 /* start xfsbufd as it is about to have something to do */
1605 wake_up_process(bp->b_target->bt_task);
1608 bp->b_flags |= _XBF_DELWRI_Q;
1609 list_add_tail(&bp->b_list, dwq);
1610 bp->b_queuetime = jiffies;
1618 xfs_buf_delwri_dequeue(
1621 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1625 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1626 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1627 list_del_init(&bp->b_list);
1630 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1636 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1640 * If a delwri buffer needs to be pushed before it has aged out, then promote
1641 * it to the head of the delwri queue so that it will be flushed on the next
1642 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1643 * than the age currently needed to flush the buffer. Hence the next time the
1644 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1647 xfs_buf_delwri_promote(
1650 struct xfs_buftarg *btp = bp->b_target;
1651 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1653 ASSERT(bp->b_flags & XBF_DELWRI);
1654 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1657 * Check the buffer age before locking the delayed write queue as we
1658 * don't need to promote buffers that are already past the flush age.
1660 if (bp->b_queuetime < jiffies - age)
1662 bp->b_queuetime = jiffies - age;
1663 spin_lock(&btp->bt_delwrite_lock);
1664 list_move(&bp->b_list, &btp->bt_delwrite_queue);
1665 spin_unlock(&btp->bt_delwrite_lock);
1669 xfs_buf_runall_queues(
1670 struct workqueue_struct *queue)
1672 flush_workqueue(queue);
1676 * Move as many buffers as specified to the supplied list
1677 * idicating if we skipped any buffers to prevent deadlocks.
1680 xfs_buf_delwri_split(
1681 xfs_buftarg_t *target,
1682 struct list_head *list,
1686 struct list_head *dwq = &target->bt_delwrite_queue;
1687 spinlock_t *dwlk = &target->bt_delwrite_lock;
1691 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1692 INIT_LIST_HEAD(list);
1694 list_for_each_entry_safe(bp, n, dwq, b_list) {
1695 ASSERT(bp->b_flags & XBF_DELWRI);
1697 if (!XFS_BUF_ISPINNED(bp) && !xfs_buf_cond_lock(bp)) {
1699 time_before(jiffies, bp->b_queuetime + age)) {
1704 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1706 bp->b_flags |= XBF_WRITE;
1707 list_move_tail(&bp->b_list, list);
1708 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1719 * Compare function is more complex than it needs to be because
1720 * the return value is only 32 bits and we are doing comparisons
1726 struct list_head *a,
1727 struct list_head *b)
1729 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1730 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1733 diff = ap->b_bn - bp->b_bn;
1742 xfs_buf_delwri_sort(
1743 xfs_buftarg_t *target,
1744 struct list_head *list)
1746 list_sort(NULL, list, xfs_buf_cmp);
1753 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1755 current->flags |= PF_MEMALLOC;
1760 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1761 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1762 struct list_head tmp;
1763 struct blk_plug plug;
1765 if (unlikely(freezing(current))) {
1766 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1769 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1772 /* sleep for a long time if there is nothing to do. */
1773 if (list_empty(&target->bt_delwrite_queue))
1774 tout = MAX_SCHEDULE_TIMEOUT;
1775 schedule_timeout_interruptible(tout);
1777 xfs_buf_delwri_split(target, &tmp, age);
1778 list_sort(NULL, &tmp, xfs_buf_cmp);
1780 blk_start_plug(&plug);
1781 while (!list_empty(&tmp)) {
1783 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1784 list_del_init(&bp->b_list);
1787 blk_finish_plug(&plug);
1788 } while (!kthread_should_stop());
1794 * Go through all incore buffers, and release buffers if they belong to
1795 * the given device. This is used in filesystem error handling to
1796 * preserve the consistency of its metadata.
1800 xfs_buftarg_t *target,
1805 LIST_HEAD(tmp_list);
1806 LIST_HEAD(wait_list);
1807 struct blk_plug plug;
1809 xfs_buf_runall_queues(xfsconvertd_workqueue);
1810 xfs_buf_runall_queues(xfsdatad_workqueue);
1811 xfs_buf_runall_queues(xfslogd_workqueue);
1813 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1814 pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1817 * Dropped the delayed write list lock, now walk the temporary list.
1818 * All I/O is issued async and then if we need to wait for completion
1819 * we do that after issuing all the IO.
1821 list_sort(NULL, &tmp_list, xfs_buf_cmp);
1823 blk_start_plug(&plug);
1824 while (!list_empty(&tmp_list)) {
1825 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1826 ASSERT(target == bp->b_target);
1827 list_del_init(&bp->b_list);
1829 bp->b_flags &= ~XBF_ASYNC;
1830 list_add(&bp->b_list, &wait_list);
1834 blk_finish_plug(&plug);
1837 /* Wait for IO to complete. */
1838 while (!list_empty(&wait_list)) {
1839 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1841 list_del_init(&bp->b_list);
1853 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1854 KM_ZONE_HWALIGN, NULL);
1858 xfslogd_workqueue = alloc_workqueue("xfslogd",
1859 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1860 if (!xfslogd_workqueue)
1861 goto out_free_buf_zone;
1863 xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1864 if (!xfsdatad_workqueue)
1865 goto out_destroy_xfslogd_workqueue;
1867 xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1869 if (!xfsconvertd_workqueue)
1870 goto out_destroy_xfsdatad_workqueue;
1874 out_destroy_xfsdatad_workqueue:
1875 destroy_workqueue(xfsdatad_workqueue);
1876 out_destroy_xfslogd_workqueue:
1877 destroy_workqueue(xfslogd_workqueue);
1879 kmem_zone_destroy(xfs_buf_zone);
1885 xfs_buf_terminate(void)
1887 destroy_workqueue(xfsconvertd_workqueue);
1888 destroy_workqueue(xfsdatad_workqueue);
1889 destroy_workqueue(xfslogd_workqueue);
1890 kmem_zone_destroy(xfs_buf_zone);
1893 #ifdef CONFIG_KDB_MODULES
1895 xfs_get_buftarg_list(void)
1897 return &xfs_buftarg_list;