hyperv: Add processing of MTU reduced by the host
[linux/fpc-iii.git] / fs / xfs / xfs_buf.c
blob24b4ebea0d4d114c924838c3d50614920c0a923d
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>
37 #include "xfs_log_format.h"
38 #include "xfs_trans_resv.h"
39 #include "xfs_sb.h"
40 #include "xfs_ag.h"
41 #include "xfs_mount.h"
42 #include "xfs_trace.h"
43 #include "xfs_log.h"
45 static kmem_zone_t *xfs_buf_zone;
47 static struct workqueue_struct *xfslogd_workqueue;
49 #ifdef XFS_BUF_LOCK_TRACKING
50 # define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
51 # define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
52 # define XB_GET_OWNER(bp) ((bp)->b_last_holder)
53 #else
54 # define XB_SET_OWNER(bp) do { } while (0)
55 # define XB_CLEAR_OWNER(bp) do { } while (0)
56 # define XB_GET_OWNER(bp) do { } while (0)
57 #endif
59 #define xb_to_gfp(flags) \
60 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
63 static inline int
64 xfs_buf_is_vmapped(
65 struct xfs_buf *bp)
68 * Return true if the buffer is vmapped.
70 * b_addr is null if the buffer is not mapped, but the code is clever
71 * enough to know it doesn't have to map a single page, so the check has
72 * to be both for b_addr and bp->b_page_count > 1.
74 return bp->b_addr && bp->b_page_count > 1;
77 static inline int
78 xfs_buf_vmap_len(
79 struct xfs_buf *bp)
81 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
85 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
86 * b_lru_ref count so that the buffer is freed immediately when the buffer
87 * reference count falls to zero. If the buffer is already on the LRU, we need
88 * to remove the reference that LRU holds on the buffer.
90 * This prevents build-up of stale buffers on the LRU.
92 void
93 xfs_buf_stale(
94 struct xfs_buf *bp)
96 ASSERT(xfs_buf_islocked(bp));
98 bp->b_flags |= XBF_STALE;
101 * Clear the delwri status so that a delwri queue walker will not
102 * flush this buffer to disk now that it is stale. The delwri queue has
103 * a reference to the buffer, so this is safe to do.
105 bp->b_flags &= ~_XBF_DELWRI_Q;
107 spin_lock(&bp->b_lock);
108 atomic_set(&bp->b_lru_ref, 0);
109 if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
110 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
111 atomic_dec(&bp->b_hold);
113 ASSERT(atomic_read(&bp->b_hold) >= 1);
114 spin_unlock(&bp->b_lock);
117 static int
118 xfs_buf_get_maps(
119 struct xfs_buf *bp,
120 int map_count)
122 ASSERT(bp->b_maps == NULL);
123 bp->b_map_count = map_count;
125 if (map_count == 1) {
126 bp->b_maps = &bp->__b_map;
127 return 0;
130 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
131 KM_NOFS);
132 if (!bp->b_maps)
133 return -ENOMEM;
134 return 0;
138 * Frees b_pages if it was allocated.
140 static void
141 xfs_buf_free_maps(
142 struct xfs_buf *bp)
144 if (bp->b_maps != &bp->__b_map) {
145 kmem_free(bp->b_maps);
146 bp->b_maps = NULL;
150 struct xfs_buf *
151 _xfs_buf_alloc(
152 struct xfs_buftarg *target,
153 struct xfs_buf_map *map,
154 int nmaps,
155 xfs_buf_flags_t flags)
157 struct xfs_buf *bp;
158 int error;
159 int i;
161 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS);
162 if (unlikely(!bp))
163 return NULL;
166 * We don't want certain flags to appear in b_flags unless they are
167 * specifically set by later operations on the buffer.
169 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
171 atomic_set(&bp->b_hold, 1);
172 atomic_set(&bp->b_lru_ref, 1);
173 init_completion(&bp->b_iowait);
174 INIT_LIST_HEAD(&bp->b_lru);
175 INIT_LIST_HEAD(&bp->b_list);
176 RB_CLEAR_NODE(&bp->b_rbnode);
177 sema_init(&bp->b_sema, 0); /* held, no waiters */
178 spin_lock_init(&bp->b_lock);
179 XB_SET_OWNER(bp);
180 bp->b_target = target;
181 bp->b_flags = flags;
184 * Set length and io_length to the same value initially.
185 * I/O routines should use io_length, which will be the same in
186 * most cases but may be reset (e.g. XFS recovery).
188 error = xfs_buf_get_maps(bp, nmaps);
189 if (error) {
190 kmem_zone_free(xfs_buf_zone, bp);
191 return NULL;
194 bp->b_bn = map[0].bm_bn;
195 bp->b_length = 0;
196 for (i = 0; i < nmaps; i++) {
197 bp->b_maps[i].bm_bn = map[i].bm_bn;
198 bp->b_maps[i].bm_len = map[i].bm_len;
199 bp->b_length += map[i].bm_len;
201 bp->b_io_length = bp->b_length;
203 atomic_set(&bp->b_pin_count, 0);
204 init_waitqueue_head(&bp->b_waiters);
206 XFS_STATS_INC(xb_create);
207 trace_xfs_buf_init(bp, _RET_IP_);
209 return bp;
213 * Allocate a page array capable of holding a specified number
214 * of pages, and point the page buf at it.
216 STATIC int
217 _xfs_buf_get_pages(
218 xfs_buf_t *bp,
219 int page_count)
221 /* Make sure that we have a page list */
222 if (bp->b_pages == NULL) {
223 bp->b_page_count = page_count;
224 if (page_count <= XB_PAGES) {
225 bp->b_pages = bp->b_page_array;
226 } else {
227 bp->b_pages = kmem_alloc(sizeof(struct page *) *
228 page_count, KM_NOFS);
229 if (bp->b_pages == NULL)
230 return -ENOMEM;
232 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
234 return 0;
238 * Frees b_pages if it was allocated.
240 STATIC void
241 _xfs_buf_free_pages(
242 xfs_buf_t *bp)
244 if (bp->b_pages != bp->b_page_array) {
245 kmem_free(bp->b_pages);
246 bp->b_pages = NULL;
251 * Releases the specified buffer.
253 * The modification state of any associated pages is left unchanged.
254 * The buffer must not be on any hash - use xfs_buf_rele instead for
255 * hashed and refcounted buffers
257 void
258 xfs_buf_free(
259 xfs_buf_t *bp)
261 trace_xfs_buf_free(bp, _RET_IP_);
263 ASSERT(list_empty(&bp->b_lru));
265 if (bp->b_flags & _XBF_PAGES) {
266 uint i;
268 if (xfs_buf_is_vmapped(bp))
269 vm_unmap_ram(bp->b_addr - bp->b_offset,
270 bp->b_page_count);
272 for (i = 0; i < bp->b_page_count; i++) {
273 struct page *page = bp->b_pages[i];
275 __free_page(page);
277 } else if (bp->b_flags & _XBF_KMEM)
278 kmem_free(bp->b_addr);
279 _xfs_buf_free_pages(bp);
280 xfs_buf_free_maps(bp);
281 kmem_zone_free(xfs_buf_zone, 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;
293 size_t nbytes, offset;
294 gfp_t gfp_mask = xb_to_gfp(flags);
295 unsigned short page_count, i;
296 xfs_off_t start, end;
297 int error;
300 * for buffers that are contained within a single page, just allocate
301 * the memory from the heap - there's no need for the complexity of
302 * page arrays to keep allocation down to order 0.
304 size = BBTOB(bp->b_length);
305 if (size < PAGE_SIZE) {
306 bp->b_addr = kmem_alloc(size, KM_NOFS);
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 + size - 1) & PAGE_MASK) !=
313 ((unsigned long)bp->b_addr & PAGE_MASK)) {
314 /* b_addr spans two pages - use alloc_page instead */
315 kmem_free(bp->b_addr);
316 bp->b_addr = NULL;
317 goto use_alloc_page;
319 bp->b_offset = offset_in_page(bp->b_addr);
320 bp->b_pages = bp->b_page_array;
321 bp->b_pages[0] = virt_to_page(bp->b_addr);
322 bp->b_page_count = 1;
323 bp->b_flags |= _XBF_KMEM;
324 return 0;
327 use_alloc_page:
328 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT;
329 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1)
330 >> PAGE_SHIFT;
331 page_count = end - start;
332 error = _xfs_buf_get_pages(bp, page_count);
333 if (unlikely(error))
334 return error;
336 offset = bp->b_offset;
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 necessary.
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 } else if (flags & XBF_UNMAPPED) {
395 bp->b_addr = NULL;
396 } else {
397 int retried = 0;
398 unsigned noio_flag;
401 * vm_map_ram() will allocate auxillary structures (e.g.
402 * pagetables) with GFP_KERNEL, yet we are likely to be under
403 * GFP_NOFS context here. Hence we need to tell memory reclaim
404 * that we are in such a context via PF_MEMALLOC_NOIO to prevent
405 * memory reclaim re-entering the filesystem here and
406 * potentially deadlocking.
408 noio_flag = memalloc_noio_save();
409 do {
410 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
411 -1, PAGE_KERNEL);
412 if (bp->b_addr)
413 break;
414 vm_unmap_aliases();
415 } while (retried++ <= 1);
416 memalloc_noio_restore(noio_flag);
418 if (!bp->b_addr)
419 return -ENOMEM;
420 bp->b_addr += bp->b_offset;
423 return 0;
427 * Finding and Reading Buffers
431 * Look up, and creates if absent, a lockable buffer for
432 * a given range of an inode. The buffer is returned
433 * locked. No I/O is implied by this call.
435 xfs_buf_t *
436 _xfs_buf_find(
437 struct xfs_buftarg *btp,
438 struct xfs_buf_map *map,
439 int nmaps,
440 xfs_buf_flags_t flags,
441 xfs_buf_t *new_bp)
443 size_t numbytes;
444 struct xfs_perag *pag;
445 struct rb_node **rbp;
446 struct rb_node *parent;
447 xfs_buf_t *bp;
448 xfs_daddr_t blkno = map[0].bm_bn;
449 xfs_daddr_t eofs;
450 int numblks = 0;
451 int i;
453 for (i = 0; i < nmaps; i++)
454 numblks += map[i].bm_len;
455 numbytes = BBTOB(numblks);
457 /* Check for IOs smaller than the sector size / not sector aligned */
458 ASSERT(!(numbytes < btp->bt_meta_sectorsize));
459 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_meta_sectormask));
462 * Corrupted block numbers can get through to here, unfortunately, so we
463 * have to check that the buffer falls within the filesystem bounds.
465 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
466 if (blkno >= eofs) {
468 * XXX (dgc): we should really be returning -EFSCORRUPTED here,
469 * but none of the higher level infrastructure supports
470 * returning a specific error on buffer lookup failures.
472 xfs_alert(btp->bt_mount,
473 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ",
474 __func__, blkno, eofs);
475 WARN_ON(1);
476 return NULL;
479 /* get tree root */
480 pag = xfs_perag_get(btp->bt_mount,
481 xfs_daddr_to_agno(btp->bt_mount, blkno));
483 /* walk tree */
484 spin_lock(&pag->pag_buf_lock);
485 rbp = &pag->pag_buf_tree.rb_node;
486 parent = NULL;
487 bp = NULL;
488 while (*rbp) {
489 parent = *rbp;
490 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
492 if (blkno < bp->b_bn)
493 rbp = &(*rbp)->rb_left;
494 else if (blkno > bp->b_bn)
495 rbp = &(*rbp)->rb_right;
496 else {
498 * found a block number match. If the range doesn't
499 * match, the only way this is allowed is if the buffer
500 * in the cache is stale and the transaction that made
501 * it stale has not yet committed. i.e. we are
502 * reallocating a busy extent. Skip this buffer and
503 * continue searching to the right for an exact match.
505 if (bp->b_length != numblks) {
506 ASSERT(bp->b_flags & XBF_STALE);
507 rbp = &(*rbp)->rb_right;
508 continue;
510 atomic_inc(&bp->b_hold);
511 goto found;
515 /* No match found */
516 if (new_bp) {
517 rb_link_node(&new_bp->b_rbnode, parent, rbp);
518 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
519 /* the buffer keeps the perag reference until it is freed */
520 new_bp->b_pag = pag;
521 spin_unlock(&pag->pag_buf_lock);
522 } else {
523 XFS_STATS_INC(xb_miss_locked);
524 spin_unlock(&pag->pag_buf_lock);
525 xfs_perag_put(pag);
527 return new_bp;
529 found:
530 spin_unlock(&pag->pag_buf_lock);
531 xfs_perag_put(pag);
533 if (!xfs_buf_trylock(bp)) {
534 if (flags & XBF_TRYLOCK) {
535 xfs_buf_rele(bp);
536 XFS_STATS_INC(xb_busy_locked);
537 return NULL;
539 xfs_buf_lock(bp);
540 XFS_STATS_INC(xb_get_locked_waited);
544 * if the buffer is stale, clear all the external state associated with
545 * it. We need to keep flags such as how we allocated the buffer memory
546 * intact here.
548 if (bp->b_flags & XBF_STALE) {
549 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
550 ASSERT(bp->b_iodone == NULL);
551 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
552 bp->b_ops = NULL;
555 trace_xfs_buf_find(bp, flags, _RET_IP_);
556 XFS_STATS_INC(xb_get_locked);
557 return bp;
561 * Assembles a buffer covering the specified range. The code is optimised for
562 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
563 * more hits than misses.
565 struct xfs_buf *
566 xfs_buf_get_map(
567 struct xfs_buftarg *target,
568 struct xfs_buf_map *map,
569 int nmaps,
570 xfs_buf_flags_t flags)
572 struct xfs_buf *bp;
573 struct xfs_buf *new_bp;
574 int error = 0;
576 bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
577 if (likely(bp))
578 goto found;
580 new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
581 if (unlikely(!new_bp))
582 return NULL;
584 error = xfs_buf_allocate_memory(new_bp, flags);
585 if (error) {
586 xfs_buf_free(new_bp);
587 return NULL;
590 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
591 if (!bp) {
592 xfs_buf_free(new_bp);
593 return NULL;
596 if (bp != new_bp)
597 xfs_buf_free(new_bp);
599 found:
600 if (!bp->b_addr) {
601 error = _xfs_buf_map_pages(bp, flags);
602 if (unlikely(error)) {
603 xfs_warn(target->bt_mount,
604 "%s: failed to map pagesn", __func__);
605 xfs_buf_relse(bp);
606 return NULL;
610 XFS_STATS_INC(xb_get);
611 trace_xfs_buf_get(bp, flags, _RET_IP_);
612 return bp;
615 STATIC int
616 _xfs_buf_read(
617 xfs_buf_t *bp,
618 xfs_buf_flags_t flags)
620 ASSERT(!(flags & XBF_WRITE));
621 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
623 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
624 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
626 if (flags & XBF_ASYNC) {
627 xfs_buf_submit(bp);
628 return 0;
630 return xfs_buf_submit_wait(bp);
633 xfs_buf_t *
634 xfs_buf_read_map(
635 struct xfs_buftarg *target,
636 struct xfs_buf_map *map,
637 int nmaps,
638 xfs_buf_flags_t flags,
639 const struct xfs_buf_ops *ops)
641 struct xfs_buf *bp;
643 flags |= XBF_READ;
645 bp = xfs_buf_get_map(target, map, nmaps, flags);
646 if (bp) {
647 trace_xfs_buf_read(bp, flags, _RET_IP_);
649 if (!XFS_BUF_ISDONE(bp)) {
650 XFS_STATS_INC(xb_get_read);
651 bp->b_ops = ops;
652 _xfs_buf_read(bp, flags);
653 } else if (flags & XBF_ASYNC) {
655 * Read ahead call which is already satisfied,
656 * drop the buffer
658 xfs_buf_relse(bp);
659 return NULL;
660 } else {
661 /* We do not want read in the flags */
662 bp->b_flags &= ~XBF_READ;
666 return bp;
670 * If we are not low on memory then do the readahead in a deadlock
671 * safe manner.
673 void
674 xfs_buf_readahead_map(
675 struct xfs_buftarg *target,
676 struct xfs_buf_map *map,
677 int nmaps,
678 const struct xfs_buf_ops *ops)
680 if (bdi_read_congested(target->bt_bdi))
681 return;
683 xfs_buf_read_map(target, map, nmaps,
684 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops);
688 * Read an uncached buffer from disk. Allocates and returns a locked
689 * buffer containing the disk contents or nothing.
692 xfs_buf_read_uncached(
693 struct xfs_buftarg *target,
694 xfs_daddr_t daddr,
695 size_t numblks,
696 int flags,
697 struct xfs_buf **bpp,
698 const struct xfs_buf_ops *ops)
700 struct xfs_buf *bp;
702 *bpp = NULL;
704 bp = xfs_buf_get_uncached(target, numblks, flags);
705 if (!bp)
706 return -ENOMEM;
708 /* set up the buffer for a read IO */
709 ASSERT(bp->b_map_count == 1);
710 bp->b_bn = XFS_BUF_DADDR_NULL; /* always null for uncached buffers */
711 bp->b_maps[0].bm_bn = daddr;
712 bp->b_flags |= XBF_READ;
713 bp->b_ops = ops;
715 xfs_buf_submit_wait(bp);
716 if (bp->b_error) {
717 int error = bp->b_error;
718 xfs_buf_relse(bp);
719 return error;
722 *bpp = bp;
723 return 0;
727 * Return a buffer allocated as an empty buffer and associated to external
728 * memory via xfs_buf_associate_memory() back to it's empty state.
730 void
731 xfs_buf_set_empty(
732 struct xfs_buf *bp,
733 size_t numblks)
735 if (bp->b_pages)
736 _xfs_buf_free_pages(bp);
738 bp->b_pages = NULL;
739 bp->b_page_count = 0;
740 bp->b_addr = NULL;
741 bp->b_length = numblks;
742 bp->b_io_length = numblks;
744 ASSERT(bp->b_map_count == 1);
745 bp->b_bn = XFS_BUF_DADDR_NULL;
746 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
747 bp->b_maps[0].bm_len = bp->b_length;
750 static inline struct page *
751 mem_to_page(
752 void *addr)
754 if ((!is_vmalloc_addr(addr))) {
755 return virt_to_page(addr);
756 } else {
757 return vmalloc_to_page(addr);
762 xfs_buf_associate_memory(
763 xfs_buf_t *bp,
764 void *mem,
765 size_t len)
767 int rval;
768 int i = 0;
769 unsigned long pageaddr;
770 unsigned long offset;
771 size_t buflen;
772 int page_count;
774 pageaddr = (unsigned long)mem & PAGE_MASK;
775 offset = (unsigned long)mem - pageaddr;
776 buflen = PAGE_ALIGN(len + offset);
777 page_count = buflen >> PAGE_SHIFT;
779 /* Free any previous set of page pointers */
780 if (bp->b_pages)
781 _xfs_buf_free_pages(bp);
783 bp->b_pages = NULL;
784 bp->b_addr = mem;
786 rval = _xfs_buf_get_pages(bp, page_count);
787 if (rval)
788 return rval;
790 bp->b_offset = offset;
792 for (i = 0; i < bp->b_page_count; i++) {
793 bp->b_pages[i] = mem_to_page((void *)pageaddr);
794 pageaddr += PAGE_SIZE;
797 bp->b_io_length = BTOBB(len);
798 bp->b_length = BTOBB(buflen);
800 return 0;
803 xfs_buf_t *
804 xfs_buf_get_uncached(
805 struct xfs_buftarg *target,
806 size_t numblks,
807 int flags)
809 unsigned long page_count;
810 int error, i;
811 struct xfs_buf *bp;
812 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
814 bp = _xfs_buf_alloc(target, &map, 1, 0);
815 if (unlikely(bp == NULL))
816 goto fail;
818 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
819 error = _xfs_buf_get_pages(bp, page_count);
820 if (error)
821 goto fail_free_buf;
823 for (i = 0; i < page_count; i++) {
824 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
825 if (!bp->b_pages[i])
826 goto fail_free_mem;
828 bp->b_flags |= _XBF_PAGES;
830 error = _xfs_buf_map_pages(bp, 0);
831 if (unlikely(error)) {
832 xfs_warn(target->bt_mount,
833 "%s: failed to map pages", __func__);
834 goto fail_free_mem;
837 trace_xfs_buf_get_uncached(bp, _RET_IP_);
838 return bp;
840 fail_free_mem:
841 while (--i >= 0)
842 __free_page(bp->b_pages[i]);
843 _xfs_buf_free_pages(bp);
844 fail_free_buf:
845 xfs_buf_free_maps(bp);
846 kmem_zone_free(xfs_buf_zone, bp);
847 fail:
848 return NULL;
852 * Increment reference count on buffer, to hold the buffer concurrently
853 * with another thread which may release (free) the buffer asynchronously.
854 * Must hold the buffer already to call this function.
856 void
857 xfs_buf_hold(
858 xfs_buf_t *bp)
860 trace_xfs_buf_hold(bp, _RET_IP_);
861 atomic_inc(&bp->b_hold);
865 * Releases a hold on the specified buffer. If the
866 * the hold count is 1, calls xfs_buf_free.
868 void
869 xfs_buf_rele(
870 xfs_buf_t *bp)
872 struct xfs_perag *pag = bp->b_pag;
874 trace_xfs_buf_rele(bp, _RET_IP_);
876 if (!pag) {
877 ASSERT(list_empty(&bp->b_lru));
878 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
879 if (atomic_dec_and_test(&bp->b_hold))
880 xfs_buf_free(bp);
881 return;
884 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
886 ASSERT(atomic_read(&bp->b_hold) > 0);
887 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
888 spin_lock(&bp->b_lock);
889 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
891 * If the buffer is added to the LRU take a new
892 * reference to the buffer for the LRU and clear the
893 * (now stale) dispose list state flag
895 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
896 bp->b_state &= ~XFS_BSTATE_DISPOSE;
897 atomic_inc(&bp->b_hold);
899 spin_unlock(&bp->b_lock);
900 spin_unlock(&pag->pag_buf_lock);
901 } else {
903 * most of the time buffers will already be removed from
904 * the LRU, so optimise that case by checking for the
905 * XFS_BSTATE_DISPOSE flag indicating the last list the
906 * buffer was on was the disposal list
908 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
909 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
910 } else {
911 ASSERT(list_empty(&bp->b_lru));
913 spin_unlock(&bp->b_lock);
915 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
916 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
917 spin_unlock(&pag->pag_buf_lock);
918 xfs_perag_put(pag);
919 xfs_buf_free(bp);
926 * Lock a buffer object, if it is not already locked.
928 * If we come across a stale, pinned, locked buffer, we know that we are
929 * being asked to lock a buffer that has been reallocated. Because it is
930 * pinned, we know that the log has not been pushed to disk and hence it
931 * will still be locked. Rather than continuing to have trylock attempts
932 * fail until someone else pushes the log, push it ourselves before
933 * returning. This means that the xfsaild will not get stuck trying
934 * to push on stale inode buffers.
937 xfs_buf_trylock(
938 struct xfs_buf *bp)
940 int locked;
942 locked = down_trylock(&bp->b_sema) == 0;
943 if (locked)
944 XB_SET_OWNER(bp);
946 trace_xfs_buf_trylock(bp, _RET_IP_);
947 return locked;
951 * Lock a buffer object.
953 * If we come across a stale, pinned, locked buffer, we know that we
954 * are being asked to lock a buffer that has been reallocated. Because
955 * it is pinned, we know that the log has not been pushed to disk and
956 * hence it will still be locked. Rather than sleeping until someone
957 * else pushes the log, push it ourselves before trying to get the lock.
959 void
960 xfs_buf_lock(
961 struct xfs_buf *bp)
963 trace_xfs_buf_lock(bp, _RET_IP_);
965 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
966 xfs_log_force(bp->b_target->bt_mount, 0);
967 down(&bp->b_sema);
968 XB_SET_OWNER(bp);
970 trace_xfs_buf_lock_done(bp, _RET_IP_);
973 void
974 xfs_buf_unlock(
975 struct xfs_buf *bp)
977 XB_CLEAR_OWNER(bp);
978 up(&bp->b_sema);
980 trace_xfs_buf_unlock(bp, _RET_IP_);
983 STATIC void
984 xfs_buf_wait_unpin(
985 xfs_buf_t *bp)
987 DECLARE_WAITQUEUE (wait, current);
989 if (atomic_read(&bp->b_pin_count) == 0)
990 return;
992 add_wait_queue(&bp->b_waiters, &wait);
993 for (;;) {
994 set_current_state(TASK_UNINTERRUPTIBLE);
995 if (atomic_read(&bp->b_pin_count) == 0)
996 break;
997 io_schedule();
999 remove_wait_queue(&bp->b_waiters, &wait);
1000 set_current_state(TASK_RUNNING);
1004 * Buffer Utility Routines
1007 void
1008 xfs_buf_ioend(
1009 struct xfs_buf *bp)
1011 bool read = bp->b_flags & XBF_READ;
1013 trace_xfs_buf_iodone(bp, _RET_IP_);
1015 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1018 * Pull in IO completion errors now. We are guaranteed to be running
1019 * single threaded, so we don't need the lock to read b_io_error.
1021 if (!bp->b_error && bp->b_io_error)
1022 xfs_buf_ioerror(bp, bp->b_io_error);
1024 /* Only validate buffers that were read without errors */
1025 if (read && !bp->b_error && bp->b_ops) {
1026 ASSERT(!bp->b_iodone);
1027 bp->b_ops->verify_read(bp);
1030 if (!bp->b_error)
1031 bp->b_flags |= XBF_DONE;
1033 if (bp->b_iodone)
1034 (*(bp->b_iodone))(bp);
1035 else if (bp->b_flags & XBF_ASYNC)
1036 xfs_buf_relse(bp);
1037 else
1038 complete(&bp->b_iowait);
1041 static void
1042 xfs_buf_ioend_work(
1043 struct work_struct *work)
1045 struct xfs_buf *bp =
1046 container_of(work, xfs_buf_t, b_iodone_work);
1048 xfs_buf_ioend(bp);
1051 void
1052 xfs_buf_ioend_async(
1053 struct xfs_buf *bp)
1055 INIT_WORK(&bp->b_iodone_work, xfs_buf_ioend_work);
1056 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1059 void
1060 xfs_buf_ioerror(
1061 xfs_buf_t *bp,
1062 int error)
1064 ASSERT(error <= 0 && error >= -1000);
1065 bp->b_error = error;
1066 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1069 void
1070 xfs_buf_ioerror_alert(
1071 struct xfs_buf *bp,
1072 const char *func)
1074 xfs_alert(bp->b_target->bt_mount,
1075 "metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1076 (__uint64_t)XFS_BUF_ADDR(bp), func, -bp->b_error, bp->b_length);
1080 xfs_bwrite(
1081 struct xfs_buf *bp)
1083 int error;
1085 ASSERT(xfs_buf_islocked(bp));
1087 bp->b_flags |= XBF_WRITE;
1088 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1089 XBF_WRITE_FAIL | XBF_DONE);
1091 error = xfs_buf_submit_wait(bp);
1092 if (error) {
1093 xfs_force_shutdown(bp->b_target->bt_mount,
1094 SHUTDOWN_META_IO_ERROR);
1096 return error;
1099 STATIC void
1100 xfs_buf_bio_end_io(
1101 struct bio *bio,
1102 int error)
1104 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1107 * don't overwrite existing errors - otherwise we can lose errors on
1108 * buffers that require multiple bios to complete.
1110 if (error) {
1111 spin_lock(&bp->b_lock);
1112 if (!bp->b_io_error)
1113 bp->b_io_error = error;
1114 spin_unlock(&bp->b_lock);
1117 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1118 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1120 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1121 xfs_buf_ioend_async(bp);
1122 bio_put(bio);
1125 static void
1126 xfs_buf_ioapply_map(
1127 struct xfs_buf *bp,
1128 int map,
1129 int *buf_offset,
1130 int *count,
1131 int rw)
1133 int page_index;
1134 int total_nr_pages = bp->b_page_count;
1135 int nr_pages;
1136 struct bio *bio;
1137 sector_t sector = bp->b_maps[map].bm_bn;
1138 int size;
1139 int offset;
1141 total_nr_pages = bp->b_page_count;
1143 /* skip the pages in the buffer before the start offset */
1144 page_index = 0;
1145 offset = *buf_offset;
1146 while (offset >= PAGE_SIZE) {
1147 page_index++;
1148 offset -= PAGE_SIZE;
1152 * Limit the IO size to the length of the current vector, and update the
1153 * remaining IO count for the next time around.
1155 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1156 *count -= size;
1157 *buf_offset += size;
1159 next_chunk:
1160 atomic_inc(&bp->b_io_remaining);
1161 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1162 if (nr_pages > total_nr_pages)
1163 nr_pages = total_nr_pages;
1165 bio = bio_alloc(GFP_NOIO, nr_pages);
1166 bio->bi_bdev = bp->b_target->bt_bdev;
1167 bio->bi_iter.bi_sector = sector;
1168 bio->bi_end_io = xfs_buf_bio_end_io;
1169 bio->bi_private = bp;
1172 for (; size && nr_pages; nr_pages--, page_index++) {
1173 int rbytes, nbytes = PAGE_SIZE - offset;
1175 if (nbytes > size)
1176 nbytes = size;
1178 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1179 offset);
1180 if (rbytes < nbytes)
1181 break;
1183 offset = 0;
1184 sector += BTOBB(nbytes);
1185 size -= nbytes;
1186 total_nr_pages--;
1189 if (likely(bio->bi_iter.bi_size)) {
1190 if (xfs_buf_is_vmapped(bp)) {
1191 flush_kernel_vmap_range(bp->b_addr,
1192 xfs_buf_vmap_len(bp));
1194 submit_bio(rw, bio);
1195 if (size)
1196 goto next_chunk;
1197 } else {
1199 * This is guaranteed not to be the last io reference count
1200 * because the caller (xfs_buf_submit) holds a count itself.
1202 atomic_dec(&bp->b_io_remaining);
1203 xfs_buf_ioerror(bp, -EIO);
1204 bio_put(bio);
1209 STATIC void
1210 _xfs_buf_ioapply(
1211 struct xfs_buf *bp)
1213 struct blk_plug plug;
1214 int rw;
1215 int offset;
1216 int size;
1217 int i;
1220 * Make sure we capture only current IO errors rather than stale errors
1221 * left over from previous use of the buffer (e.g. failed readahead).
1223 bp->b_error = 0;
1225 if (bp->b_flags & XBF_WRITE) {
1226 if (bp->b_flags & XBF_SYNCIO)
1227 rw = WRITE_SYNC;
1228 else
1229 rw = WRITE;
1230 if (bp->b_flags & XBF_FUA)
1231 rw |= REQ_FUA;
1232 if (bp->b_flags & XBF_FLUSH)
1233 rw |= REQ_FLUSH;
1236 * Run the write verifier callback function if it exists. If
1237 * this function fails it will mark the buffer with an error and
1238 * the IO should not be dispatched.
1240 if (bp->b_ops) {
1241 bp->b_ops->verify_write(bp);
1242 if (bp->b_error) {
1243 xfs_force_shutdown(bp->b_target->bt_mount,
1244 SHUTDOWN_CORRUPT_INCORE);
1245 return;
1247 } else if (bp->b_bn != XFS_BUF_DADDR_NULL) {
1248 struct xfs_mount *mp = bp->b_target->bt_mount;
1251 * non-crc filesystems don't attach verifiers during
1252 * log recovery, so don't warn for such filesystems.
1254 if (xfs_sb_version_hascrc(&mp->m_sb)) {
1255 xfs_warn(mp,
1256 "%s: no ops on block 0x%llx/0x%x",
1257 __func__, bp->b_bn, bp->b_length);
1258 xfs_hex_dump(bp->b_addr, 64);
1259 dump_stack();
1262 } else if (bp->b_flags & XBF_READ_AHEAD) {
1263 rw = READA;
1264 } else {
1265 rw = READ;
1268 /* we only use the buffer cache for meta-data */
1269 rw |= REQ_META;
1272 * Walk all the vectors issuing IO on them. Set up the initial offset
1273 * into the buffer and the desired IO size before we start -
1274 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1275 * subsequent call.
1277 offset = bp->b_offset;
1278 size = BBTOB(bp->b_io_length);
1279 blk_start_plug(&plug);
1280 for (i = 0; i < bp->b_map_count; i++) {
1281 xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1282 if (bp->b_error)
1283 break;
1284 if (size <= 0)
1285 break; /* all done */
1287 blk_finish_plug(&plug);
1291 * Asynchronous IO submission path. This transfers the buffer lock ownership and
1292 * the current reference to the IO. It is not safe to reference the buffer after
1293 * a call to this function unless the caller holds an additional reference
1294 * itself.
1296 void
1297 xfs_buf_submit(
1298 struct xfs_buf *bp)
1300 trace_xfs_buf_submit(bp, _RET_IP_);
1302 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1303 ASSERT(bp->b_flags & XBF_ASYNC);
1305 /* on shutdown we stale and complete the buffer immediately */
1306 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1307 xfs_buf_ioerror(bp, -EIO);
1308 bp->b_flags &= ~XBF_DONE;
1309 xfs_buf_stale(bp);
1310 xfs_buf_ioend(bp);
1311 return;
1314 if (bp->b_flags & XBF_WRITE)
1315 xfs_buf_wait_unpin(bp);
1317 /* clear the internal error state to avoid spurious errors */
1318 bp->b_io_error = 0;
1321 * The caller's reference is released during I/O completion.
1322 * This occurs some time after the last b_io_remaining reference is
1323 * released, so after we drop our Io reference we have to have some
1324 * other reference to ensure the buffer doesn't go away from underneath
1325 * us. Take a direct reference to ensure we have safe access to the
1326 * buffer until we are finished with it.
1328 xfs_buf_hold(bp);
1331 * Set the count to 1 initially, this will stop an I/O completion
1332 * callout which happens before we have started all the I/O from calling
1333 * xfs_buf_ioend too early.
1335 atomic_set(&bp->b_io_remaining, 1);
1336 _xfs_buf_ioapply(bp);
1339 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1340 * reference we took above. If we drop it to zero, run completion so
1341 * that we don't return to the caller with completion still pending.
1343 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1344 if (bp->b_error)
1345 xfs_buf_ioend(bp);
1346 else
1347 xfs_buf_ioend_async(bp);
1350 xfs_buf_rele(bp);
1351 /* Note: it is not safe to reference bp now we've dropped our ref */
1355 * Synchronous buffer IO submission path, read or write.
1358 xfs_buf_submit_wait(
1359 struct xfs_buf *bp)
1361 int error;
1363 trace_xfs_buf_submit_wait(bp, _RET_IP_);
1365 ASSERT(!(bp->b_flags & (_XBF_DELWRI_Q | XBF_ASYNC)));
1367 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1368 xfs_buf_ioerror(bp, -EIO);
1369 xfs_buf_stale(bp);
1370 bp->b_flags &= ~XBF_DONE;
1371 return -EIO;
1374 if (bp->b_flags & XBF_WRITE)
1375 xfs_buf_wait_unpin(bp);
1377 /* clear the internal error state to avoid spurious errors */
1378 bp->b_io_error = 0;
1381 * For synchronous IO, the IO does not inherit the submitters reference
1382 * count, nor the buffer lock. Hence we cannot release the reference we
1383 * are about to take until we've waited for all IO completion to occur,
1384 * including any xfs_buf_ioend_async() work that may be pending.
1386 xfs_buf_hold(bp);
1389 * Set the count to 1 initially, this will stop an I/O completion
1390 * callout which happens before we have started all the I/O from calling
1391 * xfs_buf_ioend too early.
1393 atomic_set(&bp->b_io_remaining, 1);
1394 _xfs_buf_ioapply(bp);
1397 * make sure we run completion synchronously if it raced with us and is
1398 * already complete.
1400 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1401 xfs_buf_ioend(bp);
1403 /* wait for completion before gathering the error from the buffer */
1404 trace_xfs_buf_iowait(bp, _RET_IP_);
1405 wait_for_completion(&bp->b_iowait);
1406 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1407 error = bp->b_error;
1410 * all done now, we can release the hold that keeps the buffer
1411 * referenced for the entire IO.
1413 xfs_buf_rele(bp);
1414 return error;
1417 xfs_caddr_t
1418 xfs_buf_offset(
1419 xfs_buf_t *bp,
1420 size_t offset)
1422 struct page *page;
1424 if (bp->b_addr)
1425 return bp->b_addr + offset;
1427 offset += bp->b_offset;
1428 page = bp->b_pages[offset >> PAGE_SHIFT];
1429 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1433 * Move data into or out of a buffer.
1435 void
1436 xfs_buf_iomove(
1437 xfs_buf_t *bp, /* buffer to process */
1438 size_t boff, /* starting buffer offset */
1439 size_t bsize, /* length to copy */
1440 void *data, /* data address */
1441 xfs_buf_rw_t mode) /* read/write/zero flag */
1443 size_t bend;
1445 bend = boff + bsize;
1446 while (boff < bend) {
1447 struct page *page;
1448 int page_index, page_offset, csize;
1450 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1451 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1452 page = bp->b_pages[page_index];
1453 csize = min_t(size_t, PAGE_SIZE - page_offset,
1454 BBTOB(bp->b_io_length) - boff);
1456 ASSERT((csize + page_offset) <= PAGE_SIZE);
1458 switch (mode) {
1459 case XBRW_ZERO:
1460 memset(page_address(page) + page_offset, 0, csize);
1461 break;
1462 case XBRW_READ:
1463 memcpy(data, page_address(page) + page_offset, csize);
1464 break;
1465 case XBRW_WRITE:
1466 memcpy(page_address(page) + page_offset, data, csize);
1469 boff += csize;
1470 data += csize;
1475 * Handling of buffer targets (buftargs).
1479 * Wait for any bufs with callbacks that have been submitted but have not yet
1480 * returned. These buffers will have an elevated hold count, so wait on those
1481 * while freeing all the buffers only held by the LRU.
1483 static enum lru_status
1484 xfs_buftarg_wait_rele(
1485 struct list_head *item,
1486 spinlock_t *lru_lock,
1487 void *arg)
1490 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1491 struct list_head *dispose = arg;
1493 if (atomic_read(&bp->b_hold) > 1) {
1494 /* need to wait, so skip it this pass */
1495 trace_xfs_buf_wait_buftarg(bp, _RET_IP_);
1496 return LRU_SKIP;
1498 if (!spin_trylock(&bp->b_lock))
1499 return LRU_SKIP;
1502 * clear the LRU reference count so the buffer doesn't get
1503 * ignored in xfs_buf_rele().
1505 atomic_set(&bp->b_lru_ref, 0);
1506 bp->b_state |= XFS_BSTATE_DISPOSE;
1507 list_move(item, dispose);
1508 spin_unlock(&bp->b_lock);
1509 return LRU_REMOVED;
1512 void
1513 xfs_wait_buftarg(
1514 struct xfs_buftarg *btp)
1516 LIST_HEAD(dispose);
1517 int loop = 0;
1519 /* loop until there is nothing left on the lru list. */
1520 while (list_lru_count(&btp->bt_lru)) {
1521 list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele,
1522 &dispose, LONG_MAX);
1524 while (!list_empty(&dispose)) {
1525 struct xfs_buf *bp;
1526 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1527 list_del_init(&bp->b_lru);
1528 if (bp->b_flags & XBF_WRITE_FAIL) {
1529 xfs_alert(btp->bt_mount,
1530 "Corruption Alert: Buffer at block 0x%llx had permanent write failures!\n"
1531 "Please run xfs_repair to determine the extent of the problem.",
1532 (long long)bp->b_bn);
1534 xfs_buf_rele(bp);
1536 if (loop++ != 0)
1537 delay(100);
1541 static enum lru_status
1542 xfs_buftarg_isolate(
1543 struct list_head *item,
1544 spinlock_t *lru_lock,
1545 void *arg)
1547 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1548 struct list_head *dispose = arg;
1551 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1552 * If we fail to get the lock, just skip it.
1554 if (!spin_trylock(&bp->b_lock))
1555 return LRU_SKIP;
1557 * Decrement the b_lru_ref count unless the value is already
1558 * zero. If the value is already zero, we need to reclaim the
1559 * buffer, otherwise it gets another trip through the LRU.
1561 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1562 spin_unlock(&bp->b_lock);
1563 return LRU_ROTATE;
1566 bp->b_state |= XFS_BSTATE_DISPOSE;
1567 list_move(item, dispose);
1568 spin_unlock(&bp->b_lock);
1569 return LRU_REMOVED;
1572 static unsigned long
1573 xfs_buftarg_shrink_scan(
1574 struct shrinker *shrink,
1575 struct shrink_control *sc)
1577 struct xfs_buftarg *btp = container_of(shrink,
1578 struct xfs_buftarg, bt_shrinker);
1579 LIST_HEAD(dispose);
1580 unsigned long freed;
1581 unsigned long nr_to_scan = sc->nr_to_scan;
1583 freed = list_lru_walk_node(&btp->bt_lru, sc->nid, xfs_buftarg_isolate,
1584 &dispose, &nr_to_scan);
1586 while (!list_empty(&dispose)) {
1587 struct xfs_buf *bp;
1588 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1589 list_del_init(&bp->b_lru);
1590 xfs_buf_rele(bp);
1593 return freed;
1596 static unsigned long
1597 xfs_buftarg_shrink_count(
1598 struct shrinker *shrink,
1599 struct shrink_control *sc)
1601 struct xfs_buftarg *btp = container_of(shrink,
1602 struct xfs_buftarg, bt_shrinker);
1603 return list_lru_count_node(&btp->bt_lru, sc->nid);
1606 void
1607 xfs_free_buftarg(
1608 struct xfs_mount *mp,
1609 struct xfs_buftarg *btp)
1611 unregister_shrinker(&btp->bt_shrinker);
1612 list_lru_destroy(&btp->bt_lru);
1614 if (mp->m_flags & XFS_MOUNT_BARRIER)
1615 xfs_blkdev_issue_flush(btp);
1617 kmem_free(btp);
1621 xfs_setsize_buftarg(
1622 xfs_buftarg_t *btp,
1623 unsigned int sectorsize)
1625 /* Set up metadata sector size info */
1626 btp->bt_meta_sectorsize = sectorsize;
1627 btp->bt_meta_sectormask = sectorsize - 1;
1629 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1630 char name[BDEVNAME_SIZE];
1632 bdevname(btp->bt_bdev, name);
1634 xfs_warn(btp->bt_mount,
1635 "Cannot set_blocksize to %u on device %s",
1636 sectorsize, name);
1637 return -EINVAL;
1640 /* Set up device logical sector size mask */
1641 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1642 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1644 return 0;
1648 * When allocating the initial buffer target we have not yet
1649 * read in the superblock, so don't know what sized sectors
1650 * are being used at this early stage. Play safe.
1652 STATIC int
1653 xfs_setsize_buftarg_early(
1654 xfs_buftarg_t *btp,
1655 struct block_device *bdev)
1657 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1660 xfs_buftarg_t *
1661 xfs_alloc_buftarg(
1662 struct xfs_mount *mp,
1663 struct block_device *bdev)
1665 xfs_buftarg_t *btp;
1667 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS);
1669 btp->bt_mount = mp;
1670 btp->bt_dev = bdev->bd_dev;
1671 btp->bt_bdev = bdev;
1672 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1674 if (xfs_setsize_buftarg_early(btp, bdev))
1675 goto error;
1677 if (list_lru_init(&btp->bt_lru))
1678 goto error;
1680 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1681 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1682 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1683 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1684 register_shrinker(&btp->bt_shrinker);
1685 return btp;
1687 error:
1688 kmem_free(btp);
1689 return NULL;
1693 * Add a buffer to the delayed write list.
1695 * This queues a buffer for writeout if it hasn't already been. Note that
1696 * neither this routine nor the buffer list submission functions perform
1697 * any internal synchronization. It is expected that the lists are thread-local
1698 * to the callers.
1700 * Returns true if we queued up the buffer, or false if it already had
1701 * been on the buffer list.
1703 bool
1704 xfs_buf_delwri_queue(
1705 struct xfs_buf *bp,
1706 struct list_head *list)
1708 ASSERT(xfs_buf_islocked(bp));
1709 ASSERT(!(bp->b_flags & XBF_READ));
1712 * If the buffer is already marked delwri it already is queued up
1713 * by someone else for imediate writeout. Just ignore it in that
1714 * case.
1716 if (bp->b_flags & _XBF_DELWRI_Q) {
1717 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1718 return false;
1721 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1724 * If a buffer gets written out synchronously or marked stale while it
1725 * is on a delwri list we lazily remove it. To do this, the other party
1726 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1727 * It remains referenced and on the list. In a rare corner case it
1728 * might get readded to a delwri list after the synchronous writeout, in
1729 * which case we need just need to re-add the flag here.
1731 bp->b_flags |= _XBF_DELWRI_Q;
1732 if (list_empty(&bp->b_list)) {
1733 atomic_inc(&bp->b_hold);
1734 list_add_tail(&bp->b_list, list);
1737 return true;
1741 * Compare function is more complex than it needs to be because
1742 * the return value is only 32 bits and we are doing comparisons
1743 * on 64 bit values
1745 static int
1746 xfs_buf_cmp(
1747 void *priv,
1748 struct list_head *a,
1749 struct list_head *b)
1751 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1752 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1753 xfs_daddr_t diff;
1755 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
1756 if (diff < 0)
1757 return -1;
1758 if (diff > 0)
1759 return 1;
1760 return 0;
1763 static int
1764 __xfs_buf_delwri_submit(
1765 struct list_head *buffer_list,
1766 struct list_head *io_list,
1767 bool wait)
1769 struct blk_plug plug;
1770 struct xfs_buf *bp, *n;
1771 int pinned = 0;
1773 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1774 if (!wait) {
1775 if (xfs_buf_ispinned(bp)) {
1776 pinned++;
1777 continue;
1779 if (!xfs_buf_trylock(bp))
1780 continue;
1781 } else {
1782 xfs_buf_lock(bp);
1786 * Someone else might have written the buffer synchronously or
1787 * marked it stale in the meantime. In that case only the
1788 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1789 * reference and remove it from the list here.
1791 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1792 list_del_init(&bp->b_list);
1793 xfs_buf_relse(bp);
1794 continue;
1797 list_move_tail(&bp->b_list, io_list);
1798 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1801 list_sort(NULL, io_list, xfs_buf_cmp);
1803 blk_start_plug(&plug);
1804 list_for_each_entry_safe(bp, n, io_list, b_list) {
1805 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC | XBF_WRITE_FAIL);
1806 bp->b_flags |= XBF_WRITE | XBF_ASYNC;
1809 * we do all Io submission async. This means if we need to wait
1810 * for IO completion we need to take an extra reference so the
1811 * buffer is still valid on the other side.
1813 if (wait)
1814 xfs_buf_hold(bp);
1815 else
1816 list_del_init(&bp->b_list);
1818 xfs_buf_submit(bp);
1820 blk_finish_plug(&plug);
1822 return pinned;
1826 * Write out a buffer list asynchronously.
1828 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1829 * out and not wait for I/O completion on any of the buffers. This interface
1830 * is only safely useable for callers that can track I/O completion by higher
1831 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1832 * function.
1835 xfs_buf_delwri_submit_nowait(
1836 struct list_head *buffer_list)
1838 LIST_HEAD (io_list);
1839 return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1843 * Write out a buffer list synchronously.
1845 * This will take the @buffer_list, write all buffers out and wait for I/O
1846 * completion on all of the buffers. @buffer_list is consumed by the function,
1847 * so callers must have some other way of tracking buffers if they require such
1848 * functionality.
1851 xfs_buf_delwri_submit(
1852 struct list_head *buffer_list)
1854 LIST_HEAD (io_list);
1855 int error = 0, error2;
1856 struct xfs_buf *bp;
1858 __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1860 /* Wait for IO to complete. */
1861 while (!list_empty(&io_list)) {
1862 bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1864 list_del_init(&bp->b_list);
1866 /* locking the buffer will wait for async IO completion. */
1867 xfs_buf_lock(bp);
1868 error2 = bp->b_error;
1869 xfs_buf_relse(bp);
1870 if (!error)
1871 error = error2;
1874 return error;
1877 int __init
1878 xfs_buf_init(void)
1880 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1881 KM_ZONE_HWALIGN, NULL);
1882 if (!xfs_buf_zone)
1883 goto out;
1885 xfslogd_workqueue = alloc_workqueue("xfslogd",
1886 WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_FREEZABLE, 1);
1887 if (!xfslogd_workqueue)
1888 goto out_free_buf_zone;
1890 return 0;
1892 out_free_buf_zone:
1893 kmem_zone_destroy(xfs_buf_zone);
1894 out:
1895 return -ENOMEM;
1898 void
1899 xfs_buf_terminate(void)
1901 destroy_workqueue(xfslogd_workqueue);
1902 kmem_zone_destroy(xfs_buf_zone);