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