Loading...
1/*
2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
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.
8 *
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.
13 *
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
17 */
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
37#include "xfs_sb.h"
38#include "xfs_inum.h"
39#include "xfs_log.h"
40#include "xfs_ag.h"
41#include "xfs_mount.h"
42#include "xfs_trace.h"
43
44static kmem_zone_t *xfs_buf_zone;
45STATIC int xfsbufd(void *);
46STATIC void xfs_buf_delwri_queue(xfs_buf_t *, int);
47
48static struct workqueue_struct *xfslogd_workqueue;
49struct workqueue_struct *xfsdatad_workqueue;
50struct workqueue_struct *xfsconvertd_workqueue;
51
52#ifdef XFS_BUF_LOCK_TRACKING
53# define XB_SET_OWNER(bp) ((bp)->b_last_holder = current->pid)
54# define XB_CLEAR_OWNER(bp) ((bp)->b_last_holder = -1)
55# define XB_GET_OWNER(bp) ((bp)->b_last_holder)
56#else
57# define XB_SET_OWNER(bp) do { } while (0)
58# define XB_CLEAR_OWNER(bp) do { } while (0)
59# define XB_GET_OWNER(bp) do { } while (0)
60#endif
61
62#define xb_to_gfp(flags) \
63 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : \
64 ((flags) & XBF_DONT_BLOCK) ? GFP_NOFS : GFP_KERNEL) | __GFP_NOWARN)
65
66#define xb_to_km(flags) \
67 (((flags) & XBF_DONT_BLOCK) ? KM_NOFS : KM_SLEEP)
68
69#define xfs_buf_allocate(flags) \
70 kmem_zone_alloc(xfs_buf_zone, xb_to_km(flags))
71#define xfs_buf_deallocate(bp) \
72 kmem_zone_free(xfs_buf_zone, (bp));
73
74static inline int
75xfs_buf_is_vmapped(
76 struct xfs_buf *bp)
77{
78 /*
79 * Return true if the buffer is vmapped.
80 *
81 * The XBF_MAPPED flag is set if the buffer should be mapped, but the
82 * code is clever enough to know it doesn't have to map a single page,
83 * so the check has to be both for XBF_MAPPED and bp->b_page_count > 1.
84 */
85 return (bp->b_flags & XBF_MAPPED) && bp->b_page_count > 1;
86}
87
88static inline int
89xfs_buf_vmap_len(
90 struct xfs_buf *bp)
91{
92 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
93}
94
95/*
96 * xfs_buf_lru_add - add a buffer to the LRU.
97 *
98 * The LRU takes a new reference to the buffer so that it will only be freed
99 * once the shrinker takes the buffer off the LRU.
100 */
101STATIC void
102xfs_buf_lru_add(
103 struct xfs_buf *bp)
104{
105 struct xfs_buftarg *btp = bp->b_target;
106
107 spin_lock(&btp->bt_lru_lock);
108 if (list_empty(&bp->b_lru)) {
109 atomic_inc(&bp->b_hold);
110 list_add_tail(&bp->b_lru, &btp->bt_lru);
111 btp->bt_lru_nr++;
112 }
113 spin_unlock(&btp->bt_lru_lock);
114}
115
116/*
117 * xfs_buf_lru_del - remove a buffer from the LRU
118 *
119 * The unlocked check is safe here because it only occurs when there are not
120 * b_lru_ref counts left on the inode under the pag->pag_buf_lock. it is there
121 * to optimise the shrinker removing the buffer from the LRU and calling
122 * xfs_buf_free(). i.e. it removes an unnecessary round trip on the
123 * bt_lru_lock.
124 */
125STATIC void
126xfs_buf_lru_del(
127 struct xfs_buf *bp)
128{
129 struct xfs_buftarg *btp = bp->b_target;
130
131 if (list_empty(&bp->b_lru))
132 return;
133
134 spin_lock(&btp->bt_lru_lock);
135 if (!list_empty(&bp->b_lru)) {
136 list_del_init(&bp->b_lru);
137 btp->bt_lru_nr--;
138 }
139 spin_unlock(&btp->bt_lru_lock);
140}
141
142/*
143 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
144 * b_lru_ref count so that the buffer is freed immediately when the buffer
145 * reference count falls to zero. If the buffer is already on the LRU, we need
146 * to remove the reference that LRU holds on the buffer.
147 *
148 * This prevents build-up of stale buffers on the LRU.
149 */
150void
151xfs_buf_stale(
152 struct xfs_buf *bp)
153{
154 bp->b_flags |= XBF_STALE;
155 atomic_set(&(bp)->b_lru_ref, 0);
156 if (!list_empty(&bp->b_lru)) {
157 struct xfs_buftarg *btp = bp->b_target;
158
159 spin_lock(&btp->bt_lru_lock);
160 if (!list_empty(&bp->b_lru)) {
161 list_del_init(&bp->b_lru);
162 btp->bt_lru_nr--;
163 atomic_dec(&bp->b_hold);
164 }
165 spin_unlock(&btp->bt_lru_lock);
166 }
167 ASSERT(atomic_read(&bp->b_hold) >= 1);
168}
169
170STATIC void
171_xfs_buf_initialize(
172 xfs_buf_t *bp,
173 xfs_buftarg_t *target,
174 xfs_off_t range_base,
175 size_t range_length,
176 xfs_buf_flags_t flags)
177{
178 /*
179 * We don't want certain flags to appear in b_flags.
180 */
181 flags &= ~(XBF_LOCK|XBF_MAPPED|XBF_DONT_BLOCK|XBF_READ_AHEAD);
182
183 memset(bp, 0, sizeof(xfs_buf_t));
184 atomic_set(&bp->b_hold, 1);
185 atomic_set(&bp->b_lru_ref, 1);
186 init_completion(&bp->b_iowait);
187 INIT_LIST_HEAD(&bp->b_lru);
188 INIT_LIST_HEAD(&bp->b_list);
189 RB_CLEAR_NODE(&bp->b_rbnode);
190 sema_init(&bp->b_sema, 0); /* held, no waiters */
191 XB_SET_OWNER(bp);
192 bp->b_target = target;
193 bp->b_file_offset = range_base;
194 /*
195 * Set buffer_length and count_desired to the same value initially.
196 * I/O routines should use count_desired, which will be the same in
197 * most cases but may be reset (e.g. XFS recovery).
198 */
199 bp->b_buffer_length = bp->b_count_desired = range_length;
200 bp->b_flags = flags;
201 bp->b_bn = XFS_BUF_DADDR_NULL;
202 atomic_set(&bp->b_pin_count, 0);
203 init_waitqueue_head(&bp->b_waiters);
204
205 XFS_STATS_INC(xb_create);
206
207 trace_xfs_buf_init(bp, _RET_IP_);
208}
209
210/*
211 * Allocate a page array capable of holding a specified number
212 * of pages, and point the page buf at it.
213 */
214STATIC int
215_xfs_buf_get_pages(
216 xfs_buf_t *bp,
217 int page_count,
218 xfs_buf_flags_t flags)
219{
220 /* Make sure that we have a page list */
221 if (bp->b_pages == NULL) {
222 bp->b_offset = xfs_buf_poff(bp->b_file_offset);
223 bp->b_page_count = page_count;
224 if (page_count <= XB_PAGES) {
225 bp->b_pages = bp->b_page_array;
226 } else {
227 bp->b_pages = kmem_alloc(sizeof(struct page *) *
228 page_count, xb_to_km(flags));
229 if (bp->b_pages == NULL)
230 return -ENOMEM;
231 }
232 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
233 }
234 return 0;
235}
236
237/*
238 * Frees b_pages if it was allocated.
239 */
240STATIC void
241_xfs_buf_free_pages(
242 xfs_buf_t *bp)
243{
244 if (bp->b_pages != bp->b_page_array) {
245 kmem_free(bp->b_pages);
246 bp->b_pages = NULL;
247 }
248}
249
250/*
251 * Releases the specified buffer.
252 *
253 * The modification state of any associated pages is left unchanged.
254 * The buffer most not be on any hash - use xfs_buf_rele instead for
255 * hashed and refcounted buffers
256 */
257void
258xfs_buf_free(
259 xfs_buf_t *bp)
260{
261 trace_xfs_buf_free(bp, _RET_IP_);
262
263 ASSERT(list_empty(&bp->b_lru));
264
265 if (bp->b_flags & _XBF_PAGES) {
266 uint i;
267
268 if (xfs_buf_is_vmapped(bp))
269 vm_unmap_ram(bp->b_addr - bp->b_offset,
270 bp->b_page_count);
271
272 for (i = 0; i < bp->b_page_count; i++) {
273 struct page *page = bp->b_pages[i];
274
275 __free_page(page);
276 }
277 } else if (bp->b_flags & _XBF_KMEM)
278 kmem_free(bp->b_addr);
279 _xfs_buf_free_pages(bp);
280 xfs_buf_deallocate(bp);
281}
282
283/*
284 * Allocates all the pages for buffer in question and builds it's page list.
285 */
286STATIC int
287xfs_buf_allocate_memory(
288 xfs_buf_t *bp,
289 uint flags)
290{
291 size_t size = bp->b_count_desired;
292 size_t nbytes, offset;
293 gfp_t gfp_mask = xb_to_gfp(flags);
294 unsigned short page_count, i;
295 xfs_off_t end;
296 int error;
297
298 /*
299 * for buffers that are contained within a single page, just allocate
300 * the memory from the heap - there's no need for the complexity of
301 * page arrays to keep allocation down to order 0.
302 */
303 if (bp->b_buffer_length < PAGE_SIZE) {
304 bp->b_addr = kmem_alloc(bp->b_buffer_length, xb_to_km(flags));
305 if (!bp->b_addr) {
306 /* low memory - use alloc_page loop instead */
307 goto use_alloc_page;
308 }
309
310 if (((unsigned long)(bp->b_addr + bp->b_buffer_length - 1) &
311 PAGE_MASK) !=
312 ((unsigned long)bp->b_addr & PAGE_MASK)) {
313 /* b_addr spans two pages - use alloc_page instead */
314 kmem_free(bp->b_addr);
315 bp->b_addr = NULL;
316 goto use_alloc_page;
317 }
318 bp->b_offset = offset_in_page(bp->b_addr);
319 bp->b_pages = bp->b_page_array;
320 bp->b_pages[0] = virt_to_page(bp->b_addr);
321 bp->b_page_count = 1;
322 bp->b_flags |= XBF_MAPPED | _XBF_KMEM;
323 return 0;
324 }
325
326use_alloc_page:
327 end = bp->b_file_offset + bp->b_buffer_length;
328 page_count = xfs_buf_btoc(end) - xfs_buf_btoct(bp->b_file_offset);
329 error = _xfs_buf_get_pages(bp, page_count, flags);
330 if (unlikely(error))
331 return error;
332
333 offset = bp->b_offset;
334 bp->b_flags |= _XBF_PAGES;
335
336 for (i = 0; i < bp->b_page_count; i++) {
337 struct page *page;
338 uint retries = 0;
339retry:
340 page = alloc_page(gfp_mask);
341 if (unlikely(page == NULL)) {
342 if (flags & XBF_READ_AHEAD) {
343 bp->b_page_count = i;
344 error = ENOMEM;
345 goto out_free_pages;
346 }
347
348 /*
349 * This could deadlock.
350 *
351 * But until all the XFS lowlevel code is revamped to
352 * handle buffer allocation failures we can't do much.
353 */
354 if (!(++retries % 100))
355 xfs_err(NULL,
356 "possible memory allocation deadlock in %s (mode:0x%x)",
357 __func__, gfp_mask);
358
359 XFS_STATS_INC(xb_page_retries);
360 congestion_wait(BLK_RW_ASYNC, HZ/50);
361 goto retry;
362 }
363
364 XFS_STATS_INC(xb_page_found);
365
366 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
367 size -= nbytes;
368 bp->b_pages[i] = page;
369 offset = 0;
370 }
371 return 0;
372
373out_free_pages:
374 for (i = 0; i < bp->b_page_count; i++)
375 __free_page(bp->b_pages[i]);
376 return error;
377}
378
379/*
380 * Map buffer into kernel address-space if necessary.
381 */
382STATIC int
383_xfs_buf_map_pages(
384 xfs_buf_t *bp,
385 uint flags)
386{
387 ASSERT(bp->b_flags & _XBF_PAGES);
388 if (bp->b_page_count == 1) {
389 /* A single page buffer is always mappable */
390 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
391 bp->b_flags |= XBF_MAPPED;
392 } else if (flags & XBF_MAPPED) {
393 int retried = 0;
394
395 do {
396 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
397 -1, PAGE_KERNEL);
398 if (bp->b_addr)
399 break;
400 vm_unmap_aliases();
401 } while (retried++ <= 1);
402
403 if (!bp->b_addr)
404 return -ENOMEM;
405 bp->b_addr += bp->b_offset;
406 bp->b_flags |= XBF_MAPPED;
407 }
408
409 return 0;
410}
411
412/*
413 * Finding and Reading Buffers
414 */
415
416/*
417 * Look up, and creates if absent, a lockable buffer for
418 * a given range of an inode. The buffer is returned
419 * locked. If other overlapping buffers exist, they are
420 * released before the new buffer is created and locked,
421 * which may imply that this call will block until those buffers
422 * are unlocked. No I/O is implied by this call.
423 */
424xfs_buf_t *
425_xfs_buf_find(
426 xfs_buftarg_t *btp, /* block device target */
427 xfs_off_t ioff, /* starting offset of range */
428 size_t isize, /* length of range */
429 xfs_buf_flags_t flags,
430 xfs_buf_t *new_bp)
431{
432 xfs_off_t range_base;
433 size_t range_length;
434 struct xfs_perag *pag;
435 struct rb_node **rbp;
436 struct rb_node *parent;
437 xfs_buf_t *bp;
438
439 range_base = (ioff << BBSHIFT);
440 range_length = (isize << BBSHIFT);
441
442 /* Check for IOs smaller than the sector size / not sector aligned */
443 ASSERT(!(range_length < (1 << btp->bt_sshift)));
444 ASSERT(!(range_base & (xfs_off_t)btp->bt_smask));
445
446 /* get tree root */
447 pag = xfs_perag_get(btp->bt_mount,
448 xfs_daddr_to_agno(btp->bt_mount, ioff));
449
450 /* walk tree */
451 spin_lock(&pag->pag_buf_lock);
452 rbp = &pag->pag_buf_tree.rb_node;
453 parent = NULL;
454 bp = NULL;
455 while (*rbp) {
456 parent = *rbp;
457 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
458
459 if (range_base < bp->b_file_offset)
460 rbp = &(*rbp)->rb_left;
461 else if (range_base > bp->b_file_offset)
462 rbp = &(*rbp)->rb_right;
463 else {
464 /*
465 * found a block offset match. If the range doesn't
466 * match, the only way this is allowed is if the buffer
467 * in the cache is stale and the transaction that made
468 * it stale has not yet committed. i.e. we are
469 * reallocating a busy extent. Skip this buffer and
470 * continue searching to the right for an exact match.
471 */
472 if (bp->b_buffer_length != range_length) {
473 ASSERT(bp->b_flags & XBF_STALE);
474 rbp = &(*rbp)->rb_right;
475 continue;
476 }
477 atomic_inc(&bp->b_hold);
478 goto found;
479 }
480 }
481
482 /* No match found */
483 if (new_bp) {
484 _xfs_buf_initialize(new_bp, btp, range_base,
485 range_length, flags);
486 rb_link_node(&new_bp->b_rbnode, parent, rbp);
487 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
488 /* the buffer keeps the perag reference until it is freed */
489 new_bp->b_pag = pag;
490 spin_unlock(&pag->pag_buf_lock);
491 } else {
492 XFS_STATS_INC(xb_miss_locked);
493 spin_unlock(&pag->pag_buf_lock);
494 xfs_perag_put(pag);
495 }
496 return new_bp;
497
498found:
499 spin_unlock(&pag->pag_buf_lock);
500 xfs_perag_put(pag);
501
502 if (!xfs_buf_trylock(bp)) {
503 if (flags & XBF_TRYLOCK) {
504 xfs_buf_rele(bp);
505 XFS_STATS_INC(xb_busy_locked);
506 return NULL;
507 }
508 xfs_buf_lock(bp);
509 XFS_STATS_INC(xb_get_locked_waited);
510 }
511
512 /*
513 * if the buffer is stale, clear all the external state associated with
514 * it. We need to keep flags such as how we allocated the buffer memory
515 * intact here.
516 */
517 if (bp->b_flags & XBF_STALE) {
518 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
519 bp->b_flags &= XBF_MAPPED | _XBF_KMEM | _XBF_PAGES;
520 }
521
522 trace_xfs_buf_find(bp, flags, _RET_IP_);
523 XFS_STATS_INC(xb_get_locked);
524 return bp;
525}
526
527/*
528 * Assembles a buffer covering the specified range.
529 * Storage in memory for all portions of the buffer will be allocated,
530 * although backing storage may not be.
531 */
532xfs_buf_t *
533xfs_buf_get(
534 xfs_buftarg_t *target,/* target for buffer */
535 xfs_off_t ioff, /* starting offset of range */
536 size_t isize, /* length of range */
537 xfs_buf_flags_t flags)
538{
539 xfs_buf_t *bp, *new_bp;
540 int error = 0;
541
542 new_bp = xfs_buf_allocate(flags);
543 if (unlikely(!new_bp))
544 return NULL;
545
546 bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
547 if (bp == new_bp) {
548 error = xfs_buf_allocate_memory(bp, flags);
549 if (error)
550 goto no_buffer;
551 } else {
552 xfs_buf_deallocate(new_bp);
553 if (unlikely(bp == NULL))
554 return NULL;
555 }
556
557 if (!(bp->b_flags & XBF_MAPPED)) {
558 error = _xfs_buf_map_pages(bp, flags);
559 if (unlikely(error)) {
560 xfs_warn(target->bt_mount,
561 "%s: failed to map pages\n", __func__);
562 goto no_buffer;
563 }
564 }
565
566 XFS_STATS_INC(xb_get);
567
568 /*
569 * Always fill in the block number now, the mapped cases can do
570 * their own overlay of this later.
571 */
572 bp->b_bn = ioff;
573 bp->b_count_desired = bp->b_buffer_length;
574
575 trace_xfs_buf_get(bp, flags, _RET_IP_);
576 return bp;
577
578 no_buffer:
579 if (flags & (XBF_LOCK | XBF_TRYLOCK))
580 xfs_buf_unlock(bp);
581 xfs_buf_rele(bp);
582 return NULL;
583}
584
585STATIC int
586_xfs_buf_read(
587 xfs_buf_t *bp,
588 xfs_buf_flags_t flags)
589{
590 int status;
591
592 ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
593 ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
594
595 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | XBF_READ_AHEAD);
596 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
597
598 status = xfs_buf_iorequest(bp);
599 if (status || bp->b_error || (flags & XBF_ASYNC))
600 return status;
601 return xfs_buf_iowait(bp);
602}
603
604xfs_buf_t *
605xfs_buf_read(
606 xfs_buftarg_t *target,
607 xfs_off_t ioff,
608 size_t isize,
609 xfs_buf_flags_t flags)
610{
611 xfs_buf_t *bp;
612
613 flags |= XBF_READ;
614
615 bp = xfs_buf_get(target, ioff, isize, flags);
616 if (bp) {
617 trace_xfs_buf_read(bp, flags, _RET_IP_);
618
619 if (!XFS_BUF_ISDONE(bp)) {
620 XFS_STATS_INC(xb_get_read);
621 _xfs_buf_read(bp, flags);
622 } else if (flags & XBF_ASYNC) {
623 /*
624 * Read ahead call which is already satisfied,
625 * drop the buffer
626 */
627 goto no_buffer;
628 } else {
629 /* We do not want read in the flags */
630 bp->b_flags &= ~XBF_READ;
631 }
632 }
633
634 return bp;
635
636 no_buffer:
637 if (flags & (XBF_LOCK | XBF_TRYLOCK))
638 xfs_buf_unlock(bp);
639 xfs_buf_rele(bp);
640 return NULL;
641}
642
643/*
644 * If we are not low on memory then do the readahead in a deadlock
645 * safe manner.
646 */
647void
648xfs_buf_readahead(
649 xfs_buftarg_t *target,
650 xfs_off_t ioff,
651 size_t isize)
652{
653 if (bdi_read_congested(target->bt_bdi))
654 return;
655
656 xfs_buf_read(target, ioff, isize,
657 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD|XBF_DONT_BLOCK);
658}
659
660/*
661 * Read an uncached buffer from disk. Allocates and returns a locked
662 * buffer containing the disk contents or nothing.
663 */
664struct xfs_buf *
665xfs_buf_read_uncached(
666 struct xfs_mount *mp,
667 struct xfs_buftarg *target,
668 xfs_daddr_t daddr,
669 size_t length,
670 int flags)
671{
672 xfs_buf_t *bp;
673 int error;
674
675 bp = xfs_buf_get_uncached(target, length, flags);
676 if (!bp)
677 return NULL;
678
679 /* set up the buffer for a read IO */
680 XFS_BUF_SET_ADDR(bp, daddr);
681 XFS_BUF_READ(bp);
682
683 xfsbdstrat(mp, bp);
684 error = xfs_buf_iowait(bp);
685 if (error || bp->b_error) {
686 xfs_buf_relse(bp);
687 return NULL;
688 }
689 return bp;
690}
691
692xfs_buf_t *
693xfs_buf_get_empty(
694 size_t len,
695 xfs_buftarg_t *target)
696{
697 xfs_buf_t *bp;
698
699 bp = xfs_buf_allocate(0);
700 if (bp)
701 _xfs_buf_initialize(bp, target, 0, len, 0);
702 return bp;
703}
704
705/*
706 * Return a buffer allocated as an empty buffer and associated to external
707 * memory via xfs_buf_associate_memory() back to it's empty state.
708 */
709void
710xfs_buf_set_empty(
711 struct xfs_buf *bp,
712 size_t len)
713{
714 if (bp->b_pages)
715 _xfs_buf_free_pages(bp);
716
717 bp->b_pages = NULL;
718 bp->b_page_count = 0;
719 bp->b_addr = NULL;
720 bp->b_file_offset = 0;
721 bp->b_buffer_length = bp->b_count_desired = len;
722 bp->b_bn = XFS_BUF_DADDR_NULL;
723 bp->b_flags &= ~XBF_MAPPED;
724}
725
726static inline struct page *
727mem_to_page(
728 void *addr)
729{
730 if ((!is_vmalloc_addr(addr))) {
731 return virt_to_page(addr);
732 } else {
733 return vmalloc_to_page(addr);
734 }
735}
736
737int
738xfs_buf_associate_memory(
739 xfs_buf_t *bp,
740 void *mem,
741 size_t len)
742{
743 int rval;
744 int i = 0;
745 unsigned long pageaddr;
746 unsigned long offset;
747 size_t buflen;
748 int page_count;
749
750 pageaddr = (unsigned long)mem & PAGE_MASK;
751 offset = (unsigned long)mem - pageaddr;
752 buflen = PAGE_ALIGN(len + offset);
753 page_count = buflen >> PAGE_SHIFT;
754
755 /* Free any previous set of page pointers */
756 if (bp->b_pages)
757 _xfs_buf_free_pages(bp);
758
759 bp->b_pages = NULL;
760 bp->b_addr = mem;
761
762 rval = _xfs_buf_get_pages(bp, page_count, XBF_DONT_BLOCK);
763 if (rval)
764 return rval;
765
766 bp->b_offset = offset;
767
768 for (i = 0; i < bp->b_page_count; i++) {
769 bp->b_pages[i] = mem_to_page((void *)pageaddr);
770 pageaddr += PAGE_SIZE;
771 }
772
773 bp->b_count_desired = len;
774 bp->b_buffer_length = buflen;
775 bp->b_flags |= XBF_MAPPED;
776
777 return 0;
778}
779
780xfs_buf_t *
781xfs_buf_get_uncached(
782 struct xfs_buftarg *target,
783 size_t len,
784 int flags)
785{
786 unsigned long page_count = PAGE_ALIGN(len) >> PAGE_SHIFT;
787 int error, i;
788 xfs_buf_t *bp;
789
790 bp = xfs_buf_allocate(0);
791 if (unlikely(bp == NULL))
792 goto fail;
793 _xfs_buf_initialize(bp, target, 0, len, 0);
794
795 error = _xfs_buf_get_pages(bp, page_count, 0);
796 if (error)
797 goto fail_free_buf;
798
799 for (i = 0; i < page_count; i++) {
800 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
801 if (!bp->b_pages[i])
802 goto fail_free_mem;
803 }
804 bp->b_flags |= _XBF_PAGES;
805
806 error = _xfs_buf_map_pages(bp, XBF_MAPPED);
807 if (unlikely(error)) {
808 xfs_warn(target->bt_mount,
809 "%s: failed to map pages\n", __func__);
810 goto fail_free_mem;
811 }
812
813 trace_xfs_buf_get_uncached(bp, _RET_IP_);
814 return bp;
815
816 fail_free_mem:
817 while (--i >= 0)
818 __free_page(bp->b_pages[i]);
819 _xfs_buf_free_pages(bp);
820 fail_free_buf:
821 xfs_buf_deallocate(bp);
822 fail:
823 return NULL;
824}
825
826/*
827 * Increment reference count on buffer, to hold the buffer concurrently
828 * with another thread which may release (free) the buffer asynchronously.
829 * Must hold the buffer already to call this function.
830 */
831void
832xfs_buf_hold(
833 xfs_buf_t *bp)
834{
835 trace_xfs_buf_hold(bp, _RET_IP_);
836 atomic_inc(&bp->b_hold);
837}
838
839/*
840 * Releases a hold on the specified buffer. If the
841 * the hold count is 1, calls xfs_buf_free.
842 */
843void
844xfs_buf_rele(
845 xfs_buf_t *bp)
846{
847 struct xfs_perag *pag = bp->b_pag;
848
849 trace_xfs_buf_rele(bp, _RET_IP_);
850
851 if (!pag) {
852 ASSERT(list_empty(&bp->b_lru));
853 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
854 if (atomic_dec_and_test(&bp->b_hold))
855 xfs_buf_free(bp);
856 return;
857 }
858
859 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
860
861 ASSERT(atomic_read(&bp->b_hold) > 0);
862 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
863 if (!(bp->b_flags & XBF_STALE) &&
864 atomic_read(&bp->b_lru_ref)) {
865 xfs_buf_lru_add(bp);
866 spin_unlock(&pag->pag_buf_lock);
867 } else {
868 xfs_buf_lru_del(bp);
869 ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
870 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
871 spin_unlock(&pag->pag_buf_lock);
872 xfs_perag_put(pag);
873 xfs_buf_free(bp);
874 }
875 }
876}
877
878
879/*
880 * Lock a buffer object, if it is not already locked.
881 *
882 * If we come across a stale, pinned, locked buffer, we know that we are
883 * being asked to lock a buffer that has been reallocated. Because it is
884 * pinned, we know that the log has not been pushed to disk and hence it
885 * will still be locked. Rather than continuing to have trylock attempts
886 * fail until someone else pushes the log, push it ourselves before
887 * returning. This means that the xfsaild will not get stuck trying
888 * to push on stale inode buffers.
889 */
890int
891xfs_buf_trylock(
892 struct xfs_buf *bp)
893{
894 int locked;
895
896 locked = down_trylock(&bp->b_sema) == 0;
897 if (locked)
898 XB_SET_OWNER(bp);
899 else if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
900 xfs_log_force(bp->b_target->bt_mount, 0);
901
902 trace_xfs_buf_trylock(bp, _RET_IP_);
903 return locked;
904}
905
906/*
907 * Lock a buffer object.
908 *
909 * If we come across a stale, pinned, locked buffer, we know that we
910 * are being asked to lock a buffer that has been reallocated. Because
911 * it is pinned, we know that the log has not been pushed to disk and
912 * hence it will still be locked. Rather than sleeping until someone
913 * else pushes the log, push it ourselves before trying to get the lock.
914 */
915void
916xfs_buf_lock(
917 struct xfs_buf *bp)
918{
919 trace_xfs_buf_lock(bp, _RET_IP_);
920
921 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
922 xfs_log_force(bp->b_target->bt_mount, 0);
923 down(&bp->b_sema);
924 XB_SET_OWNER(bp);
925
926 trace_xfs_buf_lock_done(bp, _RET_IP_);
927}
928
929/*
930 * Releases the lock on the buffer object.
931 * If the buffer is marked delwri but is not queued, do so before we
932 * unlock the buffer as we need to set flags correctly. We also need to
933 * take a reference for the delwri queue because the unlocker is going to
934 * drop their's and they don't know we just queued it.
935 */
936void
937xfs_buf_unlock(
938 struct xfs_buf *bp)
939{
940 if ((bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)) == XBF_DELWRI) {
941 atomic_inc(&bp->b_hold);
942 bp->b_flags |= XBF_ASYNC;
943 xfs_buf_delwri_queue(bp, 0);
944 }
945
946 XB_CLEAR_OWNER(bp);
947 up(&bp->b_sema);
948
949 trace_xfs_buf_unlock(bp, _RET_IP_);
950}
951
952STATIC void
953xfs_buf_wait_unpin(
954 xfs_buf_t *bp)
955{
956 DECLARE_WAITQUEUE (wait, current);
957
958 if (atomic_read(&bp->b_pin_count) == 0)
959 return;
960
961 add_wait_queue(&bp->b_waiters, &wait);
962 for (;;) {
963 set_current_state(TASK_UNINTERRUPTIBLE);
964 if (atomic_read(&bp->b_pin_count) == 0)
965 break;
966 io_schedule();
967 }
968 remove_wait_queue(&bp->b_waiters, &wait);
969 set_current_state(TASK_RUNNING);
970}
971
972/*
973 * Buffer Utility Routines
974 */
975
976STATIC void
977xfs_buf_iodone_work(
978 struct work_struct *work)
979{
980 xfs_buf_t *bp =
981 container_of(work, xfs_buf_t, b_iodone_work);
982
983 if (bp->b_iodone)
984 (*(bp->b_iodone))(bp);
985 else if (bp->b_flags & XBF_ASYNC)
986 xfs_buf_relse(bp);
987}
988
989void
990xfs_buf_ioend(
991 xfs_buf_t *bp,
992 int schedule)
993{
994 trace_xfs_buf_iodone(bp, _RET_IP_);
995
996 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
997 if (bp->b_error == 0)
998 bp->b_flags |= XBF_DONE;
999
1000 if ((bp->b_iodone) || (bp->b_flags & XBF_ASYNC)) {
1001 if (schedule) {
1002 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1003 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1004 } else {
1005 xfs_buf_iodone_work(&bp->b_iodone_work);
1006 }
1007 } else {
1008 complete(&bp->b_iowait);
1009 }
1010}
1011
1012void
1013xfs_buf_ioerror(
1014 xfs_buf_t *bp,
1015 int error)
1016{
1017 ASSERT(error >= 0 && error <= 0xffff);
1018 bp->b_error = (unsigned short)error;
1019 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1020}
1021
1022int
1023xfs_bwrite(
1024 struct xfs_mount *mp,
1025 struct xfs_buf *bp)
1026{
1027 int error;
1028
1029 bp->b_flags |= XBF_WRITE;
1030 bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
1031
1032 xfs_buf_delwri_dequeue(bp);
1033 xfs_bdstrat_cb(bp);
1034
1035 error = xfs_buf_iowait(bp);
1036 if (error)
1037 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1038 xfs_buf_relse(bp);
1039 return error;
1040}
1041
1042void
1043xfs_bdwrite(
1044 void *mp,
1045 struct xfs_buf *bp)
1046{
1047 trace_xfs_buf_bdwrite(bp, _RET_IP_);
1048
1049 bp->b_flags &= ~XBF_READ;
1050 bp->b_flags |= (XBF_DELWRI | XBF_ASYNC);
1051
1052 xfs_buf_delwri_queue(bp, 1);
1053}
1054
1055/*
1056 * Called when we want to stop a buffer from getting written or read.
1057 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1058 * so that the proper iodone callbacks get called.
1059 */
1060STATIC int
1061xfs_bioerror(
1062 xfs_buf_t *bp)
1063{
1064#ifdef XFSERRORDEBUG
1065 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1066#endif
1067
1068 /*
1069 * No need to wait until the buffer is unpinned, we aren't flushing it.
1070 */
1071 xfs_buf_ioerror(bp, EIO);
1072
1073 /*
1074 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1075 */
1076 XFS_BUF_UNREAD(bp);
1077 XFS_BUF_UNDELAYWRITE(bp);
1078 XFS_BUF_UNDONE(bp);
1079 XFS_BUF_STALE(bp);
1080
1081 xfs_buf_ioend(bp, 0);
1082
1083 return EIO;
1084}
1085
1086/*
1087 * Same as xfs_bioerror, except that we are releasing the buffer
1088 * here ourselves, and avoiding the xfs_buf_ioend call.
1089 * This is meant for userdata errors; metadata bufs come with
1090 * iodone functions attached, so that we can track down errors.
1091 */
1092STATIC int
1093xfs_bioerror_relse(
1094 struct xfs_buf *bp)
1095{
1096 int64_t fl = bp->b_flags;
1097 /*
1098 * No need to wait until the buffer is unpinned.
1099 * We aren't flushing it.
1100 *
1101 * chunkhold expects B_DONE to be set, whether
1102 * we actually finish the I/O or not. We don't want to
1103 * change that interface.
1104 */
1105 XFS_BUF_UNREAD(bp);
1106 XFS_BUF_UNDELAYWRITE(bp);
1107 XFS_BUF_DONE(bp);
1108 XFS_BUF_STALE(bp);
1109 bp->b_iodone = NULL;
1110 if (!(fl & XBF_ASYNC)) {
1111 /*
1112 * Mark b_error and B_ERROR _both_.
1113 * Lot's of chunkcache code assumes that.
1114 * There's no reason to mark error for
1115 * ASYNC buffers.
1116 */
1117 xfs_buf_ioerror(bp, EIO);
1118 XFS_BUF_FINISH_IOWAIT(bp);
1119 } else {
1120 xfs_buf_relse(bp);
1121 }
1122
1123 return EIO;
1124}
1125
1126
1127/*
1128 * All xfs metadata buffers except log state machine buffers
1129 * get this attached as their b_bdstrat callback function.
1130 * This is so that we can catch a buffer
1131 * after prematurely unpinning it to forcibly shutdown the filesystem.
1132 */
1133int
1134xfs_bdstrat_cb(
1135 struct xfs_buf *bp)
1136{
1137 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1138 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1139 /*
1140 * Metadata write that didn't get logged but
1141 * written delayed anyway. These aren't associated
1142 * with a transaction, and can be ignored.
1143 */
1144 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1145 return xfs_bioerror_relse(bp);
1146 else
1147 return xfs_bioerror(bp);
1148 }
1149
1150 xfs_buf_iorequest(bp);
1151 return 0;
1152}
1153
1154/*
1155 * Wrapper around bdstrat so that we can stop data from going to disk in case
1156 * we are shutting down the filesystem. Typically user data goes thru this
1157 * path; one of the exceptions is the superblock.
1158 */
1159void
1160xfsbdstrat(
1161 struct xfs_mount *mp,
1162 struct xfs_buf *bp)
1163{
1164 if (XFS_FORCED_SHUTDOWN(mp)) {
1165 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1166 xfs_bioerror_relse(bp);
1167 return;
1168 }
1169
1170 xfs_buf_iorequest(bp);
1171}
1172
1173STATIC void
1174_xfs_buf_ioend(
1175 xfs_buf_t *bp,
1176 int schedule)
1177{
1178 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1179 xfs_buf_ioend(bp, schedule);
1180}
1181
1182STATIC void
1183xfs_buf_bio_end_io(
1184 struct bio *bio,
1185 int error)
1186{
1187 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1188
1189 xfs_buf_ioerror(bp, -error);
1190
1191 if (!error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1192 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1193
1194 _xfs_buf_ioend(bp, 1);
1195 bio_put(bio);
1196}
1197
1198STATIC void
1199_xfs_buf_ioapply(
1200 xfs_buf_t *bp)
1201{
1202 int rw, map_i, total_nr_pages, nr_pages;
1203 struct bio *bio;
1204 int offset = bp->b_offset;
1205 int size = bp->b_count_desired;
1206 sector_t sector = bp->b_bn;
1207
1208 total_nr_pages = bp->b_page_count;
1209 map_i = 0;
1210
1211 if (bp->b_flags & XBF_WRITE) {
1212 if (bp->b_flags & XBF_SYNCIO)
1213 rw = WRITE_SYNC;
1214 else
1215 rw = WRITE;
1216 if (bp->b_flags & XBF_FUA)
1217 rw |= REQ_FUA;
1218 if (bp->b_flags & XBF_FLUSH)
1219 rw |= REQ_FLUSH;
1220 } else if (bp->b_flags & XBF_READ_AHEAD) {
1221 rw = READA;
1222 } else {
1223 rw = READ;
1224 }
1225
1226 /* we only use the buffer cache for meta-data */
1227 rw |= REQ_META;
1228
1229next_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;
1234
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;
1240
1241
1242 for (; size && nr_pages; nr_pages--, map_i++) {
1243 int rbytes, nbytes = PAGE_SIZE - offset;
1244
1245 if (nbytes > size)
1246 nbytes = size;
1247
1248 rbytes = bio_add_page(bio, bp->b_pages[map_i], nbytes, offset);
1249 if (rbytes < nbytes)
1250 break;
1251
1252 offset = 0;
1253 sector += nbytes >> BBSHIFT;
1254 size -= nbytes;
1255 total_nr_pages--;
1256 }
1257
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));
1262 }
1263 submit_bio(rw, bio);
1264 if (size)
1265 goto next_chunk;
1266 } else {
1267 xfs_buf_ioerror(bp, EIO);
1268 bio_put(bio);
1269 }
1270}
1271
1272int
1273xfs_buf_iorequest(
1274 xfs_buf_t *bp)
1275{
1276 trace_xfs_buf_iorequest(bp, _RET_IP_);
1277
1278 if (bp->b_flags & XBF_DELWRI) {
1279 xfs_buf_delwri_queue(bp, 1);
1280 return 0;
1281 }
1282
1283 if (bp->b_flags & XBF_WRITE) {
1284 xfs_buf_wait_unpin(bp);
1285 }
1286
1287 xfs_buf_hold(bp);
1288
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.
1292 */
1293 atomic_set(&bp->b_io_remaining, 1);
1294 _xfs_buf_ioapply(bp);
1295 _xfs_buf_ioend(bp, 0);
1296
1297 xfs_buf_rele(bp);
1298 return 0;
1299}
1300
1301/*
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.
1305 */
1306int
1307xfs_buf_iowait(
1308 xfs_buf_t *bp)
1309{
1310 trace_xfs_buf_iowait(bp, _RET_IP_);
1311
1312 wait_for_completion(&bp->b_iowait);
1313
1314 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1315 return bp->b_error;
1316}
1317
1318xfs_caddr_t
1319xfs_buf_offset(
1320 xfs_buf_t *bp,
1321 size_t offset)
1322{
1323 struct page *page;
1324
1325 if (bp->b_flags & XBF_MAPPED)
1326 return bp->b_addr + offset;
1327
1328 offset += bp->b_offset;
1329 page = bp->b_pages[offset >> PAGE_SHIFT];
1330 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1331}
1332
1333/*
1334 * Move data into or out of a buffer.
1335 */
1336void
1337xfs_buf_iomove(
1338 xfs_buf_t *bp, /* buffer to process */
1339 size_t boff, /* starting buffer offset */
1340 size_t bsize, /* length to copy */
1341 void *data, /* data address */
1342 xfs_buf_rw_t mode) /* read/write/zero flag */
1343{
1344 size_t bend, cpoff, csize;
1345 struct page *page;
1346
1347 bend = boff + bsize;
1348 while (boff < bend) {
1349 page = bp->b_pages[xfs_buf_btoct(boff + bp->b_offset)];
1350 cpoff = xfs_buf_poff(boff + bp->b_offset);
1351 csize = min_t(size_t,
1352 PAGE_SIZE-cpoff, bp->b_count_desired-boff);
1353
1354 ASSERT(((csize + cpoff) <= PAGE_SIZE));
1355
1356 switch (mode) {
1357 case XBRW_ZERO:
1358 memset(page_address(page) + cpoff, 0, csize);
1359 break;
1360 case XBRW_READ:
1361 memcpy(data, page_address(page) + cpoff, csize);
1362 break;
1363 case XBRW_WRITE:
1364 memcpy(page_address(page) + cpoff, data, csize);
1365 }
1366
1367 boff += csize;
1368 data += csize;
1369 }
1370}
1371
1372/*
1373 * Handling of buffer targets (buftargs).
1374 */
1375
1376/*
1377 * Wait for any bufs with callbacks that have been submitted but have not yet
1378 * returned. These buffers will have an elevated hold count, so wait on those
1379 * while freeing all the buffers only held by the LRU.
1380 */
1381void
1382xfs_wait_buftarg(
1383 struct xfs_buftarg *btp)
1384{
1385 struct xfs_buf *bp;
1386
1387restart:
1388 spin_lock(&btp->bt_lru_lock);
1389 while (!list_empty(&btp->bt_lru)) {
1390 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1391 if (atomic_read(&bp->b_hold) > 1) {
1392 spin_unlock(&btp->bt_lru_lock);
1393 delay(100);
1394 goto restart;
1395 }
1396 /*
1397 * clear the LRU reference count so the bufer doesn't get
1398 * ignored in xfs_buf_rele().
1399 */
1400 atomic_set(&bp->b_lru_ref, 0);
1401 spin_unlock(&btp->bt_lru_lock);
1402 xfs_buf_rele(bp);
1403 spin_lock(&btp->bt_lru_lock);
1404 }
1405 spin_unlock(&btp->bt_lru_lock);
1406}
1407
1408int
1409xfs_buftarg_shrink(
1410 struct shrinker *shrink,
1411 struct shrink_control *sc)
1412{
1413 struct xfs_buftarg *btp = container_of(shrink,
1414 struct xfs_buftarg, bt_shrinker);
1415 struct xfs_buf *bp;
1416 int nr_to_scan = sc->nr_to_scan;
1417 LIST_HEAD(dispose);
1418
1419 if (!nr_to_scan)
1420 return btp->bt_lru_nr;
1421
1422 spin_lock(&btp->bt_lru_lock);
1423 while (!list_empty(&btp->bt_lru)) {
1424 if (nr_to_scan-- <= 0)
1425 break;
1426
1427 bp = list_first_entry(&btp->bt_lru, struct xfs_buf, b_lru);
1428
1429 /*
1430 * Decrement the b_lru_ref count unless the value is already
1431 * zero. If the value is already zero, we need to reclaim the
1432 * buffer, otherwise it gets another trip through the LRU.
1433 */
1434 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1435 list_move_tail(&bp->b_lru, &btp->bt_lru);
1436 continue;
1437 }
1438
1439 /*
1440 * remove the buffer from the LRU now to avoid needing another
1441 * lock round trip inside xfs_buf_rele().
1442 */
1443 list_move(&bp->b_lru, &dispose);
1444 btp->bt_lru_nr--;
1445 }
1446 spin_unlock(&btp->bt_lru_lock);
1447
1448 while (!list_empty(&dispose)) {
1449 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1450 list_del_init(&bp->b_lru);
1451 xfs_buf_rele(bp);
1452 }
1453
1454 return btp->bt_lru_nr;
1455}
1456
1457void
1458xfs_free_buftarg(
1459 struct xfs_mount *mp,
1460 struct xfs_buftarg *btp)
1461{
1462 unregister_shrinker(&btp->bt_shrinker);
1463
1464 xfs_flush_buftarg(btp, 1);
1465 if (mp->m_flags & XFS_MOUNT_BARRIER)
1466 xfs_blkdev_issue_flush(btp);
1467
1468 kthread_stop(btp->bt_task);
1469 kmem_free(btp);
1470}
1471
1472STATIC int
1473xfs_setsize_buftarg_flags(
1474 xfs_buftarg_t *btp,
1475 unsigned int blocksize,
1476 unsigned int sectorsize,
1477 int verbose)
1478{
1479 btp->bt_bsize = blocksize;
1480 btp->bt_sshift = ffs(sectorsize) - 1;
1481 btp->bt_smask = sectorsize - 1;
1482
1483 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1484 xfs_warn(btp->bt_mount,
1485 "Cannot set_blocksize to %u on device %s\n",
1486 sectorsize, xfs_buf_target_name(btp));
1487 return EINVAL;
1488 }
1489
1490 return 0;
1491}
1492
1493/*
1494 * When allocating the initial buffer target we have not yet
1495 * read in the superblock, so don't know what sized sectors
1496 * are being used is at this early stage. Play safe.
1497 */
1498STATIC int
1499xfs_setsize_buftarg_early(
1500 xfs_buftarg_t *btp,
1501 struct block_device *bdev)
1502{
1503 return xfs_setsize_buftarg_flags(btp,
1504 PAGE_SIZE, bdev_logical_block_size(bdev), 0);
1505}
1506
1507int
1508xfs_setsize_buftarg(
1509 xfs_buftarg_t *btp,
1510 unsigned int blocksize,
1511 unsigned int sectorsize)
1512{
1513 return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
1514}
1515
1516STATIC int
1517xfs_alloc_delwrite_queue(
1518 xfs_buftarg_t *btp,
1519 const char *fsname)
1520{
1521 INIT_LIST_HEAD(&btp->bt_delwrite_queue);
1522 spin_lock_init(&btp->bt_delwrite_lock);
1523 btp->bt_flags = 0;
1524 btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
1525 if (IS_ERR(btp->bt_task))
1526 return PTR_ERR(btp->bt_task);
1527 return 0;
1528}
1529
1530xfs_buftarg_t *
1531xfs_alloc_buftarg(
1532 struct xfs_mount *mp,
1533 struct block_device *bdev,
1534 int external,
1535 const char *fsname)
1536{
1537 xfs_buftarg_t *btp;
1538
1539 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP);
1540
1541 btp->bt_mount = mp;
1542 btp->bt_dev = bdev->bd_dev;
1543 btp->bt_bdev = bdev;
1544 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1545 if (!btp->bt_bdi)
1546 goto error;
1547
1548 INIT_LIST_HEAD(&btp->bt_lru);
1549 spin_lock_init(&btp->bt_lru_lock);
1550 if (xfs_setsize_buftarg_early(btp, bdev))
1551 goto error;
1552 if (xfs_alloc_delwrite_queue(btp, fsname))
1553 goto error;
1554 btp->bt_shrinker.shrink = xfs_buftarg_shrink;
1555 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1556 register_shrinker(&btp->bt_shrinker);
1557 return btp;
1558
1559error:
1560 kmem_free(btp);
1561 return NULL;
1562}
1563
1564
1565/*
1566 * Delayed write buffer handling
1567 */
1568STATIC void
1569xfs_buf_delwri_queue(
1570 xfs_buf_t *bp,
1571 int unlock)
1572{
1573 struct list_head *dwq = &bp->b_target->bt_delwrite_queue;
1574 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1575
1576 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1577
1578 ASSERT((bp->b_flags&(XBF_DELWRI|XBF_ASYNC)) == (XBF_DELWRI|XBF_ASYNC));
1579
1580 spin_lock(dwlk);
1581 /* If already in the queue, dequeue and place at tail */
1582 if (!list_empty(&bp->b_list)) {
1583 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1584 if (unlock)
1585 atomic_dec(&bp->b_hold);
1586 list_del(&bp->b_list);
1587 }
1588
1589 if (list_empty(dwq)) {
1590 /* start xfsbufd as it is about to have something to do */
1591 wake_up_process(bp->b_target->bt_task);
1592 }
1593
1594 bp->b_flags |= _XBF_DELWRI_Q;
1595 list_add_tail(&bp->b_list, dwq);
1596 bp->b_queuetime = jiffies;
1597 spin_unlock(dwlk);
1598
1599 if (unlock)
1600 xfs_buf_unlock(bp);
1601}
1602
1603void
1604xfs_buf_delwri_dequeue(
1605 xfs_buf_t *bp)
1606{
1607 spinlock_t *dwlk = &bp->b_target->bt_delwrite_lock;
1608 int dequeued = 0;
1609
1610 spin_lock(dwlk);
1611 if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
1612 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1613 list_del_init(&bp->b_list);
1614 dequeued = 1;
1615 }
1616 bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
1617 spin_unlock(dwlk);
1618
1619 if (dequeued)
1620 xfs_buf_rele(bp);
1621
1622 trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
1623}
1624
1625/*
1626 * If a delwri buffer needs to be pushed before it has aged out, then promote
1627 * it to the head of the delwri queue so that it will be flushed on the next
1628 * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
1629 * than the age currently needed to flush the buffer. Hence the next time the
1630 * xfsbufd sees it is guaranteed to be considered old enough to flush.
1631 */
1632void
1633xfs_buf_delwri_promote(
1634 struct xfs_buf *bp)
1635{
1636 struct xfs_buftarg *btp = bp->b_target;
1637 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
1638
1639 ASSERT(bp->b_flags & XBF_DELWRI);
1640 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
1641
1642 /*
1643 * Check the buffer age before locking the delayed write queue as we
1644 * don't need to promote buffers that are already past the flush age.
1645 */
1646 if (bp->b_queuetime < jiffies - age)
1647 return;
1648 bp->b_queuetime = jiffies - age;
1649 spin_lock(&btp->bt_delwrite_lock);
1650 list_move(&bp->b_list, &btp->bt_delwrite_queue);
1651 spin_unlock(&btp->bt_delwrite_lock);
1652}
1653
1654STATIC void
1655xfs_buf_runall_queues(
1656 struct workqueue_struct *queue)
1657{
1658 flush_workqueue(queue);
1659}
1660
1661/*
1662 * Move as many buffers as specified to the supplied list
1663 * idicating if we skipped any buffers to prevent deadlocks.
1664 */
1665STATIC int
1666xfs_buf_delwri_split(
1667 xfs_buftarg_t *target,
1668 struct list_head *list,
1669 unsigned long age)
1670{
1671 xfs_buf_t *bp, *n;
1672 struct list_head *dwq = &target->bt_delwrite_queue;
1673 spinlock_t *dwlk = &target->bt_delwrite_lock;
1674 int skipped = 0;
1675 int force;
1676
1677 force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1678 INIT_LIST_HEAD(list);
1679 spin_lock(dwlk);
1680 list_for_each_entry_safe(bp, n, dwq, b_list) {
1681 ASSERT(bp->b_flags & XBF_DELWRI);
1682
1683 if (!xfs_buf_ispinned(bp) && xfs_buf_trylock(bp)) {
1684 if (!force &&
1685 time_before(jiffies, bp->b_queuetime + age)) {
1686 xfs_buf_unlock(bp);
1687 break;
1688 }
1689
1690 bp->b_flags &= ~(XBF_DELWRI | _XBF_DELWRI_Q);
1691 bp->b_flags |= XBF_WRITE;
1692 list_move_tail(&bp->b_list, list);
1693 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1694 } else
1695 skipped++;
1696 }
1697 spin_unlock(dwlk);
1698
1699 return skipped;
1700
1701}
1702
1703/*
1704 * Compare function is more complex than it needs to be because
1705 * the return value is only 32 bits and we are doing comparisons
1706 * on 64 bit values
1707 */
1708static int
1709xfs_buf_cmp(
1710 void *priv,
1711 struct list_head *a,
1712 struct list_head *b)
1713{
1714 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1715 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1716 xfs_daddr_t diff;
1717
1718 diff = ap->b_bn - bp->b_bn;
1719 if (diff < 0)
1720 return -1;
1721 if (diff > 0)
1722 return 1;
1723 return 0;
1724}
1725
1726STATIC int
1727xfsbufd(
1728 void *data)
1729{
1730 xfs_buftarg_t *target = (xfs_buftarg_t *)data;
1731
1732 current->flags |= PF_MEMALLOC;
1733
1734 set_freezable();
1735
1736 do {
1737 long age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
1738 long tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
1739 struct list_head tmp;
1740 struct blk_plug plug;
1741
1742 if (unlikely(freezing(current))) {
1743 set_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1744 refrigerator();
1745 } else {
1746 clear_bit(XBT_FORCE_SLEEP, &target->bt_flags);
1747 }
1748
1749 /* sleep for a long time if there is nothing to do. */
1750 if (list_empty(&target->bt_delwrite_queue))
1751 tout = MAX_SCHEDULE_TIMEOUT;
1752 schedule_timeout_interruptible(tout);
1753
1754 xfs_buf_delwri_split(target, &tmp, age);
1755 list_sort(NULL, &tmp, xfs_buf_cmp);
1756
1757 blk_start_plug(&plug);
1758 while (!list_empty(&tmp)) {
1759 struct xfs_buf *bp;
1760 bp = list_first_entry(&tmp, struct xfs_buf, b_list);
1761 list_del_init(&bp->b_list);
1762 xfs_bdstrat_cb(bp);
1763 }
1764 blk_finish_plug(&plug);
1765 } while (!kthread_should_stop());
1766
1767 return 0;
1768}
1769
1770/*
1771 * Go through all incore buffers, and release buffers if they belong to
1772 * the given device. This is used in filesystem error handling to
1773 * preserve the consistency of its metadata.
1774 */
1775int
1776xfs_flush_buftarg(
1777 xfs_buftarg_t *target,
1778 int wait)
1779{
1780 xfs_buf_t *bp;
1781 int pincount = 0;
1782 LIST_HEAD(tmp_list);
1783 LIST_HEAD(wait_list);
1784 struct blk_plug plug;
1785
1786 xfs_buf_runall_queues(xfsconvertd_workqueue);
1787 xfs_buf_runall_queues(xfsdatad_workqueue);
1788 xfs_buf_runall_queues(xfslogd_workqueue);
1789
1790 set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
1791 pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
1792
1793 /*
1794 * Dropped the delayed write list lock, now walk the temporary list.
1795 * All I/O is issued async and then if we need to wait for completion
1796 * we do that after issuing all the IO.
1797 */
1798 list_sort(NULL, &tmp_list, xfs_buf_cmp);
1799
1800 blk_start_plug(&plug);
1801 while (!list_empty(&tmp_list)) {
1802 bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
1803 ASSERT(target == bp->b_target);
1804 list_del_init(&bp->b_list);
1805 if (wait) {
1806 bp->b_flags &= ~XBF_ASYNC;
1807 list_add(&bp->b_list, &wait_list);
1808 }
1809 xfs_bdstrat_cb(bp);
1810 }
1811 blk_finish_plug(&plug);
1812
1813 if (wait) {
1814 /* Wait for IO to complete. */
1815 while (!list_empty(&wait_list)) {
1816 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
1817
1818 list_del_init(&bp->b_list);
1819 xfs_buf_iowait(bp);
1820 xfs_buf_relse(bp);
1821 }
1822 }
1823
1824 return pincount;
1825}
1826
1827int __init
1828xfs_buf_init(void)
1829{
1830 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1831 KM_ZONE_HWALIGN, NULL);
1832 if (!xfs_buf_zone)
1833 goto out;
1834
1835 xfslogd_workqueue = alloc_workqueue("xfslogd",
1836 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1837 if (!xfslogd_workqueue)
1838 goto out_free_buf_zone;
1839
1840 xfsdatad_workqueue = alloc_workqueue("xfsdatad", WQ_MEM_RECLAIM, 1);
1841 if (!xfsdatad_workqueue)
1842 goto out_destroy_xfslogd_workqueue;
1843
1844 xfsconvertd_workqueue = alloc_workqueue("xfsconvertd",
1845 WQ_MEM_RECLAIM, 1);
1846 if (!xfsconvertd_workqueue)
1847 goto out_destroy_xfsdatad_workqueue;
1848
1849 return 0;
1850
1851 out_destroy_xfsdatad_workqueue:
1852 destroy_workqueue(xfsdatad_workqueue);
1853 out_destroy_xfslogd_workqueue:
1854 destroy_workqueue(xfslogd_workqueue);
1855 out_free_buf_zone:
1856 kmem_zone_destroy(xfs_buf_zone);
1857 out:
1858 return -ENOMEM;
1859}
1860
1861void
1862xfs_buf_terminate(void)
1863{
1864 destroy_workqueue(xfsconvertd_workqueue);
1865 destroy_workqueue(xfsdatad_workqueue);
1866 destroy_workqueue(xfslogd_workqueue);
1867 kmem_zone_destroy(xfs_buf_zone);
1868}
1869
1870#ifdef CONFIG_KDB_MODULES
1871struct list_head *
1872xfs_get_buftarg_list(void)
1873{
1874 return &xfs_buftarg_list;
1875}
1876#endif
1/*
2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
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.
8 *
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.
13 *
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
17 */
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
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"
44
45static kmem_zone_t *xfs_buf_zone;
46
47static struct workqueue_struct *xfslogd_workqueue;
48
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
58
59#define xb_to_gfp(flags) \
60 ((((flags) & XBF_READ_AHEAD) ? __GFP_NORETRY : GFP_NOFS) | __GFP_NOWARN)
61
62
63static inline int
64xfs_buf_is_vmapped(
65 struct xfs_buf *bp)
66{
67 /*
68 * Return true if the buffer is vmapped.
69 *
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.
73 */
74 return bp->b_addr && bp->b_page_count > 1;
75}
76
77static inline int
78xfs_buf_vmap_len(
79 struct xfs_buf *bp)
80{
81 return (bp->b_page_count * PAGE_SIZE) - bp->b_offset;
82}
83
84/*
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.
89 *
90 * This prevents build-up of stale buffers on the LRU.
91 */
92void
93xfs_buf_stale(
94 struct xfs_buf *bp)
95{
96 ASSERT(xfs_buf_islocked(bp));
97
98 bp->b_flags |= XBF_STALE;
99
100 /*
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.
104 */
105 bp->b_flags &= ~_XBF_DELWRI_Q;
106
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);
112
113 ASSERT(atomic_read(&bp->b_hold) >= 1);
114 spin_unlock(&bp->b_lock);
115}
116
117static int
118xfs_buf_get_maps(
119 struct xfs_buf *bp,
120 int map_count)
121{
122 ASSERT(bp->b_maps == NULL);
123 bp->b_map_count = map_count;
124
125 if (map_count == 1) {
126 bp->b_maps = &bp->__b_map;
127 return 0;
128 }
129
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;
135}
136
137/*
138 * Frees b_pages if it was allocated.
139 */
140static void
141xfs_buf_free_maps(
142 struct xfs_buf *bp)
143{
144 if (bp->b_maps != &bp->__b_map) {
145 kmem_free(bp->b_maps);
146 bp->b_maps = NULL;
147 }
148}
149
150struct 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)
156{
157 struct xfs_buf *bp;
158 int error;
159 int i;
160
161 bp = kmem_zone_zalloc(xfs_buf_zone, KM_NOFS);
162 if (unlikely(!bp))
163 return NULL;
164
165 /*
166 * We don't want certain flags to appear in b_flags unless they are
167 * specifically set by later operations on the buffer.
168 */
169 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
170
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;
182
183 /*
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).
187 */
188 error = xfs_buf_get_maps(bp, nmaps);
189 if (error) {
190 kmem_zone_free(xfs_buf_zone, bp);
191 return NULL;
192 }
193
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;
200 }
201 bp->b_io_length = bp->b_length;
202
203 atomic_set(&bp->b_pin_count, 0);
204 init_waitqueue_head(&bp->b_waiters);
205
206 XFS_STATS_INC(xb_create);
207 trace_xfs_buf_init(bp, _RET_IP_);
208
209 return bp;
210}
211
212/*
213 * Allocate a page array capable of holding a specified number
214 * of pages, and point the page buf at it.
215 */
216STATIC int
217_xfs_buf_get_pages(
218 xfs_buf_t *bp,
219 int page_count,
220 xfs_buf_flags_t flags)
221{
222 /* Make sure that we have a page list */
223 if (bp->b_pages == NULL) {
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, KM_NOFS);
230 if (bp->b_pages == NULL)
231 return -ENOMEM;
232 }
233 memset(bp->b_pages, 0, sizeof(struct page *) * page_count);
234 }
235 return 0;
236}
237
238/*
239 * Frees b_pages if it was allocated.
240 */
241STATIC void
242_xfs_buf_free_pages(
243 xfs_buf_t *bp)
244{
245 if (bp->b_pages != bp->b_page_array) {
246 kmem_free(bp->b_pages);
247 bp->b_pages = NULL;
248 }
249}
250
251/*
252 * Releases the specified buffer.
253 *
254 * The modification state of any associated pages is left unchanged.
255 * The buffer must not be on any hash - use xfs_buf_rele instead for
256 * hashed and refcounted buffers
257 */
258void
259xfs_buf_free(
260 xfs_buf_t *bp)
261{
262 trace_xfs_buf_free(bp, _RET_IP_);
263
264 ASSERT(list_empty(&bp->b_lru));
265
266 if (bp->b_flags & _XBF_PAGES) {
267 uint i;
268
269 if (xfs_buf_is_vmapped(bp))
270 vm_unmap_ram(bp->b_addr - bp->b_offset,
271 bp->b_page_count);
272
273 for (i = 0; i < bp->b_page_count; i++) {
274 struct page *page = bp->b_pages[i];
275
276 __free_page(page);
277 }
278 } else if (bp->b_flags & _XBF_KMEM)
279 kmem_free(bp->b_addr);
280 _xfs_buf_free_pages(bp);
281 xfs_buf_free_maps(bp);
282 kmem_zone_free(xfs_buf_zone, bp);
283}
284
285/*
286 * Allocates all the pages for buffer in question and builds it's page list.
287 */
288STATIC int
289xfs_buf_allocate_memory(
290 xfs_buf_t *bp,
291 uint flags)
292{
293 size_t size;
294 size_t nbytes, offset;
295 gfp_t gfp_mask = xb_to_gfp(flags);
296 unsigned short page_count, i;
297 xfs_off_t start, end;
298 int error;
299
300 /*
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.
304 */
305 size = BBTOB(bp->b_length);
306 if (size < PAGE_SIZE) {
307 bp->b_addr = kmem_alloc(size, KM_NOFS);
308 if (!bp->b_addr) {
309 /* low memory - use alloc_page loop instead */
310 goto use_alloc_page;
311 }
312
313 if (((unsigned long)(bp->b_addr + size - 1) & 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;
319 }
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_KMEM;
325 return 0;
326 }
327
328use_alloc_page:
329 start = BBTOB(bp->b_maps[0].bm_bn) >> PAGE_SHIFT;
330 end = (BBTOB(bp->b_maps[0].bm_bn + bp->b_length) + PAGE_SIZE - 1)
331 >> PAGE_SHIFT;
332 page_count = end - start;
333 error = _xfs_buf_get_pages(bp, page_count, flags);
334 if (unlikely(error))
335 return error;
336
337 offset = bp->b_offset;
338 bp->b_flags |= _XBF_PAGES;
339
340 for (i = 0; i < bp->b_page_count; i++) {
341 struct page *page;
342 uint retries = 0;
343retry:
344 page = alloc_page(gfp_mask);
345 if (unlikely(page == NULL)) {
346 if (flags & XBF_READ_AHEAD) {
347 bp->b_page_count = i;
348 error = ENOMEM;
349 goto out_free_pages;
350 }
351
352 /*
353 * This could deadlock.
354 *
355 * But until all the XFS lowlevel code is revamped to
356 * handle buffer allocation failures we can't do much.
357 */
358 if (!(++retries % 100))
359 xfs_err(NULL,
360 "possible memory allocation deadlock in %s (mode:0x%x)",
361 __func__, gfp_mask);
362
363 XFS_STATS_INC(xb_page_retries);
364 congestion_wait(BLK_RW_ASYNC, HZ/50);
365 goto retry;
366 }
367
368 XFS_STATS_INC(xb_page_found);
369
370 nbytes = min_t(size_t, size, PAGE_SIZE - offset);
371 size -= nbytes;
372 bp->b_pages[i] = page;
373 offset = 0;
374 }
375 return 0;
376
377out_free_pages:
378 for (i = 0; i < bp->b_page_count; i++)
379 __free_page(bp->b_pages[i]);
380 return error;
381}
382
383/*
384 * Map buffer into kernel address-space if necessary.
385 */
386STATIC int
387_xfs_buf_map_pages(
388 xfs_buf_t *bp,
389 uint flags)
390{
391 ASSERT(bp->b_flags & _XBF_PAGES);
392 if (bp->b_page_count == 1) {
393 /* A single page buffer is always mappable */
394 bp->b_addr = page_address(bp->b_pages[0]) + bp->b_offset;
395 } else if (flags & XBF_UNMAPPED) {
396 bp->b_addr = NULL;
397 } else {
398 int retried = 0;
399 unsigned noio_flag;
400
401 /*
402 * vm_map_ram() will allocate auxillary structures (e.g.
403 * pagetables) with GFP_KERNEL, yet we are likely to be under
404 * GFP_NOFS context here. Hence we need to tell memory reclaim
405 * that we are in such a context via PF_MEMALLOC_NOIO to prevent
406 * memory reclaim re-entering the filesystem here and
407 * potentially deadlocking.
408 */
409 noio_flag = memalloc_noio_save();
410 do {
411 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
412 -1, PAGE_KERNEL);
413 if (bp->b_addr)
414 break;
415 vm_unmap_aliases();
416 } while (retried++ <= 1);
417 memalloc_noio_restore(noio_flag);
418
419 if (!bp->b_addr)
420 return -ENOMEM;
421 bp->b_addr += bp->b_offset;
422 }
423
424 return 0;
425}
426
427/*
428 * Finding and Reading Buffers
429 */
430
431/*
432 * Look up, and creates if absent, a lockable buffer for
433 * a given range of an inode. The buffer is returned
434 * locked. No I/O is implied by this call.
435 */
436xfs_buf_t *
437_xfs_buf_find(
438 struct xfs_buftarg *btp,
439 struct xfs_buf_map *map,
440 int nmaps,
441 xfs_buf_flags_t flags,
442 xfs_buf_t *new_bp)
443{
444 size_t numbytes;
445 struct xfs_perag *pag;
446 struct rb_node **rbp;
447 struct rb_node *parent;
448 xfs_buf_t *bp;
449 xfs_daddr_t blkno = map[0].bm_bn;
450 xfs_daddr_t eofs;
451 int numblks = 0;
452 int i;
453
454 for (i = 0; i < nmaps; i++)
455 numblks += map[i].bm_len;
456 numbytes = BBTOB(numblks);
457
458 /* Check for IOs smaller than the sector size / not sector aligned */
459 ASSERT(!(numbytes < btp->bt_meta_sectorsize));
460 ASSERT(!(BBTOB(blkno) & (xfs_off_t)btp->bt_meta_sectormask));
461
462 /*
463 * Corrupted block numbers can get through to here, unfortunately, so we
464 * have to check that the buffer falls within the filesystem bounds.
465 */
466 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
467 if (blkno >= eofs) {
468 /*
469 * XXX (dgc): we should really be returning EFSCORRUPTED here,
470 * but none of the higher level infrastructure supports
471 * returning a specific error on buffer lookup failures.
472 */
473 xfs_alert(btp->bt_mount,
474 "%s: Block out of range: block 0x%llx, EOFS 0x%llx ",
475 __func__, blkno, eofs);
476 WARN_ON(1);
477 return NULL;
478 }
479
480 /* get tree root */
481 pag = xfs_perag_get(btp->bt_mount,
482 xfs_daddr_to_agno(btp->bt_mount, blkno));
483
484 /* walk tree */
485 spin_lock(&pag->pag_buf_lock);
486 rbp = &pag->pag_buf_tree.rb_node;
487 parent = NULL;
488 bp = NULL;
489 while (*rbp) {
490 parent = *rbp;
491 bp = rb_entry(parent, struct xfs_buf, b_rbnode);
492
493 if (blkno < bp->b_bn)
494 rbp = &(*rbp)->rb_left;
495 else if (blkno > bp->b_bn)
496 rbp = &(*rbp)->rb_right;
497 else {
498 /*
499 * found a block number match. If the range doesn't
500 * match, the only way this is allowed is if the buffer
501 * in the cache is stale and the transaction that made
502 * it stale has not yet committed. i.e. we are
503 * reallocating a busy extent. Skip this buffer and
504 * continue searching to the right for an exact match.
505 */
506 if (bp->b_length != numblks) {
507 ASSERT(bp->b_flags & XBF_STALE);
508 rbp = &(*rbp)->rb_right;
509 continue;
510 }
511 atomic_inc(&bp->b_hold);
512 goto found;
513 }
514 }
515
516 /* No match found */
517 if (new_bp) {
518 rb_link_node(&new_bp->b_rbnode, parent, rbp);
519 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
520 /* the buffer keeps the perag reference until it is freed */
521 new_bp->b_pag = pag;
522 spin_unlock(&pag->pag_buf_lock);
523 } else {
524 XFS_STATS_INC(xb_miss_locked);
525 spin_unlock(&pag->pag_buf_lock);
526 xfs_perag_put(pag);
527 }
528 return new_bp;
529
530found:
531 spin_unlock(&pag->pag_buf_lock);
532 xfs_perag_put(pag);
533
534 if (!xfs_buf_trylock(bp)) {
535 if (flags & XBF_TRYLOCK) {
536 xfs_buf_rele(bp);
537 XFS_STATS_INC(xb_busy_locked);
538 return NULL;
539 }
540 xfs_buf_lock(bp);
541 XFS_STATS_INC(xb_get_locked_waited);
542 }
543
544 /*
545 * if the buffer is stale, clear all the external state associated with
546 * it. We need to keep flags such as how we allocated the buffer memory
547 * intact here.
548 */
549 if (bp->b_flags & XBF_STALE) {
550 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
551 ASSERT(bp->b_iodone == NULL);
552 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
553 bp->b_ops = NULL;
554 }
555
556 trace_xfs_buf_find(bp, flags, _RET_IP_);
557 XFS_STATS_INC(xb_get_locked);
558 return bp;
559}
560
561/*
562 * Assembles a buffer covering the specified range. The code is optimised for
563 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
564 * more hits than misses.
565 */
566struct xfs_buf *
567xfs_buf_get_map(
568 struct xfs_buftarg *target,
569 struct xfs_buf_map *map,
570 int nmaps,
571 xfs_buf_flags_t flags)
572{
573 struct xfs_buf *bp;
574 struct xfs_buf *new_bp;
575 int error = 0;
576
577 bp = _xfs_buf_find(target, map, nmaps, flags, NULL);
578 if (likely(bp))
579 goto found;
580
581 new_bp = _xfs_buf_alloc(target, map, nmaps, flags);
582 if (unlikely(!new_bp))
583 return NULL;
584
585 error = xfs_buf_allocate_memory(new_bp, flags);
586 if (error) {
587 xfs_buf_free(new_bp);
588 return NULL;
589 }
590
591 bp = _xfs_buf_find(target, map, nmaps, flags, new_bp);
592 if (!bp) {
593 xfs_buf_free(new_bp);
594 return NULL;
595 }
596
597 if (bp != new_bp)
598 xfs_buf_free(new_bp);
599
600found:
601 if (!bp->b_addr) {
602 error = _xfs_buf_map_pages(bp, flags);
603 if (unlikely(error)) {
604 xfs_warn(target->bt_mount,
605 "%s: failed to map pagesn", __func__);
606 xfs_buf_relse(bp);
607 return NULL;
608 }
609 }
610
611 XFS_STATS_INC(xb_get);
612 trace_xfs_buf_get(bp, flags, _RET_IP_);
613 return bp;
614}
615
616STATIC int
617_xfs_buf_read(
618 xfs_buf_t *bp,
619 xfs_buf_flags_t flags)
620{
621 ASSERT(!(flags & XBF_WRITE));
622 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
623
624 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
625 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
626
627 xfs_buf_iorequest(bp);
628 if (flags & XBF_ASYNC)
629 return 0;
630 return xfs_buf_iowait(bp);
631}
632
633xfs_buf_t *
634xfs_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)
640{
641 struct xfs_buf *bp;
642
643 flags |= XBF_READ;
644
645 bp = xfs_buf_get_map(target, map, nmaps, flags);
646 if (bp) {
647 trace_xfs_buf_read(bp, flags, _RET_IP_);
648
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) {
654 /*
655 * Read ahead call which is already satisfied,
656 * drop the buffer
657 */
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;
663 }
664 }
665
666 return bp;
667}
668
669/*
670 * If we are not low on memory then do the readahead in a deadlock
671 * safe manner.
672 */
673void
674xfs_buf_readahead_map(
675 struct xfs_buftarg *target,
676 struct xfs_buf_map *map,
677 int nmaps,
678 const struct xfs_buf_ops *ops)
679{
680 if (bdi_read_congested(target->bt_bdi))
681 return;
682
683 xfs_buf_read_map(target, map, nmaps,
684 XBF_TRYLOCK|XBF_ASYNC|XBF_READ_AHEAD, ops);
685}
686
687/*
688 * Read an uncached buffer from disk. Allocates and returns a locked
689 * buffer containing the disk contents or nothing.
690 */
691struct xfs_buf *
692xfs_buf_read_uncached(
693 struct xfs_buftarg *target,
694 xfs_daddr_t daddr,
695 size_t numblks,
696 int flags,
697 const struct xfs_buf_ops *ops)
698{
699 struct xfs_buf *bp;
700
701 bp = xfs_buf_get_uncached(target, numblks, flags);
702 if (!bp)
703 return NULL;
704
705 /* set up the buffer for a read IO */
706 ASSERT(bp->b_map_count == 1);
707 bp->b_bn = daddr;
708 bp->b_maps[0].bm_bn = daddr;
709 bp->b_flags |= XBF_READ;
710 bp->b_ops = ops;
711
712 if (XFS_FORCED_SHUTDOWN(target->bt_mount)) {
713 xfs_buf_relse(bp);
714 return NULL;
715 }
716 xfs_buf_iorequest(bp);
717 xfs_buf_iowait(bp);
718 return bp;
719}
720
721/*
722 * Return a buffer allocated as an empty buffer and associated to external
723 * memory via xfs_buf_associate_memory() back to it's empty state.
724 */
725void
726xfs_buf_set_empty(
727 struct xfs_buf *bp,
728 size_t numblks)
729{
730 if (bp->b_pages)
731 _xfs_buf_free_pages(bp);
732
733 bp->b_pages = NULL;
734 bp->b_page_count = 0;
735 bp->b_addr = NULL;
736 bp->b_length = numblks;
737 bp->b_io_length = numblks;
738
739 ASSERT(bp->b_map_count == 1);
740 bp->b_bn = XFS_BUF_DADDR_NULL;
741 bp->b_maps[0].bm_bn = XFS_BUF_DADDR_NULL;
742 bp->b_maps[0].bm_len = bp->b_length;
743}
744
745static inline struct page *
746mem_to_page(
747 void *addr)
748{
749 if ((!is_vmalloc_addr(addr))) {
750 return virt_to_page(addr);
751 } else {
752 return vmalloc_to_page(addr);
753 }
754}
755
756int
757xfs_buf_associate_memory(
758 xfs_buf_t *bp,
759 void *mem,
760 size_t len)
761{
762 int rval;
763 int i = 0;
764 unsigned long pageaddr;
765 unsigned long offset;
766 size_t buflen;
767 int page_count;
768
769 pageaddr = (unsigned long)mem & PAGE_MASK;
770 offset = (unsigned long)mem - pageaddr;
771 buflen = PAGE_ALIGN(len + offset);
772 page_count = buflen >> PAGE_SHIFT;
773
774 /* Free any previous set of page pointers */
775 if (bp->b_pages)
776 _xfs_buf_free_pages(bp);
777
778 bp->b_pages = NULL;
779 bp->b_addr = mem;
780
781 rval = _xfs_buf_get_pages(bp, page_count, 0);
782 if (rval)
783 return rval;
784
785 bp->b_offset = offset;
786
787 for (i = 0; i < bp->b_page_count; i++) {
788 bp->b_pages[i] = mem_to_page((void *)pageaddr);
789 pageaddr += PAGE_SIZE;
790 }
791
792 bp->b_io_length = BTOBB(len);
793 bp->b_length = BTOBB(buflen);
794
795 return 0;
796}
797
798xfs_buf_t *
799xfs_buf_get_uncached(
800 struct xfs_buftarg *target,
801 size_t numblks,
802 int flags)
803{
804 unsigned long page_count;
805 int error, i;
806 struct xfs_buf *bp;
807 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
808
809 bp = _xfs_buf_alloc(target, &map, 1, 0);
810 if (unlikely(bp == NULL))
811 goto fail;
812
813 page_count = PAGE_ALIGN(numblks << BBSHIFT) >> PAGE_SHIFT;
814 error = _xfs_buf_get_pages(bp, page_count, 0);
815 if (error)
816 goto fail_free_buf;
817
818 for (i = 0; i < page_count; i++) {
819 bp->b_pages[i] = alloc_page(xb_to_gfp(flags));
820 if (!bp->b_pages[i])
821 goto fail_free_mem;
822 }
823 bp->b_flags |= _XBF_PAGES;
824
825 error = _xfs_buf_map_pages(bp, 0);
826 if (unlikely(error)) {
827 xfs_warn(target->bt_mount,
828 "%s: failed to map pages", __func__);
829 goto fail_free_mem;
830 }
831
832 trace_xfs_buf_get_uncached(bp, _RET_IP_);
833 return bp;
834
835 fail_free_mem:
836 while (--i >= 0)
837 __free_page(bp->b_pages[i]);
838 _xfs_buf_free_pages(bp);
839 fail_free_buf:
840 xfs_buf_free_maps(bp);
841 kmem_zone_free(xfs_buf_zone, bp);
842 fail:
843 return NULL;
844}
845
846/*
847 * Increment reference count on buffer, to hold the buffer concurrently
848 * with another thread which may release (free) the buffer asynchronously.
849 * Must hold the buffer already to call this function.
850 */
851void
852xfs_buf_hold(
853 xfs_buf_t *bp)
854{
855 trace_xfs_buf_hold(bp, _RET_IP_);
856 atomic_inc(&bp->b_hold);
857}
858
859/*
860 * Releases a hold on the specified buffer. If the
861 * the hold count is 1, calls xfs_buf_free.
862 */
863void
864xfs_buf_rele(
865 xfs_buf_t *bp)
866{
867 struct xfs_perag *pag = bp->b_pag;
868
869 trace_xfs_buf_rele(bp, _RET_IP_);
870
871 if (!pag) {
872 ASSERT(list_empty(&bp->b_lru));
873 ASSERT(RB_EMPTY_NODE(&bp->b_rbnode));
874 if (atomic_dec_and_test(&bp->b_hold))
875 xfs_buf_free(bp);
876 return;
877 }
878
879 ASSERT(!RB_EMPTY_NODE(&bp->b_rbnode));
880
881 ASSERT(atomic_read(&bp->b_hold) > 0);
882 if (atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock)) {
883 spin_lock(&bp->b_lock);
884 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
885 /*
886 * If the buffer is added to the LRU take a new
887 * reference to the buffer for the LRU and clear the
888 * (now stale) dispose list state flag
889 */
890 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
891 bp->b_state &= ~XFS_BSTATE_DISPOSE;
892 atomic_inc(&bp->b_hold);
893 }
894 spin_unlock(&bp->b_lock);
895 spin_unlock(&pag->pag_buf_lock);
896 } else {
897 /*
898 * most of the time buffers will already be removed from
899 * the LRU, so optimise that case by checking for the
900 * XFS_BSTATE_DISPOSE flag indicating the last list the
901 * buffer was on was the disposal list
902 */
903 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
904 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
905 } else {
906 ASSERT(list_empty(&bp->b_lru));
907 }
908 spin_unlock(&bp->b_lock);
909
910 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
911 rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
912 spin_unlock(&pag->pag_buf_lock);
913 xfs_perag_put(pag);
914 xfs_buf_free(bp);
915 }
916 }
917}
918
919
920/*
921 * Lock a buffer object, if it is not already locked.
922 *
923 * If we come across a stale, pinned, locked buffer, we know that we are
924 * being asked to lock a buffer that has been reallocated. Because it is
925 * pinned, we know that the log has not been pushed to disk and hence it
926 * will still be locked. Rather than continuing to have trylock attempts
927 * fail until someone else pushes the log, push it ourselves before
928 * returning. This means that the xfsaild will not get stuck trying
929 * to push on stale inode buffers.
930 */
931int
932xfs_buf_trylock(
933 struct xfs_buf *bp)
934{
935 int locked;
936
937 locked = down_trylock(&bp->b_sema) == 0;
938 if (locked)
939 XB_SET_OWNER(bp);
940
941 trace_xfs_buf_trylock(bp, _RET_IP_);
942 return locked;
943}
944
945/*
946 * Lock a buffer object.
947 *
948 * If we come across a stale, pinned, locked buffer, we know that we
949 * are being asked to lock a buffer that has been reallocated. Because
950 * it is pinned, we know that the log has not been pushed to disk and
951 * hence it will still be locked. Rather than sleeping until someone
952 * else pushes the log, push it ourselves before trying to get the lock.
953 */
954void
955xfs_buf_lock(
956 struct xfs_buf *bp)
957{
958 trace_xfs_buf_lock(bp, _RET_IP_);
959
960 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
961 xfs_log_force(bp->b_target->bt_mount, 0);
962 down(&bp->b_sema);
963 XB_SET_OWNER(bp);
964
965 trace_xfs_buf_lock_done(bp, _RET_IP_);
966}
967
968void
969xfs_buf_unlock(
970 struct xfs_buf *bp)
971{
972 XB_CLEAR_OWNER(bp);
973 up(&bp->b_sema);
974
975 trace_xfs_buf_unlock(bp, _RET_IP_);
976}
977
978STATIC void
979xfs_buf_wait_unpin(
980 xfs_buf_t *bp)
981{
982 DECLARE_WAITQUEUE (wait, current);
983
984 if (atomic_read(&bp->b_pin_count) == 0)
985 return;
986
987 add_wait_queue(&bp->b_waiters, &wait);
988 for (;;) {
989 set_current_state(TASK_UNINTERRUPTIBLE);
990 if (atomic_read(&bp->b_pin_count) == 0)
991 break;
992 io_schedule();
993 }
994 remove_wait_queue(&bp->b_waiters, &wait);
995 set_current_state(TASK_RUNNING);
996}
997
998/*
999 * Buffer Utility Routines
1000 */
1001
1002STATIC void
1003xfs_buf_iodone_work(
1004 struct work_struct *work)
1005{
1006 struct xfs_buf *bp =
1007 container_of(work, xfs_buf_t, b_iodone_work);
1008 bool read = !!(bp->b_flags & XBF_READ);
1009
1010 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1011
1012 /* only validate buffers that were read without errors */
1013 if (read && bp->b_ops && !bp->b_error && (bp->b_flags & XBF_DONE))
1014 bp->b_ops->verify_read(bp);
1015
1016 if (bp->b_iodone)
1017 (*(bp->b_iodone))(bp);
1018 else if (bp->b_flags & XBF_ASYNC)
1019 xfs_buf_relse(bp);
1020 else {
1021 ASSERT(read && bp->b_ops);
1022 complete(&bp->b_iowait);
1023 }
1024}
1025
1026void
1027xfs_buf_ioend(
1028 struct xfs_buf *bp,
1029 int schedule)
1030{
1031 bool read = !!(bp->b_flags & XBF_READ);
1032
1033 trace_xfs_buf_iodone(bp, _RET_IP_);
1034
1035 if (bp->b_error == 0)
1036 bp->b_flags |= XBF_DONE;
1037
1038 if (bp->b_iodone || (read && bp->b_ops) || (bp->b_flags & XBF_ASYNC)) {
1039 if (schedule) {
1040 INIT_WORK(&bp->b_iodone_work, xfs_buf_iodone_work);
1041 queue_work(xfslogd_workqueue, &bp->b_iodone_work);
1042 } else {
1043 xfs_buf_iodone_work(&bp->b_iodone_work);
1044 }
1045 } else {
1046 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
1047 complete(&bp->b_iowait);
1048 }
1049}
1050
1051void
1052xfs_buf_ioerror(
1053 xfs_buf_t *bp,
1054 int error)
1055{
1056 ASSERT(error >= 0 && error <= 0xffff);
1057 bp->b_error = (unsigned short)error;
1058 trace_xfs_buf_ioerror(bp, error, _RET_IP_);
1059}
1060
1061void
1062xfs_buf_ioerror_alert(
1063 struct xfs_buf *bp,
1064 const char *func)
1065{
1066 xfs_alert(bp->b_target->bt_mount,
1067"metadata I/O error: block 0x%llx (\"%s\") error %d numblks %d",
1068 (__uint64_t)XFS_BUF_ADDR(bp), func, bp->b_error, bp->b_length);
1069}
1070
1071/*
1072 * Called when we want to stop a buffer from getting written or read.
1073 * We attach the EIO error, muck with its flags, and call xfs_buf_ioend
1074 * so that the proper iodone callbacks get called.
1075 */
1076STATIC int
1077xfs_bioerror(
1078 xfs_buf_t *bp)
1079{
1080#ifdef XFSERRORDEBUG
1081 ASSERT(XFS_BUF_ISREAD(bp) || bp->b_iodone);
1082#endif
1083
1084 /*
1085 * No need to wait until the buffer is unpinned, we aren't flushing it.
1086 */
1087 xfs_buf_ioerror(bp, EIO);
1088
1089 /*
1090 * We're calling xfs_buf_ioend, so delete XBF_DONE flag.
1091 */
1092 XFS_BUF_UNREAD(bp);
1093 XFS_BUF_UNDONE(bp);
1094 xfs_buf_stale(bp);
1095
1096 xfs_buf_ioend(bp, 0);
1097
1098 return EIO;
1099}
1100
1101/*
1102 * Same as xfs_bioerror, except that we are releasing the buffer
1103 * here ourselves, and avoiding the xfs_buf_ioend call.
1104 * This is meant for userdata errors; metadata bufs come with
1105 * iodone functions attached, so that we can track down errors.
1106 */
1107int
1108xfs_bioerror_relse(
1109 struct xfs_buf *bp)
1110{
1111 int64_t fl = bp->b_flags;
1112 /*
1113 * No need to wait until the buffer is unpinned.
1114 * We aren't flushing it.
1115 *
1116 * chunkhold expects B_DONE to be set, whether
1117 * we actually finish the I/O or not. We don't want to
1118 * change that interface.
1119 */
1120 XFS_BUF_UNREAD(bp);
1121 XFS_BUF_DONE(bp);
1122 xfs_buf_stale(bp);
1123 bp->b_iodone = NULL;
1124 if (!(fl & XBF_ASYNC)) {
1125 /*
1126 * Mark b_error and B_ERROR _both_.
1127 * Lot's of chunkcache code assumes that.
1128 * There's no reason to mark error for
1129 * ASYNC buffers.
1130 */
1131 xfs_buf_ioerror(bp, EIO);
1132 complete(&bp->b_iowait);
1133 } else {
1134 xfs_buf_relse(bp);
1135 }
1136
1137 return EIO;
1138}
1139
1140STATIC int
1141xfs_bdstrat_cb(
1142 struct xfs_buf *bp)
1143{
1144 if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
1145 trace_xfs_bdstrat_shut(bp, _RET_IP_);
1146 /*
1147 * Metadata write that didn't get logged but
1148 * written delayed anyway. These aren't associated
1149 * with a transaction, and can be ignored.
1150 */
1151 if (!bp->b_iodone && !XFS_BUF_ISREAD(bp))
1152 return xfs_bioerror_relse(bp);
1153 else
1154 return xfs_bioerror(bp);
1155 }
1156
1157 xfs_buf_iorequest(bp);
1158 return 0;
1159}
1160
1161int
1162xfs_bwrite(
1163 struct xfs_buf *bp)
1164{
1165 int error;
1166
1167 ASSERT(xfs_buf_islocked(bp));
1168
1169 bp->b_flags |= XBF_WRITE;
1170 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q | XBF_WRITE_FAIL);
1171
1172 xfs_bdstrat_cb(bp);
1173
1174 error = xfs_buf_iowait(bp);
1175 if (error) {
1176 xfs_force_shutdown(bp->b_target->bt_mount,
1177 SHUTDOWN_META_IO_ERROR);
1178 }
1179 return error;
1180}
1181
1182STATIC void
1183_xfs_buf_ioend(
1184 xfs_buf_t *bp,
1185 int schedule)
1186{
1187 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1188 xfs_buf_ioend(bp, schedule);
1189}
1190
1191STATIC void
1192xfs_buf_bio_end_io(
1193 struct bio *bio,
1194 int error)
1195{
1196 xfs_buf_t *bp = (xfs_buf_t *)bio->bi_private;
1197
1198 /*
1199 * don't overwrite existing errors - otherwise we can lose errors on
1200 * buffers that require multiple bios to complete.
1201 */
1202 if (!bp->b_error)
1203 xfs_buf_ioerror(bp, -error);
1204
1205 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1206 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1207
1208 _xfs_buf_ioend(bp, 1);
1209 bio_put(bio);
1210}
1211
1212static void
1213xfs_buf_ioapply_map(
1214 struct xfs_buf *bp,
1215 int map,
1216 int *buf_offset,
1217 int *count,
1218 int rw)
1219{
1220 int page_index;
1221 int total_nr_pages = bp->b_page_count;
1222 int nr_pages;
1223 struct bio *bio;
1224 sector_t sector = bp->b_maps[map].bm_bn;
1225 int size;
1226 int offset;
1227
1228 total_nr_pages = bp->b_page_count;
1229
1230 /* skip the pages in the buffer before the start offset */
1231 page_index = 0;
1232 offset = *buf_offset;
1233 while (offset >= PAGE_SIZE) {
1234 page_index++;
1235 offset -= PAGE_SIZE;
1236 }
1237
1238 /*
1239 * Limit the IO size to the length of the current vector, and update the
1240 * remaining IO count for the next time around.
1241 */
1242 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1243 *count -= size;
1244 *buf_offset += size;
1245
1246next_chunk:
1247 atomic_inc(&bp->b_io_remaining);
1248 nr_pages = BIO_MAX_SECTORS >> (PAGE_SHIFT - BBSHIFT);
1249 if (nr_pages > total_nr_pages)
1250 nr_pages = total_nr_pages;
1251
1252 bio = bio_alloc(GFP_NOIO, nr_pages);
1253 bio->bi_bdev = bp->b_target->bt_bdev;
1254 bio->bi_iter.bi_sector = sector;
1255 bio->bi_end_io = xfs_buf_bio_end_io;
1256 bio->bi_private = bp;
1257
1258
1259 for (; size && nr_pages; nr_pages--, page_index++) {
1260 int rbytes, nbytes = PAGE_SIZE - offset;
1261
1262 if (nbytes > size)
1263 nbytes = size;
1264
1265 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1266 offset);
1267 if (rbytes < nbytes)
1268 break;
1269
1270 offset = 0;
1271 sector += BTOBB(nbytes);
1272 size -= nbytes;
1273 total_nr_pages--;
1274 }
1275
1276 if (likely(bio->bi_iter.bi_size)) {
1277 if (xfs_buf_is_vmapped(bp)) {
1278 flush_kernel_vmap_range(bp->b_addr,
1279 xfs_buf_vmap_len(bp));
1280 }
1281 submit_bio(rw, bio);
1282 if (size)
1283 goto next_chunk;
1284 } else {
1285 /*
1286 * This is guaranteed not to be the last io reference count
1287 * because the caller (xfs_buf_iorequest) holds a count itself.
1288 */
1289 atomic_dec(&bp->b_io_remaining);
1290 xfs_buf_ioerror(bp, EIO);
1291 bio_put(bio);
1292 }
1293
1294}
1295
1296STATIC void
1297_xfs_buf_ioapply(
1298 struct xfs_buf *bp)
1299{
1300 struct blk_plug plug;
1301 int rw;
1302 int offset;
1303 int size;
1304 int i;
1305
1306 /*
1307 * Make sure we capture only current IO errors rather than stale errors
1308 * left over from previous use of the buffer (e.g. failed readahead).
1309 */
1310 bp->b_error = 0;
1311
1312 if (bp->b_flags & XBF_WRITE) {
1313 if (bp->b_flags & XBF_SYNCIO)
1314 rw = WRITE_SYNC;
1315 else
1316 rw = WRITE;
1317 if (bp->b_flags & XBF_FUA)
1318 rw |= REQ_FUA;
1319 if (bp->b_flags & XBF_FLUSH)
1320 rw |= REQ_FLUSH;
1321
1322 /*
1323 * Run the write verifier callback function if it exists. If
1324 * this function fails it will mark the buffer with an error and
1325 * the IO should not be dispatched.
1326 */
1327 if (bp->b_ops) {
1328 bp->b_ops->verify_write(bp);
1329 if (bp->b_error) {
1330 xfs_force_shutdown(bp->b_target->bt_mount,
1331 SHUTDOWN_CORRUPT_INCORE);
1332 return;
1333 }
1334 }
1335 } else if (bp->b_flags & XBF_READ_AHEAD) {
1336 rw = READA;
1337 } else {
1338 rw = READ;
1339 }
1340
1341 /* we only use the buffer cache for meta-data */
1342 rw |= REQ_META;
1343
1344 /*
1345 * Walk all the vectors issuing IO on them. Set up the initial offset
1346 * into the buffer and the desired IO size before we start -
1347 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1348 * subsequent call.
1349 */
1350 offset = bp->b_offset;
1351 size = BBTOB(bp->b_io_length);
1352 blk_start_plug(&plug);
1353 for (i = 0; i < bp->b_map_count; i++) {
1354 xfs_buf_ioapply_map(bp, i, &offset, &size, rw);
1355 if (bp->b_error)
1356 break;
1357 if (size <= 0)
1358 break; /* all done */
1359 }
1360 blk_finish_plug(&plug);
1361}
1362
1363void
1364xfs_buf_iorequest(
1365 xfs_buf_t *bp)
1366{
1367 trace_xfs_buf_iorequest(bp, _RET_IP_);
1368
1369 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1370
1371 if (bp->b_flags & XBF_WRITE)
1372 xfs_buf_wait_unpin(bp);
1373 xfs_buf_hold(bp);
1374
1375 /*
1376 * Set the count to 1 initially, this will stop an I/O
1377 * completion callout which happens before we have started
1378 * all the I/O from calling xfs_buf_ioend too early.
1379 */
1380 atomic_set(&bp->b_io_remaining, 1);
1381 _xfs_buf_ioapply(bp);
1382 /*
1383 * If _xfs_buf_ioapply failed, we'll get back here with
1384 * only the reference we took above. _xfs_buf_ioend will
1385 * drop it to zero, so we'd better not queue it for later,
1386 * or we'll free it before it's done.
1387 */
1388 _xfs_buf_ioend(bp, bp->b_error ? 0 : 1);
1389
1390 xfs_buf_rele(bp);
1391}
1392
1393/*
1394 * Waits for I/O to complete on the buffer supplied. It returns immediately if
1395 * no I/O is pending or there is already a pending error on the buffer, in which
1396 * case nothing will ever complete. It returns the I/O error code, if any, or
1397 * 0 if there was no error.
1398 */
1399int
1400xfs_buf_iowait(
1401 xfs_buf_t *bp)
1402{
1403 trace_xfs_buf_iowait(bp, _RET_IP_);
1404
1405 if (!bp->b_error)
1406 wait_for_completion(&bp->b_iowait);
1407
1408 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1409 return bp->b_error;
1410}
1411
1412xfs_caddr_t
1413xfs_buf_offset(
1414 xfs_buf_t *bp,
1415 size_t offset)
1416{
1417 struct page *page;
1418
1419 if (bp->b_addr)
1420 return bp->b_addr + offset;
1421
1422 offset += bp->b_offset;
1423 page = bp->b_pages[offset >> PAGE_SHIFT];
1424 return (xfs_caddr_t)page_address(page) + (offset & (PAGE_SIZE-1));
1425}
1426
1427/*
1428 * Move data into or out of a buffer.
1429 */
1430void
1431xfs_buf_iomove(
1432 xfs_buf_t *bp, /* buffer to process */
1433 size_t boff, /* starting buffer offset */
1434 size_t bsize, /* length to copy */
1435 void *data, /* data address */
1436 xfs_buf_rw_t mode) /* read/write/zero flag */
1437{
1438 size_t bend;
1439
1440 bend = boff + bsize;
1441 while (boff < bend) {
1442 struct page *page;
1443 int page_index, page_offset, csize;
1444
1445 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1446 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1447 page = bp->b_pages[page_index];
1448 csize = min_t(size_t, PAGE_SIZE - page_offset,
1449 BBTOB(bp->b_io_length) - boff);
1450
1451 ASSERT((csize + page_offset) <= PAGE_SIZE);
1452
1453 switch (mode) {
1454 case XBRW_ZERO:
1455 memset(page_address(page) + page_offset, 0, csize);
1456 break;
1457 case XBRW_READ:
1458 memcpy(data, page_address(page) + page_offset, csize);
1459 break;
1460 case XBRW_WRITE:
1461 memcpy(page_address(page) + page_offset, data, csize);
1462 }
1463
1464 boff += csize;
1465 data += csize;
1466 }
1467}
1468
1469/*
1470 * Handling of buffer targets (buftargs).
1471 */
1472
1473/*
1474 * Wait for any bufs with callbacks that have been submitted but have not yet
1475 * returned. These buffers will have an elevated hold count, so wait on those
1476 * while freeing all the buffers only held by the LRU.
1477 */
1478static enum lru_status
1479xfs_buftarg_wait_rele(
1480 struct list_head *item,
1481 spinlock_t *lru_lock,
1482 void *arg)
1483
1484{
1485 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1486 struct list_head *dispose = arg;
1487
1488 if (atomic_read(&bp->b_hold) > 1) {
1489 /* need to wait, so skip it this pass */
1490 trace_xfs_buf_wait_buftarg(bp, _RET_IP_);
1491 return LRU_SKIP;
1492 }
1493 if (!spin_trylock(&bp->b_lock))
1494 return LRU_SKIP;
1495
1496 /*
1497 * clear the LRU reference count so the buffer doesn't get
1498 * ignored in xfs_buf_rele().
1499 */
1500 atomic_set(&bp->b_lru_ref, 0);
1501 bp->b_state |= XFS_BSTATE_DISPOSE;
1502 list_move(item, dispose);
1503 spin_unlock(&bp->b_lock);
1504 return LRU_REMOVED;
1505}
1506
1507void
1508xfs_wait_buftarg(
1509 struct xfs_buftarg *btp)
1510{
1511 LIST_HEAD(dispose);
1512 int loop = 0;
1513
1514 /* loop until there is nothing left on the lru list. */
1515 while (list_lru_count(&btp->bt_lru)) {
1516 list_lru_walk(&btp->bt_lru, xfs_buftarg_wait_rele,
1517 &dispose, LONG_MAX);
1518
1519 while (!list_empty(&dispose)) {
1520 struct xfs_buf *bp;
1521 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1522 list_del_init(&bp->b_lru);
1523 if (bp->b_flags & XBF_WRITE_FAIL) {
1524 xfs_alert(btp->bt_mount,
1525"Corruption Alert: Buffer at block 0x%llx had permanent write failures!\n"
1526"Please run xfs_repair to determine the extent of the problem.",
1527 (long long)bp->b_bn);
1528 }
1529 xfs_buf_rele(bp);
1530 }
1531 if (loop++ != 0)
1532 delay(100);
1533 }
1534}
1535
1536static enum lru_status
1537xfs_buftarg_isolate(
1538 struct list_head *item,
1539 spinlock_t *lru_lock,
1540 void *arg)
1541{
1542 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1543 struct list_head *dispose = arg;
1544
1545 /*
1546 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1547 * If we fail to get the lock, just skip it.
1548 */
1549 if (!spin_trylock(&bp->b_lock))
1550 return LRU_SKIP;
1551 /*
1552 * Decrement the b_lru_ref count unless the value is already
1553 * zero. If the value is already zero, we need to reclaim the
1554 * buffer, otherwise it gets another trip through the LRU.
1555 */
1556 if (!atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1557 spin_unlock(&bp->b_lock);
1558 return LRU_ROTATE;
1559 }
1560
1561 bp->b_state |= XFS_BSTATE_DISPOSE;
1562 list_move(item, dispose);
1563 spin_unlock(&bp->b_lock);
1564 return LRU_REMOVED;
1565}
1566
1567static unsigned long
1568xfs_buftarg_shrink_scan(
1569 struct shrinker *shrink,
1570 struct shrink_control *sc)
1571{
1572 struct xfs_buftarg *btp = container_of(shrink,
1573 struct xfs_buftarg, bt_shrinker);
1574 LIST_HEAD(dispose);
1575 unsigned long freed;
1576 unsigned long nr_to_scan = sc->nr_to_scan;
1577
1578 freed = list_lru_walk_node(&btp->bt_lru, sc->nid, xfs_buftarg_isolate,
1579 &dispose, &nr_to_scan);
1580
1581 while (!list_empty(&dispose)) {
1582 struct xfs_buf *bp;
1583 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1584 list_del_init(&bp->b_lru);
1585 xfs_buf_rele(bp);
1586 }
1587
1588 return freed;
1589}
1590
1591static unsigned long
1592xfs_buftarg_shrink_count(
1593 struct shrinker *shrink,
1594 struct shrink_control *sc)
1595{
1596 struct xfs_buftarg *btp = container_of(shrink,
1597 struct xfs_buftarg, bt_shrinker);
1598 return list_lru_count_node(&btp->bt_lru, sc->nid);
1599}
1600
1601void
1602xfs_free_buftarg(
1603 struct xfs_mount *mp,
1604 struct xfs_buftarg *btp)
1605{
1606 unregister_shrinker(&btp->bt_shrinker);
1607 list_lru_destroy(&btp->bt_lru);
1608
1609 if (mp->m_flags & XFS_MOUNT_BARRIER)
1610 xfs_blkdev_issue_flush(btp);
1611
1612 kmem_free(btp);
1613}
1614
1615int
1616xfs_setsize_buftarg(
1617 xfs_buftarg_t *btp,
1618 unsigned int blocksize,
1619 unsigned int sectorsize)
1620{
1621 /* Set up metadata sector size info */
1622 btp->bt_meta_sectorsize = sectorsize;
1623 btp->bt_meta_sectormask = sectorsize - 1;
1624
1625 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1626 char name[BDEVNAME_SIZE];
1627
1628 bdevname(btp->bt_bdev, name);
1629
1630 xfs_warn(btp->bt_mount,
1631 "Cannot set_blocksize to %u on device %s",
1632 sectorsize, name);
1633 return EINVAL;
1634 }
1635
1636 /* Set up device logical sector size mask */
1637 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1638 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1639
1640 return 0;
1641}
1642
1643/*
1644 * When allocating the initial buffer target we have not yet
1645 * read in the superblock, so don't know what sized sectors
1646 * are being used at this early stage. Play safe.
1647 */
1648STATIC int
1649xfs_setsize_buftarg_early(
1650 xfs_buftarg_t *btp,
1651 struct block_device *bdev)
1652{
1653 return xfs_setsize_buftarg(btp, PAGE_SIZE,
1654 bdev_logical_block_size(bdev));
1655}
1656
1657xfs_buftarg_t *
1658xfs_alloc_buftarg(
1659 struct xfs_mount *mp,
1660 struct block_device *bdev,
1661 int external,
1662 const char *fsname)
1663{
1664 xfs_buftarg_t *btp;
1665
1666 btp = kmem_zalloc(sizeof(*btp), KM_SLEEP | KM_NOFS);
1667
1668 btp->bt_mount = mp;
1669 btp->bt_dev = bdev->bd_dev;
1670 btp->bt_bdev = bdev;
1671 btp->bt_bdi = blk_get_backing_dev_info(bdev);
1672 if (!btp->bt_bdi)
1673 goto error;
1674
1675 if (xfs_setsize_buftarg_early(btp, bdev))
1676 goto error;
1677
1678 if (list_lru_init(&btp->bt_lru))
1679 goto error;
1680
1681 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1682 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1683 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1684 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1685 register_shrinker(&btp->bt_shrinker);
1686 return btp;
1687
1688error:
1689 kmem_free(btp);
1690 return NULL;
1691}
1692
1693/*
1694 * Add a buffer to the delayed write list.
1695 *
1696 * This queues a buffer for writeout if it hasn't already been. Note that
1697 * neither this routine nor the buffer list submission functions perform
1698 * any internal synchronization. It is expected that the lists are thread-local
1699 * to the callers.
1700 *
1701 * Returns true if we queued up the buffer, or false if it already had
1702 * been on the buffer list.
1703 */
1704bool
1705xfs_buf_delwri_queue(
1706 struct xfs_buf *bp,
1707 struct list_head *list)
1708{
1709 ASSERT(xfs_buf_islocked(bp));
1710 ASSERT(!(bp->b_flags & XBF_READ));
1711
1712 /*
1713 * If the buffer is already marked delwri it already is queued up
1714 * by someone else for imediate writeout. Just ignore it in that
1715 * case.
1716 */
1717 if (bp->b_flags & _XBF_DELWRI_Q) {
1718 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
1719 return false;
1720 }
1721
1722 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
1723
1724 /*
1725 * If a buffer gets written out synchronously or marked stale while it
1726 * is on a delwri list we lazily remove it. To do this, the other party
1727 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
1728 * It remains referenced and on the list. In a rare corner case it
1729 * might get readded to a delwri list after the synchronous writeout, in
1730 * which case we need just need to re-add the flag here.
1731 */
1732 bp->b_flags |= _XBF_DELWRI_Q;
1733 if (list_empty(&bp->b_list)) {
1734 atomic_inc(&bp->b_hold);
1735 list_add_tail(&bp->b_list, list);
1736 }
1737
1738 return true;
1739}
1740
1741/*
1742 * Compare function is more complex than it needs to be because
1743 * the return value is only 32 bits and we are doing comparisons
1744 * on 64 bit values
1745 */
1746static int
1747xfs_buf_cmp(
1748 void *priv,
1749 struct list_head *a,
1750 struct list_head *b)
1751{
1752 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
1753 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
1754 xfs_daddr_t diff;
1755
1756 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
1757 if (diff < 0)
1758 return -1;
1759 if (diff > 0)
1760 return 1;
1761 return 0;
1762}
1763
1764static int
1765__xfs_buf_delwri_submit(
1766 struct list_head *buffer_list,
1767 struct list_head *io_list,
1768 bool wait)
1769{
1770 struct blk_plug plug;
1771 struct xfs_buf *bp, *n;
1772 int pinned = 0;
1773
1774 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
1775 if (!wait) {
1776 if (xfs_buf_ispinned(bp)) {
1777 pinned++;
1778 continue;
1779 }
1780 if (!xfs_buf_trylock(bp))
1781 continue;
1782 } else {
1783 xfs_buf_lock(bp);
1784 }
1785
1786 /*
1787 * Someone else might have written the buffer synchronously or
1788 * marked it stale in the meantime. In that case only the
1789 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
1790 * reference and remove it from the list here.
1791 */
1792 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
1793 list_del_init(&bp->b_list);
1794 xfs_buf_relse(bp);
1795 continue;
1796 }
1797
1798 list_move_tail(&bp->b_list, io_list);
1799 trace_xfs_buf_delwri_split(bp, _RET_IP_);
1800 }
1801
1802 list_sort(NULL, io_list, xfs_buf_cmp);
1803
1804 blk_start_plug(&plug);
1805 list_for_each_entry_safe(bp, n, io_list, b_list) {
1806 bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC | XBF_WRITE_FAIL);
1807 bp->b_flags |= XBF_WRITE;
1808
1809 if (!wait) {
1810 bp->b_flags |= XBF_ASYNC;
1811 list_del_init(&bp->b_list);
1812 }
1813 xfs_bdstrat_cb(bp);
1814 }
1815 blk_finish_plug(&plug);
1816
1817 return pinned;
1818}
1819
1820/*
1821 * Write out a buffer list asynchronously.
1822 *
1823 * This will take the @buffer_list, write all non-locked and non-pinned buffers
1824 * out and not wait for I/O completion on any of the buffers. This interface
1825 * is only safely useable for callers that can track I/O completion by higher
1826 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
1827 * function.
1828 */
1829int
1830xfs_buf_delwri_submit_nowait(
1831 struct list_head *buffer_list)
1832{
1833 LIST_HEAD (io_list);
1834 return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
1835}
1836
1837/*
1838 * Write out a buffer list synchronously.
1839 *
1840 * This will take the @buffer_list, write all buffers out and wait for I/O
1841 * completion on all of the buffers. @buffer_list is consumed by the function,
1842 * so callers must have some other way of tracking buffers if they require such
1843 * functionality.
1844 */
1845int
1846xfs_buf_delwri_submit(
1847 struct list_head *buffer_list)
1848{
1849 LIST_HEAD (io_list);
1850 int error = 0, error2;
1851 struct xfs_buf *bp;
1852
1853 __xfs_buf_delwri_submit(buffer_list, &io_list, true);
1854
1855 /* Wait for IO to complete. */
1856 while (!list_empty(&io_list)) {
1857 bp = list_first_entry(&io_list, struct xfs_buf, b_list);
1858
1859 list_del_init(&bp->b_list);
1860 error2 = xfs_buf_iowait(bp);
1861 xfs_buf_relse(bp);
1862 if (!error)
1863 error = error2;
1864 }
1865
1866 return error;
1867}
1868
1869int __init
1870xfs_buf_init(void)
1871{
1872 xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
1873 KM_ZONE_HWALIGN, NULL);
1874 if (!xfs_buf_zone)
1875 goto out;
1876
1877 xfslogd_workqueue = alloc_workqueue("xfslogd",
1878 WQ_MEM_RECLAIM | WQ_HIGHPRI, 1);
1879 if (!xfslogd_workqueue)
1880 goto out_free_buf_zone;
1881
1882 return 0;
1883
1884 out_free_buf_zone:
1885 kmem_zone_destroy(xfs_buf_zone);
1886 out:
1887 return -ENOMEM;
1888}
1889
1890void
1891xfs_buf_terminate(void)
1892{
1893 destroy_workqueue(xfslogd_workqueue);
1894 kmem_zone_destroy(xfs_buf_zone);
1895}