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// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include <linux/backing-dev.h>
8
9#include "xfs_shared.h"
10#include "xfs_format.h"
11#include "xfs_log_format.h"
12#include "xfs_trans_resv.h"
13#include "xfs_mount.h"
14#include "xfs_trace.h"
15#include "xfs_log.h"
16#include "xfs_log_recover.h"
17#include "xfs_trans.h"
18#include "xfs_buf_item.h"
19#include "xfs_errortag.h"
20#include "xfs_error.h"
21#include "xfs_ag.h"
22
23static kmem_zone_t *xfs_buf_zone;
24
25/*
26 * Locking orders
27 *
28 * xfs_buf_ioacct_inc:
29 * xfs_buf_ioacct_dec:
30 * b_sema (caller holds)
31 * b_lock
32 *
33 * xfs_buf_stale:
34 * b_sema (caller holds)
35 * b_lock
36 * lru_lock
37 *
38 * xfs_buf_rele:
39 * b_lock
40 * pag_buf_lock
41 * lru_lock
42 *
43 * xfs_buftarg_drain_rele
44 * lru_lock
45 * b_lock (trylock due to inversion)
46 *
47 * xfs_buftarg_isolate
48 * lru_lock
49 * b_lock (trylock due to inversion)
50 */
51
52static int __xfs_buf_submit(struct xfs_buf *bp, bool wait);
53
54static inline int
55xfs_buf_submit(
56 struct xfs_buf *bp)
57{
58 return __xfs_buf_submit(bp, !(bp->b_flags & XBF_ASYNC));
59}
60
61static inline int
62xfs_buf_is_vmapped(
63 struct xfs_buf *bp)
64{
65 /*
66 * Return true if the buffer is vmapped.
67 *
68 * b_addr is null if the buffer is not mapped, but the code is clever
69 * enough to know it doesn't have to map a single page, so the check has
70 * to be both for b_addr and bp->b_page_count > 1.
71 */
72 return bp->b_addr && bp->b_page_count > 1;
73}
74
75static inline int
76xfs_buf_vmap_len(
77 struct xfs_buf *bp)
78{
79 return (bp->b_page_count * PAGE_SIZE);
80}
81
82/*
83 * Bump the I/O in flight count on the buftarg if we haven't yet done so for
84 * this buffer. The count is incremented once per buffer (per hold cycle)
85 * because the corresponding decrement is deferred to buffer release. Buffers
86 * can undergo I/O multiple times in a hold-release cycle and per buffer I/O
87 * tracking adds unnecessary overhead. This is used for sychronization purposes
88 * with unmount (see xfs_buftarg_drain()), so all we really need is a count of
89 * in-flight buffers.
90 *
91 * Buffers that are never released (e.g., superblock, iclog buffers) must set
92 * the XBF_NO_IOACCT flag before I/O submission. Otherwise, the buftarg count
93 * never reaches zero and unmount hangs indefinitely.
94 */
95static inline void
96xfs_buf_ioacct_inc(
97 struct xfs_buf *bp)
98{
99 if (bp->b_flags & XBF_NO_IOACCT)
100 return;
101
102 ASSERT(bp->b_flags & XBF_ASYNC);
103 spin_lock(&bp->b_lock);
104 if (!(bp->b_state & XFS_BSTATE_IN_FLIGHT)) {
105 bp->b_state |= XFS_BSTATE_IN_FLIGHT;
106 percpu_counter_inc(&bp->b_target->bt_io_count);
107 }
108 spin_unlock(&bp->b_lock);
109}
110
111/*
112 * Clear the in-flight state on a buffer about to be released to the LRU or
113 * freed and unaccount from the buftarg.
114 */
115static inline void
116__xfs_buf_ioacct_dec(
117 struct xfs_buf *bp)
118{
119 lockdep_assert_held(&bp->b_lock);
120
121 if (bp->b_state & XFS_BSTATE_IN_FLIGHT) {
122 bp->b_state &= ~XFS_BSTATE_IN_FLIGHT;
123 percpu_counter_dec(&bp->b_target->bt_io_count);
124 }
125}
126
127static inline void
128xfs_buf_ioacct_dec(
129 struct xfs_buf *bp)
130{
131 spin_lock(&bp->b_lock);
132 __xfs_buf_ioacct_dec(bp);
133 spin_unlock(&bp->b_lock);
134}
135
136/*
137 * When we mark a buffer stale, we remove the buffer from the LRU and clear the
138 * b_lru_ref count so that the buffer is freed immediately when the buffer
139 * reference count falls to zero. If the buffer is already on the LRU, we need
140 * to remove the reference that LRU holds on the buffer.
141 *
142 * This prevents build-up of stale buffers on the LRU.
143 */
144void
145xfs_buf_stale(
146 struct xfs_buf *bp)
147{
148 ASSERT(xfs_buf_islocked(bp));
149
150 bp->b_flags |= XBF_STALE;
151
152 /*
153 * Clear the delwri status so that a delwri queue walker will not
154 * flush this buffer to disk now that it is stale. The delwri queue has
155 * a reference to the buffer, so this is safe to do.
156 */
157 bp->b_flags &= ~_XBF_DELWRI_Q;
158
159 /*
160 * Once the buffer is marked stale and unlocked, a subsequent lookup
161 * could reset b_flags. There is no guarantee that the buffer is
162 * unaccounted (released to LRU) before that occurs. Drop in-flight
163 * status now to preserve accounting consistency.
164 */
165 spin_lock(&bp->b_lock);
166 __xfs_buf_ioacct_dec(bp);
167
168 atomic_set(&bp->b_lru_ref, 0);
169 if (!(bp->b_state & XFS_BSTATE_DISPOSE) &&
170 (list_lru_del(&bp->b_target->bt_lru, &bp->b_lru)))
171 atomic_dec(&bp->b_hold);
172
173 ASSERT(atomic_read(&bp->b_hold) >= 1);
174 spin_unlock(&bp->b_lock);
175}
176
177static int
178xfs_buf_get_maps(
179 struct xfs_buf *bp,
180 int map_count)
181{
182 ASSERT(bp->b_maps == NULL);
183 bp->b_map_count = map_count;
184
185 if (map_count == 1) {
186 bp->b_maps = &bp->__b_map;
187 return 0;
188 }
189
190 bp->b_maps = kmem_zalloc(map_count * sizeof(struct xfs_buf_map),
191 KM_NOFS);
192 if (!bp->b_maps)
193 return -ENOMEM;
194 return 0;
195}
196
197/*
198 * Frees b_pages if it was allocated.
199 */
200static void
201xfs_buf_free_maps(
202 struct xfs_buf *bp)
203{
204 if (bp->b_maps != &bp->__b_map) {
205 kmem_free(bp->b_maps);
206 bp->b_maps = NULL;
207 }
208}
209
210static int
211_xfs_buf_alloc(
212 struct xfs_buftarg *target,
213 struct xfs_buf_map *map,
214 int nmaps,
215 xfs_buf_flags_t flags,
216 struct xfs_buf **bpp)
217{
218 struct xfs_buf *bp;
219 int error;
220 int i;
221
222 *bpp = NULL;
223 bp = kmem_cache_zalloc(xfs_buf_zone, GFP_NOFS | __GFP_NOFAIL);
224
225 /*
226 * We don't want certain flags to appear in b_flags unless they are
227 * specifically set by later operations on the buffer.
228 */
229 flags &= ~(XBF_UNMAPPED | XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD);
230
231 atomic_set(&bp->b_hold, 1);
232 atomic_set(&bp->b_lru_ref, 1);
233 init_completion(&bp->b_iowait);
234 INIT_LIST_HEAD(&bp->b_lru);
235 INIT_LIST_HEAD(&bp->b_list);
236 INIT_LIST_HEAD(&bp->b_li_list);
237 sema_init(&bp->b_sema, 0); /* held, no waiters */
238 spin_lock_init(&bp->b_lock);
239 bp->b_target = target;
240 bp->b_mount = target->bt_mount;
241 bp->b_flags = flags;
242
243 /*
244 * Set length and io_length to the same value initially.
245 * I/O routines should use io_length, which will be the same in
246 * most cases but may be reset (e.g. XFS recovery).
247 */
248 error = xfs_buf_get_maps(bp, nmaps);
249 if (error) {
250 kmem_cache_free(xfs_buf_zone, bp);
251 return error;
252 }
253
254 bp->b_bn = map[0].bm_bn;
255 bp->b_length = 0;
256 for (i = 0; i < nmaps; i++) {
257 bp->b_maps[i].bm_bn = map[i].bm_bn;
258 bp->b_maps[i].bm_len = map[i].bm_len;
259 bp->b_length += map[i].bm_len;
260 }
261
262 atomic_set(&bp->b_pin_count, 0);
263 init_waitqueue_head(&bp->b_waiters);
264
265 XFS_STATS_INC(bp->b_mount, xb_create);
266 trace_xfs_buf_init(bp, _RET_IP_);
267
268 *bpp = bp;
269 return 0;
270}
271
272static void
273xfs_buf_free_pages(
274 struct xfs_buf *bp)
275{
276 uint i;
277
278 ASSERT(bp->b_flags & _XBF_PAGES);
279
280 if (xfs_buf_is_vmapped(bp))
281 vm_unmap_ram(bp->b_addr, bp->b_page_count);
282
283 for (i = 0; i < bp->b_page_count; i++) {
284 if (bp->b_pages[i])
285 __free_page(bp->b_pages[i]);
286 }
287 if (current->reclaim_state)
288 current->reclaim_state->reclaimed_slab += bp->b_page_count;
289
290 if (bp->b_pages != bp->b_page_array)
291 kmem_free(bp->b_pages);
292 bp->b_pages = NULL;
293 bp->b_flags &= ~_XBF_PAGES;
294}
295
296static void
297xfs_buf_free(
298 struct xfs_buf *bp)
299{
300 trace_xfs_buf_free(bp, _RET_IP_);
301
302 ASSERT(list_empty(&bp->b_lru));
303
304 if (bp->b_flags & _XBF_PAGES)
305 xfs_buf_free_pages(bp);
306 else if (bp->b_flags & _XBF_KMEM)
307 kmem_free(bp->b_addr);
308
309 xfs_buf_free_maps(bp);
310 kmem_cache_free(xfs_buf_zone, bp);
311}
312
313static int
314xfs_buf_alloc_kmem(
315 struct xfs_buf *bp,
316 xfs_buf_flags_t flags)
317{
318 int align_mask = xfs_buftarg_dma_alignment(bp->b_target);
319 xfs_km_flags_t kmflag_mask = KM_NOFS;
320 size_t size = BBTOB(bp->b_length);
321
322 /* Assure zeroed buffer for non-read cases. */
323 if (!(flags & XBF_READ))
324 kmflag_mask |= KM_ZERO;
325
326 bp->b_addr = kmem_alloc_io(size, align_mask, kmflag_mask);
327 if (!bp->b_addr)
328 return -ENOMEM;
329
330 if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) !=
331 ((unsigned long)bp->b_addr & PAGE_MASK)) {
332 /* b_addr spans two pages - use alloc_page instead */
333 kmem_free(bp->b_addr);
334 bp->b_addr = NULL;
335 return -ENOMEM;
336 }
337 bp->b_offset = offset_in_page(bp->b_addr);
338 bp->b_pages = bp->b_page_array;
339 bp->b_pages[0] = kmem_to_page(bp->b_addr);
340 bp->b_page_count = 1;
341 bp->b_flags |= _XBF_KMEM;
342 return 0;
343}
344
345static int
346xfs_buf_alloc_pages(
347 struct xfs_buf *bp,
348 xfs_buf_flags_t flags)
349{
350 gfp_t gfp_mask = __GFP_NOWARN;
351 long filled = 0;
352
353 if (flags & XBF_READ_AHEAD)
354 gfp_mask |= __GFP_NORETRY;
355 else
356 gfp_mask |= GFP_NOFS;
357
358 /* Make sure that we have a page list */
359 bp->b_page_count = DIV_ROUND_UP(BBTOB(bp->b_length), PAGE_SIZE);
360 if (bp->b_page_count <= XB_PAGES) {
361 bp->b_pages = bp->b_page_array;
362 } else {
363 bp->b_pages = kzalloc(sizeof(struct page *) * bp->b_page_count,
364 gfp_mask);
365 if (!bp->b_pages)
366 return -ENOMEM;
367 }
368 bp->b_flags |= _XBF_PAGES;
369
370 /* Assure zeroed buffer for non-read cases. */
371 if (!(flags & XBF_READ))
372 gfp_mask |= __GFP_ZERO;
373
374 /*
375 * Bulk filling of pages can take multiple calls. Not filling the entire
376 * array is not an allocation failure, so don't back off if we get at
377 * least one extra page.
378 */
379 for (;;) {
380 long last = filled;
381
382 filled = alloc_pages_bulk_array(gfp_mask, bp->b_page_count,
383 bp->b_pages);
384 if (filled == bp->b_page_count) {
385 XFS_STATS_INC(bp->b_mount, xb_page_found);
386 break;
387 }
388
389 if (filled != last)
390 continue;
391
392 if (flags & XBF_READ_AHEAD) {
393 xfs_buf_free_pages(bp);
394 return -ENOMEM;
395 }
396
397 XFS_STATS_INC(bp->b_mount, xb_page_retries);
398 congestion_wait(BLK_RW_ASYNC, HZ / 50);
399 }
400 return 0;
401}
402
403/*
404 * Map buffer into kernel address-space if necessary.
405 */
406STATIC int
407_xfs_buf_map_pages(
408 struct xfs_buf *bp,
409 uint flags)
410{
411 ASSERT(bp->b_flags & _XBF_PAGES);
412 if (bp->b_page_count == 1) {
413 /* A single page buffer is always mappable */
414 bp->b_addr = page_address(bp->b_pages[0]);
415 } else if (flags & XBF_UNMAPPED) {
416 bp->b_addr = NULL;
417 } else {
418 int retried = 0;
419 unsigned nofs_flag;
420
421 /*
422 * vm_map_ram() will allocate auxiliary structures (e.g.
423 * pagetables) with GFP_KERNEL, yet we are likely to be under
424 * GFP_NOFS context here. Hence we need to tell memory reclaim
425 * that we are in such a context via PF_MEMALLOC_NOFS to prevent
426 * memory reclaim re-entering the filesystem here and
427 * potentially deadlocking.
428 */
429 nofs_flag = memalloc_nofs_save();
430 do {
431 bp->b_addr = vm_map_ram(bp->b_pages, bp->b_page_count,
432 -1);
433 if (bp->b_addr)
434 break;
435 vm_unmap_aliases();
436 } while (retried++ <= 1);
437 memalloc_nofs_restore(nofs_flag);
438
439 if (!bp->b_addr)
440 return -ENOMEM;
441 }
442
443 return 0;
444}
445
446/*
447 * Finding and Reading Buffers
448 */
449static int
450_xfs_buf_obj_cmp(
451 struct rhashtable_compare_arg *arg,
452 const void *obj)
453{
454 const struct xfs_buf_map *map = arg->key;
455 const struct xfs_buf *bp = obj;
456
457 /*
458 * The key hashing in the lookup path depends on the key being the
459 * first element of the compare_arg, make sure to assert this.
460 */
461 BUILD_BUG_ON(offsetof(struct xfs_buf_map, bm_bn) != 0);
462
463 if (bp->b_bn != map->bm_bn)
464 return 1;
465
466 if (unlikely(bp->b_length != map->bm_len)) {
467 /*
468 * found a block number match. If the range doesn't
469 * match, the only way this is allowed is if the buffer
470 * in the cache is stale and the transaction that made
471 * it stale has not yet committed. i.e. we are
472 * reallocating a busy extent. Skip this buffer and
473 * continue searching for an exact match.
474 */
475 ASSERT(bp->b_flags & XBF_STALE);
476 return 1;
477 }
478 return 0;
479}
480
481static const struct rhashtable_params xfs_buf_hash_params = {
482 .min_size = 32, /* empty AGs have minimal footprint */
483 .nelem_hint = 16,
484 .key_len = sizeof(xfs_daddr_t),
485 .key_offset = offsetof(struct xfs_buf, b_bn),
486 .head_offset = offsetof(struct xfs_buf, b_rhash_head),
487 .automatic_shrinking = true,
488 .obj_cmpfn = _xfs_buf_obj_cmp,
489};
490
491int
492xfs_buf_hash_init(
493 struct xfs_perag *pag)
494{
495 spin_lock_init(&pag->pag_buf_lock);
496 return rhashtable_init(&pag->pag_buf_hash, &xfs_buf_hash_params);
497}
498
499void
500xfs_buf_hash_destroy(
501 struct xfs_perag *pag)
502{
503 rhashtable_destroy(&pag->pag_buf_hash);
504}
505
506/*
507 * Look up a buffer in the buffer cache and return it referenced and locked
508 * in @found_bp.
509 *
510 * If @new_bp is supplied and we have a lookup miss, insert @new_bp into the
511 * cache.
512 *
513 * If XBF_TRYLOCK is set in @flags, only try to lock the buffer and return
514 * -EAGAIN if we fail to lock it.
515 *
516 * Return values are:
517 * -EFSCORRUPTED if have been supplied with an invalid address
518 * -EAGAIN on trylock failure
519 * -ENOENT if we fail to find a match and @new_bp was NULL
520 * 0, with @found_bp:
521 * - @new_bp if we inserted it into the cache
522 * - the buffer we found and locked.
523 */
524static int
525xfs_buf_find(
526 struct xfs_buftarg *btp,
527 struct xfs_buf_map *map,
528 int nmaps,
529 xfs_buf_flags_t flags,
530 struct xfs_buf *new_bp,
531 struct xfs_buf **found_bp)
532{
533 struct xfs_perag *pag;
534 struct xfs_buf *bp;
535 struct xfs_buf_map cmap = { .bm_bn = map[0].bm_bn };
536 xfs_daddr_t eofs;
537 int i;
538
539 *found_bp = NULL;
540
541 for (i = 0; i < nmaps; i++)
542 cmap.bm_len += map[i].bm_len;
543
544 /* Check for IOs smaller than the sector size / not sector aligned */
545 ASSERT(!(BBTOB(cmap.bm_len) < btp->bt_meta_sectorsize));
546 ASSERT(!(BBTOB(cmap.bm_bn) & (xfs_off_t)btp->bt_meta_sectormask));
547
548 /*
549 * Corrupted block numbers can get through to here, unfortunately, so we
550 * have to check that the buffer falls within the filesystem bounds.
551 */
552 eofs = XFS_FSB_TO_BB(btp->bt_mount, btp->bt_mount->m_sb.sb_dblocks);
553 if (cmap.bm_bn < 0 || cmap.bm_bn >= eofs) {
554 xfs_alert(btp->bt_mount,
555 "%s: daddr 0x%llx out of range, EOFS 0x%llx",
556 __func__, cmap.bm_bn, eofs);
557 WARN_ON(1);
558 return -EFSCORRUPTED;
559 }
560
561 pag = xfs_perag_get(btp->bt_mount,
562 xfs_daddr_to_agno(btp->bt_mount, cmap.bm_bn));
563
564 spin_lock(&pag->pag_buf_lock);
565 bp = rhashtable_lookup_fast(&pag->pag_buf_hash, &cmap,
566 xfs_buf_hash_params);
567 if (bp) {
568 atomic_inc(&bp->b_hold);
569 goto found;
570 }
571
572 /* No match found */
573 if (!new_bp) {
574 XFS_STATS_INC(btp->bt_mount, xb_miss_locked);
575 spin_unlock(&pag->pag_buf_lock);
576 xfs_perag_put(pag);
577 return -ENOENT;
578 }
579
580 /* the buffer keeps the perag reference until it is freed */
581 new_bp->b_pag = pag;
582 rhashtable_insert_fast(&pag->pag_buf_hash, &new_bp->b_rhash_head,
583 xfs_buf_hash_params);
584 spin_unlock(&pag->pag_buf_lock);
585 *found_bp = new_bp;
586 return 0;
587
588found:
589 spin_unlock(&pag->pag_buf_lock);
590 xfs_perag_put(pag);
591
592 if (!xfs_buf_trylock(bp)) {
593 if (flags & XBF_TRYLOCK) {
594 xfs_buf_rele(bp);
595 XFS_STATS_INC(btp->bt_mount, xb_busy_locked);
596 return -EAGAIN;
597 }
598 xfs_buf_lock(bp);
599 XFS_STATS_INC(btp->bt_mount, xb_get_locked_waited);
600 }
601
602 /*
603 * if the buffer is stale, clear all the external state associated with
604 * it. We need to keep flags such as how we allocated the buffer memory
605 * intact here.
606 */
607 if (bp->b_flags & XBF_STALE) {
608 ASSERT((bp->b_flags & _XBF_DELWRI_Q) == 0);
609 bp->b_flags &= _XBF_KMEM | _XBF_PAGES;
610 bp->b_ops = NULL;
611 }
612
613 trace_xfs_buf_find(bp, flags, _RET_IP_);
614 XFS_STATS_INC(btp->bt_mount, xb_get_locked);
615 *found_bp = bp;
616 return 0;
617}
618
619struct xfs_buf *
620xfs_buf_incore(
621 struct xfs_buftarg *target,
622 xfs_daddr_t blkno,
623 size_t numblks,
624 xfs_buf_flags_t flags)
625{
626 struct xfs_buf *bp;
627 int error;
628 DEFINE_SINGLE_BUF_MAP(map, blkno, numblks);
629
630 error = xfs_buf_find(target, &map, 1, flags, NULL, &bp);
631 if (error)
632 return NULL;
633 return bp;
634}
635
636/*
637 * Assembles a buffer covering the specified range. The code is optimised for
638 * cache hits, as metadata intensive workloads will see 3 orders of magnitude
639 * more hits than misses.
640 */
641int
642xfs_buf_get_map(
643 struct xfs_buftarg *target,
644 struct xfs_buf_map *map,
645 int nmaps,
646 xfs_buf_flags_t flags,
647 struct xfs_buf **bpp)
648{
649 struct xfs_buf *bp;
650 struct xfs_buf *new_bp;
651 int error;
652
653 *bpp = NULL;
654 error = xfs_buf_find(target, map, nmaps, flags, NULL, &bp);
655 if (!error)
656 goto found;
657 if (error != -ENOENT)
658 return error;
659
660 error = _xfs_buf_alloc(target, map, nmaps, flags, &new_bp);
661 if (error)
662 return error;
663
664 /*
665 * For buffers that fit entirely within a single page, first attempt to
666 * allocate the memory from the heap to minimise memory usage. If we
667 * can't get heap memory for these small buffers, we fall back to using
668 * the page allocator.
669 */
670 if (BBTOB(new_bp->b_length) >= PAGE_SIZE ||
671 xfs_buf_alloc_kmem(new_bp, flags) < 0) {
672 error = xfs_buf_alloc_pages(new_bp, flags);
673 if (error)
674 goto out_free_buf;
675 }
676
677 error = xfs_buf_find(target, map, nmaps, flags, new_bp, &bp);
678 if (error)
679 goto out_free_buf;
680
681 if (bp != new_bp)
682 xfs_buf_free(new_bp);
683
684found:
685 if (!bp->b_addr) {
686 error = _xfs_buf_map_pages(bp, flags);
687 if (unlikely(error)) {
688 xfs_warn_ratelimited(target->bt_mount,
689 "%s: failed to map %u pages", __func__,
690 bp->b_page_count);
691 xfs_buf_relse(bp);
692 return error;
693 }
694 }
695
696 /*
697 * Clear b_error if this is a lookup from a caller that doesn't expect
698 * valid data to be found in the buffer.
699 */
700 if (!(flags & XBF_READ))
701 xfs_buf_ioerror(bp, 0);
702
703 XFS_STATS_INC(target->bt_mount, xb_get);
704 trace_xfs_buf_get(bp, flags, _RET_IP_);
705 *bpp = bp;
706 return 0;
707out_free_buf:
708 xfs_buf_free(new_bp);
709 return error;
710}
711
712int
713_xfs_buf_read(
714 struct xfs_buf *bp,
715 xfs_buf_flags_t flags)
716{
717 ASSERT(!(flags & XBF_WRITE));
718 ASSERT(bp->b_maps[0].bm_bn != XFS_BUF_DADDR_NULL);
719
720 bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD | XBF_DONE);
721 bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
722
723 return xfs_buf_submit(bp);
724}
725
726/*
727 * Reverify a buffer found in cache without an attached ->b_ops.
728 *
729 * If the caller passed an ops structure and the buffer doesn't have ops
730 * assigned, set the ops and use it to verify the contents. If verification
731 * fails, clear XBF_DONE. We assume the buffer has no recorded errors and is
732 * already in XBF_DONE state on entry.
733 *
734 * Under normal operations, every in-core buffer is verified on read I/O
735 * completion. There are two scenarios that can lead to in-core buffers without
736 * an assigned ->b_ops. The first is during log recovery of buffers on a V4
737 * filesystem, though these buffers are purged at the end of recovery. The
738 * other is online repair, which intentionally reads with a NULL buffer ops to
739 * run several verifiers across an in-core buffer in order to establish buffer
740 * type. If repair can't establish that, the buffer will be left in memory
741 * with NULL buffer ops.
742 */
743int
744xfs_buf_reverify(
745 struct xfs_buf *bp,
746 const struct xfs_buf_ops *ops)
747{
748 ASSERT(bp->b_flags & XBF_DONE);
749 ASSERT(bp->b_error == 0);
750
751 if (!ops || bp->b_ops)
752 return 0;
753
754 bp->b_ops = ops;
755 bp->b_ops->verify_read(bp);
756 if (bp->b_error)
757 bp->b_flags &= ~XBF_DONE;
758 return bp->b_error;
759}
760
761int
762xfs_buf_read_map(
763 struct xfs_buftarg *target,
764 struct xfs_buf_map *map,
765 int nmaps,
766 xfs_buf_flags_t flags,
767 struct xfs_buf **bpp,
768 const struct xfs_buf_ops *ops,
769 xfs_failaddr_t fa)
770{
771 struct xfs_buf *bp;
772 int error;
773
774 flags |= XBF_READ;
775 *bpp = NULL;
776
777 error = xfs_buf_get_map(target, map, nmaps, flags, &bp);
778 if (error)
779 return error;
780
781 trace_xfs_buf_read(bp, flags, _RET_IP_);
782
783 if (!(bp->b_flags & XBF_DONE)) {
784 /* Initiate the buffer read and wait. */
785 XFS_STATS_INC(target->bt_mount, xb_get_read);
786 bp->b_ops = ops;
787 error = _xfs_buf_read(bp, flags);
788
789 /* Readahead iodone already dropped the buffer, so exit. */
790 if (flags & XBF_ASYNC)
791 return 0;
792 } else {
793 /* Buffer already read; all we need to do is check it. */
794 error = xfs_buf_reverify(bp, ops);
795
796 /* Readahead already finished; drop the buffer and exit. */
797 if (flags & XBF_ASYNC) {
798 xfs_buf_relse(bp);
799 return 0;
800 }
801
802 /* We do not want read in the flags */
803 bp->b_flags &= ~XBF_READ;
804 ASSERT(bp->b_ops != NULL || ops == NULL);
805 }
806
807 /*
808 * If we've had a read error, then the contents of the buffer are
809 * invalid and should not be used. To ensure that a followup read tries
810 * to pull the buffer from disk again, we clear the XBF_DONE flag and
811 * mark the buffer stale. This ensures that anyone who has a current
812 * reference to the buffer will interpret it's contents correctly and
813 * future cache lookups will also treat it as an empty, uninitialised
814 * buffer.
815 */
816 if (error) {
817 if (!XFS_FORCED_SHUTDOWN(target->bt_mount))
818 xfs_buf_ioerror_alert(bp, fa);
819
820 bp->b_flags &= ~XBF_DONE;
821 xfs_buf_stale(bp);
822 xfs_buf_relse(bp);
823
824 /* bad CRC means corrupted metadata */
825 if (error == -EFSBADCRC)
826 error = -EFSCORRUPTED;
827 return error;
828 }
829
830 *bpp = bp;
831 return 0;
832}
833
834/*
835 * If we are not low on memory then do the readahead in a deadlock
836 * safe manner.
837 */
838void
839xfs_buf_readahead_map(
840 struct xfs_buftarg *target,
841 struct xfs_buf_map *map,
842 int nmaps,
843 const struct xfs_buf_ops *ops)
844{
845 struct xfs_buf *bp;
846
847 if (bdi_read_congested(target->bt_bdev->bd_bdi))
848 return;
849
850 xfs_buf_read_map(target, map, nmaps,
851 XBF_TRYLOCK | XBF_ASYNC | XBF_READ_AHEAD, &bp, ops,
852 __this_address);
853}
854
855/*
856 * Read an uncached buffer from disk. Allocates and returns a locked
857 * buffer containing the disk contents or nothing.
858 */
859int
860xfs_buf_read_uncached(
861 struct xfs_buftarg *target,
862 xfs_daddr_t daddr,
863 size_t numblks,
864 int flags,
865 struct xfs_buf **bpp,
866 const struct xfs_buf_ops *ops)
867{
868 struct xfs_buf *bp;
869 int error;
870
871 *bpp = NULL;
872
873 error = xfs_buf_get_uncached(target, numblks, flags, &bp);
874 if (error)
875 return error;
876
877 /* set up the buffer for a read IO */
878 ASSERT(bp->b_map_count == 1);
879 bp->b_bn = XFS_BUF_DADDR_NULL; /* always null for uncached buffers */
880 bp->b_maps[0].bm_bn = daddr;
881 bp->b_flags |= XBF_READ;
882 bp->b_ops = ops;
883
884 xfs_buf_submit(bp);
885 if (bp->b_error) {
886 error = bp->b_error;
887 xfs_buf_relse(bp);
888 return error;
889 }
890
891 *bpp = bp;
892 return 0;
893}
894
895int
896xfs_buf_get_uncached(
897 struct xfs_buftarg *target,
898 size_t numblks,
899 int flags,
900 struct xfs_buf **bpp)
901{
902 int error;
903 struct xfs_buf *bp;
904 DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks);
905
906 *bpp = NULL;
907
908 /* flags might contain irrelevant bits, pass only what we care about */
909 error = _xfs_buf_alloc(target, &map, 1, flags & XBF_NO_IOACCT, &bp);
910 if (error)
911 return error;
912
913 error = xfs_buf_alloc_pages(bp, flags);
914 if (error)
915 goto fail_free_buf;
916
917 error = _xfs_buf_map_pages(bp, 0);
918 if (unlikely(error)) {
919 xfs_warn(target->bt_mount,
920 "%s: failed to map pages", __func__);
921 goto fail_free_buf;
922 }
923
924 trace_xfs_buf_get_uncached(bp, _RET_IP_);
925 *bpp = bp;
926 return 0;
927
928fail_free_buf:
929 xfs_buf_free(bp);
930 return error;
931}
932
933/*
934 * Increment reference count on buffer, to hold the buffer concurrently
935 * with another thread which may release (free) the buffer asynchronously.
936 * Must hold the buffer already to call this function.
937 */
938void
939xfs_buf_hold(
940 struct xfs_buf *bp)
941{
942 trace_xfs_buf_hold(bp, _RET_IP_);
943 atomic_inc(&bp->b_hold);
944}
945
946/*
947 * Release a hold on the specified buffer. If the hold count is 1, the buffer is
948 * placed on LRU or freed (depending on b_lru_ref).
949 */
950void
951xfs_buf_rele(
952 struct xfs_buf *bp)
953{
954 struct xfs_perag *pag = bp->b_pag;
955 bool release;
956 bool freebuf = false;
957
958 trace_xfs_buf_rele(bp, _RET_IP_);
959
960 if (!pag) {
961 ASSERT(list_empty(&bp->b_lru));
962 if (atomic_dec_and_test(&bp->b_hold)) {
963 xfs_buf_ioacct_dec(bp);
964 xfs_buf_free(bp);
965 }
966 return;
967 }
968
969 ASSERT(atomic_read(&bp->b_hold) > 0);
970
971 /*
972 * We grab the b_lock here first to serialise racing xfs_buf_rele()
973 * calls. The pag_buf_lock being taken on the last reference only
974 * serialises against racing lookups in xfs_buf_find(). IOWs, the second
975 * to last reference we drop here is not serialised against the last
976 * reference until we take bp->b_lock. Hence if we don't grab b_lock
977 * first, the last "release" reference can win the race to the lock and
978 * free the buffer before the second-to-last reference is processed,
979 * leading to a use-after-free scenario.
980 */
981 spin_lock(&bp->b_lock);
982 release = atomic_dec_and_lock(&bp->b_hold, &pag->pag_buf_lock);
983 if (!release) {
984 /*
985 * Drop the in-flight state if the buffer is already on the LRU
986 * and it holds the only reference. This is racy because we
987 * haven't acquired the pag lock, but the use of _XBF_IN_FLIGHT
988 * ensures the decrement occurs only once per-buf.
989 */
990 if ((atomic_read(&bp->b_hold) == 1) && !list_empty(&bp->b_lru))
991 __xfs_buf_ioacct_dec(bp);
992 goto out_unlock;
993 }
994
995 /* the last reference has been dropped ... */
996 __xfs_buf_ioacct_dec(bp);
997 if (!(bp->b_flags & XBF_STALE) && atomic_read(&bp->b_lru_ref)) {
998 /*
999 * If the buffer is added to the LRU take a new reference to the
1000 * buffer for the LRU and clear the (now stale) dispose list
1001 * state flag
1002 */
1003 if (list_lru_add(&bp->b_target->bt_lru, &bp->b_lru)) {
1004 bp->b_state &= ~XFS_BSTATE_DISPOSE;
1005 atomic_inc(&bp->b_hold);
1006 }
1007 spin_unlock(&pag->pag_buf_lock);
1008 } else {
1009 /*
1010 * most of the time buffers will already be removed from the
1011 * LRU, so optimise that case by checking for the
1012 * XFS_BSTATE_DISPOSE flag indicating the last list the buffer
1013 * was on was the disposal list
1014 */
1015 if (!(bp->b_state & XFS_BSTATE_DISPOSE)) {
1016 list_lru_del(&bp->b_target->bt_lru, &bp->b_lru);
1017 } else {
1018 ASSERT(list_empty(&bp->b_lru));
1019 }
1020
1021 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1022 rhashtable_remove_fast(&pag->pag_buf_hash, &bp->b_rhash_head,
1023 xfs_buf_hash_params);
1024 spin_unlock(&pag->pag_buf_lock);
1025 xfs_perag_put(pag);
1026 freebuf = true;
1027 }
1028
1029out_unlock:
1030 spin_unlock(&bp->b_lock);
1031
1032 if (freebuf)
1033 xfs_buf_free(bp);
1034}
1035
1036
1037/*
1038 * Lock a buffer object, if it is not already locked.
1039 *
1040 * If we come across a stale, pinned, locked buffer, we know that we are
1041 * being asked to lock a buffer that has been reallocated. Because it is
1042 * pinned, we know that the log has not been pushed to disk and hence it
1043 * will still be locked. Rather than continuing to have trylock attempts
1044 * fail until someone else pushes the log, push it ourselves before
1045 * returning. This means that the xfsaild will not get stuck trying
1046 * to push on stale inode buffers.
1047 */
1048int
1049xfs_buf_trylock(
1050 struct xfs_buf *bp)
1051{
1052 int locked;
1053
1054 locked = down_trylock(&bp->b_sema) == 0;
1055 if (locked)
1056 trace_xfs_buf_trylock(bp, _RET_IP_);
1057 else
1058 trace_xfs_buf_trylock_fail(bp, _RET_IP_);
1059 return locked;
1060}
1061
1062/*
1063 * Lock a buffer object.
1064 *
1065 * If we come across a stale, pinned, locked buffer, we know that we
1066 * are being asked to lock a buffer that has been reallocated. Because
1067 * it is pinned, we know that the log has not been pushed to disk and
1068 * hence it will still be locked. Rather than sleeping until someone
1069 * else pushes the log, push it ourselves before trying to get the lock.
1070 */
1071void
1072xfs_buf_lock(
1073 struct xfs_buf *bp)
1074{
1075 trace_xfs_buf_lock(bp, _RET_IP_);
1076
1077 if (atomic_read(&bp->b_pin_count) && (bp->b_flags & XBF_STALE))
1078 xfs_log_force(bp->b_mount, 0);
1079 down(&bp->b_sema);
1080
1081 trace_xfs_buf_lock_done(bp, _RET_IP_);
1082}
1083
1084void
1085xfs_buf_unlock(
1086 struct xfs_buf *bp)
1087{
1088 ASSERT(xfs_buf_islocked(bp));
1089
1090 up(&bp->b_sema);
1091 trace_xfs_buf_unlock(bp, _RET_IP_);
1092}
1093
1094STATIC void
1095xfs_buf_wait_unpin(
1096 struct xfs_buf *bp)
1097{
1098 DECLARE_WAITQUEUE (wait, current);
1099
1100 if (atomic_read(&bp->b_pin_count) == 0)
1101 return;
1102
1103 add_wait_queue(&bp->b_waiters, &wait);
1104 for (;;) {
1105 set_current_state(TASK_UNINTERRUPTIBLE);
1106 if (atomic_read(&bp->b_pin_count) == 0)
1107 break;
1108 io_schedule();
1109 }
1110 remove_wait_queue(&bp->b_waiters, &wait);
1111 set_current_state(TASK_RUNNING);
1112}
1113
1114static void
1115xfs_buf_ioerror_alert_ratelimited(
1116 struct xfs_buf *bp)
1117{
1118 static unsigned long lasttime;
1119 static struct xfs_buftarg *lasttarg;
1120
1121 if (bp->b_target != lasttarg ||
1122 time_after(jiffies, (lasttime + 5*HZ))) {
1123 lasttime = jiffies;
1124 xfs_buf_ioerror_alert(bp, __this_address);
1125 }
1126 lasttarg = bp->b_target;
1127}
1128
1129/*
1130 * Account for this latest trip around the retry handler, and decide if
1131 * we've failed enough times to constitute a permanent failure.
1132 */
1133static bool
1134xfs_buf_ioerror_permanent(
1135 struct xfs_buf *bp,
1136 struct xfs_error_cfg *cfg)
1137{
1138 struct xfs_mount *mp = bp->b_mount;
1139
1140 if (cfg->max_retries != XFS_ERR_RETRY_FOREVER &&
1141 ++bp->b_retries > cfg->max_retries)
1142 return true;
1143 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1144 time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time))
1145 return true;
1146
1147 /* At unmount we may treat errors differently */
1148 if ((mp->m_flags & XFS_MOUNT_UNMOUNTING) && mp->m_fail_unmount)
1149 return true;
1150
1151 return false;
1152}
1153
1154/*
1155 * On a sync write or shutdown we just want to stale the buffer and let the
1156 * caller handle the error in bp->b_error appropriately.
1157 *
1158 * If the write was asynchronous then no one will be looking for the error. If
1159 * this is the first failure of this type, clear the error state and write the
1160 * buffer out again. This means we always retry an async write failure at least
1161 * once, but we also need to set the buffer up to behave correctly now for
1162 * repeated failures.
1163 *
1164 * If we get repeated async write failures, then we take action according to the
1165 * error configuration we have been set up to use.
1166 *
1167 * Returns true if this function took care of error handling and the caller must
1168 * not touch the buffer again. Return false if the caller should proceed with
1169 * normal I/O completion handling.
1170 */
1171static bool
1172xfs_buf_ioend_handle_error(
1173 struct xfs_buf *bp)
1174{
1175 struct xfs_mount *mp = bp->b_mount;
1176 struct xfs_error_cfg *cfg;
1177
1178 /*
1179 * If we've already decided to shutdown the filesystem because of I/O
1180 * errors, there's no point in giving this a retry.
1181 */
1182 if (XFS_FORCED_SHUTDOWN(mp))
1183 goto out_stale;
1184
1185 xfs_buf_ioerror_alert_ratelimited(bp);
1186
1187 /*
1188 * We're not going to bother about retrying this during recovery.
1189 * One strike!
1190 */
1191 if (bp->b_flags & _XBF_LOGRECOVERY) {
1192 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1193 return false;
1194 }
1195
1196 /*
1197 * Synchronous writes will have callers process the error.
1198 */
1199 if (!(bp->b_flags & XBF_ASYNC))
1200 goto out_stale;
1201
1202 trace_xfs_buf_iodone_async(bp, _RET_IP_);
1203
1204 cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error);
1205 if (bp->b_last_error != bp->b_error ||
1206 !(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL))) {
1207 bp->b_last_error = bp->b_error;
1208 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1209 !bp->b_first_retry_time)
1210 bp->b_first_retry_time = jiffies;
1211 goto resubmit;
1212 }
1213
1214 /*
1215 * Permanent error - we need to trigger a shutdown if we haven't already
1216 * to indicate that inconsistency will result from this action.
1217 */
1218 if (xfs_buf_ioerror_permanent(bp, cfg)) {
1219 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1220 goto out_stale;
1221 }
1222
1223 /* Still considered a transient error. Caller will schedule retries. */
1224 if (bp->b_flags & _XBF_INODES)
1225 xfs_buf_inode_io_fail(bp);
1226 else if (bp->b_flags & _XBF_DQUOTS)
1227 xfs_buf_dquot_io_fail(bp);
1228 else
1229 ASSERT(list_empty(&bp->b_li_list));
1230 xfs_buf_ioerror(bp, 0);
1231 xfs_buf_relse(bp);
1232 return true;
1233
1234resubmit:
1235 xfs_buf_ioerror(bp, 0);
1236 bp->b_flags |= (XBF_DONE | XBF_WRITE_FAIL);
1237 xfs_buf_submit(bp);
1238 return true;
1239out_stale:
1240 xfs_buf_stale(bp);
1241 bp->b_flags |= XBF_DONE;
1242 bp->b_flags &= ~XBF_WRITE;
1243 trace_xfs_buf_error_relse(bp, _RET_IP_);
1244 return false;
1245}
1246
1247static void
1248xfs_buf_ioend(
1249 struct xfs_buf *bp)
1250{
1251 trace_xfs_buf_iodone(bp, _RET_IP_);
1252
1253 /*
1254 * Pull in IO completion errors now. We are guaranteed to be running
1255 * single threaded, so we don't need the lock to read b_io_error.
1256 */
1257 if (!bp->b_error && bp->b_io_error)
1258 xfs_buf_ioerror(bp, bp->b_io_error);
1259
1260 if (bp->b_flags & XBF_READ) {
1261 if (!bp->b_error && bp->b_ops)
1262 bp->b_ops->verify_read(bp);
1263 if (!bp->b_error)
1264 bp->b_flags |= XBF_DONE;
1265 } else {
1266 if (!bp->b_error) {
1267 bp->b_flags &= ~XBF_WRITE_FAIL;
1268 bp->b_flags |= XBF_DONE;
1269 }
1270
1271 if (unlikely(bp->b_error) && xfs_buf_ioend_handle_error(bp))
1272 return;
1273
1274 /* clear the retry state */
1275 bp->b_last_error = 0;
1276 bp->b_retries = 0;
1277 bp->b_first_retry_time = 0;
1278
1279 /*
1280 * Note that for things like remote attribute buffers, there may
1281 * not be a buffer log item here, so processing the buffer log
1282 * item must remain optional.
1283 */
1284 if (bp->b_log_item)
1285 xfs_buf_item_done(bp);
1286
1287 if (bp->b_flags & _XBF_INODES)
1288 xfs_buf_inode_iodone(bp);
1289 else if (bp->b_flags & _XBF_DQUOTS)
1290 xfs_buf_dquot_iodone(bp);
1291
1292 }
1293
1294 bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD |
1295 _XBF_LOGRECOVERY);
1296
1297 if (bp->b_flags & XBF_ASYNC)
1298 xfs_buf_relse(bp);
1299 else
1300 complete(&bp->b_iowait);
1301}
1302
1303static void
1304xfs_buf_ioend_work(
1305 struct work_struct *work)
1306{
1307 struct xfs_buf *bp =
1308 container_of(work, struct xfs_buf, b_ioend_work);
1309
1310 xfs_buf_ioend(bp);
1311}
1312
1313static void
1314xfs_buf_ioend_async(
1315 struct xfs_buf *bp)
1316{
1317 INIT_WORK(&bp->b_ioend_work, xfs_buf_ioend_work);
1318 queue_work(bp->b_mount->m_buf_workqueue, &bp->b_ioend_work);
1319}
1320
1321void
1322__xfs_buf_ioerror(
1323 struct xfs_buf *bp,
1324 int error,
1325 xfs_failaddr_t failaddr)
1326{
1327 ASSERT(error <= 0 && error >= -1000);
1328 bp->b_error = error;
1329 trace_xfs_buf_ioerror(bp, error, failaddr);
1330}
1331
1332void
1333xfs_buf_ioerror_alert(
1334 struct xfs_buf *bp,
1335 xfs_failaddr_t func)
1336{
1337 xfs_buf_alert_ratelimited(bp, "XFS: metadata IO error",
1338 "metadata I/O error in \"%pS\" at daddr 0x%llx len %d error %d",
1339 func, (uint64_t)XFS_BUF_ADDR(bp),
1340 bp->b_length, -bp->b_error);
1341}
1342
1343/*
1344 * To simulate an I/O failure, the buffer must be locked and held with at least
1345 * three references. The LRU reference is dropped by the stale call. The buf
1346 * item reference is dropped via ioend processing. The third reference is owned
1347 * by the caller and is dropped on I/O completion if the buffer is XBF_ASYNC.
1348 */
1349void
1350xfs_buf_ioend_fail(
1351 struct xfs_buf *bp)
1352{
1353 bp->b_flags &= ~XBF_DONE;
1354 xfs_buf_stale(bp);
1355 xfs_buf_ioerror(bp, -EIO);
1356 xfs_buf_ioend(bp);
1357}
1358
1359int
1360xfs_bwrite(
1361 struct xfs_buf *bp)
1362{
1363 int error;
1364
1365 ASSERT(xfs_buf_islocked(bp));
1366
1367 bp->b_flags |= XBF_WRITE;
1368 bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
1369 XBF_DONE);
1370
1371 error = xfs_buf_submit(bp);
1372 if (error)
1373 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
1374 return error;
1375}
1376
1377static void
1378xfs_buf_bio_end_io(
1379 struct bio *bio)
1380{
1381 struct xfs_buf *bp = (struct xfs_buf *)bio->bi_private;
1382
1383 if (!bio->bi_status &&
1384 (bp->b_flags & XBF_WRITE) && (bp->b_flags & XBF_ASYNC) &&
1385 XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_IOERROR))
1386 bio->bi_status = BLK_STS_IOERR;
1387
1388 /*
1389 * don't overwrite existing errors - otherwise we can lose errors on
1390 * buffers that require multiple bios to complete.
1391 */
1392 if (bio->bi_status) {
1393 int error = blk_status_to_errno(bio->bi_status);
1394
1395 cmpxchg(&bp->b_io_error, 0, error);
1396 }
1397
1398 if (!bp->b_error && xfs_buf_is_vmapped(bp) && (bp->b_flags & XBF_READ))
1399 invalidate_kernel_vmap_range(bp->b_addr, xfs_buf_vmap_len(bp));
1400
1401 if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
1402 xfs_buf_ioend_async(bp);
1403 bio_put(bio);
1404}
1405
1406static void
1407xfs_buf_ioapply_map(
1408 struct xfs_buf *bp,
1409 int map,
1410 int *buf_offset,
1411 int *count,
1412 int op)
1413{
1414 int page_index;
1415 unsigned int total_nr_pages = bp->b_page_count;
1416 int nr_pages;
1417 struct bio *bio;
1418 sector_t sector = bp->b_maps[map].bm_bn;
1419 int size;
1420 int offset;
1421
1422 /* skip the pages in the buffer before the start offset */
1423 page_index = 0;
1424 offset = *buf_offset;
1425 while (offset >= PAGE_SIZE) {
1426 page_index++;
1427 offset -= PAGE_SIZE;
1428 }
1429
1430 /*
1431 * Limit the IO size to the length of the current vector, and update the
1432 * remaining IO count for the next time around.
1433 */
1434 size = min_t(int, BBTOB(bp->b_maps[map].bm_len), *count);
1435 *count -= size;
1436 *buf_offset += size;
1437
1438next_chunk:
1439 atomic_inc(&bp->b_io_remaining);
1440 nr_pages = bio_max_segs(total_nr_pages);
1441
1442 bio = bio_alloc(GFP_NOIO, nr_pages);
1443 bio_set_dev(bio, bp->b_target->bt_bdev);
1444 bio->bi_iter.bi_sector = sector;
1445 bio->bi_end_io = xfs_buf_bio_end_io;
1446 bio->bi_private = bp;
1447 bio->bi_opf = op;
1448
1449 for (; size && nr_pages; nr_pages--, page_index++) {
1450 int rbytes, nbytes = PAGE_SIZE - offset;
1451
1452 if (nbytes > size)
1453 nbytes = size;
1454
1455 rbytes = bio_add_page(bio, bp->b_pages[page_index], nbytes,
1456 offset);
1457 if (rbytes < nbytes)
1458 break;
1459
1460 offset = 0;
1461 sector += BTOBB(nbytes);
1462 size -= nbytes;
1463 total_nr_pages--;
1464 }
1465
1466 if (likely(bio->bi_iter.bi_size)) {
1467 if (xfs_buf_is_vmapped(bp)) {
1468 flush_kernel_vmap_range(bp->b_addr,
1469 xfs_buf_vmap_len(bp));
1470 }
1471 submit_bio(bio);
1472 if (size)
1473 goto next_chunk;
1474 } else {
1475 /*
1476 * This is guaranteed not to be the last io reference count
1477 * because the caller (xfs_buf_submit) holds a count itself.
1478 */
1479 atomic_dec(&bp->b_io_remaining);
1480 xfs_buf_ioerror(bp, -EIO);
1481 bio_put(bio);
1482 }
1483
1484}
1485
1486STATIC void
1487_xfs_buf_ioapply(
1488 struct xfs_buf *bp)
1489{
1490 struct blk_plug plug;
1491 int op;
1492 int offset;
1493 int size;
1494 int i;
1495
1496 /*
1497 * Make sure we capture only current IO errors rather than stale errors
1498 * left over from previous use of the buffer (e.g. failed readahead).
1499 */
1500 bp->b_error = 0;
1501
1502 if (bp->b_flags & XBF_WRITE) {
1503 op = REQ_OP_WRITE;
1504
1505 /*
1506 * Run the write verifier callback function if it exists. If
1507 * this function fails it will mark the buffer with an error and
1508 * the IO should not be dispatched.
1509 */
1510 if (bp->b_ops) {
1511 bp->b_ops->verify_write(bp);
1512 if (bp->b_error) {
1513 xfs_force_shutdown(bp->b_mount,
1514 SHUTDOWN_CORRUPT_INCORE);
1515 return;
1516 }
1517 } else if (bp->b_bn != XFS_BUF_DADDR_NULL) {
1518 struct xfs_mount *mp = bp->b_mount;
1519
1520 /*
1521 * non-crc filesystems don't attach verifiers during
1522 * log recovery, so don't warn for such filesystems.
1523 */
1524 if (xfs_sb_version_hascrc(&mp->m_sb)) {
1525 xfs_warn(mp,
1526 "%s: no buf ops on daddr 0x%llx len %d",
1527 __func__, bp->b_bn, bp->b_length);
1528 xfs_hex_dump(bp->b_addr,
1529 XFS_CORRUPTION_DUMP_LEN);
1530 dump_stack();
1531 }
1532 }
1533 } else {
1534 op = REQ_OP_READ;
1535 if (bp->b_flags & XBF_READ_AHEAD)
1536 op |= REQ_RAHEAD;
1537 }
1538
1539 /* we only use the buffer cache for meta-data */
1540 op |= REQ_META;
1541
1542 /*
1543 * Walk all the vectors issuing IO on them. Set up the initial offset
1544 * into the buffer and the desired IO size before we start -
1545 * _xfs_buf_ioapply_vec() will modify them appropriately for each
1546 * subsequent call.
1547 */
1548 offset = bp->b_offset;
1549 size = BBTOB(bp->b_length);
1550 blk_start_plug(&plug);
1551 for (i = 0; i < bp->b_map_count; i++) {
1552 xfs_buf_ioapply_map(bp, i, &offset, &size, op);
1553 if (bp->b_error)
1554 break;
1555 if (size <= 0)
1556 break; /* all done */
1557 }
1558 blk_finish_plug(&plug);
1559}
1560
1561/*
1562 * Wait for I/O completion of a sync buffer and return the I/O error code.
1563 */
1564static int
1565xfs_buf_iowait(
1566 struct xfs_buf *bp)
1567{
1568 ASSERT(!(bp->b_flags & XBF_ASYNC));
1569
1570 trace_xfs_buf_iowait(bp, _RET_IP_);
1571 wait_for_completion(&bp->b_iowait);
1572 trace_xfs_buf_iowait_done(bp, _RET_IP_);
1573
1574 return bp->b_error;
1575}
1576
1577/*
1578 * Buffer I/O submission path, read or write. Asynchronous submission transfers
1579 * the buffer lock ownership and the current reference to the IO. It is not
1580 * safe to reference the buffer after a call to this function unless the caller
1581 * holds an additional reference itself.
1582 */
1583static int
1584__xfs_buf_submit(
1585 struct xfs_buf *bp,
1586 bool wait)
1587{
1588 int error = 0;
1589
1590 trace_xfs_buf_submit(bp, _RET_IP_);
1591
1592 ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
1593
1594 /* on shutdown we stale and complete the buffer immediately */
1595 if (XFS_FORCED_SHUTDOWN(bp->b_mount)) {
1596 xfs_buf_ioend_fail(bp);
1597 return -EIO;
1598 }
1599
1600 /*
1601 * Grab a reference so the buffer does not go away underneath us. For
1602 * async buffers, I/O completion drops the callers reference, which
1603 * could occur before submission returns.
1604 */
1605 xfs_buf_hold(bp);
1606
1607 if (bp->b_flags & XBF_WRITE)
1608 xfs_buf_wait_unpin(bp);
1609
1610 /* clear the internal error state to avoid spurious errors */
1611 bp->b_io_error = 0;
1612
1613 /*
1614 * Set the count to 1 initially, this will stop an I/O completion
1615 * callout which happens before we have started all the I/O from calling
1616 * xfs_buf_ioend too early.
1617 */
1618 atomic_set(&bp->b_io_remaining, 1);
1619 if (bp->b_flags & XBF_ASYNC)
1620 xfs_buf_ioacct_inc(bp);
1621 _xfs_buf_ioapply(bp);
1622
1623 /*
1624 * If _xfs_buf_ioapply failed, we can get back here with only the IO
1625 * reference we took above. If we drop it to zero, run completion so
1626 * that we don't return to the caller with completion still pending.
1627 */
1628 if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
1629 if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
1630 xfs_buf_ioend(bp);
1631 else
1632 xfs_buf_ioend_async(bp);
1633 }
1634
1635 if (wait)
1636 error = xfs_buf_iowait(bp);
1637
1638 /*
1639 * Release the hold that keeps the buffer referenced for the entire
1640 * I/O. Note that if the buffer is async, it is not safe to reference
1641 * after this release.
1642 */
1643 xfs_buf_rele(bp);
1644 return error;
1645}
1646
1647void *
1648xfs_buf_offset(
1649 struct xfs_buf *bp,
1650 size_t offset)
1651{
1652 struct page *page;
1653
1654 if (bp->b_addr)
1655 return bp->b_addr + offset;
1656
1657 page = bp->b_pages[offset >> PAGE_SHIFT];
1658 return page_address(page) + (offset & (PAGE_SIZE-1));
1659}
1660
1661void
1662xfs_buf_zero(
1663 struct xfs_buf *bp,
1664 size_t boff,
1665 size_t bsize)
1666{
1667 size_t bend;
1668
1669 bend = boff + bsize;
1670 while (boff < bend) {
1671 struct page *page;
1672 int page_index, page_offset, csize;
1673
1674 page_index = (boff + bp->b_offset) >> PAGE_SHIFT;
1675 page_offset = (boff + bp->b_offset) & ~PAGE_MASK;
1676 page = bp->b_pages[page_index];
1677 csize = min_t(size_t, PAGE_SIZE - page_offset,
1678 BBTOB(bp->b_length) - boff);
1679
1680 ASSERT((csize + page_offset) <= PAGE_SIZE);
1681
1682 memset(page_address(page) + page_offset, 0, csize);
1683
1684 boff += csize;
1685 }
1686}
1687
1688/*
1689 * Log a message about and stale a buffer that a caller has decided is corrupt.
1690 *
1691 * This function should be called for the kinds of metadata corruption that
1692 * cannot be detect from a verifier, such as incorrect inter-block relationship
1693 * data. Do /not/ call this function from a verifier function.
1694 *
1695 * The buffer must be XBF_DONE prior to the call. Afterwards, the buffer will
1696 * be marked stale, but b_error will not be set. The caller is responsible for
1697 * releasing the buffer or fixing it.
1698 */
1699void
1700__xfs_buf_mark_corrupt(
1701 struct xfs_buf *bp,
1702 xfs_failaddr_t fa)
1703{
1704 ASSERT(bp->b_flags & XBF_DONE);
1705
1706 xfs_buf_corruption_error(bp, fa);
1707 xfs_buf_stale(bp);
1708}
1709
1710/*
1711 * Handling of buffer targets (buftargs).
1712 */
1713
1714/*
1715 * Wait for any bufs with callbacks that have been submitted but have not yet
1716 * returned. These buffers will have an elevated hold count, so wait on those
1717 * while freeing all the buffers only held by the LRU.
1718 */
1719static enum lru_status
1720xfs_buftarg_drain_rele(
1721 struct list_head *item,
1722 struct list_lru_one *lru,
1723 spinlock_t *lru_lock,
1724 void *arg)
1725
1726{
1727 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1728 struct list_head *dispose = arg;
1729
1730 if (atomic_read(&bp->b_hold) > 1) {
1731 /* need to wait, so skip it this pass */
1732 trace_xfs_buf_drain_buftarg(bp, _RET_IP_);
1733 return LRU_SKIP;
1734 }
1735 if (!spin_trylock(&bp->b_lock))
1736 return LRU_SKIP;
1737
1738 /*
1739 * clear the LRU reference count so the buffer doesn't get
1740 * ignored in xfs_buf_rele().
1741 */
1742 atomic_set(&bp->b_lru_ref, 0);
1743 bp->b_state |= XFS_BSTATE_DISPOSE;
1744 list_lru_isolate_move(lru, item, dispose);
1745 spin_unlock(&bp->b_lock);
1746 return LRU_REMOVED;
1747}
1748
1749/*
1750 * Wait for outstanding I/O on the buftarg to complete.
1751 */
1752void
1753xfs_buftarg_wait(
1754 struct xfs_buftarg *btp)
1755{
1756 /*
1757 * First wait on the buftarg I/O count for all in-flight buffers to be
1758 * released. This is critical as new buffers do not make the LRU until
1759 * they are released.
1760 *
1761 * Next, flush the buffer workqueue to ensure all completion processing
1762 * has finished. Just waiting on buffer locks is not sufficient for
1763 * async IO as the reference count held over IO is not released until
1764 * after the buffer lock is dropped. Hence we need to ensure here that
1765 * all reference counts have been dropped before we start walking the
1766 * LRU list.
1767 */
1768 while (percpu_counter_sum(&btp->bt_io_count))
1769 delay(100);
1770 flush_workqueue(btp->bt_mount->m_buf_workqueue);
1771}
1772
1773void
1774xfs_buftarg_drain(
1775 struct xfs_buftarg *btp)
1776{
1777 LIST_HEAD(dispose);
1778 int loop = 0;
1779 bool write_fail = false;
1780
1781 xfs_buftarg_wait(btp);
1782
1783 /* loop until there is nothing left on the lru list. */
1784 while (list_lru_count(&btp->bt_lru)) {
1785 list_lru_walk(&btp->bt_lru, xfs_buftarg_drain_rele,
1786 &dispose, LONG_MAX);
1787
1788 while (!list_empty(&dispose)) {
1789 struct xfs_buf *bp;
1790 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1791 list_del_init(&bp->b_lru);
1792 if (bp->b_flags & XBF_WRITE_FAIL) {
1793 write_fail = true;
1794 xfs_buf_alert_ratelimited(bp,
1795 "XFS: Corruption Alert",
1796"Corruption Alert: Buffer at daddr 0x%llx had permanent write failures!",
1797 (long long)bp->b_bn);
1798 }
1799 xfs_buf_rele(bp);
1800 }
1801 if (loop++ != 0)
1802 delay(100);
1803 }
1804
1805 /*
1806 * If one or more failed buffers were freed, that means dirty metadata
1807 * was thrown away. This should only ever happen after I/O completion
1808 * handling has elevated I/O error(s) to permanent failures and shuts
1809 * down the fs.
1810 */
1811 if (write_fail) {
1812 ASSERT(XFS_FORCED_SHUTDOWN(btp->bt_mount));
1813 xfs_alert(btp->bt_mount,
1814 "Please run xfs_repair to determine the extent of the problem.");
1815 }
1816}
1817
1818static enum lru_status
1819xfs_buftarg_isolate(
1820 struct list_head *item,
1821 struct list_lru_one *lru,
1822 spinlock_t *lru_lock,
1823 void *arg)
1824{
1825 struct xfs_buf *bp = container_of(item, struct xfs_buf, b_lru);
1826 struct list_head *dispose = arg;
1827
1828 /*
1829 * we are inverting the lru lock/bp->b_lock here, so use a trylock.
1830 * If we fail to get the lock, just skip it.
1831 */
1832 if (!spin_trylock(&bp->b_lock))
1833 return LRU_SKIP;
1834 /*
1835 * Decrement the b_lru_ref count unless the value is already
1836 * zero. If the value is already zero, we need to reclaim the
1837 * buffer, otherwise it gets another trip through the LRU.
1838 */
1839 if (atomic_add_unless(&bp->b_lru_ref, -1, 0)) {
1840 spin_unlock(&bp->b_lock);
1841 return LRU_ROTATE;
1842 }
1843
1844 bp->b_state |= XFS_BSTATE_DISPOSE;
1845 list_lru_isolate_move(lru, item, dispose);
1846 spin_unlock(&bp->b_lock);
1847 return LRU_REMOVED;
1848}
1849
1850static unsigned long
1851xfs_buftarg_shrink_scan(
1852 struct shrinker *shrink,
1853 struct shrink_control *sc)
1854{
1855 struct xfs_buftarg *btp = container_of(shrink,
1856 struct xfs_buftarg, bt_shrinker);
1857 LIST_HEAD(dispose);
1858 unsigned long freed;
1859
1860 freed = list_lru_shrink_walk(&btp->bt_lru, sc,
1861 xfs_buftarg_isolate, &dispose);
1862
1863 while (!list_empty(&dispose)) {
1864 struct xfs_buf *bp;
1865 bp = list_first_entry(&dispose, struct xfs_buf, b_lru);
1866 list_del_init(&bp->b_lru);
1867 xfs_buf_rele(bp);
1868 }
1869
1870 return freed;
1871}
1872
1873static unsigned long
1874xfs_buftarg_shrink_count(
1875 struct shrinker *shrink,
1876 struct shrink_control *sc)
1877{
1878 struct xfs_buftarg *btp = container_of(shrink,
1879 struct xfs_buftarg, bt_shrinker);
1880 return list_lru_shrink_count(&btp->bt_lru, sc);
1881}
1882
1883void
1884xfs_free_buftarg(
1885 struct xfs_buftarg *btp)
1886{
1887 unregister_shrinker(&btp->bt_shrinker);
1888 ASSERT(percpu_counter_sum(&btp->bt_io_count) == 0);
1889 percpu_counter_destroy(&btp->bt_io_count);
1890 list_lru_destroy(&btp->bt_lru);
1891
1892 blkdev_issue_flush(btp->bt_bdev);
1893
1894 kmem_free(btp);
1895}
1896
1897int
1898xfs_setsize_buftarg(
1899 xfs_buftarg_t *btp,
1900 unsigned int sectorsize)
1901{
1902 /* Set up metadata sector size info */
1903 btp->bt_meta_sectorsize = sectorsize;
1904 btp->bt_meta_sectormask = sectorsize - 1;
1905
1906 if (set_blocksize(btp->bt_bdev, sectorsize)) {
1907 xfs_warn(btp->bt_mount,
1908 "Cannot set_blocksize to %u on device %pg",
1909 sectorsize, btp->bt_bdev);
1910 return -EINVAL;
1911 }
1912
1913 /* Set up device logical sector size mask */
1914 btp->bt_logical_sectorsize = bdev_logical_block_size(btp->bt_bdev);
1915 btp->bt_logical_sectormask = bdev_logical_block_size(btp->bt_bdev) - 1;
1916
1917 return 0;
1918}
1919
1920/*
1921 * When allocating the initial buffer target we have not yet
1922 * read in the superblock, so don't know what sized sectors
1923 * are being used at this early stage. Play safe.
1924 */
1925STATIC int
1926xfs_setsize_buftarg_early(
1927 xfs_buftarg_t *btp,
1928 struct block_device *bdev)
1929{
1930 return xfs_setsize_buftarg(btp, bdev_logical_block_size(bdev));
1931}
1932
1933xfs_buftarg_t *
1934xfs_alloc_buftarg(
1935 struct xfs_mount *mp,
1936 struct block_device *bdev,
1937 struct dax_device *dax_dev)
1938{
1939 xfs_buftarg_t *btp;
1940
1941 btp = kmem_zalloc(sizeof(*btp), KM_NOFS);
1942
1943 btp->bt_mount = mp;
1944 btp->bt_dev = bdev->bd_dev;
1945 btp->bt_bdev = bdev;
1946 btp->bt_daxdev = dax_dev;
1947
1948 /*
1949 * Buffer IO error rate limiting. Limit it to no more than 10 messages
1950 * per 30 seconds so as to not spam logs too much on repeated errors.
1951 */
1952 ratelimit_state_init(&btp->bt_ioerror_rl, 30 * HZ,
1953 DEFAULT_RATELIMIT_BURST);
1954
1955 if (xfs_setsize_buftarg_early(btp, bdev))
1956 goto error_free;
1957
1958 if (list_lru_init(&btp->bt_lru))
1959 goto error_free;
1960
1961 if (percpu_counter_init(&btp->bt_io_count, 0, GFP_KERNEL))
1962 goto error_lru;
1963
1964 btp->bt_shrinker.count_objects = xfs_buftarg_shrink_count;
1965 btp->bt_shrinker.scan_objects = xfs_buftarg_shrink_scan;
1966 btp->bt_shrinker.seeks = DEFAULT_SEEKS;
1967 btp->bt_shrinker.flags = SHRINKER_NUMA_AWARE;
1968 if (register_shrinker(&btp->bt_shrinker))
1969 goto error_pcpu;
1970 return btp;
1971
1972error_pcpu:
1973 percpu_counter_destroy(&btp->bt_io_count);
1974error_lru:
1975 list_lru_destroy(&btp->bt_lru);
1976error_free:
1977 kmem_free(btp);
1978 return NULL;
1979}
1980
1981/*
1982 * Cancel a delayed write list.
1983 *
1984 * Remove each buffer from the list, clear the delwri queue flag and drop the
1985 * associated buffer reference.
1986 */
1987void
1988xfs_buf_delwri_cancel(
1989 struct list_head *list)
1990{
1991 struct xfs_buf *bp;
1992
1993 while (!list_empty(list)) {
1994 bp = list_first_entry(list, struct xfs_buf, b_list);
1995
1996 xfs_buf_lock(bp);
1997 bp->b_flags &= ~_XBF_DELWRI_Q;
1998 list_del_init(&bp->b_list);
1999 xfs_buf_relse(bp);
2000 }
2001}
2002
2003/*
2004 * Add a buffer to the delayed write list.
2005 *
2006 * This queues a buffer for writeout if it hasn't already been. Note that
2007 * neither this routine nor the buffer list submission functions perform
2008 * any internal synchronization. It is expected that the lists are thread-local
2009 * to the callers.
2010 *
2011 * Returns true if we queued up the buffer, or false if it already had
2012 * been on the buffer list.
2013 */
2014bool
2015xfs_buf_delwri_queue(
2016 struct xfs_buf *bp,
2017 struct list_head *list)
2018{
2019 ASSERT(xfs_buf_islocked(bp));
2020 ASSERT(!(bp->b_flags & XBF_READ));
2021
2022 /*
2023 * If the buffer is already marked delwri it already is queued up
2024 * by someone else for imediate writeout. Just ignore it in that
2025 * case.
2026 */
2027 if (bp->b_flags & _XBF_DELWRI_Q) {
2028 trace_xfs_buf_delwri_queued(bp, _RET_IP_);
2029 return false;
2030 }
2031
2032 trace_xfs_buf_delwri_queue(bp, _RET_IP_);
2033
2034 /*
2035 * If a buffer gets written out synchronously or marked stale while it
2036 * is on a delwri list we lazily remove it. To do this, the other party
2037 * clears the _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
2038 * It remains referenced and on the list. In a rare corner case it
2039 * might get readded to a delwri list after the synchronous writeout, in
2040 * which case we need just need to re-add the flag here.
2041 */
2042 bp->b_flags |= _XBF_DELWRI_Q;
2043 if (list_empty(&bp->b_list)) {
2044 atomic_inc(&bp->b_hold);
2045 list_add_tail(&bp->b_list, list);
2046 }
2047
2048 return true;
2049}
2050
2051/*
2052 * Compare function is more complex than it needs to be because
2053 * the return value is only 32 bits and we are doing comparisons
2054 * on 64 bit values
2055 */
2056static int
2057xfs_buf_cmp(
2058 void *priv,
2059 const struct list_head *a,
2060 const struct list_head *b)
2061{
2062 struct xfs_buf *ap = container_of(a, struct xfs_buf, b_list);
2063 struct xfs_buf *bp = container_of(b, struct xfs_buf, b_list);
2064 xfs_daddr_t diff;
2065
2066 diff = ap->b_maps[0].bm_bn - bp->b_maps[0].bm_bn;
2067 if (diff < 0)
2068 return -1;
2069 if (diff > 0)
2070 return 1;
2071 return 0;
2072}
2073
2074/*
2075 * Submit buffers for write. If wait_list is specified, the buffers are
2076 * submitted using sync I/O and placed on the wait list such that the caller can
2077 * iowait each buffer. Otherwise async I/O is used and the buffers are released
2078 * at I/O completion time. In either case, buffers remain locked until I/O
2079 * completes and the buffer is released from the queue.
2080 */
2081static int
2082xfs_buf_delwri_submit_buffers(
2083 struct list_head *buffer_list,
2084 struct list_head *wait_list)
2085{
2086 struct xfs_buf *bp, *n;
2087 int pinned = 0;
2088 struct blk_plug plug;
2089
2090 list_sort(NULL, buffer_list, xfs_buf_cmp);
2091
2092 blk_start_plug(&plug);
2093 list_for_each_entry_safe(bp, n, buffer_list, b_list) {
2094 if (!wait_list) {
2095 if (xfs_buf_ispinned(bp)) {
2096 pinned++;
2097 continue;
2098 }
2099 if (!xfs_buf_trylock(bp))
2100 continue;
2101 } else {
2102 xfs_buf_lock(bp);
2103 }
2104
2105 /*
2106 * Someone else might have written the buffer synchronously or
2107 * marked it stale in the meantime. In that case only the
2108 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
2109 * reference and remove it from the list here.
2110 */
2111 if (!(bp->b_flags & _XBF_DELWRI_Q)) {
2112 list_del_init(&bp->b_list);
2113 xfs_buf_relse(bp);
2114 continue;
2115 }
2116
2117 trace_xfs_buf_delwri_split(bp, _RET_IP_);
2118
2119 /*
2120 * If we have a wait list, each buffer (and associated delwri
2121 * queue reference) transfers to it and is submitted
2122 * synchronously. Otherwise, drop the buffer from the delwri
2123 * queue and submit async.
2124 */
2125 bp->b_flags &= ~_XBF_DELWRI_Q;
2126 bp->b_flags |= XBF_WRITE;
2127 if (wait_list) {
2128 bp->b_flags &= ~XBF_ASYNC;
2129 list_move_tail(&bp->b_list, wait_list);
2130 } else {
2131 bp->b_flags |= XBF_ASYNC;
2132 list_del_init(&bp->b_list);
2133 }
2134 __xfs_buf_submit(bp, false);
2135 }
2136 blk_finish_plug(&plug);
2137
2138 return pinned;
2139}
2140
2141/*
2142 * Write out a buffer list asynchronously.
2143 *
2144 * This will take the @buffer_list, write all non-locked and non-pinned buffers
2145 * out and not wait for I/O completion on any of the buffers. This interface
2146 * is only safely useable for callers that can track I/O completion by higher
2147 * level means, e.g. AIL pushing as the @buffer_list is consumed in this
2148 * function.
2149 *
2150 * Note: this function will skip buffers it would block on, and in doing so
2151 * leaves them on @buffer_list so they can be retried on a later pass. As such,
2152 * it is up to the caller to ensure that the buffer list is fully submitted or
2153 * cancelled appropriately when they are finished with the list. Failure to
2154 * cancel or resubmit the list until it is empty will result in leaked buffers
2155 * at unmount time.
2156 */
2157int
2158xfs_buf_delwri_submit_nowait(
2159 struct list_head *buffer_list)
2160{
2161 return xfs_buf_delwri_submit_buffers(buffer_list, NULL);
2162}
2163
2164/*
2165 * Write out a buffer list synchronously.
2166 *
2167 * This will take the @buffer_list, write all buffers out and wait for I/O
2168 * completion on all of the buffers. @buffer_list is consumed by the function,
2169 * so callers must have some other way of tracking buffers if they require such
2170 * functionality.
2171 */
2172int
2173xfs_buf_delwri_submit(
2174 struct list_head *buffer_list)
2175{
2176 LIST_HEAD (wait_list);
2177 int error = 0, error2;
2178 struct xfs_buf *bp;
2179
2180 xfs_buf_delwri_submit_buffers(buffer_list, &wait_list);
2181
2182 /* Wait for IO to complete. */
2183 while (!list_empty(&wait_list)) {
2184 bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
2185
2186 list_del_init(&bp->b_list);
2187
2188 /*
2189 * Wait on the locked buffer, check for errors and unlock and
2190 * release the delwri queue reference.
2191 */
2192 error2 = xfs_buf_iowait(bp);
2193 xfs_buf_relse(bp);
2194 if (!error)
2195 error = error2;
2196 }
2197
2198 return error;
2199}
2200
2201/*
2202 * Push a single buffer on a delwri queue.
2203 *
2204 * The purpose of this function is to submit a single buffer of a delwri queue
2205 * and return with the buffer still on the original queue. The waiting delwri
2206 * buffer submission infrastructure guarantees transfer of the delwri queue
2207 * buffer reference to a temporary wait list. We reuse this infrastructure to
2208 * transfer the buffer back to the original queue.
2209 *
2210 * Note the buffer transitions from the queued state, to the submitted and wait
2211 * listed state and back to the queued state during this call. The buffer
2212 * locking and queue management logic between _delwri_pushbuf() and
2213 * _delwri_queue() guarantee that the buffer cannot be queued to another list
2214 * before returning.
2215 */
2216int
2217xfs_buf_delwri_pushbuf(
2218 struct xfs_buf *bp,
2219 struct list_head *buffer_list)
2220{
2221 LIST_HEAD (submit_list);
2222 int error;
2223
2224 ASSERT(bp->b_flags & _XBF_DELWRI_Q);
2225
2226 trace_xfs_buf_delwri_pushbuf(bp, _RET_IP_);
2227
2228 /*
2229 * Isolate the buffer to a new local list so we can submit it for I/O
2230 * independently from the rest of the original list.
2231 */
2232 xfs_buf_lock(bp);
2233 list_move(&bp->b_list, &submit_list);
2234 xfs_buf_unlock(bp);
2235
2236 /*
2237 * Delwri submission clears the DELWRI_Q buffer flag and returns with
2238 * the buffer on the wait list with the original reference. Rather than
2239 * bounce the buffer from a local wait list back to the original list
2240 * after I/O completion, reuse the original list as the wait list.
2241 */
2242 xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
2243
2244 /*
2245 * The buffer is now locked, under I/O and wait listed on the original
2246 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
2247 * return with the buffer unlocked and on the original queue.
2248 */
2249 error = xfs_buf_iowait(bp);
2250 bp->b_flags |= _XBF_DELWRI_Q;
2251 xfs_buf_unlock(bp);
2252
2253 return error;
2254}
2255
2256int __init
2257xfs_buf_init(void)
2258{
2259 xfs_buf_zone = kmem_cache_create("xfs_buf", sizeof(struct xfs_buf), 0,
2260 SLAB_HWCACHE_ALIGN |
2261 SLAB_RECLAIM_ACCOUNT |
2262 SLAB_MEM_SPREAD,
2263 NULL);
2264 if (!xfs_buf_zone)
2265 goto out;
2266
2267 return 0;
2268
2269 out:
2270 return -ENOMEM;
2271}
2272
2273void
2274xfs_buf_terminate(void)
2275{
2276 kmem_cache_destroy(xfs_buf_zone);
2277}
2278
2279void xfs_buf_set_ref(struct xfs_buf *bp, int lru_ref)
2280{
2281 /*
2282 * Set the lru reference count to 0 based on the error injection tag.
2283 * This allows userspace to disrupt buffer caching for debug/testing
2284 * purposes.
2285 */
2286 if (XFS_TEST_ERROR(false, bp->b_mount, XFS_ERRTAG_BUF_LRU_REF))
2287 lru_ref = 0;
2288
2289 atomic_set(&bp->b_lru_ref, lru_ref);
2290}
2291
2292/*
2293 * Verify an on-disk magic value against the magic value specified in the
2294 * verifier structure. The verifier magic is in disk byte order so the caller is
2295 * expected to pass the value directly from disk.
2296 */
2297bool
2298xfs_verify_magic(
2299 struct xfs_buf *bp,
2300 __be32 dmagic)
2301{
2302 struct xfs_mount *mp = bp->b_mount;
2303 int idx;
2304
2305 idx = xfs_sb_version_hascrc(&mp->m_sb);
2306 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic[idx]))
2307 return false;
2308 return dmagic == bp->b_ops->magic[idx];
2309}
2310/*
2311 * Verify an on-disk magic value against the magic value specified in the
2312 * verifier structure. The verifier magic is in disk byte order so the caller is
2313 * expected to pass the value directly from disk.
2314 */
2315bool
2316xfs_verify_magic16(
2317 struct xfs_buf *bp,
2318 __be16 dmagic)
2319{
2320 struct xfs_mount *mp = bp->b_mount;
2321 int idx;
2322
2323 idx = xfs_sb_version_hascrc(&mp->m_sb);
2324 if (WARN_ON(!bp->b_ops || !bp->b_ops->magic16[idx]))
2325 return false;
2326 return dmagic == bp->b_ops->magic16[idx];
2327}