1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_bit.h"
13 #include "xfs_mount.h"
14 #include "xfs_trans.h"
15 #include "xfs_buf_item.h"
16 #include "xfs_trans_priv.h"
17 #include "xfs_trace.h"
18 #include "xfs_log.h"
19
20
21 kmem_zone_t *xfs_buf_item_zone;
22
23 static inline struct xfs_buf_log_item *BUF_ITEM(struct xfs_log_item *lip)
24 {
25 return container_of(lip, struct xfs_buf_log_item, bli_item);
26 }
27
28 STATIC void xfs_buf_do_callbacks(struct xfs_buf *bp);
29
30 static inline int
31 xfs_buf_log_format_size(
32 struct xfs_buf_log_format *blfp)
33 {
34 return offsetof(struct xfs_buf_log_format, blf_data_map) +
35 (blfp->blf_map_size * sizeof(blfp->blf_data_map[0]));
36 }
37
38 /*
39 * This returns the number of log iovecs needed to log the
40 * given buf log item.
41 *
42 * It calculates this as 1 iovec for the buf log format structure
43 * and 1 for each stretch of non-contiguous chunks to be logged.
44 * Contiguous chunks are logged in a single iovec.
45 *
46 * If the XFS_BLI_STALE flag has been set, then log nothing.
47 */
48 STATIC void
49 xfs_buf_item_size_segment(
50 struct xfs_buf_log_item *bip,
51 struct xfs_buf_log_format *blfp,
52 int *nvecs,
53 int *nbytes)
54 {
55 struct xfs_buf *bp = bip->bli_buf;
56 int next_bit;
57 int last_bit;
58
59 last_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size, 0);
60 if (last_bit == -1)
61 return;
62
63 /*
64 * initial count for a dirty buffer is 2 vectors - the format structure
65 * and the first dirty region.
66 */
67 *nvecs += 2;
68 *nbytes += xfs_buf_log_format_size(blfp) + XFS_BLF_CHUNK;
69
70 while (last_bit != -1) {
71 /*
72 * This takes the bit number to start looking from and
73 * returns the next set bit from there. It returns -1
74 * if there are no more bits set or the start bit is
75 * beyond the end of the bitmap.
76 */
77 next_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size,
78 last_bit + 1);
79 /*
80 * If we run out of bits, leave the loop,
81 * else if we find a new set of bits bump the number of vecs,
82 * else keep scanning the current set of bits.
83 */
84 if (next_bit == -1) {
85 break;
86 } else if (next_bit != last_bit + 1) {
87 last_bit = next_bit;
88 (*nvecs)++;
89 } else if (xfs_buf_offset(bp, next_bit * XFS_BLF_CHUNK) !=
90 (xfs_buf_offset(bp, last_bit * XFS_BLF_CHUNK) +
91 XFS_BLF_CHUNK)) {
92 last_bit = next_bit;
93 (*nvecs)++;
94 } else {
95 last_bit++;
96 }
97 *nbytes += XFS_BLF_CHUNK;
98 }
99 }
100
101 /*
102 * This returns the number of log iovecs needed to log the given buf log item.
103 *
104 * It calculates this as 1 iovec for the buf log format structure and 1 for each
105 * stretch of non-contiguous chunks to be logged. Contiguous chunks are logged
106 * in a single iovec.
107 *
108 * Discontiguous buffers need a format structure per region that that is being
109 * logged. This makes the changes in the buffer appear to log recovery as though
110 * they came from separate buffers, just like would occur if multiple buffers
111 * were used instead of a single discontiguous buffer. This enables
112 * discontiguous buffers to be in-memory constructs, completely transparent to
113 * what ends up on disk.
114 *
115 * If the XFS_BLI_STALE flag has been set, then log nothing but the buf log
116 * format structures.
117 */
118 STATIC void
119 xfs_buf_item_size(
120 struct xfs_log_item *lip,
121 int *nvecs,
122 int *nbytes)
123 {
124 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
125 int i;
126
127 ASSERT(atomic_read(&bip->bli_refcount) > 0);
128 if (bip->bli_flags & XFS_BLI_STALE) {
129 /*
130 * The buffer is stale, so all we need to log
131 * is the buf log format structure with the
132 * cancel flag in it.
133 */
134 trace_xfs_buf_item_size_stale(bip);
135 ASSERT(bip->__bli_format.blf_flags & XFS_BLF_CANCEL);
136 *nvecs += bip->bli_format_count;
137 for (i = 0; i < bip->bli_format_count; i++) {
138 *nbytes += xfs_buf_log_format_size(&bip->bli_formats[i]);
139 }
140 return;
141 }
142
143 ASSERT(bip->bli_flags & XFS_BLI_LOGGED);
144
145 if (bip->bli_flags & XFS_BLI_ORDERED) {
146 /*
147 * The buffer has been logged just to order it.
148 * It is not being included in the transaction
149 * commit, so no vectors are used at all.
150 */
151 trace_xfs_buf_item_size_ordered(bip);
152 *nvecs = XFS_LOG_VEC_ORDERED;
153 return;
154 }
155
156 /*
157 * the vector count is based on the number of buffer vectors we have
158 * dirty bits in. This will only be greater than one when we have a
159 * compound buffer with more than one segment dirty. Hence for compound
160 * buffers we need to track which segment the dirty bits correspond to,
161 * and when we move from one segment to the next increment the vector
162 * count for the extra buf log format structure that will need to be
163 * written.
164 */
165 for (i = 0; i < bip->bli_format_count; i++) {
166 xfs_buf_item_size_segment(bip, &bip->bli_formats[i],
167 nvecs, nbytes);
168 }
169 trace_xfs_buf_item_size(bip);
170 }
171
172 static inline void
173 xfs_buf_item_copy_iovec(
174 struct xfs_log_vec *lv,
175 struct xfs_log_iovec **vecp,
176 struct xfs_buf *bp,
177 uint offset,
178 int first_bit,
179 uint nbits)
180 {
181 offset += first_bit * XFS_BLF_CHUNK;
182 xlog_copy_iovec(lv, vecp, XLOG_REG_TYPE_BCHUNK,
183 xfs_buf_offset(bp, offset),
184 nbits * XFS_BLF_CHUNK);
185 }
186
187 static inline bool
188 xfs_buf_item_straddle(
189 struct xfs_buf *bp,
190 uint offset,
191 int next_bit,
192 int last_bit)
193 {
194 return xfs_buf_offset(bp, offset + (next_bit << XFS_BLF_SHIFT)) !=
195 (xfs_buf_offset(bp, offset + (last_bit << XFS_BLF_SHIFT)) +
196 XFS_BLF_CHUNK);
197 }
198
199 static void
200 xfs_buf_item_format_segment(
201 struct xfs_buf_log_item *bip,
202 struct xfs_log_vec *lv,
203 struct xfs_log_iovec **vecp,
204 uint offset,
205 struct xfs_buf_log_format *blfp)
206 {
207 struct xfs_buf *bp = bip->bli_buf;
208 uint base_size;
209 int first_bit;
210 int last_bit;
211 int next_bit;
212 uint nbits;
213
214 /* copy the flags across from the base format item */
215 blfp->blf_flags = bip->__bli_format.blf_flags;
216
217 /*
218 * Base size is the actual size of the ondisk structure - it reflects
219 * the actual size of the dirty bitmap rather than the size of the in
220 * memory structure.
221 */
222 base_size = xfs_buf_log_format_size(blfp);
223
224 first_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size, 0);
225 if (!(bip->bli_flags & XFS_BLI_STALE) && first_bit == -1) {
226 /*
227 * If the map is not be dirty in the transaction, mark
228 * the size as zero and do not advance the vector pointer.
229 */
230 return;
231 }
232
233 blfp = xlog_copy_iovec(lv, vecp, XLOG_REG_TYPE_BFORMAT, blfp, base_size);
234 blfp->blf_size = 1;
235
236 if (bip->bli_flags & XFS_BLI_STALE) {
237 /*
238 * The buffer is stale, so all we need to log
239 * is the buf log format structure with the
240 * cancel flag in it.
241 */
242 trace_xfs_buf_item_format_stale(bip);
243 ASSERT(blfp->blf_flags & XFS_BLF_CANCEL);
244 return;
245 }
246
247
248 /*
249 * Fill in an iovec for each set of contiguous chunks.
250 */
251 last_bit = first_bit;
252 nbits = 1;
253 for (;;) {
254 /*
255 * This takes the bit number to start looking from and
256 * returns the next set bit from there. It returns -1
257 * if there are no more bits set or the start bit is
258 * beyond the end of the bitmap.
259 */
260 next_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size,
261 (uint)last_bit + 1);
262 /*
263 * If we run out of bits fill in the last iovec and get out of
264 * the loop. Else if we start a new set of bits then fill in
265 * the iovec for the series we were looking at and start
266 * counting the bits in the new one. Else we're still in the
267 * same set of bits so just keep counting and scanning.
268 */
269 if (next_bit == -1) {
270 xfs_buf_item_copy_iovec(lv, vecp, bp, offset,
271 first_bit, nbits);
272 blfp->blf_size++;
273 break;
274 } else if (next_bit != last_bit + 1 ||
275 xfs_buf_item_straddle(bp, offset, next_bit, last_bit)) {
276 xfs_buf_item_copy_iovec(lv, vecp, bp, offset,
277 first_bit, nbits);
278 blfp->blf_size++;
279 first_bit = next_bit;
280 last_bit = next_bit;
281 nbits = 1;
282 } else {
283 last_bit++;
284 nbits++;
285 }
286 }
287 }
288
289 /*
290 * This is called to fill in the vector of log iovecs for the
291 * given log buf item. It fills the first entry with a buf log
292 * format structure, and the rest point to contiguous chunks
293 * within the buffer.
294 */
295 STATIC void
296 xfs_buf_item_format(
297 struct xfs_log_item *lip,
298 struct xfs_log_vec *lv)
299 {
300 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
301 struct xfs_buf *bp = bip->bli_buf;
302 struct xfs_log_iovec *vecp = NULL;
303 uint offset = 0;
304 int i;
305
306 ASSERT(atomic_read(&bip->bli_refcount) > 0);
307 ASSERT((bip->bli_flags & XFS_BLI_LOGGED) ||
308 (bip->bli_flags & XFS_BLI_STALE));
309 ASSERT((bip->bli_flags & XFS_BLI_STALE) ||
310 (xfs_blft_from_flags(&bip->__bli_format) > XFS_BLFT_UNKNOWN_BUF
311 && xfs_blft_from_flags(&bip->__bli_format) < XFS_BLFT_MAX_BUF));
312 ASSERT(!(bip->bli_flags & XFS_BLI_ORDERED) ||
313 (bip->bli_flags & XFS_BLI_STALE));
314
315
316 /*
317 * If it is an inode buffer, transfer the in-memory state to the
318 * format flags and clear the in-memory state.
319 *
320 * For buffer based inode allocation, we do not transfer
321 * this state if the inode buffer allocation has not yet been committed
322 * to the log as setting the XFS_BLI_INODE_BUF flag will prevent
323 * correct replay of the inode allocation.
324 *
325 * For icreate item based inode allocation, the buffers aren't written
326 * to the journal during allocation, and hence we should always tag the
327 * buffer as an inode buffer so that the correct unlinked list replay
328 * occurs during recovery.
329 */
330 if (bip->bli_flags & XFS_BLI_INODE_BUF) {
331 if (xfs_sb_version_hascrc(&lip->li_mountp->m_sb) ||
332 !((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) &&
333 xfs_log_item_in_current_chkpt(lip)))
334 bip->__bli_format.blf_flags |= XFS_BLF_INODE_BUF;
335 bip->bli_flags &= ~XFS_BLI_INODE_BUF;
336 }
337
338 for (i = 0; i < bip->bli_format_count; i++) {
339 xfs_buf_item_format_segment(bip, lv, &vecp, offset,
340 &bip->bli_formats[i]);
341 offset += BBTOB(bp->b_maps[i].bm_len);
342 }
343
344 /*
345 * Check to make sure everything is consistent.
346 */
347 trace_xfs_buf_item_format(bip);
348 }
349
350 /*
351 * This is called to pin the buffer associated with the buf log item in memory
352 * so it cannot be written out.
353 *
354 * We also always take a reference to the buffer log item here so that the bli
355 * is held while the item is pinned in memory. This means that we can
356 * unconditionally drop the reference count a transaction holds when the
357 * transaction is completed.
358 */
359 STATIC void
360 xfs_buf_item_pin(
361 struct xfs_log_item *lip)
362 {
363 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
364
365 ASSERT(atomic_read(&bip->bli_refcount) > 0);
366 ASSERT((bip->bli_flags & XFS_BLI_LOGGED) ||
367 (bip->bli_flags & XFS_BLI_ORDERED) ||
368 (bip->bli_flags & XFS_BLI_STALE));
369
370 trace_xfs_buf_item_pin(bip);
371
372 atomic_inc(&bip->bli_refcount);
373 atomic_inc(&bip->bli_buf->b_pin_count);
374 }
375
376 /*
377 * This is called to unpin the buffer associated with the buf log
378 * item which was previously pinned with a call to xfs_buf_item_pin().
379 *
380 * Also drop the reference to the buf item for the current transaction.
381 * If the XFS_BLI_STALE flag is set and we are the last reference,
382 * then free up the buf log item and unlock the buffer.
383 *
384 * If the remove flag is set we are called from uncommit in the
385 * forced-shutdown path. If that is true and the reference count on
386 * the log item is going to drop to zero we need to free the item's
387 * descriptor in the transaction.
388 */
389 STATIC void
390 xfs_buf_item_unpin(
391 struct xfs_log_item *lip,
392 int remove)
393 {
394 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
395 xfs_buf_t *bp = bip->bli_buf;
396 struct xfs_ail *ailp = lip->li_ailp;
397 int stale = bip->bli_flags & XFS_BLI_STALE;
398 int freed;
399
400 ASSERT(bp->b_log_item == bip);
401 ASSERT(atomic_read(&bip->bli_refcount) > 0);
402
403 trace_xfs_buf_item_unpin(bip);
404
405 freed = atomic_dec_and_test(&bip->bli_refcount);
406
407 if (atomic_dec_and_test(&bp->b_pin_count))
408 wake_up_all(&bp->b_waiters);
409
410 if (freed && stale) {
411 ASSERT(bip->bli_flags & XFS_BLI_STALE);
412 ASSERT(xfs_buf_islocked(bp));
413 ASSERT(bp->b_flags & XBF_STALE);
414 ASSERT(bip->__bli_format.blf_flags & XFS_BLF_CANCEL);
415
416 trace_xfs_buf_item_unpin_stale(bip);
417
418 if (remove) {
419 /*
420 * If we are in a transaction context, we have to
421 * remove the log item from the transaction as we are
422 * about to release our reference to the buffer. If we
423 * don't, the unlock that occurs later in
424 * xfs_trans_uncommit() will try to reference the
425 * buffer which we no longer have a hold on.
426 */
427 if (!list_empty(&lip->li_trans))
428 xfs_trans_del_item(lip);
429
430 /*
431 * Since the transaction no longer refers to the buffer,
432 * the buffer should no longer refer to the transaction.
433 */
434 bp->b_transp = NULL;
435 }
436
437 /*
438 * If we get called here because of an IO error, we may
439 * or may not have the item on the AIL. xfs_trans_ail_delete()
440 * will take care of that situation.
441 * xfs_trans_ail_delete() drops the AIL lock.
442 */
443 if (bip->bli_flags & XFS_BLI_STALE_INODE) {
444 xfs_buf_do_callbacks(bp);
445 bp->b_log_item = NULL;
446 list_del_init(&bp->b_li_list);
447 bp->b_iodone = NULL;
448 } else {
449 spin_lock(&ailp->ail_lock);
450 xfs_trans_ail_delete(ailp, lip, SHUTDOWN_LOG_IO_ERROR);
451 xfs_buf_item_relse(bp);
452 ASSERT(bp->b_log_item == NULL);
453 }
454 xfs_buf_relse(bp);
455 } else if (freed && remove) {
456 /*
457 * There are currently two references to the buffer - the active
458 * LRU reference and the buf log item. What we are about to do
459 * here - simulate a failed IO completion - requires 3
460 * references.
461 *
462 * The LRU reference is removed by the xfs_buf_stale() call. The
463 * buf item reference is removed by the xfs_buf_iodone()
464 * callback that is run by xfs_buf_do_callbacks() during ioend
465 * processing (via the bp->b_iodone callback), and then finally
466 * the ioend processing will drop the IO reference if the buffer
467 * is marked XBF_ASYNC.
468 *
469 * Hence we need to take an additional reference here so that IO
470 * completion processing doesn't free the buffer prematurely.
471 */
472 xfs_buf_lock(bp);
473 xfs_buf_hold(bp);
474 bp->b_flags |= XBF_ASYNC;
475 xfs_buf_ioerror(bp, -EIO);
476 bp->b_flags &= ~XBF_DONE;
477 xfs_buf_stale(bp);
478 xfs_buf_ioend(bp);
479 }
480 }
481
482 /*
483 * Buffer IO error rate limiting. Limit it to no more than 10 messages per 30
484 * seconds so as to not spam logs too much on repeated detection of the same
485 * buffer being bad..
486 */
487
488 static DEFINE_RATELIMIT_STATE(xfs_buf_write_fail_rl_state, 30 * HZ, 10);
489
490 STATIC uint
491 xfs_buf_item_push(
492 struct xfs_log_item *lip,
493 struct list_head *buffer_list)
494 {
495 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
496 struct xfs_buf *bp = bip->bli_buf;
497 uint rval = XFS_ITEM_SUCCESS;
498
499 if (xfs_buf_ispinned(bp))
500 return XFS_ITEM_PINNED;
501 if (!xfs_buf_trylock(bp)) {
502 /*
503 * If we have just raced with a buffer being pinned and it has
504 * been marked stale, we could end up stalling until someone else
505 * issues a log force to unpin the stale buffer. Check for the
506 * race condition here so xfsaild recognizes the buffer is pinned
507 * and queues a log force to move it along.
508 */
509 if (xfs_buf_ispinned(bp))
510 return XFS_ITEM_PINNED;
511 return XFS_ITEM_LOCKED;
512 }
513
514 ASSERT(!(bip->bli_flags & XFS_BLI_STALE));
515
516 trace_xfs_buf_item_push(bip);
517
518 /* has a previous flush failed due to IO errors? */
519 if ((bp->b_flags & XBF_WRITE_FAIL) &&
520 ___ratelimit(&xfs_buf_write_fail_rl_state, "XFS: Failing async write")) {
521 xfs_warn(bp->b_mount,
522 "Failing async write on buffer block 0x%llx. Retrying async write.",
523 (long long)bp->b_bn);
524 }
525
526 if (!xfs_buf_delwri_queue(bp, buffer_list))
527 rval = XFS_ITEM_FLUSHING;
528 xfs_buf_unlock(bp);
529 return rval;
530 }
531
532 /*
533 * Drop the buffer log item refcount and take appropriate action. This helper
534 * determines whether the bli must be freed or not, since a decrement to zero
535 * does not necessarily mean the bli is unused.
536 *
537 * Return true if the bli is freed, false otherwise.
538 */
539 bool
540 xfs_buf_item_put(
541 struct xfs_buf_log_item *bip)
542 {
543 struct xfs_log_item *lip = &bip->bli_item;
544 bool aborted;
545 bool dirty;
546
547 /* drop the bli ref and return if it wasn't the last one */
548 if (!atomic_dec_and_test(&bip->bli_refcount))
549 return false;
550
551 /*
552 * We dropped the last ref and must free the item if clean or aborted.
553 * If the bli is dirty and non-aborted, the buffer was clean in the
554 * transaction but still awaiting writeback from previous changes. In
555 * that case, the bli is freed on buffer writeback completion.
556 */
557 aborted = test_bit(XFS_LI_ABORTED, &lip->li_flags) ||
558 XFS_FORCED_SHUTDOWN(lip->li_mountp);
559 dirty = bip->bli_flags & XFS_BLI_DIRTY;
560 if (dirty && !aborted)
561 return false;
562
563 /*
564 * The bli is aborted or clean. An aborted item may be in the AIL
565 * regardless of dirty state. For example, consider an aborted
566 * transaction that invalidated a dirty bli and cleared the dirty
567 * state.
568 */
569 if (aborted)
570 xfs_trans_ail_remove(lip, SHUTDOWN_LOG_IO_ERROR);
571 xfs_buf_item_relse(bip->bli_buf);
572 return true;
573 }
574
575 /*
576 * Release the buffer associated with the buf log item. If there is no dirty
577 * logged data associated with the buffer recorded in the buf log item, then
578 * free the buf log item and remove the reference to it in the buffer.
579 *
580 * This call ignores the recursion count. It is only called when the buffer
581 * should REALLY be unlocked, regardless of the recursion count.
582 *
583 * We unconditionally drop the transaction's reference to the log item. If the
584 * item was logged, then another reference was taken when it was pinned, so we
585 * can safely drop the transaction reference now. This also allows us to avoid
586 * potential races with the unpin code freeing the bli by not referencing the
587 * bli after we've dropped the reference count.
588 *
589 * If the XFS_BLI_HOLD flag is set in the buf log item, then free the log item
590 * if necessary but do not unlock the buffer. This is for support of
591 * xfs_trans_bhold(). Make sure the XFS_BLI_HOLD field is cleared if we don't
592 * free the item.
593 */
594 STATIC void
595 xfs_buf_item_release(
596 struct xfs_log_item *lip)
597 {
598 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
599 struct xfs_buf *bp = bip->bli_buf;
600 bool released;
601 bool hold = bip->bli_flags & XFS_BLI_HOLD;
602 bool stale = bip->bli_flags & XFS_BLI_STALE;
603 #if defined(DEBUG) || defined(XFS_WARN)
604 bool ordered = bip->bli_flags & XFS_BLI_ORDERED;
605 bool dirty = bip->bli_flags & XFS_BLI_DIRTY;
606 bool aborted = test_bit(XFS_LI_ABORTED,
607 &lip->li_flags);
608 #endif
609
610 trace_xfs_buf_item_release(bip);
611
612 /*
613 * The bli dirty state should match whether the blf has logged segments
614 * except for ordered buffers, where only the bli should be dirty.
615 */
616 ASSERT((!ordered && dirty == xfs_buf_item_dirty_format(bip)) ||
617 (ordered && dirty && !xfs_buf_item_dirty_format(bip)));
618 ASSERT(!stale || (bip->__bli_format.blf_flags & XFS_BLF_CANCEL));
619
620 /*
621 * Clear the buffer's association with this transaction and
622 * per-transaction state from the bli, which has been copied above.
623 */
624 bp->b_transp = NULL;
625 bip->bli_flags &= ~(XFS_BLI_LOGGED | XFS_BLI_HOLD | XFS_BLI_ORDERED);
626
627 /*
628 * Unref the item and unlock the buffer unless held or stale. Stale
629 * buffers remain locked until final unpin unless the bli is freed by
630 * the unref call. The latter implies shutdown because buffer
631 * invalidation dirties the bli and transaction.
632 */
633 released = xfs_buf_item_put(bip);
634 if (hold || (stale && !released))
635 return;
636 ASSERT(!stale || aborted);
637 xfs_buf_relse(bp);
638 }
639
640 STATIC void
641 xfs_buf_item_committing(
642 struct xfs_log_item *lip,
643 xfs_lsn_t commit_lsn)
644 {
645 return xfs_buf_item_release(lip);
646 }
647
648 /*
649 * This is called to find out where the oldest active copy of the
650 * buf log item in the on disk log resides now that the last log
651 * write of it completed at the given lsn.
652 * We always re-log all the dirty data in a buffer, so usually the
653 * latest copy in the on disk log is the only one that matters. For
654 * those cases we simply return the given lsn.
655 *
656 * The one exception to this is for buffers full of newly allocated
657 * inodes. These buffers are only relogged with the XFS_BLI_INODE_BUF
658 * flag set, indicating that only the di_next_unlinked fields from the
659 * inodes in the buffers will be replayed during recovery. If the
660 * original newly allocated inode images have not yet been flushed
661 * when the buffer is so relogged, then we need to make sure that we
662 * keep the old images in the 'active' portion of the log. We do this
663 * by returning the original lsn of that transaction here rather than
664 * the current one.
665 */
666 STATIC xfs_lsn_t
667 xfs_buf_item_committed(
668 struct xfs_log_item *lip,
669 xfs_lsn_t lsn)
670 {
671 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
672
673 trace_xfs_buf_item_committed(bip);
674
675 if ((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) && lip->li_lsn != 0)
676 return lip->li_lsn;
677 return lsn;
678 }
679
680 static const struct xfs_item_ops xfs_buf_item_ops = {
681 .iop_size = xfs_buf_item_size,
682 .iop_format = xfs_buf_item_format,
683 .iop_pin = xfs_buf_item_pin,
684 .iop_unpin = xfs_buf_item_unpin,
685 .iop_release = xfs_buf_item_release,
686 .iop_committing = xfs_buf_item_committing,
687 .iop_committed = xfs_buf_item_committed,
688 .iop_push = xfs_buf_item_push,
689 };
690
691 STATIC int
692 xfs_buf_item_get_format(
693 struct xfs_buf_log_item *bip,
694 int count)
695 {
696 ASSERT(bip->bli_formats == NULL);
697 bip->bli_format_count = count;
698
699 if (count == 1) {
700 bip->bli_formats = &bip->__bli_format;
701 return 0;
702 }
703
704 bip->bli_formats = kmem_zalloc(count * sizeof(struct xfs_buf_log_format),
705 0);
706 if (!bip->bli_formats)
707 return -ENOMEM;
708 return 0;
709 }
710
711 STATIC void
712 xfs_buf_item_free_format(
713 struct xfs_buf_log_item *bip)
714 {
715 if (bip->bli_formats != &bip->__bli_format) {
716 kmem_free(bip->bli_formats);
717 bip->bli_formats = NULL;
718 }
719 }
720
721 /*
722 * Allocate a new buf log item to go with the given buffer.
723 * Set the buffer's b_log_item field to point to the new
724 * buf log item.
725 */
726 int
727 xfs_buf_item_init(
728 struct xfs_buf *bp,
729 struct xfs_mount *mp)
730 {
731 struct xfs_buf_log_item *bip = bp->b_log_item;
732 int chunks;
733 int map_size;
734 int error;
735 int i;
736
737 /*
738 * Check to see if there is already a buf log item for
739 * this buffer. If we do already have one, there is
740 * nothing to do here so return.
741 */
742 ASSERT(bp->b_mount == mp);
743 if (bip) {
744 ASSERT(bip->bli_item.li_type == XFS_LI_BUF);
745 ASSERT(!bp->b_transp);
746 ASSERT(bip->bli_buf == bp);
747 return 0;
748 }
749
750 bip = kmem_zone_zalloc(xfs_buf_item_zone, 0);
751 xfs_log_item_init(mp, &bip->bli_item, XFS_LI_BUF, &xfs_buf_item_ops);
752 bip->bli_buf = bp;
753
754 /*
755 * chunks is the number of XFS_BLF_CHUNK size pieces the buffer
756 * can be divided into. Make sure not to truncate any pieces.
757 * map_size is the size of the bitmap needed to describe the
758 * chunks of the buffer.
759 *
760 * Discontiguous buffer support follows the layout of the underlying
761 * buffer. This makes the implementation as simple as possible.
762 */
763 error = xfs_buf_item_get_format(bip, bp->b_map_count);
764 ASSERT(error == 0);
765 if (error) { /* to stop gcc throwing set-but-unused warnings */
766 kmem_zone_free(xfs_buf_item_zone, bip);
767 return error;
768 }
769
770
771 for (i = 0; i < bip->bli_format_count; i++) {
772 chunks = DIV_ROUND_UP(BBTOB(bp->b_maps[i].bm_len),
773 XFS_BLF_CHUNK);
774 map_size = DIV_ROUND_UP(chunks, NBWORD);
775
776 bip->bli_formats[i].blf_type = XFS_LI_BUF;
777 bip->bli_formats[i].blf_blkno = bp->b_maps[i].bm_bn;
778 bip->bli_formats[i].blf_len = bp->b_maps[i].bm_len;
779 bip->bli_formats[i].blf_map_size = map_size;
780 }
781
782 bp->b_log_item = bip;
783 xfs_buf_hold(bp);
784 return 0;
785 }
786
787
788 /*
789 * Mark bytes first through last inclusive as dirty in the buf
790 * item's bitmap.
791 */
792 static void
793 xfs_buf_item_log_segment(
794 uint first,
795 uint last,
796 uint *map)
797 {
798 uint first_bit;
799 uint last_bit;
800 uint bits_to_set;
801 uint bits_set;
802 uint word_num;
803 uint *wordp;
804 uint bit;
805 uint end_bit;
806 uint mask;
807
808 /*
809 * Convert byte offsets to bit numbers.
810 */
811 first_bit = first >> XFS_BLF_SHIFT;
812 last_bit = last >> XFS_BLF_SHIFT;
813
814 /*
815 * Calculate the total number of bits to be set.
816 */
817 bits_to_set = last_bit - first_bit + 1;
818
819 /*
820 * Get a pointer to the first word in the bitmap
821 * to set a bit in.
822 */
823 word_num = first_bit >> BIT_TO_WORD_SHIFT;
824 wordp = &map[word_num];
825
826 /*
827 * Calculate the starting bit in the first word.
828 */
829 bit = first_bit & (uint)(NBWORD - 1);
830
831 /*
832 * First set any bits in the first word of our range.
833 * If it starts at bit 0 of the word, it will be
834 * set below rather than here. That is what the variable
835 * bit tells us. The variable bits_set tracks the number
836 * of bits that have been set so far. End_bit is the number
837 * of the last bit to be set in this word plus one.
838 */
839 if (bit) {
840 end_bit = min(bit + bits_to_set, (uint)NBWORD);
841 mask = ((1U << (end_bit - bit)) - 1) << bit;
842 *wordp |= mask;
843 wordp++;
844 bits_set = end_bit - bit;
845 } else {
846 bits_set = 0;
847 }
848
849 /*
850 * Now set bits a whole word at a time that are between
851 * first_bit and last_bit.
852 */
853 while ((bits_to_set - bits_set) >= NBWORD) {
854 *wordp |= 0xffffffff;
855 bits_set += NBWORD;
856 wordp++;
857 }
858
859 /*
860 * Finally, set any bits left to be set in one last partial word.
861 */
862 end_bit = bits_to_set - bits_set;
863 if (end_bit) {
864 mask = (1U << end_bit) - 1;
865 *wordp |= mask;
866 }
867 }
868
869 /*
870 * Mark bytes first through last inclusive as dirty in the buf
871 * item's bitmap.
872 */
873 void
874 xfs_buf_item_log(
875 struct xfs_buf_log_item *bip,
876 uint first,
877 uint last)
878 {
879 int i;
880 uint start;
881 uint end;
882 struct xfs_buf *bp = bip->bli_buf;
883
884 /*
885 * walk each buffer segment and mark them dirty appropriately.
886 */
887 start = 0;
888 for (i = 0; i < bip->bli_format_count; i++) {
889 if (start > last)
890 break;
891 end = start + BBTOB(bp->b_maps[i].bm_len) - 1;
892
893 /* skip to the map that includes the first byte to log */
894 if (first > end) {
895 start += BBTOB(bp->b_maps[i].bm_len);
896 continue;
897 }
898
899 /*
900 * Trim the range to this segment and mark it in the bitmap.
901 * Note that we must convert buffer offsets to segment relative
902 * offsets (e.g., the first byte of each segment is byte 0 of
903 * that segment).
904 */
905 if (first < start)
906 first = start;
907 if (end > last)
908 end = last;
909 xfs_buf_item_log_segment(first - start, end - start,
910 &bip->bli_formats[i].blf_data_map[0]);
911
912 start += BBTOB(bp->b_maps[i].bm_len);
913 }
914 }
915
916
917 /*
918 * Return true if the buffer has any ranges logged/dirtied by a transaction,
919 * false otherwise.
920 */
921 bool
922 xfs_buf_item_dirty_format(
923 struct xfs_buf_log_item *bip)
924 {
925 int i;
926
927 for (i = 0; i < bip->bli_format_count; i++) {
928 if (!xfs_bitmap_empty(bip->bli_formats[i].blf_data_map,
929 bip->bli_formats[i].blf_map_size))
930 return true;
931 }
932
933 return false;
934 }
935
936 STATIC void
937 xfs_buf_item_free(
938 struct xfs_buf_log_item *bip)
939 {
940 xfs_buf_item_free_format(bip);
941 kmem_free(bip->bli_item.li_lv_shadow);
942 kmem_zone_free(xfs_buf_item_zone, bip);
943 }
944
945 /*
946 * This is called when the buf log item is no longer needed. It should
947 * free the buf log item associated with the given buffer and clear
948 * the buffer's pointer to the buf log item. If there are no more
949 * items in the list, clear the b_iodone field of the buffer (see
950 * xfs_buf_attach_iodone() below).
951 */
952 void
953 xfs_buf_item_relse(
954 xfs_buf_t *bp)
955 {
956 struct xfs_buf_log_item *bip = bp->b_log_item;
957
958 trace_xfs_buf_item_relse(bp, _RET_IP_);
959 ASSERT(!(bip->bli_item.li_flags & XFS_LI_IN_AIL));
960
961 bp->b_log_item = NULL;
962 if (list_empty(&bp->b_li_list))
963 bp->b_iodone = NULL;
964
965 xfs_buf_rele(bp);
966 xfs_buf_item_free(bip);
967 }
968
969
970 /*
971 * Add the given log item with its callback to the list of callbacks
972 * to be called when the buffer's I/O completes. If it is not set
973 * already, set the buffer's b_iodone() routine to be
974 * xfs_buf_iodone_callbacks() and link the log item into the list of
975 * items rooted at b_li_list.
976 */
977 void
978 xfs_buf_attach_iodone(
979 struct xfs_buf *bp,
980 void (*cb)(struct xfs_buf *, struct xfs_log_item *),
981 struct xfs_log_item *lip)
982 {
983 ASSERT(xfs_buf_islocked(bp));
984
985 lip->li_cb = cb;
986 list_add_tail(&lip->li_bio_list, &bp->b_li_list);
987
988 ASSERT(bp->b_iodone == NULL ||
989 bp->b_iodone == xfs_buf_iodone_callbacks);
990 bp->b_iodone = xfs_buf_iodone_callbacks;
991 }
992
993 /*
994 * We can have many callbacks on a buffer. Running the callbacks individually
995 * can cause a lot of contention on the AIL lock, so we allow for a single
996 * callback to be able to scan the remaining items in bp->b_li_list for other
997 * items of the same type and callback to be processed in the first call.
998 *
999 * As a result, the loop walking the callback list below will also modify the
1000 * list. it removes the first item from the list and then runs the callback.
1001 * The loop then restarts from the new first item int the list. This allows the
1002 * callback to scan and modify the list attached to the buffer and we don't
1003 * have to care about maintaining a next item pointer.
1004 */
1005 STATIC void
1006 xfs_buf_do_callbacks(
1007 struct xfs_buf *bp)
1008 {
1009 struct xfs_buf_log_item *blip = bp->b_log_item;
1010 struct xfs_log_item *lip;
1011
1012 /* If there is a buf_log_item attached, run its callback */
1013 if (blip) {
1014 lip = &blip->bli_item;
1015 lip->li_cb(bp, lip);
1016 }
1017
1018 while (!list_empty(&bp->b_li_list)) {
1019 lip = list_first_entry(&bp->b_li_list, struct xfs_log_item,
1020 li_bio_list);
1021
1022 /*
1023 * Remove the item from the list, so we don't have any
1024 * confusion if the item is added to another buf.
1025 * Don't touch the log item after calling its
1026 * callback, because it could have freed itself.
1027 */
1028 list_del_init(&lip->li_bio_list);
1029 lip->li_cb(bp, lip);
1030 }
1031 }
1032
1033 /*
1034 * Invoke the error state callback for each log item affected by the failed I/O.
1035 *
1036 * If a metadata buffer write fails with a non-permanent error, the buffer is
1037 * eventually resubmitted and so the completion callbacks are not run. The error
1038 * state may need to be propagated to the log items attached to the buffer,
1039 * however, so the next AIL push of the item knows hot to handle it correctly.
1040 */
1041 STATIC void
1042 xfs_buf_do_callbacks_fail(
1043 struct xfs_buf *bp)
1044 {
1045 struct xfs_log_item *lip;
1046 struct xfs_ail *ailp;
1047
1048 /*
1049 * Buffer log item errors are handled directly by xfs_buf_item_push()
1050 * and xfs_buf_iodone_callback_error, and they have no IO error
1051 * callbacks. Check only for items in b_li_list.
1052 */
1053 if (list_empty(&bp->b_li_list))
1054 return;
1055
1056 lip = list_first_entry(&bp->b_li_list, struct xfs_log_item,
1057 li_bio_list);
1058 ailp = lip->li_ailp;
1059 spin_lock(&ailp->ail_lock);
1060 list_for_each_entry(lip, &bp->b_li_list, li_bio_list) {
1061 if (lip->li_ops->iop_error)
1062 lip->li_ops->iop_error(lip, bp);
1063 }
1064 spin_unlock(&ailp->ail_lock);
1065 }
1066
1067 static bool
1068 xfs_buf_iodone_callback_error(
1069 struct xfs_buf *bp)
1070 {
1071 struct xfs_buf_log_item *bip = bp->b_log_item;
1072 struct xfs_log_item *lip;
1073 struct xfs_mount *mp;
1074 static ulong lasttime;
1075 static xfs_buftarg_t *lasttarg;
1076 struct xfs_error_cfg *cfg;
1077
1078 /*
1079 * The failed buffer might not have a buf_log_item attached or the
1080 * log_item list might be empty. Get the mp from the available
1081 * xfs_log_item
1082 */
1083 lip = list_first_entry_or_null(&bp->b_li_list, struct xfs_log_item,
1084 li_bio_list);
1085 mp = lip ? lip->li_mountp : bip->bli_item.li_mountp;
1086
1087 /*
1088 * If we've already decided to shutdown the filesystem because of
1089 * I/O errors, there's no point in giving this a retry.
1090 */
1091 if (XFS_FORCED_SHUTDOWN(mp))
1092 goto out_stale;
1093
1094 if (bp->b_target != lasttarg ||
1095 time_after(jiffies, (lasttime + 5*HZ))) {
1096 lasttime = jiffies;
1097 xfs_buf_ioerror_alert(bp, __func__);
1098 }
1099 lasttarg = bp->b_target;
1100
1101 /* synchronous writes will have callers process the error */
1102 if (!(bp->b_flags & XBF_ASYNC))
1103 goto out_stale;
1104
1105 trace_xfs_buf_item_iodone_async(bp, _RET_IP_);
1106 ASSERT(bp->b_iodone != NULL);
1107
1108 cfg = xfs_error_get_cfg(mp, XFS_ERR_METADATA, bp->b_error);
1109
1110 /*
1111 * If the write was asynchronous then no one will be looking for the
1112 * error. If this is the first failure of this type, clear the error
1113 * state and write the buffer out again. This means we always retry an
1114 * async write failure at least once, but we also need to set the buffer
1115 * up to behave correctly now for repeated failures.
1116 */
1117 if (!(bp->b_flags & (XBF_STALE | XBF_WRITE_FAIL)) ||
1118 bp->b_last_error != bp->b_error) {
1119 bp->b_flags |= (XBF_WRITE | XBF_DONE | XBF_WRITE_FAIL);
1120 bp->b_last_error = bp->b_error;
1121 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1122 !bp->b_first_retry_time)
1123 bp->b_first_retry_time = jiffies;
1124
1125 xfs_buf_ioerror(bp, 0);
1126 xfs_buf_submit(bp);
1127 return true;
1128 }
1129
1130 /*
1131 * Repeated failure on an async write. Take action according to the
1132 * error configuration we have been set up to use.
1133 */
1134
1135 if (cfg->max_retries != XFS_ERR_RETRY_FOREVER &&
1136 ++bp->b_retries > cfg->max_retries)
1137 goto permanent_error;
1138 if (cfg->retry_timeout != XFS_ERR_RETRY_FOREVER &&
1139 time_after(jiffies, cfg->retry_timeout + bp->b_first_retry_time))
1140 goto permanent_error;
1141
1142 /* At unmount we may treat errors differently */
1143 if ((mp->m_flags & XFS_MOUNT_UNMOUNTING) && mp->m_fail_unmount)
1144 goto permanent_error;
1145
1146 /*
1147 * Still a transient error, run IO completion failure callbacks and let
1148 * the higher layers retry the buffer.
1149 */
1150 xfs_buf_do_callbacks_fail(bp);
1151 xfs_buf_ioerror(bp, 0);
1152 xfs_buf_relse(bp);
1153 return true;
1154
1155 /*
1156 * Permanent error - we need to trigger a shutdown if we haven't already
1157 * to indicate that inconsistency will result from this action.
1158 */
1159 permanent_error:
1160 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1161 out_stale:
1162 xfs_buf_stale(bp);
1163 bp->b_flags |= XBF_DONE;
1164 trace_xfs_buf_error_relse(bp, _RET_IP_);
1165 return false;
1166 }
1167
1168 /*
1169 * This is the iodone() function for buffers which have had callbacks attached
1170 * to them by xfs_buf_attach_iodone(). We need to iterate the items on the
1171 * callback list, mark the buffer as having no more callbacks and then push the
1172 * buffer through IO completion processing.
1173 */
1174 void
1175 xfs_buf_iodone_callbacks(
1176 struct xfs_buf *bp)
1177 {
1178 /*
1179 * If there is an error, process it. Some errors require us
1180 * to run callbacks after failure processing is done so we
1181 * detect that and take appropriate action.
1182 */
1183 if (bp->b_error && xfs_buf_iodone_callback_error(bp))
1184 return;
1185
1186 /*
1187 * Successful IO or permanent error. Either way, we can clear the
1188 * retry state here in preparation for the next error that may occur.
1189 */
1190 bp->b_last_error = 0;
1191 bp->b_retries = 0;
1192 bp->b_first_retry_time = 0;
1193
1194 xfs_buf_do_callbacks(bp);
1195 bp->b_log_item = NULL;
1196 list_del_init(&bp->b_li_list);
1197 bp->b_iodone = NULL;
1198 xfs_buf_ioend(bp);
1199 }
1200
1201 /*
1202 * This is the iodone() function for buffers which have been
1203 * logged. It is called when they are eventually flushed out.
1204 * It should remove the buf item from the AIL, and free the buf item.
1205 * It is called by xfs_buf_iodone_callbacks() above which will take
1206 * care of cleaning up the buffer itself.
1207 */
1208 void
1209 xfs_buf_iodone(
1210 struct xfs_buf *bp,
1211 struct xfs_log_item *lip)
1212 {
1213 struct xfs_ail *ailp = lip->li_ailp;
1214
1215 ASSERT(BUF_ITEM(lip)->bli_buf == bp);
1216
1217 xfs_buf_rele(bp);
1218
1219 /*
1220 * If we are forcibly shutting down, this may well be
1221 * off the AIL already. That's because we simulate the
1222 * log-committed callbacks to unpin these buffers. Or we may never
1223 * have put this item on AIL because of the transaction was
1224 * aborted forcibly. xfs_trans_ail_delete() takes care of these.
1225 *
1226 * Either way, AIL is useless if we're forcing a shutdown.
1227 */
1228 spin_lock(&ailp->ail_lock);
1229 xfs_trans_ail_delete(ailp, lip, SHUTDOWN_CORRUPT_INCORE);
1230 xfs_buf_item_free(BUF_ITEM(lip));
1231 }
1232
1233 /*
1234 * Requeue a failed buffer for writeback.
1235 *
1236 * We clear the log item failed state here as well, but we have to be careful
1237 * about reference counts because the only active reference counts on the buffer
1238 * may be the failed log items. Hence if we clear the log item failed state
1239 * before queuing the buffer for IO we can release all active references to
1240 * the buffer and free it, leading to use after free problems in
1241 * xfs_buf_delwri_queue. It makes no difference to the buffer or log items which
1242 * order we process them in - the buffer is locked, and we own the buffer list
1243 * so nothing on them is going to change while we are performing this action.
1244 *
1245 * Hence we can safely queue the buffer for IO before we clear the failed log
1246 * item state, therefore always having an active reference to the buffer and
1247 * avoiding the transient zero-reference state that leads to use-after-free.
1248 *
1249 * Return true if the buffer was added to the buffer list, false if it was
1250 * already on the buffer list.
1251 */
1252 bool
1253 xfs_buf_resubmit_failed_buffers(
1254 struct xfs_buf *bp,
1255 struct list_head *buffer_list)
1256 {
1257 struct xfs_log_item *lip;
1258 bool ret;
1259
1260 ret = xfs_buf_delwri_queue(bp, buffer_list);
1261
1262 /*
1263 * XFS_LI_FAILED set/clear is protected by ail_lock, caller of this
1264 * function already have it acquired
1265 */
1266 list_for_each_entry(lip, &bp->b_li_list, li_bio_list)
1267 xfs_clear_li_failed(lip);
1268
1269 return ret;
1270 }