Loading...
1/*
2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#include "xfs.h"
19#include "xfs_fs.h"
20#include "xfs_types.h"
21#include "xfs_bit.h"
22#include "xfs_log.h"
23#include "xfs_inum.h"
24#include "xfs_trans.h"
25#include "xfs_sb.h"
26#include "xfs_ag.h"
27#include "xfs_mount.h"
28#include "xfs_error.h"
29#include "xfs_bmap_btree.h"
30#include "xfs_alloc_btree.h"
31#include "xfs_ialloc_btree.h"
32#include "xfs_dinode.h"
33#include "xfs_inode.h"
34#include "xfs_inode_item.h"
35#include "xfs_alloc.h"
36#include "xfs_ialloc.h"
37#include "xfs_log_priv.h"
38#include "xfs_buf_item.h"
39#include "xfs_log_recover.h"
40#include "xfs_extfree_item.h"
41#include "xfs_trans_priv.h"
42#include "xfs_quota.h"
43#include "xfs_rw.h"
44#include "xfs_utils.h"
45#include "xfs_trace.h"
46
47STATIC int xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
48STATIC int xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
49#if defined(DEBUG)
50STATIC void xlog_recover_check_summary(xlog_t *);
51#else
52#define xlog_recover_check_summary(log)
53#endif
54
55/*
56 * This structure is used during recovery to record the buf log items which
57 * have been canceled and should not be replayed.
58 */
59struct xfs_buf_cancel {
60 xfs_daddr_t bc_blkno;
61 uint bc_len;
62 int bc_refcount;
63 struct list_head bc_list;
64};
65
66/*
67 * Sector aligned buffer routines for buffer create/read/write/access
68 */
69
70/*
71 * Verify the given count of basic blocks is valid number of blocks
72 * to specify for an operation involving the given XFS log buffer.
73 * Returns nonzero if the count is valid, 0 otherwise.
74 */
75
76static inline int
77xlog_buf_bbcount_valid(
78 xlog_t *log,
79 int bbcount)
80{
81 return bbcount > 0 && bbcount <= log->l_logBBsize;
82}
83
84/*
85 * Allocate a buffer to hold log data. The buffer needs to be able
86 * to map to a range of nbblks basic blocks at any valid (basic
87 * block) offset within the log.
88 */
89STATIC xfs_buf_t *
90xlog_get_bp(
91 xlog_t *log,
92 int nbblks)
93{
94 struct xfs_buf *bp;
95
96 if (!xlog_buf_bbcount_valid(log, nbblks)) {
97 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
98 nbblks);
99 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
100 return NULL;
101 }
102
103 /*
104 * We do log I/O in units of log sectors (a power-of-2
105 * multiple of the basic block size), so we round up the
106 * requested size to accommodate the basic blocks required
107 * for complete log sectors.
108 *
109 * In addition, the buffer may be used for a non-sector-
110 * aligned block offset, in which case an I/O of the
111 * requested size could extend beyond the end of the
112 * buffer. If the requested size is only 1 basic block it
113 * will never straddle a sector boundary, so this won't be
114 * an issue. Nor will this be a problem if the log I/O is
115 * done in basic blocks (sector size 1). But otherwise we
116 * extend the buffer by one extra log sector to ensure
117 * there's space to accommodate this possibility.
118 */
119 if (nbblks > 1 && log->l_sectBBsize > 1)
120 nbblks += log->l_sectBBsize;
121 nbblks = round_up(nbblks, log->l_sectBBsize);
122
123 bp = xfs_buf_get_uncached(log->l_mp->m_logdev_targp, BBTOB(nbblks), 0);
124 if (bp)
125 xfs_buf_unlock(bp);
126 return bp;
127}
128
129STATIC void
130xlog_put_bp(
131 xfs_buf_t *bp)
132{
133 xfs_buf_free(bp);
134}
135
136/*
137 * Return the address of the start of the given block number's data
138 * in a log buffer. The buffer covers a log sector-aligned region.
139 */
140STATIC xfs_caddr_t
141xlog_align(
142 xlog_t *log,
143 xfs_daddr_t blk_no,
144 int nbblks,
145 xfs_buf_t *bp)
146{
147 xfs_daddr_t offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
148
149 ASSERT(BBTOB(offset + nbblks) <= XFS_BUF_SIZE(bp));
150 return bp->b_addr + BBTOB(offset);
151}
152
153
154/*
155 * nbblks should be uint, but oh well. Just want to catch that 32-bit length.
156 */
157STATIC int
158xlog_bread_noalign(
159 xlog_t *log,
160 xfs_daddr_t blk_no,
161 int nbblks,
162 xfs_buf_t *bp)
163{
164 int error;
165
166 if (!xlog_buf_bbcount_valid(log, nbblks)) {
167 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
168 nbblks);
169 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
170 return EFSCORRUPTED;
171 }
172
173 blk_no = round_down(blk_no, log->l_sectBBsize);
174 nbblks = round_up(nbblks, log->l_sectBBsize);
175
176 ASSERT(nbblks > 0);
177 ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
178
179 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
180 XFS_BUF_READ(bp);
181 XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
182
183 xfsbdstrat(log->l_mp, bp);
184 error = xfs_buf_iowait(bp);
185 if (error)
186 xfs_ioerror_alert("xlog_bread", log->l_mp,
187 bp, XFS_BUF_ADDR(bp));
188 return error;
189}
190
191STATIC int
192xlog_bread(
193 xlog_t *log,
194 xfs_daddr_t blk_no,
195 int nbblks,
196 xfs_buf_t *bp,
197 xfs_caddr_t *offset)
198{
199 int error;
200
201 error = xlog_bread_noalign(log, blk_no, nbblks, bp);
202 if (error)
203 return error;
204
205 *offset = xlog_align(log, blk_no, nbblks, bp);
206 return 0;
207}
208
209/*
210 * Read at an offset into the buffer. Returns with the buffer in it's original
211 * state regardless of the result of the read.
212 */
213STATIC int
214xlog_bread_offset(
215 xlog_t *log,
216 xfs_daddr_t blk_no, /* block to read from */
217 int nbblks, /* blocks to read */
218 xfs_buf_t *bp,
219 xfs_caddr_t offset)
220{
221 xfs_caddr_t orig_offset = bp->b_addr;
222 int orig_len = bp->b_buffer_length;
223 int error, error2;
224
225 error = xfs_buf_associate_memory(bp, offset, BBTOB(nbblks));
226 if (error)
227 return error;
228
229 error = xlog_bread_noalign(log, blk_no, nbblks, bp);
230
231 /* must reset buffer pointer even on error */
232 error2 = xfs_buf_associate_memory(bp, orig_offset, orig_len);
233 if (error)
234 return error;
235 return error2;
236}
237
238/*
239 * Write out the buffer at the given block for the given number of blocks.
240 * The buffer is kept locked across the write and is returned locked.
241 * This can only be used for synchronous log writes.
242 */
243STATIC int
244xlog_bwrite(
245 xlog_t *log,
246 xfs_daddr_t blk_no,
247 int nbblks,
248 xfs_buf_t *bp)
249{
250 int error;
251
252 if (!xlog_buf_bbcount_valid(log, nbblks)) {
253 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
254 nbblks);
255 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
256 return EFSCORRUPTED;
257 }
258
259 blk_no = round_down(blk_no, log->l_sectBBsize);
260 nbblks = round_up(nbblks, log->l_sectBBsize);
261
262 ASSERT(nbblks > 0);
263 ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
264
265 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
266 XFS_BUF_ZEROFLAGS(bp);
267 xfs_buf_hold(bp);
268 xfs_buf_lock(bp);
269 XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
270
271 if ((error = xfs_bwrite(log->l_mp, bp)))
272 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
273 bp, XFS_BUF_ADDR(bp));
274 return error;
275}
276
277#ifdef DEBUG
278/*
279 * dump debug superblock and log record information
280 */
281STATIC void
282xlog_header_check_dump(
283 xfs_mount_t *mp,
284 xlog_rec_header_t *head)
285{
286 xfs_debug(mp, "%s: SB : uuid = %pU, fmt = %d\n",
287 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
288 xfs_debug(mp, " log : uuid = %pU, fmt = %d\n",
289 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
290}
291#else
292#define xlog_header_check_dump(mp, head)
293#endif
294
295/*
296 * check log record header for recovery
297 */
298STATIC int
299xlog_header_check_recover(
300 xfs_mount_t *mp,
301 xlog_rec_header_t *head)
302{
303 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
304
305 /*
306 * IRIX doesn't write the h_fmt field and leaves it zeroed
307 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
308 * a dirty log created in IRIX.
309 */
310 if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
311 xfs_warn(mp,
312 "dirty log written in incompatible format - can't recover");
313 xlog_header_check_dump(mp, head);
314 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
315 XFS_ERRLEVEL_HIGH, mp);
316 return XFS_ERROR(EFSCORRUPTED);
317 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
318 xfs_warn(mp,
319 "dirty log entry has mismatched uuid - can't recover");
320 xlog_header_check_dump(mp, head);
321 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
322 XFS_ERRLEVEL_HIGH, mp);
323 return XFS_ERROR(EFSCORRUPTED);
324 }
325 return 0;
326}
327
328/*
329 * read the head block of the log and check the header
330 */
331STATIC int
332xlog_header_check_mount(
333 xfs_mount_t *mp,
334 xlog_rec_header_t *head)
335{
336 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
337
338 if (uuid_is_nil(&head->h_fs_uuid)) {
339 /*
340 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
341 * h_fs_uuid is nil, we assume this log was last mounted
342 * by IRIX and continue.
343 */
344 xfs_warn(mp, "nil uuid in log - IRIX style log");
345 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
346 xfs_warn(mp, "log has mismatched uuid - can't recover");
347 xlog_header_check_dump(mp, head);
348 XFS_ERROR_REPORT("xlog_header_check_mount",
349 XFS_ERRLEVEL_HIGH, mp);
350 return XFS_ERROR(EFSCORRUPTED);
351 }
352 return 0;
353}
354
355STATIC void
356xlog_recover_iodone(
357 struct xfs_buf *bp)
358{
359 if (bp->b_error) {
360 /*
361 * We're not going to bother about retrying
362 * this during recovery. One strike!
363 */
364 xfs_ioerror_alert("xlog_recover_iodone",
365 bp->b_target->bt_mount, bp,
366 XFS_BUF_ADDR(bp));
367 xfs_force_shutdown(bp->b_target->bt_mount,
368 SHUTDOWN_META_IO_ERROR);
369 }
370 bp->b_iodone = NULL;
371 xfs_buf_ioend(bp, 0);
372}
373
374/*
375 * This routine finds (to an approximation) the first block in the physical
376 * log which contains the given cycle. It uses a binary search algorithm.
377 * Note that the algorithm can not be perfect because the disk will not
378 * necessarily be perfect.
379 */
380STATIC int
381xlog_find_cycle_start(
382 xlog_t *log,
383 xfs_buf_t *bp,
384 xfs_daddr_t first_blk,
385 xfs_daddr_t *last_blk,
386 uint cycle)
387{
388 xfs_caddr_t offset;
389 xfs_daddr_t mid_blk;
390 xfs_daddr_t end_blk;
391 uint mid_cycle;
392 int error;
393
394 end_blk = *last_blk;
395 mid_blk = BLK_AVG(first_blk, end_blk);
396 while (mid_blk != first_blk && mid_blk != end_blk) {
397 error = xlog_bread(log, mid_blk, 1, bp, &offset);
398 if (error)
399 return error;
400 mid_cycle = xlog_get_cycle(offset);
401 if (mid_cycle == cycle)
402 end_blk = mid_blk; /* last_half_cycle == mid_cycle */
403 else
404 first_blk = mid_blk; /* first_half_cycle == mid_cycle */
405 mid_blk = BLK_AVG(first_blk, end_blk);
406 }
407 ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
408 (mid_blk == end_blk && mid_blk-1 == first_blk));
409
410 *last_blk = end_blk;
411
412 return 0;
413}
414
415/*
416 * Check that a range of blocks does not contain stop_on_cycle_no.
417 * Fill in *new_blk with the block offset where such a block is
418 * found, or with -1 (an invalid block number) if there is no such
419 * block in the range. The scan needs to occur from front to back
420 * and the pointer into the region must be updated since a later
421 * routine will need to perform another test.
422 */
423STATIC int
424xlog_find_verify_cycle(
425 xlog_t *log,
426 xfs_daddr_t start_blk,
427 int nbblks,
428 uint stop_on_cycle_no,
429 xfs_daddr_t *new_blk)
430{
431 xfs_daddr_t i, j;
432 uint cycle;
433 xfs_buf_t *bp;
434 xfs_daddr_t bufblks;
435 xfs_caddr_t buf = NULL;
436 int error = 0;
437
438 /*
439 * Greedily allocate a buffer big enough to handle the full
440 * range of basic blocks we'll be examining. If that fails,
441 * try a smaller size. We need to be able to read at least
442 * a log sector, or we're out of luck.
443 */
444 bufblks = 1 << ffs(nbblks);
445 while (!(bp = xlog_get_bp(log, bufblks))) {
446 bufblks >>= 1;
447 if (bufblks < log->l_sectBBsize)
448 return ENOMEM;
449 }
450
451 for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
452 int bcount;
453
454 bcount = min(bufblks, (start_blk + nbblks - i));
455
456 error = xlog_bread(log, i, bcount, bp, &buf);
457 if (error)
458 goto out;
459
460 for (j = 0; j < bcount; j++) {
461 cycle = xlog_get_cycle(buf);
462 if (cycle == stop_on_cycle_no) {
463 *new_blk = i+j;
464 goto out;
465 }
466
467 buf += BBSIZE;
468 }
469 }
470
471 *new_blk = -1;
472
473out:
474 xlog_put_bp(bp);
475 return error;
476}
477
478/*
479 * Potentially backup over partial log record write.
480 *
481 * In the typical case, last_blk is the number of the block directly after
482 * a good log record. Therefore, we subtract one to get the block number
483 * of the last block in the given buffer. extra_bblks contains the number
484 * of blocks we would have read on a previous read. This happens when the
485 * last log record is split over the end of the physical log.
486 *
487 * extra_bblks is the number of blocks potentially verified on a previous
488 * call to this routine.
489 */
490STATIC int
491xlog_find_verify_log_record(
492 xlog_t *log,
493 xfs_daddr_t start_blk,
494 xfs_daddr_t *last_blk,
495 int extra_bblks)
496{
497 xfs_daddr_t i;
498 xfs_buf_t *bp;
499 xfs_caddr_t offset = NULL;
500 xlog_rec_header_t *head = NULL;
501 int error = 0;
502 int smallmem = 0;
503 int num_blks = *last_blk - start_blk;
504 int xhdrs;
505
506 ASSERT(start_blk != 0 || *last_blk != start_blk);
507
508 if (!(bp = xlog_get_bp(log, num_blks))) {
509 if (!(bp = xlog_get_bp(log, 1)))
510 return ENOMEM;
511 smallmem = 1;
512 } else {
513 error = xlog_bread(log, start_blk, num_blks, bp, &offset);
514 if (error)
515 goto out;
516 offset += ((num_blks - 1) << BBSHIFT);
517 }
518
519 for (i = (*last_blk) - 1; i >= 0; i--) {
520 if (i < start_blk) {
521 /* valid log record not found */
522 xfs_warn(log->l_mp,
523 "Log inconsistent (didn't find previous header)");
524 ASSERT(0);
525 error = XFS_ERROR(EIO);
526 goto out;
527 }
528
529 if (smallmem) {
530 error = xlog_bread(log, i, 1, bp, &offset);
531 if (error)
532 goto out;
533 }
534
535 head = (xlog_rec_header_t *)offset;
536
537 if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
538 break;
539
540 if (!smallmem)
541 offset -= BBSIZE;
542 }
543
544 /*
545 * We hit the beginning of the physical log & still no header. Return
546 * to caller. If caller can handle a return of -1, then this routine
547 * will be called again for the end of the physical log.
548 */
549 if (i == -1) {
550 error = -1;
551 goto out;
552 }
553
554 /*
555 * We have the final block of the good log (the first block
556 * of the log record _before_ the head. So we check the uuid.
557 */
558 if ((error = xlog_header_check_mount(log->l_mp, head)))
559 goto out;
560
561 /*
562 * We may have found a log record header before we expected one.
563 * last_blk will be the 1st block # with a given cycle #. We may end
564 * up reading an entire log record. In this case, we don't want to
565 * reset last_blk. Only when last_blk points in the middle of a log
566 * record do we update last_blk.
567 */
568 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
569 uint h_size = be32_to_cpu(head->h_size);
570
571 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
572 if (h_size % XLOG_HEADER_CYCLE_SIZE)
573 xhdrs++;
574 } else {
575 xhdrs = 1;
576 }
577
578 if (*last_blk - i + extra_bblks !=
579 BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
580 *last_blk = i;
581
582out:
583 xlog_put_bp(bp);
584 return error;
585}
586
587/*
588 * Head is defined to be the point of the log where the next log write
589 * write could go. This means that incomplete LR writes at the end are
590 * eliminated when calculating the head. We aren't guaranteed that previous
591 * LR have complete transactions. We only know that a cycle number of
592 * current cycle number -1 won't be present in the log if we start writing
593 * from our current block number.
594 *
595 * last_blk contains the block number of the first block with a given
596 * cycle number.
597 *
598 * Return: zero if normal, non-zero if error.
599 */
600STATIC int
601xlog_find_head(
602 xlog_t *log,
603 xfs_daddr_t *return_head_blk)
604{
605 xfs_buf_t *bp;
606 xfs_caddr_t offset;
607 xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
608 int num_scan_bblks;
609 uint first_half_cycle, last_half_cycle;
610 uint stop_on_cycle;
611 int error, log_bbnum = log->l_logBBsize;
612
613 /* Is the end of the log device zeroed? */
614 if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
615 *return_head_blk = first_blk;
616
617 /* Is the whole lot zeroed? */
618 if (!first_blk) {
619 /* Linux XFS shouldn't generate totally zeroed logs -
620 * mkfs etc write a dummy unmount record to a fresh
621 * log so we can store the uuid in there
622 */
623 xfs_warn(log->l_mp, "totally zeroed log");
624 }
625
626 return 0;
627 } else if (error) {
628 xfs_warn(log->l_mp, "empty log check failed");
629 return error;
630 }
631
632 first_blk = 0; /* get cycle # of 1st block */
633 bp = xlog_get_bp(log, 1);
634 if (!bp)
635 return ENOMEM;
636
637 error = xlog_bread(log, 0, 1, bp, &offset);
638 if (error)
639 goto bp_err;
640
641 first_half_cycle = xlog_get_cycle(offset);
642
643 last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
644 error = xlog_bread(log, last_blk, 1, bp, &offset);
645 if (error)
646 goto bp_err;
647
648 last_half_cycle = xlog_get_cycle(offset);
649 ASSERT(last_half_cycle != 0);
650
651 /*
652 * If the 1st half cycle number is equal to the last half cycle number,
653 * then the entire log is stamped with the same cycle number. In this
654 * case, head_blk can't be set to zero (which makes sense). The below
655 * math doesn't work out properly with head_blk equal to zero. Instead,
656 * we set it to log_bbnum which is an invalid block number, but this
657 * value makes the math correct. If head_blk doesn't changed through
658 * all the tests below, *head_blk is set to zero at the very end rather
659 * than log_bbnum. In a sense, log_bbnum and zero are the same block
660 * in a circular file.
661 */
662 if (first_half_cycle == last_half_cycle) {
663 /*
664 * In this case we believe that the entire log should have
665 * cycle number last_half_cycle. We need to scan backwards
666 * from the end verifying that there are no holes still
667 * containing last_half_cycle - 1. If we find such a hole,
668 * then the start of that hole will be the new head. The
669 * simple case looks like
670 * x | x ... | x - 1 | x
671 * Another case that fits this picture would be
672 * x | x + 1 | x ... | x
673 * In this case the head really is somewhere at the end of the
674 * log, as one of the latest writes at the beginning was
675 * incomplete.
676 * One more case is
677 * x | x + 1 | x ... | x - 1 | x
678 * This is really the combination of the above two cases, and
679 * the head has to end up at the start of the x-1 hole at the
680 * end of the log.
681 *
682 * In the 256k log case, we will read from the beginning to the
683 * end of the log and search for cycle numbers equal to x-1.
684 * We don't worry about the x+1 blocks that we encounter,
685 * because we know that they cannot be the head since the log
686 * started with x.
687 */
688 head_blk = log_bbnum;
689 stop_on_cycle = last_half_cycle - 1;
690 } else {
691 /*
692 * In this case we want to find the first block with cycle
693 * number matching last_half_cycle. We expect the log to be
694 * some variation on
695 * x + 1 ... | x ... | x
696 * The first block with cycle number x (last_half_cycle) will
697 * be where the new head belongs. First we do a binary search
698 * for the first occurrence of last_half_cycle. The binary
699 * search may not be totally accurate, so then we scan back
700 * from there looking for occurrences of last_half_cycle before
701 * us. If that backwards scan wraps around the beginning of
702 * the log, then we look for occurrences of last_half_cycle - 1
703 * at the end of the log. The cases we're looking for look
704 * like
705 * v binary search stopped here
706 * x + 1 ... | x | x + 1 | x ... | x
707 * ^ but we want to locate this spot
708 * or
709 * <---------> less than scan distance
710 * x + 1 ... | x ... | x - 1 | x
711 * ^ we want to locate this spot
712 */
713 stop_on_cycle = last_half_cycle;
714 if ((error = xlog_find_cycle_start(log, bp, first_blk,
715 &head_blk, last_half_cycle)))
716 goto bp_err;
717 }
718
719 /*
720 * Now validate the answer. Scan back some number of maximum possible
721 * blocks and make sure each one has the expected cycle number. The
722 * maximum is determined by the total possible amount of buffering
723 * in the in-core log. The following number can be made tighter if
724 * we actually look at the block size of the filesystem.
725 */
726 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
727 if (head_blk >= num_scan_bblks) {
728 /*
729 * We are guaranteed that the entire check can be performed
730 * in one buffer.
731 */
732 start_blk = head_blk - num_scan_bblks;
733 if ((error = xlog_find_verify_cycle(log,
734 start_blk, num_scan_bblks,
735 stop_on_cycle, &new_blk)))
736 goto bp_err;
737 if (new_blk != -1)
738 head_blk = new_blk;
739 } else { /* need to read 2 parts of log */
740 /*
741 * We are going to scan backwards in the log in two parts.
742 * First we scan the physical end of the log. In this part
743 * of the log, we are looking for blocks with cycle number
744 * last_half_cycle - 1.
745 * If we find one, then we know that the log starts there, as
746 * we've found a hole that didn't get written in going around
747 * the end of the physical log. The simple case for this is
748 * x + 1 ... | x ... | x - 1 | x
749 * <---------> less than scan distance
750 * If all of the blocks at the end of the log have cycle number
751 * last_half_cycle, then we check the blocks at the start of
752 * the log looking for occurrences of last_half_cycle. If we
753 * find one, then our current estimate for the location of the
754 * first occurrence of last_half_cycle is wrong and we move
755 * back to the hole we've found. This case looks like
756 * x + 1 ... | x | x + 1 | x ...
757 * ^ binary search stopped here
758 * Another case we need to handle that only occurs in 256k
759 * logs is
760 * x + 1 ... | x ... | x+1 | x ...
761 * ^ binary search stops here
762 * In a 256k log, the scan at the end of the log will see the
763 * x + 1 blocks. We need to skip past those since that is
764 * certainly not the head of the log. By searching for
765 * last_half_cycle-1 we accomplish that.
766 */
767 ASSERT(head_blk <= INT_MAX &&
768 (xfs_daddr_t) num_scan_bblks >= head_blk);
769 start_blk = log_bbnum - (num_scan_bblks - head_blk);
770 if ((error = xlog_find_verify_cycle(log, start_blk,
771 num_scan_bblks - (int)head_blk,
772 (stop_on_cycle - 1), &new_blk)))
773 goto bp_err;
774 if (new_blk != -1) {
775 head_blk = new_blk;
776 goto validate_head;
777 }
778
779 /*
780 * Scan beginning of log now. The last part of the physical
781 * log is good. This scan needs to verify that it doesn't find
782 * the last_half_cycle.
783 */
784 start_blk = 0;
785 ASSERT(head_blk <= INT_MAX);
786 if ((error = xlog_find_verify_cycle(log,
787 start_blk, (int)head_blk,
788 stop_on_cycle, &new_blk)))
789 goto bp_err;
790 if (new_blk != -1)
791 head_blk = new_blk;
792 }
793
794validate_head:
795 /*
796 * Now we need to make sure head_blk is not pointing to a block in
797 * the middle of a log record.
798 */
799 num_scan_bblks = XLOG_REC_SHIFT(log);
800 if (head_blk >= num_scan_bblks) {
801 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
802
803 /* start ptr at last block ptr before head_blk */
804 if ((error = xlog_find_verify_log_record(log, start_blk,
805 &head_blk, 0)) == -1) {
806 error = XFS_ERROR(EIO);
807 goto bp_err;
808 } else if (error)
809 goto bp_err;
810 } else {
811 start_blk = 0;
812 ASSERT(head_blk <= INT_MAX);
813 if ((error = xlog_find_verify_log_record(log, start_blk,
814 &head_blk, 0)) == -1) {
815 /* We hit the beginning of the log during our search */
816 start_blk = log_bbnum - (num_scan_bblks - head_blk);
817 new_blk = log_bbnum;
818 ASSERT(start_blk <= INT_MAX &&
819 (xfs_daddr_t) log_bbnum-start_blk >= 0);
820 ASSERT(head_blk <= INT_MAX);
821 if ((error = xlog_find_verify_log_record(log,
822 start_blk, &new_blk,
823 (int)head_blk)) == -1) {
824 error = XFS_ERROR(EIO);
825 goto bp_err;
826 } else if (error)
827 goto bp_err;
828 if (new_blk != log_bbnum)
829 head_blk = new_blk;
830 } else if (error)
831 goto bp_err;
832 }
833
834 xlog_put_bp(bp);
835 if (head_blk == log_bbnum)
836 *return_head_blk = 0;
837 else
838 *return_head_blk = head_blk;
839 /*
840 * When returning here, we have a good block number. Bad block
841 * means that during a previous crash, we didn't have a clean break
842 * from cycle number N to cycle number N-1. In this case, we need
843 * to find the first block with cycle number N-1.
844 */
845 return 0;
846
847 bp_err:
848 xlog_put_bp(bp);
849
850 if (error)
851 xfs_warn(log->l_mp, "failed to find log head");
852 return error;
853}
854
855/*
856 * Find the sync block number or the tail of the log.
857 *
858 * This will be the block number of the last record to have its
859 * associated buffers synced to disk. Every log record header has
860 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
861 * to get a sync block number. The only concern is to figure out which
862 * log record header to believe.
863 *
864 * The following algorithm uses the log record header with the largest
865 * lsn. The entire log record does not need to be valid. We only care
866 * that the header is valid.
867 *
868 * We could speed up search by using current head_blk buffer, but it is not
869 * available.
870 */
871STATIC int
872xlog_find_tail(
873 xlog_t *log,
874 xfs_daddr_t *head_blk,
875 xfs_daddr_t *tail_blk)
876{
877 xlog_rec_header_t *rhead;
878 xlog_op_header_t *op_head;
879 xfs_caddr_t offset = NULL;
880 xfs_buf_t *bp;
881 int error, i, found;
882 xfs_daddr_t umount_data_blk;
883 xfs_daddr_t after_umount_blk;
884 xfs_lsn_t tail_lsn;
885 int hblks;
886
887 found = 0;
888
889 /*
890 * Find previous log record
891 */
892 if ((error = xlog_find_head(log, head_blk)))
893 return error;
894
895 bp = xlog_get_bp(log, 1);
896 if (!bp)
897 return ENOMEM;
898 if (*head_blk == 0) { /* special case */
899 error = xlog_bread(log, 0, 1, bp, &offset);
900 if (error)
901 goto done;
902
903 if (xlog_get_cycle(offset) == 0) {
904 *tail_blk = 0;
905 /* leave all other log inited values alone */
906 goto done;
907 }
908 }
909
910 /*
911 * Search backwards looking for log record header block
912 */
913 ASSERT(*head_blk < INT_MAX);
914 for (i = (int)(*head_blk) - 1; i >= 0; i--) {
915 error = xlog_bread(log, i, 1, bp, &offset);
916 if (error)
917 goto done;
918
919 if (*(__be32 *)offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
920 found = 1;
921 break;
922 }
923 }
924 /*
925 * If we haven't found the log record header block, start looking
926 * again from the end of the physical log. XXXmiken: There should be
927 * a check here to make sure we didn't search more than N blocks in
928 * the previous code.
929 */
930 if (!found) {
931 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
932 error = xlog_bread(log, i, 1, bp, &offset);
933 if (error)
934 goto done;
935
936 if (*(__be32 *)offset ==
937 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
938 found = 2;
939 break;
940 }
941 }
942 }
943 if (!found) {
944 xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
945 ASSERT(0);
946 return XFS_ERROR(EIO);
947 }
948
949 /* find blk_no of tail of log */
950 rhead = (xlog_rec_header_t *)offset;
951 *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
952
953 /*
954 * Reset log values according to the state of the log when we
955 * crashed. In the case where head_blk == 0, we bump curr_cycle
956 * one because the next write starts a new cycle rather than
957 * continuing the cycle of the last good log record. At this
958 * point we have guaranteed that all partial log records have been
959 * accounted for. Therefore, we know that the last good log record
960 * written was complete and ended exactly on the end boundary
961 * of the physical log.
962 */
963 log->l_prev_block = i;
964 log->l_curr_block = (int)*head_blk;
965 log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
966 if (found == 2)
967 log->l_curr_cycle++;
968 atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
969 atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
970 xlog_assign_grant_head(&log->l_grant_reserve_head, log->l_curr_cycle,
971 BBTOB(log->l_curr_block));
972 xlog_assign_grant_head(&log->l_grant_write_head, log->l_curr_cycle,
973 BBTOB(log->l_curr_block));
974
975 /*
976 * Look for unmount record. If we find it, then we know there
977 * was a clean unmount. Since 'i' could be the last block in
978 * the physical log, we convert to a log block before comparing
979 * to the head_blk.
980 *
981 * Save the current tail lsn to use to pass to
982 * xlog_clear_stale_blocks() below. We won't want to clear the
983 * unmount record if there is one, so we pass the lsn of the
984 * unmount record rather than the block after it.
985 */
986 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
987 int h_size = be32_to_cpu(rhead->h_size);
988 int h_version = be32_to_cpu(rhead->h_version);
989
990 if ((h_version & XLOG_VERSION_2) &&
991 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
992 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
993 if (h_size % XLOG_HEADER_CYCLE_SIZE)
994 hblks++;
995 } else {
996 hblks = 1;
997 }
998 } else {
999 hblks = 1;
1000 }
1001 after_umount_blk = (i + hblks + (int)
1002 BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
1003 tail_lsn = atomic64_read(&log->l_tail_lsn);
1004 if (*head_blk == after_umount_blk &&
1005 be32_to_cpu(rhead->h_num_logops) == 1) {
1006 umount_data_blk = (i + hblks) % log->l_logBBsize;
1007 error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
1008 if (error)
1009 goto done;
1010
1011 op_head = (xlog_op_header_t *)offset;
1012 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1013 /*
1014 * Set tail and last sync so that newly written
1015 * log records will point recovery to after the
1016 * current unmount record.
1017 */
1018 xlog_assign_atomic_lsn(&log->l_tail_lsn,
1019 log->l_curr_cycle, after_umount_blk);
1020 xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1021 log->l_curr_cycle, after_umount_blk);
1022 *tail_blk = after_umount_blk;
1023
1024 /*
1025 * Note that the unmount was clean. If the unmount
1026 * was not clean, we need to know this to rebuild the
1027 * superblock counters from the perag headers if we
1028 * have a filesystem using non-persistent counters.
1029 */
1030 log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1031 }
1032 }
1033
1034 /*
1035 * Make sure that there are no blocks in front of the head
1036 * with the same cycle number as the head. This can happen
1037 * because we allow multiple outstanding log writes concurrently,
1038 * and the later writes might make it out before earlier ones.
1039 *
1040 * We use the lsn from before modifying it so that we'll never
1041 * overwrite the unmount record after a clean unmount.
1042 *
1043 * Do this only if we are going to recover the filesystem
1044 *
1045 * NOTE: This used to say "if (!readonly)"
1046 * However on Linux, we can & do recover a read-only filesystem.
1047 * We only skip recovery if NORECOVERY is specified on mount,
1048 * in which case we would not be here.
1049 *
1050 * But... if the -device- itself is readonly, just skip this.
1051 * We can't recover this device anyway, so it won't matter.
1052 */
1053 if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
1054 error = xlog_clear_stale_blocks(log, tail_lsn);
1055
1056done:
1057 xlog_put_bp(bp);
1058
1059 if (error)
1060 xfs_warn(log->l_mp, "failed to locate log tail");
1061 return error;
1062}
1063
1064/*
1065 * Is the log zeroed at all?
1066 *
1067 * The last binary search should be changed to perform an X block read
1068 * once X becomes small enough. You can then search linearly through
1069 * the X blocks. This will cut down on the number of reads we need to do.
1070 *
1071 * If the log is partially zeroed, this routine will pass back the blkno
1072 * of the first block with cycle number 0. It won't have a complete LR
1073 * preceding it.
1074 *
1075 * Return:
1076 * 0 => the log is completely written to
1077 * -1 => use *blk_no as the first block of the log
1078 * >0 => error has occurred
1079 */
1080STATIC int
1081xlog_find_zeroed(
1082 xlog_t *log,
1083 xfs_daddr_t *blk_no)
1084{
1085 xfs_buf_t *bp;
1086 xfs_caddr_t offset;
1087 uint first_cycle, last_cycle;
1088 xfs_daddr_t new_blk, last_blk, start_blk;
1089 xfs_daddr_t num_scan_bblks;
1090 int error, log_bbnum = log->l_logBBsize;
1091
1092 *blk_no = 0;
1093
1094 /* check totally zeroed log */
1095 bp = xlog_get_bp(log, 1);
1096 if (!bp)
1097 return ENOMEM;
1098 error = xlog_bread(log, 0, 1, bp, &offset);
1099 if (error)
1100 goto bp_err;
1101
1102 first_cycle = xlog_get_cycle(offset);
1103 if (first_cycle == 0) { /* completely zeroed log */
1104 *blk_no = 0;
1105 xlog_put_bp(bp);
1106 return -1;
1107 }
1108
1109 /* check partially zeroed log */
1110 error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1111 if (error)
1112 goto bp_err;
1113
1114 last_cycle = xlog_get_cycle(offset);
1115 if (last_cycle != 0) { /* log completely written to */
1116 xlog_put_bp(bp);
1117 return 0;
1118 } else if (first_cycle != 1) {
1119 /*
1120 * If the cycle of the last block is zero, the cycle of
1121 * the first block must be 1. If it's not, maybe we're
1122 * not looking at a log... Bail out.
1123 */
1124 xfs_warn(log->l_mp,
1125 "Log inconsistent or not a log (last==0, first!=1)");
1126 return XFS_ERROR(EINVAL);
1127 }
1128
1129 /* we have a partially zeroed log */
1130 last_blk = log_bbnum-1;
1131 if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1132 goto bp_err;
1133
1134 /*
1135 * Validate the answer. Because there is no way to guarantee that
1136 * the entire log is made up of log records which are the same size,
1137 * we scan over the defined maximum blocks. At this point, the maximum
1138 * is not chosen to mean anything special. XXXmiken
1139 */
1140 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1141 ASSERT(num_scan_bblks <= INT_MAX);
1142
1143 if (last_blk < num_scan_bblks)
1144 num_scan_bblks = last_blk;
1145 start_blk = last_blk - num_scan_bblks;
1146
1147 /*
1148 * We search for any instances of cycle number 0 that occur before
1149 * our current estimate of the head. What we're trying to detect is
1150 * 1 ... | 0 | 1 | 0...
1151 * ^ binary search ends here
1152 */
1153 if ((error = xlog_find_verify_cycle(log, start_blk,
1154 (int)num_scan_bblks, 0, &new_blk)))
1155 goto bp_err;
1156 if (new_blk != -1)
1157 last_blk = new_blk;
1158
1159 /*
1160 * Potentially backup over partial log record write. We don't need
1161 * to search the end of the log because we know it is zero.
1162 */
1163 if ((error = xlog_find_verify_log_record(log, start_blk,
1164 &last_blk, 0)) == -1) {
1165 error = XFS_ERROR(EIO);
1166 goto bp_err;
1167 } else if (error)
1168 goto bp_err;
1169
1170 *blk_no = last_blk;
1171bp_err:
1172 xlog_put_bp(bp);
1173 if (error)
1174 return error;
1175 return -1;
1176}
1177
1178/*
1179 * These are simple subroutines used by xlog_clear_stale_blocks() below
1180 * to initialize a buffer full of empty log record headers and write
1181 * them into the log.
1182 */
1183STATIC void
1184xlog_add_record(
1185 xlog_t *log,
1186 xfs_caddr_t buf,
1187 int cycle,
1188 int block,
1189 int tail_cycle,
1190 int tail_block)
1191{
1192 xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
1193
1194 memset(buf, 0, BBSIZE);
1195 recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1196 recp->h_cycle = cpu_to_be32(cycle);
1197 recp->h_version = cpu_to_be32(
1198 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1199 recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1200 recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1201 recp->h_fmt = cpu_to_be32(XLOG_FMT);
1202 memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1203}
1204
1205STATIC int
1206xlog_write_log_records(
1207 xlog_t *log,
1208 int cycle,
1209 int start_block,
1210 int blocks,
1211 int tail_cycle,
1212 int tail_block)
1213{
1214 xfs_caddr_t offset;
1215 xfs_buf_t *bp;
1216 int balign, ealign;
1217 int sectbb = log->l_sectBBsize;
1218 int end_block = start_block + blocks;
1219 int bufblks;
1220 int error = 0;
1221 int i, j = 0;
1222
1223 /*
1224 * Greedily allocate a buffer big enough to handle the full
1225 * range of basic blocks to be written. If that fails, try
1226 * a smaller size. We need to be able to write at least a
1227 * log sector, or we're out of luck.
1228 */
1229 bufblks = 1 << ffs(blocks);
1230 while (!(bp = xlog_get_bp(log, bufblks))) {
1231 bufblks >>= 1;
1232 if (bufblks < sectbb)
1233 return ENOMEM;
1234 }
1235
1236 /* We may need to do a read at the start to fill in part of
1237 * the buffer in the starting sector not covered by the first
1238 * write below.
1239 */
1240 balign = round_down(start_block, sectbb);
1241 if (balign != start_block) {
1242 error = xlog_bread_noalign(log, start_block, 1, bp);
1243 if (error)
1244 goto out_put_bp;
1245
1246 j = start_block - balign;
1247 }
1248
1249 for (i = start_block; i < end_block; i += bufblks) {
1250 int bcount, endcount;
1251
1252 bcount = min(bufblks, end_block - start_block);
1253 endcount = bcount - j;
1254
1255 /* We may need to do a read at the end to fill in part of
1256 * the buffer in the final sector not covered by the write.
1257 * If this is the same sector as the above read, skip it.
1258 */
1259 ealign = round_down(end_block, sectbb);
1260 if (j == 0 && (start_block + endcount > ealign)) {
1261 offset = bp->b_addr + BBTOB(ealign - start_block);
1262 error = xlog_bread_offset(log, ealign, sectbb,
1263 bp, offset);
1264 if (error)
1265 break;
1266
1267 }
1268
1269 offset = xlog_align(log, start_block, endcount, bp);
1270 for (; j < endcount; j++) {
1271 xlog_add_record(log, offset, cycle, i+j,
1272 tail_cycle, tail_block);
1273 offset += BBSIZE;
1274 }
1275 error = xlog_bwrite(log, start_block, endcount, bp);
1276 if (error)
1277 break;
1278 start_block += endcount;
1279 j = 0;
1280 }
1281
1282 out_put_bp:
1283 xlog_put_bp(bp);
1284 return error;
1285}
1286
1287/*
1288 * This routine is called to blow away any incomplete log writes out
1289 * in front of the log head. We do this so that we won't become confused
1290 * if we come up, write only a little bit more, and then crash again.
1291 * If we leave the partial log records out there, this situation could
1292 * cause us to think those partial writes are valid blocks since they
1293 * have the current cycle number. We get rid of them by overwriting them
1294 * with empty log records with the old cycle number rather than the
1295 * current one.
1296 *
1297 * The tail lsn is passed in rather than taken from
1298 * the log so that we will not write over the unmount record after a
1299 * clean unmount in a 512 block log. Doing so would leave the log without
1300 * any valid log records in it until a new one was written. If we crashed
1301 * during that time we would not be able to recover.
1302 */
1303STATIC int
1304xlog_clear_stale_blocks(
1305 xlog_t *log,
1306 xfs_lsn_t tail_lsn)
1307{
1308 int tail_cycle, head_cycle;
1309 int tail_block, head_block;
1310 int tail_distance, max_distance;
1311 int distance;
1312 int error;
1313
1314 tail_cycle = CYCLE_LSN(tail_lsn);
1315 tail_block = BLOCK_LSN(tail_lsn);
1316 head_cycle = log->l_curr_cycle;
1317 head_block = log->l_curr_block;
1318
1319 /*
1320 * Figure out the distance between the new head of the log
1321 * and the tail. We want to write over any blocks beyond the
1322 * head that we may have written just before the crash, but
1323 * we don't want to overwrite the tail of the log.
1324 */
1325 if (head_cycle == tail_cycle) {
1326 /*
1327 * The tail is behind the head in the physical log,
1328 * so the distance from the head to the tail is the
1329 * distance from the head to the end of the log plus
1330 * the distance from the beginning of the log to the
1331 * tail.
1332 */
1333 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1334 XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1335 XFS_ERRLEVEL_LOW, log->l_mp);
1336 return XFS_ERROR(EFSCORRUPTED);
1337 }
1338 tail_distance = tail_block + (log->l_logBBsize - head_block);
1339 } else {
1340 /*
1341 * The head is behind the tail in the physical log,
1342 * so the distance from the head to the tail is just
1343 * the tail block minus the head block.
1344 */
1345 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1346 XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1347 XFS_ERRLEVEL_LOW, log->l_mp);
1348 return XFS_ERROR(EFSCORRUPTED);
1349 }
1350 tail_distance = tail_block - head_block;
1351 }
1352
1353 /*
1354 * If the head is right up against the tail, we can't clear
1355 * anything.
1356 */
1357 if (tail_distance <= 0) {
1358 ASSERT(tail_distance == 0);
1359 return 0;
1360 }
1361
1362 max_distance = XLOG_TOTAL_REC_SHIFT(log);
1363 /*
1364 * Take the smaller of the maximum amount of outstanding I/O
1365 * we could have and the distance to the tail to clear out.
1366 * We take the smaller so that we don't overwrite the tail and
1367 * we don't waste all day writing from the head to the tail
1368 * for no reason.
1369 */
1370 max_distance = MIN(max_distance, tail_distance);
1371
1372 if ((head_block + max_distance) <= log->l_logBBsize) {
1373 /*
1374 * We can stomp all the blocks we need to without
1375 * wrapping around the end of the log. Just do it
1376 * in a single write. Use the cycle number of the
1377 * current cycle minus one so that the log will look like:
1378 * n ... | n - 1 ...
1379 */
1380 error = xlog_write_log_records(log, (head_cycle - 1),
1381 head_block, max_distance, tail_cycle,
1382 tail_block);
1383 if (error)
1384 return error;
1385 } else {
1386 /*
1387 * We need to wrap around the end of the physical log in
1388 * order to clear all the blocks. Do it in two separate
1389 * I/Os. The first write should be from the head to the
1390 * end of the physical log, and it should use the current
1391 * cycle number minus one just like above.
1392 */
1393 distance = log->l_logBBsize - head_block;
1394 error = xlog_write_log_records(log, (head_cycle - 1),
1395 head_block, distance, tail_cycle,
1396 tail_block);
1397
1398 if (error)
1399 return error;
1400
1401 /*
1402 * Now write the blocks at the start of the physical log.
1403 * This writes the remainder of the blocks we want to clear.
1404 * It uses the current cycle number since we're now on the
1405 * same cycle as the head so that we get:
1406 * n ... n ... | n - 1 ...
1407 * ^^^^^ blocks we're writing
1408 */
1409 distance = max_distance - (log->l_logBBsize - head_block);
1410 error = xlog_write_log_records(log, head_cycle, 0, distance,
1411 tail_cycle, tail_block);
1412 if (error)
1413 return error;
1414 }
1415
1416 return 0;
1417}
1418
1419/******************************************************************************
1420 *
1421 * Log recover routines
1422 *
1423 ******************************************************************************
1424 */
1425
1426STATIC xlog_recover_t *
1427xlog_recover_find_tid(
1428 struct hlist_head *head,
1429 xlog_tid_t tid)
1430{
1431 xlog_recover_t *trans;
1432 struct hlist_node *n;
1433
1434 hlist_for_each_entry(trans, n, head, r_list) {
1435 if (trans->r_log_tid == tid)
1436 return trans;
1437 }
1438 return NULL;
1439}
1440
1441STATIC void
1442xlog_recover_new_tid(
1443 struct hlist_head *head,
1444 xlog_tid_t tid,
1445 xfs_lsn_t lsn)
1446{
1447 xlog_recover_t *trans;
1448
1449 trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1450 trans->r_log_tid = tid;
1451 trans->r_lsn = lsn;
1452 INIT_LIST_HEAD(&trans->r_itemq);
1453
1454 INIT_HLIST_NODE(&trans->r_list);
1455 hlist_add_head(&trans->r_list, head);
1456}
1457
1458STATIC void
1459xlog_recover_add_item(
1460 struct list_head *head)
1461{
1462 xlog_recover_item_t *item;
1463
1464 item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1465 INIT_LIST_HEAD(&item->ri_list);
1466 list_add_tail(&item->ri_list, head);
1467}
1468
1469STATIC int
1470xlog_recover_add_to_cont_trans(
1471 struct log *log,
1472 xlog_recover_t *trans,
1473 xfs_caddr_t dp,
1474 int len)
1475{
1476 xlog_recover_item_t *item;
1477 xfs_caddr_t ptr, old_ptr;
1478 int old_len;
1479
1480 if (list_empty(&trans->r_itemq)) {
1481 /* finish copying rest of trans header */
1482 xlog_recover_add_item(&trans->r_itemq);
1483 ptr = (xfs_caddr_t) &trans->r_theader +
1484 sizeof(xfs_trans_header_t) - len;
1485 memcpy(ptr, dp, len); /* d, s, l */
1486 return 0;
1487 }
1488 /* take the tail entry */
1489 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1490
1491 old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1492 old_len = item->ri_buf[item->ri_cnt-1].i_len;
1493
1494 ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1495 memcpy(&ptr[old_len], dp, len); /* d, s, l */
1496 item->ri_buf[item->ri_cnt-1].i_len += len;
1497 item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1498 trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
1499 return 0;
1500}
1501
1502/*
1503 * The next region to add is the start of a new region. It could be
1504 * a whole region or it could be the first part of a new region. Because
1505 * of this, the assumption here is that the type and size fields of all
1506 * format structures fit into the first 32 bits of the structure.
1507 *
1508 * This works because all regions must be 32 bit aligned. Therefore, we
1509 * either have both fields or we have neither field. In the case we have
1510 * neither field, the data part of the region is zero length. We only have
1511 * a log_op_header and can throw away the header since a new one will appear
1512 * later. If we have at least 4 bytes, then we can determine how many regions
1513 * will appear in the current log item.
1514 */
1515STATIC int
1516xlog_recover_add_to_trans(
1517 struct log *log,
1518 xlog_recover_t *trans,
1519 xfs_caddr_t dp,
1520 int len)
1521{
1522 xfs_inode_log_format_t *in_f; /* any will do */
1523 xlog_recover_item_t *item;
1524 xfs_caddr_t ptr;
1525
1526 if (!len)
1527 return 0;
1528 if (list_empty(&trans->r_itemq)) {
1529 /* we need to catch log corruptions here */
1530 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1531 xfs_warn(log->l_mp, "%s: bad header magic number",
1532 __func__);
1533 ASSERT(0);
1534 return XFS_ERROR(EIO);
1535 }
1536 if (len == sizeof(xfs_trans_header_t))
1537 xlog_recover_add_item(&trans->r_itemq);
1538 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1539 return 0;
1540 }
1541
1542 ptr = kmem_alloc(len, KM_SLEEP);
1543 memcpy(ptr, dp, len);
1544 in_f = (xfs_inode_log_format_t *)ptr;
1545
1546 /* take the tail entry */
1547 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1548 if (item->ri_total != 0 &&
1549 item->ri_total == item->ri_cnt) {
1550 /* tail item is in use, get a new one */
1551 xlog_recover_add_item(&trans->r_itemq);
1552 item = list_entry(trans->r_itemq.prev,
1553 xlog_recover_item_t, ri_list);
1554 }
1555
1556 if (item->ri_total == 0) { /* first region to be added */
1557 if (in_f->ilf_size == 0 ||
1558 in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
1559 xfs_warn(log->l_mp,
1560 "bad number of regions (%d) in inode log format",
1561 in_f->ilf_size);
1562 ASSERT(0);
1563 return XFS_ERROR(EIO);
1564 }
1565
1566 item->ri_total = in_f->ilf_size;
1567 item->ri_buf =
1568 kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
1569 KM_SLEEP);
1570 }
1571 ASSERT(item->ri_total > item->ri_cnt);
1572 /* Description region is ri_buf[0] */
1573 item->ri_buf[item->ri_cnt].i_addr = ptr;
1574 item->ri_buf[item->ri_cnt].i_len = len;
1575 item->ri_cnt++;
1576 trace_xfs_log_recover_item_add(log, trans, item, 0);
1577 return 0;
1578}
1579
1580/*
1581 * Sort the log items in the transaction. Cancelled buffers need
1582 * to be put first so they are processed before any items that might
1583 * modify the buffers. If they are cancelled, then the modifications
1584 * don't need to be replayed.
1585 */
1586STATIC int
1587xlog_recover_reorder_trans(
1588 struct log *log,
1589 xlog_recover_t *trans,
1590 int pass)
1591{
1592 xlog_recover_item_t *item, *n;
1593 LIST_HEAD(sort_list);
1594
1595 list_splice_init(&trans->r_itemq, &sort_list);
1596 list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1597 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1598
1599 switch (ITEM_TYPE(item)) {
1600 case XFS_LI_BUF:
1601 if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1602 trace_xfs_log_recover_item_reorder_head(log,
1603 trans, item, pass);
1604 list_move(&item->ri_list, &trans->r_itemq);
1605 break;
1606 }
1607 case XFS_LI_INODE:
1608 case XFS_LI_DQUOT:
1609 case XFS_LI_QUOTAOFF:
1610 case XFS_LI_EFD:
1611 case XFS_LI_EFI:
1612 trace_xfs_log_recover_item_reorder_tail(log,
1613 trans, item, pass);
1614 list_move_tail(&item->ri_list, &trans->r_itemq);
1615 break;
1616 default:
1617 xfs_warn(log->l_mp,
1618 "%s: unrecognized type of log operation",
1619 __func__);
1620 ASSERT(0);
1621 return XFS_ERROR(EIO);
1622 }
1623 }
1624 ASSERT(list_empty(&sort_list));
1625 return 0;
1626}
1627
1628/*
1629 * Build up the table of buf cancel records so that we don't replay
1630 * cancelled data in the second pass. For buffer records that are
1631 * not cancel records, there is nothing to do here so we just return.
1632 *
1633 * If we get a cancel record which is already in the table, this indicates
1634 * that the buffer was cancelled multiple times. In order to ensure
1635 * that during pass 2 we keep the record in the table until we reach its
1636 * last occurrence in the log, we keep a reference count in the cancel
1637 * record in the table to tell us how many times we expect to see this
1638 * record during the second pass.
1639 */
1640STATIC int
1641xlog_recover_buffer_pass1(
1642 struct log *log,
1643 xlog_recover_item_t *item)
1644{
1645 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1646 struct list_head *bucket;
1647 struct xfs_buf_cancel *bcp;
1648
1649 /*
1650 * If this isn't a cancel buffer item, then just return.
1651 */
1652 if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1653 trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1654 return 0;
1655 }
1656
1657 /*
1658 * Insert an xfs_buf_cancel record into the hash table of them.
1659 * If there is already an identical record, bump its reference count.
1660 */
1661 bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1662 list_for_each_entry(bcp, bucket, bc_list) {
1663 if (bcp->bc_blkno == buf_f->blf_blkno &&
1664 bcp->bc_len == buf_f->blf_len) {
1665 bcp->bc_refcount++;
1666 trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1667 return 0;
1668 }
1669 }
1670
1671 bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
1672 bcp->bc_blkno = buf_f->blf_blkno;
1673 bcp->bc_len = buf_f->blf_len;
1674 bcp->bc_refcount = 1;
1675 list_add_tail(&bcp->bc_list, bucket);
1676
1677 trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1678 return 0;
1679}
1680
1681/*
1682 * Check to see whether the buffer being recovered has a corresponding
1683 * entry in the buffer cancel record table. If it does then return 1
1684 * so that it will be cancelled, otherwise return 0. If the buffer is
1685 * actually a buffer cancel item (XFS_BLF_CANCEL is set), then decrement
1686 * the refcount on the entry in the table and remove it from the table
1687 * if this is the last reference.
1688 *
1689 * We remove the cancel record from the table when we encounter its
1690 * last occurrence in the log so that if the same buffer is re-used
1691 * again after its last cancellation we actually replay the changes
1692 * made at that point.
1693 */
1694STATIC int
1695xlog_check_buffer_cancelled(
1696 struct log *log,
1697 xfs_daddr_t blkno,
1698 uint len,
1699 ushort flags)
1700{
1701 struct list_head *bucket;
1702 struct xfs_buf_cancel *bcp;
1703
1704 if (log->l_buf_cancel_table == NULL) {
1705 /*
1706 * There is nothing in the table built in pass one,
1707 * so this buffer must not be cancelled.
1708 */
1709 ASSERT(!(flags & XFS_BLF_CANCEL));
1710 return 0;
1711 }
1712
1713 /*
1714 * Search for an entry in the cancel table that matches our buffer.
1715 */
1716 bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
1717 list_for_each_entry(bcp, bucket, bc_list) {
1718 if (bcp->bc_blkno == blkno && bcp->bc_len == len)
1719 goto found;
1720 }
1721
1722 /*
1723 * We didn't find a corresponding entry in the table, so return 0 so
1724 * that the buffer is NOT cancelled.
1725 */
1726 ASSERT(!(flags & XFS_BLF_CANCEL));
1727 return 0;
1728
1729found:
1730 /*
1731 * We've go a match, so return 1 so that the recovery of this buffer
1732 * is cancelled. If this buffer is actually a buffer cancel log
1733 * item, then decrement the refcount on the one in the table and
1734 * remove it if this is the last reference.
1735 */
1736 if (flags & XFS_BLF_CANCEL) {
1737 if (--bcp->bc_refcount == 0) {
1738 list_del(&bcp->bc_list);
1739 kmem_free(bcp);
1740 }
1741 }
1742 return 1;
1743}
1744
1745/*
1746 * Perform recovery for a buffer full of inodes. In these buffers, the only
1747 * data which should be recovered is that which corresponds to the
1748 * di_next_unlinked pointers in the on disk inode structures. The rest of the
1749 * data for the inodes is always logged through the inodes themselves rather
1750 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
1751 *
1752 * The only time when buffers full of inodes are fully recovered is when the
1753 * buffer is full of newly allocated inodes. In this case the buffer will
1754 * not be marked as an inode buffer and so will be sent to
1755 * xlog_recover_do_reg_buffer() below during recovery.
1756 */
1757STATIC int
1758xlog_recover_do_inode_buffer(
1759 struct xfs_mount *mp,
1760 xlog_recover_item_t *item,
1761 struct xfs_buf *bp,
1762 xfs_buf_log_format_t *buf_f)
1763{
1764 int i;
1765 int item_index = 0;
1766 int bit = 0;
1767 int nbits = 0;
1768 int reg_buf_offset = 0;
1769 int reg_buf_bytes = 0;
1770 int next_unlinked_offset;
1771 int inodes_per_buf;
1772 xfs_agino_t *logged_nextp;
1773 xfs_agino_t *buffer_nextp;
1774
1775 trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
1776
1777 inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1778 for (i = 0; i < inodes_per_buf; i++) {
1779 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1780 offsetof(xfs_dinode_t, di_next_unlinked);
1781
1782 while (next_unlinked_offset >=
1783 (reg_buf_offset + reg_buf_bytes)) {
1784 /*
1785 * The next di_next_unlinked field is beyond
1786 * the current logged region. Find the next
1787 * logged region that contains or is beyond
1788 * the current di_next_unlinked field.
1789 */
1790 bit += nbits;
1791 bit = xfs_next_bit(buf_f->blf_data_map,
1792 buf_f->blf_map_size, bit);
1793
1794 /*
1795 * If there are no more logged regions in the
1796 * buffer, then we're done.
1797 */
1798 if (bit == -1)
1799 return 0;
1800
1801 nbits = xfs_contig_bits(buf_f->blf_data_map,
1802 buf_f->blf_map_size, bit);
1803 ASSERT(nbits > 0);
1804 reg_buf_offset = bit << XFS_BLF_SHIFT;
1805 reg_buf_bytes = nbits << XFS_BLF_SHIFT;
1806 item_index++;
1807 }
1808
1809 /*
1810 * If the current logged region starts after the current
1811 * di_next_unlinked field, then move on to the next
1812 * di_next_unlinked field.
1813 */
1814 if (next_unlinked_offset < reg_buf_offset)
1815 continue;
1816
1817 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1818 ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
1819 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1820
1821 /*
1822 * The current logged region contains a copy of the
1823 * current di_next_unlinked field. Extract its value
1824 * and copy it to the buffer copy.
1825 */
1826 logged_nextp = item->ri_buf[item_index].i_addr +
1827 next_unlinked_offset - reg_buf_offset;
1828 if (unlikely(*logged_nextp == 0)) {
1829 xfs_alert(mp,
1830 "Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
1831 "Trying to replay bad (0) inode di_next_unlinked field.",
1832 item, bp);
1833 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1834 XFS_ERRLEVEL_LOW, mp);
1835 return XFS_ERROR(EFSCORRUPTED);
1836 }
1837
1838 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1839 next_unlinked_offset);
1840 *buffer_nextp = *logged_nextp;
1841 }
1842
1843 return 0;
1844}
1845
1846/*
1847 * Perform a 'normal' buffer recovery. Each logged region of the
1848 * buffer should be copied over the corresponding region in the
1849 * given buffer. The bitmap in the buf log format structure indicates
1850 * where to place the logged data.
1851 */
1852STATIC void
1853xlog_recover_do_reg_buffer(
1854 struct xfs_mount *mp,
1855 xlog_recover_item_t *item,
1856 struct xfs_buf *bp,
1857 xfs_buf_log_format_t *buf_f)
1858{
1859 int i;
1860 int bit;
1861 int nbits;
1862 int error;
1863
1864 trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
1865
1866 bit = 0;
1867 i = 1; /* 0 is the buf format structure */
1868 while (1) {
1869 bit = xfs_next_bit(buf_f->blf_data_map,
1870 buf_f->blf_map_size, bit);
1871 if (bit == -1)
1872 break;
1873 nbits = xfs_contig_bits(buf_f->blf_data_map,
1874 buf_f->blf_map_size, bit);
1875 ASSERT(nbits > 0);
1876 ASSERT(item->ri_buf[i].i_addr != NULL);
1877 ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
1878 ASSERT(XFS_BUF_COUNT(bp) >=
1879 ((uint)bit << XFS_BLF_SHIFT)+(nbits<<XFS_BLF_SHIFT));
1880
1881 /*
1882 * Do a sanity check if this is a dquot buffer. Just checking
1883 * the first dquot in the buffer should do. XXXThis is
1884 * probably a good thing to do for other buf types also.
1885 */
1886 error = 0;
1887 if (buf_f->blf_flags &
1888 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
1889 if (item->ri_buf[i].i_addr == NULL) {
1890 xfs_alert(mp,
1891 "XFS: NULL dquot in %s.", __func__);
1892 goto next;
1893 }
1894 if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
1895 xfs_alert(mp,
1896 "XFS: dquot too small (%d) in %s.",
1897 item->ri_buf[i].i_len, __func__);
1898 goto next;
1899 }
1900 error = xfs_qm_dqcheck(mp, item->ri_buf[i].i_addr,
1901 -1, 0, XFS_QMOPT_DOWARN,
1902 "dquot_buf_recover");
1903 if (error)
1904 goto next;
1905 }
1906
1907 memcpy(xfs_buf_offset(bp,
1908 (uint)bit << XFS_BLF_SHIFT), /* dest */
1909 item->ri_buf[i].i_addr, /* source */
1910 nbits<<XFS_BLF_SHIFT); /* length */
1911 next:
1912 i++;
1913 bit += nbits;
1914 }
1915
1916 /* Shouldn't be any more regions */
1917 ASSERT(i == item->ri_total);
1918}
1919
1920/*
1921 * Do some primitive error checking on ondisk dquot data structures.
1922 */
1923int
1924xfs_qm_dqcheck(
1925 struct xfs_mount *mp,
1926 xfs_disk_dquot_t *ddq,
1927 xfs_dqid_t id,
1928 uint type, /* used only when IO_dorepair is true */
1929 uint flags,
1930 char *str)
1931{
1932 xfs_dqblk_t *d = (xfs_dqblk_t *)ddq;
1933 int errs = 0;
1934
1935 /*
1936 * We can encounter an uninitialized dquot buffer for 2 reasons:
1937 * 1. If we crash while deleting the quotainode(s), and those blks got
1938 * used for user data. This is because we take the path of regular
1939 * file deletion; however, the size field of quotainodes is never
1940 * updated, so all the tricks that we play in itruncate_finish
1941 * don't quite matter.
1942 *
1943 * 2. We don't play the quota buffers when there's a quotaoff logitem.
1944 * But the allocation will be replayed so we'll end up with an
1945 * uninitialized quota block.
1946 *
1947 * This is all fine; things are still consistent, and we haven't lost
1948 * any quota information. Just don't complain about bad dquot blks.
1949 */
1950 if (ddq->d_magic != cpu_to_be16(XFS_DQUOT_MAGIC)) {
1951 if (flags & XFS_QMOPT_DOWARN)
1952 xfs_alert(mp,
1953 "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1954 str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
1955 errs++;
1956 }
1957 if (ddq->d_version != XFS_DQUOT_VERSION) {
1958 if (flags & XFS_QMOPT_DOWARN)
1959 xfs_alert(mp,
1960 "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1961 str, id, ddq->d_version, XFS_DQUOT_VERSION);
1962 errs++;
1963 }
1964
1965 if (ddq->d_flags != XFS_DQ_USER &&
1966 ddq->d_flags != XFS_DQ_PROJ &&
1967 ddq->d_flags != XFS_DQ_GROUP) {
1968 if (flags & XFS_QMOPT_DOWARN)
1969 xfs_alert(mp,
1970 "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1971 str, id, ddq->d_flags);
1972 errs++;
1973 }
1974
1975 if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
1976 if (flags & XFS_QMOPT_DOWARN)
1977 xfs_alert(mp,
1978 "%s : ondisk-dquot 0x%p, ID mismatch: "
1979 "0x%x expected, found id 0x%x",
1980 str, ddq, id, be32_to_cpu(ddq->d_id));
1981 errs++;
1982 }
1983
1984 if (!errs && ddq->d_id) {
1985 if (ddq->d_blk_softlimit &&
1986 be64_to_cpu(ddq->d_bcount) >=
1987 be64_to_cpu(ddq->d_blk_softlimit)) {
1988 if (!ddq->d_btimer) {
1989 if (flags & XFS_QMOPT_DOWARN)
1990 xfs_alert(mp,
1991 "%s : Dquot ID 0x%x (0x%p) BLK TIMER NOT STARTED",
1992 str, (int)be32_to_cpu(ddq->d_id), ddq);
1993 errs++;
1994 }
1995 }
1996 if (ddq->d_ino_softlimit &&
1997 be64_to_cpu(ddq->d_icount) >=
1998 be64_to_cpu(ddq->d_ino_softlimit)) {
1999 if (!ddq->d_itimer) {
2000 if (flags & XFS_QMOPT_DOWARN)
2001 xfs_alert(mp,
2002 "%s : Dquot ID 0x%x (0x%p) INODE TIMER NOT STARTED",
2003 str, (int)be32_to_cpu(ddq->d_id), ddq);
2004 errs++;
2005 }
2006 }
2007 if (ddq->d_rtb_softlimit &&
2008 be64_to_cpu(ddq->d_rtbcount) >=
2009 be64_to_cpu(ddq->d_rtb_softlimit)) {
2010 if (!ddq->d_rtbtimer) {
2011 if (flags & XFS_QMOPT_DOWARN)
2012 xfs_alert(mp,
2013 "%s : Dquot ID 0x%x (0x%p) RTBLK TIMER NOT STARTED",
2014 str, (int)be32_to_cpu(ddq->d_id), ddq);
2015 errs++;
2016 }
2017 }
2018 }
2019
2020 if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2021 return errs;
2022
2023 if (flags & XFS_QMOPT_DOWARN)
2024 xfs_notice(mp, "Re-initializing dquot ID 0x%x", id);
2025
2026 /*
2027 * Typically, a repair is only requested by quotacheck.
2028 */
2029 ASSERT(id != -1);
2030 ASSERT(flags & XFS_QMOPT_DQREPAIR);
2031 memset(d, 0, sizeof(xfs_dqblk_t));
2032
2033 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2034 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2035 d->dd_diskdq.d_flags = type;
2036 d->dd_diskdq.d_id = cpu_to_be32(id);
2037
2038 return errs;
2039}
2040
2041/*
2042 * Perform a dquot buffer recovery.
2043 * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2044 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2045 * Else, treat it as a regular buffer and do recovery.
2046 */
2047STATIC void
2048xlog_recover_do_dquot_buffer(
2049 xfs_mount_t *mp,
2050 xlog_t *log,
2051 xlog_recover_item_t *item,
2052 xfs_buf_t *bp,
2053 xfs_buf_log_format_t *buf_f)
2054{
2055 uint type;
2056
2057 trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2058
2059 /*
2060 * Filesystems are required to send in quota flags at mount time.
2061 */
2062 if (mp->m_qflags == 0) {
2063 return;
2064 }
2065
2066 type = 0;
2067 if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2068 type |= XFS_DQ_USER;
2069 if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2070 type |= XFS_DQ_PROJ;
2071 if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2072 type |= XFS_DQ_GROUP;
2073 /*
2074 * This type of quotas was turned off, so ignore this buffer
2075 */
2076 if (log->l_quotaoffs_flag & type)
2077 return;
2078
2079 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2080}
2081
2082/*
2083 * This routine replays a modification made to a buffer at runtime.
2084 * There are actually two types of buffer, regular and inode, which
2085 * are handled differently. Inode buffers are handled differently
2086 * in that we only recover a specific set of data from them, namely
2087 * the inode di_next_unlinked fields. This is because all other inode
2088 * data is actually logged via inode records and any data we replay
2089 * here which overlaps that may be stale.
2090 *
2091 * When meta-data buffers are freed at run time we log a buffer item
2092 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2093 * of the buffer in the log should not be replayed at recovery time.
2094 * This is so that if the blocks covered by the buffer are reused for
2095 * file data before we crash we don't end up replaying old, freed
2096 * meta-data into a user's file.
2097 *
2098 * To handle the cancellation of buffer log items, we make two passes
2099 * over the log during recovery. During the first we build a table of
2100 * those buffers which have been cancelled, and during the second we
2101 * only replay those buffers which do not have corresponding cancel
2102 * records in the table. See xlog_recover_do_buffer_pass[1,2] above
2103 * for more details on the implementation of the table of cancel records.
2104 */
2105STATIC int
2106xlog_recover_buffer_pass2(
2107 xlog_t *log,
2108 xlog_recover_item_t *item)
2109{
2110 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
2111 xfs_mount_t *mp = log->l_mp;
2112 xfs_buf_t *bp;
2113 int error;
2114 uint buf_flags;
2115
2116 /*
2117 * In this pass we only want to recover all the buffers which have
2118 * not been cancelled and are not cancellation buffers themselves.
2119 */
2120 if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2121 buf_f->blf_len, buf_f->blf_flags)) {
2122 trace_xfs_log_recover_buf_cancel(log, buf_f);
2123 return 0;
2124 }
2125
2126 trace_xfs_log_recover_buf_recover(log, buf_f);
2127
2128 buf_flags = XBF_LOCK;
2129 if (!(buf_f->blf_flags & XFS_BLF_INODE_BUF))
2130 buf_flags |= XBF_MAPPED;
2131
2132 bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2133 buf_flags);
2134 if (!bp)
2135 return XFS_ERROR(ENOMEM);
2136 error = bp->b_error;
2137 if (error) {
2138 xfs_ioerror_alert("xlog_recover_do..(read#1)", mp,
2139 bp, buf_f->blf_blkno);
2140 xfs_buf_relse(bp);
2141 return error;
2142 }
2143
2144 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2145 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2146 } else if (buf_f->blf_flags &
2147 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2148 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2149 } else {
2150 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2151 }
2152 if (error)
2153 return XFS_ERROR(error);
2154
2155 /*
2156 * Perform delayed write on the buffer. Asynchronous writes will be
2157 * slower when taking into account all the buffers to be flushed.
2158 *
2159 * Also make sure that only inode buffers with good sizes stay in
2160 * the buffer cache. The kernel moves inodes in buffers of 1 block
2161 * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger. The inode
2162 * buffers in the log can be a different size if the log was generated
2163 * by an older kernel using unclustered inode buffers or a newer kernel
2164 * running with a different inode cluster size. Regardless, if the
2165 * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2166 * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2167 * the buffer out of the buffer cache so that the buffer won't
2168 * overlap with future reads of those inodes.
2169 */
2170 if (XFS_DINODE_MAGIC ==
2171 be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2172 (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2173 (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2174 XFS_BUF_STALE(bp);
2175 error = xfs_bwrite(mp, bp);
2176 } else {
2177 ASSERT(bp->b_target->bt_mount == mp);
2178 bp->b_iodone = xlog_recover_iodone;
2179 xfs_bdwrite(mp, bp);
2180 }
2181
2182 return (error);
2183}
2184
2185STATIC int
2186xlog_recover_inode_pass2(
2187 xlog_t *log,
2188 xlog_recover_item_t *item)
2189{
2190 xfs_inode_log_format_t *in_f;
2191 xfs_mount_t *mp = log->l_mp;
2192 xfs_buf_t *bp;
2193 xfs_dinode_t *dip;
2194 int len;
2195 xfs_caddr_t src;
2196 xfs_caddr_t dest;
2197 int error;
2198 int attr_index;
2199 uint fields;
2200 xfs_icdinode_t *dicp;
2201 int need_free = 0;
2202
2203 if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2204 in_f = item->ri_buf[0].i_addr;
2205 } else {
2206 in_f = kmem_alloc(sizeof(xfs_inode_log_format_t), KM_SLEEP);
2207 need_free = 1;
2208 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2209 if (error)
2210 goto error;
2211 }
2212
2213 /*
2214 * Inode buffers can be freed, look out for it,
2215 * and do not replay the inode.
2216 */
2217 if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2218 in_f->ilf_len, 0)) {
2219 error = 0;
2220 trace_xfs_log_recover_inode_cancel(log, in_f);
2221 goto error;
2222 }
2223 trace_xfs_log_recover_inode_recover(log, in_f);
2224
2225 bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len,
2226 XBF_LOCK);
2227 if (!bp) {
2228 error = ENOMEM;
2229 goto error;
2230 }
2231 error = bp->b_error;
2232 if (error) {
2233 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2234 bp, in_f->ilf_blkno);
2235 xfs_buf_relse(bp);
2236 goto error;
2237 }
2238 ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2239 dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
2240
2241 /*
2242 * Make sure the place we're flushing out to really looks
2243 * like an inode!
2244 */
2245 if (unlikely(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC))) {
2246 xfs_buf_relse(bp);
2247 xfs_alert(mp,
2248 "%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
2249 __func__, dip, bp, in_f->ilf_ino);
2250 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2251 XFS_ERRLEVEL_LOW, mp);
2252 error = EFSCORRUPTED;
2253 goto error;
2254 }
2255 dicp = item->ri_buf[1].i_addr;
2256 if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2257 xfs_buf_relse(bp);
2258 xfs_alert(mp,
2259 "%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
2260 __func__, item, in_f->ilf_ino);
2261 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2262 XFS_ERRLEVEL_LOW, mp);
2263 error = EFSCORRUPTED;
2264 goto error;
2265 }
2266
2267 /* Skip replay when the on disk inode is newer than the log one */
2268 if (dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2269 /*
2270 * Deal with the wrap case, DI_MAX_FLUSH is less
2271 * than smaller numbers
2272 */
2273 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
2274 dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2275 /* do nothing */
2276 } else {
2277 xfs_buf_relse(bp);
2278 trace_xfs_log_recover_inode_skip(log, in_f);
2279 error = 0;
2280 goto error;
2281 }
2282 }
2283 /* Take the opportunity to reset the flush iteration count */
2284 dicp->di_flushiter = 0;
2285
2286 if (unlikely(S_ISREG(dicp->di_mode))) {
2287 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2288 (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2289 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
2290 XFS_ERRLEVEL_LOW, mp, dicp);
2291 xfs_buf_relse(bp);
2292 xfs_alert(mp,
2293 "%s: Bad regular inode log record, rec ptr 0x%p, "
2294 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2295 __func__, item, dip, bp, in_f->ilf_ino);
2296 error = EFSCORRUPTED;
2297 goto error;
2298 }
2299 } else if (unlikely(S_ISDIR(dicp->di_mode))) {
2300 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2301 (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2302 (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2303 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
2304 XFS_ERRLEVEL_LOW, mp, dicp);
2305 xfs_buf_relse(bp);
2306 xfs_alert(mp,
2307 "%s: Bad dir inode log record, rec ptr 0x%p, "
2308 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2309 __func__, item, dip, bp, in_f->ilf_ino);
2310 error = EFSCORRUPTED;
2311 goto error;
2312 }
2313 }
2314 if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2315 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
2316 XFS_ERRLEVEL_LOW, mp, dicp);
2317 xfs_buf_relse(bp);
2318 xfs_alert(mp,
2319 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2320 "dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2321 __func__, item, dip, bp, in_f->ilf_ino,
2322 dicp->di_nextents + dicp->di_anextents,
2323 dicp->di_nblocks);
2324 error = EFSCORRUPTED;
2325 goto error;
2326 }
2327 if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2328 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
2329 XFS_ERRLEVEL_LOW, mp, dicp);
2330 xfs_buf_relse(bp);
2331 xfs_alert(mp,
2332 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
2333 "dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__,
2334 item, dip, bp, in_f->ilf_ino, dicp->di_forkoff);
2335 error = EFSCORRUPTED;
2336 goto error;
2337 }
2338 if (unlikely(item->ri_buf[1].i_len > sizeof(struct xfs_icdinode))) {
2339 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
2340 XFS_ERRLEVEL_LOW, mp, dicp);
2341 xfs_buf_relse(bp);
2342 xfs_alert(mp,
2343 "%s: Bad inode log record length %d, rec ptr 0x%p",
2344 __func__, item->ri_buf[1].i_len, item);
2345 error = EFSCORRUPTED;
2346 goto error;
2347 }
2348
2349 /* The core is in in-core format */
2350 xfs_dinode_to_disk(dip, item->ri_buf[1].i_addr);
2351
2352 /* the rest is in on-disk format */
2353 if (item->ri_buf[1].i_len > sizeof(struct xfs_icdinode)) {
2354 memcpy((xfs_caddr_t) dip + sizeof(struct xfs_icdinode),
2355 item->ri_buf[1].i_addr + sizeof(struct xfs_icdinode),
2356 item->ri_buf[1].i_len - sizeof(struct xfs_icdinode));
2357 }
2358
2359 fields = in_f->ilf_fields;
2360 switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2361 case XFS_ILOG_DEV:
2362 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
2363 break;
2364 case XFS_ILOG_UUID:
2365 memcpy(XFS_DFORK_DPTR(dip),
2366 &in_f->ilf_u.ilfu_uuid,
2367 sizeof(uuid_t));
2368 break;
2369 }
2370
2371 if (in_f->ilf_size == 2)
2372 goto write_inode_buffer;
2373 len = item->ri_buf[2].i_len;
2374 src = item->ri_buf[2].i_addr;
2375 ASSERT(in_f->ilf_size <= 4);
2376 ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2377 ASSERT(!(fields & XFS_ILOG_DFORK) ||
2378 (len == in_f->ilf_dsize));
2379
2380 switch (fields & XFS_ILOG_DFORK) {
2381 case XFS_ILOG_DDATA:
2382 case XFS_ILOG_DEXT:
2383 memcpy(XFS_DFORK_DPTR(dip), src, len);
2384 break;
2385
2386 case XFS_ILOG_DBROOT:
2387 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2388 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
2389 XFS_DFORK_DSIZE(dip, mp));
2390 break;
2391
2392 default:
2393 /*
2394 * There are no data fork flags set.
2395 */
2396 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2397 break;
2398 }
2399
2400 /*
2401 * If we logged any attribute data, recover it. There may or
2402 * may not have been any other non-core data logged in this
2403 * transaction.
2404 */
2405 if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2406 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2407 attr_index = 3;
2408 } else {
2409 attr_index = 2;
2410 }
2411 len = item->ri_buf[attr_index].i_len;
2412 src = item->ri_buf[attr_index].i_addr;
2413 ASSERT(len == in_f->ilf_asize);
2414
2415 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2416 case XFS_ILOG_ADATA:
2417 case XFS_ILOG_AEXT:
2418 dest = XFS_DFORK_APTR(dip);
2419 ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2420 memcpy(dest, src, len);
2421 break;
2422
2423 case XFS_ILOG_ABROOT:
2424 dest = XFS_DFORK_APTR(dip);
2425 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2426 len, (xfs_bmdr_block_t*)dest,
2427 XFS_DFORK_ASIZE(dip, mp));
2428 break;
2429
2430 default:
2431 xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
2432 ASSERT(0);
2433 xfs_buf_relse(bp);
2434 error = EIO;
2435 goto error;
2436 }
2437 }
2438
2439write_inode_buffer:
2440 ASSERT(bp->b_target->bt_mount == mp);
2441 bp->b_iodone = xlog_recover_iodone;
2442 xfs_bdwrite(mp, bp);
2443error:
2444 if (need_free)
2445 kmem_free(in_f);
2446 return XFS_ERROR(error);
2447}
2448
2449/*
2450 * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2451 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2452 * of that type.
2453 */
2454STATIC int
2455xlog_recover_quotaoff_pass1(
2456 xlog_t *log,
2457 xlog_recover_item_t *item)
2458{
2459 xfs_qoff_logformat_t *qoff_f = item->ri_buf[0].i_addr;
2460 ASSERT(qoff_f);
2461
2462 /*
2463 * The logitem format's flag tells us if this was user quotaoff,
2464 * group/project quotaoff or both.
2465 */
2466 if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2467 log->l_quotaoffs_flag |= XFS_DQ_USER;
2468 if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2469 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2470 if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2471 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2472
2473 return (0);
2474}
2475
2476/*
2477 * Recover a dquot record
2478 */
2479STATIC int
2480xlog_recover_dquot_pass2(
2481 xlog_t *log,
2482 xlog_recover_item_t *item)
2483{
2484 xfs_mount_t *mp = log->l_mp;
2485 xfs_buf_t *bp;
2486 struct xfs_disk_dquot *ddq, *recddq;
2487 int error;
2488 xfs_dq_logformat_t *dq_f;
2489 uint type;
2490
2491
2492 /*
2493 * Filesystems are required to send in quota flags at mount time.
2494 */
2495 if (mp->m_qflags == 0)
2496 return (0);
2497
2498 recddq = item->ri_buf[1].i_addr;
2499 if (recddq == NULL) {
2500 xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
2501 return XFS_ERROR(EIO);
2502 }
2503 if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
2504 xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
2505 item->ri_buf[1].i_len, __func__);
2506 return XFS_ERROR(EIO);
2507 }
2508
2509 /*
2510 * This type of quotas was turned off, so ignore this record.
2511 */
2512 type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2513 ASSERT(type);
2514 if (log->l_quotaoffs_flag & type)
2515 return (0);
2516
2517 /*
2518 * At this point we know that quota was _not_ turned off.
2519 * Since the mount flags are not indicating to us otherwise, this
2520 * must mean that quota is on, and the dquot needs to be replayed.
2521 * Remember that we may not have fully recovered the superblock yet,
2522 * so we can't do the usual trick of looking at the SB quota bits.
2523 *
2524 * The other possibility, of course, is that the quota subsystem was
2525 * removed since the last mount - ENOSYS.
2526 */
2527 dq_f = item->ri_buf[0].i_addr;
2528 ASSERT(dq_f);
2529 error = xfs_qm_dqcheck(mp, recddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2530 "xlog_recover_dquot_pass2 (log copy)");
2531 if (error)
2532 return XFS_ERROR(EIO);
2533 ASSERT(dq_f->qlf_len == 1);
2534
2535 error = xfs_read_buf(mp, mp->m_ddev_targp,
2536 dq_f->qlf_blkno,
2537 XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2538 0, &bp);
2539 if (error) {
2540 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2541 bp, dq_f->qlf_blkno);
2542 return error;
2543 }
2544 ASSERT(bp);
2545 ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2546
2547 /*
2548 * At least the magic num portion should be on disk because this
2549 * was among a chunk of dquots created earlier, and we did some
2550 * minimal initialization then.
2551 */
2552 error = xfs_qm_dqcheck(mp, ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2553 "xlog_recover_dquot_pass2");
2554 if (error) {
2555 xfs_buf_relse(bp);
2556 return XFS_ERROR(EIO);
2557 }
2558
2559 memcpy(ddq, recddq, item->ri_buf[1].i_len);
2560
2561 ASSERT(dq_f->qlf_size == 2);
2562 ASSERT(bp->b_target->bt_mount == mp);
2563 bp->b_iodone = xlog_recover_iodone;
2564 xfs_bdwrite(mp, bp);
2565
2566 return (0);
2567}
2568
2569/*
2570 * This routine is called to create an in-core extent free intent
2571 * item from the efi format structure which was logged on disk.
2572 * It allocates an in-core efi, copies the extents from the format
2573 * structure into it, and adds the efi to the AIL with the given
2574 * LSN.
2575 */
2576STATIC int
2577xlog_recover_efi_pass2(
2578 xlog_t *log,
2579 xlog_recover_item_t *item,
2580 xfs_lsn_t lsn)
2581{
2582 int error;
2583 xfs_mount_t *mp = log->l_mp;
2584 xfs_efi_log_item_t *efip;
2585 xfs_efi_log_format_t *efi_formatp;
2586
2587 efi_formatp = item->ri_buf[0].i_addr;
2588
2589 efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2590 if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2591 &(efip->efi_format)))) {
2592 xfs_efi_item_free(efip);
2593 return error;
2594 }
2595 atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
2596
2597 spin_lock(&log->l_ailp->xa_lock);
2598 /*
2599 * xfs_trans_ail_update() drops the AIL lock.
2600 */
2601 xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
2602 return 0;
2603}
2604
2605
2606/*
2607 * This routine is called when an efd format structure is found in
2608 * a committed transaction in the log. It's purpose is to cancel
2609 * the corresponding efi if it was still in the log. To do this
2610 * it searches the AIL for the efi with an id equal to that in the
2611 * efd format structure. If we find it, we remove the efi from the
2612 * AIL and free it.
2613 */
2614STATIC int
2615xlog_recover_efd_pass2(
2616 xlog_t *log,
2617 xlog_recover_item_t *item)
2618{
2619 xfs_efd_log_format_t *efd_formatp;
2620 xfs_efi_log_item_t *efip = NULL;
2621 xfs_log_item_t *lip;
2622 __uint64_t efi_id;
2623 struct xfs_ail_cursor cur;
2624 struct xfs_ail *ailp = log->l_ailp;
2625
2626 efd_formatp = item->ri_buf[0].i_addr;
2627 ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2628 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2629 (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2630 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2631 efi_id = efd_formatp->efd_efi_id;
2632
2633 /*
2634 * Search for the efi with the id in the efd format structure
2635 * in the AIL.
2636 */
2637 spin_lock(&ailp->xa_lock);
2638 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2639 while (lip != NULL) {
2640 if (lip->li_type == XFS_LI_EFI) {
2641 efip = (xfs_efi_log_item_t *)lip;
2642 if (efip->efi_format.efi_id == efi_id) {
2643 /*
2644 * xfs_trans_ail_delete() drops the
2645 * AIL lock.
2646 */
2647 xfs_trans_ail_delete(ailp, lip);
2648 xfs_efi_item_free(efip);
2649 spin_lock(&ailp->xa_lock);
2650 break;
2651 }
2652 }
2653 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2654 }
2655 xfs_trans_ail_cursor_done(ailp, &cur);
2656 spin_unlock(&ailp->xa_lock);
2657
2658 return 0;
2659}
2660
2661/*
2662 * Free up any resources allocated by the transaction
2663 *
2664 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2665 */
2666STATIC void
2667xlog_recover_free_trans(
2668 struct xlog_recover *trans)
2669{
2670 xlog_recover_item_t *item, *n;
2671 int i;
2672
2673 list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2674 /* Free the regions in the item. */
2675 list_del(&item->ri_list);
2676 for (i = 0; i < item->ri_cnt; i++)
2677 kmem_free(item->ri_buf[i].i_addr);
2678 /* Free the item itself */
2679 kmem_free(item->ri_buf);
2680 kmem_free(item);
2681 }
2682 /* Free the transaction recover structure */
2683 kmem_free(trans);
2684}
2685
2686STATIC int
2687xlog_recover_commit_pass1(
2688 struct log *log,
2689 struct xlog_recover *trans,
2690 xlog_recover_item_t *item)
2691{
2692 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
2693
2694 switch (ITEM_TYPE(item)) {
2695 case XFS_LI_BUF:
2696 return xlog_recover_buffer_pass1(log, item);
2697 case XFS_LI_QUOTAOFF:
2698 return xlog_recover_quotaoff_pass1(log, item);
2699 case XFS_LI_INODE:
2700 case XFS_LI_EFI:
2701 case XFS_LI_EFD:
2702 case XFS_LI_DQUOT:
2703 /* nothing to do in pass 1 */
2704 return 0;
2705 default:
2706 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
2707 __func__, ITEM_TYPE(item));
2708 ASSERT(0);
2709 return XFS_ERROR(EIO);
2710 }
2711}
2712
2713STATIC int
2714xlog_recover_commit_pass2(
2715 struct log *log,
2716 struct xlog_recover *trans,
2717 xlog_recover_item_t *item)
2718{
2719 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
2720
2721 switch (ITEM_TYPE(item)) {
2722 case XFS_LI_BUF:
2723 return xlog_recover_buffer_pass2(log, item);
2724 case XFS_LI_INODE:
2725 return xlog_recover_inode_pass2(log, item);
2726 case XFS_LI_EFI:
2727 return xlog_recover_efi_pass2(log, item, trans->r_lsn);
2728 case XFS_LI_EFD:
2729 return xlog_recover_efd_pass2(log, item);
2730 case XFS_LI_DQUOT:
2731 return xlog_recover_dquot_pass2(log, item);
2732 case XFS_LI_QUOTAOFF:
2733 /* nothing to do in pass2 */
2734 return 0;
2735 default:
2736 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
2737 __func__, ITEM_TYPE(item));
2738 ASSERT(0);
2739 return XFS_ERROR(EIO);
2740 }
2741}
2742
2743/*
2744 * Perform the transaction.
2745 *
2746 * If the transaction modifies a buffer or inode, do it now. Otherwise,
2747 * EFIs and EFDs get queued up by adding entries into the AIL for them.
2748 */
2749STATIC int
2750xlog_recover_commit_trans(
2751 struct log *log,
2752 struct xlog_recover *trans,
2753 int pass)
2754{
2755 int error = 0;
2756 xlog_recover_item_t *item;
2757
2758 hlist_del(&trans->r_list);
2759
2760 error = xlog_recover_reorder_trans(log, trans, pass);
2761 if (error)
2762 return error;
2763
2764 list_for_each_entry(item, &trans->r_itemq, ri_list) {
2765 if (pass == XLOG_RECOVER_PASS1)
2766 error = xlog_recover_commit_pass1(log, trans, item);
2767 else
2768 error = xlog_recover_commit_pass2(log, trans, item);
2769 if (error)
2770 return error;
2771 }
2772
2773 xlog_recover_free_trans(trans);
2774 return 0;
2775}
2776
2777STATIC int
2778xlog_recover_unmount_trans(
2779 struct log *log,
2780 xlog_recover_t *trans)
2781{
2782 /* Do nothing now */
2783 xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
2784 return 0;
2785}
2786
2787/*
2788 * There are two valid states of the r_state field. 0 indicates that the
2789 * transaction structure is in a normal state. We have either seen the
2790 * start of the transaction or the last operation we added was not a partial
2791 * operation. If the last operation we added to the transaction was a
2792 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2793 *
2794 * NOTE: skip LRs with 0 data length.
2795 */
2796STATIC int
2797xlog_recover_process_data(
2798 xlog_t *log,
2799 struct hlist_head rhash[],
2800 xlog_rec_header_t *rhead,
2801 xfs_caddr_t dp,
2802 int pass)
2803{
2804 xfs_caddr_t lp;
2805 int num_logops;
2806 xlog_op_header_t *ohead;
2807 xlog_recover_t *trans;
2808 xlog_tid_t tid;
2809 int error;
2810 unsigned long hash;
2811 uint flags;
2812
2813 lp = dp + be32_to_cpu(rhead->h_len);
2814 num_logops = be32_to_cpu(rhead->h_num_logops);
2815
2816 /* check the log format matches our own - else we can't recover */
2817 if (xlog_header_check_recover(log->l_mp, rhead))
2818 return (XFS_ERROR(EIO));
2819
2820 while ((dp < lp) && num_logops) {
2821 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2822 ohead = (xlog_op_header_t *)dp;
2823 dp += sizeof(xlog_op_header_t);
2824 if (ohead->oh_clientid != XFS_TRANSACTION &&
2825 ohead->oh_clientid != XFS_LOG) {
2826 xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
2827 __func__, ohead->oh_clientid);
2828 ASSERT(0);
2829 return (XFS_ERROR(EIO));
2830 }
2831 tid = be32_to_cpu(ohead->oh_tid);
2832 hash = XLOG_RHASH(tid);
2833 trans = xlog_recover_find_tid(&rhash[hash], tid);
2834 if (trans == NULL) { /* not found; add new tid */
2835 if (ohead->oh_flags & XLOG_START_TRANS)
2836 xlog_recover_new_tid(&rhash[hash], tid,
2837 be64_to_cpu(rhead->h_lsn));
2838 } else {
2839 if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2840 xfs_warn(log->l_mp, "%s: bad length 0x%x",
2841 __func__, be32_to_cpu(ohead->oh_len));
2842 WARN_ON(1);
2843 return (XFS_ERROR(EIO));
2844 }
2845 flags = ohead->oh_flags & ~XLOG_END_TRANS;
2846 if (flags & XLOG_WAS_CONT_TRANS)
2847 flags &= ~XLOG_CONTINUE_TRANS;
2848 switch (flags) {
2849 case XLOG_COMMIT_TRANS:
2850 error = xlog_recover_commit_trans(log,
2851 trans, pass);
2852 break;
2853 case XLOG_UNMOUNT_TRANS:
2854 error = xlog_recover_unmount_trans(log, trans);
2855 break;
2856 case XLOG_WAS_CONT_TRANS:
2857 error = xlog_recover_add_to_cont_trans(log,
2858 trans, dp,
2859 be32_to_cpu(ohead->oh_len));
2860 break;
2861 case XLOG_START_TRANS:
2862 xfs_warn(log->l_mp, "%s: bad transaction",
2863 __func__);
2864 ASSERT(0);
2865 error = XFS_ERROR(EIO);
2866 break;
2867 case 0:
2868 case XLOG_CONTINUE_TRANS:
2869 error = xlog_recover_add_to_trans(log, trans,
2870 dp, be32_to_cpu(ohead->oh_len));
2871 break;
2872 default:
2873 xfs_warn(log->l_mp, "%s: bad flag 0x%x",
2874 __func__, flags);
2875 ASSERT(0);
2876 error = XFS_ERROR(EIO);
2877 break;
2878 }
2879 if (error)
2880 return error;
2881 }
2882 dp += be32_to_cpu(ohead->oh_len);
2883 num_logops--;
2884 }
2885 return 0;
2886}
2887
2888/*
2889 * Process an extent free intent item that was recovered from
2890 * the log. We need to free the extents that it describes.
2891 */
2892STATIC int
2893xlog_recover_process_efi(
2894 xfs_mount_t *mp,
2895 xfs_efi_log_item_t *efip)
2896{
2897 xfs_efd_log_item_t *efdp;
2898 xfs_trans_t *tp;
2899 int i;
2900 int error = 0;
2901 xfs_extent_t *extp;
2902 xfs_fsblock_t startblock_fsb;
2903
2904 ASSERT(!test_bit(XFS_EFI_RECOVERED, &efip->efi_flags));
2905
2906 /*
2907 * First check the validity of the extents described by the
2908 * EFI. If any are bad, then assume that all are bad and
2909 * just toss the EFI.
2910 */
2911 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2912 extp = &(efip->efi_format.efi_extents[i]);
2913 startblock_fsb = XFS_BB_TO_FSB(mp,
2914 XFS_FSB_TO_DADDR(mp, extp->ext_start));
2915 if ((startblock_fsb == 0) ||
2916 (extp->ext_len == 0) ||
2917 (startblock_fsb >= mp->m_sb.sb_dblocks) ||
2918 (extp->ext_len >= mp->m_sb.sb_agblocks)) {
2919 /*
2920 * This will pull the EFI from the AIL and
2921 * free the memory associated with it.
2922 */
2923 xfs_efi_release(efip, efip->efi_format.efi_nextents);
2924 return XFS_ERROR(EIO);
2925 }
2926 }
2927
2928 tp = xfs_trans_alloc(mp, 0);
2929 error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
2930 if (error)
2931 goto abort_error;
2932 efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
2933
2934 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2935 extp = &(efip->efi_format.efi_extents[i]);
2936 error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
2937 if (error)
2938 goto abort_error;
2939 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
2940 extp->ext_len);
2941 }
2942
2943 set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
2944 error = xfs_trans_commit(tp, 0);
2945 return error;
2946
2947abort_error:
2948 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
2949 return error;
2950}
2951
2952/*
2953 * When this is called, all of the EFIs which did not have
2954 * corresponding EFDs should be in the AIL. What we do now
2955 * is free the extents associated with each one.
2956 *
2957 * Since we process the EFIs in normal transactions, they
2958 * will be removed at some point after the commit. This prevents
2959 * us from just walking down the list processing each one.
2960 * We'll use a flag in the EFI to skip those that we've already
2961 * processed and use the AIL iteration mechanism's generation
2962 * count to try to speed this up at least a bit.
2963 *
2964 * When we start, we know that the EFIs are the only things in
2965 * the AIL. As we process them, however, other items are added
2966 * to the AIL. Since everything added to the AIL must come after
2967 * everything already in the AIL, we stop processing as soon as
2968 * we see something other than an EFI in the AIL.
2969 */
2970STATIC int
2971xlog_recover_process_efis(
2972 xlog_t *log)
2973{
2974 xfs_log_item_t *lip;
2975 xfs_efi_log_item_t *efip;
2976 int error = 0;
2977 struct xfs_ail_cursor cur;
2978 struct xfs_ail *ailp;
2979
2980 ailp = log->l_ailp;
2981 spin_lock(&ailp->xa_lock);
2982 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2983 while (lip != NULL) {
2984 /*
2985 * We're done when we see something other than an EFI.
2986 * There should be no EFIs left in the AIL now.
2987 */
2988 if (lip->li_type != XFS_LI_EFI) {
2989#ifdef DEBUG
2990 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2991 ASSERT(lip->li_type != XFS_LI_EFI);
2992#endif
2993 break;
2994 }
2995
2996 /*
2997 * Skip EFIs that we've already processed.
2998 */
2999 efip = (xfs_efi_log_item_t *)lip;
3000 if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags)) {
3001 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3002 continue;
3003 }
3004
3005 spin_unlock(&ailp->xa_lock);
3006 error = xlog_recover_process_efi(log->l_mp, efip);
3007 spin_lock(&ailp->xa_lock);
3008 if (error)
3009 goto out;
3010 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3011 }
3012out:
3013 xfs_trans_ail_cursor_done(ailp, &cur);
3014 spin_unlock(&ailp->xa_lock);
3015 return error;
3016}
3017
3018/*
3019 * This routine performs a transaction to null out a bad inode pointer
3020 * in an agi unlinked inode hash bucket.
3021 */
3022STATIC void
3023xlog_recover_clear_agi_bucket(
3024 xfs_mount_t *mp,
3025 xfs_agnumber_t agno,
3026 int bucket)
3027{
3028 xfs_trans_t *tp;
3029 xfs_agi_t *agi;
3030 xfs_buf_t *agibp;
3031 int offset;
3032 int error;
3033
3034 tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3035 error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp),
3036 0, 0, 0);
3037 if (error)
3038 goto out_abort;
3039
3040 error = xfs_read_agi(mp, tp, agno, &agibp);
3041 if (error)
3042 goto out_abort;
3043
3044 agi = XFS_BUF_TO_AGI(agibp);
3045 agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3046 offset = offsetof(xfs_agi_t, agi_unlinked) +
3047 (sizeof(xfs_agino_t) * bucket);
3048 xfs_trans_log_buf(tp, agibp, offset,
3049 (offset + sizeof(xfs_agino_t) - 1));
3050
3051 error = xfs_trans_commit(tp, 0);
3052 if (error)
3053 goto out_error;
3054 return;
3055
3056out_abort:
3057 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3058out_error:
3059 xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
3060 return;
3061}
3062
3063STATIC xfs_agino_t
3064xlog_recover_process_one_iunlink(
3065 struct xfs_mount *mp,
3066 xfs_agnumber_t agno,
3067 xfs_agino_t agino,
3068 int bucket)
3069{
3070 struct xfs_buf *ibp;
3071 struct xfs_dinode *dip;
3072 struct xfs_inode *ip;
3073 xfs_ino_t ino;
3074 int error;
3075
3076 ino = XFS_AGINO_TO_INO(mp, agno, agino);
3077 error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
3078 if (error)
3079 goto fail;
3080
3081 /*
3082 * Get the on disk inode to find the next inode in the bucket.
3083 */
3084 error = xfs_itobp(mp, NULL, ip, &dip, &ibp, XBF_LOCK);
3085 if (error)
3086 goto fail_iput;
3087
3088 ASSERT(ip->i_d.di_nlink == 0);
3089 ASSERT(ip->i_d.di_mode != 0);
3090
3091 /* setup for the next pass */
3092 agino = be32_to_cpu(dip->di_next_unlinked);
3093 xfs_buf_relse(ibp);
3094
3095 /*
3096 * Prevent any DMAPI event from being sent when the reference on
3097 * the inode is dropped.
3098 */
3099 ip->i_d.di_dmevmask = 0;
3100
3101 IRELE(ip);
3102 return agino;
3103
3104 fail_iput:
3105 IRELE(ip);
3106 fail:
3107 /*
3108 * We can't read in the inode this bucket points to, or this inode
3109 * is messed up. Just ditch this bucket of inodes. We will lose
3110 * some inodes and space, but at least we won't hang.
3111 *
3112 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3113 * clear the inode pointer in the bucket.
3114 */
3115 xlog_recover_clear_agi_bucket(mp, agno, bucket);
3116 return NULLAGINO;
3117}
3118
3119/*
3120 * xlog_iunlink_recover
3121 *
3122 * This is called during recovery to process any inodes which
3123 * we unlinked but not freed when the system crashed. These
3124 * inodes will be on the lists in the AGI blocks. What we do
3125 * here is scan all the AGIs and fully truncate and free any
3126 * inodes found on the lists. Each inode is removed from the
3127 * lists when it has been fully truncated and is freed. The
3128 * freeing of the inode and its removal from the list must be
3129 * atomic.
3130 */
3131STATIC void
3132xlog_recover_process_iunlinks(
3133 xlog_t *log)
3134{
3135 xfs_mount_t *mp;
3136 xfs_agnumber_t agno;
3137 xfs_agi_t *agi;
3138 xfs_buf_t *agibp;
3139 xfs_agino_t agino;
3140 int bucket;
3141 int error;
3142 uint mp_dmevmask;
3143
3144 mp = log->l_mp;
3145
3146 /*
3147 * Prevent any DMAPI event from being sent while in this function.
3148 */
3149 mp_dmevmask = mp->m_dmevmask;
3150 mp->m_dmevmask = 0;
3151
3152 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3153 /*
3154 * Find the agi for this ag.
3155 */
3156 error = xfs_read_agi(mp, NULL, agno, &agibp);
3157 if (error) {
3158 /*
3159 * AGI is b0rked. Don't process it.
3160 *
3161 * We should probably mark the filesystem as corrupt
3162 * after we've recovered all the ag's we can....
3163 */
3164 continue;
3165 }
3166 agi = XFS_BUF_TO_AGI(agibp);
3167
3168 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3169 agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3170 while (agino != NULLAGINO) {
3171 /*
3172 * Release the agi buffer so that it can
3173 * be acquired in the normal course of the
3174 * transaction to truncate and free the inode.
3175 */
3176 xfs_buf_relse(agibp);
3177
3178 agino = xlog_recover_process_one_iunlink(mp,
3179 agno, agino, bucket);
3180
3181 /*
3182 * Reacquire the agibuffer and continue around
3183 * the loop. This should never fail as we know
3184 * the buffer was good earlier on.
3185 */
3186 error = xfs_read_agi(mp, NULL, agno, &agibp);
3187 ASSERT(error == 0);
3188 agi = XFS_BUF_TO_AGI(agibp);
3189 }
3190 }
3191
3192 /*
3193 * Release the buffer for the current agi so we can
3194 * go on to the next one.
3195 */
3196 xfs_buf_relse(agibp);
3197 }
3198
3199 mp->m_dmevmask = mp_dmevmask;
3200}
3201
3202
3203#ifdef DEBUG
3204STATIC void
3205xlog_pack_data_checksum(
3206 xlog_t *log,
3207 xlog_in_core_t *iclog,
3208 int size)
3209{
3210 int i;
3211 __be32 *up;
3212 uint chksum = 0;
3213
3214 up = (__be32 *)iclog->ic_datap;
3215 /* divide length by 4 to get # words */
3216 for (i = 0; i < (size >> 2); i++) {
3217 chksum ^= be32_to_cpu(*up);
3218 up++;
3219 }
3220 iclog->ic_header.h_chksum = cpu_to_be32(chksum);
3221}
3222#else
3223#define xlog_pack_data_checksum(log, iclog, size)
3224#endif
3225
3226/*
3227 * Stamp cycle number in every block
3228 */
3229void
3230xlog_pack_data(
3231 xlog_t *log,
3232 xlog_in_core_t *iclog,
3233 int roundoff)
3234{
3235 int i, j, k;
3236 int size = iclog->ic_offset + roundoff;
3237 __be32 cycle_lsn;
3238 xfs_caddr_t dp;
3239
3240 xlog_pack_data_checksum(log, iclog, size);
3241
3242 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3243
3244 dp = iclog->ic_datap;
3245 for (i = 0; i < BTOBB(size) &&
3246 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3247 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
3248 *(__be32 *)dp = cycle_lsn;
3249 dp += BBSIZE;
3250 }
3251
3252 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3253 xlog_in_core_2_t *xhdr = iclog->ic_data;
3254
3255 for ( ; i < BTOBB(size); i++) {
3256 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3257 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3258 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
3259 *(__be32 *)dp = cycle_lsn;
3260 dp += BBSIZE;
3261 }
3262
3263 for (i = 1; i < log->l_iclog_heads; i++) {
3264 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3265 }
3266 }
3267}
3268
3269STATIC void
3270xlog_unpack_data(
3271 xlog_rec_header_t *rhead,
3272 xfs_caddr_t dp,
3273 xlog_t *log)
3274{
3275 int i, j, k;
3276
3277 for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3278 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3279 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3280 dp += BBSIZE;
3281 }
3282
3283 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3284 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
3285 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3286 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3287 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3288 *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3289 dp += BBSIZE;
3290 }
3291 }
3292}
3293
3294STATIC int
3295xlog_valid_rec_header(
3296 xlog_t *log,
3297 xlog_rec_header_t *rhead,
3298 xfs_daddr_t blkno)
3299{
3300 int hlen;
3301
3302 if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
3303 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3304 XFS_ERRLEVEL_LOW, log->l_mp);
3305 return XFS_ERROR(EFSCORRUPTED);
3306 }
3307 if (unlikely(
3308 (!rhead->h_version ||
3309 (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3310 xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
3311 __func__, be32_to_cpu(rhead->h_version));
3312 return XFS_ERROR(EIO);
3313 }
3314
3315 /* LR body must have data or it wouldn't have been written */
3316 hlen = be32_to_cpu(rhead->h_len);
3317 if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3318 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3319 XFS_ERRLEVEL_LOW, log->l_mp);
3320 return XFS_ERROR(EFSCORRUPTED);
3321 }
3322 if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3323 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3324 XFS_ERRLEVEL_LOW, log->l_mp);
3325 return XFS_ERROR(EFSCORRUPTED);
3326 }
3327 return 0;
3328}
3329
3330/*
3331 * Read the log from tail to head and process the log records found.
3332 * Handle the two cases where the tail and head are in the same cycle
3333 * and where the active portion of the log wraps around the end of
3334 * the physical log separately. The pass parameter is passed through
3335 * to the routines called to process the data and is not looked at
3336 * here.
3337 */
3338STATIC int
3339xlog_do_recovery_pass(
3340 xlog_t *log,
3341 xfs_daddr_t head_blk,
3342 xfs_daddr_t tail_blk,
3343 int pass)
3344{
3345 xlog_rec_header_t *rhead;
3346 xfs_daddr_t blk_no;
3347 xfs_caddr_t offset;
3348 xfs_buf_t *hbp, *dbp;
3349 int error = 0, h_size;
3350 int bblks, split_bblks;
3351 int hblks, split_hblks, wrapped_hblks;
3352 struct hlist_head rhash[XLOG_RHASH_SIZE];
3353
3354 ASSERT(head_blk != tail_blk);
3355
3356 /*
3357 * Read the header of the tail block and get the iclog buffer size from
3358 * h_size. Use this to tell how many sectors make up the log header.
3359 */
3360 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3361 /*
3362 * When using variable length iclogs, read first sector of
3363 * iclog header and extract the header size from it. Get a
3364 * new hbp that is the correct size.
3365 */
3366 hbp = xlog_get_bp(log, 1);
3367 if (!hbp)
3368 return ENOMEM;
3369
3370 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
3371 if (error)
3372 goto bread_err1;
3373
3374 rhead = (xlog_rec_header_t *)offset;
3375 error = xlog_valid_rec_header(log, rhead, tail_blk);
3376 if (error)
3377 goto bread_err1;
3378 h_size = be32_to_cpu(rhead->h_size);
3379 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3380 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3381 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3382 if (h_size % XLOG_HEADER_CYCLE_SIZE)
3383 hblks++;
3384 xlog_put_bp(hbp);
3385 hbp = xlog_get_bp(log, hblks);
3386 } else {
3387 hblks = 1;
3388 }
3389 } else {
3390 ASSERT(log->l_sectBBsize == 1);
3391 hblks = 1;
3392 hbp = xlog_get_bp(log, 1);
3393 h_size = XLOG_BIG_RECORD_BSIZE;
3394 }
3395
3396 if (!hbp)
3397 return ENOMEM;
3398 dbp = xlog_get_bp(log, BTOBB(h_size));
3399 if (!dbp) {
3400 xlog_put_bp(hbp);
3401 return ENOMEM;
3402 }
3403
3404 memset(rhash, 0, sizeof(rhash));
3405 if (tail_blk <= head_blk) {
3406 for (blk_no = tail_blk; blk_no < head_blk; ) {
3407 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3408 if (error)
3409 goto bread_err2;
3410
3411 rhead = (xlog_rec_header_t *)offset;
3412 error = xlog_valid_rec_header(log, rhead, blk_no);
3413 if (error)
3414 goto bread_err2;
3415
3416 /* blocks in data section */
3417 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3418 error = xlog_bread(log, blk_no + hblks, bblks, dbp,
3419 &offset);
3420 if (error)
3421 goto bread_err2;
3422
3423 xlog_unpack_data(rhead, offset, log);
3424 if ((error = xlog_recover_process_data(log,
3425 rhash, rhead, offset, pass)))
3426 goto bread_err2;
3427 blk_no += bblks + hblks;
3428 }
3429 } else {
3430 /*
3431 * Perform recovery around the end of the physical log.
3432 * When the head is not on the same cycle number as the tail,
3433 * we can't do a sequential recovery as above.
3434 */
3435 blk_no = tail_blk;
3436 while (blk_no < log->l_logBBsize) {
3437 /*
3438 * Check for header wrapping around physical end-of-log
3439 */
3440 offset = hbp->b_addr;
3441 split_hblks = 0;
3442 wrapped_hblks = 0;
3443 if (blk_no + hblks <= log->l_logBBsize) {
3444 /* Read header in one read */
3445 error = xlog_bread(log, blk_no, hblks, hbp,
3446 &offset);
3447 if (error)
3448 goto bread_err2;
3449 } else {
3450 /* This LR is split across physical log end */
3451 if (blk_no != log->l_logBBsize) {
3452 /* some data before physical log end */
3453 ASSERT(blk_no <= INT_MAX);
3454 split_hblks = log->l_logBBsize - (int)blk_no;
3455 ASSERT(split_hblks > 0);
3456 error = xlog_bread(log, blk_no,
3457 split_hblks, hbp,
3458 &offset);
3459 if (error)
3460 goto bread_err2;
3461 }
3462
3463 /*
3464 * Note: this black magic still works with
3465 * large sector sizes (non-512) only because:
3466 * - we increased the buffer size originally
3467 * by 1 sector giving us enough extra space
3468 * for the second read;
3469 * - the log start is guaranteed to be sector
3470 * aligned;
3471 * - we read the log end (LR header start)
3472 * _first_, then the log start (LR header end)
3473 * - order is important.
3474 */
3475 wrapped_hblks = hblks - split_hblks;
3476 error = xlog_bread_offset(log, 0,
3477 wrapped_hblks, hbp,
3478 offset + BBTOB(split_hblks));
3479 if (error)
3480 goto bread_err2;
3481 }
3482 rhead = (xlog_rec_header_t *)offset;
3483 error = xlog_valid_rec_header(log, rhead,
3484 split_hblks ? blk_no : 0);
3485 if (error)
3486 goto bread_err2;
3487
3488 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3489 blk_no += hblks;
3490
3491 /* Read in data for log record */
3492 if (blk_no + bblks <= log->l_logBBsize) {
3493 error = xlog_bread(log, blk_no, bblks, dbp,
3494 &offset);
3495 if (error)
3496 goto bread_err2;
3497 } else {
3498 /* This log record is split across the
3499 * physical end of log */
3500 offset = dbp->b_addr;
3501 split_bblks = 0;
3502 if (blk_no != log->l_logBBsize) {
3503 /* some data is before the physical
3504 * end of log */
3505 ASSERT(!wrapped_hblks);
3506 ASSERT(blk_no <= INT_MAX);
3507 split_bblks =
3508 log->l_logBBsize - (int)blk_no;
3509 ASSERT(split_bblks > 0);
3510 error = xlog_bread(log, blk_no,
3511 split_bblks, dbp,
3512 &offset);
3513 if (error)
3514 goto bread_err2;
3515 }
3516
3517 /*
3518 * Note: this black magic still works with
3519 * large sector sizes (non-512) only because:
3520 * - we increased the buffer size originally
3521 * by 1 sector giving us enough extra space
3522 * for the second read;
3523 * - the log start is guaranteed to be sector
3524 * aligned;
3525 * - we read the log end (LR header start)
3526 * _first_, then the log start (LR header end)
3527 * - order is important.
3528 */
3529 error = xlog_bread_offset(log, 0,
3530 bblks - split_bblks, hbp,
3531 offset + BBTOB(split_bblks));
3532 if (error)
3533 goto bread_err2;
3534 }
3535 xlog_unpack_data(rhead, offset, log);
3536 if ((error = xlog_recover_process_data(log, rhash,
3537 rhead, offset, pass)))
3538 goto bread_err2;
3539 blk_no += bblks;
3540 }
3541
3542 ASSERT(blk_no >= log->l_logBBsize);
3543 blk_no -= log->l_logBBsize;
3544
3545 /* read first part of physical log */
3546 while (blk_no < head_blk) {
3547 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3548 if (error)
3549 goto bread_err2;
3550
3551 rhead = (xlog_rec_header_t *)offset;
3552 error = xlog_valid_rec_header(log, rhead, blk_no);
3553 if (error)
3554 goto bread_err2;
3555
3556 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3557 error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3558 &offset);
3559 if (error)
3560 goto bread_err2;
3561
3562 xlog_unpack_data(rhead, offset, log);
3563 if ((error = xlog_recover_process_data(log, rhash,
3564 rhead, offset, pass)))
3565 goto bread_err2;
3566 blk_no += bblks + hblks;
3567 }
3568 }
3569
3570 bread_err2:
3571 xlog_put_bp(dbp);
3572 bread_err1:
3573 xlog_put_bp(hbp);
3574 return error;
3575}
3576
3577/*
3578 * Do the recovery of the log. We actually do this in two phases.
3579 * The two passes are necessary in order to implement the function
3580 * of cancelling a record written into the log. The first pass
3581 * determines those things which have been cancelled, and the
3582 * second pass replays log items normally except for those which
3583 * have been cancelled. The handling of the replay and cancellations
3584 * takes place in the log item type specific routines.
3585 *
3586 * The table of items which have cancel records in the log is allocated
3587 * and freed at this level, since only here do we know when all of
3588 * the log recovery has been completed.
3589 */
3590STATIC int
3591xlog_do_log_recovery(
3592 xlog_t *log,
3593 xfs_daddr_t head_blk,
3594 xfs_daddr_t tail_blk)
3595{
3596 int error, i;
3597
3598 ASSERT(head_blk != tail_blk);
3599
3600 /*
3601 * First do a pass to find all of the cancelled buf log items.
3602 * Store them in the buf_cancel_table for use in the second pass.
3603 */
3604 log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
3605 sizeof(struct list_head),
3606 KM_SLEEP);
3607 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3608 INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
3609
3610 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3611 XLOG_RECOVER_PASS1);
3612 if (error != 0) {
3613 kmem_free(log->l_buf_cancel_table);
3614 log->l_buf_cancel_table = NULL;
3615 return error;
3616 }
3617 /*
3618 * Then do a second pass to actually recover the items in the log.
3619 * When it is complete free the table of buf cancel items.
3620 */
3621 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3622 XLOG_RECOVER_PASS2);
3623#ifdef DEBUG
3624 if (!error) {
3625 int i;
3626
3627 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3628 ASSERT(list_empty(&log->l_buf_cancel_table[i]));
3629 }
3630#endif /* DEBUG */
3631
3632 kmem_free(log->l_buf_cancel_table);
3633 log->l_buf_cancel_table = NULL;
3634
3635 return error;
3636}
3637
3638/*
3639 * Do the actual recovery
3640 */
3641STATIC int
3642xlog_do_recover(
3643 xlog_t *log,
3644 xfs_daddr_t head_blk,
3645 xfs_daddr_t tail_blk)
3646{
3647 int error;
3648 xfs_buf_t *bp;
3649 xfs_sb_t *sbp;
3650
3651 /*
3652 * First replay the images in the log.
3653 */
3654 error = xlog_do_log_recovery(log, head_blk, tail_blk);
3655 if (error) {
3656 return error;
3657 }
3658
3659 XFS_bflush(log->l_mp->m_ddev_targp);
3660
3661 /*
3662 * If IO errors happened during recovery, bail out.
3663 */
3664 if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3665 return (EIO);
3666 }
3667
3668 /*
3669 * We now update the tail_lsn since much of the recovery has completed
3670 * and there may be space available to use. If there were no extent
3671 * or iunlinks, we can free up the entire log and set the tail_lsn to
3672 * be the last_sync_lsn. This was set in xlog_find_tail to be the
3673 * lsn of the last known good LR on disk. If there are extent frees
3674 * or iunlinks they will have some entries in the AIL; so we look at
3675 * the AIL to determine how to set the tail_lsn.
3676 */
3677 xlog_assign_tail_lsn(log->l_mp);
3678
3679 /*
3680 * Now that we've finished replaying all buffer and inode
3681 * updates, re-read in the superblock.
3682 */
3683 bp = xfs_getsb(log->l_mp, 0);
3684 XFS_BUF_UNDONE(bp);
3685 ASSERT(!(XFS_BUF_ISWRITE(bp)));
3686 ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
3687 XFS_BUF_READ(bp);
3688 XFS_BUF_UNASYNC(bp);
3689 xfsbdstrat(log->l_mp, bp);
3690 error = xfs_buf_iowait(bp);
3691 if (error) {
3692 xfs_ioerror_alert("xlog_do_recover",
3693 log->l_mp, bp, XFS_BUF_ADDR(bp));
3694 ASSERT(0);
3695 xfs_buf_relse(bp);
3696 return error;
3697 }
3698
3699 /* Convert superblock from on-disk format */
3700 sbp = &log->l_mp->m_sb;
3701 xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3702 ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3703 ASSERT(xfs_sb_good_version(sbp));
3704 xfs_buf_relse(bp);
3705
3706 /* We've re-read the superblock so re-initialize per-cpu counters */
3707 xfs_icsb_reinit_counters(log->l_mp);
3708
3709 xlog_recover_check_summary(log);
3710
3711 /* Normal transactions can now occur */
3712 log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3713 return 0;
3714}
3715
3716/*
3717 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3718 *
3719 * Return error or zero.
3720 */
3721int
3722xlog_recover(
3723 xlog_t *log)
3724{
3725 xfs_daddr_t head_blk, tail_blk;
3726 int error;
3727
3728 /* find the tail of the log */
3729 if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3730 return error;
3731
3732 if (tail_blk != head_blk) {
3733 /* There used to be a comment here:
3734 *
3735 * disallow recovery on read-only mounts. note -- mount
3736 * checks for ENOSPC and turns it into an intelligent
3737 * error message.
3738 * ...but this is no longer true. Now, unless you specify
3739 * NORECOVERY (in which case this function would never be
3740 * called), we just go ahead and recover. We do this all
3741 * under the vfs layer, so we can get away with it unless
3742 * the device itself is read-only, in which case we fail.
3743 */
3744 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3745 return error;
3746 }
3747
3748 xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
3749 log->l_mp->m_logname ? log->l_mp->m_logname
3750 : "internal");
3751
3752 error = xlog_do_recover(log, head_blk, tail_blk);
3753 log->l_flags |= XLOG_RECOVERY_NEEDED;
3754 }
3755 return error;
3756}
3757
3758/*
3759 * In the first part of recovery we replay inodes and buffers and build
3760 * up the list of extent free items which need to be processed. Here
3761 * we process the extent free items and clean up the on disk unlinked
3762 * inode lists. This is separated from the first part of recovery so
3763 * that the root and real-time bitmap inodes can be read in from disk in
3764 * between the two stages. This is necessary so that we can free space
3765 * in the real-time portion of the file system.
3766 */
3767int
3768xlog_recover_finish(
3769 xlog_t *log)
3770{
3771 /*
3772 * Now we're ready to do the transactions needed for the
3773 * rest of recovery. Start with completing all the extent
3774 * free intent records and then process the unlinked inode
3775 * lists. At this point, we essentially run in normal mode
3776 * except that we're still performing recovery actions
3777 * rather than accepting new requests.
3778 */
3779 if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3780 int error;
3781 error = xlog_recover_process_efis(log);
3782 if (error) {
3783 xfs_alert(log->l_mp, "Failed to recover EFIs");
3784 return error;
3785 }
3786 /*
3787 * Sync the log to get all the EFIs out of the AIL.
3788 * This isn't absolutely necessary, but it helps in
3789 * case the unlink transactions would have problems
3790 * pushing the EFIs out of the way.
3791 */
3792 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3793
3794 xlog_recover_process_iunlinks(log);
3795
3796 xlog_recover_check_summary(log);
3797
3798 xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
3799 log->l_mp->m_logname ? log->l_mp->m_logname
3800 : "internal");
3801 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3802 } else {
3803 xfs_info(log->l_mp, "Ending clean mount");
3804 }
3805 return 0;
3806}
3807
3808
3809#if defined(DEBUG)
3810/*
3811 * Read all of the agf and agi counters and check that they
3812 * are consistent with the superblock counters.
3813 */
3814void
3815xlog_recover_check_summary(
3816 xlog_t *log)
3817{
3818 xfs_mount_t *mp;
3819 xfs_agf_t *agfp;
3820 xfs_buf_t *agfbp;
3821 xfs_buf_t *agibp;
3822 xfs_agnumber_t agno;
3823 __uint64_t freeblks;
3824 __uint64_t itotal;
3825 __uint64_t ifree;
3826 int error;
3827
3828 mp = log->l_mp;
3829
3830 freeblks = 0LL;
3831 itotal = 0LL;
3832 ifree = 0LL;
3833 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3834 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
3835 if (error) {
3836 xfs_alert(mp, "%s agf read failed agno %d error %d",
3837 __func__, agno, error);
3838 } else {
3839 agfp = XFS_BUF_TO_AGF(agfbp);
3840 freeblks += be32_to_cpu(agfp->agf_freeblks) +
3841 be32_to_cpu(agfp->agf_flcount);
3842 xfs_buf_relse(agfbp);
3843 }
3844
3845 error = xfs_read_agi(mp, NULL, agno, &agibp);
3846 if (error) {
3847 xfs_alert(mp, "%s agi read failed agno %d error %d",
3848 __func__, agno, error);
3849 } else {
3850 struct xfs_agi *agi = XFS_BUF_TO_AGI(agibp);
3851
3852 itotal += be32_to_cpu(agi->agi_count);
3853 ifree += be32_to_cpu(agi->agi_freecount);
3854 xfs_buf_relse(agibp);
3855 }
3856 }
3857}
3858#endif /* DEBUG */
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_sb.h"
14#include "xfs_mount.h"
15#include "xfs_defer.h"
16#include "xfs_inode.h"
17#include "xfs_trans.h"
18#include "xfs_log.h"
19#include "xfs_log_priv.h"
20#include "xfs_log_recover.h"
21#include "xfs_inode_item.h"
22#include "xfs_extfree_item.h"
23#include "xfs_trans_priv.h"
24#include "xfs_alloc.h"
25#include "xfs_ialloc.h"
26#include "xfs_quota.h"
27#include "xfs_trace.h"
28#include "xfs_icache.h"
29#include "xfs_bmap_btree.h"
30#include "xfs_error.h"
31#include "xfs_dir2.h"
32#include "xfs_rmap_item.h"
33#include "xfs_buf_item.h"
34#include "xfs_refcount_item.h"
35#include "xfs_bmap_item.h"
36
37#define BLK_AVG(blk1, blk2) ((blk1+blk2) >> 1)
38
39STATIC int
40xlog_find_zeroed(
41 struct xlog *,
42 xfs_daddr_t *);
43STATIC int
44xlog_clear_stale_blocks(
45 struct xlog *,
46 xfs_lsn_t);
47#if defined(DEBUG)
48STATIC void
49xlog_recover_check_summary(
50 struct xlog *);
51#else
52#define xlog_recover_check_summary(log)
53#endif
54STATIC int
55xlog_do_recovery_pass(
56 struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
57
58/*
59 * This structure is used during recovery to record the buf log items which
60 * have been canceled and should not be replayed.
61 */
62struct xfs_buf_cancel {
63 xfs_daddr_t bc_blkno;
64 uint bc_len;
65 int bc_refcount;
66 struct list_head bc_list;
67};
68
69/*
70 * Sector aligned buffer routines for buffer create/read/write/access
71 */
72
73/*
74 * Verify the log-relative block number and length in basic blocks are valid for
75 * an operation involving the given XFS log buffer. Returns true if the fields
76 * are valid, false otherwise.
77 */
78static inline bool
79xlog_verify_bno(
80 struct xlog *log,
81 xfs_daddr_t blk_no,
82 int bbcount)
83{
84 if (blk_no < 0 || blk_no >= log->l_logBBsize)
85 return false;
86 if (bbcount <= 0 || (blk_no + bbcount) > log->l_logBBsize)
87 return false;
88 return true;
89}
90
91/*
92 * Allocate a buffer to hold log data. The buffer needs to be able to map to
93 * a range of nbblks basic blocks at any valid offset within the log.
94 */
95static char *
96xlog_alloc_buffer(
97 struct xlog *log,
98 int nbblks)
99{
100 int align_mask = xfs_buftarg_dma_alignment(log->l_targ);
101
102 /*
103 * Pass log block 0 since we don't have an addr yet, buffer will be
104 * verified on read.
105 */
106 if (!xlog_verify_bno(log, 0, nbblks)) {
107 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
108 nbblks);
109 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
110 return NULL;
111 }
112
113 /*
114 * We do log I/O in units of log sectors (a power-of-2 multiple of the
115 * basic block size), so we round up the requested size to accommodate
116 * the basic blocks required for complete log sectors.
117 *
118 * In addition, the buffer may be used for a non-sector-aligned block
119 * offset, in which case an I/O of the requested size could extend
120 * beyond the end of the buffer. If the requested size is only 1 basic
121 * block it will never straddle a sector boundary, so this won't be an
122 * issue. Nor will this be a problem if the log I/O is done in basic
123 * blocks (sector size 1). But otherwise we extend the buffer by one
124 * extra log sector to ensure there's space to accommodate this
125 * possibility.
126 */
127 if (nbblks > 1 && log->l_sectBBsize > 1)
128 nbblks += log->l_sectBBsize;
129 nbblks = round_up(nbblks, log->l_sectBBsize);
130 return kmem_alloc_io(BBTOB(nbblks), align_mask, KM_MAYFAIL | KM_ZERO);
131}
132
133/*
134 * Return the address of the start of the given block number's data
135 * in a log buffer. The buffer covers a log sector-aligned region.
136 */
137static inline unsigned int
138xlog_align(
139 struct xlog *log,
140 xfs_daddr_t blk_no)
141{
142 return BBTOB(blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1));
143}
144
145static int
146xlog_do_io(
147 struct xlog *log,
148 xfs_daddr_t blk_no,
149 unsigned int nbblks,
150 char *data,
151 unsigned int op)
152{
153 int error;
154
155 if (!xlog_verify_bno(log, blk_no, nbblks)) {
156 xfs_warn(log->l_mp,
157 "Invalid log block/length (0x%llx, 0x%x) for buffer",
158 blk_no, nbblks);
159 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
160 return -EFSCORRUPTED;
161 }
162
163 blk_no = round_down(blk_no, log->l_sectBBsize);
164 nbblks = round_up(nbblks, log->l_sectBBsize);
165 ASSERT(nbblks > 0);
166
167 error = xfs_rw_bdev(log->l_targ->bt_bdev, log->l_logBBstart + blk_no,
168 BBTOB(nbblks), data, op);
169 if (error && !XFS_FORCED_SHUTDOWN(log->l_mp)) {
170 xfs_alert(log->l_mp,
171 "log recovery %s I/O error at daddr 0x%llx len %d error %d",
172 op == REQ_OP_WRITE ? "write" : "read",
173 blk_no, nbblks, error);
174 }
175 return error;
176}
177
178STATIC int
179xlog_bread_noalign(
180 struct xlog *log,
181 xfs_daddr_t blk_no,
182 int nbblks,
183 char *data)
184{
185 return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
186}
187
188STATIC int
189xlog_bread(
190 struct xlog *log,
191 xfs_daddr_t blk_no,
192 int nbblks,
193 char *data,
194 char **offset)
195{
196 int error;
197
198 error = xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
199 if (!error)
200 *offset = data + xlog_align(log, blk_no);
201 return error;
202}
203
204STATIC int
205xlog_bwrite(
206 struct xlog *log,
207 xfs_daddr_t blk_no,
208 int nbblks,
209 char *data)
210{
211 return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_WRITE);
212}
213
214#ifdef DEBUG
215/*
216 * dump debug superblock and log record information
217 */
218STATIC void
219xlog_header_check_dump(
220 xfs_mount_t *mp,
221 xlog_rec_header_t *head)
222{
223 xfs_debug(mp, "%s: SB : uuid = %pU, fmt = %d",
224 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
225 xfs_debug(mp, " log : uuid = %pU, fmt = %d",
226 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
227}
228#else
229#define xlog_header_check_dump(mp, head)
230#endif
231
232/*
233 * check log record header for recovery
234 */
235STATIC int
236xlog_header_check_recover(
237 xfs_mount_t *mp,
238 xlog_rec_header_t *head)
239{
240 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
241
242 /*
243 * IRIX doesn't write the h_fmt field and leaves it zeroed
244 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
245 * a dirty log created in IRIX.
246 */
247 if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
248 xfs_warn(mp,
249 "dirty log written in incompatible format - can't recover");
250 xlog_header_check_dump(mp, head);
251 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
252 XFS_ERRLEVEL_HIGH, mp);
253 return -EFSCORRUPTED;
254 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
255 xfs_warn(mp,
256 "dirty log entry has mismatched uuid - can't recover");
257 xlog_header_check_dump(mp, head);
258 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
259 XFS_ERRLEVEL_HIGH, mp);
260 return -EFSCORRUPTED;
261 }
262 return 0;
263}
264
265/*
266 * read the head block of the log and check the header
267 */
268STATIC int
269xlog_header_check_mount(
270 xfs_mount_t *mp,
271 xlog_rec_header_t *head)
272{
273 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
274
275 if (uuid_is_null(&head->h_fs_uuid)) {
276 /*
277 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
278 * h_fs_uuid is null, we assume this log was last mounted
279 * by IRIX and continue.
280 */
281 xfs_warn(mp, "null uuid in log - IRIX style log");
282 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
283 xfs_warn(mp, "log has mismatched uuid - can't recover");
284 xlog_header_check_dump(mp, head);
285 XFS_ERROR_REPORT("xlog_header_check_mount",
286 XFS_ERRLEVEL_HIGH, mp);
287 return -EFSCORRUPTED;
288 }
289 return 0;
290}
291
292STATIC void
293xlog_recover_iodone(
294 struct xfs_buf *bp)
295{
296 if (bp->b_error) {
297 /*
298 * We're not going to bother about retrying
299 * this during recovery. One strike!
300 */
301 if (!XFS_FORCED_SHUTDOWN(bp->b_mount)) {
302 xfs_buf_ioerror_alert(bp, __func__);
303 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
304 }
305 }
306
307 /*
308 * On v5 supers, a bli could be attached to update the metadata LSN.
309 * Clean it up.
310 */
311 if (bp->b_log_item)
312 xfs_buf_item_relse(bp);
313 ASSERT(bp->b_log_item == NULL);
314
315 bp->b_iodone = NULL;
316 xfs_buf_ioend(bp);
317}
318
319/*
320 * This routine finds (to an approximation) the first block in the physical
321 * log which contains the given cycle. It uses a binary search algorithm.
322 * Note that the algorithm can not be perfect because the disk will not
323 * necessarily be perfect.
324 */
325STATIC int
326xlog_find_cycle_start(
327 struct xlog *log,
328 char *buffer,
329 xfs_daddr_t first_blk,
330 xfs_daddr_t *last_blk,
331 uint cycle)
332{
333 char *offset;
334 xfs_daddr_t mid_blk;
335 xfs_daddr_t end_blk;
336 uint mid_cycle;
337 int error;
338
339 end_blk = *last_blk;
340 mid_blk = BLK_AVG(first_blk, end_blk);
341 while (mid_blk != first_blk && mid_blk != end_blk) {
342 error = xlog_bread(log, mid_blk, 1, buffer, &offset);
343 if (error)
344 return error;
345 mid_cycle = xlog_get_cycle(offset);
346 if (mid_cycle == cycle)
347 end_blk = mid_blk; /* last_half_cycle == mid_cycle */
348 else
349 first_blk = mid_blk; /* first_half_cycle == mid_cycle */
350 mid_blk = BLK_AVG(first_blk, end_blk);
351 }
352 ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
353 (mid_blk == end_blk && mid_blk-1 == first_blk));
354
355 *last_blk = end_blk;
356
357 return 0;
358}
359
360/*
361 * Check that a range of blocks does not contain stop_on_cycle_no.
362 * Fill in *new_blk with the block offset where such a block is
363 * found, or with -1 (an invalid block number) if there is no such
364 * block in the range. The scan needs to occur from front to back
365 * and the pointer into the region must be updated since a later
366 * routine will need to perform another test.
367 */
368STATIC int
369xlog_find_verify_cycle(
370 struct xlog *log,
371 xfs_daddr_t start_blk,
372 int nbblks,
373 uint stop_on_cycle_no,
374 xfs_daddr_t *new_blk)
375{
376 xfs_daddr_t i, j;
377 uint cycle;
378 char *buffer;
379 xfs_daddr_t bufblks;
380 char *buf = NULL;
381 int error = 0;
382
383 /*
384 * Greedily allocate a buffer big enough to handle the full
385 * range of basic blocks we'll be examining. If that fails,
386 * try a smaller size. We need to be able to read at least
387 * a log sector, or we're out of luck.
388 */
389 bufblks = 1 << ffs(nbblks);
390 while (bufblks > log->l_logBBsize)
391 bufblks >>= 1;
392 while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
393 bufblks >>= 1;
394 if (bufblks < log->l_sectBBsize)
395 return -ENOMEM;
396 }
397
398 for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
399 int bcount;
400
401 bcount = min(bufblks, (start_blk + nbblks - i));
402
403 error = xlog_bread(log, i, bcount, buffer, &buf);
404 if (error)
405 goto out;
406
407 for (j = 0; j < bcount; j++) {
408 cycle = xlog_get_cycle(buf);
409 if (cycle == stop_on_cycle_no) {
410 *new_blk = i+j;
411 goto out;
412 }
413
414 buf += BBSIZE;
415 }
416 }
417
418 *new_blk = -1;
419
420out:
421 kmem_free(buffer);
422 return error;
423}
424
425/*
426 * Potentially backup over partial log record write.
427 *
428 * In the typical case, last_blk is the number of the block directly after
429 * a good log record. Therefore, we subtract one to get the block number
430 * of the last block in the given buffer. extra_bblks contains the number
431 * of blocks we would have read on a previous read. This happens when the
432 * last log record is split over the end of the physical log.
433 *
434 * extra_bblks is the number of blocks potentially verified on a previous
435 * call to this routine.
436 */
437STATIC int
438xlog_find_verify_log_record(
439 struct xlog *log,
440 xfs_daddr_t start_blk,
441 xfs_daddr_t *last_blk,
442 int extra_bblks)
443{
444 xfs_daddr_t i;
445 char *buffer;
446 char *offset = NULL;
447 xlog_rec_header_t *head = NULL;
448 int error = 0;
449 int smallmem = 0;
450 int num_blks = *last_blk - start_blk;
451 int xhdrs;
452
453 ASSERT(start_blk != 0 || *last_blk != start_blk);
454
455 buffer = xlog_alloc_buffer(log, num_blks);
456 if (!buffer) {
457 buffer = xlog_alloc_buffer(log, 1);
458 if (!buffer)
459 return -ENOMEM;
460 smallmem = 1;
461 } else {
462 error = xlog_bread(log, start_blk, num_blks, buffer, &offset);
463 if (error)
464 goto out;
465 offset += ((num_blks - 1) << BBSHIFT);
466 }
467
468 for (i = (*last_blk) - 1; i >= 0; i--) {
469 if (i < start_blk) {
470 /* valid log record not found */
471 xfs_warn(log->l_mp,
472 "Log inconsistent (didn't find previous header)");
473 ASSERT(0);
474 error = -EIO;
475 goto out;
476 }
477
478 if (smallmem) {
479 error = xlog_bread(log, i, 1, buffer, &offset);
480 if (error)
481 goto out;
482 }
483
484 head = (xlog_rec_header_t *)offset;
485
486 if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
487 break;
488
489 if (!smallmem)
490 offset -= BBSIZE;
491 }
492
493 /*
494 * We hit the beginning of the physical log & still no header. Return
495 * to caller. If caller can handle a return of -1, then this routine
496 * will be called again for the end of the physical log.
497 */
498 if (i == -1) {
499 error = 1;
500 goto out;
501 }
502
503 /*
504 * We have the final block of the good log (the first block
505 * of the log record _before_ the head. So we check the uuid.
506 */
507 if ((error = xlog_header_check_mount(log->l_mp, head)))
508 goto out;
509
510 /*
511 * We may have found a log record header before we expected one.
512 * last_blk will be the 1st block # with a given cycle #. We may end
513 * up reading an entire log record. In this case, we don't want to
514 * reset last_blk. Only when last_blk points in the middle of a log
515 * record do we update last_blk.
516 */
517 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
518 uint h_size = be32_to_cpu(head->h_size);
519
520 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
521 if (h_size % XLOG_HEADER_CYCLE_SIZE)
522 xhdrs++;
523 } else {
524 xhdrs = 1;
525 }
526
527 if (*last_blk - i + extra_bblks !=
528 BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
529 *last_blk = i;
530
531out:
532 kmem_free(buffer);
533 return error;
534}
535
536/*
537 * Head is defined to be the point of the log where the next log write
538 * could go. This means that incomplete LR writes at the end are
539 * eliminated when calculating the head. We aren't guaranteed that previous
540 * LR have complete transactions. We only know that a cycle number of
541 * current cycle number -1 won't be present in the log if we start writing
542 * from our current block number.
543 *
544 * last_blk contains the block number of the first block with a given
545 * cycle number.
546 *
547 * Return: zero if normal, non-zero if error.
548 */
549STATIC int
550xlog_find_head(
551 struct xlog *log,
552 xfs_daddr_t *return_head_blk)
553{
554 char *buffer;
555 char *offset;
556 xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
557 int num_scan_bblks;
558 uint first_half_cycle, last_half_cycle;
559 uint stop_on_cycle;
560 int error, log_bbnum = log->l_logBBsize;
561
562 /* Is the end of the log device zeroed? */
563 error = xlog_find_zeroed(log, &first_blk);
564 if (error < 0) {
565 xfs_warn(log->l_mp, "empty log check failed");
566 return error;
567 }
568 if (error == 1) {
569 *return_head_blk = first_blk;
570
571 /* Is the whole lot zeroed? */
572 if (!first_blk) {
573 /* Linux XFS shouldn't generate totally zeroed logs -
574 * mkfs etc write a dummy unmount record to a fresh
575 * log so we can store the uuid in there
576 */
577 xfs_warn(log->l_mp, "totally zeroed log");
578 }
579
580 return 0;
581 }
582
583 first_blk = 0; /* get cycle # of 1st block */
584 buffer = xlog_alloc_buffer(log, 1);
585 if (!buffer)
586 return -ENOMEM;
587
588 error = xlog_bread(log, 0, 1, buffer, &offset);
589 if (error)
590 goto out_free_buffer;
591
592 first_half_cycle = xlog_get_cycle(offset);
593
594 last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
595 error = xlog_bread(log, last_blk, 1, buffer, &offset);
596 if (error)
597 goto out_free_buffer;
598
599 last_half_cycle = xlog_get_cycle(offset);
600 ASSERT(last_half_cycle != 0);
601
602 /*
603 * If the 1st half cycle number is equal to the last half cycle number,
604 * then the entire log is stamped with the same cycle number. In this
605 * case, head_blk can't be set to zero (which makes sense). The below
606 * math doesn't work out properly with head_blk equal to zero. Instead,
607 * we set it to log_bbnum which is an invalid block number, but this
608 * value makes the math correct. If head_blk doesn't changed through
609 * all the tests below, *head_blk is set to zero at the very end rather
610 * than log_bbnum. In a sense, log_bbnum and zero are the same block
611 * in a circular file.
612 */
613 if (first_half_cycle == last_half_cycle) {
614 /*
615 * In this case we believe that the entire log should have
616 * cycle number last_half_cycle. We need to scan backwards
617 * from the end verifying that there are no holes still
618 * containing last_half_cycle - 1. If we find such a hole,
619 * then the start of that hole will be the new head. The
620 * simple case looks like
621 * x | x ... | x - 1 | x
622 * Another case that fits this picture would be
623 * x | x + 1 | x ... | x
624 * In this case the head really is somewhere at the end of the
625 * log, as one of the latest writes at the beginning was
626 * incomplete.
627 * One more case is
628 * x | x + 1 | x ... | x - 1 | x
629 * This is really the combination of the above two cases, and
630 * the head has to end up at the start of the x-1 hole at the
631 * end of the log.
632 *
633 * In the 256k log case, we will read from the beginning to the
634 * end of the log and search for cycle numbers equal to x-1.
635 * We don't worry about the x+1 blocks that we encounter,
636 * because we know that they cannot be the head since the log
637 * started with x.
638 */
639 head_blk = log_bbnum;
640 stop_on_cycle = last_half_cycle - 1;
641 } else {
642 /*
643 * In this case we want to find the first block with cycle
644 * number matching last_half_cycle. We expect the log to be
645 * some variation on
646 * x + 1 ... | x ... | x
647 * The first block with cycle number x (last_half_cycle) will
648 * be where the new head belongs. First we do a binary search
649 * for the first occurrence of last_half_cycle. The binary
650 * search may not be totally accurate, so then we scan back
651 * from there looking for occurrences of last_half_cycle before
652 * us. If that backwards scan wraps around the beginning of
653 * the log, then we look for occurrences of last_half_cycle - 1
654 * at the end of the log. The cases we're looking for look
655 * like
656 * v binary search stopped here
657 * x + 1 ... | x | x + 1 | x ... | x
658 * ^ but we want to locate this spot
659 * or
660 * <---------> less than scan distance
661 * x + 1 ... | x ... | x - 1 | x
662 * ^ we want to locate this spot
663 */
664 stop_on_cycle = last_half_cycle;
665 error = xlog_find_cycle_start(log, buffer, first_blk, &head_blk,
666 last_half_cycle);
667 if (error)
668 goto out_free_buffer;
669 }
670
671 /*
672 * Now validate the answer. Scan back some number of maximum possible
673 * blocks and make sure each one has the expected cycle number. The
674 * maximum is determined by the total possible amount of buffering
675 * in the in-core log. The following number can be made tighter if
676 * we actually look at the block size of the filesystem.
677 */
678 num_scan_bblks = min_t(int, log_bbnum, XLOG_TOTAL_REC_SHIFT(log));
679 if (head_blk >= num_scan_bblks) {
680 /*
681 * We are guaranteed that the entire check can be performed
682 * in one buffer.
683 */
684 start_blk = head_blk - num_scan_bblks;
685 if ((error = xlog_find_verify_cycle(log,
686 start_blk, num_scan_bblks,
687 stop_on_cycle, &new_blk)))
688 goto out_free_buffer;
689 if (new_blk != -1)
690 head_blk = new_blk;
691 } else { /* need to read 2 parts of log */
692 /*
693 * We are going to scan backwards in the log in two parts.
694 * First we scan the physical end of the log. In this part
695 * of the log, we are looking for blocks with cycle number
696 * last_half_cycle - 1.
697 * If we find one, then we know that the log starts there, as
698 * we've found a hole that didn't get written in going around
699 * the end of the physical log. The simple case for this is
700 * x + 1 ... | x ... | x - 1 | x
701 * <---------> less than scan distance
702 * If all of the blocks at the end of the log have cycle number
703 * last_half_cycle, then we check the blocks at the start of
704 * the log looking for occurrences of last_half_cycle. If we
705 * find one, then our current estimate for the location of the
706 * first occurrence of last_half_cycle is wrong and we move
707 * back to the hole we've found. This case looks like
708 * x + 1 ... | x | x + 1 | x ...
709 * ^ binary search stopped here
710 * Another case we need to handle that only occurs in 256k
711 * logs is
712 * x + 1 ... | x ... | x+1 | x ...
713 * ^ binary search stops here
714 * In a 256k log, the scan at the end of the log will see the
715 * x + 1 blocks. We need to skip past those since that is
716 * certainly not the head of the log. By searching for
717 * last_half_cycle-1 we accomplish that.
718 */
719 ASSERT(head_blk <= INT_MAX &&
720 (xfs_daddr_t) num_scan_bblks >= head_blk);
721 start_blk = log_bbnum - (num_scan_bblks - head_blk);
722 if ((error = xlog_find_verify_cycle(log, start_blk,
723 num_scan_bblks - (int)head_blk,
724 (stop_on_cycle - 1), &new_blk)))
725 goto out_free_buffer;
726 if (new_blk != -1) {
727 head_blk = new_blk;
728 goto validate_head;
729 }
730
731 /*
732 * Scan beginning of log now. The last part of the physical
733 * log is good. This scan needs to verify that it doesn't find
734 * the last_half_cycle.
735 */
736 start_blk = 0;
737 ASSERT(head_blk <= INT_MAX);
738 if ((error = xlog_find_verify_cycle(log,
739 start_blk, (int)head_blk,
740 stop_on_cycle, &new_blk)))
741 goto out_free_buffer;
742 if (new_blk != -1)
743 head_blk = new_blk;
744 }
745
746validate_head:
747 /*
748 * Now we need to make sure head_blk is not pointing to a block in
749 * the middle of a log record.
750 */
751 num_scan_bblks = XLOG_REC_SHIFT(log);
752 if (head_blk >= num_scan_bblks) {
753 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
754
755 /* start ptr at last block ptr before head_blk */
756 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
757 if (error == 1)
758 error = -EIO;
759 if (error)
760 goto out_free_buffer;
761 } else {
762 start_blk = 0;
763 ASSERT(head_blk <= INT_MAX);
764 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
765 if (error < 0)
766 goto out_free_buffer;
767 if (error == 1) {
768 /* We hit the beginning of the log during our search */
769 start_blk = log_bbnum - (num_scan_bblks - head_blk);
770 new_blk = log_bbnum;
771 ASSERT(start_blk <= INT_MAX &&
772 (xfs_daddr_t) log_bbnum-start_blk >= 0);
773 ASSERT(head_blk <= INT_MAX);
774 error = xlog_find_verify_log_record(log, start_blk,
775 &new_blk, (int)head_blk);
776 if (error == 1)
777 error = -EIO;
778 if (error)
779 goto out_free_buffer;
780 if (new_blk != log_bbnum)
781 head_blk = new_blk;
782 } else if (error)
783 goto out_free_buffer;
784 }
785
786 kmem_free(buffer);
787 if (head_blk == log_bbnum)
788 *return_head_blk = 0;
789 else
790 *return_head_blk = head_blk;
791 /*
792 * When returning here, we have a good block number. Bad block
793 * means that during a previous crash, we didn't have a clean break
794 * from cycle number N to cycle number N-1. In this case, we need
795 * to find the first block with cycle number N-1.
796 */
797 return 0;
798
799out_free_buffer:
800 kmem_free(buffer);
801 if (error)
802 xfs_warn(log->l_mp, "failed to find log head");
803 return error;
804}
805
806/*
807 * Seek backwards in the log for log record headers.
808 *
809 * Given a starting log block, walk backwards until we find the provided number
810 * of records or hit the provided tail block. The return value is the number of
811 * records encountered or a negative error code. The log block and buffer
812 * pointer of the last record seen are returned in rblk and rhead respectively.
813 */
814STATIC int
815xlog_rseek_logrec_hdr(
816 struct xlog *log,
817 xfs_daddr_t head_blk,
818 xfs_daddr_t tail_blk,
819 int count,
820 char *buffer,
821 xfs_daddr_t *rblk,
822 struct xlog_rec_header **rhead,
823 bool *wrapped)
824{
825 int i;
826 int error;
827 int found = 0;
828 char *offset = NULL;
829 xfs_daddr_t end_blk;
830
831 *wrapped = false;
832
833 /*
834 * Walk backwards from the head block until we hit the tail or the first
835 * block in the log.
836 */
837 end_blk = head_blk > tail_blk ? tail_blk : 0;
838 for (i = (int) head_blk - 1; i >= end_blk; i--) {
839 error = xlog_bread(log, i, 1, buffer, &offset);
840 if (error)
841 goto out_error;
842
843 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
844 *rblk = i;
845 *rhead = (struct xlog_rec_header *) offset;
846 if (++found == count)
847 break;
848 }
849 }
850
851 /*
852 * If we haven't hit the tail block or the log record header count,
853 * start looking again from the end of the physical log. Note that
854 * callers can pass head == tail if the tail is not yet known.
855 */
856 if (tail_blk >= head_blk && found != count) {
857 for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
858 error = xlog_bread(log, i, 1, buffer, &offset);
859 if (error)
860 goto out_error;
861
862 if (*(__be32 *)offset ==
863 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
864 *wrapped = true;
865 *rblk = i;
866 *rhead = (struct xlog_rec_header *) offset;
867 if (++found == count)
868 break;
869 }
870 }
871 }
872
873 return found;
874
875out_error:
876 return error;
877}
878
879/*
880 * Seek forward in the log for log record headers.
881 *
882 * Given head and tail blocks, walk forward from the tail block until we find
883 * the provided number of records or hit the head block. The return value is the
884 * number of records encountered or a negative error code. The log block and
885 * buffer pointer of the last record seen are returned in rblk and rhead
886 * respectively.
887 */
888STATIC int
889xlog_seek_logrec_hdr(
890 struct xlog *log,
891 xfs_daddr_t head_blk,
892 xfs_daddr_t tail_blk,
893 int count,
894 char *buffer,
895 xfs_daddr_t *rblk,
896 struct xlog_rec_header **rhead,
897 bool *wrapped)
898{
899 int i;
900 int error;
901 int found = 0;
902 char *offset = NULL;
903 xfs_daddr_t end_blk;
904
905 *wrapped = false;
906
907 /*
908 * Walk forward from the tail block until we hit the head or the last
909 * block in the log.
910 */
911 end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
912 for (i = (int) tail_blk; i <= end_blk; i++) {
913 error = xlog_bread(log, i, 1, buffer, &offset);
914 if (error)
915 goto out_error;
916
917 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
918 *rblk = i;
919 *rhead = (struct xlog_rec_header *) offset;
920 if (++found == count)
921 break;
922 }
923 }
924
925 /*
926 * If we haven't hit the head block or the log record header count,
927 * start looking again from the start of the physical log.
928 */
929 if (tail_blk > head_blk && found != count) {
930 for (i = 0; i < (int) head_blk; i++) {
931 error = xlog_bread(log, i, 1, buffer, &offset);
932 if (error)
933 goto out_error;
934
935 if (*(__be32 *)offset ==
936 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
937 *wrapped = true;
938 *rblk = i;
939 *rhead = (struct xlog_rec_header *) offset;
940 if (++found == count)
941 break;
942 }
943 }
944 }
945
946 return found;
947
948out_error:
949 return error;
950}
951
952/*
953 * Calculate distance from head to tail (i.e., unused space in the log).
954 */
955static inline int
956xlog_tail_distance(
957 struct xlog *log,
958 xfs_daddr_t head_blk,
959 xfs_daddr_t tail_blk)
960{
961 if (head_blk < tail_blk)
962 return tail_blk - head_blk;
963
964 return tail_blk + (log->l_logBBsize - head_blk);
965}
966
967/*
968 * Verify the log tail. This is particularly important when torn or incomplete
969 * writes have been detected near the front of the log and the head has been
970 * walked back accordingly.
971 *
972 * We also have to handle the case where the tail was pinned and the head
973 * blocked behind the tail right before a crash. If the tail had been pushed
974 * immediately prior to the crash and the subsequent checkpoint was only
975 * partially written, it's possible it overwrote the last referenced tail in the
976 * log with garbage. This is not a coherency problem because the tail must have
977 * been pushed before it can be overwritten, but appears as log corruption to
978 * recovery because we have no way to know the tail was updated if the
979 * subsequent checkpoint didn't write successfully.
980 *
981 * Therefore, CRC check the log from tail to head. If a failure occurs and the
982 * offending record is within max iclog bufs from the head, walk the tail
983 * forward and retry until a valid tail is found or corruption is detected out
984 * of the range of a possible overwrite.
985 */
986STATIC int
987xlog_verify_tail(
988 struct xlog *log,
989 xfs_daddr_t head_blk,
990 xfs_daddr_t *tail_blk,
991 int hsize)
992{
993 struct xlog_rec_header *thead;
994 char *buffer;
995 xfs_daddr_t first_bad;
996 int error = 0;
997 bool wrapped;
998 xfs_daddr_t tmp_tail;
999 xfs_daddr_t orig_tail = *tail_blk;
1000
1001 buffer = xlog_alloc_buffer(log, 1);
1002 if (!buffer)
1003 return -ENOMEM;
1004
1005 /*
1006 * Make sure the tail points to a record (returns positive count on
1007 * success).
1008 */
1009 error = xlog_seek_logrec_hdr(log, head_blk, *tail_blk, 1, buffer,
1010 &tmp_tail, &thead, &wrapped);
1011 if (error < 0)
1012 goto out;
1013 if (*tail_blk != tmp_tail)
1014 *tail_blk = tmp_tail;
1015
1016 /*
1017 * Run a CRC check from the tail to the head. We can't just check
1018 * MAX_ICLOGS records past the tail because the tail may point to stale
1019 * blocks cleared during the search for the head/tail. These blocks are
1020 * overwritten with zero-length records and thus record count is not a
1021 * reliable indicator of the iclog state before a crash.
1022 */
1023 first_bad = 0;
1024 error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1025 XLOG_RECOVER_CRCPASS, &first_bad);
1026 while ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1027 int tail_distance;
1028
1029 /*
1030 * Is corruption within range of the head? If so, retry from
1031 * the next record. Otherwise return an error.
1032 */
1033 tail_distance = xlog_tail_distance(log, head_blk, first_bad);
1034 if (tail_distance > BTOBB(XLOG_MAX_ICLOGS * hsize))
1035 break;
1036
1037 /* skip to the next record; returns positive count on success */
1038 error = xlog_seek_logrec_hdr(log, head_blk, first_bad, 2,
1039 buffer, &tmp_tail, &thead, &wrapped);
1040 if (error < 0)
1041 goto out;
1042
1043 *tail_blk = tmp_tail;
1044 first_bad = 0;
1045 error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1046 XLOG_RECOVER_CRCPASS, &first_bad);
1047 }
1048
1049 if (!error && *tail_blk != orig_tail)
1050 xfs_warn(log->l_mp,
1051 "Tail block (0x%llx) overwrite detected. Updated to 0x%llx",
1052 orig_tail, *tail_blk);
1053out:
1054 kmem_free(buffer);
1055 return error;
1056}
1057
1058/*
1059 * Detect and trim torn writes from the head of the log.
1060 *
1061 * Storage without sector atomicity guarantees can result in torn writes in the
1062 * log in the event of a crash. Our only means to detect this scenario is via
1063 * CRC verification. While we can't always be certain that CRC verification
1064 * failure is due to a torn write vs. an unrelated corruption, we do know that
1065 * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1066 * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1067 * the log and treat failures in this range as torn writes as a matter of
1068 * policy. In the event of CRC failure, the head is walked back to the last good
1069 * record in the log and the tail is updated from that record and verified.
1070 */
1071STATIC int
1072xlog_verify_head(
1073 struct xlog *log,
1074 xfs_daddr_t *head_blk, /* in/out: unverified head */
1075 xfs_daddr_t *tail_blk, /* out: tail block */
1076 char *buffer,
1077 xfs_daddr_t *rhead_blk, /* start blk of last record */
1078 struct xlog_rec_header **rhead, /* ptr to last record */
1079 bool *wrapped) /* last rec. wraps phys. log */
1080{
1081 struct xlog_rec_header *tmp_rhead;
1082 char *tmp_buffer;
1083 xfs_daddr_t first_bad;
1084 xfs_daddr_t tmp_rhead_blk;
1085 int found;
1086 int error;
1087 bool tmp_wrapped;
1088
1089 /*
1090 * Check the head of the log for torn writes. Search backwards from the
1091 * head until we hit the tail or the maximum number of log record I/Os
1092 * that could have been in flight at one time. Use a temporary buffer so
1093 * we don't trash the rhead/buffer pointers from the caller.
1094 */
1095 tmp_buffer = xlog_alloc_buffer(log, 1);
1096 if (!tmp_buffer)
1097 return -ENOMEM;
1098 error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1099 XLOG_MAX_ICLOGS, tmp_buffer,
1100 &tmp_rhead_blk, &tmp_rhead, &tmp_wrapped);
1101 kmem_free(tmp_buffer);
1102 if (error < 0)
1103 return error;
1104
1105 /*
1106 * Now run a CRC verification pass over the records starting at the
1107 * block found above to the current head. If a CRC failure occurs, the
1108 * log block of the first bad record is saved in first_bad.
1109 */
1110 error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1111 XLOG_RECOVER_CRCPASS, &first_bad);
1112 if ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1113 /*
1114 * We've hit a potential torn write. Reset the error and warn
1115 * about it.
1116 */
1117 error = 0;
1118 xfs_warn(log->l_mp,
1119"Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1120 first_bad, *head_blk);
1121
1122 /*
1123 * Get the header block and buffer pointer for the last good
1124 * record before the bad record.
1125 *
1126 * Note that xlog_find_tail() clears the blocks at the new head
1127 * (i.e., the records with invalid CRC) if the cycle number
1128 * matches the the current cycle.
1129 */
1130 found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1,
1131 buffer, rhead_blk, rhead, wrapped);
1132 if (found < 0)
1133 return found;
1134 if (found == 0) /* XXX: right thing to do here? */
1135 return -EIO;
1136
1137 /*
1138 * Reset the head block to the starting block of the first bad
1139 * log record and set the tail block based on the last good
1140 * record.
1141 *
1142 * Bail out if the updated head/tail match as this indicates
1143 * possible corruption outside of the acceptable
1144 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1145 */
1146 *head_blk = first_bad;
1147 *tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1148 if (*head_blk == *tail_blk) {
1149 ASSERT(0);
1150 return 0;
1151 }
1152 }
1153 if (error)
1154 return error;
1155
1156 return xlog_verify_tail(log, *head_blk, tail_blk,
1157 be32_to_cpu((*rhead)->h_size));
1158}
1159
1160/*
1161 * We need to make sure we handle log wrapping properly, so we can't use the
1162 * calculated logbno directly. Make sure it wraps to the correct bno inside the
1163 * log.
1164 *
1165 * The log is limited to 32 bit sizes, so we use the appropriate modulus
1166 * operation here and cast it back to a 64 bit daddr on return.
1167 */
1168static inline xfs_daddr_t
1169xlog_wrap_logbno(
1170 struct xlog *log,
1171 xfs_daddr_t bno)
1172{
1173 int mod;
1174
1175 div_s64_rem(bno, log->l_logBBsize, &mod);
1176 return mod;
1177}
1178
1179/*
1180 * Check whether the head of the log points to an unmount record. In other
1181 * words, determine whether the log is clean. If so, update the in-core state
1182 * appropriately.
1183 */
1184static int
1185xlog_check_unmount_rec(
1186 struct xlog *log,
1187 xfs_daddr_t *head_blk,
1188 xfs_daddr_t *tail_blk,
1189 struct xlog_rec_header *rhead,
1190 xfs_daddr_t rhead_blk,
1191 char *buffer,
1192 bool *clean)
1193{
1194 struct xlog_op_header *op_head;
1195 xfs_daddr_t umount_data_blk;
1196 xfs_daddr_t after_umount_blk;
1197 int hblks;
1198 int error;
1199 char *offset;
1200
1201 *clean = false;
1202
1203 /*
1204 * Look for unmount record. If we find it, then we know there was a
1205 * clean unmount. Since 'i' could be the last block in the physical
1206 * log, we convert to a log block before comparing to the head_blk.
1207 *
1208 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1209 * below. We won't want to clear the unmount record if there is one, so
1210 * we pass the lsn of the unmount record rather than the block after it.
1211 */
1212 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1213 int h_size = be32_to_cpu(rhead->h_size);
1214 int h_version = be32_to_cpu(rhead->h_version);
1215
1216 if ((h_version & XLOG_VERSION_2) &&
1217 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1218 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1219 if (h_size % XLOG_HEADER_CYCLE_SIZE)
1220 hblks++;
1221 } else {
1222 hblks = 1;
1223 }
1224 } else {
1225 hblks = 1;
1226 }
1227
1228 after_umount_blk = xlog_wrap_logbno(log,
1229 rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len)));
1230
1231 if (*head_blk == after_umount_blk &&
1232 be32_to_cpu(rhead->h_num_logops) == 1) {
1233 umount_data_blk = xlog_wrap_logbno(log, rhead_blk + hblks);
1234 error = xlog_bread(log, umount_data_blk, 1, buffer, &offset);
1235 if (error)
1236 return error;
1237
1238 op_head = (struct xlog_op_header *)offset;
1239 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1240 /*
1241 * Set tail and last sync so that newly written log
1242 * records will point recovery to after the current
1243 * unmount record.
1244 */
1245 xlog_assign_atomic_lsn(&log->l_tail_lsn,
1246 log->l_curr_cycle, after_umount_blk);
1247 xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1248 log->l_curr_cycle, after_umount_blk);
1249 *tail_blk = after_umount_blk;
1250
1251 *clean = true;
1252 }
1253 }
1254
1255 return 0;
1256}
1257
1258static void
1259xlog_set_state(
1260 struct xlog *log,
1261 xfs_daddr_t head_blk,
1262 struct xlog_rec_header *rhead,
1263 xfs_daddr_t rhead_blk,
1264 bool bump_cycle)
1265{
1266 /*
1267 * Reset log values according to the state of the log when we
1268 * crashed. In the case where head_blk == 0, we bump curr_cycle
1269 * one because the next write starts a new cycle rather than
1270 * continuing the cycle of the last good log record. At this
1271 * point we have guaranteed that all partial log records have been
1272 * accounted for. Therefore, we know that the last good log record
1273 * written was complete and ended exactly on the end boundary
1274 * of the physical log.
1275 */
1276 log->l_prev_block = rhead_blk;
1277 log->l_curr_block = (int)head_blk;
1278 log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1279 if (bump_cycle)
1280 log->l_curr_cycle++;
1281 atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1282 atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1283 xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1284 BBTOB(log->l_curr_block));
1285 xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1286 BBTOB(log->l_curr_block));
1287}
1288
1289/*
1290 * Find the sync block number or the tail of the log.
1291 *
1292 * This will be the block number of the last record to have its
1293 * associated buffers synced to disk. Every log record header has
1294 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
1295 * to get a sync block number. The only concern is to figure out which
1296 * log record header to believe.
1297 *
1298 * The following algorithm uses the log record header with the largest
1299 * lsn. The entire log record does not need to be valid. We only care
1300 * that the header is valid.
1301 *
1302 * We could speed up search by using current head_blk buffer, but it is not
1303 * available.
1304 */
1305STATIC int
1306xlog_find_tail(
1307 struct xlog *log,
1308 xfs_daddr_t *head_blk,
1309 xfs_daddr_t *tail_blk)
1310{
1311 xlog_rec_header_t *rhead;
1312 char *offset = NULL;
1313 char *buffer;
1314 int error;
1315 xfs_daddr_t rhead_blk;
1316 xfs_lsn_t tail_lsn;
1317 bool wrapped = false;
1318 bool clean = false;
1319
1320 /*
1321 * Find previous log record
1322 */
1323 if ((error = xlog_find_head(log, head_blk)))
1324 return error;
1325 ASSERT(*head_blk < INT_MAX);
1326
1327 buffer = xlog_alloc_buffer(log, 1);
1328 if (!buffer)
1329 return -ENOMEM;
1330 if (*head_blk == 0) { /* special case */
1331 error = xlog_bread(log, 0, 1, buffer, &offset);
1332 if (error)
1333 goto done;
1334
1335 if (xlog_get_cycle(offset) == 0) {
1336 *tail_blk = 0;
1337 /* leave all other log inited values alone */
1338 goto done;
1339 }
1340 }
1341
1342 /*
1343 * Search backwards through the log looking for the log record header
1344 * block. This wraps all the way back around to the head so something is
1345 * seriously wrong if we can't find it.
1346 */
1347 error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, buffer,
1348 &rhead_blk, &rhead, &wrapped);
1349 if (error < 0)
1350 return error;
1351 if (!error) {
1352 xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1353 return -EIO;
1354 }
1355 *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1356
1357 /*
1358 * Set the log state based on the current head record.
1359 */
1360 xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1361 tail_lsn = atomic64_read(&log->l_tail_lsn);
1362
1363 /*
1364 * Look for an unmount record at the head of the log. This sets the log
1365 * state to determine whether recovery is necessary.
1366 */
1367 error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1368 rhead_blk, buffer, &clean);
1369 if (error)
1370 goto done;
1371
1372 /*
1373 * Verify the log head if the log is not clean (e.g., we have anything
1374 * but an unmount record at the head). This uses CRC verification to
1375 * detect and trim torn writes. If discovered, CRC failures are
1376 * considered torn writes and the log head is trimmed accordingly.
1377 *
1378 * Note that we can only run CRC verification when the log is dirty
1379 * because there's no guarantee that the log data behind an unmount
1380 * record is compatible with the current architecture.
1381 */
1382 if (!clean) {
1383 xfs_daddr_t orig_head = *head_blk;
1384
1385 error = xlog_verify_head(log, head_blk, tail_blk, buffer,
1386 &rhead_blk, &rhead, &wrapped);
1387 if (error)
1388 goto done;
1389
1390 /* update in-core state again if the head changed */
1391 if (*head_blk != orig_head) {
1392 xlog_set_state(log, *head_blk, rhead, rhead_blk,
1393 wrapped);
1394 tail_lsn = atomic64_read(&log->l_tail_lsn);
1395 error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1396 rhead, rhead_blk, buffer,
1397 &clean);
1398 if (error)
1399 goto done;
1400 }
1401 }
1402
1403 /*
1404 * Note that the unmount was clean. If the unmount was not clean, we
1405 * need to know this to rebuild the superblock counters from the perag
1406 * headers if we have a filesystem using non-persistent counters.
1407 */
1408 if (clean)
1409 log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1410
1411 /*
1412 * Make sure that there are no blocks in front of the head
1413 * with the same cycle number as the head. This can happen
1414 * because we allow multiple outstanding log writes concurrently,
1415 * and the later writes might make it out before earlier ones.
1416 *
1417 * We use the lsn from before modifying it so that we'll never
1418 * overwrite the unmount record after a clean unmount.
1419 *
1420 * Do this only if we are going to recover the filesystem
1421 *
1422 * NOTE: This used to say "if (!readonly)"
1423 * However on Linux, we can & do recover a read-only filesystem.
1424 * We only skip recovery if NORECOVERY is specified on mount,
1425 * in which case we would not be here.
1426 *
1427 * But... if the -device- itself is readonly, just skip this.
1428 * We can't recover this device anyway, so it won't matter.
1429 */
1430 if (!xfs_readonly_buftarg(log->l_targ))
1431 error = xlog_clear_stale_blocks(log, tail_lsn);
1432
1433done:
1434 kmem_free(buffer);
1435
1436 if (error)
1437 xfs_warn(log->l_mp, "failed to locate log tail");
1438 return error;
1439}
1440
1441/*
1442 * Is the log zeroed at all?
1443 *
1444 * The last binary search should be changed to perform an X block read
1445 * once X becomes small enough. You can then search linearly through
1446 * the X blocks. This will cut down on the number of reads we need to do.
1447 *
1448 * If the log is partially zeroed, this routine will pass back the blkno
1449 * of the first block with cycle number 0. It won't have a complete LR
1450 * preceding it.
1451 *
1452 * Return:
1453 * 0 => the log is completely written to
1454 * 1 => use *blk_no as the first block of the log
1455 * <0 => error has occurred
1456 */
1457STATIC int
1458xlog_find_zeroed(
1459 struct xlog *log,
1460 xfs_daddr_t *blk_no)
1461{
1462 char *buffer;
1463 char *offset;
1464 uint first_cycle, last_cycle;
1465 xfs_daddr_t new_blk, last_blk, start_blk;
1466 xfs_daddr_t num_scan_bblks;
1467 int error, log_bbnum = log->l_logBBsize;
1468
1469 *blk_no = 0;
1470
1471 /* check totally zeroed log */
1472 buffer = xlog_alloc_buffer(log, 1);
1473 if (!buffer)
1474 return -ENOMEM;
1475 error = xlog_bread(log, 0, 1, buffer, &offset);
1476 if (error)
1477 goto out_free_buffer;
1478
1479 first_cycle = xlog_get_cycle(offset);
1480 if (first_cycle == 0) { /* completely zeroed log */
1481 *blk_no = 0;
1482 kmem_free(buffer);
1483 return 1;
1484 }
1485
1486 /* check partially zeroed log */
1487 error = xlog_bread(log, log_bbnum-1, 1, buffer, &offset);
1488 if (error)
1489 goto out_free_buffer;
1490
1491 last_cycle = xlog_get_cycle(offset);
1492 if (last_cycle != 0) { /* log completely written to */
1493 kmem_free(buffer);
1494 return 0;
1495 }
1496
1497 /* we have a partially zeroed log */
1498 last_blk = log_bbnum-1;
1499 error = xlog_find_cycle_start(log, buffer, 0, &last_blk, 0);
1500 if (error)
1501 goto out_free_buffer;
1502
1503 /*
1504 * Validate the answer. Because there is no way to guarantee that
1505 * the entire log is made up of log records which are the same size,
1506 * we scan over the defined maximum blocks. At this point, the maximum
1507 * is not chosen to mean anything special. XXXmiken
1508 */
1509 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1510 ASSERT(num_scan_bblks <= INT_MAX);
1511
1512 if (last_blk < num_scan_bblks)
1513 num_scan_bblks = last_blk;
1514 start_blk = last_blk - num_scan_bblks;
1515
1516 /*
1517 * We search for any instances of cycle number 0 that occur before
1518 * our current estimate of the head. What we're trying to detect is
1519 * 1 ... | 0 | 1 | 0...
1520 * ^ binary search ends here
1521 */
1522 if ((error = xlog_find_verify_cycle(log, start_blk,
1523 (int)num_scan_bblks, 0, &new_blk)))
1524 goto out_free_buffer;
1525 if (new_blk != -1)
1526 last_blk = new_blk;
1527
1528 /*
1529 * Potentially backup over partial log record write. We don't need
1530 * to search the end of the log because we know it is zero.
1531 */
1532 error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1533 if (error == 1)
1534 error = -EIO;
1535 if (error)
1536 goto out_free_buffer;
1537
1538 *blk_no = last_blk;
1539out_free_buffer:
1540 kmem_free(buffer);
1541 if (error)
1542 return error;
1543 return 1;
1544}
1545
1546/*
1547 * These are simple subroutines used by xlog_clear_stale_blocks() below
1548 * to initialize a buffer full of empty log record headers and write
1549 * them into the log.
1550 */
1551STATIC void
1552xlog_add_record(
1553 struct xlog *log,
1554 char *buf,
1555 int cycle,
1556 int block,
1557 int tail_cycle,
1558 int tail_block)
1559{
1560 xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
1561
1562 memset(buf, 0, BBSIZE);
1563 recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1564 recp->h_cycle = cpu_to_be32(cycle);
1565 recp->h_version = cpu_to_be32(
1566 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1567 recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1568 recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1569 recp->h_fmt = cpu_to_be32(XLOG_FMT);
1570 memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1571}
1572
1573STATIC int
1574xlog_write_log_records(
1575 struct xlog *log,
1576 int cycle,
1577 int start_block,
1578 int blocks,
1579 int tail_cycle,
1580 int tail_block)
1581{
1582 char *offset;
1583 char *buffer;
1584 int balign, ealign;
1585 int sectbb = log->l_sectBBsize;
1586 int end_block = start_block + blocks;
1587 int bufblks;
1588 int error = 0;
1589 int i, j = 0;
1590
1591 /*
1592 * Greedily allocate a buffer big enough to handle the full
1593 * range of basic blocks to be written. If that fails, try
1594 * a smaller size. We need to be able to write at least a
1595 * log sector, or we're out of luck.
1596 */
1597 bufblks = 1 << ffs(blocks);
1598 while (bufblks > log->l_logBBsize)
1599 bufblks >>= 1;
1600 while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
1601 bufblks >>= 1;
1602 if (bufblks < sectbb)
1603 return -ENOMEM;
1604 }
1605
1606 /* We may need to do a read at the start to fill in part of
1607 * the buffer in the starting sector not covered by the first
1608 * write below.
1609 */
1610 balign = round_down(start_block, sectbb);
1611 if (balign != start_block) {
1612 error = xlog_bread_noalign(log, start_block, 1, buffer);
1613 if (error)
1614 goto out_free_buffer;
1615
1616 j = start_block - balign;
1617 }
1618
1619 for (i = start_block; i < end_block; i += bufblks) {
1620 int bcount, endcount;
1621
1622 bcount = min(bufblks, end_block - start_block);
1623 endcount = bcount - j;
1624
1625 /* We may need to do a read at the end to fill in part of
1626 * the buffer in the final sector not covered by the write.
1627 * If this is the same sector as the above read, skip it.
1628 */
1629 ealign = round_down(end_block, sectbb);
1630 if (j == 0 && (start_block + endcount > ealign)) {
1631 error = xlog_bread_noalign(log, ealign, sectbb,
1632 buffer + BBTOB(ealign - start_block));
1633 if (error)
1634 break;
1635
1636 }
1637
1638 offset = buffer + xlog_align(log, start_block);
1639 for (; j < endcount; j++) {
1640 xlog_add_record(log, offset, cycle, i+j,
1641 tail_cycle, tail_block);
1642 offset += BBSIZE;
1643 }
1644 error = xlog_bwrite(log, start_block, endcount, buffer);
1645 if (error)
1646 break;
1647 start_block += endcount;
1648 j = 0;
1649 }
1650
1651out_free_buffer:
1652 kmem_free(buffer);
1653 return error;
1654}
1655
1656/*
1657 * This routine is called to blow away any incomplete log writes out
1658 * in front of the log head. We do this so that we won't become confused
1659 * if we come up, write only a little bit more, and then crash again.
1660 * If we leave the partial log records out there, this situation could
1661 * cause us to think those partial writes are valid blocks since they
1662 * have the current cycle number. We get rid of them by overwriting them
1663 * with empty log records with the old cycle number rather than the
1664 * current one.
1665 *
1666 * The tail lsn is passed in rather than taken from
1667 * the log so that we will not write over the unmount record after a
1668 * clean unmount in a 512 block log. Doing so would leave the log without
1669 * any valid log records in it until a new one was written. If we crashed
1670 * during that time we would not be able to recover.
1671 */
1672STATIC int
1673xlog_clear_stale_blocks(
1674 struct xlog *log,
1675 xfs_lsn_t tail_lsn)
1676{
1677 int tail_cycle, head_cycle;
1678 int tail_block, head_block;
1679 int tail_distance, max_distance;
1680 int distance;
1681 int error;
1682
1683 tail_cycle = CYCLE_LSN(tail_lsn);
1684 tail_block = BLOCK_LSN(tail_lsn);
1685 head_cycle = log->l_curr_cycle;
1686 head_block = log->l_curr_block;
1687
1688 /*
1689 * Figure out the distance between the new head of the log
1690 * and the tail. We want to write over any blocks beyond the
1691 * head that we may have written just before the crash, but
1692 * we don't want to overwrite the tail of the log.
1693 */
1694 if (head_cycle == tail_cycle) {
1695 /*
1696 * The tail is behind the head in the physical log,
1697 * so the distance from the head to the tail is the
1698 * distance from the head to the end of the log plus
1699 * the distance from the beginning of the log to the
1700 * tail.
1701 */
1702 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1703 XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1704 XFS_ERRLEVEL_LOW, log->l_mp);
1705 return -EFSCORRUPTED;
1706 }
1707 tail_distance = tail_block + (log->l_logBBsize - head_block);
1708 } else {
1709 /*
1710 * The head is behind the tail in the physical log,
1711 * so the distance from the head to the tail is just
1712 * the tail block minus the head block.
1713 */
1714 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1715 XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1716 XFS_ERRLEVEL_LOW, log->l_mp);
1717 return -EFSCORRUPTED;
1718 }
1719 tail_distance = tail_block - head_block;
1720 }
1721
1722 /*
1723 * If the head is right up against the tail, we can't clear
1724 * anything.
1725 */
1726 if (tail_distance <= 0) {
1727 ASSERT(tail_distance == 0);
1728 return 0;
1729 }
1730
1731 max_distance = XLOG_TOTAL_REC_SHIFT(log);
1732 /*
1733 * Take the smaller of the maximum amount of outstanding I/O
1734 * we could have and the distance to the tail to clear out.
1735 * We take the smaller so that we don't overwrite the tail and
1736 * we don't waste all day writing from the head to the tail
1737 * for no reason.
1738 */
1739 max_distance = min(max_distance, tail_distance);
1740
1741 if ((head_block + max_distance) <= log->l_logBBsize) {
1742 /*
1743 * We can stomp all the blocks we need to without
1744 * wrapping around the end of the log. Just do it
1745 * in a single write. Use the cycle number of the
1746 * current cycle minus one so that the log will look like:
1747 * n ... | n - 1 ...
1748 */
1749 error = xlog_write_log_records(log, (head_cycle - 1),
1750 head_block, max_distance, tail_cycle,
1751 tail_block);
1752 if (error)
1753 return error;
1754 } else {
1755 /*
1756 * We need to wrap around the end of the physical log in
1757 * order to clear all the blocks. Do it in two separate
1758 * I/Os. The first write should be from the head to the
1759 * end of the physical log, and it should use the current
1760 * cycle number minus one just like above.
1761 */
1762 distance = log->l_logBBsize - head_block;
1763 error = xlog_write_log_records(log, (head_cycle - 1),
1764 head_block, distance, tail_cycle,
1765 tail_block);
1766
1767 if (error)
1768 return error;
1769
1770 /*
1771 * Now write the blocks at the start of the physical log.
1772 * This writes the remainder of the blocks we want to clear.
1773 * It uses the current cycle number since we're now on the
1774 * same cycle as the head so that we get:
1775 * n ... n ... | n - 1 ...
1776 * ^^^^^ blocks we're writing
1777 */
1778 distance = max_distance - (log->l_logBBsize - head_block);
1779 error = xlog_write_log_records(log, head_cycle, 0, distance,
1780 tail_cycle, tail_block);
1781 if (error)
1782 return error;
1783 }
1784
1785 return 0;
1786}
1787
1788/******************************************************************************
1789 *
1790 * Log recover routines
1791 *
1792 ******************************************************************************
1793 */
1794
1795/*
1796 * Sort the log items in the transaction.
1797 *
1798 * The ordering constraints are defined by the inode allocation and unlink
1799 * behaviour. The rules are:
1800 *
1801 * 1. Every item is only logged once in a given transaction. Hence it
1802 * represents the last logged state of the item. Hence ordering is
1803 * dependent on the order in which operations need to be performed so
1804 * required initial conditions are always met.
1805 *
1806 * 2. Cancelled buffers are recorded in pass 1 in a separate table and
1807 * there's nothing to replay from them so we can simply cull them
1808 * from the transaction. However, we can't do that until after we've
1809 * replayed all the other items because they may be dependent on the
1810 * cancelled buffer and replaying the cancelled buffer can remove it
1811 * form the cancelled buffer table. Hence they have tobe done last.
1812 *
1813 * 3. Inode allocation buffers must be replayed before inode items that
1814 * read the buffer and replay changes into it. For filesystems using the
1815 * ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1816 * treated the same as inode allocation buffers as they create and
1817 * initialise the buffers directly.
1818 *
1819 * 4. Inode unlink buffers must be replayed after inode items are replayed.
1820 * This ensures that inodes are completely flushed to the inode buffer
1821 * in a "free" state before we remove the unlinked inode list pointer.
1822 *
1823 * Hence the ordering needs to be inode allocation buffers first, inode items
1824 * second, inode unlink buffers third and cancelled buffers last.
1825 *
1826 * But there's a problem with that - we can't tell an inode allocation buffer
1827 * apart from a regular buffer, so we can't separate them. We can, however,
1828 * tell an inode unlink buffer from the others, and so we can separate them out
1829 * from all the other buffers and move them to last.
1830 *
1831 * Hence, 4 lists, in order from head to tail:
1832 * - buffer_list for all buffers except cancelled/inode unlink buffers
1833 * - item_list for all non-buffer items
1834 * - inode_buffer_list for inode unlink buffers
1835 * - cancel_list for the cancelled buffers
1836 *
1837 * Note that we add objects to the tail of the lists so that first-to-last
1838 * ordering is preserved within the lists. Adding objects to the head of the
1839 * list means when we traverse from the head we walk them in last-to-first
1840 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1841 * but for all other items there may be specific ordering that we need to
1842 * preserve.
1843 */
1844STATIC int
1845xlog_recover_reorder_trans(
1846 struct xlog *log,
1847 struct xlog_recover *trans,
1848 int pass)
1849{
1850 xlog_recover_item_t *item, *n;
1851 int error = 0;
1852 LIST_HEAD(sort_list);
1853 LIST_HEAD(cancel_list);
1854 LIST_HEAD(buffer_list);
1855 LIST_HEAD(inode_buffer_list);
1856 LIST_HEAD(inode_list);
1857
1858 list_splice_init(&trans->r_itemq, &sort_list);
1859 list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1860 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1861
1862 switch (ITEM_TYPE(item)) {
1863 case XFS_LI_ICREATE:
1864 list_move_tail(&item->ri_list, &buffer_list);
1865 break;
1866 case XFS_LI_BUF:
1867 if (buf_f->blf_flags & XFS_BLF_CANCEL) {
1868 trace_xfs_log_recover_item_reorder_head(log,
1869 trans, item, pass);
1870 list_move(&item->ri_list, &cancel_list);
1871 break;
1872 }
1873 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
1874 list_move(&item->ri_list, &inode_buffer_list);
1875 break;
1876 }
1877 list_move_tail(&item->ri_list, &buffer_list);
1878 break;
1879 case XFS_LI_INODE:
1880 case XFS_LI_DQUOT:
1881 case XFS_LI_QUOTAOFF:
1882 case XFS_LI_EFD:
1883 case XFS_LI_EFI:
1884 case XFS_LI_RUI:
1885 case XFS_LI_RUD:
1886 case XFS_LI_CUI:
1887 case XFS_LI_CUD:
1888 case XFS_LI_BUI:
1889 case XFS_LI_BUD:
1890 trace_xfs_log_recover_item_reorder_tail(log,
1891 trans, item, pass);
1892 list_move_tail(&item->ri_list, &inode_list);
1893 break;
1894 default:
1895 xfs_warn(log->l_mp,
1896 "%s: unrecognized type of log operation",
1897 __func__);
1898 ASSERT(0);
1899 /*
1900 * return the remaining items back to the transaction
1901 * item list so they can be freed in caller.
1902 */
1903 if (!list_empty(&sort_list))
1904 list_splice_init(&sort_list, &trans->r_itemq);
1905 error = -EIO;
1906 goto out;
1907 }
1908 }
1909out:
1910 ASSERT(list_empty(&sort_list));
1911 if (!list_empty(&buffer_list))
1912 list_splice(&buffer_list, &trans->r_itemq);
1913 if (!list_empty(&inode_list))
1914 list_splice_tail(&inode_list, &trans->r_itemq);
1915 if (!list_empty(&inode_buffer_list))
1916 list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1917 if (!list_empty(&cancel_list))
1918 list_splice_tail(&cancel_list, &trans->r_itemq);
1919 return error;
1920}
1921
1922/*
1923 * Build up the table of buf cancel records so that we don't replay
1924 * cancelled data in the second pass. For buffer records that are
1925 * not cancel records, there is nothing to do here so we just return.
1926 *
1927 * If we get a cancel record which is already in the table, this indicates
1928 * that the buffer was cancelled multiple times. In order to ensure
1929 * that during pass 2 we keep the record in the table until we reach its
1930 * last occurrence in the log, we keep a reference count in the cancel
1931 * record in the table to tell us how many times we expect to see this
1932 * record during the second pass.
1933 */
1934STATIC int
1935xlog_recover_buffer_pass1(
1936 struct xlog *log,
1937 struct xlog_recover_item *item)
1938{
1939 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1940 struct list_head *bucket;
1941 struct xfs_buf_cancel *bcp;
1942
1943 /*
1944 * If this isn't a cancel buffer item, then just return.
1945 */
1946 if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1947 trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1948 return 0;
1949 }
1950
1951 /*
1952 * Insert an xfs_buf_cancel record into the hash table of them.
1953 * If there is already an identical record, bump its reference count.
1954 */
1955 bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1956 list_for_each_entry(bcp, bucket, bc_list) {
1957 if (bcp->bc_blkno == buf_f->blf_blkno &&
1958 bcp->bc_len == buf_f->blf_len) {
1959 bcp->bc_refcount++;
1960 trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1961 return 0;
1962 }
1963 }
1964
1965 bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), 0);
1966 bcp->bc_blkno = buf_f->blf_blkno;
1967 bcp->bc_len = buf_f->blf_len;
1968 bcp->bc_refcount = 1;
1969 list_add_tail(&bcp->bc_list, bucket);
1970
1971 trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1972 return 0;
1973}
1974
1975/*
1976 * Check to see whether the buffer being recovered has a corresponding
1977 * entry in the buffer cancel record table. If it is, return the cancel
1978 * buffer structure to the caller.
1979 */
1980STATIC struct xfs_buf_cancel *
1981xlog_peek_buffer_cancelled(
1982 struct xlog *log,
1983 xfs_daddr_t blkno,
1984 uint len,
1985 unsigned short flags)
1986{
1987 struct list_head *bucket;
1988 struct xfs_buf_cancel *bcp;
1989
1990 if (!log->l_buf_cancel_table) {
1991 /* empty table means no cancelled buffers in the log */
1992 ASSERT(!(flags & XFS_BLF_CANCEL));
1993 return NULL;
1994 }
1995
1996 bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
1997 list_for_each_entry(bcp, bucket, bc_list) {
1998 if (bcp->bc_blkno == blkno && bcp->bc_len == len)
1999 return bcp;
2000 }
2001
2002 /*
2003 * We didn't find a corresponding entry in the table, so return 0 so
2004 * that the buffer is NOT cancelled.
2005 */
2006 ASSERT(!(flags & XFS_BLF_CANCEL));
2007 return NULL;
2008}
2009
2010/*
2011 * If the buffer is being cancelled then return 1 so that it will be cancelled,
2012 * otherwise return 0. If the buffer is actually a buffer cancel item
2013 * (XFS_BLF_CANCEL is set), then decrement the refcount on the entry in the
2014 * table and remove it from the table if this is the last reference.
2015 *
2016 * We remove the cancel record from the table when we encounter its last
2017 * occurrence in the log so that if the same buffer is re-used again after its
2018 * last cancellation we actually replay the changes made at that point.
2019 */
2020STATIC int
2021xlog_check_buffer_cancelled(
2022 struct xlog *log,
2023 xfs_daddr_t blkno,
2024 uint len,
2025 unsigned short flags)
2026{
2027 struct xfs_buf_cancel *bcp;
2028
2029 bcp = xlog_peek_buffer_cancelled(log, blkno, len, flags);
2030 if (!bcp)
2031 return 0;
2032
2033 /*
2034 * We've go a match, so return 1 so that the recovery of this buffer
2035 * is cancelled. If this buffer is actually a buffer cancel log
2036 * item, then decrement the refcount on the one in the table and
2037 * remove it if this is the last reference.
2038 */
2039 if (flags & XFS_BLF_CANCEL) {
2040 if (--bcp->bc_refcount == 0) {
2041 list_del(&bcp->bc_list);
2042 kmem_free(bcp);
2043 }
2044 }
2045 return 1;
2046}
2047
2048/*
2049 * Perform recovery for a buffer full of inodes. In these buffers, the only
2050 * data which should be recovered is that which corresponds to the
2051 * di_next_unlinked pointers in the on disk inode structures. The rest of the
2052 * data for the inodes is always logged through the inodes themselves rather
2053 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
2054 *
2055 * The only time when buffers full of inodes are fully recovered is when the
2056 * buffer is full of newly allocated inodes. In this case the buffer will
2057 * not be marked as an inode buffer and so will be sent to
2058 * xlog_recover_do_reg_buffer() below during recovery.
2059 */
2060STATIC int
2061xlog_recover_do_inode_buffer(
2062 struct xfs_mount *mp,
2063 xlog_recover_item_t *item,
2064 struct xfs_buf *bp,
2065 xfs_buf_log_format_t *buf_f)
2066{
2067 int i;
2068 int item_index = 0;
2069 int bit = 0;
2070 int nbits = 0;
2071 int reg_buf_offset = 0;
2072 int reg_buf_bytes = 0;
2073 int next_unlinked_offset;
2074 int inodes_per_buf;
2075 xfs_agino_t *logged_nextp;
2076 xfs_agino_t *buffer_nextp;
2077
2078 trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
2079
2080 /*
2081 * Post recovery validation only works properly on CRC enabled
2082 * filesystems.
2083 */
2084 if (xfs_sb_version_hascrc(&mp->m_sb))
2085 bp->b_ops = &xfs_inode_buf_ops;
2086
2087 inodes_per_buf = BBTOB(bp->b_length) >> mp->m_sb.sb_inodelog;
2088 for (i = 0; i < inodes_per_buf; i++) {
2089 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
2090 offsetof(xfs_dinode_t, di_next_unlinked);
2091
2092 while (next_unlinked_offset >=
2093 (reg_buf_offset + reg_buf_bytes)) {
2094 /*
2095 * The next di_next_unlinked field is beyond
2096 * the current logged region. Find the next
2097 * logged region that contains or is beyond
2098 * the current di_next_unlinked field.
2099 */
2100 bit += nbits;
2101 bit = xfs_next_bit(buf_f->blf_data_map,
2102 buf_f->blf_map_size, bit);
2103
2104 /*
2105 * If there are no more logged regions in the
2106 * buffer, then we're done.
2107 */
2108 if (bit == -1)
2109 return 0;
2110
2111 nbits = xfs_contig_bits(buf_f->blf_data_map,
2112 buf_f->blf_map_size, bit);
2113 ASSERT(nbits > 0);
2114 reg_buf_offset = bit << XFS_BLF_SHIFT;
2115 reg_buf_bytes = nbits << XFS_BLF_SHIFT;
2116 item_index++;
2117 }
2118
2119 /*
2120 * If the current logged region starts after the current
2121 * di_next_unlinked field, then move on to the next
2122 * di_next_unlinked field.
2123 */
2124 if (next_unlinked_offset < reg_buf_offset)
2125 continue;
2126
2127 ASSERT(item->ri_buf[item_index].i_addr != NULL);
2128 ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
2129 ASSERT((reg_buf_offset + reg_buf_bytes) <= BBTOB(bp->b_length));
2130
2131 /*
2132 * The current logged region contains a copy of the
2133 * current di_next_unlinked field. Extract its value
2134 * and copy it to the buffer copy.
2135 */
2136 logged_nextp = item->ri_buf[item_index].i_addr +
2137 next_unlinked_offset - reg_buf_offset;
2138 if (unlikely(*logged_nextp == 0)) {
2139 xfs_alert(mp,
2140 "Bad inode buffer log record (ptr = "PTR_FMT", bp = "PTR_FMT"). "
2141 "Trying to replay bad (0) inode di_next_unlinked field.",
2142 item, bp);
2143 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
2144 XFS_ERRLEVEL_LOW, mp);
2145 return -EFSCORRUPTED;
2146 }
2147
2148 buffer_nextp = xfs_buf_offset(bp, next_unlinked_offset);
2149 *buffer_nextp = *logged_nextp;
2150
2151 /*
2152 * If necessary, recalculate the CRC in the on-disk inode. We
2153 * have to leave the inode in a consistent state for whoever
2154 * reads it next....
2155 */
2156 xfs_dinode_calc_crc(mp,
2157 xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
2158
2159 }
2160
2161 return 0;
2162}
2163
2164/*
2165 * V5 filesystems know the age of the buffer on disk being recovered. We can
2166 * have newer objects on disk than we are replaying, and so for these cases we
2167 * don't want to replay the current change as that will make the buffer contents
2168 * temporarily invalid on disk.
2169 *
2170 * The magic number might not match the buffer type we are going to recover
2171 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags. Hence
2172 * extract the LSN of the existing object in the buffer based on it's current
2173 * magic number. If we don't recognise the magic number in the buffer, then
2174 * return a LSN of -1 so that the caller knows it was an unrecognised block and
2175 * so can recover the buffer.
2176 *
2177 * Note: we cannot rely solely on magic number matches to determine that the
2178 * buffer has a valid LSN - we also need to verify that it belongs to this
2179 * filesystem, so we need to extract the object's LSN and compare it to that
2180 * which we read from the superblock. If the UUIDs don't match, then we've got a
2181 * stale metadata block from an old filesystem instance that we need to recover
2182 * over the top of.
2183 */
2184static xfs_lsn_t
2185xlog_recover_get_buf_lsn(
2186 struct xfs_mount *mp,
2187 struct xfs_buf *bp)
2188{
2189 uint32_t magic32;
2190 uint16_t magic16;
2191 uint16_t magicda;
2192 void *blk = bp->b_addr;
2193 uuid_t *uuid;
2194 xfs_lsn_t lsn = -1;
2195
2196 /* v4 filesystems always recover immediately */
2197 if (!xfs_sb_version_hascrc(&mp->m_sb))
2198 goto recover_immediately;
2199
2200 magic32 = be32_to_cpu(*(__be32 *)blk);
2201 switch (magic32) {
2202 case XFS_ABTB_CRC_MAGIC:
2203 case XFS_ABTC_CRC_MAGIC:
2204 case XFS_ABTB_MAGIC:
2205 case XFS_ABTC_MAGIC:
2206 case XFS_RMAP_CRC_MAGIC:
2207 case XFS_REFC_CRC_MAGIC:
2208 case XFS_IBT_CRC_MAGIC:
2209 case XFS_IBT_MAGIC: {
2210 struct xfs_btree_block *btb = blk;
2211
2212 lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
2213 uuid = &btb->bb_u.s.bb_uuid;
2214 break;
2215 }
2216 case XFS_BMAP_CRC_MAGIC:
2217 case XFS_BMAP_MAGIC: {
2218 struct xfs_btree_block *btb = blk;
2219
2220 lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
2221 uuid = &btb->bb_u.l.bb_uuid;
2222 break;
2223 }
2224 case XFS_AGF_MAGIC:
2225 lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
2226 uuid = &((struct xfs_agf *)blk)->agf_uuid;
2227 break;
2228 case XFS_AGFL_MAGIC:
2229 lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
2230 uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
2231 break;
2232 case XFS_AGI_MAGIC:
2233 lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
2234 uuid = &((struct xfs_agi *)blk)->agi_uuid;
2235 break;
2236 case XFS_SYMLINK_MAGIC:
2237 lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
2238 uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
2239 break;
2240 case XFS_DIR3_BLOCK_MAGIC:
2241 case XFS_DIR3_DATA_MAGIC:
2242 case XFS_DIR3_FREE_MAGIC:
2243 lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
2244 uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
2245 break;
2246 case XFS_ATTR3_RMT_MAGIC:
2247 /*
2248 * Remote attr blocks are written synchronously, rather than
2249 * being logged. That means they do not contain a valid LSN
2250 * (i.e. transactionally ordered) in them, and hence any time we
2251 * see a buffer to replay over the top of a remote attribute
2252 * block we should simply do so.
2253 */
2254 goto recover_immediately;
2255 case XFS_SB_MAGIC:
2256 /*
2257 * superblock uuids are magic. We may or may not have a
2258 * sb_meta_uuid on disk, but it will be set in the in-core
2259 * superblock. We set the uuid pointer for verification
2260 * according to the superblock feature mask to ensure we check
2261 * the relevant UUID in the superblock.
2262 */
2263 lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
2264 if (xfs_sb_version_hasmetauuid(&mp->m_sb))
2265 uuid = &((struct xfs_dsb *)blk)->sb_meta_uuid;
2266 else
2267 uuid = &((struct xfs_dsb *)blk)->sb_uuid;
2268 break;
2269 default:
2270 break;
2271 }
2272
2273 if (lsn != (xfs_lsn_t)-1) {
2274 if (!uuid_equal(&mp->m_sb.sb_meta_uuid, uuid))
2275 goto recover_immediately;
2276 return lsn;
2277 }
2278
2279 magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
2280 switch (magicda) {
2281 case XFS_DIR3_LEAF1_MAGIC:
2282 case XFS_DIR3_LEAFN_MAGIC:
2283 case XFS_DA3_NODE_MAGIC:
2284 lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
2285 uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
2286 break;
2287 default:
2288 break;
2289 }
2290
2291 if (lsn != (xfs_lsn_t)-1) {
2292 if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
2293 goto recover_immediately;
2294 return lsn;
2295 }
2296
2297 /*
2298 * We do individual object checks on dquot and inode buffers as they
2299 * have their own individual LSN records. Also, we could have a stale
2300 * buffer here, so we have to at least recognise these buffer types.
2301 *
2302 * A notd complexity here is inode unlinked list processing - it logs
2303 * the inode directly in the buffer, but we don't know which inodes have
2304 * been modified, and there is no global buffer LSN. Hence we need to
2305 * recover all inode buffer types immediately. This problem will be
2306 * fixed by logical logging of the unlinked list modifications.
2307 */
2308 magic16 = be16_to_cpu(*(__be16 *)blk);
2309 switch (magic16) {
2310 case XFS_DQUOT_MAGIC:
2311 case XFS_DINODE_MAGIC:
2312 goto recover_immediately;
2313 default:
2314 break;
2315 }
2316
2317 /* unknown buffer contents, recover immediately */
2318
2319recover_immediately:
2320 return (xfs_lsn_t)-1;
2321
2322}
2323
2324/*
2325 * Validate the recovered buffer is of the correct type and attach the
2326 * appropriate buffer operations to them for writeback. Magic numbers are in a
2327 * few places:
2328 * the first 16 bits of the buffer (inode buffer, dquot buffer),
2329 * the first 32 bits of the buffer (most blocks),
2330 * inside a struct xfs_da_blkinfo at the start of the buffer.
2331 */
2332static void
2333xlog_recover_validate_buf_type(
2334 struct xfs_mount *mp,
2335 struct xfs_buf *bp,
2336 xfs_buf_log_format_t *buf_f,
2337 xfs_lsn_t current_lsn)
2338{
2339 struct xfs_da_blkinfo *info = bp->b_addr;
2340 uint32_t magic32;
2341 uint16_t magic16;
2342 uint16_t magicda;
2343 char *warnmsg = NULL;
2344
2345 /*
2346 * We can only do post recovery validation on items on CRC enabled
2347 * fielsystems as we need to know when the buffer was written to be able
2348 * to determine if we should have replayed the item. If we replay old
2349 * metadata over a newer buffer, then it will enter a temporarily
2350 * inconsistent state resulting in verification failures. Hence for now
2351 * just avoid the verification stage for non-crc filesystems
2352 */
2353 if (!xfs_sb_version_hascrc(&mp->m_sb))
2354 return;
2355
2356 magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
2357 magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
2358 magicda = be16_to_cpu(info->magic);
2359 switch (xfs_blft_from_flags(buf_f)) {
2360 case XFS_BLFT_BTREE_BUF:
2361 switch (magic32) {
2362 case XFS_ABTB_CRC_MAGIC:
2363 case XFS_ABTB_MAGIC:
2364 bp->b_ops = &xfs_bnobt_buf_ops;
2365 break;
2366 case XFS_ABTC_CRC_MAGIC:
2367 case XFS_ABTC_MAGIC:
2368 bp->b_ops = &xfs_cntbt_buf_ops;
2369 break;
2370 case XFS_IBT_CRC_MAGIC:
2371 case XFS_IBT_MAGIC:
2372 bp->b_ops = &xfs_inobt_buf_ops;
2373 break;
2374 case XFS_FIBT_CRC_MAGIC:
2375 case XFS_FIBT_MAGIC:
2376 bp->b_ops = &xfs_finobt_buf_ops;
2377 break;
2378 case XFS_BMAP_CRC_MAGIC:
2379 case XFS_BMAP_MAGIC:
2380 bp->b_ops = &xfs_bmbt_buf_ops;
2381 break;
2382 case XFS_RMAP_CRC_MAGIC:
2383 bp->b_ops = &xfs_rmapbt_buf_ops;
2384 break;
2385 case XFS_REFC_CRC_MAGIC:
2386 bp->b_ops = &xfs_refcountbt_buf_ops;
2387 break;
2388 default:
2389 warnmsg = "Bad btree block magic!";
2390 break;
2391 }
2392 break;
2393 case XFS_BLFT_AGF_BUF:
2394 if (magic32 != XFS_AGF_MAGIC) {
2395 warnmsg = "Bad AGF block magic!";
2396 break;
2397 }
2398 bp->b_ops = &xfs_agf_buf_ops;
2399 break;
2400 case XFS_BLFT_AGFL_BUF:
2401 if (magic32 != XFS_AGFL_MAGIC) {
2402 warnmsg = "Bad AGFL block magic!";
2403 break;
2404 }
2405 bp->b_ops = &xfs_agfl_buf_ops;
2406 break;
2407 case XFS_BLFT_AGI_BUF:
2408 if (magic32 != XFS_AGI_MAGIC) {
2409 warnmsg = "Bad AGI block magic!";
2410 break;
2411 }
2412 bp->b_ops = &xfs_agi_buf_ops;
2413 break;
2414 case XFS_BLFT_UDQUOT_BUF:
2415 case XFS_BLFT_PDQUOT_BUF:
2416 case XFS_BLFT_GDQUOT_BUF:
2417#ifdef CONFIG_XFS_QUOTA
2418 if (magic16 != XFS_DQUOT_MAGIC) {
2419 warnmsg = "Bad DQUOT block magic!";
2420 break;
2421 }
2422 bp->b_ops = &xfs_dquot_buf_ops;
2423#else
2424 xfs_alert(mp,
2425 "Trying to recover dquots without QUOTA support built in!");
2426 ASSERT(0);
2427#endif
2428 break;
2429 case XFS_BLFT_DINO_BUF:
2430 if (magic16 != XFS_DINODE_MAGIC) {
2431 warnmsg = "Bad INODE block magic!";
2432 break;
2433 }
2434 bp->b_ops = &xfs_inode_buf_ops;
2435 break;
2436 case XFS_BLFT_SYMLINK_BUF:
2437 if (magic32 != XFS_SYMLINK_MAGIC) {
2438 warnmsg = "Bad symlink block magic!";
2439 break;
2440 }
2441 bp->b_ops = &xfs_symlink_buf_ops;
2442 break;
2443 case XFS_BLFT_DIR_BLOCK_BUF:
2444 if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
2445 magic32 != XFS_DIR3_BLOCK_MAGIC) {
2446 warnmsg = "Bad dir block magic!";
2447 break;
2448 }
2449 bp->b_ops = &xfs_dir3_block_buf_ops;
2450 break;
2451 case XFS_BLFT_DIR_DATA_BUF:
2452 if (magic32 != XFS_DIR2_DATA_MAGIC &&
2453 magic32 != XFS_DIR3_DATA_MAGIC) {
2454 warnmsg = "Bad dir data magic!";
2455 break;
2456 }
2457 bp->b_ops = &xfs_dir3_data_buf_ops;
2458 break;
2459 case XFS_BLFT_DIR_FREE_BUF:
2460 if (magic32 != XFS_DIR2_FREE_MAGIC &&
2461 magic32 != XFS_DIR3_FREE_MAGIC) {
2462 warnmsg = "Bad dir3 free magic!";
2463 break;
2464 }
2465 bp->b_ops = &xfs_dir3_free_buf_ops;
2466 break;
2467 case XFS_BLFT_DIR_LEAF1_BUF:
2468 if (magicda != XFS_DIR2_LEAF1_MAGIC &&
2469 magicda != XFS_DIR3_LEAF1_MAGIC) {
2470 warnmsg = "Bad dir leaf1 magic!";
2471 break;
2472 }
2473 bp->b_ops = &xfs_dir3_leaf1_buf_ops;
2474 break;
2475 case XFS_BLFT_DIR_LEAFN_BUF:
2476 if (magicda != XFS_DIR2_LEAFN_MAGIC &&
2477 magicda != XFS_DIR3_LEAFN_MAGIC) {
2478 warnmsg = "Bad dir leafn magic!";
2479 break;
2480 }
2481 bp->b_ops = &xfs_dir3_leafn_buf_ops;
2482 break;
2483 case XFS_BLFT_DA_NODE_BUF:
2484 if (magicda != XFS_DA_NODE_MAGIC &&
2485 magicda != XFS_DA3_NODE_MAGIC) {
2486 warnmsg = "Bad da node magic!";
2487 break;
2488 }
2489 bp->b_ops = &xfs_da3_node_buf_ops;
2490 break;
2491 case XFS_BLFT_ATTR_LEAF_BUF:
2492 if (magicda != XFS_ATTR_LEAF_MAGIC &&
2493 magicda != XFS_ATTR3_LEAF_MAGIC) {
2494 warnmsg = "Bad attr leaf magic!";
2495 break;
2496 }
2497 bp->b_ops = &xfs_attr3_leaf_buf_ops;
2498 break;
2499 case XFS_BLFT_ATTR_RMT_BUF:
2500 if (magic32 != XFS_ATTR3_RMT_MAGIC) {
2501 warnmsg = "Bad attr remote magic!";
2502 break;
2503 }
2504 bp->b_ops = &xfs_attr3_rmt_buf_ops;
2505 break;
2506 case XFS_BLFT_SB_BUF:
2507 if (magic32 != XFS_SB_MAGIC) {
2508 warnmsg = "Bad SB block magic!";
2509 break;
2510 }
2511 bp->b_ops = &xfs_sb_buf_ops;
2512 break;
2513#ifdef CONFIG_XFS_RT
2514 case XFS_BLFT_RTBITMAP_BUF:
2515 case XFS_BLFT_RTSUMMARY_BUF:
2516 /* no magic numbers for verification of RT buffers */
2517 bp->b_ops = &xfs_rtbuf_ops;
2518 break;
2519#endif /* CONFIG_XFS_RT */
2520 default:
2521 xfs_warn(mp, "Unknown buffer type %d!",
2522 xfs_blft_from_flags(buf_f));
2523 break;
2524 }
2525
2526 /*
2527 * Nothing else to do in the case of a NULL current LSN as this means
2528 * the buffer is more recent than the change in the log and will be
2529 * skipped.
2530 */
2531 if (current_lsn == NULLCOMMITLSN)
2532 return;
2533
2534 if (warnmsg) {
2535 xfs_warn(mp, warnmsg);
2536 ASSERT(0);
2537 }
2538
2539 /*
2540 * We must update the metadata LSN of the buffer as it is written out to
2541 * ensure that older transactions never replay over this one and corrupt
2542 * the buffer. This can occur if log recovery is interrupted at some
2543 * point after the current transaction completes, at which point a
2544 * subsequent mount starts recovery from the beginning.
2545 *
2546 * Write verifiers update the metadata LSN from log items attached to
2547 * the buffer. Therefore, initialize a bli purely to carry the LSN to
2548 * the verifier. We'll clean it up in our ->iodone() callback.
2549 */
2550 if (bp->b_ops) {
2551 struct xfs_buf_log_item *bip;
2552
2553 ASSERT(!bp->b_iodone || bp->b_iodone == xlog_recover_iodone);
2554 bp->b_iodone = xlog_recover_iodone;
2555 xfs_buf_item_init(bp, mp);
2556 bip = bp->b_log_item;
2557 bip->bli_item.li_lsn = current_lsn;
2558 }
2559}
2560
2561/*
2562 * Perform a 'normal' buffer recovery. Each logged region of the
2563 * buffer should be copied over the corresponding region in the
2564 * given buffer. The bitmap in the buf log format structure indicates
2565 * where to place the logged data.
2566 */
2567STATIC void
2568xlog_recover_do_reg_buffer(
2569 struct xfs_mount *mp,
2570 xlog_recover_item_t *item,
2571 struct xfs_buf *bp,
2572 xfs_buf_log_format_t *buf_f,
2573 xfs_lsn_t current_lsn)
2574{
2575 int i;
2576 int bit;
2577 int nbits;
2578 xfs_failaddr_t fa;
2579
2580 trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
2581
2582 bit = 0;
2583 i = 1; /* 0 is the buf format structure */
2584 while (1) {
2585 bit = xfs_next_bit(buf_f->blf_data_map,
2586 buf_f->blf_map_size, bit);
2587 if (bit == -1)
2588 break;
2589 nbits = xfs_contig_bits(buf_f->blf_data_map,
2590 buf_f->blf_map_size, bit);
2591 ASSERT(nbits > 0);
2592 ASSERT(item->ri_buf[i].i_addr != NULL);
2593 ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
2594 ASSERT(BBTOB(bp->b_length) >=
2595 ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
2596
2597 /*
2598 * The dirty regions logged in the buffer, even though
2599 * contiguous, may span multiple chunks. This is because the
2600 * dirty region may span a physical page boundary in a buffer
2601 * and hence be split into two separate vectors for writing into
2602 * the log. Hence we need to trim nbits back to the length of
2603 * the current region being copied out of the log.
2604 */
2605 if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
2606 nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
2607
2608 /*
2609 * Do a sanity check if this is a dquot buffer. Just checking
2610 * the first dquot in the buffer should do. XXXThis is
2611 * probably a good thing to do for other buf types also.
2612 */
2613 fa = NULL;
2614 if (buf_f->blf_flags &
2615 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2616 if (item->ri_buf[i].i_addr == NULL) {
2617 xfs_alert(mp,
2618 "XFS: NULL dquot in %s.", __func__);
2619 goto next;
2620 }
2621 if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
2622 xfs_alert(mp,
2623 "XFS: dquot too small (%d) in %s.",
2624 item->ri_buf[i].i_len, __func__);
2625 goto next;
2626 }
2627 fa = xfs_dquot_verify(mp, item->ri_buf[i].i_addr,
2628 -1, 0);
2629 if (fa) {
2630 xfs_alert(mp,
2631 "dquot corrupt at %pS trying to replay into block 0x%llx",
2632 fa, bp->b_bn);
2633 goto next;
2634 }
2635 }
2636
2637 memcpy(xfs_buf_offset(bp,
2638 (uint)bit << XFS_BLF_SHIFT), /* dest */
2639 item->ri_buf[i].i_addr, /* source */
2640 nbits<<XFS_BLF_SHIFT); /* length */
2641 next:
2642 i++;
2643 bit += nbits;
2644 }
2645
2646 /* Shouldn't be any more regions */
2647 ASSERT(i == item->ri_total);
2648
2649 xlog_recover_validate_buf_type(mp, bp, buf_f, current_lsn);
2650}
2651
2652/*
2653 * Perform a dquot buffer recovery.
2654 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
2655 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2656 * Else, treat it as a regular buffer and do recovery.
2657 *
2658 * Return false if the buffer was tossed and true if we recovered the buffer to
2659 * indicate to the caller if the buffer needs writing.
2660 */
2661STATIC bool
2662xlog_recover_do_dquot_buffer(
2663 struct xfs_mount *mp,
2664 struct xlog *log,
2665 struct xlog_recover_item *item,
2666 struct xfs_buf *bp,
2667 struct xfs_buf_log_format *buf_f)
2668{
2669 uint type;
2670
2671 trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2672
2673 /*
2674 * Filesystems are required to send in quota flags at mount time.
2675 */
2676 if (!mp->m_qflags)
2677 return false;
2678
2679 type = 0;
2680 if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2681 type |= XFS_DQ_USER;
2682 if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2683 type |= XFS_DQ_PROJ;
2684 if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2685 type |= XFS_DQ_GROUP;
2686 /*
2687 * This type of quotas was turned off, so ignore this buffer
2688 */
2689 if (log->l_quotaoffs_flag & type)
2690 return false;
2691
2692 xlog_recover_do_reg_buffer(mp, item, bp, buf_f, NULLCOMMITLSN);
2693 return true;
2694}
2695
2696/*
2697 * This routine replays a modification made to a buffer at runtime.
2698 * There are actually two types of buffer, regular and inode, which
2699 * are handled differently. Inode buffers are handled differently
2700 * in that we only recover a specific set of data from them, namely
2701 * the inode di_next_unlinked fields. This is because all other inode
2702 * data is actually logged via inode records and any data we replay
2703 * here which overlaps that may be stale.
2704 *
2705 * When meta-data buffers are freed at run time we log a buffer item
2706 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2707 * of the buffer in the log should not be replayed at recovery time.
2708 * This is so that if the blocks covered by the buffer are reused for
2709 * file data before we crash we don't end up replaying old, freed
2710 * meta-data into a user's file.
2711 *
2712 * To handle the cancellation of buffer log items, we make two passes
2713 * over the log during recovery. During the first we build a table of
2714 * those buffers which have been cancelled, and during the second we
2715 * only replay those buffers which do not have corresponding cancel
2716 * records in the table. See xlog_recover_buffer_pass[1,2] above
2717 * for more details on the implementation of the table of cancel records.
2718 */
2719STATIC int
2720xlog_recover_buffer_pass2(
2721 struct xlog *log,
2722 struct list_head *buffer_list,
2723 struct xlog_recover_item *item,
2724 xfs_lsn_t current_lsn)
2725{
2726 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
2727 xfs_mount_t *mp = log->l_mp;
2728 xfs_buf_t *bp;
2729 int error;
2730 uint buf_flags;
2731 xfs_lsn_t lsn;
2732
2733 /*
2734 * In this pass we only want to recover all the buffers which have
2735 * not been cancelled and are not cancellation buffers themselves.
2736 */
2737 if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2738 buf_f->blf_len, buf_f->blf_flags)) {
2739 trace_xfs_log_recover_buf_cancel(log, buf_f);
2740 return 0;
2741 }
2742
2743 trace_xfs_log_recover_buf_recover(log, buf_f);
2744
2745 buf_flags = 0;
2746 if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
2747 buf_flags |= XBF_UNMAPPED;
2748
2749 bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2750 buf_flags, NULL);
2751 if (!bp)
2752 return -ENOMEM;
2753 error = bp->b_error;
2754 if (error) {
2755 xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#1)");
2756 goto out_release;
2757 }
2758
2759 /*
2760 * Recover the buffer only if we get an LSN from it and it's less than
2761 * the lsn of the transaction we are replaying.
2762 *
2763 * Note that we have to be extremely careful of readahead here.
2764 * Readahead does not attach verfiers to the buffers so if we don't
2765 * actually do any replay after readahead because of the LSN we found
2766 * in the buffer if more recent than that current transaction then we
2767 * need to attach the verifier directly. Failure to do so can lead to
2768 * future recovery actions (e.g. EFI and unlinked list recovery) can
2769 * operate on the buffers and they won't get the verifier attached. This
2770 * can lead to blocks on disk having the correct content but a stale
2771 * CRC.
2772 *
2773 * It is safe to assume these clean buffers are currently up to date.
2774 * If the buffer is dirtied by a later transaction being replayed, then
2775 * the verifier will be reset to match whatever recover turns that
2776 * buffer into.
2777 */
2778 lsn = xlog_recover_get_buf_lsn(mp, bp);
2779 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
2780 trace_xfs_log_recover_buf_skip(log, buf_f);
2781 xlog_recover_validate_buf_type(mp, bp, buf_f, NULLCOMMITLSN);
2782 goto out_release;
2783 }
2784
2785 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2786 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2787 if (error)
2788 goto out_release;
2789 } else if (buf_f->blf_flags &
2790 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2791 bool dirty;
2792
2793 dirty = xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2794 if (!dirty)
2795 goto out_release;
2796 } else {
2797 xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn);
2798 }
2799
2800 /*
2801 * Perform delayed write on the buffer. Asynchronous writes will be
2802 * slower when taking into account all the buffers to be flushed.
2803 *
2804 * Also make sure that only inode buffers with good sizes stay in
2805 * the buffer cache. The kernel moves inodes in buffers of 1 block
2806 * or inode_cluster_size bytes, whichever is bigger. The inode
2807 * buffers in the log can be a different size if the log was generated
2808 * by an older kernel using unclustered inode buffers or a newer kernel
2809 * running with a different inode cluster size. Regardless, if the
2810 * the inode buffer size isn't max(blocksize, inode_cluster_size)
2811 * for *our* value of inode_cluster_size, then we need to keep
2812 * the buffer out of the buffer cache so that the buffer won't
2813 * overlap with future reads of those inodes.
2814 */
2815 if (XFS_DINODE_MAGIC ==
2816 be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2817 (BBTOB(bp->b_length) != M_IGEO(log->l_mp)->inode_cluster_size)) {
2818 xfs_buf_stale(bp);
2819 error = xfs_bwrite(bp);
2820 } else {
2821 ASSERT(bp->b_mount == mp);
2822 bp->b_iodone = xlog_recover_iodone;
2823 xfs_buf_delwri_queue(bp, buffer_list);
2824 }
2825
2826out_release:
2827 xfs_buf_relse(bp);
2828 return error;
2829}
2830
2831/*
2832 * Inode fork owner changes
2833 *
2834 * If we have been told that we have to reparent the inode fork, it's because an
2835 * extent swap operation on a CRC enabled filesystem has been done and we are
2836 * replaying it. We need to walk the BMBT of the appropriate fork and change the
2837 * owners of it.
2838 *
2839 * The complexity here is that we don't have an inode context to work with, so
2840 * after we've replayed the inode we need to instantiate one. This is where the
2841 * fun begins.
2842 *
2843 * We are in the middle of log recovery, so we can't run transactions. That
2844 * means we cannot use cache coherent inode instantiation via xfs_iget(), as
2845 * that will result in the corresponding iput() running the inode through
2846 * xfs_inactive(). If we've just replayed an inode core that changes the link
2847 * count to zero (i.e. it's been unlinked), then xfs_inactive() will run
2848 * transactions (bad!).
2849 *
2850 * So, to avoid this, we instantiate an inode directly from the inode core we've
2851 * just recovered. We have the buffer still locked, and all we really need to
2852 * instantiate is the inode core and the forks being modified. We can do this
2853 * manually, then run the inode btree owner change, and then tear down the
2854 * xfs_inode without having to run any transactions at all.
2855 *
2856 * Also, because we don't have a transaction context available here but need to
2857 * gather all the buffers we modify for writeback so we pass the buffer_list
2858 * instead for the operation to use.
2859 */
2860
2861STATIC int
2862xfs_recover_inode_owner_change(
2863 struct xfs_mount *mp,
2864 struct xfs_dinode *dip,
2865 struct xfs_inode_log_format *in_f,
2866 struct list_head *buffer_list)
2867{
2868 struct xfs_inode *ip;
2869 int error;
2870
2871 ASSERT(in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER));
2872
2873 ip = xfs_inode_alloc(mp, in_f->ilf_ino);
2874 if (!ip)
2875 return -ENOMEM;
2876
2877 /* instantiate the inode */
2878 xfs_inode_from_disk(ip, dip);
2879 ASSERT(ip->i_d.di_version >= 3);
2880
2881 error = xfs_iformat_fork(ip, dip);
2882 if (error)
2883 goto out_free_ip;
2884
2885 if (!xfs_inode_verify_forks(ip)) {
2886 error = -EFSCORRUPTED;
2887 goto out_free_ip;
2888 }
2889
2890 if (in_f->ilf_fields & XFS_ILOG_DOWNER) {
2891 ASSERT(in_f->ilf_fields & XFS_ILOG_DBROOT);
2892 error = xfs_bmbt_change_owner(NULL, ip, XFS_DATA_FORK,
2893 ip->i_ino, buffer_list);
2894 if (error)
2895 goto out_free_ip;
2896 }
2897
2898 if (in_f->ilf_fields & XFS_ILOG_AOWNER) {
2899 ASSERT(in_f->ilf_fields & XFS_ILOG_ABROOT);
2900 error = xfs_bmbt_change_owner(NULL, ip, XFS_ATTR_FORK,
2901 ip->i_ino, buffer_list);
2902 if (error)
2903 goto out_free_ip;
2904 }
2905
2906out_free_ip:
2907 xfs_inode_free(ip);
2908 return error;
2909}
2910
2911STATIC int
2912xlog_recover_inode_pass2(
2913 struct xlog *log,
2914 struct list_head *buffer_list,
2915 struct xlog_recover_item *item,
2916 xfs_lsn_t current_lsn)
2917{
2918 struct xfs_inode_log_format *in_f;
2919 xfs_mount_t *mp = log->l_mp;
2920 xfs_buf_t *bp;
2921 xfs_dinode_t *dip;
2922 int len;
2923 char *src;
2924 char *dest;
2925 int error;
2926 int attr_index;
2927 uint fields;
2928 struct xfs_log_dinode *ldip;
2929 uint isize;
2930 int need_free = 0;
2931
2932 if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
2933 in_f = item->ri_buf[0].i_addr;
2934 } else {
2935 in_f = kmem_alloc(sizeof(struct xfs_inode_log_format), 0);
2936 need_free = 1;
2937 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2938 if (error)
2939 goto error;
2940 }
2941
2942 /*
2943 * Inode buffers can be freed, look out for it,
2944 * and do not replay the inode.
2945 */
2946 if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2947 in_f->ilf_len, 0)) {
2948 error = 0;
2949 trace_xfs_log_recover_inode_cancel(log, in_f);
2950 goto error;
2951 }
2952 trace_xfs_log_recover_inode_recover(log, in_f);
2953
2954 bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len, 0,
2955 &xfs_inode_buf_ops);
2956 if (!bp) {
2957 error = -ENOMEM;
2958 goto error;
2959 }
2960 error = bp->b_error;
2961 if (error) {
2962 xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#2)");
2963 goto out_release;
2964 }
2965 ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2966 dip = xfs_buf_offset(bp, in_f->ilf_boffset);
2967
2968 /*
2969 * Make sure the place we're flushing out to really looks
2970 * like an inode!
2971 */
2972 if (unlikely(!xfs_verify_magic16(bp, dip->di_magic))) {
2973 xfs_alert(mp,
2974 "%s: Bad inode magic number, dip = "PTR_FMT", dino bp = "PTR_FMT", ino = %Ld",
2975 __func__, dip, bp, in_f->ilf_ino);
2976 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2977 XFS_ERRLEVEL_LOW, mp);
2978 error = -EFSCORRUPTED;
2979 goto out_release;
2980 }
2981 ldip = item->ri_buf[1].i_addr;
2982 if (unlikely(ldip->di_magic != XFS_DINODE_MAGIC)) {
2983 xfs_alert(mp,
2984 "%s: Bad inode log record, rec ptr "PTR_FMT", ino %Ld",
2985 __func__, item, in_f->ilf_ino);
2986 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2987 XFS_ERRLEVEL_LOW, mp);
2988 error = -EFSCORRUPTED;
2989 goto out_release;
2990 }
2991
2992 /*
2993 * If the inode has an LSN in it, recover the inode only if it's less
2994 * than the lsn of the transaction we are replaying. Note: we still
2995 * need to replay an owner change even though the inode is more recent
2996 * than the transaction as there is no guarantee that all the btree
2997 * blocks are more recent than this transaction, too.
2998 */
2999 if (dip->di_version >= 3) {
3000 xfs_lsn_t lsn = be64_to_cpu(dip->di_lsn);
3001
3002 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3003 trace_xfs_log_recover_inode_skip(log, in_f);
3004 error = 0;
3005 goto out_owner_change;
3006 }
3007 }
3008
3009 /*
3010 * di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
3011 * are transactional and if ordering is necessary we can determine that
3012 * more accurately by the LSN field in the V3 inode core. Don't trust
3013 * the inode versions we might be changing them here - use the
3014 * superblock flag to determine whether we need to look at di_flushiter
3015 * to skip replay when the on disk inode is newer than the log one
3016 */
3017 if (!xfs_sb_version_hascrc(&mp->m_sb) &&
3018 ldip->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
3019 /*
3020 * Deal with the wrap case, DI_MAX_FLUSH is less
3021 * than smaller numbers
3022 */
3023 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
3024 ldip->di_flushiter < (DI_MAX_FLUSH >> 1)) {
3025 /* do nothing */
3026 } else {
3027 trace_xfs_log_recover_inode_skip(log, in_f);
3028 error = 0;
3029 goto out_release;
3030 }
3031 }
3032
3033 /* Take the opportunity to reset the flush iteration count */
3034 ldip->di_flushiter = 0;
3035
3036 if (unlikely(S_ISREG(ldip->di_mode))) {
3037 if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3038 (ldip->di_format != XFS_DINODE_FMT_BTREE)) {
3039 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
3040 XFS_ERRLEVEL_LOW, mp, ldip,
3041 sizeof(*ldip));
3042 xfs_alert(mp,
3043 "%s: Bad regular inode log record, rec ptr "PTR_FMT", "
3044 "ino ptr = "PTR_FMT", ino bp = "PTR_FMT", ino %Ld",
3045 __func__, item, dip, bp, in_f->ilf_ino);
3046 error = -EFSCORRUPTED;
3047 goto out_release;
3048 }
3049 } else if (unlikely(S_ISDIR(ldip->di_mode))) {
3050 if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3051 (ldip->di_format != XFS_DINODE_FMT_BTREE) &&
3052 (ldip->di_format != XFS_DINODE_FMT_LOCAL)) {
3053 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
3054 XFS_ERRLEVEL_LOW, mp, ldip,
3055 sizeof(*ldip));
3056 xfs_alert(mp,
3057 "%s: Bad dir inode log record, rec ptr "PTR_FMT", "
3058 "ino ptr = "PTR_FMT", ino bp = "PTR_FMT", ino %Ld",
3059 __func__, item, dip, bp, in_f->ilf_ino);
3060 error = -EFSCORRUPTED;
3061 goto out_release;
3062 }
3063 }
3064 if (unlikely(ldip->di_nextents + ldip->di_anextents > ldip->di_nblocks)){
3065 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
3066 XFS_ERRLEVEL_LOW, mp, ldip,
3067 sizeof(*ldip));
3068 xfs_alert(mp,
3069 "%s: Bad inode log record, rec ptr "PTR_FMT", dino ptr "PTR_FMT", "
3070 "dino bp "PTR_FMT", ino %Ld, total extents = %d, nblocks = %Ld",
3071 __func__, item, dip, bp, in_f->ilf_ino,
3072 ldip->di_nextents + ldip->di_anextents,
3073 ldip->di_nblocks);
3074 error = -EFSCORRUPTED;
3075 goto out_release;
3076 }
3077 if (unlikely(ldip->di_forkoff > mp->m_sb.sb_inodesize)) {
3078 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
3079 XFS_ERRLEVEL_LOW, mp, ldip,
3080 sizeof(*ldip));
3081 xfs_alert(mp,
3082 "%s: Bad inode log record, rec ptr "PTR_FMT", dino ptr "PTR_FMT", "
3083 "dino bp "PTR_FMT", ino %Ld, forkoff 0x%x", __func__,
3084 item, dip, bp, in_f->ilf_ino, ldip->di_forkoff);
3085 error = -EFSCORRUPTED;
3086 goto out_release;
3087 }
3088 isize = xfs_log_dinode_size(ldip->di_version);
3089 if (unlikely(item->ri_buf[1].i_len > isize)) {
3090 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
3091 XFS_ERRLEVEL_LOW, mp, ldip,
3092 sizeof(*ldip));
3093 xfs_alert(mp,
3094 "%s: Bad inode log record length %d, rec ptr "PTR_FMT,
3095 __func__, item->ri_buf[1].i_len, item);
3096 error = -EFSCORRUPTED;
3097 goto out_release;
3098 }
3099
3100 /* recover the log dinode inode into the on disk inode */
3101 xfs_log_dinode_to_disk(ldip, dip);
3102
3103 fields = in_f->ilf_fields;
3104 if (fields & XFS_ILOG_DEV)
3105 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
3106
3107 if (in_f->ilf_size == 2)
3108 goto out_owner_change;
3109 len = item->ri_buf[2].i_len;
3110 src = item->ri_buf[2].i_addr;
3111 ASSERT(in_f->ilf_size <= 4);
3112 ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
3113 ASSERT(!(fields & XFS_ILOG_DFORK) ||
3114 (len == in_f->ilf_dsize));
3115
3116 switch (fields & XFS_ILOG_DFORK) {
3117 case XFS_ILOG_DDATA:
3118 case XFS_ILOG_DEXT:
3119 memcpy(XFS_DFORK_DPTR(dip), src, len);
3120 break;
3121
3122 case XFS_ILOG_DBROOT:
3123 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
3124 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
3125 XFS_DFORK_DSIZE(dip, mp));
3126 break;
3127
3128 default:
3129 /*
3130 * There are no data fork flags set.
3131 */
3132 ASSERT((fields & XFS_ILOG_DFORK) == 0);
3133 break;
3134 }
3135
3136 /*
3137 * If we logged any attribute data, recover it. There may or
3138 * may not have been any other non-core data logged in this
3139 * transaction.
3140 */
3141 if (in_f->ilf_fields & XFS_ILOG_AFORK) {
3142 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
3143 attr_index = 3;
3144 } else {
3145 attr_index = 2;
3146 }
3147 len = item->ri_buf[attr_index].i_len;
3148 src = item->ri_buf[attr_index].i_addr;
3149 ASSERT(len == in_f->ilf_asize);
3150
3151 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
3152 case XFS_ILOG_ADATA:
3153 case XFS_ILOG_AEXT:
3154 dest = XFS_DFORK_APTR(dip);
3155 ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
3156 memcpy(dest, src, len);
3157 break;
3158
3159 case XFS_ILOG_ABROOT:
3160 dest = XFS_DFORK_APTR(dip);
3161 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
3162 len, (xfs_bmdr_block_t*)dest,
3163 XFS_DFORK_ASIZE(dip, mp));
3164 break;
3165
3166 default:
3167 xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
3168 ASSERT(0);
3169 error = -EIO;
3170 goto out_release;
3171 }
3172 }
3173
3174out_owner_change:
3175 /* Recover the swapext owner change unless inode has been deleted */
3176 if ((in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER)) &&
3177 (dip->di_mode != 0))
3178 error = xfs_recover_inode_owner_change(mp, dip, in_f,
3179 buffer_list);
3180 /* re-generate the checksum. */
3181 xfs_dinode_calc_crc(log->l_mp, dip);
3182
3183 ASSERT(bp->b_mount == mp);
3184 bp->b_iodone = xlog_recover_iodone;
3185 xfs_buf_delwri_queue(bp, buffer_list);
3186
3187out_release:
3188 xfs_buf_relse(bp);
3189error:
3190 if (need_free)
3191 kmem_free(in_f);
3192 return error;
3193}
3194
3195/*
3196 * Recover QUOTAOFF records. We simply make a note of it in the xlog
3197 * structure, so that we know not to do any dquot item or dquot buffer recovery,
3198 * of that type.
3199 */
3200STATIC int
3201xlog_recover_quotaoff_pass1(
3202 struct xlog *log,
3203 struct xlog_recover_item *item)
3204{
3205 xfs_qoff_logformat_t *qoff_f = item->ri_buf[0].i_addr;
3206 ASSERT(qoff_f);
3207
3208 /*
3209 * The logitem format's flag tells us if this was user quotaoff,
3210 * group/project quotaoff or both.
3211 */
3212 if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
3213 log->l_quotaoffs_flag |= XFS_DQ_USER;
3214 if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
3215 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
3216 if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
3217 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
3218
3219 return 0;
3220}
3221
3222/*
3223 * Recover a dquot record
3224 */
3225STATIC int
3226xlog_recover_dquot_pass2(
3227 struct xlog *log,
3228 struct list_head *buffer_list,
3229 struct xlog_recover_item *item,
3230 xfs_lsn_t current_lsn)
3231{
3232 xfs_mount_t *mp = log->l_mp;
3233 xfs_buf_t *bp;
3234 struct xfs_disk_dquot *ddq, *recddq;
3235 xfs_failaddr_t fa;
3236 int error;
3237 xfs_dq_logformat_t *dq_f;
3238 uint type;
3239
3240
3241 /*
3242 * Filesystems are required to send in quota flags at mount time.
3243 */
3244 if (mp->m_qflags == 0)
3245 return 0;
3246
3247 recddq = item->ri_buf[1].i_addr;
3248 if (recddq == NULL) {
3249 xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
3250 return -EIO;
3251 }
3252 if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
3253 xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
3254 item->ri_buf[1].i_len, __func__);
3255 return -EIO;
3256 }
3257
3258 /*
3259 * This type of quotas was turned off, so ignore this record.
3260 */
3261 type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3262 ASSERT(type);
3263 if (log->l_quotaoffs_flag & type)
3264 return 0;
3265
3266 /*
3267 * At this point we know that quota was _not_ turned off.
3268 * Since the mount flags are not indicating to us otherwise, this
3269 * must mean that quota is on, and the dquot needs to be replayed.
3270 * Remember that we may not have fully recovered the superblock yet,
3271 * so we can't do the usual trick of looking at the SB quota bits.
3272 *
3273 * The other possibility, of course, is that the quota subsystem was
3274 * removed since the last mount - ENOSYS.
3275 */
3276 dq_f = item->ri_buf[0].i_addr;
3277 ASSERT(dq_f);
3278 fa = xfs_dquot_verify(mp, recddq, dq_f->qlf_id, 0);
3279 if (fa) {
3280 xfs_alert(mp, "corrupt dquot ID 0x%x in log at %pS",
3281 dq_f->qlf_id, fa);
3282 return -EIO;
3283 }
3284 ASSERT(dq_f->qlf_len == 1);
3285
3286 /*
3287 * At this point we are assuming that the dquots have been allocated
3288 * and hence the buffer has valid dquots stamped in it. It should,
3289 * therefore, pass verifier validation. If the dquot is bad, then the
3290 * we'll return an error here, so we don't need to specifically check
3291 * the dquot in the buffer after the verifier has run.
3292 */
3293 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dq_f->qlf_blkno,
3294 XFS_FSB_TO_BB(mp, dq_f->qlf_len), 0, &bp,
3295 &xfs_dquot_buf_ops);
3296 if (error)
3297 return error;
3298
3299 ASSERT(bp);
3300 ddq = xfs_buf_offset(bp, dq_f->qlf_boffset);
3301
3302 /*
3303 * If the dquot has an LSN in it, recover the dquot only if it's less
3304 * than the lsn of the transaction we are replaying.
3305 */
3306 if (xfs_sb_version_hascrc(&mp->m_sb)) {
3307 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddq;
3308 xfs_lsn_t lsn = be64_to_cpu(dqb->dd_lsn);
3309
3310 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3311 goto out_release;
3312 }
3313 }
3314
3315 memcpy(ddq, recddq, item->ri_buf[1].i_len);
3316 if (xfs_sb_version_hascrc(&mp->m_sb)) {
3317 xfs_update_cksum((char *)ddq, sizeof(struct xfs_dqblk),
3318 XFS_DQUOT_CRC_OFF);
3319 }
3320
3321 ASSERT(dq_f->qlf_size == 2);
3322 ASSERT(bp->b_mount == mp);
3323 bp->b_iodone = xlog_recover_iodone;
3324 xfs_buf_delwri_queue(bp, buffer_list);
3325
3326out_release:
3327 xfs_buf_relse(bp);
3328 return 0;
3329}
3330
3331/*
3332 * This routine is called to create an in-core extent free intent
3333 * item from the efi format structure which was logged on disk.
3334 * It allocates an in-core efi, copies the extents from the format
3335 * structure into it, and adds the efi to the AIL with the given
3336 * LSN.
3337 */
3338STATIC int
3339xlog_recover_efi_pass2(
3340 struct xlog *log,
3341 struct xlog_recover_item *item,
3342 xfs_lsn_t lsn)
3343{
3344 int error;
3345 struct xfs_mount *mp = log->l_mp;
3346 struct xfs_efi_log_item *efip;
3347 struct xfs_efi_log_format *efi_formatp;
3348
3349 efi_formatp = item->ri_buf[0].i_addr;
3350
3351 efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
3352 error = xfs_efi_copy_format(&item->ri_buf[0], &efip->efi_format);
3353 if (error) {
3354 xfs_efi_item_free(efip);
3355 return error;
3356 }
3357 atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
3358
3359 spin_lock(&log->l_ailp->ail_lock);
3360 /*
3361 * The EFI has two references. One for the EFD and one for EFI to ensure
3362 * it makes it into the AIL. Insert the EFI into the AIL directly and
3363 * drop the EFI reference. Note that xfs_trans_ail_update() drops the
3364 * AIL lock.
3365 */
3366 xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
3367 xfs_efi_release(efip);
3368 return 0;
3369}
3370
3371
3372/*
3373 * This routine is called when an EFD format structure is found in a committed
3374 * transaction in the log. Its purpose is to cancel the corresponding EFI if it
3375 * was still in the log. To do this it searches the AIL for the EFI with an id
3376 * equal to that in the EFD format structure. If we find it we drop the EFD
3377 * reference, which removes the EFI from the AIL and frees it.
3378 */
3379STATIC int
3380xlog_recover_efd_pass2(
3381 struct xlog *log,
3382 struct xlog_recover_item *item)
3383{
3384 xfs_efd_log_format_t *efd_formatp;
3385 xfs_efi_log_item_t *efip = NULL;
3386 struct xfs_log_item *lip;
3387 uint64_t efi_id;
3388 struct xfs_ail_cursor cur;
3389 struct xfs_ail *ailp = log->l_ailp;
3390
3391 efd_formatp = item->ri_buf[0].i_addr;
3392 ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
3393 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
3394 (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
3395 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
3396 efi_id = efd_formatp->efd_efi_id;
3397
3398 /*
3399 * Search for the EFI with the id in the EFD format structure in the
3400 * AIL.
3401 */
3402 spin_lock(&ailp->ail_lock);
3403 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3404 while (lip != NULL) {
3405 if (lip->li_type == XFS_LI_EFI) {
3406 efip = (xfs_efi_log_item_t *)lip;
3407 if (efip->efi_format.efi_id == efi_id) {
3408 /*
3409 * Drop the EFD reference to the EFI. This
3410 * removes the EFI from the AIL and frees it.
3411 */
3412 spin_unlock(&ailp->ail_lock);
3413 xfs_efi_release(efip);
3414 spin_lock(&ailp->ail_lock);
3415 break;
3416 }
3417 }
3418 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3419 }
3420
3421 xfs_trans_ail_cursor_done(&cur);
3422 spin_unlock(&ailp->ail_lock);
3423
3424 return 0;
3425}
3426
3427/*
3428 * This routine is called to create an in-core extent rmap update
3429 * item from the rui format structure which was logged on disk.
3430 * It allocates an in-core rui, copies the extents from the format
3431 * structure into it, and adds the rui to the AIL with the given
3432 * LSN.
3433 */
3434STATIC int
3435xlog_recover_rui_pass2(
3436 struct xlog *log,
3437 struct xlog_recover_item *item,
3438 xfs_lsn_t lsn)
3439{
3440 int error;
3441 struct xfs_mount *mp = log->l_mp;
3442 struct xfs_rui_log_item *ruip;
3443 struct xfs_rui_log_format *rui_formatp;
3444
3445 rui_formatp = item->ri_buf[0].i_addr;
3446
3447 ruip = xfs_rui_init(mp, rui_formatp->rui_nextents);
3448 error = xfs_rui_copy_format(&item->ri_buf[0], &ruip->rui_format);
3449 if (error) {
3450 xfs_rui_item_free(ruip);
3451 return error;
3452 }
3453 atomic_set(&ruip->rui_next_extent, rui_formatp->rui_nextents);
3454
3455 spin_lock(&log->l_ailp->ail_lock);
3456 /*
3457 * The RUI has two references. One for the RUD and one for RUI to ensure
3458 * it makes it into the AIL. Insert the RUI into the AIL directly and
3459 * drop the RUI reference. Note that xfs_trans_ail_update() drops the
3460 * AIL lock.
3461 */
3462 xfs_trans_ail_update(log->l_ailp, &ruip->rui_item, lsn);
3463 xfs_rui_release(ruip);
3464 return 0;
3465}
3466
3467
3468/*
3469 * This routine is called when an RUD format structure is found in a committed
3470 * transaction in the log. Its purpose is to cancel the corresponding RUI if it
3471 * was still in the log. To do this it searches the AIL for the RUI with an id
3472 * equal to that in the RUD format structure. If we find it we drop the RUD
3473 * reference, which removes the RUI from the AIL and frees it.
3474 */
3475STATIC int
3476xlog_recover_rud_pass2(
3477 struct xlog *log,
3478 struct xlog_recover_item *item)
3479{
3480 struct xfs_rud_log_format *rud_formatp;
3481 struct xfs_rui_log_item *ruip = NULL;
3482 struct xfs_log_item *lip;
3483 uint64_t rui_id;
3484 struct xfs_ail_cursor cur;
3485 struct xfs_ail *ailp = log->l_ailp;
3486
3487 rud_formatp = item->ri_buf[0].i_addr;
3488 ASSERT(item->ri_buf[0].i_len == sizeof(struct xfs_rud_log_format));
3489 rui_id = rud_formatp->rud_rui_id;
3490
3491 /*
3492 * Search for the RUI with the id in the RUD format structure in the
3493 * AIL.
3494 */
3495 spin_lock(&ailp->ail_lock);
3496 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3497 while (lip != NULL) {
3498 if (lip->li_type == XFS_LI_RUI) {
3499 ruip = (struct xfs_rui_log_item *)lip;
3500 if (ruip->rui_format.rui_id == rui_id) {
3501 /*
3502 * Drop the RUD reference to the RUI. This
3503 * removes the RUI from the AIL and frees it.
3504 */
3505 spin_unlock(&ailp->ail_lock);
3506 xfs_rui_release(ruip);
3507 spin_lock(&ailp->ail_lock);
3508 break;
3509 }
3510 }
3511 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3512 }
3513
3514 xfs_trans_ail_cursor_done(&cur);
3515 spin_unlock(&ailp->ail_lock);
3516
3517 return 0;
3518}
3519
3520/*
3521 * Copy an CUI format buffer from the given buf, and into the destination
3522 * CUI format structure. The CUI/CUD items were designed not to need any
3523 * special alignment handling.
3524 */
3525static int
3526xfs_cui_copy_format(
3527 struct xfs_log_iovec *buf,
3528 struct xfs_cui_log_format *dst_cui_fmt)
3529{
3530 struct xfs_cui_log_format *src_cui_fmt;
3531 uint len;
3532
3533 src_cui_fmt = buf->i_addr;
3534 len = xfs_cui_log_format_sizeof(src_cui_fmt->cui_nextents);
3535
3536 if (buf->i_len == len) {
3537 memcpy(dst_cui_fmt, src_cui_fmt, len);
3538 return 0;
3539 }
3540 return -EFSCORRUPTED;
3541}
3542
3543/*
3544 * This routine is called to create an in-core extent refcount update
3545 * item from the cui format structure which was logged on disk.
3546 * It allocates an in-core cui, copies the extents from the format
3547 * structure into it, and adds the cui to the AIL with the given
3548 * LSN.
3549 */
3550STATIC int
3551xlog_recover_cui_pass2(
3552 struct xlog *log,
3553 struct xlog_recover_item *item,
3554 xfs_lsn_t lsn)
3555{
3556 int error;
3557 struct xfs_mount *mp = log->l_mp;
3558 struct xfs_cui_log_item *cuip;
3559 struct xfs_cui_log_format *cui_formatp;
3560
3561 cui_formatp = item->ri_buf[0].i_addr;
3562
3563 cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
3564 error = xfs_cui_copy_format(&item->ri_buf[0], &cuip->cui_format);
3565 if (error) {
3566 xfs_cui_item_free(cuip);
3567 return error;
3568 }
3569 atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
3570
3571 spin_lock(&log->l_ailp->ail_lock);
3572 /*
3573 * The CUI has two references. One for the CUD and one for CUI to ensure
3574 * it makes it into the AIL. Insert the CUI into the AIL directly and
3575 * drop the CUI reference. Note that xfs_trans_ail_update() drops the
3576 * AIL lock.
3577 */
3578 xfs_trans_ail_update(log->l_ailp, &cuip->cui_item, lsn);
3579 xfs_cui_release(cuip);
3580 return 0;
3581}
3582
3583
3584/*
3585 * This routine is called when an CUD format structure is found in a committed
3586 * transaction in the log. Its purpose is to cancel the corresponding CUI if it
3587 * was still in the log. To do this it searches the AIL for the CUI with an id
3588 * equal to that in the CUD format structure. If we find it we drop the CUD
3589 * reference, which removes the CUI from the AIL and frees it.
3590 */
3591STATIC int
3592xlog_recover_cud_pass2(
3593 struct xlog *log,
3594 struct xlog_recover_item *item)
3595{
3596 struct xfs_cud_log_format *cud_formatp;
3597 struct xfs_cui_log_item *cuip = NULL;
3598 struct xfs_log_item *lip;
3599 uint64_t cui_id;
3600 struct xfs_ail_cursor cur;
3601 struct xfs_ail *ailp = log->l_ailp;
3602
3603 cud_formatp = item->ri_buf[0].i_addr;
3604 if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format))
3605 return -EFSCORRUPTED;
3606 cui_id = cud_formatp->cud_cui_id;
3607
3608 /*
3609 * Search for the CUI with the id in the CUD format structure in the
3610 * AIL.
3611 */
3612 spin_lock(&ailp->ail_lock);
3613 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3614 while (lip != NULL) {
3615 if (lip->li_type == XFS_LI_CUI) {
3616 cuip = (struct xfs_cui_log_item *)lip;
3617 if (cuip->cui_format.cui_id == cui_id) {
3618 /*
3619 * Drop the CUD reference to the CUI. This
3620 * removes the CUI from the AIL and frees it.
3621 */
3622 spin_unlock(&ailp->ail_lock);
3623 xfs_cui_release(cuip);
3624 spin_lock(&ailp->ail_lock);
3625 break;
3626 }
3627 }
3628 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3629 }
3630
3631 xfs_trans_ail_cursor_done(&cur);
3632 spin_unlock(&ailp->ail_lock);
3633
3634 return 0;
3635}
3636
3637/*
3638 * Copy an BUI format buffer from the given buf, and into the destination
3639 * BUI format structure. The BUI/BUD items were designed not to need any
3640 * special alignment handling.
3641 */
3642static int
3643xfs_bui_copy_format(
3644 struct xfs_log_iovec *buf,
3645 struct xfs_bui_log_format *dst_bui_fmt)
3646{
3647 struct xfs_bui_log_format *src_bui_fmt;
3648 uint len;
3649
3650 src_bui_fmt = buf->i_addr;
3651 len = xfs_bui_log_format_sizeof(src_bui_fmt->bui_nextents);
3652
3653 if (buf->i_len == len) {
3654 memcpy(dst_bui_fmt, src_bui_fmt, len);
3655 return 0;
3656 }
3657 return -EFSCORRUPTED;
3658}
3659
3660/*
3661 * This routine is called to create an in-core extent bmap update
3662 * item from the bui format structure which was logged on disk.
3663 * It allocates an in-core bui, copies the extents from the format
3664 * structure into it, and adds the bui to the AIL with the given
3665 * LSN.
3666 */
3667STATIC int
3668xlog_recover_bui_pass2(
3669 struct xlog *log,
3670 struct xlog_recover_item *item,
3671 xfs_lsn_t lsn)
3672{
3673 int error;
3674 struct xfs_mount *mp = log->l_mp;
3675 struct xfs_bui_log_item *buip;
3676 struct xfs_bui_log_format *bui_formatp;
3677
3678 bui_formatp = item->ri_buf[0].i_addr;
3679
3680 if (bui_formatp->bui_nextents != XFS_BUI_MAX_FAST_EXTENTS)
3681 return -EFSCORRUPTED;
3682 buip = xfs_bui_init(mp);
3683 error = xfs_bui_copy_format(&item->ri_buf[0], &buip->bui_format);
3684 if (error) {
3685 xfs_bui_item_free(buip);
3686 return error;
3687 }
3688 atomic_set(&buip->bui_next_extent, bui_formatp->bui_nextents);
3689
3690 spin_lock(&log->l_ailp->ail_lock);
3691 /*
3692 * The RUI has two references. One for the RUD and one for RUI to ensure
3693 * it makes it into the AIL. Insert the RUI into the AIL directly and
3694 * drop the RUI reference. Note that xfs_trans_ail_update() drops the
3695 * AIL lock.
3696 */
3697 xfs_trans_ail_update(log->l_ailp, &buip->bui_item, lsn);
3698 xfs_bui_release(buip);
3699 return 0;
3700}
3701
3702
3703/*
3704 * This routine is called when an BUD format structure is found in a committed
3705 * transaction in the log. Its purpose is to cancel the corresponding BUI if it
3706 * was still in the log. To do this it searches the AIL for the BUI with an id
3707 * equal to that in the BUD format structure. If we find it we drop the BUD
3708 * reference, which removes the BUI from the AIL and frees it.
3709 */
3710STATIC int
3711xlog_recover_bud_pass2(
3712 struct xlog *log,
3713 struct xlog_recover_item *item)
3714{
3715 struct xfs_bud_log_format *bud_formatp;
3716 struct xfs_bui_log_item *buip = NULL;
3717 struct xfs_log_item *lip;
3718 uint64_t bui_id;
3719 struct xfs_ail_cursor cur;
3720 struct xfs_ail *ailp = log->l_ailp;
3721
3722 bud_formatp = item->ri_buf[0].i_addr;
3723 if (item->ri_buf[0].i_len != sizeof(struct xfs_bud_log_format))
3724 return -EFSCORRUPTED;
3725 bui_id = bud_formatp->bud_bui_id;
3726
3727 /*
3728 * Search for the BUI with the id in the BUD format structure in the
3729 * AIL.
3730 */
3731 spin_lock(&ailp->ail_lock);
3732 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3733 while (lip != NULL) {
3734 if (lip->li_type == XFS_LI_BUI) {
3735 buip = (struct xfs_bui_log_item *)lip;
3736 if (buip->bui_format.bui_id == bui_id) {
3737 /*
3738 * Drop the BUD reference to the BUI. This
3739 * removes the BUI from the AIL and frees it.
3740 */
3741 spin_unlock(&ailp->ail_lock);
3742 xfs_bui_release(buip);
3743 spin_lock(&ailp->ail_lock);
3744 break;
3745 }
3746 }
3747 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3748 }
3749
3750 xfs_trans_ail_cursor_done(&cur);
3751 spin_unlock(&ailp->ail_lock);
3752
3753 return 0;
3754}
3755
3756/*
3757 * This routine is called when an inode create format structure is found in a
3758 * committed transaction in the log. It's purpose is to initialise the inodes
3759 * being allocated on disk. This requires us to get inode cluster buffers that
3760 * match the range to be initialised, stamped with inode templates and written
3761 * by delayed write so that subsequent modifications will hit the cached buffer
3762 * and only need writing out at the end of recovery.
3763 */
3764STATIC int
3765xlog_recover_do_icreate_pass2(
3766 struct xlog *log,
3767 struct list_head *buffer_list,
3768 xlog_recover_item_t *item)
3769{
3770 struct xfs_mount *mp = log->l_mp;
3771 struct xfs_icreate_log *icl;
3772 struct xfs_ino_geometry *igeo = M_IGEO(mp);
3773 xfs_agnumber_t agno;
3774 xfs_agblock_t agbno;
3775 unsigned int count;
3776 unsigned int isize;
3777 xfs_agblock_t length;
3778 int bb_per_cluster;
3779 int cancel_count;
3780 int nbufs;
3781 int i;
3782
3783 icl = (struct xfs_icreate_log *)item->ri_buf[0].i_addr;
3784 if (icl->icl_type != XFS_LI_ICREATE) {
3785 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad type");
3786 return -EINVAL;
3787 }
3788
3789 if (icl->icl_size != 1) {
3790 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad icl size");
3791 return -EINVAL;
3792 }
3793
3794 agno = be32_to_cpu(icl->icl_ag);
3795 if (agno >= mp->m_sb.sb_agcount) {
3796 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agno");
3797 return -EINVAL;
3798 }
3799 agbno = be32_to_cpu(icl->icl_agbno);
3800 if (!agbno || agbno == NULLAGBLOCK || agbno >= mp->m_sb.sb_agblocks) {
3801 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agbno");
3802 return -EINVAL;
3803 }
3804 isize = be32_to_cpu(icl->icl_isize);
3805 if (isize != mp->m_sb.sb_inodesize) {
3806 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad isize");
3807 return -EINVAL;
3808 }
3809 count = be32_to_cpu(icl->icl_count);
3810 if (!count) {
3811 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count");
3812 return -EINVAL;
3813 }
3814 length = be32_to_cpu(icl->icl_length);
3815 if (!length || length >= mp->m_sb.sb_agblocks) {
3816 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad length");
3817 return -EINVAL;
3818 }
3819
3820 /*
3821 * The inode chunk is either full or sparse and we only support
3822 * m_ino_geo.ialloc_min_blks sized sparse allocations at this time.
3823 */
3824 if (length != igeo->ialloc_blks &&
3825 length != igeo->ialloc_min_blks) {
3826 xfs_warn(log->l_mp,
3827 "%s: unsupported chunk length", __FUNCTION__);
3828 return -EINVAL;
3829 }
3830
3831 /* verify inode count is consistent with extent length */
3832 if ((count >> mp->m_sb.sb_inopblog) != length) {
3833 xfs_warn(log->l_mp,
3834 "%s: inconsistent inode count and chunk length",
3835 __FUNCTION__);
3836 return -EINVAL;
3837 }
3838
3839 /*
3840 * The icreate transaction can cover multiple cluster buffers and these
3841 * buffers could have been freed and reused. Check the individual
3842 * buffers for cancellation so we don't overwrite anything written after
3843 * a cancellation.
3844 */
3845 bb_per_cluster = XFS_FSB_TO_BB(mp, igeo->blocks_per_cluster);
3846 nbufs = length / igeo->blocks_per_cluster;
3847 for (i = 0, cancel_count = 0; i < nbufs; i++) {
3848 xfs_daddr_t daddr;
3849
3850 daddr = XFS_AGB_TO_DADDR(mp, agno,
3851 agbno + i * igeo->blocks_per_cluster);
3852 if (xlog_check_buffer_cancelled(log, daddr, bb_per_cluster, 0))
3853 cancel_count++;
3854 }
3855
3856 /*
3857 * We currently only use icreate for a single allocation at a time. This
3858 * means we should expect either all or none of the buffers to be
3859 * cancelled. Be conservative and skip replay if at least one buffer is
3860 * cancelled, but warn the user that something is awry if the buffers
3861 * are not consistent.
3862 *
3863 * XXX: This must be refined to only skip cancelled clusters once we use
3864 * icreate for multiple chunk allocations.
3865 */
3866 ASSERT(!cancel_count || cancel_count == nbufs);
3867 if (cancel_count) {
3868 if (cancel_count != nbufs)
3869 xfs_warn(mp,
3870 "WARNING: partial inode chunk cancellation, skipped icreate.");
3871 trace_xfs_log_recover_icreate_cancel(log, icl);
3872 return 0;
3873 }
3874
3875 trace_xfs_log_recover_icreate_recover(log, icl);
3876 return xfs_ialloc_inode_init(mp, NULL, buffer_list, count, agno, agbno,
3877 length, be32_to_cpu(icl->icl_gen));
3878}
3879
3880STATIC void
3881xlog_recover_buffer_ra_pass2(
3882 struct xlog *log,
3883 struct xlog_recover_item *item)
3884{
3885 struct xfs_buf_log_format *buf_f = item->ri_buf[0].i_addr;
3886 struct xfs_mount *mp = log->l_mp;
3887
3888 if (xlog_peek_buffer_cancelled(log, buf_f->blf_blkno,
3889 buf_f->blf_len, buf_f->blf_flags)) {
3890 return;
3891 }
3892
3893 xfs_buf_readahead(mp->m_ddev_targp, buf_f->blf_blkno,
3894 buf_f->blf_len, NULL);
3895}
3896
3897STATIC void
3898xlog_recover_inode_ra_pass2(
3899 struct xlog *log,
3900 struct xlog_recover_item *item)
3901{
3902 struct xfs_inode_log_format ilf_buf;
3903 struct xfs_inode_log_format *ilfp;
3904 struct xfs_mount *mp = log->l_mp;
3905 int error;
3906
3907 if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
3908 ilfp = item->ri_buf[0].i_addr;
3909 } else {
3910 ilfp = &ilf_buf;
3911 memset(ilfp, 0, sizeof(*ilfp));
3912 error = xfs_inode_item_format_convert(&item->ri_buf[0], ilfp);
3913 if (error)
3914 return;
3915 }
3916
3917 if (xlog_peek_buffer_cancelled(log, ilfp->ilf_blkno, ilfp->ilf_len, 0))
3918 return;
3919
3920 xfs_buf_readahead(mp->m_ddev_targp, ilfp->ilf_blkno,
3921 ilfp->ilf_len, &xfs_inode_buf_ra_ops);
3922}
3923
3924STATIC void
3925xlog_recover_dquot_ra_pass2(
3926 struct xlog *log,
3927 struct xlog_recover_item *item)
3928{
3929 struct xfs_mount *mp = log->l_mp;
3930 struct xfs_disk_dquot *recddq;
3931 struct xfs_dq_logformat *dq_f;
3932 uint type;
3933 int len;
3934
3935
3936 if (mp->m_qflags == 0)
3937 return;
3938
3939 recddq = item->ri_buf[1].i_addr;
3940 if (recddq == NULL)
3941 return;
3942 if (item->ri_buf[1].i_len < sizeof(struct xfs_disk_dquot))
3943 return;
3944
3945 type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3946 ASSERT(type);
3947 if (log->l_quotaoffs_flag & type)
3948 return;
3949
3950 dq_f = item->ri_buf[0].i_addr;
3951 ASSERT(dq_f);
3952 ASSERT(dq_f->qlf_len == 1);
3953
3954 len = XFS_FSB_TO_BB(mp, dq_f->qlf_len);
3955 if (xlog_peek_buffer_cancelled(log, dq_f->qlf_blkno, len, 0))
3956 return;
3957
3958 xfs_buf_readahead(mp->m_ddev_targp, dq_f->qlf_blkno, len,
3959 &xfs_dquot_buf_ra_ops);
3960}
3961
3962STATIC void
3963xlog_recover_ra_pass2(
3964 struct xlog *log,
3965 struct xlog_recover_item *item)
3966{
3967 switch (ITEM_TYPE(item)) {
3968 case XFS_LI_BUF:
3969 xlog_recover_buffer_ra_pass2(log, item);
3970 break;
3971 case XFS_LI_INODE:
3972 xlog_recover_inode_ra_pass2(log, item);
3973 break;
3974 case XFS_LI_DQUOT:
3975 xlog_recover_dquot_ra_pass2(log, item);
3976 break;
3977 case XFS_LI_EFI:
3978 case XFS_LI_EFD:
3979 case XFS_LI_QUOTAOFF:
3980 case XFS_LI_RUI:
3981 case XFS_LI_RUD:
3982 case XFS_LI_CUI:
3983 case XFS_LI_CUD:
3984 case XFS_LI_BUI:
3985 case XFS_LI_BUD:
3986 default:
3987 break;
3988 }
3989}
3990
3991STATIC int
3992xlog_recover_commit_pass1(
3993 struct xlog *log,
3994 struct xlog_recover *trans,
3995 struct xlog_recover_item *item)
3996{
3997 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
3998
3999 switch (ITEM_TYPE(item)) {
4000 case XFS_LI_BUF:
4001 return xlog_recover_buffer_pass1(log, item);
4002 case XFS_LI_QUOTAOFF:
4003 return xlog_recover_quotaoff_pass1(log, item);
4004 case XFS_LI_INODE:
4005 case XFS_LI_EFI:
4006 case XFS_LI_EFD:
4007 case XFS_LI_DQUOT:
4008 case XFS_LI_ICREATE:
4009 case XFS_LI_RUI:
4010 case XFS_LI_RUD:
4011 case XFS_LI_CUI:
4012 case XFS_LI_CUD:
4013 case XFS_LI_BUI:
4014 case XFS_LI_BUD:
4015 /* nothing to do in pass 1 */
4016 return 0;
4017 default:
4018 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
4019 __func__, ITEM_TYPE(item));
4020 ASSERT(0);
4021 return -EIO;
4022 }
4023}
4024
4025STATIC int
4026xlog_recover_commit_pass2(
4027 struct xlog *log,
4028 struct xlog_recover *trans,
4029 struct list_head *buffer_list,
4030 struct xlog_recover_item *item)
4031{
4032 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
4033
4034 switch (ITEM_TYPE(item)) {
4035 case XFS_LI_BUF:
4036 return xlog_recover_buffer_pass2(log, buffer_list, item,
4037 trans->r_lsn);
4038 case XFS_LI_INODE:
4039 return xlog_recover_inode_pass2(log, buffer_list, item,
4040 trans->r_lsn);
4041 case XFS_LI_EFI:
4042 return xlog_recover_efi_pass2(log, item, trans->r_lsn);
4043 case XFS_LI_EFD:
4044 return xlog_recover_efd_pass2(log, item);
4045 case XFS_LI_RUI:
4046 return xlog_recover_rui_pass2(log, item, trans->r_lsn);
4047 case XFS_LI_RUD:
4048 return xlog_recover_rud_pass2(log, item);
4049 case XFS_LI_CUI:
4050 return xlog_recover_cui_pass2(log, item, trans->r_lsn);
4051 case XFS_LI_CUD:
4052 return xlog_recover_cud_pass2(log, item);
4053 case XFS_LI_BUI:
4054 return xlog_recover_bui_pass2(log, item, trans->r_lsn);
4055 case XFS_LI_BUD:
4056 return xlog_recover_bud_pass2(log, item);
4057 case XFS_LI_DQUOT:
4058 return xlog_recover_dquot_pass2(log, buffer_list, item,
4059 trans->r_lsn);
4060 case XFS_LI_ICREATE:
4061 return xlog_recover_do_icreate_pass2(log, buffer_list, item);
4062 case XFS_LI_QUOTAOFF:
4063 /* nothing to do in pass2 */
4064 return 0;
4065 default:
4066 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
4067 __func__, ITEM_TYPE(item));
4068 ASSERT(0);
4069 return -EIO;
4070 }
4071}
4072
4073STATIC int
4074xlog_recover_items_pass2(
4075 struct xlog *log,
4076 struct xlog_recover *trans,
4077 struct list_head *buffer_list,
4078 struct list_head *item_list)
4079{
4080 struct xlog_recover_item *item;
4081 int error = 0;
4082
4083 list_for_each_entry(item, item_list, ri_list) {
4084 error = xlog_recover_commit_pass2(log, trans,
4085 buffer_list, item);
4086 if (error)
4087 return error;
4088 }
4089
4090 return error;
4091}
4092
4093/*
4094 * Perform the transaction.
4095 *
4096 * If the transaction modifies a buffer or inode, do it now. Otherwise,
4097 * EFIs and EFDs get queued up by adding entries into the AIL for them.
4098 */
4099STATIC int
4100xlog_recover_commit_trans(
4101 struct xlog *log,
4102 struct xlog_recover *trans,
4103 int pass,
4104 struct list_head *buffer_list)
4105{
4106 int error = 0;
4107 int items_queued = 0;
4108 struct xlog_recover_item *item;
4109 struct xlog_recover_item *next;
4110 LIST_HEAD (ra_list);
4111 LIST_HEAD (done_list);
4112
4113 #define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
4114
4115 hlist_del_init(&trans->r_list);
4116
4117 error = xlog_recover_reorder_trans(log, trans, pass);
4118 if (error)
4119 return error;
4120
4121 list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
4122 switch (pass) {
4123 case XLOG_RECOVER_PASS1:
4124 error = xlog_recover_commit_pass1(log, trans, item);
4125 break;
4126 case XLOG_RECOVER_PASS2:
4127 xlog_recover_ra_pass2(log, item);
4128 list_move_tail(&item->ri_list, &ra_list);
4129 items_queued++;
4130 if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
4131 error = xlog_recover_items_pass2(log, trans,
4132 buffer_list, &ra_list);
4133 list_splice_tail_init(&ra_list, &done_list);
4134 items_queued = 0;
4135 }
4136
4137 break;
4138 default:
4139 ASSERT(0);
4140 }
4141
4142 if (error)
4143 goto out;
4144 }
4145
4146out:
4147 if (!list_empty(&ra_list)) {
4148 if (!error)
4149 error = xlog_recover_items_pass2(log, trans,
4150 buffer_list, &ra_list);
4151 list_splice_tail_init(&ra_list, &done_list);
4152 }
4153
4154 if (!list_empty(&done_list))
4155 list_splice_init(&done_list, &trans->r_itemq);
4156
4157 return error;
4158}
4159
4160STATIC void
4161xlog_recover_add_item(
4162 struct list_head *head)
4163{
4164 xlog_recover_item_t *item;
4165
4166 item = kmem_zalloc(sizeof(xlog_recover_item_t), 0);
4167 INIT_LIST_HEAD(&item->ri_list);
4168 list_add_tail(&item->ri_list, head);
4169}
4170
4171STATIC int
4172xlog_recover_add_to_cont_trans(
4173 struct xlog *log,
4174 struct xlog_recover *trans,
4175 char *dp,
4176 int len)
4177{
4178 xlog_recover_item_t *item;
4179 char *ptr, *old_ptr;
4180 int old_len;
4181
4182 /*
4183 * If the transaction is empty, the header was split across this and the
4184 * previous record. Copy the rest of the header.
4185 */
4186 if (list_empty(&trans->r_itemq)) {
4187 ASSERT(len <= sizeof(struct xfs_trans_header));
4188 if (len > sizeof(struct xfs_trans_header)) {
4189 xfs_warn(log->l_mp, "%s: bad header length", __func__);
4190 return -EIO;
4191 }
4192
4193 xlog_recover_add_item(&trans->r_itemq);
4194 ptr = (char *)&trans->r_theader +
4195 sizeof(struct xfs_trans_header) - len;
4196 memcpy(ptr, dp, len);
4197 return 0;
4198 }
4199
4200 /* take the tail entry */
4201 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
4202
4203 old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
4204 old_len = item->ri_buf[item->ri_cnt-1].i_len;
4205
4206 ptr = kmem_realloc(old_ptr, len + old_len, 0);
4207 memcpy(&ptr[old_len], dp, len);
4208 item->ri_buf[item->ri_cnt-1].i_len += len;
4209 item->ri_buf[item->ri_cnt-1].i_addr = ptr;
4210 trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
4211 return 0;
4212}
4213
4214/*
4215 * The next region to add is the start of a new region. It could be
4216 * a whole region or it could be the first part of a new region. Because
4217 * of this, the assumption here is that the type and size fields of all
4218 * format structures fit into the first 32 bits of the structure.
4219 *
4220 * This works because all regions must be 32 bit aligned. Therefore, we
4221 * either have both fields or we have neither field. In the case we have
4222 * neither field, the data part of the region is zero length. We only have
4223 * a log_op_header and can throw away the header since a new one will appear
4224 * later. If we have at least 4 bytes, then we can determine how many regions
4225 * will appear in the current log item.
4226 */
4227STATIC int
4228xlog_recover_add_to_trans(
4229 struct xlog *log,
4230 struct xlog_recover *trans,
4231 char *dp,
4232 int len)
4233{
4234 struct xfs_inode_log_format *in_f; /* any will do */
4235 xlog_recover_item_t *item;
4236 char *ptr;
4237
4238 if (!len)
4239 return 0;
4240 if (list_empty(&trans->r_itemq)) {
4241 /* we need to catch log corruptions here */
4242 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
4243 xfs_warn(log->l_mp, "%s: bad header magic number",
4244 __func__);
4245 ASSERT(0);
4246 return -EIO;
4247 }
4248
4249 if (len > sizeof(struct xfs_trans_header)) {
4250 xfs_warn(log->l_mp, "%s: bad header length", __func__);
4251 ASSERT(0);
4252 return -EIO;
4253 }
4254
4255 /*
4256 * The transaction header can be arbitrarily split across op
4257 * records. If we don't have the whole thing here, copy what we
4258 * do have and handle the rest in the next record.
4259 */
4260 if (len == sizeof(struct xfs_trans_header))
4261 xlog_recover_add_item(&trans->r_itemq);
4262 memcpy(&trans->r_theader, dp, len);
4263 return 0;
4264 }
4265
4266 ptr = kmem_alloc(len, 0);
4267 memcpy(ptr, dp, len);
4268 in_f = (struct xfs_inode_log_format *)ptr;
4269
4270 /* take the tail entry */
4271 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
4272 if (item->ri_total != 0 &&
4273 item->ri_total == item->ri_cnt) {
4274 /* tail item is in use, get a new one */
4275 xlog_recover_add_item(&trans->r_itemq);
4276 item = list_entry(trans->r_itemq.prev,
4277 xlog_recover_item_t, ri_list);
4278 }
4279
4280 if (item->ri_total == 0) { /* first region to be added */
4281 if (in_f->ilf_size == 0 ||
4282 in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
4283 xfs_warn(log->l_mp,
4284 "bad number of regions (%d) in inode log format",
4285 in_f->ilf_size);
4286 ASSERT(0);
4287 kmem_free(ptr);
4288 return -EIO;
4289 }
4290
4291 item->ri_total = in_f->ilf_size;
4292 item->ri_buf =
4293 kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
4294 0);
4295 }
4296 ASSERT(item->ri_total > item->ri_cnt);
4297 /* Description region is ri_buf[0] */
4298 item->ri_buf[item->ri_cnt].i_addr = ptr;
4299 item->ri_buf[item->ri_cnt].i_len = len;
4300 item->ri_cnt++;
4301 trace_xfs_log_recover_item_add(log, trans, item, 0);
4302 return 0;
4303}
4304
4305/*
4306 * Free up any resources allocated by the transaction
4307 *
4308 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
4309 */
4310STATIC void
4311xlog_recover_free_trans(
4312 struct xlog_recover *trans)
4313{
4314 xlog_recover_item_t *item, *n;
4315 int i;
4316
4317 hlist_del_init(&trans->r_list);
4318
4319 list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
4320 /* Free the regions in the item. */
4321 list_del(&item->ri_list);
4322 for (i = 0; i < item->ri_cnt; i++)
4323 kmem_free(item->ri_buf[i].i_addr);
4324 /* Free the item itself */
4325 kmem_free(item->ri_buf);
4326 kmem_free(item);
4327 }
4328 /* Free the transaction recover structure */
4329 kmem_free(trans);
4330}
4331
4332/*
4333 * On error or completion, trans is freed.
4334 */
4335STATIC int
4336xlog_recovery_process_trans(
4337 struct xlog *log,
4338 struct xlog_recover *trans,
4339 char *dp,
4340 unsigned int len,
4341 unsigned int flags,
4342 int pass,
4343 struct list_head *buffer_list)
4344{
4345 int error = 0;
4346 bool freeit = false;
4347
4348 /* mask off ophdr transaction container flags */
4349 flags &= ~XLOG_END_TRANS;
4350 if (flags & XLOG_WAS_CONT_TRANS)
4351 flags &= ~XLOG_CONTINUE_TRANS;
4352
4353 /*
4354 * Callees must not free the trans structure. We'll decide if we need to
4355 * free it or not based on the operation being done and it's result.
4356 */
4357 switch (flags) {
4358 /* expected flag values */
4359 case 0:
4360 case XLOG_CONTINUE_TRANS:
4361 error = xlog_recover_add_to_trans(log, trans, dp, len);
4362 break;
4363 case XLOG_WAS_CONT_TRANS:
4364 error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
4365 break;
4366 case XLOG_COMMIT_TRANS:
4367 error = xlog_recover_commit_trans(log, trans, pass,
4368 buffer_list);
4369 /* success or fail, we are now done with this transaction. */
4370 freeit = true;
4371 break;
4372
4373 /* unexpected flag values */
4374 case XLOG_UNMOUNT_TRANS:
4375 /* just skip trans */
4376 xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
4377 freeit = true;
4378 break;
4379 case XLOG_START_TRANS:
4380 default:
4381 xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
4382 ASSERT(0);
4383 error = -EIO;
4384 break;
4385 }
4386 if (error || freeit)
4387 xlog_recover_free_trans(trans);
4388 return error;
4389}
4390
4391/*
4392 * Lookup the transaction recovery structure associated with the ID in the
4393 * current ophdr. If the transaction doesn't exist and the start flag is set in
4394 * the ophdr, then allocate a new transaction for future ID matches to find.
4395 * Either way, return what we found during the lookup - an existing transaction
4396 * or nothing.
4397 */
4398STATIC struct xlog_recover *
4399xlog_recover_ophdr_to_trans(
4400 struct hlist_head rhash[],
4401 struct xlog_rec_header *rhead,
4402 struct xlog_op_header *ohead)
4403{
4404 struct xlog_recover *trans;
4405 xlog_tid_t tid;
4406 struct hlist_head *rhp;
4407
4408 tid = be32_to_cpu(ohead->oh_tid);
4409 rhp = &rhash[XLOG_RHASH(tid)];
4410 hlist_for_each_entry(trans, rhp, r_list) {
4411 if (trans->r_log_tid == tid)
4412 return trans;
4413 }
4414
4415 /*
4416 * skip over non-start transaction headers - we could be
4417 * processing slack space before the next transaction starts
4418 */
4419 if (!(ohead->oh_flags & XLOG_START_TRANS))
4420 return NULL;
4421
4422 ASSERT(be32_to_cpu(ohead->oh_len) == 0);
4423
4424 /*
4425 * This is a new transaction so allocate a new recovery container to
4426 * hold the recovery ops that will follow.
4427 */
4428 trans = kmem_zalloc(sizeof(struct xlog_recover), 0);
4429 trans->r_log_tid = tid;
4430 trans->r_lsn = be64_to_cpu(rhead->h_lsn);
4431 INIT_LIST_HEAD(&trans->r_itemq);
4432 INIT_HLIST_NODE(&trans->r_list);
4433 hlist_add_head(&trans->r_list, rhp);
4434
4435 /*
4436 * Nothing more to do for this ophdr. Items to be added to this new
4437 * transaction will be in subsequent ophdr containers.
4438 */
4439 return NULL;
4440}
4441
4442STATIC int
4443xlog_recover_process_ophdr(
4444 struct xlog *log,
4445 struct hlist_head rhash[],
4446 struct xlog_rec_header *rhead,
4447 struct xlog_op_header *ohead,
4448 char *dp,
4449 char *end,
4450 int pass,
4451 struct list_head *buffer_list)
4452{
4453 struct xlog_recover *trans;
4454 unsigned int len;
4455 int error;
4456
4457 /* Do we understand who wrote this op? */
4458 if (ohead->oh_clientid != XFS_TRANSACTION &&
4459 ohead->oh_clientid != XFS_LOG) {
4460 xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
4461 __func__, ohead->oh_clientid);
4462 ASSERT(0);
4463 return -EIO;
4464 }
4465
4466 /*
4467 * Check the ophdr contains all the data it is supposed to contain.
4468 */
4469 len = be32_to_cpu(ohead->oh_len);
4470 if (dp + len > end) {
4471 xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
4472 WARN_ON(1);
4473 return -EIO;
4474 }
4475
4476 trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
4477 if (!trans) {
4478 /* nothing to do, so skip over this ophdr */
4479 return 0;
4480 }
4481
4482 /*
4483 * The recovered buffer queue is drained only once we know that all
4484 * recovery items for the current LSN have been processed. This is
4485 * required because:
4486 *
4487 * - Buffer write submission updates the metadata LSN of the buffer.
4488 * - Log recovery skips items with a metadata LSN >= the current LSN of
4489 * the recovery item.
4490 * - Separate recovery items against the same metadata buffer can share
4491 * a current LSN. I.e., consider that the LSN of a recovery item is
4492 * defined as the starting LSN of the first record in which its
4493 * transaction appears, that a record can hold multiple transactions,
4494 * and/or that a transaction can span multiple records.
4495 *
4496 * In other words, we are allowed to submit a buffer from log recovery
4497 * once per current LSN. Otherwise, we may incorrectly skip recovery
4498 * items and cause corruption.
4499 *
4500 * We don't know up front whether buffers are updated multiple times per
4501 * LSN. Therefore, track the current LSN of each commit log record as it
4502 * is processed and drain the queue when it changes. Use commit records
4503 * because they are ordered correctly by the logging code.
4504 */
4505 if (log->l_recovery_lsn != trans->r_lsn &&
4506 ohead->oh_flags & XLOG_COMMIT_TRANS) {
4507 error = xfs_buf_delwri_submit(buffer_list);
4508 if (error)
4509 return error;
4510 log->l_recovery_lsn = trans->r_lsn;
4511 }
4512
4513 return xlog_recovery_process_trans(log, trans, dp, len,
4514 ohead->oh_flags, pass, buffer_list);
4515}
4516
4517/*
4518 * There are two valid states of the r_state field. 0 indicates that the
4519 * transaction structure is in a normal state. We have either seen the
4520 * start of the transaction or the last operation we added was not a partial
4521 * operation. If the last operation we added to the transaction was a
4522 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
4523 *
4524 * NOTE: skip LRs with 0 data length.
4525 */
4526STATIC int
4527xlog_recover_process_data(
4528 struct xlog *log,
4529 struct hlist_head rhash[],
4530 struct xlog_rec_header *rhead,
4531 char *dp,
4532 int pass,
4533 struct list_head *buffer_list)
4534{
4535 struct xlog_op_header *ohead;
4536 char *end;
4537 int num_logops;
4538 int error;
4539
4540 end = dp + be32_to_cpu(rhead->h_len);
4541 num_logops = be32_to_cpu(rhead->h_num_logops);
4542
4543 /* check the log format matches our own - else we can't recover */
4544 if (xlog_header_check_recover(log->l_mp, rhead))
4545 return -EIO;
4546
4547 trace_xfs_log_recover_record(log, rhead, pass);
4548 while ((dp < end) && num_logops) {
4549
4550 ohead = (struct xlog_op_header *)dp;
4551 dp += sizeof(*ohead);
4552 ASSERT(dp <= end);
4553
4554 /* errors will abort recovery */
4555 error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
4556 dp, end, pass, buffer_list);
4557 if (error)
4558 return error;
4559
4560 dp += be32_to_cpu(ohead->oh_len);
4561 num_logops--;
4562 }
4563 return 0;
4564}
4565
4566/* Recover the EFI if necessary. */
4567STATIC int
4568xlog_recover_process_efi(
4569 struct xfs_mount *mp,
4570 struct xfs_ail *ailp,
4571 struct xfs_log_item *lip)
4572{
4573 struct xfs_efi_log_item *efip;
4574 int error;
4575
4576 /*
4577 * Skip EFIs that we've already processed.
4578 */
4579 efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4580 if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags))
4581 return 0;
4582
4583 spin_unlock(&ailp->ail_lock);
4584 error = xfs_efi_recover(mp, efip);
4585 spin_lock(&ailp->ail_lock);
4586
4587 return error;
4588}
4589
4590/* Release the EFI since we're cancelling everything. */
4591STATIC void
4592xlog_recover_cancel_efi(
4593 struct xfs_mount *mp,
4594 struct xfs_ail *ailp,
4595 struct xfs_log_item *lip)
4596{
4597 struct xfs_efi_log_item *efip;
4598
4599 efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4600
4601 spin_unlock(&ailp->ail_lock);
4602 xfs_efi_release(efip);
4603 spin_lock(&ailp->ail_lock);
4604}
4605
4606/* Recover the RUI if necessary. */
4607STATIC int
4608xlog_recover_process_rui(
4609 struct xfs_mount *mp,
4610 struct xfs_ail *ailp,
4611 struct xfs_log_item *lip)
4612{
4613 struct xfs_rui_log_item *ruip;
4614 int error;
4615
4616 /*
4617 * Skip RUIs that we've already processed.
4618 */
4619 ruip = container_of(lip, struct xfs_rui_log_item, rui_item);
4620 if (test_bit(XFS_RUI_RECOVERED, &ruip->rui_flags))
4621 return 0;
4622
4623 spin_unlock(&ailp->ail_lock);
4624 error = xfs_rui_recover(mp, ruip);
4625 spin_lock(&ailp->ail_lock);
4626
4627 return error;
4628}
4629
4630/* Release the RUI since we're cancelling everything. */
4631STATIC void
4632xlog_recover_cancel_rui(
4633 struct xfs_mount *mp,
4634 struct xfs_ail *ailp,
4635 struct xfs_log_item *lip)
4636{
4637 struct xfs_rui_log_item *ruip;
4638
4639 ruip = container_of(lip, struct xfs_rui_log_item, rui_item);
4640
4641 spin_unlock(&ailp->ail_lock);
4642 xfs_rui_release(ruip);
4643 spin_lock(&ailp->ail_lock);
4644}
4645
4646/* Recover the CUI if necessary. */
4647STATIC int
4648xlog_recover_process_cui(
4649 struct xfs_trans *parent_tp,
4650 struct xfs_ail *ailp,
4651 struct xfs_log_item *lip)
4652{
4653 struct xfs_cui_log_item *cuip;
4654 int error;
4655
4656 /*
4657 * Skip CUIs that we've already processed.
4658 */
4659 cuip = container_of(lip, struct xfs_cui_log_item, cui_item);
4660 if (test_bit(XFS_CUI_RECOVERED, &cuip->cui_flags))
4661 return 0;
4662
4663 spin_unlock(&ailp->ail_lock);
4664 error = xfs_cui_recover(parent_tp, cuip);
4665 spin_lock(&ailp->ail_lock);
4666
4667 return error;
4668}
4669
4670/* Release the CUI since we're cancelling everything. */
4671STATIC void
4672xlog_recover_cancel_cui(
4673 struct xfs_mount *mp,
4674 struct xfs_ail *ailp,
4675 struct xfs_log_item *lip)
4676{
4677 struct xfs_cui_log_item *cuip;
4678
4679 cuip = container_of(lip, struct xfs_cui_log_item, cui_item);
4680
4681 spin_unlock(&ailp->ail_lock);
4682 xfs_cui_release(cuip);
4683 spin_lock(&ailp->ail_lock);
4684}
4685
4686/* Recover the BUI if necessary. */
4687STATIC int
4688xlog_recover_process_bui(
4689 struct xfs_trans *parent_tp,
4690 struct xfs_ail *ailp,
4691 struct xfs_log_item *lip)
4692{
4693 struct xfs_bui_log_item *buip;
4694 int error;
4695
4696 /*
4697 * Skip BUIs that we've already processed.
4698 */
4699 buip = container_of(lip, struct xfs_bui_log_item, bui_item);
4700 if (test_bit(XFS_BUI_RECOVERED, &buip->bui_flags))
4701 return 0;
4702
4703 spin_unlock(&ailp->ail_lock);
4704 error = xfs_bui_recover(parent_tp, buip);
4705 spin_lock(&ailp->ail_lock);
4706
4707 return error;
4708}
4709
4710/* Release the BUI since we're cancelling everything. */
4711STATIC void
4712xlog_recover_cancel_bui(
4713 struct xfs_mount *mp,
4714 struct xfs_ail *ailp,
4715 struct xfs_log_item *lip)
4716{
4717 struct xfs_bui_log_item *buip;
4718
4719 buip = container_of(lip, struct xfs_bui_log_item, bui_item);
4720
4721 spin_unlock(&ailp->ail_lock);
4722 xfs_bui_release(buip);
4723 spin_lock(&ailp->ail_lock);
4724}
4725
4726/* Is this log item a deferred action intent? */
4727static inline bool xlog_item_is_intent(struct xfs_log_item *lip)
4728{
4729 switch (lip->li_type) {
4730 case XFS_LI_EFI:
4731 case XFS_LI_RUI:
4732 case XFS_LI_CUI:
4733 case XFS_LI_BUI:
4734 return true;
4735 default:
4736 return false;
4737 }
4738}
4739
4740/* Take all the collected deferred ops and finish them in order. */
4741static int
4742xlog_finish_defer_ops(
4743 struct xfs_trans *parent_tp)
4744{
4745 struct xfs_mount *mp = parent_tp->t_mountp;
4746 struct xfs_trans *tp;
4747 int64_t freeblks;
4748 uint resblks;
4749 int error;
4750
4751 /*
4752 * We're finishing the defer_ops that accumulated as a result of
4753 * recovering unfinished intent items during log recovery. We
4754 * reserve an itruncate transaction because it is the largest
4755 * permanent transaction type. Since we're the only user of the fs
4756 * right now, take 93% (15/16) of the available free blocks. Use
4757 * weird math to avoid a 64-bit division.
4758 */
4759 freeblks = percpu_counter_sum(&mp->m_fdblocks);
4760 if (freeblks <= 0)
4761 return -ENOSPC;
4762 resblks = min_t(int64_t, UINT_MAX, freeblks);
4763 resblks = (resblks * 15) >> 4;
4764 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, resblks,
4765 0, XFS_TRANS_RESERVE, &tp);
4766 if (error)
4767 return error;
4768 /* transfer all collected dfops to this transaction */
4769 xfs_defer_move(tp, parent_tp);
4770
4771 return xfs_trans_commit(tp);
4772}
4773
4774/*
4775 * When this is called, all of the log intent items which did not have
4776 * corresponding log done items should be in the AIL. What we do now
4777 * is update the data structures associated with each one.
4778 *
4779 * Since we process the log intent items in normal transactions, they
4780 * will be removed at some point after the commit. This prevents us
4781 * from just walking down the list processing each one. We'll use a
4782 * flag in the intent item to skip those that we've already processed
4783 * and use the AIL iteration mechanism's generation count to try to
4784 * speed this up at least a bit.
4785 *
4786 * When we start, we know that the intents are the only things in the
4787 * AIL. As we process them, however, other items are added to the
4788 * AIL.
4789 */
4790STATIC int
4791xlog_recover_process_intents(
4792 struct xlog *log)
4793{
4794 struct xfs_trans *parent_tp;
4795 struct xfs_ail_cursor cur;
4796 struct xfs_log_item *lip;
4797 struct xfs_ail *ailp;
4798 int error;
4799#if defined(DEBUG) || defined(XFS_WARN)
4800 xfs_lsn_t last_lsn;
4801#endif
4802
4803 /*
4804 * The intent recovery handlers commit transactions to complete recovery
4805 * for individual intents, but any new deferred operations that are
4806 * queued during that process are held off until the very end. The
4807 * purpose of this transaction is to serve as a container for deferred
4808 * operations. Each intent recovery handler must transfer dfops here
4809 * before its local transaction commits, and we'll finish the entire
4810 * list below.
4811 */
4812 error = xfs_trans_alloc_empty(log->l_mp, &parent_tp);
4813 if (error)
4814 return error;
4815
4816 ailp = log->l_ailp;
4817 spin_lock(&ailp->ail_lock);
4818 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4819#if defined(DEBUG) || defined(XFS_WARN)
4820 last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
4821#endif
4822 while (lip != NULL) {
4823 /*
4824 * We're done when we see something other than an intent.
4825 * There should be no intents left in the AIL now.
4826 */
4827 if (!xlog_item_is_intent(lip)) {
4828#ifdef DEBUG
4829 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4830 ASSERT(!xlog_item_is_intent(lip));
4831#endif
4832 break;
4833 }
4834
4835 /*
4836 * We should never see a redo item with a LSN higher than
4837 * the last transaction we found in the log at the start
4838 * of recovery.
4839 */
4840 ASSERT(XFS_LSN_CMP(last_lsn, lip->li_lsn) >= 0);
4841
4842 /*
4843 * NOTE: If your intent processing routine can create more
4844 * deferred ops, you /must/ attach them to the dfops in this
4845 * routine or else those subsequent intents will get
4846 * replayed in the wrong order!
4847 */
4848 switch (lip->li_type) {
4849 case XFS_LI_EFI:
4850 error = xlog_recover_process_efi(log->l_mp, ailp, lip);
4851 break;
4852 case XFS_LI_RUI:
4853 error = xlog_recover_process_rui(log->l_mp, ailp, lip);
4854 break;
4855 case XFS_LI_CUI:
4856 error = xlog_recover_process_cui(parent_tp, ailp, lip);
4857 break;
4858 case XFS_LI_BUI:
4859 error = xlog_recover_process_bui(parent_tp, ailp, lip);
4860 break;
4861 }
4862 if (error)
4863 goto out;
4864 lip = xfs_trans_ail_cursor_next(ailp, &cur);
4865 }
4866out:
4867 xfs_trans_ail_cursor_done(&cur);
4868 spin_unlock(&ailp->ail_lock);
4869 if (!error)
4870 error = xlog_finish_defer_ops(parent_tp);
4871 xfs_trans_cancel(parent_tp);
4872
4873 return error;
4874}
4875
4876/*
4877 * A cancel occurs when the mount has failed and we're bailing out.
4878 * Release all pending log intent items so they don't pin the AIL.
4879 */
4880STATIC void
4881xlog_recover_cancel_intents(
4882 struct xlog *log)
4883{
4884 struct xfs_log_item *lip;
4885 struct xfs_ail_cursor cur;
4886 struct xfs_ail *ailp;
4887
4888 ailp = log->l_ailp;
4889 spin_lock(&ailp->ail_lock);
4890 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4891 while (lip != NULL) {
4892 /*
4893 * We're done when we see something other than an intent.
4894 * There should be no intents left in the AIL now.
4895 */
4896 if (!xlog_item_is_intent(lip)) {
4897#ifdef DEBUG
4898 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4899 ASSERT(!xlog_item_is_intent(lip));
4900#endif
4901 break;
4902 }
4903
4904 switch (lip->li_type) {
4905 case XFS_LI_EFI:
4906 xlog_recover_cancel_efi(log->l_mp, ailp, lip);
4907 break;
4908 case XFS_LI_RUI:
4909 xlog_recover_cancel_rui(log->l_mp, ailp, lip);
4910 break;
4911 case XFS_LI_CUI:
4912 xlog_recover_cancel_cui(log->l_mp, ailp, lip);
4913 break;
4914 case XFS_LI_BUI:
4915 xlog_recover_cancel_bui(log->l_mp, ailp, lip);
4916 break;
4917 }
4918
4919 lip = xfs_trans_ail_cursor_next(ailp, &cur);
4920 }
4921
4922 xfs_trans_ail_cursor_done(&cur);
4923 spin_unlock(&ailp->ail_lock);
4924}
4925
4926/*
4927 * This routine performs a transaction to null out a bad inode pointer
4928 * in an agi unlinked inode hash bucket.
4929 */
4930STATIC void
4931xlog_recover_clear_agi_bucket(
4932 xfs_mount_t *mp,
4933 xfs_agnumber_t agno,
4934 int bucket)
4935{
4936 xfs_trans_t *tp;
4937 xfs_agi_t *agi;
4938 xfs_buf_t *agibp;
4939 int offset;
4940 int error;
4941
4942 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
4943 if (error)
4944 goto out_error;
4945
4946 error = xfs_read_agi(mp, tp, agno, &agibp);
4947 if (error)
4948 goto out_abort;
4949
4950 agi = XFS_BUF_TO_AGI(agibp);
4951 agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
4952 offset = offsetof(xfs_agi_t, agi_unlinked) +
4953 (sizeof(xfs_agino_t) * bucket);
4954 xfs_trans_log_buf(tp, agibp, offset,
4955 (offset + sizeof(xfs_agino_t) - 1));
4956
4957 error = xfs_trans_commit(tp);
4958 if (error)
4959 goto out_error;
4960 return;
4961
4962out_abort:
4963 xfs_trans_cancel(tp);
4964out_error:
4965 xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
4966 return;
4967}
4968
4969STATIC xfs_agino_t
4970xlog_recover_process_one_iunlink(
4971 struct xfs_mount *mp,
4972 xfs_agnumber_t agno,
4973 xfs_agino_t agino,
4974 int bucket)
4975{
4976 struct xfs_buf *ibp;
4977 struct xfs_dinode *dip;
4978 struct xfs_inode *ip;
4979 xfs_ino_t ino;
4980 int error;
4981
4982 ino = XFS_AGINO_TO_INO(mp, agno, agino);
4983 error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
4984 if (error)
4985 goto fail;
4986
4987 /*
4988 * Get the on disk inode to find the next inode in the bucket.
4989 */
4990 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0, 0);
4991 if (error)
4992 goto fail_iput;
4993
4994 xfs_iflags_clear(ip, XFS_IRECOVERY);
4995 ASSERT(VFS_I(ip)->i_nlink == 0);
4996 ASSERT(VFS_I(ip)->i_mode != 0);
4997
4998 /* setup for the next pass */
4999 agino = be32_to_cpu(dip->di_next_unlinked);
5000 xfs_buf_relse(ibp);
5001
5002 /*
5003 * Prevent any DMAPI event from being sent when the reference on
5004 * the inode is dropped.
5005 */
5006 ip->i_d.di_dmevmask = 0;
5007
5008 xfs_irele(ip);
5009 return agino;
5010
5011 fail_iput:
5012 xfs_irele(ip);
5013 fail:
5014 /*
5015 * We can't read in the inode this bucket points to, or this inode
5016 * is messed up. Just ditch this bucket of inodes. We will lose
5017 * some inodes and space, but at least we won't hang.
5018 *
5019 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
5020 * clear the inode pointer in the bucket.
5021 */
5022 xlog_recover_clear_agi_bucket(mp, agno, bucket);
5023 return NULLAGINO;
5024}
5025
5026/*
5027 * Recover AGI unlinked lists
5028 *
5029 * This is called during recovery to process any inodes which we unlinked but
5030 * not freed when the system crashed. These inodes will be on the lists in the
5031 * AGI blocks. What we do here is scan all the AGIs and fully truncate and free
5032 * any inodes found on the lists. Each inode is removed from the lists when it
5033 * has been fully truncated and is freed. The freeing of the inode and its
5034 * removal from the list must be atomic.
5035 *
5036 * If everything we touch in the agi processing loop is already in memory, this
5037 * loop can hold the cpu for a long time. It runs without lock contention,
5038 * memory allocation contention, the need wait for IO, etc, and so will run
5039 * until we either run out of inodes to process, run low on memory or we run out
5040 * of log space.
5041 *
5042 * This behaviour is bad for latency on single CPU and non-preemptible kernels,
5043 * and can prevent other filesytem work (such as CIL pushes) from running. This
5044 * can lead to deadlocks if the recovery process runs out of log reservation
5045 * space. Hence we need to yield the CPU when there is other kernel work
5046 * scheduled on this CPU to ensure other scheduled work can run without undue
5047 * latency.
5048 */
5049STATIC void
5050xlog_recover_process_iunlinks(
5051 struct xlog *log)
5052{
5053 xfs_mount_t *mp;
5054 xfs_agnumber_t agno;
5055 xfs_agi_t *agi;
5056 xfs_buf_t *agibp;
5057 xfs_agino_t agino;
5058 int bucket;
5059 int error;
5060
5061 mp = log->l_mp;
5062
5063 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5064 /*
5065 * Find the agi for this ag.
5066 */
5067 error = xfs_read_agi(mp, NULL, agno, &agibp);
5068 if (error) {
5069 /*
5070 * AGI is b0rked. Don't process it.
5071 *
5072 * We should probably mark the filesystem as corrupt
5073 * after we've recovered all the ag's we can....
5074 */
5075 continue;
5076 }
5077 /*
5078 * Unlock the buffer so that it can be acquired in the normal
5079 * course of the transaction to truncate and free each inode.
5080 * Because we are not racing with anyone else here for the AGI
5081 * buffer, we don't even need to hold it locked to read the
5082 * initial unlinked bucket entries out of the buffer. We keep
5083 * buffer reference though, so that it stays pinned in memory
5084 * while we need the buffer.
5085 */
5086 agi = XFS_BUF_TO_AGI(agibp);
5087 xfs_buf_unlock(agibp);
5088
5089 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
5090 agino = be32_to_cpu(agi->agi_unlinked[bucket]);
5091 while (agino != NULLAGINO) {
5092 agino = xlog_recover_process_one_iunlink(mp,
5093 agno, agino, bucket);
5094 cond_resched();
5095 }
5096 }
5097 xfs_buf_rele(agibp);
5098 }
5099}
5100
5101STATIC void
5102xlog_unpack_data(
5103 struct xlog_rec_header *rhead,
5104 char *dp,
5105 struct xlog *log)
5106{
5107 int i, j, k;
5108
5109 for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
5110 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
5111 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
5112 dp += BBSIZE;
5113 }
5114
5115 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
5116 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
5117 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
5118 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
5119 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
5120 *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
5121 dp += BBSIZE;
5122 }
5123 }
5124}
5125
5126/*
5127 * CRC check, unpack and process a log record.
5128 */
5129STATIC int
5130xlog_recover_process(
5131 struct xlog *log,
5132 struct hlist_head rhash[],
5133 struct xlog_rec_header *rhead,
5134 char *dp,
5135 int pass,
5136 struct list_head *buffer_list)
5137{
5138 __le32 old_crc = rhead->h_crc;
5139 __le32 crc;
5140
5141 crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
5142
5143 /*
5144 * Nothing else to do if this is a CRC verification pass. Just return
5145 * if this a record with a non-zero crc. Unfortunately, mkfs always
5146 * sets old_crc to 0 so we must consider this valid even on v5 supers.
5147 * Otherwise, return EFSBADCRC on failure so the callers up the stack
5148 * know precisely what failed.
5149 */
5150 if (pass == XLOG_RECOVER_CRCPASS) {
5151 if (old_crc && crc != old_crc)
5152 return -EFSBADCRC;
5153 return 0;
5154 }
5155
5156 /*
5157 * We're in the normal recovery path. Issue a warning if and only if the
5158 * CRC in the header is non-zero. This is an advisory warning and the
5159 * zero CRC check prevents warnings from being emitted when upgrading
5160 * the kernel from one that does not add CRCs by default.
5161 */
5162 if (crc != old_crc) {
5163 if (old_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
5164 xfs_alert(log->l_mp,
5165 "log record CRC mismatch: found 0x%x, expected 0x%x.",
5166 le32_to_cpu(old_crc),
5167 le32_to_cpu(crc));
5168 xfs_hex_dump(dp, 32);
5169 }
5170
5171 /*
5172 * If the filesystem is CRC enabled, this mismatch becomes a
5173 * fatal log corruption failure.
5174 */
5175 if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
5176 return -EFSCORRUPTED;
5177 }
5178
5179 xlog_unpack_data(rhead, dp, log);
5180
5181 return xlog_recover_process_data(log, rhash, rhead, dp, pass,
5182 buffer_list);
5183}
5184
5185STATIC int
5186xlog_valid_rec_header(
5187 struct xlog *log,
5188 struct xlog_rec_header *rhead,
5189 xfs_daddr_t blkno)
5190{
5191 int hlen;
5192
5193 if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
5194 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
5195 XFS_ERRLEVEL_LOW, log->l_mp);
5196 return -EFSCORRUPTED;
5197 }
5198 if (unlikely(
5199 (!rhead->h_version ||
5200 (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
5201 xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
5202 __func__, be32_to_cpu(rhead->h_version));
5203 return -EIO;
5204 }
5205
5206 /* LR body must have data or it wouldn't have been written */
5207 hlen = be32_to_cpu(rhead->h_len);
5208 if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
5209 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
5210 XFS_ERRLEVEL_LOW, log->l_mp);
5211 return -EFSCORRUPTED;
5212 }
5213 if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
5214 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
5215 XFS_ERRLEVEL_LOW, log->l_mp);
5216 return -EFSCORRUPTED;
5217 }
5218 return 0;
5219}
5220
5221/*
5222 * Read the log from tail to head and process the log records found.
5223 * Handle the two cases where the tail and head are in the same cycle
5224 * and where the active portion of the log wraps around the end of
5225 * the physical log separately. The pass parameter is passed through
5226 * to the routines called to process the data and is not looked at
5227 * here.
5228 */
5229STATIC int
5230xlog_do_recovery_pass(
5231 struct xlog *log,
5232 xfs_daddr_t head_blk,
5233 xfs_daddr_t tail_blk,
5234 int pass,
5235 xfs_daddr_t *first_bad) /* out: first bad log rec */
5236{
5237 xlog_rec_header_t *rhead;
5238 xfs_daddr_t blk_no, rblk_no;
5239 xfs_daddr_t rhead_blk;
5240 char *offset;
5241 char *hbp, *dbp;
5242 int error = 0, h_size, h_len;
5243 int error2 = 0;
5244 int bblks, split_bblks;
5245 int hblks, split_hblks, wrapped_hblks;
5246 int i;
5247 struct hlist_head rhash[XLOG_RHASH_SIZE];
5248 LIST_HEAD (buffer_list);
5249
5250 ASSERT(head_blk != tail_blk);
5251 blk_no = rhead_blk = tail_blk;
5252
5253 for (i = 0; i < XLOG_RHASH_SIZE; i++)
5254 INIT_HLIST_HEAD(&rhash[i]);
5255
5256 /*
5257 * Read the header of the tail block and get the iclog buffer size from
5258 * h_size. Use this to tell how many sectors make up the log header.
5259 */
5260 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
5261 /*
5262 * When using variable length iclogs, read first sector of
5263 * iclog header and extract the header size from it. Get a
5264 * new hbp that is the correct size.
5265 */
5266 hbp = xlog_alloc_buffer(log, 1);
5267 if (!hbp)
5268 return -ENOMEM;
5269
5270 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
5271 if (error)
5272 goto bread_err1;
5273
5274 rhead = (xlog_rec_header_t *)offset;
5275 error = xlog_valid_rec_header(log, rhead, tail_blk);
5276 if (error)
5277 goto bread_err1;
5278
5279 /*
5280 * xfsprogs has a bug where record length is based on lsunit but
5281 * h_size (iclog size) is hardcoded to 32k. Now that we
5282 * unconditionally CRC verify the unmount record, this means the
5283 * log buffer can be too small for the record and cause an
5284 * overrun.
5285 *
5286 * Detect this condition here. Use lsunit for the buffer size as
5287 * long as this looks like the mkfs case. Otherwise, return an
5288 * error to avoid a buffer overrun.
5289 */
5290 h_size = be32_to_cpu(rhead->h_size);
5291 h_len = be32_to_cpu(rhead->h_len);
5292 if (h_len > h_size) {
5293 if (h_len <= log->l_mp->m_logbsize &&
5294 be32_to_cpu(rhead->h_num_logops) == 1) {
5295 xfs_warn(log->l_mp,
5296 "invalid iclog size (%d bytes), using lsunit (%d bytes)",
5297 h_size, log->l_mp->m_logbsize);
5298 h_size = log->l_mp->m_logbsize;
5299 } else
5300 return -EFSCORRUPTED;
5301 }
5302
5303 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
5304 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
5305 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
5306 if (h_size % XLOG_HEADER_CYCLE_SIZE)
5307 hblks++;
5308 kmem_free(hbp);
5309 hbp = xlog_alloc_buffer(log, hblks);
5310 } else {
5311 hblks = 1;
5312 }
5313 } else {
5314 ASSERT(log->l_sectBBsize == 1);
5315 hblks = 1;
5316 hbp = xlog_alloc_buffer(log, 1);
5317 h_size = XLOG_BIG_RECORD_BSIZE;
5318 }
5319
5320 if (!hbp)
5321 return -ENOMEM;
5322 dbp = xlog_alloc_buffer(log, BTOBB(h_size));
5323 if (!dbp) {
5324 kmem_free(hbp);
5325 return -ENOMEM;
5326 }
5327
5328 memset(rhash, 0, sizeof(rhash));
5329 if (tail_blk > head_blk) {
5330 /*
5331 * Perform recovery around the end of the physical log.
5332 * When the head is not on the same cycle number as the tail,
5333 * we can't do a sequential recovery.
5334 */
5335 while (blk_no < log->l_logBBsize) {
5336 /*
5337 * Check for header wrapping around physical end-of-log
5338 */
5339 offset = hbp;
5340 split_hblks = 0;
5341 wrapped_hblks = 0;
5342 if (blk_no + hblks <= log->l_logBBsize) {
5343 /* Read header in one read */
5344 error = xlog_bread(log, blk_no, hblks, hbp,
5345 &offset);
5346 if (error)
5347 goto bread_err2;
5348 } else {
5349 /* This LR is split across physical log end */
5350 if (blk_no != log->l_logBBsize) {
5351 /* some data before physical log end */
5352 ASSERT(blk_no <= INT_MAX);
5353 split_hblks = log->l_logBBsize - (int)blk_no;
5354 ASSERT(split_hblks > 0);
5355 error = xlog_bread(log, blk_no,
5356 split_hblks, hbp,
5357 &offset);
5358 if (error)
5359 goto bread_err2;
5360 }
5361
5362 /*
5363 * Note: this black magic still works with
5364 * large sector sizes (non-512) only because:
5365 * - we increased the buffer size originally
5366 * by 1 sector giving us enough extra space
5367 * for the second read;
5368 * - the log start is guaranteed to be sector
5369 * aligned;
5370 * - we read the log end (LR header start)
5371 * _first_, then the log start (LR header end)
5372 * - order is important.
5373 */
5374 wrapped_hblks = hblks - split_hblks;
5375 error = xlog_bread_noalign(log, 0,
5376 wrapped_hblks,
5377 offset + BBTOB(split_hblks));
5378 if (error)
5379 goto bread_err2;
5380 }
5381 rhead = (xlog_rec_header_t *)offset;
5382 error = xlog_valid_rec_header(log, rhead,
5383 split_hblks ? blk_no : 0);
5384 if (error)
5385 goto bread_err2;
5386
5387 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
5388 blk_no += hblks;
5389
5390 /*
5391 * Read the log record data in multiple reads if it
5392 * wraps around the end of the log. Note that if the
5393 * header already wrapped, blk_no could point past the
5394 * end of the log. The record data is contiguous in
5395 * that case.
5396 */
5397 if (blk_no + bblks <= log->l_logBBsize ||
5398 blk_no >= log->l_logBBsize) {
5399 rblk_no = xlog_wrap_logbno(log, blk_no);
5400 error = xlog_bread(log, rblk_no, bblks, dbp,
5401 &offset);
5402 if (error)
5403 goto bread_err2;
5404 } else {
5405 /* This log record is split across the
5406 * physical end of log */
5407 offset = dbp;
5408 split_bblks = 0;
5409 if (blk_no != log->l_logBBsize) {
5410 /* some data is before the physical
5411 * end of log */
5412 ASSERT(!wrapped_hblks);
5413 ASSERT(blk_no <= INT_MAX);
5414 split_bblks =
5415 log->l_logBBsize - (int)blk_no;
5416 ASSERT(split_bblks > 0);
5417 error = xlog_bread(log, blk_no,
5418 split_bblks, dbp,
5419 &offset);
5420 if (error)
5421 goto bread_err2;
5422 }
5423
5424 /*
5425 * Note: this black magic still works with
5426 * large sector sizes (non-512) only because:
5427 * - we increased the buffer size originally
5428 * by 1 sector giving us enough extra space
5429 * for the second read;
5430 * - the log start is guaranteed to be sector
5431 * aligned;
5432 * - we read the log end (LR header start)
5433 * _first_, then the log start (LR header end)
5434 * - order is important.
5435 */
5436 error = xlog_bread_noalign(log, 0,
5437 bblks - split_bblks,
5438 offset + BBTOB(split_bblks));
5439 if (error)
5440 goto bread_err2;
5441 }
5442
5443 error = xlog_recover_process(log, rhash, rhead, offset,
5444 pass, &buffer_list);
5445 if (error)
5446 goto bread_err2;
5447
5448 blk_no += bblks;
5449 rhead_blk = blk_no;
5450 }
5451
5452 ASSERT(blk_no >= log->l_logBBsize);
5453 blk_no -= log->l_logBBsize;
5454 rhead_blk = blk_no;
5455 }
5456
5457 /* read first part of physical log */
5458 while (blk_no < head_blk) {
5459 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
5460 if (error)
5461 goto bread_err2;
5462
5463 rhead = (xlog_rec_header_t *)offset;
5464 error = xlog_valid_rec_header(log, rhead, blk_no);
5465 if (error)
5466 goto bread_err2;
5467
5468 /* blocks in data section */
5469 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
5470 error = xlog_bread(log, blk_no+hblks, bblks, dbp,
5471 &offset);
5472 if (error)
5473 goto bread_err2;
5474
5475 error = xlog_recover_process(log, rhash, rhead, offset, pass,
5476 &buffer_list);
5477 if (error)
5478 goto bread_err2;
5479
5480 blk_no += bblks + hblks;
5481 rhead_blk = blk_no;
5482 }
5483
5484 bread_err2:
5485 kmem_free(dbp);
5486 bread_err1:
5487 kmem_free(hbp);
5488
5489 /*
5490 * Submit buffers that have been added from the last record processed,
5491 * regardless of error status.
5492 */
5493 if (!list_empty(&buffer_list))
5494 error2 = xfs_buf_delwri_submit(&buffer_list);
5495
5496 if (error && first_bad)
5497 *first_bad = rhead_blk;
5498
5499 /*
5500 * Transactions are freed at commit time but transactions without commit
5501 * records on disk are never committed. Free any that may be left in the
5502 * hash table.
5503 */
5504 for (i = 0; i < XLOG_RHASH_SIZE; i++) {
5505 struct hlist_node *tmp;
5506 struct xlog_recover *trans;
5507
5508 hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
5509 xlog_recover_free_trans(trans);
5510 }
5511
5512 return error ? error : error2;
5513}
5514
5515/*
5516 * Do the recovery of the log. We actually do this in two phases.
5517 * The two passes are necessary in order to implement the function
5518 * of cancelling a record written into the log. The first pass
5519 * determines those things which have been cancelled, and the
5520 * second pass replays log items normally except for those which
5521 * have been cancelled. The handling of the replay and cancellations
5522 * takes place in the log item type specific routines.
5523 *
5524 * The table of items which have cancel records in the log is allocated
5525 * and freed at this level, since only here do we know when all of
5526 * the log recovery has been completed.
5527 */
5528STATIC int
5529xlog_do_log_recovery(
5530 struct xlog *log,
5531 xfs_daddr_t head_blk,
5532 xfs_daddr_t tail_blk)
5533{
5534 int error, i;
5535
5536 ASSERT(head_blk != tail_blk);
5537
5538 /*
5539 * First do a pass to find all of the cancelled buf log items.
5540 * Store them in the buf_cancel_table for use in the second pass.
5541 */
5542 log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
5543 sizeof(struct list_head),
5544 0);
5545 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
5546 INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
5547
5548 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
5549 XLOG_RECOVER_PASS1, NULL);
5550 if (error != 0) {
5551 kmem_free(log->l_buf_cancel_table);
5552 log->l_buf_cancel_table = NULL;
5553 return error;
5554 }
5555 /*
5556 * Then do a second pass to actually recover the items in the log.
5557 * When it is complete free the table of buf cancel items.
5558 */
5559 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
5560 XLOG_RECOVER_PASS2, NULL);
5561#ifdef DEBUG
5562 if (!error) {
5563 int i;
5564
5565 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
5566 ASSERT(list_empty(&log->l_buf_cancel_table[i]));
5567 }
5568#endif /* DEBUG */
5569
5570 kmem_free(log->l_buf_cancel_table);
5571 log->l_buf_cancel_table = NULL;
5572
5573 return error;
5574}
5575
5576/*
5577 * Do the actual recovery
5578 */
5579STATIC int
5580xlog_do_recover(
5581 struct xlog *log,
5582 xfs_daddr_t head_blk,
5583 xfs_daddr_t tail_blk)
5584{
5585 struct xfs_mount *mp = log->l_mp;
5586 int error;
5587 xfs_buf_t *bp;
5588 xfs_sb_t *sbp;
5589
5590 trace_xfs_log_recover(log, head_blk, tail_blk);
5591
5592 /*
5593 * First replay the images in the log.
5594 */
5595 error = xlog_do_log_recovery(log, head_blk, tail_blk);
5596 if (error)
5597 return error;
5598
5599 /*
5600 * If IO errors happened during recovery, bail out.
5601 */
5602 if (XFS_FORCED_SHUTDOWN(mp)) {
5603 return -EIO;
5604 }
5605
5606 /*
5607 * We now update the tail_lsn since much of the recovery has completed
5608 * and there may be space available to use. If there were no extent
5609 * or iunlinks, we can free up the entire log and set the tail_lsn to
5610 * be the last_sync_lsn. This was set in xlog_find_tail to be the
5611 * lsn of the last known good LR on disk. If there are extent frees
5612 * or iunlinks they will have some entries in the AIL; so we look at
5613 * the AIL to determine how to set the tail_lsn.
5614 */
5615 xlog_assign_tail_lsn(mp);
5616
5617 /*
5618 * Now that we've finished replaying all buffer and inode
5619 * updates, re-read in the superblock and reverify it.
5620 */
5621 bp = xfs_getsb(mp);
5622 bp->b_flags &= ~(XBF_DONE | XBF_ASYNC);
5623 ASSERT(!(bp->b_flags & XBF_WRITE));
5624 bp->b_flags |= XBF_READ;
5625 bp->b_ops = &xfs_sb_buf_ops;
5626
5627 error = xfs_buf_submit(bp);
5628 if (error) {
5629 if (!XFS_FORCED_SHUTDOWN(mp)) {
5630 xfs_buf_ioerror_alert(bp, __func__);
5631 ASSERT(0);
5632 }
5633 xfs_buf_relse(bp);
5634 return error;
5635 }
5636
5637 /* Convert superblock from on-disk format */
5638 sbp = &mp->m_sb;
5639 xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
5640 xfs_buf_relse(bp);
5641
5642 /* re-initialise in-core superblock and geometry structures */
5643 xfs_reinit_percpu_counters(mp);
5644 error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
5645 if (error) {
5646 xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
5647 return error;
5648 }
5649 mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
5650
5651 xlog_recover_check_summary(log);
5652
5653 /* Normal transactions can now occur */
5654 log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
5655 return 0;
5656}
5657
5658/*
5659 * Perform recovery and re-initialize some log variables in xlog_find_tail.
5660 *
5661 * Return error or zero.
5662 */
5663int
5664xlog_recover(
5665 struct xlog *log)
5666{
5667 xfs_daddr_t head_blk, tail_blk;
5668 int error;
5669
5670 /* find the tail of the log */
5671 error = xlog_find_tail(log, &head_blk, &tail_blk);
5672 if (error)
5673 return error;
5674
5675 /*
5676 * The superblock was read before the log was available and thus the LSN
5677 * could not be verified. Check the superblock LSN against the current
5678 * LSN now that it's known.
5679 */
5680 if (xfs_sb_version_hascrc(&log->l_mp->m_sb) &&
5681 !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
5682 return -EINVAL;
5683
5684 if (tail_blk != head_blk) {
5685 /* There used to be a comment here:
5686 *
5687 * disallow recovery on read-only mounts. note -- mount
5688 * checks for ENOSPC and turns it into an intelligent
5689 * error message.
5690 * ...but this is no longer true. Now, unless you specify
5691 * NORECOVERY (in which case this function would never be
5692 * called), we just go ahead and recover. We do this all
5693 * under the vfs layer, so we can get away with it unless
5694 * the device itself is read-only, in which case we fail.
5695 */
5696 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
5697 return error;
5698 }
5699
5700 /*
5701 * Version 5 superblock log feature mask validation. We know the
5702 * log is dirty so check if there are any unknown log features
5703 * in what we need to recover. If there are unknown features
5704 * (e.g. unsupported transactions, then simply reject the
5705 * attempt at recovery before touching anything.
5706 */
5707 if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
5708 xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
5709 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
5710 xfs_warn(log->l_mp,
5711"Superblock has unknown incompatible log features (0x%x) enabled.",
5712 (log->l_mp->m_sb.sb_features_log_incompat &
5713 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
5714 xfs_warn(log->l_mp,
5715"The log can not be fully and/or safely recovered by this kernel.");
5716 xfs_warn(log->l_mp,
5717"Please recover the log on a kernel that supports the unknown features.");
5718 return -EINVAL;
5719 }
5720
5721 /*
5722 * Delay log recovery if the debug hook is set. This is debug
5723 * instrumention to coordinate simulation of I/O failures with
5724 * log recovery.
5725 */
5726 if (xfs_globals.log_recovery_delay) {
5727 xfs_notice(log->l_mp,
5728 "Delaying log recovery for %d seconds.",
5729 xfs_globals.log_recovery_delay);
5730 msleep(xfs_globals.log_recovery_delay * 1000);
5731 }
5732
5733 xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
5734 log->l_mp->m_logname ? log->l_mp->m_logname
5735 : "internal");
5736
5737 error = xlog_do_recover(log, head_blk, tail_blk);
5738 log->l_flags |= XLOG_RECOVERY_NEEDED;
5739 }
5740 return error;
5741}
5742
5743/*
5744 * In the first part of recovery we replay inodes and buffers and build
5745 * up the list of extent free items which need to be processed. Here
5746 * we process the extent free items and clean up the on disk unlinked
5747 * inode lists. This is separated from the first part of recovery so
5748 * that the root and real-time bitmap inodes can be read in from disk in
5749 * between the two stages. This is necessary so that we can free space
5750 * in the real-time portion of the file system.
5751 */
5752int
5753xlog_recover_finish(
5754 struct xlog *log)
5755{
5756 /*
5757 * Now we're ready to do the transactions needed for the
5758 * rest of recovery. Start with completing all the extent
5759 * free intent records and then process the unlinked inode
5760 * lists. At this point, we essentially run in normal mode
5761 * except that we're still performing recovery actions
5762 * rather than accepting new requests.
5763 */
5764 if (log->l_flags & XLOG_RECOVERY_NEEDED) {
5765 int error;
5766 error = xlog_recover_process_intents(log);
5767 if (error) {
5768 xfs_alert(log->l_mp, "Failed to recover intents");
5769 return error;
5770 }
5771
5772 /*
5773 * Sync the log to get all the intents out of the AIL.
5774 * This isn't absolutely necessary, but it helps in
5775 * case the unlink transactions would have problems
5776 * pushing the intents out of the way.
5777 */
5778 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
5779
5780 xlog_recover_process_iunlinks(log);
5781
5782 xlog_recover_check_summary(log);
5783
5784 xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
5785 log->l_mp->m_logname ? log->l_mp->m_logname
5786 : "internal");
5787 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
5788 } else {
5789 xfs_info(log->l_mp, "Ending clean mount");
5790 }
5791 return 0;
5792}
5793
5794void
5795xlog_recover_cancel(
5796 struct xlog *log)
5797{
5798 if (log->l_flags & XLOG_RECOVERY_NEEDED)
5799 xlog_recover_cancel_intents(log);
5800}
5801
5802#if defined(DEBUG)
5803/*
5804 * Read all of the agf and agi counters and check that they
5805 * are consistent with the superblock counters.
5806 */
5807STATIC void
5808xlog_recover_check_summary(
5809 struct xlog *log)
5810{
5811 xfs_mount_t *mp;
5812 xfs_agf_t *agfp;
5813 xfs_buf_t *agfbp;
5814 xfs_buf_t *agibp;
5815 xfs_agnumber_t agno;
5816 uint64_t freeblks;
5817 uint64_t itotal;
5818 uint64_t ifree;
5819 int error;
5820
5821 mp = log->l_mp;
5822
5823 freeblks = 0LL;
5824 itotal = 0LL;
5825 ifree = 0LL;
5826 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5827 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
5828 if (error) {
5829 xfs_alert(mp, "%s agf read failed agno %d error %d",
5830 __func__, agno, error);
5831 } else {
5832 agfp = XFS_BUF_TO_AGF(agfbp);
5833 freeblks += be32_to_cpu(agfp->agf_freeblks) +
5834 be32_to_cpu(agfp->agf_flcount);
5835 xfs_buf_relse(agfbp);
5836 }
5837
5838 error = xfs_read_agi(mp, NULL, agno, &agibp);
5839 if (error) {
5840 xfs_alert(mp, "%s agi read failed agno %d error %d",
5841 __func__, agno, error);
5842 } else {
5843 struct xfs_agi *agi = XFS_BUF_TO_AGI(agibp);
5844
5845 itotal += be32_to_cpu(agi->agi_count);
5846 ifree += be32_to_cpu(agi->agi_freecount);
5847 xfs_buf_relse(agibp);
5848 }
5849 }
5850}
5851#endif /* DEBUG */