Merge branch 'for-linus' of git://oss.sgi.com/xfs/xfs
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / fs / xfs / linux-2.6 / xfs_buf.c
blob596bb2c9de429f76e5c1642aa545906a673956a2
1 /*
2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3 * All Rights Reserved.
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
18 #include "xfs.h"
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>
38 #include "xfs_sb.h"
39 #include "xfs_inum.h"
40 #include "xfs_log.h"
41 #include "xfs_ag.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)
57 #else
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)
61 #endif
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));
75 static inline int
76 xfs_buf_is_vmapped(
77 struct xfs_buf *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;
89 static inline int
90 xfs_buf_vmap_len(
91 struct xfs_buf *bp)
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.
102 STATIC void
103 xfs_buf_lru_add(
104 struct xfs_buf *bp)
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);
112 btp->bt_lru_nr++;
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
124 * bt_lru_lock.
126 STATIC void
127 xfs_buf_lru_del(
128 struct xfs_buf *bp)
130 struct xfs_buftarg *btp = bp->b_target;
132 if (list_empty(&bp->b_lru))
133 return;
135 spin_lock(&btp->bt_lru_lock);
136 if (!list_empty(&bp->b_lru)) {
137 list_del_init(&bp->b_lru);
138 btp->bt_lru_nr--;
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.
151 void
152 xfs_buf_stale(
153 struct xfs_buf *bp)
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);
163 btp->bt_lru_nr--;
164 atomic_dec(&bp->b_hold);
166 spin_unlock(&btp->bt_lru_lock);
168 ASSERT(atomic_read(&bp->b_hold) >= 1);
171 STATIC void
172 _xfs_buf_initialize(
173 xfs_buf_t *bp,
174 xfs_buftarg_t *target,
175 xfs_off_t range_base,
176 size_t range_length,
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 */
192 XB_SET_OWNER(bp);
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;
201 bp->b_flags = flags;
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.
215 STATIC int
216 _xfs_buf_get_pages(
217 xfs_buf_t *bp,
218 int page_count,
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;
227 } else {
228 bp->b_pages = kmem_alloc(sizeof(struct page *) *
229 page_count, xb_to_km(flags));
230 if (bp->b_pages == NULL)
231 return -ENOMEM;
233 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
235 return 0;
239 * Frees b_pages if it was allocated.
241 STATIC void
242 _xfs_buf_free_pages(
243 xfs_buf_t *bp)
245 if (bp->b_pages != bp->b_page_array) {
246 kmem_free(bp->b_pages);
247 bp->b_pages = NULL;
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
258 void
259 xfs_buf_free(
260 xfs_buf_t *bp)
262 trace_xfs_buf_free(bp, _RET_IP_);
264 ASSERT(list_empty(&bp->b_lru));
266 if (bp->b_flags & _XBF_PAGES) {
267 uint i;
269 if (xfs_buf_is_vmapped(bp))
270 vm_unmap_ram(bp->b_addr - bp->b_offset,
271 bp->b_page_count);
273 for (i = 0; i < bp->b_page_count; i++) {
274 struct page *page = bp->b_pages[i];
276 __free_page(page);
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.
287 STATIC int
288 xfs_buf_allocate_memory(
289 xfs_buf_t *bp,
290 uint flags)
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;
296 pgoff_t first;
297 xfs_off_t end;
298 int error;
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));
307 if (!bp->b_addr) {
308 /* low memory - use alloc_page loop instead */
309 goto use_alloc_page;
312 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
313 PAGE_MASK) !=
314 ((unsigned long)bp->b_addr & PAGE_MASK)) {
315 /* b_addr spans two pages - use alloc_page instead */
316 kmem_free(bp->b_addr);
317 bp->b_addr = NULL;
318 goto use_alloc_page;
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;
325 return 0;
328 use_alloc_page:
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);
332 if (unlikely(error))
333 return error;
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++) {
340 struct page *page;
341 uint retries = 0;
342 retry:
343 page = alloc_page(gfp_mask);
344 if (unlikely(page == NULL)) {
345 if (flags & XBF_READ_AHEAD) {
346 bp->b_page_count = i;
347 error = ENOMEM;
348 goto out_free_pages;
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))
358 xfs_err(NULL,
359 "possible memory allocation deadlock in %s (mode:0x%x)",
360 __func__, gfp_mask);
362 XFS_STATS_INC(xb_page_retries);
363 congestion_wait(BLK_RW_ASYNC, HZ/50);
364 goto retry;
367 XFS_STATS_INC(xb_page_found);
369 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
370 size -= nbytes;
371 bp->b_pages[i] = page;
372 offset = 0;
374 return 0;
376 out_free_pages:
377 for (i = 0; i < bp->b_page_count; i++)
378 __free_page(bp->b_pages[i]);
379 return error;
383 * Map buffer into kernel address-space if nessecary.
385 STATIC int
386 _xfs_buf_map_pages(
387 xfs_buf_t *bp,
388 uint flags)
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) {
396 int retried = 0;
398 do {
399 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
400 -1, PAGE_KERNEL);
401 if (bp->b_addr)
402 break;
403 vm_unmap_aliases();
404 } while (retried++ <= 1);
406 if (!bp->b_addr)
407 return -ENOMEM;
408 bp->b_addr += bp->b_offset;
409 bp->b_flags |= XBF_MAPPED;
412 return 0;
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.
427 xfs_buf_t *
428 _xfs_buf_find(
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,
433 xfs_buf_t *new_bp)
435 xfs_off_t range_base;
436 size_t range_length;
437 struct xfs_perag *pag;
438 struct rb_node **rbp;
439 struct rb_node *parent;
440 xfs_buf_t *bp;
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));
449 /* get tree root */
450 pag = xfs_perag_get(btp->bt_mount,
451 xfs_daddr_to_agno(btp->bt_mount, ioff));
453 /* walk tree */
454 spin_lock(&pag->pag_buf_lock);
455 rbp = &pag->pag_buf_tree.rb_node;
456 parent = NULL;
457 bp = NULL;
458 while (*rbp) {
459 parent = *rbp;
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;
466 else {
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;
478 continue;
480 atomic_inc(&bp->b_hold);
481 goto found;
485 /* No match found */
486 if (new_bp) {
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 */
492 new_bp->b_pag = pag;
493 spin_unlock(&pag->pag_buf_lock);
494 } else {
495 XFS_STATS_INC(xb_miss_locked);
496 spin_unlock(&pag->pag_buf_lock);
497 xfs_perag_put(pag);
499 return new_bp;
501 found:
502 spin_unlock(&pag->pag_buf_lock);
503 xfs_perag_put(pag);
505 if (xfs_buf_cond_lock(bp)) {
506 /* failed, so wait for the lock if requested. */
507 if (!(flags & XBF_TRYLOCK)) {
508 xfs_buf_lock(bp);
509 XFS_STATS_INC(xb_get_locked_waited);
510 } else {
511 xfs_buf_rele(bp);
512 XFS_STATS_INC(xb_busy_locked);
513 return NULL;
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
520 * intact here.
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);
529 return bp;
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.
537 xfs_buf_t *
538 xfs_buf_get(
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;
545 int error = 0;
547 new_bp = xfs_buf_allocate(flags);
548 if (unlikely(!new_bp))
549 return NULL;
551 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
552 if (bp == new_bp) {
553 error = xfs_buf_allocate_memory(bp, flags);
554 if (error)
555 goto no_buffer;
556 } else {
557 xfs_buf_deallocate(new_bp);
558 if (unlikely(bp == NULL))
559 return 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__);
567 goto no_buffer;
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.
577 bp->b_bn = ioff;
578 bp->b_count_desired = bp->b_buffer_length;
580 trace_xfs_buf_get(bp, flags, _RET_IP_);
581 return bp;
583 no_buffer:
584 if (flags & (XBF_LOCK | XBF_TRYLOCK))
585 xfs_buf_unlock(bp);
586 xfs_buf_rele(bp);
587 return NULL;
590 STATIC int
591 _xfs_buf_read(
592 xfs_buf_t *bp,
593 xfs_buf_flags_t flags)
595 int status;
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))
607 return status;
608 return xfs_buf_iowait(bp);
611 xfs_buf_t *
612 xfs_buf_read(
613 xfs_buftarg_t *target,
614 xfs_off_t ioff,
615 size_t isize,
616 xfs_buf_flags_t flags)
618 xfs_buf_t *bp;
620 flags |= XBF_READ;
622 bp = xfs_buf_get(target, ioff, isize, flags);
623 if (bp) {
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,
632 * drop the buffer
634 goto no_buffer;
635 } else {
636 /* We do not want read in the flags */
637 bp->b_flags &= ~XBF_READ;
641 return bp;
643 no_buffer:
644 if (flags & (XBF_LOCK | XBF_TRYLOCK))
645 xfs_buf_unlock(bp);
646 xfs_buf_rele(bp);
647 return NULL;
651 * If we are not low on memory then do the readahead in a deadlock
652 * safe manner.
654 void
655 xfs_buf_readahead(
656 xfs_buftarg_t *target,
657 xfs_off_t ioff,
658 size_t isize)
660 struct backing_dev_info *bdi;
662 if (bdi_read_congested(target->bt_bdi))
663 return;
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.
673 struct xfs_buf *
674 xfs_buf_read_uncached(
675 struct xfs_mount *mp,
676 struct xfs_buftarg *target,
677 xfs_daddr_t daddr,
678 size_t length,
679 int flags)
681 xfs_buf_t *bp;
682 int error;
684 bp = xfs_buf_get_uncached(target, length, flags);
685 if (!bp)
686 return NULL;
688 /* set up the buffer for a read IO */
689 xfs_buf_lock(bp);
690 XFS_BUF_SET_ADDR(bp, daddr);
691 XFS_BUF_READ(bp);
692 XFS_BUF_BUSY(bp);
694 xfsbdstrat(mp, bp);
695 error = xfs_buf_iowait(bp);
696 if (error || bp->b_error) {
697 xfs_buf_relse(bp);
698 return NULL;
700 return bp;
703 xfs_buf_t *
704 xfs_buf_get_empty(
705 size_t len,
706 xfs_buftarg_t *target)
708 xfs_buf_t *bp;
710 bp = xfs_buf_allocate(0);
711 if (bp)
712 _xfs_buf_initialize(bp, target, 0, len, 0);
713 return bp;
716 static inline struct page *
717 mem_to_page(
718 void *addr)
720 if ((!is_vmalloc_addr(addr))) {
721 return virt_to_page(addr);
722 } else {
723 return vmalloc_to_page(addr);
728 xfs_buf_associate_memory(
729 xfs_buf_t *bp,
730 void *mem,
731 size_t len)
733 int rval;
734 int i = 0;
735 unsigned long pageaddr;
736 unsigned long offset;
737 size_t buflen;
738 int page_count;
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 */
746 if (bp->b_pages)
747 _xfs_buf_free_pages(bp);
749 bp->b_pages = NULL;
750 bp->b_addr = mem;
752 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
753 if (rval)
754 return rval;
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;
767 return 0;
770 xfs_buf_t *
771 xfs_buf_get_uncached(
772 struct xfs_buftarg *target,
773 size_t len,
774 int flags)
776 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
777 int error, i;
778 xfs_buf_t *bp;
780 bp = xfs_buf_allocate(0);
781 if (unlikely(bp == NULL))
782 goto fail;
783 _xfs_buf_initialize(bp, target, 0, len, 0);
785 error = _xfs_buf_get_pages(bp, page_count, 0);
786 if (error)
787 goto fail_free_buf;
789 for (i = 0; i < page_count; i++) {
790 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
791 if (!bp->b_pages[i])
792 goto fail_free_mem;
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__);
800 goto fail_free_mem;
803 xfs_buf_unlock(bp);
805 trace_xfs_buf_get_uncached(bp, _RET_IP_);
806 return bp;
808 fail_free_mem:
809 while (--i >= 0)
810 __free_page(bp->b_pages[i]);
811 _xfs_buf_free_pages(bp);
812 fail_free_buf:
813 xfs_buf_deallocate(bp);
814 fail:
815 return NULL;
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.
823 void
824 xfs_buf_hold(
825 xfs_buf_t *bp)
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.
835 void
836 xfs_buf_rele(
837 xfs_buf_t *bp)
839 struct xfs_perag *pag = bp->b_pag;
841 trace_xfs_buf_rele(bp, _RET_IP_);
843 if (!pag) {
844 ASSERT(list_empty(&bp->b_lru));
845 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
846 if (atomic_dec_and_test(&bp->b_hold))
847 xfs_buf_free(bp);
848 return;
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)) {
857 xfs_buf_lru_add(bp);
858 spin_unlock(&pag->pag_buf_lock);
859 } else {
860 xfs_buf_lru_del(bp);
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);
864 xfs_perag_put(pag);
865 xfs_buf_free(bp);
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.
883 xfs_buf_cond_lock(
884 xfs_buf_t *bp)
886 int locked;
888 locked = down_trylock(&bp->b_sema) == 0;
889 if (locked)
890 XB_SET_OWNER(bp);
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;
899 xfs_buf_lock_value(
900 xfs_buf_t *bp)
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.
914 void
915 xfs_buf_lock(
916 xfs_buf_t *bp)
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);
924 down(&bp->b_sema);
925 XB_SET_OWNER(bp);
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.
937 void
938 xfs_buf_unlock(
939 xfs_buf_t *bp)
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);
947 XB_CLEAR_OWNER(bp);
948 up(&bp->b_sema);
950 trace_xfs_buf_unlock(bp, _RET_IP_);
953 STATIC void
954 xfs_buf_wait_unpin(
955 xfs_buf_t *bp)
957 DECLARE_WAITQUEUE (wait, current);
959 if (atomic_read(&bp->b_pin_count) == 0)
960 return;
962 add_wait_queue(&bp->b_waiters, &wait);
963 for (;;) {
964 set_current_state(TASK_UNINTERRUPTIBLE);
965 if (atomic_read(&bp->b_pin_count) == 0)
966 break;
967 io_schedule();
969 remove_wait_queue(&bp->b_waiters, &wait);
970 set_current_state(TASK_RUNNING);
974 * Buffer Utility Routines
977 STATIC void
978 xfs_buf_iodone_work(
979 struct work_struct *work)
981 xfs_buf_t *bp =
982 container_of(work, xfs_buf_t, b_iodone_work);
984 if (bp->b_iodone)
985 (*(bp->b_iodone))(bp);
986 else if (bp->b_flags & XBF_ASYNC)
987 xfs_buf_relse(bp);
990 void
991 xfs_buf_ioend(
992 xfs_buf_t *bp,
993 int schedule)
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)) {
1002 if (schedule) {
1003 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1004 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1005 } else {
1006 xfs_buf_iodone_work(&bp->b_iodone_work);
1008 } else {
1009 complete(&bp->b_iowait);
1013 void
1014 xfs_buf_ioerror(
1015 xfs_buf_t *bp,
1016 int error)
1018 ASSERT(error >= 0 && error <= 0xffff);
1019 bp->b_error = (unsigned short)error;
1020 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1024 xfs_bwrite(
1025 struct xfs_mount *mp,
1026 struct xfs_buf *bp)
1028 int error;
1030 bp->b_flags |= XBF_WRITE;
1031 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1033 xfs_buf_delwri_dequeue(bp);
1034 xfs_bdstrat_cb(bp);
1036 error = xfs_buf_iowait(bp);
1037 if (error)
1038 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1039 xfs_buf_relse(bp);
1040 return error;
1043 void
1044 xfs_bdwrite(
1045 void *mp,
1046 struct xfs_buf *bp)
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.
1061 STATIC int
1062 xfs_bioerror(
1063 xfs_buf_t *bp)
1065 #ifdef XFSERRORDEBUG
1066 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1067 #endif
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.
1077 XFS_BUF_UNREAD(bp);
1078 XFS_BUF_UNDELAYWRITE(bp);
1079 XFS_BUF_UNDONE(bp);
1080 XFS_BUF_STALE(bp);
1082 xfs_buf_ioend(bp, 0);
1084 return EIO;
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.
1093 STATIC int
1094 xfs_bioerror_relse(
1095 struct xfs_buf *bp)
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.
1106 XFS_BUF_UNREAD(bp);
1107 XFS_BUF_UNDELAYWRITE(bp);
1108 XFS_BUF_DONE(bp);
1109 XFS_BUF_STALE(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
1116 * ASYNC buffers.
1118 XFS_BUF_ERROR(bp, EIO);
1119 XFS_BUF_FINISH_IOWAIT(bp);
1120 } else {
1121 xfs_buf_relse(bp);
1124 return EIO;
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.
1135 xfs_bdstrat_cb(
1136 struct xfs_buf *bp)
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);
1147 else
1148 return xfs_bioerror(bp);
1151 xfs_buf_iorequest(bp);
1152 return 0;
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.
1160 void
1161 xfsbdstrat(
1162 struct xfs_mount *mp,
1163 struct xfs_buf *bp)
1165 if (XFS_FORCED_SHUTDOWN(mp)) {
1166 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1167 xfs_bioerror_relse(bp);
1168 return;
1171 xfs_buf_iorequest(bp);
1174 STATIC void
1175 _xfs_buf_ioend(
1176 xfs_buf_t *bp,
1177 int schedule)
1179 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1180 xfs_buf_ioend(bp, schedule);
1183 STATIC void
1184 xfs_buf_bio_end_io(
1185 struct bio *bio,
1186 int error)
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);
1196 bio_put(bio);
1199 STATIC void
1200 _xfs_buf_ioapply(
1201 xfs_buf_t *bp)
1203 int rw, map_i, total_nr_pages, nr_pages;
1204 struct bio *bio;
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;
1210 map_i = 0;
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;
1223 } else {
1224 rw = (bp->b_flags & XBF_WRITE) ? WRITE :
1225 (bp->b_flags & XBF_READ_AHEAD) ? READA : READ;
1229 next_chunk:
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;
1245 if (nbytes > size)
1246 nbytes = size;
1248 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1249 if (rbytes < nbytes)
1250 break;
1252 offset = 0;
1253 sector += nbytes >> BBSHIFT;
1254 size -= nbytes;
1255 total_nr_pages--;
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);
1264 if (size)
1265 goto next_chunk;
1266 } else {
1267 xfs_buf_ioerror(bp, EIO);
1268 bio_put(bio);
1273 xfs_buf_iorequest(
1274 xfs_buf_t *bp)
1276 trace_xfs_buf_iorequest(bp, _RET_IP_);
1278 if (bp->b_flags & XBF_DELWRI) {
1279 xfs_buf_delwri_queue(bp, 1);
1280 return 0;
1283 if (bp->b_flags & XBF_WRITE) {
1284 xfs_buf_wait_unpin(bp);
1287 xfs_buf_hold(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);
1297 xfs_buf_rele(bp);
1298 return 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.
1307 xfs_buf_iowait(
1308 xfs_buf_t *bp)
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_);
1317 return bp->b_error;
1320 xfs_caddr_t
1321 xfs_buf_offset(
1322 xfs_buf_t *bp,
1323 size_t offset)
1325 struct page *page;
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.
1338 void
1339 xfs_buf_iomove(
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;
1347 struct page *page;
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));
1358 switch (mode) {
1359 case XBRW_ZERO:
1360 memset(page_address(page) + cpoff, 0, csize);
1361 break;
1362 case XBRW_READ:
1363 memcpy(data, page_address(page) + cpoff, csize);
1364 break;
1365 case XBRW_WRITE:
1366 memcpy(page_address(page) + cpoff, data, csize);
1369 boff += csize;
1370 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.
1383 void
1384 xfs_wait_buftarg(
1385 struct xfs_buftarg *btp)
1387 struct xfs_buf *bp;
1389 restart:
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);
1395 delay(100);
1396 goto restart;
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);
1404 xfs_buf_rele(bp);
1405 spin_lock(&btp->bt_lru_lock);
1407 spin_unlock(&btp->bt_lru_lock);
1411 xfs_buftarg_shrink(
1412 struct shrinker *shrink,
1413 int nr_to_scan,
1414 gfp_t mask)
1416 struct xfs_buftarg *btp = container_of(shrink,
1417 struct xfs_buftarg, bt_shrinker);
1418 struct xfs_buf *bp;
1419 LIST_HEAD(dispose);
1421 if (!nr_to_scan)
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)
1427 break;
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);
1438 continue;
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);
1446 btp->bt_lru_nr--;
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);
1453 xfs_buf_rele(bp);
1456 return btp->bt_lru_nr;
1459 void
1460 xfs_free_buftarg(
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);
1471 kmem_free(btp);
1474 STATIC int
1475 xfs_setsize_buftarg_flags(
1476 xfs_buftarg_t *btp,
1477 unsigned int blocksize,
1478 unsigned int sectorsize,
1479 int verbose)
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));
1489 return EINVAL;
1492 return 0;
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.
1500 STATIC int
1501 xfs_setsize_buftarg_early(
1502 xfs_buftarg_t *btp,
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(
1511 xfs_buftarg_t *btp,
1512 unsigned int blocksize,
1513 unsigned int sectorsize)
1515 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1518 STATIC int
1519 xfs_alloc_delwrite_queue(
1520 xfs_buftarg_t *btp,
1521 const char *fsname)
1523 INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1524 spin_lock_init(&btp->bt_delwrite_lock);
1525 btp->bt_flags = 0;
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);
1529 return 0;
1532 xfs_buftarg_t *
1533 xfs_alloc_buftarg(
1534 struct xfs_mount *mp,
1535 struct block_device *bdev,
1536 int external,
1537 const char *fsname)
1539 xfs_buftarg_t *btp;
1541 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1543 btp->bt_mount = mp;
1544 btp->bt_dev = bdev->bd_dev;
1545 btp->bt_bdev = bdev;
1546 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1547 if (!btp->bt_bdi)
1548 goto error;
1550 INIT_LIST_HEAD(&btp->bt_lru);
1551 spin_lock_init(&btp->bt_lru_lock);
1552 if (xfs_setsize_buftarg_early(btp, bdev))
1553 goto error;
1554 if (xfs_alloc_delwrite_queue(btp, fsname))
1555 goto error;
1556 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1557 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1558 register_shrinker(&btp->bt_shrinker);
1559 return btp;
1561 error:
1562 kmem_free(btp);
1563 return NULL;
1568 * Delayed write buffer handling
1570 STATIC void
1571 xfs_buf_delwri_queue(
1572 xfs_buf_t *bp,
1573 int unlock)
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));
1582 spin_lock(dwlk);
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);
1586 if (unlock)
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;
1599 spin_unlock(dwlk);
1601 if (unlock)
1602 xfs_buf_unlock(bp);
1605 void
1606 xfs_buf_delwri_dequeue(
1607 xfs_buf_t *bp)
1609 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1610 int dequeued = 0;
1612 spin_lock(dwlk);
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);
1616 dequeued = 1;
1618 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1619 spin_unlock(dwlk);
1621 if (dequeued)
1622 xfs_buf_rele(bp);
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.
1634 void
1635 xfs_buf_delwri_promote(
1636 struct xfs_buf *bp)
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)
1649 return;
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);
1656 STATIC void
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.
1667 STATIC int
1668 xfs_buf_delwri_split(
1669 xfs_buftarg_t *target,
1670 struct list_head *list,
1671 unsigned long age)
1673 xfs_buf_t *bp, *n;
1674 struct list_head *dwq = &target->bt_delwrite_queue;
1675 spinlock_t *dwlk = &target->bt_delwrite_lock;
1676 int skipped = 0;
1677 int force;
1679 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1680 INIT_LIST_HEAD(list);
1681 spin_lock(dwlk);
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)) {
1686 if (!force &&
1687 time_before(jiffies, bp->b_queuetime + age)) {
1688 xfs_buf_unlock(bp);
1689 break;
1692 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q|
1693 _XBF_RUN_QUEUES);
1694 bp->b_flags |= XBF_WRITE;
1695 list_move_tail(&bp->b_list, list);
1696 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1697 } else
1698 skipped++;
1700 spin_unlock(dwlk);
1702 return skipped;
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
1709 * on 64 bit values
1711 static int
1712 xfs_buf_cmp(
1713 void *priv,
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);
1719 xfs_daddr_t diff;
1721 diff = ap->b_bn - bp->b_bn;
1722 if (diff < 0)
1723 return -1;
1724 if (diff > 0)
1725 return 1;
1726 return 0;
1729 void
1730 xfs_buf_delwri_sort(
1731 xfs_buftarg_t *target,
1732 struct list_head *list)
1734 list_sort(NULL, list, xfs_buf_cmp);
1737 STATIC int
1738 xfsbufd(
1739 void *data)
1741 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1743 current->flags |= PF_MEMALLOC;
1745 set_freezable();
1747 do {
1748 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1749 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1750 int count = 0;
1751 struct list_head tmp;
1753 if (unlikely(freezing(current))) {
1754 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1755 refrigerator();
1756 } else {
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)) {
1768 struct xfs_buf *bp;
1769 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1770 list_del_init(&bp->b_list);
1771 xfs_bdstrat_cb(bp);
1772 count++;
1774 if (count)
1775 blk_flush_plug(current);
1777 } while (!kthread_should_stop());
1779 return 0;
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.
1788 xfs_flush_buftarg(
1789 xfs_buftarg_t *target,
1790 int wait)
1792 xfs_buf_t *bp;
1793 int pincount = 0;
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);
1814 if (wait) {
1815 bp->b_flags &= ~XBF_ASYNC;
1816 list_add(&bp->b_list, &wait_list);
1818 xfs_bdstrat_cb(bp);
1821 if (wait) {
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);
1828 xfs_buf_iowait(bp);
1829 xfs_buf_relse(bp);
1833 return pincount;
1836 int __init
1837 xfs_buf_init(void)
1839 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1840 KM_ZONE_HWALIGN, NULL);
1841 if (!xfs_buf_zone)
1842 goto out;
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",
1854 WQ_MEM_RECLAIM, 1);
1855 if (!xfsconvertd_workqueue)
1856 goto out_destroy_xfsdatad_workqueue;
1858 return 0;
1860 out_destroy_xfsdatad_workqueue:
1861 destroy_workqueue(xfsdatad_workqueue);
1862 out_destroy_xfslogd_workqueue:
1863 destroy_workqueue(xfslogd_workqueue);
1864 out_free_buf_zone:
1865 kmem_zone_destroy(xfs_buf_zone);
1866 out:
1867 return -ENOMEM;
1870 void
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
1880 struct list_head *
1881 xfs_get_buftarg_list(void)
1883 return &xfs_buftarg_list;
1885 #endif