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