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_trans_priv.h"
16 #include "xfs_buf_item.h"
17 #include "xfs_inode.h"
18 #include "xfs_inode_item.h"
19 #include "xfs_quota.h"
20 #include "xfs_dquot_item.h"
21 #include "xfs_dquot.h"
22 #include "xfs_trace.h"
23 #include "xfs_log.h"
24
25
26 kmem_zone_t *xfs_buf_item_zone;
27
BUF_ITEM(struct xfs_log_item * lip)28 static inline struct xfs_buf_log_item *BUF_ITEM(struct xfs_log_item *lip)
29 {
30 return container_of(lip, struct xfs_buf_log_item, bli_item);
31 }
32
33 /* Is this log iovec plausibly large enough to contain the buffer log format? */
34 bool
xfs_buf_log_check_iovec(struct xfs_log_iovec * iovec)35 xfs_buf_log_check_iovec(
36 struct xfs_log_iovec *iovec)
37 {
38 struct xfs_buf_log_format *blfp = iovec->i_addr;
39 char *bmp_end;
40 char *item_end;
41
42 if (offsetof(struct xfs_buf_log_format, blf_data_map) > iovec->i_len)
43 return false;
44
45 item_end = (char *)iovec->i_addr + iovec->i_len;
46 bmp_end = (char *)&blfp->blf_data_map[blfp->blf_map_size];
47 return bmp_end <= item_end;
48 }
49
50 static inline int
xfs_buf_log_format_size(struct xfs_buf_log_format * blfp)51 xfs_buf_log_format_size(
52 struct xfs_buf_log_format *blfp)
53 {
54 return offsetof(struct xfs_buf_log_format, blf_data_map) +
55 (blfp->blf_map_size * sizeof(blfp->blf_data_map[0]));
56 }
57
58 static inline bool
xfs_buf_item_straddle(struct xfs_buf * bp,uint offset,int first_bit,int nbits)59 xfs_buf_item_straddle(
60 struct xfs_buf *bp,
61 uint offset,
62 int first_bit,
63 int nbits)
64 {
65 void *first, *last;
66
67 first = xfs_buf_offset(bp, offset + (first_bit << XFS_BLF_SHIFT));
68 last = xfs_buf_offset(bp,
69 offset + ((first_bit + nbits) << XFS_BLF_SHIFT));
70
71 if (last - first != nbits * XFS_BLF_CHUNK)
72 return true;
73 return false;
74 }
75
76 /*
77 * This returns the number of log iovecs needed to log the
78 * given buf log item.
79 *
80 * It calculates this as 1 iovec for the buf log format structure
81 * and 1 for each stretch of non-contiguous chunks to be logged.
82 * Contiguous chunks are logged in a single iovec.
83 *
84 * If the XFS_BLI_STALE flag has been set, then log nothing.
85 */
86 STATIC void
xfs_buf_item_size_segment(struct xfs_buf_log_item * bip,struct xfs_buf_log_format * blfp,uint offset,int * nvecs,int * nbytes)87 xfs_buf_item_size_segment(
88 struct xfs_buf_log_item *bip,
89 struct xfs_buf_log_format *blfp,
90 uint offset,
91 int *nvecs,
92 int *nbytes)
93 {
94 struct xfs_buf *bp = bip->bli_buf;
95 int first_bit;
96 int nbits;
97 int next_bit;
98 int last_bit;
99
100 first_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size, 0);
101 if (first_bit == -1)
102 return;
103
104 (*nvecs)++;
105 *nbytes += xfs_buf_log_format_size(blfp);
106
107 do {
108 nbits = xfs_contig_bits(blfp->blf_data_map,
109 blfp->blf_map_size, first_bit);
110 ASSERT(nbits > 0);
111
112 /*
113 * Straddling a page is rare because we don't log contiguous
114 * chunks of unmapped buffers anywhere.
115 */
116 if (nbits > 1 &&
117 xfs_buf_item_straddle(bp, offset, first_bit, nbits))
118 goto slow_scan;
119
120 (*nvecs)++;
121 *nbytes += nbits * XFS_BLF_CHUNK;
122
123 /*
124 * This takes the bit number to start looking from and
125 * returns the next set bit from there. It returns -1
126 * if there are no more bits set or the start bit is
127 * beyond the end of the bitmap.
128 */
129 first_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size,
130 (uint)first_bit + nbits + 1);
131 } while (first_bit != -1);
132
133 return;
134
135 slow_scan:
136 /* Count the first bit we jumped out of the above loop from */
137 (*nvecs)++;
138 *nbytes += XFS_BLF_CHUNK;
139 last_bit = first_bit;
140 while (last_bit != -1) {
141 /*
142 * This takes the bit number to start looking from and
143 * returns the next set bit from there. It returns -1
144 * if there are no more bits set or the start bit is
145 * beyond the end of the bitmap.
146 */
147 next_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size,
148 last_bit + 1);
149 /*
150 * If we run out of bits, leave the loop,
151 * else if we find a new set of bits bump the number of vecs,
152 * else keep scanning the current set of bits.
153 */
154 if (next_bit == -1) {
155 break;
156 } else if (next_bit != last_bit + 1 ||
157 xfs_buf_item_straddle(bp, offset, first_bit, nbits)) {
158 last_bit = next_bit;
159 first_bit = next_bit;
160 (*nvecs)++;
161 nbits = 1;
162 } else {
163 last_bit++;
164 nbits++;
165 }
166 *nbytes += XFS_BLF_CHUNK;
167 }
168 }
169
170 /*
171 * This returns the number of log iovecs needed to log the given buf log item.
172 *
173 * It calculates this as 1 iovec for the buf log format structure and 1 for each
174 * stretch of non-contiguous chunks to be logged. Contiguous chunks are logged
175 * in a single iovec.
176 *
177 * Discontiguous buffers need a format structure per region that is being
178 * logged. This makes the changes in the buffer appear to log recovery as though
179 * they came from separate buffers, just like would occur if multiple buffers
180 * were used instead of a single discontiguous buffer. This enables
181 * discontiguous buffers to be in-memory constructs, completely transparent to
182 * what ends up on disk.
183 *
184 * If the XFS_BLI_STALE flag has been set, then log nothing but the buf log
185 * format structures.
186 */
187 STATIC void
xfs_buf_item_size(struct xfs_log_item * lip,int * nvecs,int * nbytes)188 xfs_buf_item_size(
189 struct xfs_log_item *lip,
190 int *nvecs,
191 int *nbytes)
192 {
193 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
194 struct xfs_buf *bp = bip->bli_buf;
195 int i;
196 int bytes;
197 uint offset = 0;
198
199 ASSERT(atomic_read(&bip->bli_refcount) > 0);
200 if (bip->bli_flags & XFS_BLI_STALE) {
201 /*
202 * The buffer is stale, so all we need to log
203 * is the buf log format structure with the
204 * cancel flag in it.
205 */
206 trace_xfs_buf_item_size_stale(bip);
207 ASSERT(bip->__bli_format.blf_flags & XFS_BLF_CANCEL);
208 *nvecs += bip->bli_format_count;
209 for (i = 0; i < bip->bli_format_count; i++) {
210 *nbytes += xfs_buf_log_format_size(&bip->bli_formats[i]);
211 }
212 return;
213 }
214
215 ASSERT(bip->bli_flags & XFS_BLI_LOGGED);
216
217 if (bip->bli_flags & XFS_BLI_ORDERED) {
218 /*
219 * The buffer has been logged just to order it.
220 * It is not being included in the transaction
221 * commit, so no vectors are used at all.
222 */
223 trace_xfs_buf_item_size_ordered(bip);
224 *nvecs = XFS_LOG_VEC_ORDERED;
225 return;
226 }
227
228 /*
229 * The vector count is based on the number of buffer vectors we have
230 * dirty bits in. This will only be greater than one when we have a
231 * compound buffer with more than one segment dirty. Hence for compound
232 * buffers we need to track which segment the dirty bits correspond to,
233 * and when we move from one segment to the next increment the vector
234 * count for the extra buf log format structure that will need to be
235 * written.
236 */
237 bytes = 0;
238 for (i = 0; i < bip->bli_format_count; i++) {
239 xfs_buf_item_size_segment(bip, &bip->bli_formats[i], offset,
240 nvecs, &bytes);
241 offset += BBTOB(bp->b_maps[i].bm_len);
242 }
243
244 /*
245 * Round up the buffer size required to minimise the number of memory
246 * allocations that need to be done as this item grows when relogged by
247 * repeated modifications.
248 */
249 *nbytes = round_up(bytes, 512);
250 trace_xfs_buf_item_size(bip);
251 }
252
253 static inline void
xfs_buf_item_copy_iovec(struct xfs_log_vec * lv,struct xfs_log_iovec ** vecp,struct xfs_buf * bp,uint offset,int first_bit,uint nbits)254 xfs_buf_item_copy_iovec(
255 struct xfs_log_vec *lv,
256 struct xfs_log_iovec **vecp,
257 struct xfs_buf *bp,
258 uint offset,
259 int first_bit,
260 uint nbits)
261 {
262 offset += first_bit * XFS_BLF_CHUNK;
263 xlog_copy_iovec(lv, vecp, XLOG_REG_TYPE_BCHUNK,
264 xfs_buf_offset(bp, offset),
265 nbits * XFS_BLF_CHUNK);
266 }
267
268 static void
xfs_buf_item_format_segment(struct xfs_buf_log_item * bip,struct xfs_log_vec * lv,struct xfs_log_iovec ** vecp,uint offset,struct xfs_buf_log_format * blfp)269 xfs_buf_item_format_segment(
270 struct xfs_buf_log_item *bip,
271 struct xfs_log_vec *lv,
272 struct xfs_log_iovec **vecp,
273 uint offset,
274 struct xfs_buf_log_format *blfp)
275 {
276 struct xfs_buf *bp = bip->bli_buf;
277 uint base_size;
278 int first_bit;
279 int last_bit;
280 int next_bit;
281 uint nbits;
282
283 /* copy the flags across from the base format item */
284 blfp->blf_flags = bip->__bli_format.blf_flags;
285
286 /*
287 * Base size is the actual size of the ondisk structure - it reflects
288 * the actual size of the dirty bitmap rather than the size of the in
289 * memory structure.
290 */
291 base_size = xfs_buf_log_format_size(blfp);
292
293 first_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size, 0);
294 if (!(bip->bli_flags & XFS_BLI_STALE) && first_bit == -1) {
295 /*
296 * If the map is not be dirty in the transaction, mark
297 * the size as zero and do not advance the vector pointer.
298 */
299 return;
300 }
301
302 blfp = xlog_copy_iovec(lv, vecp, XLOG_REG_TYPE_BFORMAT, blfp, base_size);
303 blfp->blf_size = 1;
304
305 if (bip->bli_flags & XFS_BLI_STALE) {
306 /*
307 * The buffer is stale, so all we need to log
308 * is the buf log format structure with the
309 * cancel flag in it.
310 */
311 trace_xfs_buf_item_format_stale(bip);
312 ASSERT(blfp->blf_flags & XFS_BLF_CANCEL);
313 return;
314 }
315
316
317 /*
318 * Fill in an iovec for each set of contiguous chunks.
319 */
320 do {
321 ASSERT(first_bit >= 0);
322 nbits = xfs_contig_bits(blfp->blf_data_map,
323 blfp->blf_map_size, first_bit);
324 ASSERT(nbits > 0);
325
326 /*
327 * Straddling a page is rare because we don't log contiguous
328 * chunks of unmapped buffers anywhere.
329 */
330 if (nbits > 1 &&
331 xfs_buf_item_straddle(bp, offset, first_bit, nbits))
332 goto slow_scan;
333
334 xfs_buf_item_copy_iovec(lv, vecp, bp, offset,
335 first_bit, nbits);
336 blfp->blf_size++;
337
338 /*
339 * This takes the bit number to start looking from and
340 * returns the next set bit from there. It returns -1
341 * if there are no more bits set or the start bit is
342 * beyond the end of the bitmap.
343 */
344 first_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size,
345 (uint)first_bit + nbits + 1);
346 } while (first_bit != -1);
347
348 return;
349
350 slow_scan:
351 ASSERT(bp->b_addr == NULL);
352 last_bit = first_bit;
353 nbits = 1;
354 for (;;) {
355 /*
356 * This takes the bit number to start looking from and
357 * returns the next set bit from there. It returns -1
358 * if there are no more bits set or the start bit is
359 * beyond the end of the bitmap.
360 */
361 next_bit = xfs_next_bit(blfp->blf_data_map, blfp->blf_map_size,
362 (uint)last_bit + 1);
363 /*
364 * If we run out of bits fill in the last iovec and get out of
365 * the loop. Else if we start a new set of bits then fill in
366 * the iovec for the series we were looking at and start
367 * counting the bits in the new one. Else we're still in the
368 * same set of bits so just keep counting and scanning.
369 */
370 if (next_bit == -1) {
371 xfs_buf_item_copy_iovec(lv, vecp, bp, offset,
372 first_bit, nbits);
373 blfp->blf_size++;
374 break;
375 } else if (next_bit != last_bit + 1 ||
376 xfs_buf_item_straddle(bp, offset, first_bit, nbits)) {
377 xfs_buf_item_copy_iovec(lv, vecp, bp, offset,
378 first_bit, nbits);
379 blfp->blf_size++;
380 first_bit = next_bit;
381 last_bit = next_bit;
382 nbits = 1;
383 } else {
384 last_bit++;
385 nbits++;
386 }
387 }
388 }
389
390 /*
391 * This is called to fill in the vector of log iovecs for the
392 * given log buf item. It fills the first entry with a buf log
393 * format structure, and the rest point to contiguous chunks
394 * within the buffer.
395 */
396 STATIC void
xfs_buf_item_format(struct xfs_log_item * lip,struct xfs_log_vec * lv)397 xfs_buf_item_format(
398 struct xfs_log_item *lip,
399 struct xfs_log_vec *lv)
400 {
401 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
402 struct xfs_buf *bp = bip->bli_buf;
403 struct xfs_log_iovec *vecp = NULL;
404 uint offset = 0;
405 int i;
406
407 ASSERT(atomic_read(&bip->bli_refcount) > 0);
408 ASSERT((bip->bli_flags & XFS_BLI_LOGGED) ||
409 (bip->bli_flags & XFS_BLI_STALE));
410 ASSERT((bip->bli_flags & XFS_BLI_STALE) ||
411 (xfs_blft_from_flags(&bip->__bli_format) > XFS_BLFT_UNKNOWN_BUF
412 && xfs_blft_from_flags(&bip->__bli_format) < XFS_BLFT_MAX_BUF));
413 ASSERT(!(bip->bli_flags & XFS_BLI_ORDERED) ||
414 (bip->bli_flags & XFS_BLI_STALE));
415
416
417 /*
418 * If it is an inode buffer, transfer the in-memory state to the
419 * format flags and clear the in-memory state.
420 *
421 * For buffer based inode allocation, we do not transfer
422 * this state if the inode buffer allocation has not yet been committed
423 * to the log as setting the XFS_BLI_INODE_BUF flag will prevent
424 * correct replay of the inode allocation.
425 *
426 * For icreate item based inode allocation, the buffers aren't written
427 * to the journal during allocation, and hence we should always tag the
428 * buffer as an inode buffer so that the correct unlinked list replay
429 * occurs during recovery.
430 */
431 if (bip->bli_flags & XFS_BLI_INODE_BUF) {
432 if (xfs_sb_version_has_v3inode(&lip->li_mountp->m_sb) ||
433 !((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) &&
434 xfs_log_item_in_current_chkpt(lip)))
435 bip->__bli_format.blf_flags |= XFS_BLF_INODE_BUF;
436 bip->bli_flags &= ~XFS_BLI_INODE_BUF;
437 }
438
439 for (i = 0; i < bip->bli_format_count; i++) {
440 xfs_buf_item_format_segment(bip, lv, &vecp, offset,
441 &bip->bli_formats[i]);
442 offset += BBTOB(bp->b_maps[i].bm_len);
443 }
444
445 /*
446 * Check to make sure everything is consistent.
447 */
448 trace_xfs_buf_item_format(bip);
449 }
450
451 /*
452 * This is called to pin the buffer associated with the buf log item in memory
453 * so it cannot be written out.
454 *
455 * We also always take a reference to the buffer log item here so that the bli
456 * is held while the item is pinned in memory. This means that we can
457 * unconditionally drop the reference count a transaction holds when the
458 * transaction is completed.
459 */
460 STATIC void
xfs_buf_item_pin(struct xfs_log_item * lip)461 xfs_buf_item_pin(
462 struct xfs_log_item *lip)
463 {
464 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
465
466 ASSERT(atomic_read(&bip->bli_refcount) > 0);
467 ASSERT((bip->bli_flags & XFS_BLI_LOGGED) ||
468 (bip->bli_flags & XFS_BLI_ORDERED) ||
469 (bip->bli_flags & XFS_BLI_STALE));
470
471 trace_xfs_buf_item_pin(bip);
472
473 atomic_inc(&bip->bli_refcount);
474 atomic_inc(&bip->bli_buf->b_pin_count);
475 }
476
477 /*
478 * This is called to unpin the buffer associated with the buf log
479 * item which was previously pinned with a call to xfs_buf_item_pin().
480 *
481 * Also drop the reference to the buf item for the current transaction.
482 * If the XFS_BLI_STALE flag is set and we are the last reference,
483 * then free up the buf log item and unlock the buffer.
484 *
485 * If the remove flag is set we are called from uncommit in the
486 * forced-shutdown path. If that is true and the reference count on
487 * the log item is going to drop to zero we need to free the item's
488 * descriptor in the transaction.
489 */
490 STATIC void
xfs_buf_item_unpin(struct xfs_log_item * lip,int remove)491 xfs_buf_item_unpin(
492 struct xfs_log_item *lip,
493 int remove)
494 {
495 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
496 struct xfs_buf *bp = bip->bli_buf;
497 int stale = bip->bli_flags & XFS_BLI_STALE;
498 int freed;
499
500 ASSERT(bp->b_log_item == bip);
501 ASSERT(atomic_read(&bip->bli_refcount) > 0);
502
503 trace_xfs_buf_item_unpin(bip);
504
505 freed = atomic_dec_and_test(&bip->bli_refcount);
506
507 if (atomic_dec_and_test(&bp->b_pin_count))
508 wake_up_all(&bp->b_waiters);
509
510 if (freed && stale) {
511 ASSERT(bip->bli_flags & XFS_BLI_STALE);
512 ASSERT(xfs_buf_islocked(bp));
513 ASSERT(bp->b_flags & XBF_STALE);
514 ASSERT(bip->__bli_format.blf_flags & XFS_BLF_CANCEL);
515
516 trace_xfs_buf_item_unpin_stale(bip);
517
518 if (remove) {
519 /*
520 * If we are in a transaction context, we have to
521 * remove the log item from the transaction as we are
522 * about to release our reference to the buffer. If we
523 * don't, the unlock that occurs later in
524 * xfs_trans_uncommit() will try to reference the
525 * buffer which we no longer have a hold on.
526 */
527 if (!list_empty(&lip->li_trans))
528 xfs_trans_del_item(lip);
529
530 /*
531 * Since the transaction no longer refers to the buffer,
532 * the buffer should no longer refer to the transaction.
533 */
534 bp->b_transp = NULL;
535 }
536
537 /*
538 * If we get called here because of an IO error, we may or may
539 * not have the item on the AIL. xfs_trans_ail_delete() will
540 * take care of that situation. xfs_trans_ail_delete() drops
541 * the AIL lock.
542 */
543 if (bip->bli_flags & XFS_BLI_STALE_INODE) {
544 xfs_buf_item_done(bp);
545 xfs_buf_inode_iodone(bp);
546 ASSERT(list_empty(&bp->b_li_list));
547 } else {
548 xfs_trans_ail_delete(lip, SHUTDOWN_LOG_IO_ERROR);
549 xfs_buf_item_relse(bp);
550 ASSERT(bp->b_log_item == NULL);
551 }
552 xfs_buf_relse(bp);
553 } else if (freed && remove) {
554 /*
555 * The buffer must be locked and held by the caller to simulate
556 * an async I/O failure.
557 */
558 xfs_buf_lock(bp);
559 xfs_buf_hold(bp);
560 bp->b_flags |= XBF_ASYNC;
561 xfs_buf_ioend_fail(bp);
562 }
563 }
564
565 STATIC uint
xfs_buf_item_push(struct xfs_log_item * lip,struct list_head * buffer_list)566 xfs_buf_item_push(
567 struct xfs_log_item *lip,
568 struct list_head *buffer_list)
569 {
570 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
571 struct xfs_buf *bp = bip->bli_buf;
572 uint rval = XFS_ITEM_SUCCESS;
573
574 if (xfs_buf_ispinned(bp))
575 return XFS_ITEM_PINNED;
576 if (!xfs_buf_trylock(bp)) {
577 /*
578 * If we have just raced with a buffer being pinned and it has
579 * been marked stale, we could end up stalling until someone else
580 * issues a log force to unpin the stale buffer. Check for the
581 * race condition here so xfsaild recognizes the buffer is pinned
582 * and queues a log force to move it along.
583 */
584 if (xfs_buf_ispinned(bp))
585 return XFS_ITEM_PINNED;
586 return XFS_ITEM_LOCKED;
587 }
588
589 ASSERT(!(bip->bli_flags & XFS_BLI_STALE));
590
591 trace_xfs_buf_item_push(bip);
592
593 /* has a previous flush failed due to IO errors? */
594 if (bp->b_flags & XBF_WRITE_FAIL) {
595 xfs_buf_alert_ratelimited(bp, "XFS: Failing async write",
596 "Failing async write on buffer block 0x%llx. Retrying async write.",
597 (long long)bp->b_bn);
598 }
599
600 if (!xfs_buf_delwri_queue(bp, buffer_list))
601 rval = XFS_ITEM_FLUSHING;
602 xfs_buf_unlock(bp);
603 return rval;
604 }
605
606 /*
607 * Drop the buffer log item refcount and take appropriate action. This helper
608 * determines whether the bli must be freed or not, since a decrement to zero
609 * does not necessarily mean the bli is unused.
610 *
611 * Return true if the bli is freed, false otherwise.
612 */
613 bool
xfs_buf_item_put(struct xfs_buf_log_item * bip)614 xfs_buf_item_put(
615 struct xfs_buf_log_item *bip)
616 {
617 struct xfs_log_item *lip = &bip->bli_item;
618 bool aborted;
619 bool dirty;
620
621 /* drop the bli ref and return if it wasn't the last one */
622 if (!atomic_dec_and_test(&bip->bli_refcount))
623 return false;
624
625 /*
626 * We dropped the last ref and must free the item if clean or aborted.
627 * If the bli is dirty and non-aborted, the buffer was clean in the
628 * transaction but still awaiting writeback from previous changes. In
629 * that case, the bli is freed on buffer writeback completion.
630 */
631 aborted = test_bit(XFS_LI_ABORTED, &lip->li_flags) ||
632 XFS_FORCED_SHUTDOWN(lip->li_mountp);
633 dirty = bip->bli_flags & XFS_BLI_DIRTY;
634 if (dirty && !aborted)
635 return false;
636
637 /*
638 * The bli is aborted or clean. An aborted item may be in the AIL
639 * regardless of dirty state. For example, consider an aborted
640 * transaction that invalidated a dirty bli and cleared the dirty
641 * state.
642 */
643 if (aborted)
644 xfs_trans_ail_delete(lip, 0);
645 xfs_buf_item_relse(bip->bli_buf);
646 return true;
647 }
648
649 /*
650 * Release the buffer associated with the buf log item. If there is no dirty
651 * logged data associated with the buffer recorded in the buf log item, then
652 * free the buf log item and remove the reference to it in the buffer.
653 *
654 * This call ignores the recursion count. It is only called when the buffer
655 * should REALLY be unlocked, regardless of the recursion count.
656 *
657 * We unconditionally drop the transaction's reference to the log item. If the
658 * item was logged, then another reference was taken when it was pinned, so we
659 * can safely drop the transaction reference now. This also allows us to avoid
660 * potential races with the unpin code freeing the bli by not referencing the
661 * bli after we've dropped the reference count.
662 *
663 * If the XFS_BLI_HOLD flag is set in the buf log item, then free the log item
664 * if necessary but do not unlock the buffer. This is for support of
665 * xfs_trans_bhold(). Make sure the XFS_BLI_HOLD field is cleared if we don't
666 * free the item.
667 */
668 STATIC void
xfs_buf_item_release(struct xfs_log_item * lip)669 xfs_buf_item_release(
670 struct xfs_log_item *lip)
671 {
672 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
673 struct xfs_buf *bp = bip->bli_buf;
674 bool released;
675 bool hold = bip->bli_flags & XFS_BLI_HOLD;
676 bool stale = bip->bli_flags & XFS_BLI_STALE;
677 #if defined(DEBUG) || defined(XFS_WARN)
678 bool ordered = bip->bli_flags & XFS_BLI_ORDERED;
679 bool dirty = bip->bli_flags & XFS_BLI_DIRTY;
680 bool aborted = test_bit(XFS_LI_ABORTED,
681 &lip->li_flags);
682 #endif
683
684 trace_xfs_buf_item_release(bip);
685
686 /*
687 * The bli dirty state should match whether the blf has logged segments
688 * except for ordered buffers, where only the bli should be dirty.
689 */
690 ASSERT((!ordered && dirty == xfs_buf_item_dirty_format(bip)) ||
691 (ordered && dirty && !xfs_buf_item_dirty_format(bip)));
692 ASSERT(!stale || (bip->__bli_format.blf_flags & XFS_BLF_CANCEL));
693
694 /*
695 * Clear the buffer's association with this transaction and
696 * per-transaction state from the bli, which has been copied above.
697 */
698 bp->b_transp = NULL;
699 bip->bli_flags &= ~(XFS_BLI_LOGGED | XFS_BLI_HOLD | XFS_BLI_ORDERED);
700
701 /*
702 * Unref the item and unlock the buffer unless held or stale. Stale
703 * buffers remain locked until final unpin unless the bli is freed by
704 * the unref call. The latter implies shutdown because buffer
705 * invalidation dirties the bli and transaction.
706 */
707 released = xfs_buf_item_put(bip);
708 if (hold || (stale && !released))
709 return;
710 ASSERT(!stale || aborted);
711 xfs_buf_relse(bp);
712 }
713
714 STATIC void
xfs_buf_item_committing(struct xfs_log_item * lip,xfs_lsn_t commit_lsn)715 xfs_buf_item_committing(
716 struct xfs_log_item *lip,
717 xfs_lsn_t commit_lsn)
718 {
719 return xfs_buf_item_release(lip);
720 }
721
722 /*
723 * This is called to find out where the oldest active copy of the
724 * buf log item in the on disk log resides now that the last log
725 * write of it completed at the given lsn.
726 * We always re-log all the dirty data in a buffer, so usually the
727 * latest copy in the on disk log is the only one that matters. For
728 * those cases we simply return the given lsn.
729 *
730 * The one exception to this is for buffers full of newly allocated
731 * inodes. These buffers are only relogged with the XFS_BLI_INODE_BUF
732 * flag set, indicating that only the di_next_unlinked fields from the
733 * inodes in the buffers will be replayed during recovery. If the
734 * original newly allocated inode images have not yet been flushed
735 * when the buffer is so relogged, then we need to make sure that we
736 * keep the old images in the 'active' portion of the log. We do this
737 * by returning the original lsn of that transaction here rather than
738 * the current one.
739 */
740 STATIC xfs_lsn_t
xfs_buf_item_committed(struct xfs_log_item * lip,xfs_lsn_t lsn)741 xfs_buf_item_committed(
742 struct xfs_log_item *lip,
743 xfs_lsn_t lsn)
744 {
745 struct xfs_buf_log_item *bip = BUF_ITEM(lip);
746
747 trace_xfs_buf_item_committed(bip);
748
749 if ((bip->bli_flags & XFS_BLI_INODE_ALLOC_BUF) && lip->li_lsn != 0)
750 return lip->li_lsn;
751 return lsn;
752 }
753
754 static const struct xfs_item_ops xfs_buf_item_ops = {
755 .iop_size = xfs_buf_item_size,
756 .iop_format = xfs_buf_item_format,
757 .iop_pin = xfs_buf_item_pin,
758 .iop_unpin = xfs_buf_item_unpin,
759 .iop_release = xfs_buf_item_release,
760 .iop_committing = xfs_buf_item_committing,
761 .iop_committed = xfs_buf_item_committed,
762 .iop_push = xfs_buf_item_push,
763 };
764
765 STATIC void
xfs_buf_item_get_format(struct xfs_buf_log_item * bip,int count)766 xfs_buf_item_get_format(
767 struct xfs_buf_log_item *bip,
768 int count)
769 {
770 ASSERT(bip->bli_formats == NULL);
771 bip->bli_format_count = count;
772
773 if (count == 1) {
774 bip->bli_formats = &bip->__bli_format;
775 return;
776 }
777
778 bip->bli_formats = kmem_zalloc(count * sizeof(struct xfs_buf_log_format),
779 0);
780 }
781
782 STATIC void
xfs_buf_item_free_format(struct xfs_buf_log_item * bip)783 xfs_buf_item_free_format(
784 struct xfs_buf_log_item *bip)
785 {
786 if (bip->bli_formats != &bip->__bli_format) {
787 kmem_free(bip->bli_formats);
788 bip->bli_formats = NULL;
789 }
790 }
791
792 /*
793 * Allocate a new buf log item to go with the given buffer.
794 * Set the buffer's b_log_item field to point to the new
795 * buf log item.
796 */
797 int
xfs_buf_item_init(struct xfs_buf * bp,struct xfs_mount * mp)798 xfs_buf_item_init(
799 struct xfs_buf *bp,
800 struct xfs_mount *mp)
801 {
802 struct xfs_buf_log_item *bip = bp->b_log_item;
803 int chunks;
804 int map_size;
805 int i;
806
807 /*
808 * Check to see if there is already a buf log item for
809 * this buffer. If we do already have one, there is
810 * nothing to do here so return.
811 */
812 ASSERT(bp->b_mount == mp);
813 if (bip) {
814 ASSERT(bip->bli_item.li_type == XFS_LI_BUF);
815 ASSERT(!bp->b_transp);
816 ASSERT(bip->bli_buf == bp);
817 return 0;
818 }
819
820 bip = kmem_cache_zalloc(xfs_buf_item_zone, GFP_KERNEL | __GFP_NOFAIL);
821 xfs_log_item_init(mp, &bip->bli_item, XFS_LI_BUF, &xfs_buf_item_ops);
822 bip->bli_buf = bp;
823
824 /*
825 * chunks is the number of XFS_BLF_CHUNK size pieces the buffer
826 * can be divided into. Make sure not to truncate any pieces.
827 * map_size is the size of the bitmap needed to describe the
828 * chunks of the buffer.
829 *
830 * Discontiguous buffer support follows the layout of the underlying
831 * buffer. This makes the implementation as simple as possible.
832 */
833 xfs_buf_item_get_format(bip, bp->b_map_count);
834
835 for (i = 0; i < bip->bli_format_count; i++) {
836 chunks = DIV_ROUND_UP(BBTOB(bp->b_maps[i].bm_len),
837 XFS_BLF_CHUNK);
838 map_size = DIV_ROUND_UP(chunks, NBWORD);
839
840 if (map_size > XFS_BLF_DATAMAP_SIZE) {
841 kmem_cache_free(xfs_buf_item_zone, bip);
842 xfs_err(mp,
843 "buffer item dirty bitmap (%u uints) too small to reflect %u bytes!",
844 map_size,
845 BBTOB(bp->b_maps[i].bm_len));
846 return -EFSCORRUPTED;
847 }
848
849 bip->bli_formats[i].blf_type = XFS_LI_BUF;
850 bip->bli_formats[i].blf_blkno = bp->b_maps[i].bm_bn;
851 bip->bli_formats[i].blf_len = bp->b_maps[i].bm_len;
852 bip->bli_formats[i].blf_map_size = map_size;
853 }
854
855 bp->b_log_item = bip;
856 xfs_buf_hold(bp);
857 return 0;
858 }
859
860
861 /*
862 * Mark bytes first through last inclusive as dirty in the buf
863 * item's bitmap.
864 */
865 static void
xfs_buf_item_log_segment(uint first,uint last,uint * map)866 xfs_buf_item_log_segment(
867 uint first,
868 uint last,
869 uint *map)
870 {
871 uint first_bit;
872 uint last_bit;
873 uint bits_to_set;
874 uint bits_set;
875 uint word_num;
876 uint *wordp;
877 uint bit;
878 uint end_bit;
879 uint mask;
880
881 ASSERT(first < XFS_BLF_DATAMAP_SIZE * XFS_BLF_CHUNK * NBWORD);
882 ASSERT(last < XFS_BLF_DATAMAP_SIZE * XFS_BLF_CHUNK * NBWORD);
883
884 /*
885 * Convert byte offsets to bit numbers.
886 */
887 first_bit = first >> XFS_BLF_SHIFT;
888 last_bit = last >> XFS_BLF_SHIFT;
889
890 /*
891 * Calculate the total number of bits to be set.
892 */
893 bits_to_set = last_bit - first_bit + 1;
894
895 /*
896 * Get a pointer to the first word in the bitmap
897 * to set a bit in.
898 */
899 word_num = first_bit >> BIT_TO_WORD_SHIFT;
900 wordp = &map[word_num];
901
902 /*
903 * Calculate the starting bit in the first word.
904 */
905 bit = first_bit & (uint)(NBWORD - 1);
906
907 /*
908 * First set any bits in the first word of our range.
909 * If it starts at bit 0 of the word, it will be
910 * set below rather than here. That is what the variable
911 * bit tells us. The variable bits_set tracks the number
912 * of bits that have been set so far. End_bit is the number
913 * of the last bit to be set in this word plus one.
914 */
915 if (bit) {
916 end_bit = min(bit + bits_to_set, (uint)NBWORD);
917 mask = ((1U << (end_bit - bit)) - 1) << bit;
918 *wordp |= mask;
919 wordp++;
920 bits_set = end_bit - bit;
921 } else {
922 bits_set = 0;
923 }
924
925 /*
926 * Now set bits a whole word at a time that are between
927 * first_bit and last_bit.
928 */
929 while ((bits_to_set - bits_set) >= NBWORD) {
930 *wordp = 0xffffffff;
931 bits_set += NBWORD;
932 wordp++;
933 }
934
935 /*
936 * Finally, set any bits left to be set in one last partial word.
937 */
938 end_bit = bits_to_set - bits_set;
939 if (end_bit) {
940 mask = (1U << end_bit) - 1;
941 *wordp |= mask;
942 }
943 }
944
945 /*
946 * Mark bytes first through last inclusive as dirty in the buf
947 * item's bitmap.
948 */
949 void
xfs_buf_item_log(struct xfs_buf_log_item * bip,uint first,uint last)950 xfs_buf_item_log(
951 struct xfs_buf_log_item *bip,
952 uint first,
953 uint last)
954 {
955 int i;
956 uint start;
957 uint end;
958 struct xfs_buf *bp = bip->bli_buf;
959
960 /*
961 * walk each buffer segment and mark them dirty appropriately.
962 */
963 start = 0;
964 for (i = 0; i < bip->bli_format_count; i++) {
965 if (start > last)
966 break;
967 end = start + BBTOB(bp->b_maps[i].bm_len) - 1;
968
969 /* skip to the map that includes the first byte to log */
970 if (first > end) {
971 start += BBTOB(bp->b_maps[i].bm_len);
972 continue;
973 }
974
975 /*
976 * Trim the range to this segment and mark it in the bitmap.
977 * Note that we must convert buffer offsets to segment relative
978 * offsets (e.g., the first byte of each segment is byte 0 of
979 * that segment).
980 */
981 if (first < start)
982 first = start;
983 if (end > last)
984 end = last;
985 xfs_buf_item_log_segment(first - start, end - start,
986 &bip->bli_formats[i].blf_data_map[0]);
987
988 start += BBTOB(bp->b_maps[i].bm_len);
989 }
990 }
991
992
993 /*
994 * Return true if the buffer has any ranges logged/dirtied by a transaction,
995 * false otherwise.
996 */
997 bool
xfs_buf_item_dirty_format(struct xfs_buf_log_item * bip)998 xfs_buf_item_dirty_format(
999 struct xfs_buf_log_item *bip)
1000 {
1001 int i;
1002
1003 for (i = 0; i < bip->bli_format_count; i++) {
1004 if (!xfs_bitmap_empty(bip->bli_formats[i].blf_data_map,
1005 bip->bli_formats[i].blf_map_size))
1006 return true;
1007 }
1008
1009 return false;
1010 }
1011
1012 STATIC void
xfs_buf_item_free(struct xfs_buf_log_item * bip)1013 xfs_buf_item_free(
1014 struct xfs_buf_log_item *bip)
1015 {
1016 xfs_buf_item_free_format(bip);
1017 kmem_free(bip->bli_item.li_lv_shadow);
1018 kmem_cache_free(xfs_buf_item_zone, bip);
1019 }
1020
1021 /*
1022 * xfs_buf_item_relse() is called when the buf log item is no longer needed.
1023 */
1024 void
xfs_buf_item_relse(struct xfs_buf * bp)1025 xfs_buf_item_relse(
1026 struct xfs_buf *bp)
1027 {
1028 struct xfs_buf_log_item *bip = bp->b_log_item;
1029
1030 trace_xfs_buf_item_relse(bp, _RET_IP_);
1031 ASSERT(!test_bit(XFS_LI_IN_AIL, &bip->bli_item.li_flags));
1032
1033 bp->b_log_item = NULL;
1034 xfs_buf_rele(bp);
1035 xfs_buf_item_free(bip);
1036 }
1037
1038 void
xfs_buf_item_done(struct xfs_buf * bp)1039 xfs_buf_item_done(
1040 struct xfs_buf *bp)
1041 {
1042 /*
1043 * If we are forcibly shutting down, this may well be off the AIL
1044 * already. That's because we simulate the log-committed callbacks to
1045 * unpin these buffers. Or we may never have put this item on AIL
1046 * because of the transaction was aborted forcibly.
1047 * xfs_trans_ail_delete() takes care of these.
1048 *
1049 * Either way, AIL is useless if we're forcing a shutdown.
1050 *
1051 * Note that log recovery writes might have buffer items that are not on
1052 * the AIL even when the file system is not shut down.
1053 */
1054 xfs_trans_ail_delete(&bp->b_log_item->bli_item,
1055 (bp->b_flags & _XBF_LOGRECOVERY) ? 0 :
1056 SHUTDOWN_CORRUPT_INCORE);
1057 xfs_buf_item_relse(bp);
1058 }
1059