WIP FPC-III support
[linux/fpc-iii.git] / fs / xfs / xfs_buf_item_recover.c
blobd44e8b4a33919d68ca9b58f2766da865e10b6829
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6 #include "xfs.h"
7 #include "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 #include "xfs_log_priv.h"
20 #include "xfs_log_recover.h"
21 #include "xfs_error.h"
22 #include "xfs_inode.h"
23 #include "xfs_dir2.h"
24 #include "xfs_quota.h"
27 * This structure is used during recovery to record the buf log items which
28 * have been canceled and should not be replayed.
30 struct xfs_buf_cancel {
31 xfs_daddr_t bc_blkno;
32 uint bc_len;
33 int bc_refcount;
34 struct list_head bc_list;
37 static struct xfs_buf_cancel *
38 xlog_find_buffer_cancelled(
39 struct xlog *log,
40 xfs_daddr_t blkno,
41 uint len)
43 struct list_head *bucket;
44 struct xfs_buf_cancel *bcp;
46 if (!log->l_buf_cancel_table)
47 return NULL;
49 bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
50 list_for_each_entry(bcp, bucket, bc_list) {
51 if (bcp->bc_blkno == blkno && bcp->bc_len == len)
52 return bcp;
55 return NULL;
58 static bool
59 xlog_add_buffer_cancelled(
60 struct xlog *log,
61 xfs_daddr_t blkno,
62 uint len)
64 struct xfs_buf_cancel *bcp;
67 * If we find an existing cancel record, this indicates that the buffer
68 * was cancelled multiple times. To ensure that during pass 2 we keep
69 * the record in the table until we reach its last occurrence in the
70 * log, a reference count is kept to tell how many times we expect to
71 * see this record during the second pass.
73 bcp = xlog_find_buffer_cancelled(log, blkno, len);
74 if (bcp) {
75 bcp->bc_refcount++;
76 return false;
79 bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), 0);
80 bcp->bc_blkno = blkno;
81 bcp->bc_len = len;
82 bcp->bc_refcount = 1;
83 list_add_tail(&bcp->bc_list, XLOG_BUF_CANCEL_BUCKET(log, blkno));
84 return true;
88 * Check if there is and entry for blkno, len in the buffer cancel record table.
90 bool
91 xlog_is_buffer_cancelled(
92 struct xlog *log,
93 xfs_daddr_t blkno,
94 uint len)
96 return xlog_find_buffer_cancelled(log, blkno, len) != NULL;
100 * Check if there is and entry for blkno, len in the buffer cancel record table,
101 * and decremented the reference count on it if there is one.
103 * Remove the cancel record once the refcount hits zero, so that if the same
104 * buffer is re-used again after its last cancellation we actually replay the
105 * changes made at that point.
107 static bool
108 xlog_put_buffer_cancelled(
109 struct xlog *log,
110 xfs_daddr_t blkno,
111 uint len)
113 struct xfs_buf_cancel *bcp;
115 bcp = xlog_find_buffer_cancelled(log, blkno, len);
116 if (!bcp) {
117 ASSERT(0);
118 return false;
121 if (--bcp->bc_refcount == 0) {
122 list_del(&bcp->bc_list);
123 kmem_free(bcp);
125 return true;
128 /* log buffer item recovery */
131 * Sort buffer items for log recovery. Most buffer items should end up on the
132 * buffer list and are recovered first, with the following exceptions:
134 * 1. XFS_BLF_CANCEL buffers must be processed last because some log items
135 * might depend on the incor ecancellation record, and replaying a cancelled
136 * buffer item can remove the incore record.
138 * 2. XFS_BLF_INODE_BUF buffers are handled after most regular items so that
139 * we replay di_next_unlinked only after flushing the inode 'free' state
140 * to the inode buffer.
142 * See xlog_recover_reorder_trans for more details.
144 STATIC enum xlog_recover_reorder
145 xlog_recover_buf_reorder(
146 struct xlog_recover_item *item)
148 struct xfs_buf_log_format *buf_f = item->ri_buf[0].i_addr;
150 if (buf_f->blf_flags & XFS_BLF_CANCEL)
151 return XLOG_REORDER_CANCEL_LIST;
152 if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
153 return XLOG_REORDER_INODE_BUFFER_LIST;
154 return XLOG_REORDER_BUFFER_LIST;
157 STATIC void
158 xlog_recover_buf_ra_pass2(
159 struct xlog *log,
160 struct xlog_recover_item *item)
162 struct xfs_buf_log_format *buf_f = item->ri_buf[0].i_addr;
164 xlog_buf_readahead(log, buf_f->blf_blkno, buf_f->blf_len, NULL);
168 * Build up the table of buf cancel records so that we don't replay cancelled
169 * data in the second pass.
171 static int
172 xlog_recover_buf_commit_pass1(
173 struct xlog *log,
174 struct xlog_recover_item *item)
176 struct xfs_buf_log_format *bf = item->ri_buf[0].i_addr;
178 if (!xfs_buf_log_check_iovec(&item->ri_buf[0])) {
179 xfs_err(log->l_mp, "bad buffer log item size (%d)",
180 item->ri_buf[0].i_len);
181 return -EFSCORRUPTED;
184 if (!(bf->blf_flags & XFS_BLF_CANCEL))
185 trace_xfs_log_recover_buf_not_cancel(log, bf);
186 else if (xlog_add_buffer_cancelled(log, bf->blf_blkno, bf->blf_len))
187 trace_xfs_log_recover_buf_cancel_add(log, bf);
188 else
189 trace_xfs_log_recover_buf_cancel_ref_inc(log, bf);
190 return 0;
194 * Validate the recovered buffer is of the correct type and attach the
195 * appropriate buffer operations to them for writeback. Magic numbers are in a
196 * few places:
197 * the first 16 bits of the buffer (inode buffer, dquot buffer),
198 * the first 32 bits of the buffer (most blocks),
199 * inside a struct xfs_da_blkinfo at the start of the buffer.
201 static void
202 xlog_recover_validate_buf_type(
203 struct xfs_mount *mp,
204 struct xfs_buf *bp,
205 struct xfs_buf_log_format *buf_f,
206 xfs_lsn_t current_lsn)
208 struct xfs_da_blkinfo *info = bp->b_addr;
209 uint32_t magic32;
210 uint16_t magic16;
211 uint16_t magicda;
212 char *warnmsg = NULL;
215 * We can only do post recovery validation on items on CRC enabled
216 * fielsystems as we need to know when the buffer was written to be able
217 * to determine if we should have replayed the item. If we replay old
218 * metadata over a newer buffer, then it will enter a temporarily
219 * inconsistent state resulting in verification failures. Hence for now
220 * just avoid the verification stage for non-crc filesystems
222 if (!xfs_sb_version_hascrc(&mp->m_sb))
223 return;
225 magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
226 magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
227 magicda = be16_to_cpu(info->magic);
228 switch (xfs_blft_from_flags(buf_f)) {
229 case XFS_BLFT_BTREE_BUF:
230 switch (magic32) {
231 case XFS_ABTB_CRC_MAGIC:
232 case XFS_ABTB_MAGIC:
233 bp->b_ops = &xfs_bnobt_buf_ops;
234 break;
235 case XFS_ABTC_CRC_MAGIC:
236 case XFS_ABTC_MAGIC:
237 bp->b_ops = &xfs_cntbt_buf_ops;
238 break;
239 case XFS_IBT_CRC_MAGIC:
240 case XFS_IBT_MAGIC:
241 bp->b_ops = &xfs_inobt_buf_ops;
242 break;
243 case XFS_FIBT_CRC_MAGIC:
244 case XFS_FIBT_MAGIC:
245 bp->b_ops = &xfs_finobt_buf_ops;
246 break;
247 case XFS_BMAP_CRC_MAGIC:
248 case XFS_BMAP_MAGIC:
249 bp->b_ops = &xfs_bmbt_buf_ops;
250 break;
251 case XFS_RMAP_CRC_MAGIC:
252 bp->b_ops = &xfs_rmapbt_buf_ops;
253 break;
254 case XFS_REFC_CRC_MAGIC:
255 bp->b_ops = &xfs_refcountbt_buf_ops;
256 break;
257 default:
258 warnmsg = "Bad btree block magic!";
259 break;
261 break;
262 case XFS_BLFT_AGF_BUF:
263 if (magic32 != XFS_AGF_MAGIC) {
264 warnmsg = "Bad AGF block magic!";
265 break;
267 bp->b_ops = &xfs_agf_buf_ops;
268 break;
269 case XFS_BLFT_AGFL_BUF:
270 if (magic32 != XFS_AGFL_MAGIC) {
271 warnmsg = "Bad AGFL block magic!";
272 break;
274 bp->b_ops = &xfs_agfl_buf_ops;
275 break;
276 case XFS_BLFT_AGI_BUF:
277 if (magic32 != XFS_AGI_MAGIC) {
278 warnmsg = "Bad AGI block magic!";
279 break;
281 bp->b_ops = &xfs_agi_buf_ops;
282 break;
283 case XFS_BLFT_UDQUOT_BUF:
284 case XFS_BLFT_PDQUOT_BUF:
285 case XFS_BLFT_GDQUOT_BUF:
286 #ifdef CONFIG_XFS_QUOTA
287 if (magic16 != XFS_DQUOT_MAGIC) {
288 warnmsg = "Bad DQUOT block magic!";
289 break;
291 bp->b_ops = &xfs_dquot_buf_ops;
292 #else
293 xfs_alert(mp,
294 "Trying to recover dquots without QUOTA support built in!");
295 ASSERT(0);
296 #endif
297 break;
298 case XFS_BLFT_DINO_BUF:
299 if (magic16 != XFS_DINODE_MAGIC) {
300 warnmsg = "Bad INODE block magic!";
301 break;
303 bp->b_ops = &xfs_inode_buf_ops;
304 break;
305 case XFS_BLFT_SYMLINK_BUF:
306 if (magic32 != XFS_SYMLINK_MAGIC) {
307 warnmsg = "Bad symlink block magic!";
308 break;
310 bp->b_ops = &xfs_symlink_buf_ops;
311 break;
312 case XFS_BLFT_DIR_BLOCK_BUF:
313 if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
314 magic32 != XFS_DIR3_BLOCK_MAGIC) {
315 warnmsg = "Bad dir block magic!";
316 break;
318 bp->b_ops = &xfs_dir3_block_buf_ops;
319 break;
320 case XFS_BLFT_DIR_DATA_BUF:
321 if (magic32 != XFS_DIR2_DATA_MAGIC &&
322 magic32 != XFS_DIR3_DATA_MAGIC) {
323 warnmsg = "Bad dir data magic!";
324 break;
326 bp->b_ops = &xfs_dir3_data_buf_ops;
327 break;
328 case XFS_BLFT_DIR_FREE_BUF:
329 if (magic32 != XFS_DIR2_FREE_MAGIC &&
330 magic32 != XFS_DIR3_FREE_MAGIC) {
331 warnmsg = "Bad dir3 free magic!";
332 break;
334 bp->b_ops = &xfs_dir3_free_buf_ops;
335 break;
336 case XFS_BLFT_DIR_LEAF1_BUF:
337 if (magicda != XFS_DIR2_LEAF1_MAGIC &&
338 magicda != XFS_DIR3_LEAF1_MAGIC) {
339 warnmsg = "Bad dir leaf1 magic!";
340 break;
342 bp->b_ops = &xfs_dir3_leaf1_buf_ops;
343 break;
344 case XFS_BLFT_DIR_LEAFN_BUF:
345 if (magicda != XFS_DIR2_LEAFN_MAGIC &&
346 magicda != XFS_DIR3_LEAFN_MAGIC) {
347 warnmsg = "Bad dir leafn magic!";
348 break;
350 bp->b_ops = &xfs_dir3_leafn_buf_ops;
351 break;
352 case XFS_BLFT_DA_NODE_BUF:
353 if (magicda != XFS_DA_NODE_MAGIC &&
354 magicda != XFS_DA3_NODE_MAGIC) {
355 warnmsg = "Bad da node magic!";
356 break;
358 bp->b_ops = &xfs_da3_node_buf_ops;
359 break;
360 case XFS_BLFT_ATTR_LEAF_BUF:
361 if (magicda != XFS_ATTR_LEAF_MAGIC &&
362 magicda != XFS_ATTR3_LEAF_MAGIC) {
363 warnmsg = "Bad attr leaf magic!";
364 break;
366 bp->b_ops = &xfs_attr3_leaf_buf_ops;
367 break;
368 case XFS_BLFT_ATTR_RMT_BUF:
369 if (magic32 != XFS_ATTR3_RMT_MAGIC) {
370 warnmsg = "Bad attr remote magic!";
371 break;
373 bp->b_ops = &xfs_attr3_rmt_buf_ops;
374 break;
375 case XFS_BLFT_SB_BUF:
376 if (magic32 != XFS_SB_MAGIC) {
377 warnmsg = "Bad SB block magic!";
378 break;
380 bp->b_ops = &xfs_sb_buf_ops;
381 break;
382 #ifdef CONFIG_XFS_RT
383 case XFS_BLFT_RTBITMAP_BUF:
384 case XFS_BLFT_RTSUMMARY_BUF:
385 /* no magic numbers for verification of RT buffers */
386 bp->b_ops = &xfs_rtbuf_ops;
387 break;
388 #endif /* CONFIG_XFS_RT */
389 default:
390 xfs_warn(mp, "Unknown buffer type %d!",
391 xfs_blft_from_flags(buf_f));
392 break;
396 * Nothing else to do in the case of a NULL current LSN as this means
397 * the buffer is more recent than the change in the log and will be
398 * skipped.
400 if (current_lsn == NULLCOMMITLSN)
401 return;
403 if (warnmsg) {
404 xfs_warn(mp, warnmsg);
405 ASSERT(0);
409 * We must update the metadata LSN of the buffer as it is written out to
410 * ensure that older transactions never replay over this one and corrupt
411 * the buffer. This can occur if log recovery is interrupted at some
412 * point after the current transaction completes, at which point a
413 * subsequent mount starts recovery from the beginning.
415 * Write verifiers update the metadata LSN from log items attached to
416 * the buffer. Therefore, initialize a bli purely to carry the LSN to
417 * the verifier.
419 if (bp->b_ops) {
420 struct xfs_buf_log_item *bip;
422 bp->b_flags |= _XBF_LOGRECOVERY;
423 xfs_buf_item_init(bp, mp);
424 bip = bp->b_log_item;
425 bip->bli_item.li_lsn = current_lsn;
430 * Perform a 'normal' buffer recovery. Each logged region of the
431 * buffer should be copied over the corresponding region in the
432 * given buffer. The bitmap in the buf log format structure indicates
433 * where to place the logged data.
435 STATIC void
436 xlog_recover_do_reg_buffer(
437 struct xfs_mount *mp,
438 struct xlog_recover_item *item,
439 struct xfs_buf *bp,
440 struct xfs_buf_log_format *buf_f,
441 xfs_lsn_t current_lsn)
443 int i;
444 int bit;
445 int nbits;
446 xfs_failaddr_t fa;
447 const size_t size_disk_dquot = sizeof(struct xfs_disk_dquot);
449 trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
451 bit = 0;
452 i = 1; /* 0 is the buf format structure */
453 while (1) {
454 bit = xfs_next_bit(buf_f->blf_data_map,
455 buf_f->blf_map_size, bit);
456 if (bit == -1)
457 break;
458 nbits = xfs_contig_bits(buf_f->blf_data_map,
459 buf_f->blf_map_size, bit);
460 ASSERT(nbits > 0);
461 ASSERT(item->ri_buf[i].i_addr != NULL);
462 ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
463 ASSERT(BBTOB(bp->b_length) >=
464 ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
467 * The dirty regions logged in the buffer, even though
468 * contiguous, may span multiple chunks. This is because the
469 * dirty region may span a physical page boundary in a buffer
470 * and hence be split into two separate vectors for writing into
471 * the log. Hence we need to trim nbits back to the length of
472 * the current region being copied out of the log.
474 if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
475 nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
478 * Do a sanity check if this is a dquot buffer. Just checking
479 * the first dquot in the buffer should do. XXXThis is
480 * probably a good thing to do for other buf types also.
482 fa = NULL;
483 if (buf_f->blf_flags &
484 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
485 if (item->ri_buf[i].i_addr == NULL) {
486 xfs_alert(mp,
487 "XFS: NULL dquot in %s.", __func__);
488 goto next;
490 if (item->ri_buf[i].i_len < size_disk_dquot) {
491 xfs_alert(mp,
492 "XFS: dquot too small (%d) in %s.",
493 item->ri_buf[i].i_len, __func__);
494 goto next;
496 fa = xfs_dquot_verify(mp, item->ri_buf[i].i_addr, -1);
497 if (fa) {
498 xfs_alert(mp,
499 "dquot corrupt at %pS trying to replay into block 0x%llx",
500 fa, bp->b_bn);
501 goto next;
505 memcpy(xfs_buf_offset(bp,
506 (uint)bit << XFS_BLF_SHIFT), /* dest */
507 item->ri_buf[i].i_addr, /* source */
508 nbits<<XFS_BLF_SHIFT); /* length */
509 next:
510 i++;
511 bit += nbits;
514 /* Shouldn't be any more regions */
515 ASSERT(i == item->ri_total);
517 xlog_recover_validate_buf_type(mp, bp, buf_f, current_lsn);
521 * Perform a dquot buffer recovery.
522 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
523 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
524 * Else, treat it as a regular buffer and do recovery.
526 * Return false if the buffer was tossed and true if we recovered the buffer to
527 * indicate to the caller if the buffer needs writing.
529 STATIC bool
530 xlog_recover_do_dquot_buffer(
531 struct xfs_mount *mp,
532 struct xlog *log,
533 struct xlog_recover_item *item,
534 struct xfs_buf *bp,
535 struct xfs_buf_log_format *buf_f)
537 uint type;
539 trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
542 * Filesystems are required to send in quota flags at mount time.
544 if (!mp->m_qflags)
545 return false;
547 type = 0;
548 if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
549 type |= XFS_DQTYPE_USER;
550 if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
551 type |= XFS_DQTYPE_PROJ;
552 if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
553 type |= XFS_DQTYPE_GROUP;
555 * This type of quotas was turned off, so ignore this buffer
557 if (log->l_quotaoffs_flag & type)
558 return false;
560 xlog_recover_do_reg_buffer(mp, item, bp, buf_f, NULLCOMMITLSN);
561 return true;
565 * Perform recovery for a buffer full of inodes. In these buffers, the only
566 * data which should be recovered is that which corresponds to the
567 * di_next_unlinked pointers in the on disk inode structures. The rest of the
568 * data for the inodes is always logged through the inodes themselves rather
569 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
571 * The only time when buffers full of inodes are fully recovered is when the
572 * buffer is full of newly allocated inodes. In this case the buffer will
573 * not be marked as an inode buffer and so will be sent to
574 * xlog_recover_do_reg_buffer() below during recovery.
576 STATIC int
577 xlog_recover_do_inode_buffer(
578 struct xfs_mount *mp,
579 struct xlog_recover_item *item,
580 struct xfs_buf *bp,
581 struct xfs_buf_log_format *buf_f)
583 int i;
584 int item_index = 0;
585 int bit = 0;
586 int nbits = 0;
587 int reg_buf_offset = 0;
588 int reg_buf_bytes = 0;
589 int next_unlinked_offset;
590 int inodes_per_buf;
591 xfs_agino_t *logged_nextp;
592 xfs_agino_t *buffer_nextp;
594 trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
597 * Post recovery validation only works properly on CRC enabled
598 * filesystems.
600 if (xfs_sb_version_hascrc(&mp->m_sb))
601 bp->b_ops = &xfs_inode_buf_ops;
603 inodes_per_buf = BBTOB(bp->b_length) >> mp->m_sb.sb_inodelog;
604 for (i = 0; i < inodes_per_buf; i++) {
605 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
606 offsetof(xfs_dinode_t, di_next_unlinked);
608 while (next_unlinked_offset >=
609 (reg_buf_offset + reg_buf_bytes)) {
611 * The next di_next_unlinked field is beyond
612 * the current logged region. Find the next
613 * logged region that contains or is beyond
614 * the current di_next_unlinked field.
616 bit += nbits;
617 bit = xfs_next_bit(buf_f->blf_data_map,
618 buf_f->blf_map_size, bit);
621 * If there are no more logged regions in the
622 * buffer, then we're done.
624 if (bit == -1)
625 return 0;
627 nbits = xfs_contig_bits(buf_f->blf_data_map,
628 buf_f->blf_map_size, bit);
629 ASSERT(nbits > 0);
630 reg_buf_offset = bit << XFS_BLF_SHIFT;
631 reg_buf_bytes = nbits << XFS_BLF_SHIFT;
632 item_index++;
636 * If the current logged region starts after the current
637 * di_next_unlinked field, then move on to the next
638 * di_next_unlinked field.
640 if (next_unlinked_offset < reg_buf_offset)
641 continue;
643 ASSERT(item->ri_buf[item_index].i_addr != NULL);
644 ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
645 ASSERT((reg_buf_offset + reg_buf_bytes) <= BBTOB(bp->b_length));
648 * The current logged region contains a copy of the
649 * current di_next_unlinked field. Extract its value
650 * and copy it to the buffer copy.
652 logged_nextp = item->ri_buf[item_index].i_addr +
653 next_unlinked_offset - reg_buf_offset;
654 if (XFS_IS_CORRUPT(mp, *logged_nextp == 0)) {
655 xfs_alert(mp,
656 "Bad inode buffer log record (ptr = "PTR_FMT", bp = "PTR_FMT"). "
657 "Trying to replay bad (0) inode di_next_unlinked field.",
658 item, bp);
659 return -EFSCORRUPTED;
662 buffer_nextp = xfs_buf_offset(bp, next_unlinked_offset);
663 *buffer_nextp = *logged_nextp;
666 * If necessary, recalculate the CRC in the on-disk inode. We
667 * have to leave the inode in a consistent state for whoever
668 * reads it next....
670 xfs_dinode_calc_crc(mp,
671 xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
675 return 0;
679 * V5 filesystems know the age of the buffer on disk being recovered. We can
680 * have newer objects on disk than we are replaying, and so for these cases we
681 * don't want to replay the current change as that will make the buffer contents
682 * temporarily invalid on disk.
684 * The magic number might not match the buffer type we are going to recover
685 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags. Hence
686 * extract the LSN of the existing object in the buffer based on it's current
687 * magic number. If we don't recognise the magic number in the buffer, then
688 * return a LSN of -1 so that the caller knows it was an unrecognised block and
689 * so can recover the buffer.
691 * Note: we cannot rely solely on magic number matches to determine that the
692 * buffer has a valid LSN - we also need to verify that it belongs to this
693 * filesystem, so we need to extract the object's LSN and compare it to that
694 * which we read from the superblock. If the UUIDs don't match, then we've got a
695 * stale metadata block from an old filesystem instance that we need to recover
696 * over the top of.
698 static xfs_lsn_t
699 xlog_recover_get_buf_lsn(
700 struct xfs_mount *mp,
701 struct xfs_buf *bp)
703 uint32_t magic32;
704 uint16_t magic16;
705 uint16_t magicda;
706 void *blk = bp->b_addr;
707 uuid_t *uuid;
708 xfs_lsn_t lsn = -1;
710 /* v4 filesystems always recover immediately */
711 if (!xfs_sb_version_hascrc(&mp->m_sb))
712 goto recover_immediately;
714 magic32 = be32_to_cpu(*(__be32 *)blk);
715 switch (magic32) {
716 case XFS_ABTB_CRC_MAGIC:
717 case XFS_ABTC_CRC_MAGIC:
718 case XFS_ABTB_MAGIC:
719 case XFS_ABTC_MAGIC:
720 case XFS_RMAP_CRC_MAGIC:
721 case XFS_REFC_CRC_MAGIC:
722 case XFS_FIBT_CRC_MAGIC:
723 case XFS_FIBT_MAGIC:
724 case XFS_IBT_CRC_MAGIC:
725 case XFS_IBT_MAGIC: {
726 struct xfs_btree_block *btb = blk;
728 lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
729 uuid = &btb->bb_u.s.bb_uuid;
730 break;
732 case XFS_BMAP_CRC_MAGIC:
733 case XFS_BMAP_MAGIC: {
734 struct xfs_btree_block *btb = blk;
736 lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
737 uuid = &btb->bb_u.l.bb_uuid;
738 break;
740 case XFS_AGF_MAGIC:
741 lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
742 uuid = &((struct xfs_agf *)blk)->agf_uuid;
743 break;
744 case XFS_AGFL_MAGIC:
745 lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
746 uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
747 break;
748 case XFS_AGI_MAGIC:
749 lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
750 uuid = &((struct xfs_agi *)blk)->agi_uuid;
751 break;
752 case XFS_SYMLINK_MAGIC:
753 lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
754 uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
755 break;
756 case XFS_DIR3_BLOCK_MAGIC:
757 case XFS_DIR3_DATA_MAGIC:
758 case XFS_DIR3_FREE_MAGIC:
759 lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
760 uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
761 break;
762 case XFS_ATTR3_RMT_MAGIC:
764 * Remote attr blocks are written synchronously, rather than
765 * being logged. That means they do not contain a valid LSN
766 * (i.e. transactionally ordered) in them, and hence any time we
767 * see a buffer to replay over the top of a remote attribute
768 * block we should simply do so.
770 goto recover_immediately;
771 case XFS_SB_MAGIC:
773 * superblock uuids are magic. We may or may not have a
774 * sb_meta_uuid on disk, but it will be set in the in-core
775 * superblock. We set the uuid pointer for verification
776 * according to the superblock feature mask to ensure we check
777 * the relevant UUID in the superblock.
779 lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
780 if (xfs_sb_version_hasmetauuid(&mp->m_sb))
781 uuid = &((struct xfs_dsb *)blk)->sb_meta_uuid;
782 else
783 uuid = &((struct xfs_dsb *)blk)->sb_uuid;
784 break;
785 default:
786 break;
789 if (lsn != (xfs_lsn_t)-1) {
790 if (!uuid_equal(&mp->m_sb.sb_meta_uuid, uuid))
791 goto recover_immediately;
792 return lsn;
795 magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
796 switch (magicda) {
797 case XFS_DIR3_LEAF1_MAGIC:
798 case XFS_DIR3_LEAFN_MAGIC:
799 case XFS_DA3_NODE_MAGIC:
800 lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
801 uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
802 break;
803 default:
804 break;
807 if (lsn != (xfs_lsn_t)-1) {
808 if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
809 goto recover_immediately;
810 return lsn;
814 * We do individual object checks on dquot and inode buffers as they
815 * have their own individual LSN records. Also, we could have a stale
816 * buffer here, so we have to at least recognise these buffer types.
818 * A notd complexity here is inode unlinked list processing - it logs
819 * the inode directly in the buffer, but we don't know which inodes have
820 * been modified, and there is no global buffer LSN. Hence we need to
821 * recover all inode buffer types immediately. This problem will be
822 * fixed by logical logging of the unlinked list modifications.
824 magic16 = be16_to_cpu(*(__be16 *)blk);
825 switch (magic16) {
826 case XFS_DQUOT_MAGIC:
827 case XFS_DINODE_MAGIC:
828 goto recover_immediately;
829 default:
830 break;
833 /* unknown buffer contents, recover immediately */
835 recover_immediately:
836 return (xfs_lsn_t)-1;
841 * This routine replays a modification made to a buffer at runtime.
842 * There are actually two types of buffer, regular and inode, which
843 * are handled differently. Inode buffers are handled differently
844 * in that we only recover a specific set of data from them, namely
845 * the inode di_next_unlinked fields. This is because all other inode
846 * data is actually logged via inode records and any data we replay
847 * here which overlaps that may be stale.
849 * When meta-data buffers are freed at run time we log a buffer item
850 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
851 * of the buffer in the log should not be replayed at recovery time.
852 * This is so that if the blocks covered by the buffer are reused for
853 * file data before we crash we don't end up replaying old, freed
854 * meta-data into a user's file.
856 * To handle the cancellation of buffer log items, we make two passes
857 * over the log during recovery. During the first we build a table of
858 * those buffers which have been cancelled, and during the second we
859 * only replay those buffers which do not have corresponding cancel
860 * records in the table. See xlog_recover_buf_pass[1,2] above
861 * for more details on the implementation of the table of cancel records.
863 STATIC int
864 xlog_recover_buf_commit_pass2(
865 struct xlog *log,
866 struct list_head *buffer_list,
867 struct xlog_recover_item *item,
868 xfs_lsn_t current_lsn)
870 struct xfs_buf_log_format *buf_f = item->ri_buf[0].i_addr;
871 struct xfs_mount *mp = log->l_mp;
872 struct xfs_buf *bp;
873 int error;
874 uint buf_flags;
875 xfs_lsn_t lsn;
878 * In this pass we only want to recover all the buffers which have
879 * not been cancelled and are not cancellation buffers themselves.
881 if (buf_f->blf_flags & XFS_BLF_CANCEL) {
882 if (xlog_put_buffer_cancelled(log, buf_f->blf_blkno,
883 buf_f->blf_len))
884 goto cancelled;
885 } else {
887 if (xlog_is_buffer_cancelled(log, buf_f->blf_blkno,
888 buf_f->blf_len))
889 goto cancelled;
892 trace_xfs_log_recover_buf_recover(log, buf_f);
894 buf_flags = 0;
895 if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
896 buf_flags |= XBF_UNMAPPED;
898 error = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
899 buf_flags, &bp, NULL);
900 if (error)
901 return error;
904 * Recover the buffer only if we get an LSN from it and it's less than
905 * the lsn of the transaction we are replaying.
907 * Note that we have to be extremely careful of readahead here.
908 * Readahead does not attach verfiers to the buffers so if we don't
909 * actually do any replay after readahead because of the LSN we found
910 * in the buffer if more recent than that current transaction then we
911 * need to attach the verifier directly. Failure to do so can lead to
912 * future recovery actions (e.g. EFI and unlinked list recovery) can
913 * operate on the buffers and they won't get the verifier attached. This
914 * can lead to blocks on disk having the correct content but a stale
915 * CRC.
917 * It is safe to assume these clean buffers are currently up to date.
918 * If the buffer is dirtied by a later transaction being replayed, then
919 * the verifier will be reset to match whatever recover turns that
920 * buffer into.
922 lsn = xlog_recover_get_buf_lsn(mp, bp);
923 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
924 trace_xfs_log_recover_buf_skip(log, buf_f);
925 xlog_recover_validate_buf_type(mp, bp, buf_f, NULLCOMMITLSN);
926 goto out_release;
929 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
930 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
931 if (error)
932 goto out_release;
933 } else if (buf_f->blf_flags &
934 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
935 bool dirty;
937 dirty = xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
938 if (!dirty)
939 goto out_release;
940 } else {
941 xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn);
945 * Perform delayed write on the buffer. Asynchronous writes will be
946 * slower when taking into account all the buffers to be flushed.
948 * Also make sure that only inode buffers with good sizes stay in
949 * the buffer cache. The kernel moves inodes in buffers of 1 block
950 * or inode_cluster_size bytes, whichever is bigger. The inode
951 * buffers in the log can be a different size if the log was generated
952 * by an older kernel using unclustered inode buffers or a newer kernel
953 * running with a different inode cluster size. Regardless, if
954 * the inode buffer size isn't max(blocksize, inode_cluster_size)
955 * for *our* value of inode_cluster_size, then we need to keep
956 * the buffer out of the buffer cache so that the buffer won't
957 * overlap with future reads of those inodes.
959 if (XFS_DINODE_MAGIC ==
960 be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
961 (BBTOB(bp->b_length) != M_IGEO(log->l_mp)->inode_cluster_size)) {
962 xfs_buf_stale(bp);
963 error = xfs_bwrite(bp);
964 } else {
965 ASSERT(bp->b_mount == mp);
966 bp->b_flags |= _XBF_LOGRECOVERY;
967 xfs_buf_delwri_queue(bp, buffer_list);
970 out_release:
971 xfs_buf_relse(bp);
972 return error;
973 cancelled:
974 trace_xfs_log_recover_buf_cancel(log, buf_f);
975 return 0;
978 const struct xlog_recover_item_ops xlog_buf_item_ops = {
979 .item_type = XFS_LI_BUF,
980 .reorder = xlog_recover_buf_reorder,
981 .ra_pass2 = xlog_recover_buf_ra_pass2,
982 .commit_pass1 = xlog_recover_buf_commit_pass1,
983 .commit_pass2 = xlog_recover_buf_commit_pass2,