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/*
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_shared.h"
21#include "xfs_format.h"
22#include "xfs_log_format.h"
23#include "xfs_trans_resv.h"
24#include "xfs_bit.h"
25#include "xfs_sb.h"
26#include "xfs_mount.h"
27#include "xfs_da_format.h"
28#include "xfs_da_btree.h"
29#include "xfs_inode.h"
30#include "xfs_trans.h"
31#include "xfs_log.h"
32#include "xfs_log_priv.h"
33#include "xfs_log_recover.h"
34#include "xfs_inode_item.h"
35#include "xfs_extfree_item.h"
36#include "xfs_trans_priv.h"
37#include "xfs_alloc.h"
38#include "xfs_ialloc.h"
39#include "xfs_quota.h"
40#include "xfs_cksum.h"
41#include "xfs_trace.h"
42#include "xfs_icache.h"
43#include "xfs_bmap_btree.h"
44#include "xfs_error.h"
45#include "xfs_dir2.h"
46
47#define BLK_AVG(blk1, blk2) ((blk1+blk2) >> 1)
48
49STATIC int
50xlog_find_zeroed(
51 struct xlog *,
52 xfs_daddr_t *);
53STATIC int
54xlog_clear_stale_blocks(
55 struct xlog *,
56 xfs_lsn_t);
57#if defined(DEBUG)
58STATIC void
59xlog_recover_check_summary(
60 struct xlog *);
61#else
62#define xlog_recover_check_summary(log)
63#endif
64STATIC int
65xlog_do_recovery_pass(
66 struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
67
68/*
69 * This structure is used during recovery to record the buf log items which
70 * have been canceled and should not be replayed.
71 */
72struct xfs_buf_cancel {
73 xfs_daddr_t bc_blkno;
74 uint bc_len;
75 int bc_refcount;
76 struct list_head bc_list;
77};
78
79/*
80 * Sector aligned buffer routines for buffer create/read/write/access
81 */
82
83/*
84 * Verify the given count of basic blocks is valid number of blocks
85 * to specify for an operation involving the given XFS log buffer.
86 * Returns nonzero if the count is valid, 0 otherwise.
87 */
88
89static inline int
90xlog_buf_bbcount_valid(
91 struct xlog *log,
92 int bbcount)
93{
94 return bbcount > 0 && bbcount <= log->l_logBBsize;
95}
96
97/*
98 * Allocate a buffer to hold log data. The buffer needs to be able
99 * to map to a range of nbblks basic blocks at any valid (basic
100 * block) offset within the log.
101 */
102STATIC xfs_buf_t *
103xlog_get_bp(
104 struct xlog *log,
105 int nbblks)
106{
107 struct xfs_buf *bp;
108
109 if (!xlog_buf_bbcount_valid(log, nbblks)) {
110 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
111 nbblks);
112 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
113 return NULL;
114 }
115
116 /*
117 * We do log I/O in units of log sectors (a power-of-2
118 * multiple of the basic block size), so we round up the
119 * requested size to accommodate the basic blocks required
120 * for complete log sectors.
121 *
122 * In addition, the buffer may be used for a non-sector-
123 * aligned block offset, in which case an I/O of the
124 * requested size could extend beyond the end of the
125 * buffer. If the requested size is only 1 basic block it
126 * will never straddle a sector boundary, so this won't be
127 * an issue. Nor will this be a problem if the log I/O is
128 * done in basic blocks (sector size 1). But otherwise we
129 * extend the buffer by one extra log sector to ensure
130 * there's space to accommodate this possibility.
131 */
132 if (nbblks > 1 && log->l_sectBBsize > 1)
133 nbblks += log->l_sectBBsize;
134 nbblks = round_up(nbblks, log->l_sectBBsize);
135
136 bp = xfs_buf_get_uncached(log->l_mp->m_logdev_targp, nbblks, 0);
137 if (bp)
138 xfs_buf_unlock(bp);
139 return bp;
140}
141
142STATIC void
143xlog_put_bp(
144 xfs_buf_t *bp)
145{
146 xfs_buf_free(bp);
147}
148
149/*
150 * Return the address of the start of the given block number's data
151 * in a log buffer. The buffer covers a log sector-aligned region.
152 */
153STATIC char *
154xlog_align(
155 struct xlog *log,
156 xfs_daddr_t blk_no,
157 int nbblks,
158 struct xfs_buf *bp)
159{
160 xfs_daddr_t offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
161
162 ASSERT(offset + nbblks <= bp->b_length);
163 return bp->b_addr + BBTOB(offset);
164}
165
166
167/*
168 * nbblks should be uint, but oh well. Just want to catch that 32-bit length.
169 */
170STATIC int
171xlog_bread_noalign(
172 struct xlog *log,
173 xfs_daddr_t blk_no,
174 int nbblks,
175 struct xfs_buf *bp)
176{
177 int error;
178
179 if (!xlog_buf_bbcount_valid(log, nbblks)) {
180 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
181 nbblks);
182 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
183 return -EFSCORRUPTED;
184 }
185
186 blk_no = round_down(blk_no, log->l_sectBBsize);
187 nbblks = round_up(nbblks, log->l_sectBBsize);
188
189 ASSERT(nbblks > 0);
190 ASSERT(nbblks <= bp->b_length);
191
192 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
193 bp->b_flags |= XBF_READ;
194 bp->b_io_length = nbblks;
195 bp->b_error = 0;
196
197 error = xfs_buf_submit_wait(bp);
198 if (error && !XFS_FORCED_SHUTDOWN(log->l_mp))
199 xfs_buf_ioerror_alert(bp, __func__);
200 return error;
201}
202
203STATIC int
204xlog_bread(
205 struct xlog *log,
206 xfs_daddr_t blk_no,
207 int nbblks,
208 struct xfs_buf *bp,
209 char **offset)
210{
211 int error;
212
213 error = xlog_bread_noalign(log, blk_no, nbblks, bp);
214 if (error)
215 return error;
216
217 *offset = xlog_align(log, blk_no, nbblks, bp);
218 return 0;
219}
220
221/*
222 * Read at an offset into the buffer. Returns with the buffer in it's original
223 * state regardless of the result of the read.
224 */
225STATIC int
226xlog_bread_offset(
227 struct xlog *log,
228 xfs_daddr_t blk_no, /* block to read from */
229 int nbblks, /* blocks to read */
230 struct xfs_buf *bp,
231 char *offset)
232{
233 char *orig_offset = bp->b_addr;
234 int orig_len = BBTOB(bp->b_length);
235 int error, error2;
236
237 error = xfs_buf_associate_memory(bp, offset, BBTOB(nbblks));
238 if (error)
239 return error;
240
241 error = xlog_bread_noalign(log, blk_no, nbblks, bp);
242
243 /* must reset buffer pointer even on error */
244 error2 = xfs_buf_associate_memory(bp, orig_offset, orig_len);
245 if (error)
246 return error;
247 return error2;
248}
249
250/*
251 * Write out the buffer at the given block for the given number of blocks.
252 * The buffer is kept locked across the write and is returned locked.
253 * This can only be used for synchronous log writes.
254 */
255STATIC int
256xlog_bwrite(
257 struct xlog *log,
258 xfs_daddr_t blk_no,
259 int nbblks,
260 struct xfs_buf *bp)
261{
262 int error;
263
264 if (!xlog_buf_bbcount_valid(log, nbblks)) {
265 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
266 nbblks);
267 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
268 return -EFSCORRUPTED;
269 }
270
271 blk_no = round_down(blk_no, log->l_sectBBsize);
272 nbblks = round_up(nbblks, log->l_sectBBsize);
273
274 ASSERT(nbblks > 0);
275 ASSERT(nbblks <= bp->b_length);
276
277 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
278 xfs_buf_hold(bp);
279 xfs_buf_lock(bp);
280 bp->b_io_length = nbblks;
281 bp->b_error = 0;
282
283 error = xfs_bwrite(bp);
284 if (error)
285 xfs_buf_ioerror_alert(bp, __func__);
286 xfs_buf_relse(bp);
287 return error;
288}
289
290#ifdef DEBUG
291/*
292 * dump debug superblock and log record information
293 */
294STATIC void
295xlog_header_check_dump(
296 xfs_mount_t *mp,
297 xlog_rec_header_t *head)
298{
299 xfs_debug(mp, "%s: SB : uuid = %pU, fmt = %d",
300 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
301 xfs_debug(mp, " log : uuid = %pU, fmt = %d",
302 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
303}
304#else
305#define xlog_header_check_dump(mp, head)
306#endif
307
308/*
309 * check log record header for recovery
310 */
311STATIC int
312xlog_header_check_recover(
313 xfs_mount_t *mp,
314 xlog_rec_header_t *head)
315{
316 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
317
318 /*
319 * IRIX doesn't write the h_fmt field and leaves it zeroed
320 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
321 * a dirty log created in IRIX.
322 */
323 if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
324 xfs_warn(mp,
325 "dirty log written in incompatible format - can't recover");
326 xlog_header_check_dump(mp, head);
327 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
328 XFS_ERRLEVEL_HIGH, mp);
329 return -EFSCORRUPTED;
330 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
331 xfs_warn(mp,
332 "dirty log entry has mismatched uuid - can't recover");
333 xlog_header_check_dump(mp, head);
334 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
335 XFS_ERRLEVEL_HIGH, mp);
336 return -EFSCORRUPTED;
337 }
338 return 0;
339}
340
341/*
342 * read the head block of the log and check the header
343 */
344STATIC int
345xlog_header_check_mount(
346 xfs_mount_t *mp,
347 xlog_rec_header_t *head)
348{
349 ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
350
351 if (uuid_is_nil(&head->h_fs_uuid)) {
352 /*
353 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
354 * h_fs_uuid is nil, we assume this log was last mounted
355 * by IRIX and continue.
356 */
357 xfs_warn(mp, "nil uuid in log - IRIX style log");
358 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
359 xfs_warn(mp, "log has mismatched uuid - can't recover");
360 xlog_header_check_dump(mp, head);
361 XFS_ERROR_REPORT("xlog_header_check_mount",
362 XFS_ERRLEVEL_HIGH, mp);
363 return -EFSCORRUPTED;
364 }
365 return 0;
366}
367
368STATIC void
369xlog_recover_iodone(
370 struct xfs_buf *bp)
371{
372 if (bp->b_error) {
373 /*
374 * We're not going to bother about retrying
375 * this during recovery. One strike!
376 */
377 if (!XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
378 xfs_buf_ioerror_alert(bp, __func__);
379 xfs_force_shutdown(bp->b_target->bt_mount,
380 SHUTDOWN_META_IO_ERROR);
381 }
382 }
383 bp->b_iodone = NULL;
384 xfs_buf_ioend(bp);
385}
386
387/*
388 * This routine finds (to an approximation) the first block in the physical
389 * log which contains the given cycle. It uses a binary search algorithm.
390 * Note that the algorithm can not be perfect because the disk will not
391 * necessarily be perfect.
392 */
393STATIC int
394xlog_find_cycle_start(
395 struct xlog *log,
396 struct xfs_buf *bp,
397 xfs_daddr_t first_blk,
398 xfs_daddr_t *last_blk,
399 uint cycle)
400{
401 char *offset;
402 xfs_daddr_t mid_blk;
403 xfs_daddr_t end_blk;
404 uint mid_cycle;
405 int error;
406
407 end_blk = *last_blk;
408 mid_blk = BLK_AVG(first_blk, end_blk);
409 while (mid_blk != first_blk && mid_blk != end_blk) {
410 error = xlog_bread(log, mid_blk, 1, bp, &offset);
411 if (error)
412 return error;
413 mid_cycle = xlog_get_cycle(offset);
414 if (mid_cycle == cycle)
415 end_blk = mid_blk; /* last_half_cycle == mid_cycle */
416 else
417 first_blk = mid_blk; /* first_half_cycle == mid_cycle */
418 mid_blk = BLK_AVG(first_blk, end_blk);
419 }
420 ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
421 (mid_blk == end_blk && mid_blk-1 == first_blk));
422
423 *last_blk = end_blk;
424
425 return 0;
426}
427
428/*
429 * Check that a range of blocks does not contain stop_on_cycle_no.
430 * Fill in *new_blk with the block offset where such a block is
431 * found, or with -1 (an invalid block number) if there is no such
432 * block in the range. The scan needs to occur from front to back
433 * and the pointer into the region must be updated since a later
434 * routine will need to perform another test.
435 */
436STATIC int
437xlog_find_verify_cycle(
438 struct xlog *log,
439 xfs_daddr_t start_blk,
440 int nbblks,
441 uint stop_on_cycle_no,
442 xfs_daddr_t *new_blk)
443{
444 xfs_daddr_t i, j;
445 uint cycle;
446 xfs_buf_t *bp;
447 xfs_daddr_t bufblks;
448 char *buf = NULL;
449 int error = 0;
450
451 /*
452 * Greedily allocate a buffer big enough to handle the full
453 * range of basic blocks we'll be examining. If that fails,
454 * try a smaller size. We need to be able to read at least
455 * a log sector, or we're out of luck.
456 */
457 bufblks = 1 << ffs(nbblks);
458 while (bufblks > log->l_logBBsize)
459 bufblks >>= 1;
460 while (!(bp = xlog_get_bp(log, bufblks))) {
461 bufblks >>= 1;
462 if (bufblks < log->l_sectBBsize)
463 return -ENOMEM;
464 }
465
466 for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
467 int bcount;
468
469 bcount = min(bufblks, (start_blk + nbblks - i));
470
471 error = xlog_bread(log, i, bcount, bp, &buf);
472 if (error)
473 goto out;
474
475 for (j = 0; j < bcount; j++) {
476 cycle = xlog_get_cycle(buf);
477 if (cycle == stop_on_cycle_no) {
478 *new_blk = i+j;
479 goto out;
480 }
481
482 buf += BBSIZE;
483 }
484 }
485
486 *new_blk = -1;
487
488out:
489 xlog_put_bp(bp);
490 return error;
491}
492
493/*
494 * Potentially backup over partial log record write.
495 *
496 * In the typical case, last_blk is the number of the block directly after
497 * a good log record. Therefore, we subtract one to get the block number
498 * of the last block in the given buffer. extra_bblks contains the number
499 * of blocks we would have read on a previous read. This happens when the
500 * last log record is split over the end of the physical log.
501 *
502 * extra_bblks is the number of blocks potentially verified on a previous
503 * call to this routine.
504 */
505STATIC int
506xlog_find_verify_log_record(
507 struct xlog *log,
508 xfs_daddr_t start_blk,
509 xfs_daddr_t *last_blk,
510 int extra_bblks)
511{
512 xfs_daddr_t i;
513 xfs_buf_t *bp;
514 char *offset = NULL;
515 xlog_rec_header_t *head = NULL;
516 int error = 0;
517 int smallmem = 0;
518 int num_blks = *last_blk - start_blk;
519 int xhdrs;
520
521 ASSERT(start_blk != 0 || *last_blk != start_blk);
522
523 if (!(bp = xlog_get_bp(log, num_blks))) {
524 if (!(bp = xlog_get_bp(log, 1)))
525 return -ENOMEM;
526 smallmem = 1;
527 } else {
528 error = xlog_bread(log, start_blk, num_blks, bp, &offset);
529 if (error)
530 goto out;
531 offset += ((num_blks - 1) << BBSHIFT);
532 }
533
534 for (i = (*last_blk) - 1; i >= 0; i--) {
535 if (i < start_blk) {
536 /* valid log record not found */
537 xfs_warn(log->l_mp,
538 "Log inconsistent (didn't find previous header)");
539 ASSERT(0);
540 error = -EIO;
541 goto out;
542 }
543
544 if (smallmem) {
545 error = xlog_bread(log, i, 1, bp, &offset);
546 if (error)
547 goto out;
548 }
549
550 head = (xlog_rec_header_t *)offset;
551
552 if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
553 break;
554
555 if (!smallmem)
556 offset -= BBSIZE;
557 }
558
559 /*
560 * We hit the beginning of the physical log & still no header. Return
561 * to caller. If caller can handle a return of -1, then this routine
562 * will be called again for the end of the physical log.
563 */
564 if (i == -1) {
565 error = 1;
566 goto out;
567 }
568
569 /*
570 * We have the final block of the good log (the first block
571 * of the log record _before_ the head. So we check the uuid.
572 */
573 if ((error = xlog_header_check_mount(log->l_mp, head)))
574 goto out;
575
576 /*
577 * We may have found a log record header before we expected one.
578 * last_blk will be the 1st block # with a given cycle #. We may end
579 * up reading an entire log record. In this case, we don't want to
580 * reset last_blk. Only when last_blk points in the middle of a log
581 * record do we update last_blk.
582 */
583 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
584 uint h_size = be32_to_cpu(head->h_size);
585
586 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
587 if (h_size % XLOG_HEADER_CYCLE_SIZE)
588 xhdrs++;
589 } else {
590 xhdrs = 1;
591 }
592
593 if (*last_blk - i + extra_bblks !=
594 BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
595 *last_blk = i;
596
597out:
598 xlog_put_bp(bp);
599 return error;
600}
601
602/*
603 * Head is defined to be the point of the log where the next log write
604 * could go. This means that incomplete LR writes at the end are
605 * eliminated when calculating the head. We aren't guaranteed that previous
606 * LR have complete transactions. We only know that a cycle number of
607 * current cycle number -1 won't be present in the log if we start writing
608 * from our current block number.
609 *
610 * last_blk contains the block number of the first block with a given
611 * cycle number.
612 *
613 * Return: zero if normal, non-zero if error.
614 */
615STATIC int
616xlog_find_head(
617 struct xlog *log,
618 xfs_daddr_t *return_head_blk)
619{
620 xfs_buf_t *bp;
621 char *offset;
622 xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
623 int num_scan_bblks;
624 uint first_half_cycle, last_half_cycle;
625 uint stop_on_cycle;
626 int error, log_bbnum = log->l_logBBsize;
627
628 /* Is the end of the log device zeroed? */
629 error = xlog_find_zeroed(log, &first_blk);
630 if (error < 0) {
631 xfs_warn(log->l_mp, "empty log check failed");
632 return error;
633 }
634 if (error == 1) {
635 *return_head_blk = first_blk;
636
637 /* Is the whole lot zeroed? */
638 if (!first_blk) {
639 /* Linux XFS shouldn't generate totally zeroed logs -
640 * mkfs etc write a dummy unmount record to a fresh
641 * log so we can store the uuid in there
642 */
643 xfs_warn(log->l_mp, "totally zeroed log");
644 }
645
646 return 0;
647 }
648
649 first_blk = 0; /* get cycle # of 1st block */
650 bp = xlog_get_bp(log, 1);
651 if (!bp)
652 return -ENOMEM;
653
654 error = xlog_bread(log, 0, 1, bp, &offset);
655 if (error)
656 goto bp_err;
657
658 first_half_cycle = xlog_get_cycle(offset);
659
660 last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
661 error = xlog_bread(log, last_blk, 1, bp, &offset);
662 if (error)
663 goto bp_err;
664
665 last_half_cycle = xlog_get_cycle(offset);
666 ASSERT(last_half_cycle != 0);
667
668 /*
669 * If the 1st half cycle number is equal to the last half cycle number,
670 * then the entire log is stamped with the same cycle number. In this
671 * case, head_blk can't be set to zero (which makes sense). The below
672 * math doesn't work out properly with head_blk equal to zero. Instead,
673 * we set it to log_bbnum which is an invalid block number, but this
674 * value makes the math correct. If head_blk doesn't changed through
675 * all the tests below, *head_blk is set to zero at the very end rather
676 * than log_bbnum. In a sense, log_bbnum and zero are the same block
677 * in a circular file.
678 */
679 if (first_half_cycle == last_half_cycle) {
680 /*
681 * In this case we believe that the entire log should have
682 * cycle number last_half_cycle. We need to scan backwards
683 * from the end verifying that there are no holes still
684 * containing last_half_cycle - 1. If we find such a hole,
685 * then the start of that hole will be the new head. The
686 * simple case looks like
687 * x | x ... | x - 1 | x
688 * Another case that fits this picture would be
689 * x | x + 1 | x ... | x
690 * In this case the head really is somewhere at the end of the
691 * log, as one of the latest writes at the beginning was
692 * incomplete.
693 * One more case is
694 * x | x + 1 | x ... | x - 1 | x
695 * This is really the combination of the above two cases, and
696 * the head has to end up at the start of the x-1 hole at the
697 * end of the log.
698 *
699 * In the 256k log case, we will read from the beginning to the
700 * end of the log and search for cycle numbers equal to x-1.
701 * We don't worry about the x+1 blocks that we encounter,
702 * because we know that they cannot be the head since the log
703 * started with x.
704 */
705 head_blk = log_bbnum;
706 stop_on_cycle = last_half_cycle - 1;
707 } else {
708 /*
709 * In this case we want to find the first block with cycle
710 * number matching last_half_cycle. We expect the log to be
711 * some variation on
712 * x + 1 ... | x ... | x
713 * The first block with cycle number x (last_half_cycle) will
714 * be where the new head belongs. First we do a binary search
715 * for the first occurrence of last_half_cycle. The binary
716 * search may not be totally accurate, so then we scan back
717 * from there looking for occurrences of last_half_cycle before
718 * us. If that backwards scan wraps around the beginning of
719 * the log, then we look for occurrences of last_half_cycle - 1
720 * at the end of the log. The cases we're looking for look
721 * like
722 * v binary search stopped here
723 * x + 1 ... | x | x + 1 | x ... | x
724 * ^ but we want to locate this spot
725 * or
726 * <---------> less than scan distance
727 * x + 1 ... | x ... | x - 1 | x
728 * ^ we want to locate this spot
729 */
730 stop_on_cycle = last_half_cycle;
731 if ((error = xlog_find_cycle_start(log, bp, first_blk,
732 &head_blk, last_half_cycle)))
733 goto bp_err;
734 }
735
736 /*
737 * Now validate the answer. Scan back some number of maximum possible
738 * blocks and make sure each one has the expected cycle number. The
739 * maximum is determined by the total possible amount of buffering
740 * in the in-core log. The following number can be made tighter if
741 * we actually look at the block size of the filesystem.
742 */
743 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
744 if (head_blk >= num_scan_bblks) {
745 /*
746 * We are guaranteed that the entire check can be performed
747 * in one buffer.
748 */
749 start_blk = head_blk - num_scan_bblks;
750 if ((error = xlog_find_verify_cycle(log,
751 start_blk, num_scan_bblks,
752 stop_on_cycle, &new_blk)))
753 goto bp_err;
754 if (new_blk != -1)
755 head_blk = new_blk;
756 } else { /* need to read 2 parts of log */
757 /*
758 * We are going to scan backwards in the log in two parts.
759 * First we scan the physical end of the log. In this part
760 * of the log, we are looking for blocks with cycle number
761 * last_half_cycle - 1.
762 * If we find one, then we know that the log starts there, as
763 * we've found a hole that didn't get written in going around
764 * the end of the physical log. The simple case for this is
765 * x + 1 ... | x ... | x - 1 | x
766 * <---------> less than scan distance
767 * If all of the blocks at the end of the log have cycle number
768 * last_half_cycle, then we check the blocks at the start of
769 * the log looking for occurrences of last_half_cycle. If we
770 * find one, then our current estimate for the location of the
771 * first occurrence of last_half_cycle is wrong and we move
772 * back to the hole we've found. This case looks like
773 * x + 1 ... | x | x + 1 | x ...
774 * ^ binary search stopped here
775 * Another case we need to handle that only occurs in 256k
776 * logs is
777 * x + 1 ... | x ... | x+1 | x ...
778 * ^ binary search stops here
779 * In a 256k log, the scan at the end of the log will see the
780 * x + 1 blocks. We need to skip past those since that is
781 * certainly not the head of the log. By searching for
782 * last_half_cycle-1 we accomplish that.
783 */
784 ASSERT(head_blk <= INT_MAX &&
785 (xfs_daddr_t) num_scan_bblks >= head_blk);
786 start_blk = log_bbnum - (num_scan_bblks - head_blk);
787 if ((error = xlog_find_verify_cycle(log, start_blk,
788 num_scan_bblks - (int)head_blk,
789 (stop_on_cycle - 1), &new_blk)))
790 goto bp_err;
791 if (new_blk != -1) {
792 head_blk = new_blk;
793 goto validate_head;
794 }
795
796 /*
797 * Scan beginning of log now. The last part of the physical
798 * log is good. This scan needs to verify that it doesn't find
799 * the last_half_cycle.
800 */
801 start_blk = 0;
802 ASSERT(head_blk <= INT_MAX);
803 if ((error = xlog_find_verify_cycle(log,
804 start_blk, (int)head_blk,
805 stop_on_cycle, &new_blk)))
806 goto bp_err;
807 if (new_blk != -1)
808 head_blk = new_blk;
809 }
810
811validate_head:
812 /*
813 * Now we need to make sure head_blk is not pointing to a block in
814 * the middle of a log record.
815 */
816 num_scan_bblks = XLOG_REC_SHIFT(log);
817 if (head_blk >= num_scan_bblks) {
818 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
819
820 /* start ptr at last block ptr before head_blk */
821 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
822 if (error == 1)
823 error = -EIO;
824 if (error)
825 goto bp_err;
826 } else {
827 start_blk = 0;
828 ASSERT(head_blk <= INT_MAX);
829 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
830 if (error < 0)
831 goto bp_err;
832 if (error == 1) {
833 /* We hit the beginning of the log during our search */
834 start_blk = log_bbnum - (num_scan_bblks - head_blk);
835 new_blk = log_bbnum;
836 ASSERT(start_blk <= INT_MAX &&
837 (xfs_daddr_t) log_bbnum-start_blk >= 0);
838 ASSERT(head_blk <= INT_MAX);
839 error = xlog_find_verify_log_record(log, start_blk,
840 &new_blk, (int)head_blk);
841 if (error == 1)
842 error = -EIO;
843 if (error)
844 goto bp_err;
845 if (new_blk != log_bbnum)
846 head_blk = new_blk;
847 } else if (error)
848 goto bp_err;
849 }
850
851 xlog_put_bp(bp);
852 if (head_blk == log_bbnum)
853 *return_head_blk = 0;
854 else
855 *return_head_blk = head_blk;
856 /*
857 * When returning here, we have a good block number. Bad block
858 * means that during a previous crash, we didn't have a clean break
859 * from cycle number N to cycle number N-1. In this case, we need
860 * to find the first block with cycle number N-1.
861 */
862 return 0;
863
864 bp_err:
865 xlog_put_bp(bp);
866
867 if (error)
868 xfs_warn(log->l_mp, "failed to find log head");
869 return error;
870}
871
872/*
873 * Seek backwards in the log for log record headers.
874 *
875 * Given a starting log block, walk backwards until we find the provided number
876 * of records or hit the provided tail block. The return value is the number of
877 * records encountered or a negative error code. The log block and buffer
878 * pointer of the last record seen are returned in rblk and rhead respectively.
879 */
880STATIC int
881xlog_rseek_logrec_hdr(
882 struct xlog *log,
883 xfs_daddr_t head_blk,
884 xfs_daddr_t tail_blk,
885 int count,
886 struct xfs_buf *bp,
887 xfs_daddr_t *rblk,
888 struct xlog_rec_header **rhead,
889 bool *wrapped)
890{
891 int i;
892 int error;
893 int found = 0;
894 char *offset = NULL;
895 xfs_daddr_t end_blk;
896
897 *wrapped = false;
898
899 /*
900 * Walk backwards from the head block until we hit the tail or the first
901 * block in the log.
902 */
903 end_blk = head_blk > tail_blk ? tail_blk : 0;
904 for (i = (int) head_blk - 1; i >= end_blk; i--) {
905 error = xlog_bread(log, i, 1, bp, &offset);
906 if (error)
907 goto out_error;
908
909 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
910 *rblk = i;
911 *rhead = (struct xlog_rec_header *) offset;
912 if (++found == count)
913 break;
914 }
915 }
916
917 /*
918 * If we haven't hit the tail block or the log record header count,
919 * start looking again from the end of the physical log. Note that
920 * callers can pass head == tail if the tail is not yet known.
921 */
922 if (tail_blk >= head_blk && found != count) {
923 for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
924 error = xlog_bread(log, i, 1, bp, &offset);
925 if (error)
926 goto out_error;
927
928 if (*(__be32 *)offset ==
929 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
930 *wrapped = true;
931 *rblk = i;
932 *rhead = (struct xlog_rec_header *) offset;
933 if (++found == count)
934 break;
935 }
936 }
937 }
938
939 return found;
940
941out_error:
942 return error;
943}
944
945/*
946 * Seek forward in the log for log record headers.
947 *
948 * Given head and tail blocks, walk forward from the tail block until we find
949 * the provided number of records or hit the head block. The return value is the
950 * number of records encountered or a negative error code. The log block and
951 * buffer pointer of the last record seen are returned in rblk and rhead
952 * respectively.
953 */
954STATIC int
955xlog_seek_logrec_hdr(
956 struct xlog *log,
957 xfs_daddr_t head_blk,
958 xfs_daddr_t tail_blk,
959 int count,
960 struct xfs_buf *bp,
961 xfs_daddr_t *rblk,
962 struct xlog_rec_header **rhead,
963 bool *wrapped)
964{
965 int i;
966 int error;
967 int found = 0;
968 char *offset = NULL;
969 xfs_daddr_t end_blk;
970
971 *wrapped = false;
972
973 /*
974 * Walk forward from the tail block until we hit the head or the last
975 * block in the log.
976 */
977 end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
978 for (i = (int) tail_blk; i <= end_blk; i++) {
979 error = xlog_bread(log, i, 1, bp, &offset);
980 if (error)
981 goto out_error;
982
983 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
984 *rblk = i;
985 *rhead = (struct xlog_rec_header *) offset;
986 if (++found == count)
987 break;
988 }
989 }
990
991 /*
992 * If we haven't hit the head block or the log record header count,
993 * start looking again from the start of the physical log.
994 */
995 if (tail_blk > head_blk && found != count) {
996 for (i = 0; i < (int) head_blk; i++) {
997 error = xlog_bread(log, i, 1, bp, &offset);
998 if (error)
999 goto out_error;
1000
1001 if (*(__be32 *)offset ==
1002 cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
1003 *wrapped = true;
1004 *rblk = i;
1005 *rhead = (struct xlog_rec_header *) offset;
1006 if (++found == count)
1007 break;
1008 }
1009 }
1010 }
1011
1012 return found;
1013
1014out_error:
1015 return error;
1016}
1017
1018/*
1019 * Check the log tail for torn writes. This is required when torn writes are
1020 * detected at the head and the head had to be walked back to a previous record.
1021 * The tail of the previous record must now be verified to ensure the torn
1022 * writes didn't corrupt the previous tail.
1023 *
1024 * Return an error if CRC verification fails as recovery cannot proceed.
1025 */
1026STATIC int
1027xlog_verify_tail(
1028 struct xlog *log,
1029 xfs_daddr_t head_blk,
1030 xfs_daddr_t tail_blk)
1031{
1032 struct xlog_rec_header *thead;
1033 struct xfs_buf *bp;
1034 xfs_daddr_t first_bad;
1035 int count;
1036 int error = 0;
1037 bool wrapped;
1038 xfs_daddr_t tmp_head;
1039
1040 bp = xlog_get_bp(log, 1);
1041 if (!bp)
1042 return -ENOMEM;
1043
1044 /*
1045 * Seek XLOG_MAX_ICLOGS + 1 records past the current tail record to get
1046 * a temporary head block that points after the last possible
1047 * concurrently written record of the tail.
1048 */
1049 count = xlog_seek_logrec_hdr(log, head_blk, tail_blk,
1050 XLOG_MAX_ICLOGS + 1, bp, &tmp_head, &thead,
1051 &wrapped);
1052 if (count < 0) {
1053 error = count;
1054 goto out;
1055 }
1056
1057 /*
1058 * If the call above didn't find XLOG_MAX_ICLOGS + 1 records, we ran
1059 * into the actual log head. tmp_head points to the start of the record
1060 * so update it to the actual head block.
1061 */
1062 if (count < XLOG_MAX_ICLOGS + 1)
1063 tmp_head = head_blk;
1064
1065 /*
1066 * We now have a tail and temporary head block that covers at least
1067 * XLOG_MAX_ICLOGS records from the tail. We need to verify that these
1068 * records were completely written. Run a CRC verification pass from
1069 * tail to head and return the result.
1070 */
1071 error = xlog_do_recovery_pass(log, tmp_head, tail_blk,
1072 XLOG_RECOVER_CRCPASS, &first_bad);
1073
1074out:
1075 xlog_put_bp(bp);
1076 return error;
1077}
1078
1079/*
1080 * Detect and trim torn writes from the head of the log.
1081 *
1082 * Storage without sector atomicity guarantees can result in torn writes in the
1083 * log in the event of a crash. Our only means to detect this scenario is via
1084 * CRC verification. While we can't always be certain that CRC verification
1085 * failure is due to a torn write vs. an unrelated corruption, we do know that
1086 * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1087 * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1088 * the log and treat failures in this range as torn writes as a matter of
1089 * policy. In the event of CRC failure, the head is walked back to the last good
1090 * record in the log and the tail is updated from that record and verified.
1091 */
1092STATIC int
1093xlog_verify_head(
1094 struct xlog *log,
1095 xfs_daddr_t *head_blk, /* in/out: unverified head */
1096 xfs_daddr_t *tail_blk, /* out: tail block */
1097 struct xfs_buf *bp,
1098 xfs_daddr_t *rhead_blk, /* start blk of last record */
1099 struct xlog_rec_header **rhead, /* ptr to last record */
1100 bool *wrapped) /* last rec. wraps phys. log */
1101{
1102 struct xlog_rec_header *tmp_rhead;
1103 struct xfs_buf *tmp_bp;
1104 xfs_daddr_t first_bad;
1105 xfs_daddr_t tmp_rhead_blk;
1106 int found;
1107 int error;
1108 bool tmp_wrapped;
1109
1110 /*
1111 * Check the head of the log for torn writes. Search backwards from the
1112 * head until we hit the tail or the maximum number of log record I/Os
1113 * that could have been in flight at one time. Use a temporary buffer so
1114 * we don't trash the rhead/bp pointers from the caller.
1115 */
1116 tmp_bp = xlog_get_bp(log, 1);
1117 if (!tmp_bp)
1118 return -ENOMEM;
1119 error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1120 XLOG_MAX_ICLOGS, tmp_bp, &tmp_rhead_blk,
1121 &tmp_rhead, &tmp_wrapped);
1122 xlog_put_bp(tmp_bp);
1123 if (error < 0)
1124 return error;
1125
1126 /*
1127 * Now run a CRC verification pass over the records starting at the
1128 * block found above to the current head. If a CRC failure occurs, the
1129 * log block of the first bad record is saved in first_bad.
1130 */
1131 error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1132 XLOG_RECOVER_CRCPASS, &first_bad);
1133 if (error == -EFSBADCRC) {
1134 /*
1135 * We've hit a potential torn write. Reset the error and warn
1136 * about it.
1137 */
1138 error = 0;
1139 xfs_warn(log->l_mp,
1140"Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1141 first_bad, *head_blk);
1142
1143 /*
1144 * Get the header block and buffer pointer for the last good
1145 * record before the bad record.
1146 *
1147 * Note that xlog_find_tail() clears the blocks at the new head
1148 * (i.e., the records with invalid CRC) if the cycle number
1149 * matches the the current cycle.
1150 */
1151 found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1, bp,
1152 rhead_blk, rhead, wrapped);
1153 if (found < 0)
1154 return found;
1155 if (found == 0) /* XXX: right thing to do here? */
1156 return -EIO;
1157
1158 /*
1159 * Reset the head block to the starting block of the first bad
1160 * log record and set the tail block based on the last good
1161 * record.
1162 *
1163 * Bail out if the updated head/tail match as this indicates
1164 * possible corruption outside of the acceptable
1165 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1166 */
1167 *head_blk = first_bad;
1168 *tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1169 if (*head_blk == *tail_blk) {
1170 ASSERT(0);
1171 return 0;
1172 }
1173
1174 /*
1175 * Now verify the tail based on the updated head. This is
1176 * required because the torn writes trimmed from the head could
1177 * have been written over the tail of a previous record. Return
1178 * any errors since recovery cannot proceed if the tail is
1179 * corrupt.
1180 *
1181 * XXX: This leaves a gap in truly robust protection from torn
1182 * writes in the log. If the head is behind the tail, the tail
1183 * pushes forward to create some space and then a crash occurs
1184 * causing the writes into the previous record's tail region to
1185 * tear, log recovery isn't able to recover.
1186 *
1187 * How likely is this to occur? If possible, can we do something
1188 * more intelligent here? Is it safe to push the tail forward if
1189 * we can determine that the tail is within the range of the
1190 * torn write (e.g., the kernel can only overwrite the tail if
1191 * it has actually been pushed forward)? Alternatively, could we
1192 * somehow prevent this condition at runtime?
1193 */
1194 error = xlog_verify_tail(log, *head_blk, *tail_blk);
1195 }
1196
1197 return error;
1198}
1199
1200/*
1201 * Check whether the head of the log points to an unmount record. In other
1202 * words, determine whether the log is clean. If so, update the in-core state
1203 * appropriately.
1204 */
1205static int
1206xlog_check_unmount_rec(
1207 struct xlog *log,
1208 xfs_daddr_t *head_blk,
1209 xfs_daddr_t *tail_blk,
1210 struct xlog_rec_header *rhead,
1211 xfs_daddr_t rhead_blk,
1212 struct xfs_buf *bp,
1213 bool *clean)
1214{
1215 struct xlog_op_header *op_head;
1216 xfs_daddr_t umount_data_blk;
1217 xfs_daddr_t after_umount_blk;
1218 int hblks;
1219 int error;
1220 char *offset;
1221
1222 *clean = false;
1223
1224 /*
1225 * Look for unmount record. If we find it, then we know there was a
1226 * clean unmount. Since 'i' could be the last block in the physical
1227 * log, we convert to a log block before comparing to the head_blk.
1228 *
1229 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1230 * below. We won't want to clear the unmount record if there is one, so
1231 * we pass the lsn of the unmount record rather than the block after it.
1232 */
1233 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1234 int h_size = be32_to_cpu(rhead->h_size);
1235 int h_version = be32_to_cpu(rhead->h_version);
1236
1237 if ((h_version & XLOG_VERSION_2) &&
1238 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1239 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1240 if (h_size % XLOG_HEADER_CYCLE_SIZE)
1241 hblks++;
1242 } else {
1243 hblks = 1;
1244 }
1245 } else {
1246 hblks = 1;
1247 }
1248 after_umount_blk = rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len));
1249 after_umount_blk = do_mod(after_umount_blk, log->l_logBBsize);
1250 if (*head_blk == after_umount_blk &&
1251 be32_to_cpu(rhead->h_num_logops) == 1) {
1252 umount_data_blk = rhead_blk + hblks;
1253 umount_data_blk = do_mod(umount_data_blk, log->l_logBBsize);
1254 error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
1255 if (error)
1256 return error;
1257
1258 op_head = (struct xlog_op_header *)offset;
1259 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1260 /*
1261 * Set tail and last sync so that newly written log
1262 * records will point recovery to after the current
1263 * unmount record.
1264 */
1265 xlog_assign_atomic_lsn(&log->l_tail_lsn,
1266 log->l_curr_cycle, after_umount_blk);
1267 xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1268 log->l_curr_cycle, after_umount_blk);
1269 *tail_blk = after_umount_blk;
1270
1271 *clean = true;
1272 }
1273 }
1274
1275 return 0;
1276}
1277
1278static void
1279xlog_set_state(
1280 struct xlog *log,
1281 xfs_daddr_t head_blk,
1282 struct xlog_rec_header *rhead,
1283 xfs_daddr_t rhead_blk,
1284 bool bump_cycle)
1285{
1286 /*
1287 * Reset log values according to the state of the log when we
1288 * crashed. In the case where head_blk == 0, we bump curr_cycle
1289 * one because the next write starts a new cycle rather than
1290 * continuing the cycle of the last good log record. At this
1291 * point we have guaranteed that all partial log records have been
1292 * accounted for. Therefore, we know that the last good log record
1293 * written was complete and ended exactly on the end boundary
1294 * of the physical log.
1295 */
1296 log->l_prev_block = rhead_blk;
1297 log->l_curr_block = (int)head_blk;
1298 log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1299 if (bump_cycle)
1300 log->l_curr_cycle++;
1301 atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1302 atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1303 xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1304 BBTOB(log->l_curr_block));
1305 xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1306 BBTOB(log->l_curr_block));
1307}
1308
1309/*
1310 * Find the sync block number or the tail of the log.
1311 *
1312 * This will be the block number of the last record to have its
1313 * associated buffers synced to disk. Every log record header has
1314 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
1315 * to get a sync block number. The only concern is to figure out which
1316 * log record header to believe.
1317 *
1318 * The following algorithm uses the log record header with the largest
1319 * lsn. The entire log record does not need to be valid. We only care
1320 * that the header is valid.
1321 *
1322 * We could speed up search by using current head_blk buffer, but it is not
1323 * available.
1324 */
1325STATIC int
1326xlog_find_tail(
1327 struct xlog *log,
1328 xfs_daddr_t *head_blk,
1329 xfs_daddr_t *tail_blk)
1330{
1331 xlog_rec_header_t *rhead;
1332 char *offset = NULL;
1333 xfs_buf_t *bp;
1334 int error;
1335 xfs_daddr_t rhead_blk;
1336 xfs_lsn_t tail_lsn;
1337 bool wrapped = false;
1338 bool clean = false;
1339
1340 /*
1341 * Find previous log record
1342 */
1343 if ((error = xlog_find_head(log, head_blk)))
1344 return error;
1345 ASSERT(*head_blk < INT_MAX);
1346
1347 bp = xlog_get_bp(log, 1);
1348 if (!bp)
1349 return -ENOMEM;
1350 if (*head_blk == 0) { /* special case */
1351 error = xlog_bread(log, 0, 1, bp, &offset);
1352 if (error)
1353 goto done;
1354
1355 if (xlog_get_cycle(offset) == 0) {
1356 *tail_blk = 0;
1357 /* leave all other log inited values alone */
1358 goto done;
1359 }
1360 }
1361
1362 /*
1363 * Search backwards through the log looking for the log record header
1364 * block. This wraps all the way back around to the head so something is
1365 * seriously wrong if we can't find it.
1366 */
1367 error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, bp,
1368 &rhead_blk, &rhead, &wrapped);
1369 if (error < 0)
1370 return error;
1371 if (!error) {
1372 xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1373 return -EIO;
1374 }
1375 *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1376
1377 /*
1378 * Set the log state based on the current head record.
1379 */
1380 xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1381 tail_lsn = atomic64_read(&log->l_tail_lsn);
1382
1383 /*
1384 * Look for an unmount record at the head of the log. This sets the log
1385 * state to determine whether recovery is necessary.
1386 */
1387 error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1388 rhead_blk, bp, &clean);
1389 if (error)
1390 goto done;
1391
1392 /*
1393 * Verify the log head if the log is not clean (e.g., we have anything
1394 * but an unmount record at the head). This uses CRC verification to
1395 * detect and trim torn writes. If discovered, CRC failures are
1396 * considered torn writes and the log head is trimmed accordingly.
1397 *
1398 * Note that we can only run CRC verification when the log is dirty
1399 * because there's no guarantee that the log data behind an unmount
1400 * record is compatible with the current architecture.
1401 */
1402 if (!clean) {
1403 xfs_daddr_t orig_head = *head_blk;
1404
1405 error = xlog_verify_head(log, head_blk, tail_blk, bp,
1406 &rhead_blk, &rhead, &wrapped);
1407 if (error)
1408 goto done;
1409
1410 /* update in-core state again if the head changed */
1411 if (*head_blk != orig_head) {
1412 xlog_set_state(log, *head_blk, rhead, rhead_blk,
1413 wrapped);
1414 tail_lsn = atomic64_read(&log->l_tail_lsn);
1415 error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1416 rhead, rhead_blk, bp,
1417 &clean);
1418 if (error)
1419 goto done;
1420 }
1421 }
1422
1423 /*
1424 * Note that the unmount was clean. If the unmount was not clean, we
1425 * need to know this to rebuild the superblock counters from the perag
1426 * headers if we have a filesystem using non-persistent counters.
1427 */
1428 if (clean)
1429 log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1430
1431 /*
1432 * Make sure that there are no blocks in front of the head
1433 * with the same cycle number as the head. This can happen
1434 * because we allow multiple outstanding log writes concurrently,
1435 * and the later writes might make it out before earlier ones.
1436 *
1437 * We use the lsn from before modifying it so that we'll never
1438 * overwrite the unmount record after a clean unmount.
1439 *
1440 * Do this only if we are going to recover the filesystem
1441 *
1442 * NOTE: This used to say "if (!readonly)"
1443 * However on Linux, we can & do recover a read-only filesystem.
1444 * We only skip recovery if NORECOVERY is specified on mount,
1445 * in which case we would not be here.
1446 *
1447 * But... if the -device- itself is readonly, just skip this.
1448 * We can't recover this device anyway, so it won't matter.
1449 */
1450 if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
1451 error = xlog_clear_stale_blocks(log, tail_lsn);
1452
1453done:
1454 xlog_put_bp(bp);
1455
1456 if (error)
1457 xfs_warn(log->l_mp, "failed to locate log tail");
1458 return error;
1459}
1460
1461/*
1462 * Is the log zeroed at all?
1463 *
1464 * The last binary search should be changed to perform an X block read
1465 * once X becomes small enough. You can then search linearly through
1466 * the X blocks. This will cut down on the number of reads we need to do.
1467 *
1468 * If the log is partially zeroed, this routine will pass back the blkno
1469 * of the first block with cycle number 0. It won't have a complete LR
1470 * preceding it.
1471 *
1472 * Return:
1473 * 0 => the log is completely written to
1474 * 1 => use *blk_no as the first block of the log
1475 * <0 => error has occurred
1476 */
1477STATIC int
1478xlog_find_zeroed(
1479 struct xlog *log,
1480 xfs_daddr_t *blk_no)
1481{
1482 xfs_buf_t *bp;
1483 char *offset;
1484 uint first_cycle, last_cycle;
1485 xfs_daddr_t new_blk, last_blk, start_blk;
1486 xfs_daddr_t num_scan_bblks;
1487 int error, log_bbnum = log->l_logBBsize;
1488
1489 *blk_no = 0;
1490
1491 /* check totally zeroed log */
1492 bp = xlog_get_bp(log, 1);
1493 if (!bp)
1494 return -ENOMEM;
1495 error = xlog_bread(log, 0, 1, bp, &offset);
1496 if (error)
1497 goto bp_err;
1498
1499 first_cycle = xlog_get_cycle(offset);
1500 if (first_cycle == 0) { /* completely zeroed log */
1501 *blk_no = 0;
1502 xlog_put_bp(bp);
1503 return 1;
1504 }
1505
1506 /* check partially zeroed log */
1507 error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1508 if (error)
1509 goto bp_err;
1510
1511 last_cycle = xlog_get_cycle(offset);
1512 if (last_cycle != 0) { /* log completely written to */
1513 xlog_put_bp(bp);
1514 return 0;
1515 } else if (first_cycle != 1) {
1516 /*
1517 * If the cycle of the last block is zero, the cycle of
1518 * the first block must be 1. If it's not, maybe we're
1519 * not looking at a log... Bail out.
1520 */
1521 xfs_warn(log->l_mp,
1522 "Log inconsistent or not a log (last==0, first!=1)");
1523 error = -EINVAL;
1524 goto bp_err;
1525 }
1526
1527 /* we have a partially zeroed log */
1528 last_blk = log_bbnum-1;
1529 if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1530 goto bp_err;
1531
1532 /*
1533 * Validate the answer. Because there is no way to guarantee that
1534 * the entire log is made up of log records which are the same size,
1535 * we scan over the defined maximum blocks. At this point, the maximum
1536 * is not chosen to mean anything special. XXXmiken
1537 */
1538 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1539 ASSERT(num_scan_bblks <= INT_MAX);
1540
1541 if (last_blk < num_scan_bblks)
1542 num_scan_bblks = last_blk;
1543 start_blk = last_blk - num_scan_bblks;
1544
1545 /*
1546 * We search for any instances of cycle number 0 that occur before
1547 * our current estimate of the head. What we're trying to detect is
1548 * 1 ... | 0 | 1 | 0...
1549 * ^ binary search ends here
1550 */
1551 if ((error = xlog_find_verify_cycle(log, start_blk,
1552 (int)num_scan_bblks, 0, &new_blk)))
1553 goto bp_err;
1554 if (new_blk != -1)
1555 last_blk = new_blk;
1556
1557 /*
1558 * Potentially backup over partial log record write. We don't need
1559 * to search the end of the log because we know it is zero.
1560 */
1561 error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1562 if (error == 1)
1563 error = -EIO;
1564 if (error)
1565 goto bp_err;
1566
1567 *blk_no = last_blk;
1568bp_err:
1569 xlog_put_bp(bp);
1570 if (error)
1571 return error;
1572 return 1;
1573}
1574
1575/*
1576 * These are simple subroutines used by xlog_clear_stale_blocks() below
1577 * to initialize a buffer full of empty log record headers and write
1578 * them into the log.
1579 */
1580STATIC void
1581xlog_add_record(
1582 struct xlog *log,
1583 char *buf,
1584 int cycle,
1585 int block,
1586 int tail_cycle,
1587 int tail_block)
1588{
1589 xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
1590
1591 memset(buf, 0, BBSIZE);
1592 recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1593 recp->h_cycle = cpu_to_be32(cycle);
1594 recp->h_version = cpu_to_be32(
1595 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1596 recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1597 recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1598 recp->h_fmt = cpu_to_be32(XLOG_FMT);
1599 memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1600}
1601
1602STATIC int
1603xlog_write_log_records(
1604 struct xlog *log,
1605 int cycle,
1606 int start_block,
1607 int blocks,
1608 int tail_cycle,
1609 int tail_block)
1610{
1611 char *offset;
1612 xfs_buf_t *bp;
1613 int balign, ealign;
1614 int sectbb = log->l_sectBBsize;
1615 int end_block = start_block + blocks;
1616 int bufblks;
1617 int error = 0;
1618 int i, j = 0;
1619
1620 /*
1621 * Greedily allocate a buffer big enough to handle the full
1622 * range of basic blocks to be written. If that fails, try
1623 * a smaller size. We need to be able to write at least a
1624 * log sector, or we're out of luck.
1625 */
1626 bufblks = 1 << ffs(blocks);
1627 while (bufblks > log->l_logBBsize)
1628 bufblks >>= 1;
1629 while (!(bp = xlog_get_bp(log, bufblks))) {
1630 bufblks >>= 1;
1631 if (bufblks < sectbb)
1632 return -ENOMEM;
1633 }
1634
1635 /* We may need to do a read at the start to fill in part of
1636 * the buffer in the starting sector not covered by the first
1637 * write below.
1638 */
1639 balign = round_down(start_block, sectbb);
1640 if (balign != start_block) {
1641 error = xlog_bread_noalign(log, start_block, 1, bp);
1642 if (error)
1643 goto out_put_bp;
1644
1645 j = start_block - balign;
1646 }
1647
1648 for (i = start_block; i < end_block; i += bufblks) {
1649 int bcount, endcount;
1650
1651 bcount = min(bufblks, end_block - start_block);
1652 endcount = bcount - j;
1653
1654 /* We may need to do a read at the end to fill in part of
1655 * the buffer in the final sector not covered by the write.
1656 * If this is the same sector as the above read, skip it.
1657 */
1658 ealign = round_down(end_block, sectbb);
1659 if (j == 0 && (start_block + endcount > ealign)) {
1660 offset = bp->b_addr + BBTOB(ealign - start_block);
1661 error = xlog_bread_offset(log, ealign, sectbb,
1662 bp, offset);
1663 if (error)
1664 break;
1665
1666 }
1667
1668 offset = xlog_align(log, start_block, endcount, bp);
1669 for (; j < endcount; j++) {
1670 xlog_add_record(log, offset, cycle, i+j,
1671 tail_cycle, tail_block);
1672 offset += BBSIZE;
1673 }
1674 error = xlog_bwrite(log, start_block, endcount, bp);
1675 if (error)
1676 break;
1677 start_block += endcount;
1678 j = 0;
1679 }
1680
1681 out_put_bp:
1682 xlog_put_bp(bp);
1683 return error;
1684}
1685
1686/*
1687 * This routine is called to blow away any incomplete log writes out
1688 * in front of the log head. We do this so that we won't become confused
1689 * if we come up, write only a little bit more, and then crash again.
1690 * If we leave the partial log records out there, this situation could
1691 * cause us to think those partial writes are valid blocks since they
1692 * have the current cycle number. We get rid of them by overwriting them
1693 * with empty log records with the old cycle number rather than the
1694 * current one.
1695 *
1696 * The tail lsn is passed in rather than taken from
1697 * the log so that we will not write over the unmount record after a
1698 * clean unmount in a 512 block log. Doing so would leave the log without
1699 * any valid log records in it until a new one was written. If we crashed
1700 * during that time we would not be able to recover.
1701 */
1702STATIC int
1703xlog_clear_stale_blocks(
1704 struct xlog *log,
1705 xfs_lsn_t tail_lsn)
1706{
1707 int tail_cycle, head_cycle;
1708 int tail_block, head_block;
1709 int tail_distance, max_distance;
1710 int distance;
1711 int error;
1712
1713 tail_cycle = CYCLE_LSN(tail_lsn);
1714 tail_block = BLOCK_LSN(tail_lsn);
1715 head_cycle = log->l_curr_cycle;
1716 head_block = log->l_curr_block;
1717
1718 /*
1719 * Figure out the distance between the new head of the log
1720 * and the tail. We want to write over any blocks beyond the
1721 * head that we may have written just before the crash, but
1722 * we don't want to overwrite the tail of the log.
1723 */
1724 if (head_cycle == tail_cycle) {
1725 /*
1726 * The tail is behind the head in the physical log,
1727 * so the distance from the head to the tail is the
1728 * distance from the head to the end of the log plus
1729 * the distance from the beginning of the log to the
1730 * tail.
1731 */
1732 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1733 XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1734 XFS_ERRLEVEL_LOW, log->l_mp);
1735 return -EFSCORRUPTED;
1736 }
1737 tail_distance = tail_block + (log->l_logBBsize - head_block);
1738 } else {
1739 /*
1740 * The head is behind the tail in the physical log,
1741 * so the distance from the head to the tail is just
1742 * the tail block minus the head block.
1743 */
1744 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1745 XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1746 XFS_ERRLEVEL_LOW, log->l_mp);
1747 return -EFSCORRUPTED;
1748 }
1749 tail_distance = tail_block - head_block;
1750 }
1751
1752 /*
1753 * If the head is right up against the tail, we can't clear
1754 * anything.
1755 */
1756 if (tail_distance <= 0) {
1757 ASSERT(tail_distance == 0);
1758 return 0;
1759 }
1760
1761 max_distance = XLOG_TOTAL_REC_SHIFT(log);
1762 /*
1763 * Take the smaller of the maximum amount of outstanding I/O
1764 * we could have and the distance to the tail to clear out.
1765 * We take the smaller so that we don't overwrite the tail and
1766 * we don't waste all day writing from the head to the tail
1767 * for no reason.
1768 */
1769 max_distance = MIN(max_distance, tail_distance);
1770
1771 if ((head_block + max_distance) <= log->l_logBBsize) {
1772 /*
1773 * We can stomp all the blocks we need to without
1774 * wrapping around the end of the log. Just do it
1775 * in a single write. Use the cycle number of the
1776 * current cycle minus one so that the log will look like:
1777 * n ... | n - 1 ...
1778 */
1779 error = xlog_write_log_records(log, (head_cycle - 1),
1780 head_block, max_distance, tail_cycle,
1781 tail_block);
1782 if (error)
1783 return error;
1784 } else {
1785 /*
1786 * We need to wrap around the end of the physical log in
1787 * order to clear all the blocks. Do it in two separate
1788 * I/Os. The first write should be from the head to the
1789 * end of the physical log, and it should use the current
1790 * cycle number minus one just like above.
1791 */
1792 distance = log->l_logBBsize - head_block;
1793 error = xlog_write_log_records(log, (head_cycle - 1),
1794 head_block, distance, tail_cycle,
1795 tail_block);
1796
1797 if (error)
1798 return error;
1799
1800 /*
1801 * Now write the blocks at the start of the physical log.
1802 * This writes the remainder of the blocks we want to clear.
1803 * It uses the current cycle number since we're now on the
1804 * same cycle as the head so that we get:
1805 * n ... n ... | n - 1 ...
1806 * ^^^^^ blocks we're writing
1807 */
1808 distance = max_distance - (log->l_logBBsize - head_block);
1809 error = xlog_write_log_records(log, head_cycle, 0, distance,
1810 tail_cycle, tail_block);
1811 if (error)
1812 return error;
1813 }
1814
1815 return 0;
1816}
1817
1818/******************************************************************************
1819 *
1820 * Log recover routines
1821 *
1822 ******************************************************************************
1823 */
1824
1825/*
1826 * Sort the log items in the transaction.
1827 *
1828 * The ordering constraints are defined by the inode allocation and unlink
1829 * behaviour. The rules are:
1830 *
1831 * 1. Every item is only logged once in a given transaction. Hence it
1832 * represents the last logged state of the item. Hence ordering is
1833 * dependent on the order in which operations need to be performed so
1834 * required initial conditions are always met.
1835 *
1836 * 2. Cancelled buffers are recorded in pass 1 in a separate table and
1837 * there's nothing to replay from them so we can simply cull them
1838 * from the transaction. However, we can't do that until after we've
1839 * replayed all the other items because they may be dependent on the
1840 * cancelled buffer and replaying the cancelled buffer can remove it
1841 * form the cancelled buffer table. Hence they have tobe done last.
1842 *
1843 * 3. Inode allocation buffers must be replayed before inode items that
1844 * read the buffer and replay changes into it. For filesystems using the
1845 * ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1846 * treated the same as inode allocation buffers as they create and
1847 * initialise the buffers directly.
1848 *
1849 * 4. Inode unlink buffers must be replayed after inode items are replayed.
1850 * This ensures that inodes are completely flushed to the inode buffer
1851 * in a "free" state before we remove the unlinked inode list pointer.
1852 *
1853 * Hence the ordering needs to be inode allocation buffers first, inode items
1854 * second, inode unlink buffers third and cancelled buffers last.
1855 *
1856 * But there's a problem with that - we can't tell an inode allocation buffer
1857 * apart from a regular buffer, so we can't separate them. We can, however,
1858 * tell an inode unlink buffer from the others, and so we can separate them out
1859 * from all the other buffers and move them to last.
1860 *
1861 * Hence, 4 lists, in order from head to tail:
1862 * - buffer_list for all buffers except cancelled/inode unlink buffers
1863 * - item_list for all non-buffer items
1864 * - inode_buffer_list for inode unlink buffers
1865 * - cancel_list for the cancelled buffers
1866 *
1867 * Note that we add objects to the tail of the lists so that first-to-last
1868 * ordering is preserved within the lists. Adding objects to the head of the
1869 * list means when we traverse from the head we walk them in last-to-first
1870 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1871 * but for all other items there may be specific ordering that we need to
1872 * preserve.
1873 */
1874STATIC int
1875xlog_recover_reorder_trans(
1876 struct xlog *log,
1877 struct xlog_recover *trans,
1878 int pass)
1879{
1880 xlog_recover_item_t *item, *n;
1881 int error = 0;
1882 LIST_HEAD(sort_list);
1883 LIST_HEAD(cancel_list);
1884 LIST_HEAD(buffer_list);
1885 LIST_HEAD(inode_buffer_list);
1886 LIST_HEAD(inode_list);
1887
1888 list_splice_init(&trans->r_itemq, &sort_list);
1889 list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1890 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1891
1892 switch (ITEM_TYPE(item)) {
1893 case XFS_LI_ICREATE:
1894 list_move_tail(&item->ri_list, &buffer_list);
1895 break;
1896 case XFS_LI_BUF:
1897 if (buf_f->blf_flags & XFS_BLF_CANCEL) {
1898 trace_xfs_log_recover_item_reorder_head(log,
1899 trans, item, pass);
1900 list_move(&item->ri_list, &cancel_list);
1901 break;
1902 }
1903 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
1904 list_move(&item->ri_list, &inode_buffer_list);
1905 break;
1906 }
1907 list_move_tail(&item->ri_list, &buffer_list);
1908 break;
1909 case XFS_LI_INODE:
1910 case XFS_LI_DQUOT:
1911 case XFS_LI_QUOTAOFF:
1912 case XFS_LI_EFD:
1913 case XFS_LI_EFI:
1914 trace_xfs_log_recover_item_reorder_tail(log,
1915 trans, item, pass);
1916 list_move_tail(&item->ri_list, &inode_list);
1917 break;
1918 default:
1919 xfs_warn(log->l_mp,
1920 "%s: unrecognized type of log operation",
1921 __func__);
1922 ASSERT(0);
1923 /*
1924 * return the remaining items back to the transaction
1925 * item list so they can be freed in caller.
1926 */
1927 if (!list_empty(&sort_list))
1928 list_splice_init(&sort_list, &trans->r_itemq);
1929 error = -EIO;
1930 goto out;
1931 }
1932 }
1933out:
1934 ASSERT(list_empty(&sort_list));
1935 if (!list_empty(&buffer_list))
1936 list_splice(&buffer_list, &trans->r_itemq);
1937 if (!list_empty(&inode_list))
1938 list_splice_tail(&inode_list, &trans->r_itemq);
1939 if (!list_empty(&inode_buffer_list))
1940 list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1941 if (!list_empty(&cancel_list))
1942 list_splice_tail(&cancel_list, &trans->r_itemq);
1943 return error;
1944}
1945
1946/*
1947 * Build up the table of buf cancel records so that we don't replay
1948 * cancelled data in the second pass. For buffer records that are
1949 * not cancel records, there is nothing to do here so we just return.
1950 *
1951 * If we get a cancel record which is already in the table, this indicates
1952 * that the buffer was cancelled multiple times. In order to ensure
1953 * that during pass 2 we keep the record in the table until we reach its
1954 * last occurrence in the log, we keep a reference count in the cancel
1955 * record in the table to tell us how many times we expect to see this
1956 * record during the second pass.
1957 */
1958STATIC int
1959xlog_recover_buffer_pass1(
1960 struct xlog *log,
1961 struct xlog_recover_item *item)
1962{
1963 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1964 struct list_head *bucket;
1965 struct xfs_buf_cancel *bcp;
1966
1967 /*
1968 * If this isn't a cancel buffer item, then just return.
1969 */
1970 if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1971 trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1972 return 0;
1973 }
1974
1975 /*
1976 * Insert an xfs_buf_cancel record into the hash table of them.
1977 * If there is already an identical record, bump its reference count.
1978 */
1979 bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1980 list_for_each_entry(bcp, bucket, bc_list) {
1981 if (bcp->bc_blkno == buf_f->blf_blkno &&
1982 bcp->bc_len == buf_f->blf_len) {
1983 bcp->bc_refcount++;
1984 trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1985 return 0;
1986 }
1987 }
1988
1989 bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
1990 bcp->bc_blkno = buf_f->blf_blkno;
1991 bcp->bc_len = buf_f->blf_len;
1992 bcp->bc_refcount = 1;
1993 list_add_tail(&bcp->bc_list, bucket);
1994
1995 trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1996 return 0;
1997}
1998
1999/*
2000 * Check to see whether the buffer being recovered has a corresponding
2001 * entry in the buffer cancel record table. If it is, return the cancel
2002 * buffer structure to the caller.
2003 */
2004STATIC struct xfs_buf_cancel *
2005xlog_peek_buffer_cancelled(
2006 struct xlog *log,
2007 xfs_daddr_t blkno,
2008 uint len,
2009 ushort flags)
2010{
2011 struct list_head *bucket;
2012 struct xfs_buf_cancel *bcp;
2013
2014 if (!log->l_buf_cancel_table) {
2015 /* empty table means no cancelled buffers in the log */
2016 ASSERT(!(flags & XFS_BLF_CANCEL));
2017 return NULL;
2018 }
2019
2020 bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
2021 list_for_each_entry(bcp, bucket, bc_list) {
2022 if (bcp->bc_blkno == blkno && bcp->bc_len == len)
2023 return bcp;
2024 }
2025
2026 /*
2027 * We didn't find a corresponding entry in the table, so return 0 so
2028 * that the buffer is NOT cancelled.
2029 */
2030 ASSERT(!(flags & XFS_BLF_CANCEL));
2031 return NULL;
2032}
2033
2034/*
2035 * If the buffer is being cancelled then return 1 so that it will be cancelled,
2036 * otherwise return 0. If the buffer is actually a buffer cancel item
2037 * (XFS_BLF_CANCEL is set), then decrement the refcount on the entry in the
2038 * table and remove it from the table if this is the last reference.
2039 *
2040 * We remove the cancel record from the table when we encounter its last
2041 * occurrence in the log so that if the same buffer is re-used again after its
2042 * last cancellation we actually replay the changes made at that point.
2043 */
2044STATIC int
2045xlog_check_buffer_cancelled(
2046 struct xlog *log,
2047 xfs_daddr_t blkno,
2048 uint len,
2049 ushort flags)
2050{
2051 struct xfs_buf_cancel *bcp;
2052
2053 bcp = xlog_peek_buffer_cancelled(log, blkno, len, flags);
2054 if (!bcp)
2055 return 0;
2056
2057 /*
2058 * We've go a match, so return 1 so that the recovery of this buffer
2059 * is cancelled. If this buffer is actually a buffer cancel log
2060 * item, then decrement the refcount on the one in the table and
2061 * remove it if this is the last reference.
2062 */
2063 if (flags & XFS_BLF_CANCEL) {
2064 if (--bcp->bc_refcount == 0) {
2065 list_del(&bcp->bc_list);
2066 kmem_free(bcp);
2067 }
2068 }
2069 return 1;
2070}
2071
2072/*
2073 * Perform recovery for a buffer full of inodes. In these buffers, the only
2074 * data which should be recovered is that which corresponds to the
2075 * di_next_unlinked pointers in the on disk inode structures. The rest of the
2076 * data for the inodes is always logged through the inodes themselves rather
2077 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
2078 *
2079 * The only time when buffers full of inodes are fully recovered is when the
2080 * buffer is full of newly allocated inodes. In this case the buffer will
2081 * not be marked as an inode buffer and so will be sent to
2082 * xlog_recover_do_reg_buffer() below during recovery.
2083 */
2084STATIC int
2085xlog_recover_do_inode_buffer(
2086 struct xfs_mount *mp,
2087 xlog_recover_item_t *item,
2088 struct xfs_buf *bp,
2089 xfs_buf_log_format_t *buf_f)
2090{
2091 int i;
2092 int item_index = 0;
2093 int bit = 0;
2094 int nbits = 0;
2095 int reg_buf_offset = 0;
2096 int reg_buf_bytes = 0;
2097 int next_unlinked_offset;
2098 int inodes_per_buf;
2099 xfs_agino_t *logged_nextp;
2100 xfs_agino_t *buffer_nextp;
2101
2102 trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
2103
2104 /*
2105 * Post recovery validation only works properly on CRC enabled
2106 * filesystems.
2107 */
2108 if (xfs_sb_version_hascrc(&mp->m_sb))
2109 bp->b_ops = &xfs_inode_buf_ops;
2110
2111 inodes_per_buf = BBTOB(bp->b_io_length) >> mp->m_sb.sb_inodelog;
2112 for (i = 0; i < inodes_per_buf; i++) {
2113 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
2114 offsetof(xfs_dinode_t, di_next_unlinked);
2115
2116 while (next_unlinked_offset >=
2117 (reg_buf_offset + reg_buf_bytes)) {
2118 /*
2119 * The next di_next_unlinked field is beyond
2120 * the current logged region. Find the next
2121 * logged region that contains or is beyond
2122 * the current di_next_unlinked field.
2123 */
2124 bit += nbits;
2125 bit = xfs_next_bit(buf_f->blf_data_map,
2126 buf_f->blf_map_size, bit);
2127
2128 /*
2129 * If there are no more logged regions in the
2130 * buffer, then we're done.
2131 */
2132 if (bit == -1)
2133 return 0;
2134
2135 nbits = xfs_contig_bits(buf_f->blf_data_map,
2136 buf_f->blf_map_size, bit);
2137 ASSERT(nbits > 0);
2138 reg_buf_offset = bit << XFS_BLF_SHIFT;
2139 reg_buf_bytes = nbits << XFS_BLF_SHIFT;
2140 item_index++;
2141 }
2142
2143 /*
2144 * If the current logged region starts after the current
2145 * di_next_unlinked field, then move on to the next
2146 * di_next_unlinked field.
2147 */
2148 if (next_unlinked_offset < reg_buf_offset)
2149 continue;
2150
2151 ASSERT(item->ri_buf[item_index].i_addr != NULL);
2152 ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
2153 ASSERT((reg_buf_offset + reg_buf_bytes) <=
2154 BBTOB(bp->b_io_length));
2155
2156 /*
2157 * The current logged region contains a copy of the
2158 * current di_next_unlinked field. Extract its value
2159 * and copy it to the buffer copy.
2160 */
2161 logged_nextp = item->ri_buf[item_index].i_addr +
2162 next_unlinked_offset - reg_buf_offset;
2163 if (unlikely(*logged_nextp == 0)) {
2164 xfs_alert(mp,
2165 "Bad inode buffer log record (ptr = 0x%p, bp = 0x%p). "
2166 "Trying to replay bad (0) inode di_next_unlinked field.",
2167 item, bp);
2168 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
2169 XFS_ERRLEVEL_LOW, mp);
2170 return -EFSCORRUPTED;
2171 }
2172
2173 buffer_nextp = xfs_buf_offset(bp, next_unlinked_offset);
2174 *buffer_nextp = *logged_nextp;
2175
2176 /*
2177 * If necessary, recalculate the CRC in the on-disk inode. We
2178 * have to leave the inode in a consistent state for whoever
2179 * reads it next....
2180 */
2181 xfs_dinode_calc_crc(mp,
2182 xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
2183
2184 }
2185
2186 return 0;
2187}
2188
2189/*
2190 * V5 filesystems know the age of the buffer on disk being recovered. We can
2191 * have newer objects on disk than we are replaying, and so for these cases we
2192 * don't want to replay the current change as that will make the buffer contents
2193 * temporarily invalid on disk.
2194 *
2195 * The magic number might not match the buffer type we are going to recover
2196 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags. Hence
2197 * extract the LSN of the existing object in the buffer based on it's current
2198 * magic number. If we don't recognise the magic number in the buffer, then
2199 * return a LSN of -1 so that the caller knows it was an unrecognised block and
2200 * so can recover the buffer.
2201 *
2202 * Note: we cannot rely solely on magic number matches to determine that the
2203 * buffer has a valid LSN - we also need to verify that it belongs to this
2204 * filesystem, so we need to extract the object's LSN and compare it to that
2205 * which we read from the superblock. If the UUIDs don't match, then we've got a
2206 * stale metadata block from an old filesystem instance that we need to recover
2207 * over the top of.
2208 */
2209static xfs_lsn_t
2210xlog_recover_get_buf_lsn(
2211 struct xfs_mount *mp,
2212 struct xfs_buf *bp)
2213{
2214 __uint32_t magic32;
2215 __uint16_t magic16;
2216 __uint16_t magicda;
2217 void *blk = bp->b_addr;
2218 uuid_t *uuid;
2219 xfs_lsn_t lsn = -1;
2220
2221 /* v4 filesystems always recover immediately */
2222 if (!xfs_sb_version_hascrc(&mp->m_sb))
2223 goto recover_immediately;
2224
2225 magic32 = be32_to_cpu(*(__be32 *)blk);
2226 switch (magic32) {
2227 case XFS_ABTB_CRC_MAGIC:
2228 case XFS_ABTC_CRC_MAGIC:
2229 case XFS_ABTB_MAGIC:
2230 case XFS_ABTC_MAGIC:
2231 case XFS_IBT_CRC_MAGIC:
2232 case XFS_IBT_MAGIC: {
2233 struct xfs_btree_block *btb = blk;
2234
2235 lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
2236 uuid = &btb->bb_u.s.bb_uuid;
2237 break;
2238 }
2239 case XFS_BMAP_CRC_MAGIC:
2240 case XFS_BMAP_MAGIC: {
2241 struct xfs_btree_block *btb = blk;
2242
2243 lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
2244 uuid = &btb->bb_u.l.bb_uuid;
2245 break;
2246 }
2247 case XFS_AGF_MAGIC:
2248 lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
2249 uuid = &((struct xfs_agf *)blk)->agf_uuid;
2250 break;
2251 case XFS_AGFL_MAGIC:
2252 lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
2253 uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
2254 break;
2255 case XFS_AGI_MAGIC:
2256 lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
2257 uuid = &((struct xfs_agi *)blk)->agi_uuid;
2258 break;
2259 case XFS_SYMLINK_MAGIC:
2260 lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
2261 uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
2262 break;
2263 case XFS_DIR3_BLOCK_MAGIC:
2264 case XFS_DIR3_DATA_MAGIC:
2265 case XFS_DIR3_FREE_MAGIC:
2266 lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
2267 uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
2268 break;
2269 case XFS_ATTR3_RMT_MAGIC:
2270 /*
2271 * Remote attr blocks are written synchronously, rather than
2272 * being logged. That means they do not contain a valid LSN
2273 * (i.e. transactionally ordered) in them, and hence any time we
2274 * see a buffer to replay over the top of a remote attribute
2275 * block we should simply do so.
2276 */
2277 goto recover_immediately;
2278 case XFS_SB_MAGIC:
2279 /*
2280 * superblock uuids are magic. We may or may not have a
2281 * sb_meta_uuid on disk, but it will be set in the in-core
2282 * superblock. We set the uuid pointer for verification
2283 * according to the superblock feature mask to ensure we check
2284 * the relevant UUID in the superblock.
2285 */
2286 lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
2287 if (xfs_sb_version_hasmetauuid(&mp->m_sb))
2288 uuid = &((struct xfs_dsb *)blk)->sb_meta_uuid;
2289 else
2290 uuid = &((struct xfs_dsb *)blk)->sb_uuid;
2291 break;
2292 default:
2293 break;
2294 }
2295
2296 if (lsn != (xfs_lsn_t)-1) {
2297 if (!uuid_equal(&mp->m_sb.sb_meta_uuid, uuid))
2298 goto recover_immediately;
2299 return lsn;
2300 }
2301
2302 magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
2303 switch (magicda) {
2304 case XFS_DIR3_LEAF1_MAGIC:
2305 case XFS_DIR3_LEAFN_MAGIC:
2306 case XFS_DA3_NODE_MAGIC:
2307 lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
2308 uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
2309 break;
2310 default:
2311 break;
2312 }
2313
2314 if (lsn != (xfs_lsn_t)-1) {
2315 if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
2316 goto recover_immediately;
2317 return lsn;
2318 }
2319
2320 /*
2321 * We do individual object checks on dquot and inode buffers as they
2322 * have their own individual LSN records. Also, we could have a stale
2323 * buffer here, so we have to at least recognise these buffer types.
2324 *
2325 * A notd complexity here is inode unlinked list processing - it logs
2326 * the inode directly in the buffer, but we don't know which inodes have
2327 * been modified, and there is no global buffer LSN. Hence we need to
2328 * recover all inode buffer types immediately. This problem will be
2329 * fixed by logical logging of the unlinked list modifications.
2330 */
2331 magic16 = be16_to_cpu(*(__be16 *)blk);
2332 switch (magic16) {
2333 case XFS_DQUOT_MAGIC:
2334 case XFS_DINODE_MAGIC:
2335 goto recover_immediately;
2336 default:
2337 break;
2338 }
2339
2340 /* unknown buffer contents, recover immediately */
2341
2342recover_immediately:
2343 return (xfs_lsn_t)-1;
2344
2345}
2346
2347/*
2348 * Validate the recovered buffer is of the correct type and attach the
2349 * appropriate buffer operations to them for writeback. Magic numbers are in a
2350 * few places:
2351 * the first 16 bits of the buffer (inode buffer, dquot buffer),
2352 * the first 32 bits of the buffer (most blocks),
2353 * inside a struct xfs_da_blkinfo at the start of the buffer.
2354 */
2355static void
2356xlog_recover_validate_buf_type(
2357 struct xfs_mount *mp,
2358 struct xfs_buf *bp,
2359 xfs_buf_log_format_t *buf_f)
2360{
2361 struct xfs_da_blkinfo *info = bp->b_addr;
2362 __uint32_t magic32;
2363 __uint16_t magic16;
2364 __uint16_t magicda;
2365
2366 /*
2367 * We can only do post recovery validation on items on CRC enabled
2368 * fielsystems as we need to know when the buffer was written to be able
2369 * to determine if we should have replayed the item. If we replay old
2370 * metadata over a newer buffer, then it will enter a temporarily
2371 * inconsistent state resulting in verification failures. Hence for now
2372 * just avoid the verification stage for non-crc filesystems
2373 */
2374 if (!xfs_sb_version_hascrc(&mp->m_sb))
2375 return;
2376
2377 magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
2378 magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
2379 magicda = be16_to_cpu(info->magic);
2380 switch (xfs_blft_from_flags(buf_f)) {
2381 case XFS_BLFT_BTREE_BUF:
2382 switch (magic32) {
2383 case XFS_ABTB_CRC_MAGIC:
2384 case XFS_ABTC_CRC_MAGIC:
2385 case XFS_ABTB_MAGIC:
2386 case XFS_ABTC_MAGIC:
2387 bp->b_ops = &xfs_allocbt_buf_ops;
2388 break;
2389 case XFS_IBT_CRC_MAGIC:
2390 case XFS_FIBT_CRC_MAGIC:
2391 case XFS_IBT_MAGIC:
2392 case XFS_FIBT_MAGIC:
2393 bp->b_ops = &xfs_inobt_buf_ops;
2394 break;
2395 case XFS_BMAP_CRC_MAGIC:
2396 case XFS_BMAP_MAGIC:
2397 bp->b_ops = &xfs_bmbt_buf_ops;
2398 break;
2399 default:
2400 xfs_warn(mp, "Bad btree block magic!");
2401 ASSERT(0);
2402 break;
2403 }
2404 break;
2405 case XFS_BLFT_AGF_BUF:
2406 if (magic32 != XFS_AGF_MAGIC) {
2407 xfs_warn(mp, "Bad AGF block magic!");
2408 ASSERT(0);
2409 break;
2410 }
2411 bp->b_ops = &xfs_agf_buf_ops;
2412 break;
2413 case XFS_BLFT_AGFL_BUF:
2414 if (magic32 != XFS_AGFL_MAGIC) {
2415 xfs_warn(mp, "Bad AGFL block magic!");
2416 ASSERT(0);
2417 break;
2418 }
2419 bp->b_ops = &xfs_agfl_buf_ops;
2420 break;
2421 case XFS_BLFT_AGI_BUF:
2422 if (magic32 != XFS_AGI_MAGIC) {
2423 xfs_warn(mp, "Bad AGI block magic!");
2424 ASSERT(0);
2425 break;
2426 }
2427 bp->b_ops = &xfs_agi_buf_ops;
2428 break;
2429 case XFS_BLFT_UDQUOT_BUF:
2430 case XFS_BLFT_PDQUOT_BUF:
2431 case XFS_BLFT_GDQUOT_BUF:
2432#ifdef CONFIG_XFS_QUOTA
2433 if (magic16 != XFS_DQUOT_MAGIC) {
2434 xfs_warn(mp, "Bad DQUOT block magic!");
2435 ASSERT(0);
2436 break;
2437 }
2438 bp->b_ops = &xfs_dquot_buf_ops;
2439#else
2440 xfs_alert(mp,
2441 "Trying to recover dquots without QUOTA support built in!");
2442 ASSERT(0);
2443#endif
2444 break;
2445 case XFS_BLFT_DINO_BUF:
2446 if (magic16 != XFS_DINODE_MAGIC) {
2447 xfs_warn(mp, "Bad INODE block magic!");
2448 ASSERT(0);
2449 break;
2450 }
2451 bp->b_ops = &xfs_inode_buf_ops;
2452 break;
2453 case XFS_BLFT_SYMLINK_BUF:
2454 if (magic32 != XFS_SYMLINK_MAGIC) {
2455 xfs_warn(mp, "Bad symlink block magic!");
2456 ASSERT(0);
2457 break;
2458 }
2459 bp->b_ops = &xfs_symlink_buf_ops;
2460 break;
2461 case XFS_BLFT_DIR_BLOCK_BUF:
2462 if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
2463 magic32 != XFS_DIR3_BLOCK_MAGIC) {
2464 xfs_warn(mp, "Bad dir block magic!");
2465 ASSERT(0);
2466 break;
2467 }
2468 bp->b_ops = &xfs_dir3_block_buf_ops;
2469 break;
2470 case XFS_BLFT_DIR_DATA_BUF:
2471 if (magic32 != XFS_DIR2_DATA_MAGIC &&
2472 magic32 != XFS_DIR3_DATA_MAGIC) {
2473 xfs_warn(mp, "Bad dir data magic!");
2474 ASSERT(0);
2475 break;
2476 }
2477 bp->b_ops = &xfs_dir3_data_buf_ops;
2478 break;
2479 case XFS_BLFT_DIR_FREE_BUF:
2480 if (magic32 != XFS_DIR2_FREE_MAGIC &&
2481 magic32 != XFS_DIR3_FREE_MAGIC) {
2482 xfs_warn(mp, "Bad dir3 free magic!");
2483 ASSERT(0);
2484 break;
2485 }
2486 bp->b_ops = &xfs_dir3_free_buf_ops;
2487 break;
2488 case XFS_BLFT_DIR_LEAF1_BUF:
2489 if (magicda != XFS_DIR2_LEAF1_MAGIC &&
2490 magicda != XFS_DIR3_LEAF1_MAGIC) {
2491 xfs_warn(mp, "Bad dir leaf1 magic!");
2492 ASSERT(0);
2493 break;
2494 }
2495 bp->b_ops = &xfs_dir3_leaf1_buf_ops;
2496 break;
2497 case XFS_BLFT_DIR_LEAFN_BUF:
2498 if (magicda != XFS_DIR2_LEAFN_MAGIC &&
2499 magicda != XFS_DIR3_LEAFN_MAGIC) {
2500 xfs_warn(mp, "Bad dir leafn magic!");
2501 ASSERT(0);
2502 break;
2503 }
2504 bp->b_ops = &xfs_dir3_leafn_buf_ops;
2505 break;
2506 case XFS_BLFT_DA_NODE_BUF:
2507 if (magicda != XFS_DA_NODE_MAGIC &&
2508 magicda != XFS_DA3_NODE_MAGIC) {
2509 xfs_warn(mp, "Bad da node magic!");
2510 ASSERT(0);
2511 break;
2512 }
2513 bp->b_ops = &xfs_da3_node_buf_ops;
2514 break;
2515 case XFS_BLFT_ATTR_LEAF_BUF:
2516 if (magicda != XFS_ATTR_LEAF_MAGIC &&
2517 magicda != XFS_ATTR3_LEAF_MAGIC) {
2518 xfs_warn(mp, "Bad attr leaf magic!");
2519 ASSERT(0);
2520 break;
2521 }
2522 bp->b_ops = &xfs_attr3_leaf_buf_ops;
2523 break;
2524 case XFS_BLFT_ATTR_RMT_BUF:
2525 if (magic32 != XFS_ATTR3_RMT_MAGIC) {
2526 xfs_warn(mp, "Bad attr remote magic!");
2527 ASSERT(0);
2528 break;
2529 }
2530 bp->b_ops = &xfs_attr3_rmt_buf_ops;
2531 break;
2532 case XFS_BLFT_SB_BUF:
2533 if (magic32 != XFS_SB_MAGIC) {
2534 xfs_warn(mp, "Bad SB block magic!");
2535 ASSERT(0);
2536 break;
2537 }
2538 bp->b_ops = &xfs_sb_buf_ops;
2539 break;
2540#ifdef CONFIG_XFS_RT
2541 case XFS_BLFT_RTBITMAP_BUF:
2542 case XFS_BLFT_RTSUMMARY_BUF:
2543 /* no magic numbers for verification of RT buffers */
2544 bp->b_ops = &xfs_rtbuf_ops;
2545 break;
2546#endif /* CONFIG_XFS_RT */
2547 default:
2548 xfs_warn(mp, "Unknown buffer type %d!",
2549 xfs_blft_from_flags(buf_f));
2550 break;
2551 }
2552}
2553
2554/*
2555 * Perform a 'normal' buffer recovery. Each logged region of the
2556 * buffer should be copied over the corresponding region in the
2557 * given buffer. The bitmap in the buf log format structure indicates
2558 * where to place the logged data.
2559 */
2560STATIC void
2561xlog_recover_do_reg_buffer(
2562 struct xfs_mount *mp,
2563 xlog_recover_item_t *item,
2564 struct xfs_buf *bp,
2565 xfs_buf_log_format_t *buf_f)
2566{
2567 int i;
2568 int bit;
2569 int nbits;
2570 int error;
2571
2572 trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
2573
2574 bit = 0;
2575 i = 1; /* 0 is the buf format structure */
2576 while (1) {
2577 bit = xfs_next_bit(buf_f->blf_data_map,
2578 buf_f->blf_map_size, bit);
2579 if (bit == -1)
2580 break;
2581 nbits = xfs_contig_bits(buf_f->blf_data_map,
2582 buf_f->blf_map_size, bit);
2583 ASSERT(nbits > 0);
2584 ASSERT(item->ri_buf[i].i_addr != NULL);
2585 ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
2586 ASSERT(BBTOB(bp->b_io_length) >=
2587 ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
2588
2589 /*
2590 * The dirty regions logged in the buffer, even though
2591 * contiguous, may span multiple chunks. This is because the
2592 * dirty region may span a physical page boundary in a buffer
2593 * and hence be split into two separate vectors for writing into
2594 * the log. Hence we need to trim nbits back to the length of
2595 * the current region being copied out of the log.
2596 */
2597 if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
2598 nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
2599
2600 /*
2601 * Do a sanity check if this is a dquot buffer. Just checking
2602 * the first dquot in the buffer should do. XXXThis is
2603 * probably a good thing to do for other buf types also.
2604 */
2605 error = 0;
2606 if (buf_f->blf_flags &
2607 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2608 if (item->ri_buf[i].i_addr == NULL) {
2609 xfs_alert(mp,
2610 "XFS: NULL dquot in %s.", __func__);
2611 goto next;
2612 }
2613 if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
2614 xfs_alert(mp,
2615 "XFS: dquot too small (%d) in %s.",
2616 item->ri_buf[i].i_len, __func__);
2617 goto next;
2618 }
2619 error = xfs_dqcheck(mp, item->ri_buf[i].i_addr,
2620 -1, 0, XFS_QMOPT_DOWARN,
2621 "dquot_buf_recover");
2622 if (error)
2623 goto next;
2624 }
2625
2626 memcpy(xfs_buf_offset(bp,
2627 (uint)bit << XFS_BLF_SHIFT), /* dest */
2628 item->ri_buf[i].i_addr, /* source */
2629 nbits<<XFS_BLF_SHIFT); /* length */
2630 next:
2631 i++;
2632 bit += nbits;
2633 }
2634
2635 /* Shouldn't be any more regions */
2636 ASSERT(i == item->ri_total);
2637
2638 xlog_recover_validate_buf_type(mp, bp, buf_f);
2639}
2640
2641/*
2642 * Perform a dquot buffer recovery.
2643 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
2644 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2645 * Else, treat it as a regular buffer and do recovery.
2646 *
2647 * Return false if the buffer was tossed and true if we recovered the buffer to
2648 * indicate to the caller if the buffer needs writing.
2649 */
2650STATIC bool
2651xlog_recover_do_dquot_buffer(
2652 struct xfs_mount *mp,
2653 struct xlog *log,
2654 struct xlog_recover_item *item,
2655 struct xfs_buf *bp,
2656 struct xfs_buf_log_format *buf_f)
2657{
2658 uint type;
2659
2660 trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2661
2662 /*
2663 * Filesystems are required to send in quota flags at mount time.
2664 */
2665 if (!mp->m_qflags)
2666 return false;
2667
2668 type = 0;
2669 if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2670 type |= XFS_DQ_USER;
2671 if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2672 type |= XFS_DQ_PROJ;
2673 if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2674 type |= XFS_DQ_GROUP;
2675 /*
2676 * This type of quotas was turned off, so ignore this buffer
2677 */
2678 if (log->l_quotaoffs_flag & type)
2679 return false;
2680
2681 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2682 return true;
2683}
2684
2685/*
2686 * This routine replays a modification made to a buffer at runtime.
2687 * There are actually two types of buffer, regular and inode, which
2688 * are handled differently. Inode buffers are handled differently
2689 * in that we only recover a specific set of data from them, namely
2690 * the inode di_next_unlinked fields. This is because all other inode
2691 * data is actually logged via inode records and any data we replay
2692 * here which overlaps that may be stale.
2693 *
2694 * When meta-data buffers are freed at run time we log a buffer item
2695 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2696 * of the buffer in the log should not be replayed at recovery time.
2697 * This is so that if the blocks covered by the buffer are reused for
2698 * file data before we crash we don't end up replaying old, freed
2699 * meta-data into a user's file.
2700 *
2701 * To handle the cancellation of buffer log items, we make two passes
2702 * over the log during recovery. During the first we build a table of
2703 * those buffers which have been cancelled, and during the second we
2704 * only replay those buffers which do not have corresponding cancel
2705 * records in the table. See xlog_recover_buffer_pass[1,2] above
2706 * for more details on the implementation of the table of cancel records.
2707 */
2708STATIC int
2709xlog_recover_buffer_pass2(
2710 struct xlog *log,
2711 struct list_head *buffer_list,
2712 struct xlog_recover_item *item,
2713 xfs_lsn_t current_lsn)
2714{
2715 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
2716 xfs_mount_t *mp = log->l_mp;
2717 xfs_buf_t *bp;
2718 int error;
2719 uint buf_flags;
2720 xfs_lsn_t lsn;
2721
2722 /*
2723 * In this pass we only want to recover all the buffers which have
2724 * not been cancelled and are not cancellation buffers themselves.
2725 */
2726 if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2727 buf_f->blf_len, buf_f->blf_flags)) {
2728 trace_xfs_log_recover_buf_cancel(log, buf_f);
2729 return 0;
2730 }
2731
2732 trace_xfs_log_recover_buf_recover(log, buf_f);
2733
2734 buf_flags = 0;
2735 if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
2736 buf_flags |= XBF_UNMAPPED;
2737
2738 bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2739 buf_flags, NULL);
2740 if (!bp)
2741 return -ENOMEM;
2742 error = bp->b_error;
2743 if (error) {
2744 xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#1)");
2745 goto out_release;
2746 }
2747
2748 /*
2749 * Recover the buffer only if we get an LSN from it and it's less than
2750 * the lsn of the transaction we are replaying.
2751 *
2752 * Note that we have to be extremely careful of readahead here.
2753 * Readahead does not attach verfiers to the buffers so if we don't
2754 * actually do any replay after readahead because of the LSN we found
2755 * in the buffer if more recent than that current transaction then we
2756 * need to attach the verifier directly. Failure to do so can lead to
2757 * future recovery actions (e.g. EFI and unlinked list recovery) can
2758 * operate on the buffers and they won't get the verifier attached. This
2759 * can lead to blocks on disk having the correct content but a stale
2760 * CRC.
2761 *
2762 * It is safe to assume these clean buffers are currently up to date.
2763 * If the buffer is dirtied by a later transaction being replayed, then
2764 * the verifier will be reset to match whatever recover turns that
2765 * buffer into.
2766 */
2767 lsn = xlog_recover_get_buf_lsn(mp, bp);
2768 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
2769 xlog_recover_validate_buf_type(mp, bp, buf_f);
2770 goto out_release;
2771 }
2772
2773 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2774 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2775 if (error)
2776 goto out_release;
2777 } else if (buf_f->blf_flags &
2778 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2779 bool dirty;
2780
2781 dirty = xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2782 if (!dirty)
2783 goto out_release;
2784 } else {
2785 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2786 }
2787
2788 /*
2789 * Perform delayed write on the buffer. Asynchronous writes will be
2790 * slower when taking into account all the buffers to be flushed.
2791 *
2792 * Also make sure that only inode buffers with good sizes stay in
2793 * the buffer cache. The kernel moves inodes in buffers of 1 block
2794 * or mp->m_inode_cluster_size bytes, whichever is bigger. The inode
2795 * buffers in the log can be a different size if the log was generated
2796 * by an older kernel using unclustered inode buffers or a newer kernel
2797 * running with a different inode cluster size. Regardless, if the
2798 * the inode buffer size isn't MAX(blocksize, mp->m_inode_cluster_size)
2799 * for *our* value of mp->m_inode_cluster_size, then we need to keep
2800 * the buffer out of the buffer cache so that the buffer won't
2801 * overlap with future reads of those inodes.
2802 */
2803 if (XFS_DINODE_MAGIC ==
2804 be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2805 (BBTOB(bp->b_io_length) != MAX(log->l_mp->m_sb.sb_blocksize,
2806 (__uint32_t)log->l_mp->m_inode_cluster_size))) {
2807 xfs_buf_stale(bp);
2808 error = xfs_bwrite(bp);
2809 } else {
2810 ASSERT(bp->b_target->bt_mount == mp);
2811 bp->b_iodone = xlog_recover_iodone;
2812 xfs_buf_delwri_queue(bp, buffer_list);
2813 }
2814
2815out_release:
2816 xfs_buf_relse(bp);
2817 return error;
2818}
2819
2820/*
2821 * Inode fork owner changes
2822 *
2823 * If we have been told that we have to reparent the inode fork, it's because an
2824 * extent swap operation on a CRC enabled filesystem has been done and we are
2825 * replaying it. We need to walk the BMBT of the appropriate fork and change the
2826 * owners of it.
2827 *
2828 * The complexity here is that we don't have an inode context to work with, so
2829 * after we've replayed the inode we need to instantiate one. This is where the
2830 * fun begins.
2831 *
2832 * We are in the middle of log recovery, so we can't run transactions. That
2833 * means we cannot use cache coherent inode instantiation via xfs_iget(), as
2834 * that will result in the corresponding iput() running the inode through
2835 * xfs_inactive(). If we've just replayed an inode core that changes the link
2836 * count to zero (i.e. it's been unlinked), then xfs_inactive() will run
2837 * transactions (bad!).
2838 *
2839 * So, to avoid this, we instantiate an inode directly from the inode core we've
2840 * just recovered. We have the buffer still locked, and all we really need to
2841 * instantiate is the inode core and the forks being modified. We can do this
2842 * manually, then run the inode btree owner change, and then tear down the
2843 * xfs_inode without having to run any transactions at all.
2844 *
2845 * Also, because we don't have a transaction context available here but need to
2846 * gather all the buffers we modify for writeback so we pass the buffer_list
2847 * instead for the operation to use.
2848 */
2849
2850STATIC int
2851xfs_recover_inode_owner_change(
2852 struct xfs_mount *mp,
2853 struct xfs_dinode *dip,
2854 struct xfs_inode_log_format *in_f,
2855 struct list_head *buffer_list)
2856{
2857 struct xfs_inode *ip;
2858 int error;
2859
2860 ASSERT(in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER));
2861
2862 ip = xfs_inode_alloc(mp, in_f->ilf_ino);
2863 if (!ip)
2864 return -ENOMEM;
2865
2866 /* instantiate the inode */
2867 xfs_inode_from_disk(ip, dip);
2868 ASSERT(ip->i_d.di_version >= 3);
2869
2870 error = xfs_iformat_fork(ip, dip);
2871 if (error)
2872 goto out_free_ip;
2873
2874
2875 if (in_f->ilf_fields & XFS_ILOG_DOWNER) {
2876 ASSERT(in_f->ilf_fields & XFS_ILOG_DBROOT);
2877 error = xfs_bmbt_change_owner(NULL, ip, XFS_DATA_FORK,
2878 ip->i_ino, buffer_list);
2879 if (error)
2880 goto out_free_ip;
2881 }
2882
2883 if (in_f->ilf_fields & XFS_ILOG_AOWNER) {
2884 ASSERT(in_f->ilf_fields & XFS_ILOG_ABROOT);
2885 error = xfs_bmbt_change_owner(NULL, ip, XFS_ATTR_FORK,
2886 ip->i_ino, buffer_list);
2887 if (error)
2888 goto out_free_ip;
2889 }
2890
2891out_free_ip:
2892 xfs_inode_free(ip);
2893 return error;
2894}
2895
2896STATIC int
2897xlog_recover_inode_pass2(
2898 struct xlog *log,
2899 struct list_head *buffer_list,
2900 struct xlog_recover_item *item,
2901 xfs_lsn_t current_lsn)
2902{
2903 xfs_inode_log_format_t *in_f;
2904 xfs_mount_t *mp = log->l_mp;
2905 xfs_buf_t *bp;
2906 xfs_dinode_t *dip;
2907 int len;
2908 char *src;
2909 char *dest;
2910 int error;
2911 int attr_index;
2912 uint fields;
2913 struct xfs_log_dinode *ldip;
2914 uint isize;
2915 int need_free = 0;
2916
2917 if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2918 in_f = item->ri_buf[0].i_addr;
2919 } else {
2920 in_f = kmem_alloc(sizeof(xfs_inode_log_format_t), KM_SLEEP);
2921 need_free = 1;
2922 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2923 if (error)
2924 goto error;
2925 }
2926
2927 /*
2928 * Inode buffers can be freed, look out for it,
2929 * and do not replay the inode.
2930 */
2931 if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2932 in_f->ilf_len, 0)) {
2933 error = 0;
2934 trace_xfs_log_recover_inode_cancel(log, in_f);
2935 goto error;
2936 }
2937 trace_xfs_log_recover_inode_recover(log, in_f);
2938
2939 bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len, 0,
2940 &xfs_inode_buf_ops);
2941 if (!bp) {
2942 error = -ENOMEM;
2943 goto error;
2944 }
2945 error = bp->b_error;
2946 if (error) {
2947 xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#2)");
2948 goto out_release;
2949 }
2950 ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2951 dip = xfs_buf_offset(bp, in_f->ilf_boffset);
2952
2953 /*
2954 * Make sure the place we're flushing out to really looks
2955 * like an inode!
2956 */
2957 if (unlikely(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC))) {
2958 xfs_alert(mp,
2959 "%s: Bad inode magic number, dip = 0x%p, dino bp = 0x%p, ino = %Ld",
2960 __func__, dip, bp, in_f->ilf_ino);
2961 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2962 XFS_ERRLEVEL_LOW, mp);
2963 error = -EFSCORRUPTED;
2964 goto out_release;
2965 }
2966 ldip = item->ri_buf[1].i_addr;
2967 if (unlikely(ldip->di_magic != XFS_DINODE_MAGIC)) {
2968 xfs_alert(mp,
2969 "%s: Bad inode log record, rec ptr 0x%p, ino %Ld",
2970 __func__, item, in_f->ilf_ino);
2971 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2972 XFS_ERRLEVEL_LOW, mp);
2973 error = -EFSCORRUPTED;
2974 goto out_release;
2975 }
2976
2977 /*
2978 * If the inode has an LSN in it, recover the inode only if it's less
2979 * than the lsn of the transaction we are replaying. Note: we still
2980 * need to replay an owner change even though the inode is more recent
2981 * than the transaction as there is no guarantee that all the btree
2982 * blocks are more recent than this transaction, too.
2983 */
2984 if (dip->di_version >= 3) {
2985 xfs_lsn_t lsn = be64_to_cpu(dip->di_lsn);
2986
2987 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
2988 trace_xfs_log_recover_inode_skip(log, in_f);
2989 error = 0;
2990 goto out_owner_change;
2991 }
2992 }
2993
2994 /*
2995 * di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
2996 * are transactional and if ordering is necessary we can determine that
2997 * more accurately by the LSN field in the V3 inode core. Don't trust
2998 * the inode versions we might be changing them here - use the
2999 * superblock flag to determine whether we need to look at di_flushiter
3000 * to skip replay when the on disk inode is newer than the log one
3001 */
3002 if (!xfs_sb_version_hascrc(&mp->m_sb) &&
3003 ldip->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
3004 /*
3005 * Deal with the wrap case, DI_MAX_FLUSH is less
3006 * than smaller numbers
3007 */
3008 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
3009 ldip->di_flushiter < (DI_MAX_FLUSH >> 1)) {
3010 /* do nothing */
3011 } else {
3012 trace_xfs_log_recover_inode_skip(log, in_f);
3013 error = 0;
3014 goto out_release;
3015 }
3016 }
3017
3018 /* Take the opportunity to reset the flush iteration count */
3019 ldip->di_flushiter = 0;
3020
3021 if (unlikely(S_ISREG(ldip->di_mode))) {
3022 if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3023 (ldip->di_format != XFS_DINODE_FMT_BTREE)) {
3024 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
3025 XFS_ERRLEVEL_LOW, mp, ldip);
3026 xfs_alert(mp,
3027 "%s: Bad regular inode log record, rec ptr 0x%p, "
3028 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
3029 __func__, item, dip, bp, in_f->ilf_ino);
3030 error = -EFSCORRUPTED;
3031 goto out_release;
3032 }
3033 } else if (unlikely(S_ISDIR(ldip->di_mode))) {
3034 if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3035 (ldip->di_format != XFS_DINODE_FMT_BTREE) &&
3036 (ldip->di_format != XFS_DINODE_FMT_LOCAL)) {
3037 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
3038 XFS_ERRLEVEL_LOW, mp, ldip);
3039 xfs_alert(mp,
3040 "%s: Bad dir inode log record, rec ptr 0x%p, "
3041 "ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
3042 __func__, item, dip, bp, in_f->ilf_ino);
3043 error = -EFSCORRUPTED;
3044 goto out_release;
3045 }
3046 }
3047 if (unlikely(ldip->di_nextents + ldip->di_anextents > ldip->di_nblocks)){
3048 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
3049 XFS_ERRLEVEL_LOW, mp, ldip);
3050 xfs_alert(mp,
3051 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
3052 "dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
3053 __func__, item, dip, bp, in_f->ilf_ino,
3054 ldip->di_nextents + ldip->di_anextents,
3055 ldip->di_nblocks);
3056 error = -EFSCORRUPTED;
3057 goto out_release;
3058 }
3059 if (unlikely(ldip->di_forkoff > mp->m_sb.sb_inodesize)) {
3060 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
3061 XFS_ERRLEVEL_LOW, mp, ldip);
3062 xfs_alert(mp,
3063 "%s: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, "
3064 "dino bp 0x%p, ino %Ld, forkoff 0x%x", __func__,
3065 item, dip, bp, in_f->ilf_ino, ldip->di_forkoff);
3066 error = -EFSCORRUPTED;
3067 goto out_release;
3068 }
3069 isize = xfs_log_dinode_size(ldip->di_version);
3070 if (unlikely(item->ri_buf[1].i_len > isize)) {
3071 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
3072 XFS_ERRLEVEL_LOW, mp, ldip);
3073 xfs_alert(mp,
3074 "%s: Bad inode log record length %d, rec ptr 0x%p",
3075 __func__, item->ri_buf[1].i_len, item);
3076 error = -EFSCORRUPTED;
3077 goto out_release;
3078 }
3079
3080 /* recover the log dinode inode into the on disk inode */
3081 xfs_log_dinode_to_disk(ldip, dip);
3082
3083 /* the rest is in on-disk format */
3084 if (item->ri_buf[1].i_len > isize) {
3085 memcpy((char *)dip + isize,
3086 item->ri_buf[1].i_addr + isize,
3087 item->ri_buf[1].i_len - isize);
3088 }
3089
3090 fields = in_f->ilf_fields;
3091 switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
3092 case XFS_ILOG_DEV:
3093 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
3094 break;
3095 case XFS_ILOG_UUID:
3096 memcpy(XFS_DFORK_DPTR(dip),
3097 &in_f->ilf_u.ilfu_uuid,
3098 sizeof(uuid_t));
3099 break;
3100 }
3101
3102 if (in_f->ilf_size == 2)
3103 goto out_owner_change;
3104 len = item->ri_buf[2].i_len;
3105 src = item->ri_buf[2].i_addr;
3106 ASSERT(in_f->ilf_size <= 4);
3107 ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
3108 ASSERT(!(fields & XFS_ILOG_DFORK) ||
3109 (len == in_f->ilf_dsize));
3110
3111 switch (fields & XFS_ILOG_DFORK) {
3112 case XFS_ILOG_DDATA:
3113 case XFS_ILOG_DEXT:
3114 memcpy(XFS_DFORK_DPTR(dip), src, len);
3115 break;
3116
3117 case XFS_ILOG_DBROOT:
3118 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
3119 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
3120 XFS_DFORK_DSIZE(dip, mp));
3121 break;
3122
3123 default:
3124 /*
3125 * There are no data fork flags set.
3126 */
3127 ASSERT((fields & XFS_ILOG_DFORK) == 0);
3128 break;
3129 }
3130
3131 /*
3132 * If we logged any attribute data, recover it. There may or
3133 * may not have been any other non-core data logged in this
3134 * transaction.
3135 */
3136 if (in_f->ilf_fields & XFS_ILOG_AFORK) {
3137 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
3138 attr_index = 3;
3139 } else {
3140 attr_index = 2;
3141 }
3142 len = item->ri_buf[attr_index].i_len;
3143 src = item->ri_buf[attr_index].i_addr;
3144 ASSERT(len == in_f->ilf_asize);
3145
3146 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
3147 case XFS_ILOG_ADATA:
3148 case XFS_ILOG_AEXT:
3149 dest = XFS_DFORK_APTR(dip);
3150 ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
3151 memcpy(dest, src, len);
3152 break;
3153
3154 case XFS_ILOG_ABROOT:
3155 dest = XFS_DFORK_APTR(dip);
3156 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
3157 len, (xfs_bmdr_block_t*)dest,
3158 XFS_DFORK_ASIZE(dip, mp));
3159 break;
3160
3161 default:
3162 xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
3163 ASSERT(0);
3164 error = -EIO;
3165 goto out_release;
3166 }
3167 }
3168
3169out_owner_change:
3170 if (in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER))
3171 error = xfs_recover_inode_owner_change(mp, dip, in_f,
3172 buffer_list);
3173 /* re-generate the checksum. */
3174 xfs_dinode_calc_crc(log->l_mp, dip);
3175
3176 ASSERT(bp->b_target->bt_mount == mp);
3177 bp->b_iodone = xlog_recover_iodone;
3178 xfs_buf_delwri_queue(bp, buffer_list);
3179
3180out_release:
3181 xfs_buf_relse(bp);
3182error:
3183 if (need_free)
3184 kmem_free(in_f);
3185 return error;
3186}
3187
3188/*
3189 * Recover QUOTAOFF records. We simply make a note of it in the xlog
3190 * structure, so that we know not to do any dquot item or dquot buffer recovery,
3191 * of that type.
3192 */
3193STATIC int
3194xlog_recover_quotaoff_pass1(
3195 struct xlog *log,
3196 struct xlog_recover_item *item)
3197{
3198 xfs_qoff_logformat_t *qoff_f = item->ri_buf[0].i_addr;
3199 ASSERT(qoff_f);
3200
3201 /*
3202 * The logitem format's flag tells us if this was user quotaoff,
3203 * group/project quotaoff or both.
3204 */
3205 if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
3206 log->l_quotaoffs_flag |= XFS_DQ_USER;
3207 if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
3208 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
3209 if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
3210 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
3211
3212 return 0;
3213}
3214
3215/*
3216 * Recover a dquot record
3217 */
3218STATIC int
3219xlog_recover_dquot_pass2(
3220 struct xlog *log,
3221 struct list_head *buffer_list,
3222 struct xlog_recover_item *item,
3223 xfs_lsn_t current_lsn)
3224{
3225 xfs_mount_t *mp = log->l_mp;
3226 xfs_buf_t *bp;
3227 struct xfs_disk_dquot *ddq, *recddq;
3228 int error;
3229 xfs_dq_logformat_t *dq_f;
3230 uint type;
3231
3232
3233 /*
3234 * Filesystems are required to send in quota flags at mount time.
3235 */
3236 if (mp->m_qflags == 0)
3237 return 0;
3238
3239 recddq = item->ri_buf[1].i_addr;
3240 if (recddq == NULL) {
3241 xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
3242 return -EIO;
3243 }
3244 if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
3245 xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
3246 item->ri_buf[1].i_len, __func__);
3247 return -EIO;
3248 }
3249
3250 /*
3251 * This type of quotas was turned off, so ignore this record.
3252 */
3253 type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3254 ASSERT(type);
3255 if (log->l_quotaoffs_flag & type)
3256 return 0;
3257
3258 /*
3259 * At this point we know that quota was _not_ turned off.
3260 * Since the mount flags are not indicating to us otherwise, this
3261 * must mean that quota is on, and the dquot needs to be replayed.
3262 * Remember that we may not have fully recovered the superblock yet,
3263 * so we can't do the usual trick of looking at the SB quota bits.
3264 *
3265 * The other possibility, of course, is that the quota subsystem was
3266 * removed since the last mount - ENOSYS.
3267 */
3268 dq_f = item->ri_buf[0].i_addr;
3269 ASSERT(dq_f);
3270 error = xfs_dqcheck(mp, recddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
3271 "xlog_recover_dquot_pass2 (log copy)");
3272 if (error)
3273 return -EIO;
3274 ASSERT(dq_f->qlf_len == 1);
3275
3276 /*
3277 * At this point we are assuming that the dquots have been allocated
3278 * and hence the buffer has valid dquots stamped in it. It should,
3279 * therefore, pass verifier validation. If the dquot is bad, then the
3280 * we'll return an error here, so we don't need to specifically check
3281 * the dquot in the buffer after the verifier has run.
3282 */
3283 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dq_f->qlf_blkno,
3284 XFS_FSB_TO_BB(mp, dq_f->qlf_len), 0, &bp,
3285 &xfs_dquot_buf_ops);
3286 if (error)
3287 return error;
3288
3289 ASSERT(bp);
3290 ddq = xfs_buf_offset(bp, dq_f->qlf_boffset);
3291
3292 /*
3293 * If the dquot has an LSN in it, recover the dquot only if it's less
3294 * than the lsn of the transaction we are replaying.
3295 */
3296 if (xfs_sb_version_hascrc(&mp->m_sb)) {
3297 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddq;
3298 xfs_lsn_t lsn = be64_to_cpu(dqb->dd_lsn);
3299
3300 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3301 goto out_release;
3302 }
3303 }
3304
3305 memcpy(ddq, recddq, item->ri_buf[1].i_len);
3306 if (xfs_sb_version_hascrc(&mp->m_sb)) {
3307 xfs_update_cksum((char *)ddq, sizeof(struct xfs_dqblk),
3308 XFS_DQUOT_CRC_OFF);
3309 }
3310
3311 ASSERT(dq_f->qlf_size == 2);
3312 ASSERT(bp->b_target->bt_mount == mp);
3313 bp->b_iodone = xlog_recover_iodone;
3314 xfs_buf_delwri_queue(bp, buffer_list);
3315
3316out_release:
3317 xfs_buf_relse(bp);
3318 return 0;
3319}
3320
3321/*
3322 * This routine is called to create an in-core extent free intent
3323 * item from the efi format structure which was logged on disk.
3324 * It allocates an in-core efi, copies the extents from the format
3325 * structure into it, and adds the efi to the AIL with the given
3326 * LSN.
3327 */
3328STATIC int
3329xlog_recover_efi_pass2(
3330 struct xlog *log,
3331 struct xlog_recover_item *item,
3332 xfs_lsn_t lsn)
3333{
3334 int error;
3335 struct xfs_mount *mp = log->l_mp;
3336 struct xfs_efi_log_item *efip;
3337 struct xfs_efi_log_format *efi_formatp;
3338
3339 efi_formatp = item->ri_buf[0].i_addr;
3340
3341 efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
3342 error = xfs_efi_copy_format(&item->ri_buf[0], &efip->efi_format);
3343 if (error) {
3344 xfs_efi_item_free(efip);
3345 return error;
3346 }
3347 atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
3348
3349 spin_lock(&log->l_ailp->xa_lock);
3350 /*
3351 * The EFI has two references. One for the EFD and one for EFI to ensure
3352 * it makes it into the AIL. Insert the EFI into the AIL directly and
3353 * drop the EFI reference. Note that xfs_trans_ail_update() drops the
3354 * AIL lock.
3355 */
3356 xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
3357 xfs_efi_release(efip);
3358 return 0;
3359}
3360
3361
3362/*
3363 * This routine is called when an EFD format structure is found in a committed
3364 * transaction in the log. Its purpose is to cancel the corresponding EFI if it
3365 * was still in the log. To do this it searches the AIL for the EFI with an id
3366 * equal to that in the EFD format structure. If we find it we drop the EFD
3367 * reference, which removes the EFI from the AIL and frees it.
3368 */
3369STATIC int
3370xlog_recover_efd_pass2(
3371 struct xlog *log,
3372 struct xlog_recover_item *item)
3373{
3374 xfs_efd_log_format_t *efd_formatp;
3375 xfs_efi_log_item_t *efip = NULL;
3376 xfs_log_item_t *lip;
3377 __uint64_t efi_id;
3378 struct xfs_ail_cursor cur;
3379 struct xfs_ail *ailp = log->l_ailp;
3380
3381 efd_formatp = item->ri_buf[0].i_addr;
3382 ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
3383 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
3384 (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
3385 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
3386 efi_id = efd_formatp->efd_efi_id;
3387
3388 /*
3389 * Search for the EFI with the id in the EFD format structure in the
3390 * AIL.
3391 */
3392 spin_lock(&ailp->xa_lock);
3393 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3394 while (lip != NULL) {
3395 if (lip->li_type == XFS_LI_EFI) {
3396 efip = (xfs_efi_log_item_t *)lip;
3397 if (efip->efi_format.efi_id == efi_id) {
3398 /*
3399 * Drop the EFD reference to the EFI. This
3400 * removes the EFI from the AIL and frees it.
3401 */
3402 spin_unlock(&ailp->xa_lock);
3403 xfs_efi_release(efip);
3404 spin_lock(&ailp->xa_lock);
3405 break;
3406 }
3407 }
3408 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3409 }
3410
3411 xfs_trans_ail_cursor_done(&cur);
3412 spin_unlock(&ailp->xa_lock);
3413
3414 return 0;
3415}
3416
3417/*
3418 * This routine is called when an inode create format structure is found in a
3419 * committed transaction in the log. It's purpose is to initialise the inodes
3420 * being allocated on disk. This requires us to get inode cluster buffers that
3421 * match the range to be intialised, stamped with inode templates and written
3422 * by delayed write so that subsequent modifications will hit the cached buffer
3423 * and only need writing out at the end of recovery.
3424 */
3425STATIC int
3426xlog_recover_do_icreate_pass2(
3427 struct xlog *log,
3428 struct list_head *buffer_list,
3429 xlog_recover_item_t *item)
3430{
3431 struct xfs_mount *mp = log->l_mp;
3432 struct xfs_icreate_log *icl;
3433 xfs_agnumber_t agno;
3434 xfs_agblock_t agbno;
3435 unsigned int count;
3436 unsigned int isize;
3437 xfs_agblock_t length;
3438 int blks_per_cluster;
3439 int bb_per_cluster;
3440 int cancel_count;
3441 int nbufs;
3442 int i;
3443
3444 icl = (struct xfs_icreate_log *)item->ri_buf[0].i_addr;
3445 if (icl->icl_type != XFS_LI_ICREATE) {
3446 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad type");
3447 return -EINVAL;
3448 }
3449
3450 if (icl->icl_size != 1) {
3451 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad icl size");
3452 return -EINVAL;
3453 }
3454
3455 agno = be32_to_cpu(icl->icl_ag);
3456 if (agno >= mp->m_sb.sb_agcount) {
3457 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agno");
3458 return -EINVAL;
3459 }
3460 agbno = be32_to_cpu(icl->icl_agbno);
3461 if (!agbno || agbno == NULLAGBLOCK || agbno >= mp->m_sb.sb_agblocks) {
3462 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agbno");
3463 return -EINVAL;
3464 }
3465 isize = be32_to_cpu(icl->icl_isize);
3466 if (isize != mp->m_sb.sb_inodesize) {
3467 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad isize");
3468 return -EINVAL;
3469 }
3470 count = be32_to_cpu(icl->icl_count);
3471 if (!count) {
3472 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count");
3473 return -EINVAL;
3474 }
3475 length = be32_to_cpu(icl->icl_length);
3476 if (!length || length >= mp->m_sb.sb_agblocks) {
3477 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad length");
3478 return -EINVAL;
3479 }
3480
3481 /*
3482 * The inode chunk is either full or sparse and we only support
3483 * m_ialloc_min_blks sized sparse allocations at this time.
3484 */
3485 if (length != mp->m_ialloc_blks &&
3486 length != mp->m_ialloc_min_blks) {
3487 xfs_warn(log->l_mp,
3488 "%s: unsupported chunk length", __FUNCTION__);
3489 return -EINVAL;
3490 }
3491
3492 /* verify inode count is consistent with extent length */
3493 if ((count >> mp->m_sb.sb_inopblog) != length) {
3494 xfs_warn(log->l_mp,
3495 "%s: inconsistent inode count and chunk length",
3496 __FUNCTION__);
3497 return -EINVAL;
3498 }
3499
3500 /*
3501 * The icreate transaction can cover multiple cluster buffers and these
3502 * buffers could have been freed and reused. Check the individual
3503 * buffers for cancellation so we don't overwrite anything written after
3504 * a cancellation.
3505 */
3506 blks_per_cluster = xfs_icluster_size_fsb(mp);
3507 bb_per_cluster = XFS_FSB_TO_BB(mp, blks_per_cluster);
3508 nbufs = length / blks_per_cluster;
3509 for (i = 0, cancel_count = 0; i < nbufs; i++) {
3510 xfs_daddr_t daddr;
3511
3512 daddr = XFS_AGB_TO_DADDR(mp, agno,
3513 agbno + i * blks_per_cluster);
3514 if (xlog_check_buffer_cancelled(log, daddr, bb_per_cluster, 0))
3515 cancel_count++;
3516 }
3517
3518 /*
3519 * We currently only use icreate for a single allocation at a time. This
3520 * means we should expect either all or none of the buffers to be
3521 * cancelled. Be conservative and skip replay if at least one buffer is
3522 * cancelled, but warn the user that something is awry if the buffers
3523 * are not consistent.
3524 *
3525 * XXX: This must be refined to only skip cancelled clusters once we use
3526 * icreate for multiple chunk allocations.
3527 */
3528 ASSERT(!cancel_count || cancel_count == nbufs);
3529 if (cancel_count) {
3530 if (cancel_count != nbufs)
3531 xfs_warn(mp,
3532 "WARNING: partial inode chunk cancellation, skipped icreate.");
3533 trace_xfs_log_recover_icreate_cancel(log, icl);
3534 return 0;
3535 }
3536
3537 trace_xfs_log_recover_icreate_recover(log, icl);
3538 return xfs_ialloc_inode_init(mp, NULL, buffer_list, count, agno, agbno,
3539 length, be32_to_cpu(icl->icl_gen));
3540}
3541
3542STATIC void
3543xlog_recover_buffer_ra_pass2(
3544 struct xlog *log,
3545 struct xlog_recover_item *item)
3546{
3547 struct xfs_buf_log_format *buf_f = item->ri_buf[0].i_addr;
3548 struct xfs_mount *mp = log->l_mp;
3549
3550 if (xlog_peek_buffer_cancelled(log, buf_f->blf_blkno,
3551 buf_f->blf_len, buf_f->blf_flags)) {
3552 return;
3553 }
3554
3555 xfs_buf_readahead(mp->m_ddev_targp, buf_f->blf_blkno,
3556 buf_f->blf_len, NULL);
3557}
3558
3559STATIC void
3560xlog_recover_inode_ra_pass2(
3561 struct xlog *log,
3562 struct xlog_recover_item *item)
3563{
3564 struct xfs_inode_log_format ilf_buf;
3565 struct xfs_inode_log_format *ilfp;
3566 struct xfs_mount *mp = log->l_mp;
3567 int error;
3568
3569 if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
3570 ilfp = item->ri_buf[0].i_addr;
3571 } else {
3572 ilfp = &ilf_buf;
3573 memset(ilfp, 0, sizeof(*ilfp));
3574 error = xfs_inode_item_format_convert(&item->ri_buf[0], ilfp);
3575 if (error)
3576 return;
3577 }
3578
3579 if (xlog_peek_buffer_cancelled(log, ilfp->ilf_blkno, ilfp->ilf_len, 0))
3580 return;
3581
3582 xfs_buf_readahead(mp->m_ddev_targp, ilfp->ilf_blkno,
3583 ilfp->ilf_len, &xfs_inode_buf_ra_ops);
3584}
3585
3586STATIC void
3587xlog_recover_dquot_ra_pass2(
3588 struct xlog *log,
3589 struct xlog_recover_item *item)
3590{
3591 struct xfs_mount *mp = log->l_mp;
3592 struct xfs_disk_dquot *recddq;
3593 struct xfs_dq_logformat *dq_f;
3594 uint type;
3595 int len;
3596
3597
3598 if (mp->m_qflags == 0)
3599 return;
3600
3601 recddq = item->ri_buf[1].i_addr;
3602 if (recddq == NULL)
3603 return;
3604 if (item->ri_buf[1].i_len < sizeof(struct xfs_disk_dquot))
3605 return;
3606
3607 type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3608 ASSERT(type);
3609 if (log->l_quotaoffs_flag & type)
3610 return;
3611
3612 dq_f = item->ri_buf[0].i_addr;
3613 ASSERT(dq_f);
3614 ASSERT(dq_f->qlf_len == 1);
3615
3616 len = XFS_FSB_TO_BB(mp, dq_f->qlf_len);
3617 if (xlog_peek_buffer_cancelled(log, dq_f->qlf_blkno, len, 0))
3618 return;
3619
3620 xfs_buf_readahead(mp->m_ddev_targp, dq_f->qlf_blkno, len,
3621 &xfs_dquot_buf_ra_ops);
3622}
3623
3624STATIC void
3625xlog_recover_ra_pass2(
3626 struct xlog *log,
3627 struct xlog_recover_item *item)
3628{
3629 switch (ITEM_TYPE(item)) {
3630 case XFS_LI_BUF:
3631 xlog_recover_buffer_ra_pass2(log, item);
3632 break;
3633 case XFS_LI_INODE:
3634 xlog_recover_inode_ra_pass2(log, item);
3635 break;
3636 case XFS_LI_DQUOT:
3637 xlog_recover_dquot_ra_pass2(log, item);
3638 break;
3639 case XFS_LI_EFI:
3640 case XFS_LI_EFD:
3641 case XFS_LI_QUOTAOFF:
3642 default:
3643 break;
3644 }
3645}
3646
3647STATIC int
3648xlog_recover_commit_pass1(
3649 struct xlog *log,
3650 struct xlog_recover *trans,
3651 struct xlog_recover_item *item)
3652{
3653 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
3654
3655 switch (ITEM_TYPE(item)) {
3656 case XFS_LI_BUF:
3657 return xlog_recover_buffer_pass1(log, item);
3658 case XFS_LI_QUOTAOFF:
3659 return xlog_recover_quotaoff_pass1(log, item);
3660 case XFS_LI_INODE:
3661 case XFS_LI_EFI:
3662 case XFS_LI_EFD:
3663 case XFS_LI_DQUOT:
3664 case XFS_LI_ICREATE:
3665 /* nothing to do in pass 1 */
3666 return 0;
3667 default:
3668 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
3669 __func__, ITEM_TYPE(item));
3670 ASSERT(0);
3671 return -EIO;
3672 }
3673}
3674
3675STATIC int
3676xlog_recover_commit_pass2(
3677 struct xlog *log,
3678 struct xlog_recover *trans,
3679 struct list_head *buffer_list,
3680 struct xlog_recover_item *item)
3681{
3682 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
3683
3684 switch (ITEM_TYPE(item)) {
3685 case XFS_LI_BUF:
3686 return xlog_recover_buffer_pass2(log, buffer_list, item,
3687 trans->r_lsn);
3688 case XFS_LI_INODE:
3689 return xlog_recover_inode_pass2(log, buffer_list, item,
3690 trans->r_lsn);
3691 case XFS_LI_EFI:
3692 return xlog_recover_efi_pass2(log, item, trans->r_lsn);
3693 case XFS_LI_EFD:
3694 return xlog_recover_efd_pass2(log, item);
3695 case XFS_LI_DQUOT:
3696 return xlog_recover_dquot_pass2(log, buffer_list, item,
3697 trans->r_lsn);
3698 case XFS_LI_ICREATE:
3699 return xlog_recover_do_icreate_pass2(log, buffer_list, item);
3700 case XFS_LI_QUOTAOFF:
3701 /* nothing to do in pass2 */
3702 return 0;
3703 default:
3704 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
3705 __func__, ITEM_TYPE(item));
3706 ASSERT(0);
3707 return -EIO;
3708 }
3709}
3710
3711STATIC int
3712xlog_recover_items_pass2(
3713 struct xlog *log,
3714 struct xlog_recover *trans,
3715 struct list_head *buffer_list,
3716 struct list_head *item_list)
3717{
3718 struct xlog_recover_item *item;
3719 int error = 0;
3720
3721 list_for_each_entry(item, item_list, ri_list) {
3722 error = xlog_recover_commit_pass2(log, trans,
3723 buffer_list, item);
3724 if (error)
3725 return error;
3726 }
3727
3728 return error;
3729}
3730
3731/*
3732 * Perform the transaction.
3733 *
3734 * If the transaction modifies a buffer or inode, do it now. Otherwise,
3735 * EFIs and EFDs get queued up by adding entries into the AIL for them.
3736 */
3737STATIC int
3738xlog_recover_commit_trans(
3739 struct xlog *log,
3740 struct xlog_recover *trans,
3741 int pass)
3742{
3743 int error = 0;
3744 int error2;
3745 int items_queued = 0;
3746 struct xlog_recover_item *item;
3747 struct xlog_recover_item *next;
3748 LIST_HEAD (buffer_list);
3749 LIST_HEAD (ra_list);
3750 LIST_HEAD (done_list);
3751
3752 #define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
3753
3754 hlist_del(&trans->r_list);
3755
3756 error = xlog_recover_reorder_trans(log, trans, pass);
3757 if (error)
3758 return error;
3759
3760 list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
3761 switch (pass) {
3762 case XLOG_RECOVER_PASS1:
3763 error = xlog_recover_commit_pass1(log, trans, item);
3764 break;
3765 case XLOG_RECOVER_PASS2:
3766 xlog_recover_ra_pass2(log, item);
3767 list_move_tail(&item->ri_list, &ra_list);
3768 items_queued++;
3769 if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
3770 error = xlog_recover_items_pass2(log, trans,
3771 &buffer_list, &ra_list);
3772 list_splice_tail_init(&ra_list, &done_list);
3773 items_queued = 0;
3774 }
3775
3776 break;
3777 default:
3778 ASSERT(0);
3779 }
3780
3781 if (error)
3782 goto out;
3783 }
3784
3785out:
3786 if (!list_empty(&ra_list)) {
3787 if (!error)
3788 error = xlog_recover_items_pass2(log, trans,
3789 &buffer_list, &ra_list);
3790 list_splice_tail_init(&ra_list, &done_list);
3791 }
3792
3793 if (!list_empty(&done_list))
3794 list_splice_init(&done_list, &trans->r_itemq);
3795
3796 error2 = xfs_buf_delwri_submit(&buffer_list);
3797 return error ? error : error2;
3798}
3799
3800STATIC void
3801xlog_recover_add_item(
3802 struct list_head *head)
3803{
3804 xlog_recover_item_t *item;
3805
3806 item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
3807 INIT_LIST_HEAD(&item->ri_list);
3808 list_add_tail(&item->ri_list, head);
3809}
3810
3811STATIC int
3812xlog_recover_add_to_cont_trans(
3813 struct xlog *log,
3814 struct xlog_recover *trans,
3815 char *dp,
3816 int len)
3817{
3818 xlog_recover_item_t *item;
3819 char *ptr, *old_ptr;
3820 int old_len;
3821
3822 /*
3823 * If the transaction is empty, the header was split across this and the
3824 * previous record. Copy the rest of the header.
3825 */
3826 if (list_empty(&trans->r_itemq)) {
3827 ASSERT(len <= sizeof(struct xfs_trans_header));
3828 if (len > sizeof(struct xfs_trans_header)) {
3829 xfs_warn(log->l_mp, "%s: bad header length", __func__);
3830 return -EIO;
3831 }
3832
3833 xlog_recover_add_item(&trans->r_itemq);
3834 ptr = (char *)&trans->r_theader +
3835 sizeof(struct xfs_trans_header) - len;
3836 memcpy(ptr, dp, len);
3837 return 0;
3838 }
3839
3840 /* take the tail entry */
3841 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
3842
3843 old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
3844 old_len = item->ri_buf[item->ri_cnt-1].i_len;
3845
3846 ptr = kmem_realloc(old_ptr, len+old_len, old_len, KM_SLEEP);
3847 memcpy(&ptr[old_len], dp, len);
3848 item->ri_buf[item->ri_cnt-1].i_len += len;
3849 item->ri_buf[item->ri_cnt-1].i_addr = ptr;
3850 trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
3851 return 0;
3852}
3853
3854/*
3855 * The next region to add is the start of a new region. It could be
3856 * a whole region or it could be the first part of a new region. Because
3857 * of this, the assumption here is that the type and size fields of all
3858 * format structures fit into the first 32 bits of the structure.
3859 *
3860 * This works because all regions must be 32 bit aligned. Therefore, we
3861 * either have both fields or we have neither field. In the case we have
3862 * neither field, the data part of the region is zero length. We only have
3863 * a log_op_header and can throw away the header since a new one will appear
3864 * later. If we have at least 4 bytes, then we can determine how many regions
3865 * will appear in the current log item.
3866 */
3867STATIC int
3868xlog_recover_add_to_trans(
3869 struct xlog *log,
3870 struct xlog_recover *trans,
3871 char *dp,
3872 int len)
3873{
3874 xfs_inode_log_format_t *in_f; /* any will do */
3875 xlog_recover_item_t *item;
3876 char *ptr;
3877
3878 if (!len)
3879 return 0;
3880 if (list_empty(&trans->r_itemq)) {
3881 /* we need to catch log corruptions here */
3882 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
3883 xfs_warn(log->l_mp, "%s: bad header magic number",
3884 __func__);
3885 ASSERT(0);
3886 return -EIO;
3887 }
3888
3889 if (len > sizeof(struct xfs_trans_header)) {
3890 xfs_warn(log->l_mp, "%s: bad header length", __func__);
3891 ASSERT(0);
3892 return -EIO;
3893 }
3894
3895 /*
3896 * The transaction header can be arbitrarily split across op
3897 * records. If we don't have the whole thing here, copy what we
3898 * do have and handle the rest in the next record.
3899 */
3900 if (len == sizeof(struct xfs_trans_header))
3901 xlog_recover_add_item(&trans->r_itemq);
3902 memcpy(&trans->r_theader, dp, len);
3903 return 0;
3904 }
3905
3906 ptr = kmem_alloc(len, KM_SLEEP);
3907 memcpy(ptr, dp, len);
3908 in_f = (xfs_inode_log_format_t *)ptr;
3909
3910 /* take the tail entry */
3911 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
3912 if (item->ri_total != 0 &&
3913 item->ri_total == item->ri_cnt) {
3914 /* tail item is in use, get a new one */
3915 xlog_recover_add_item(&trans->r_itemq);
3916 item = list_entry(trans->r_itemq.prev,
3917 xlog_recover_item_t, ri_list);
3918 }
3919
3920 if (item->ri_total == 0) { /* first region to be added */
3921 if (in_f->ilf_size == 0 ||
3922 in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
3923 xfs_warn(log->l_mp,
3924 "bad number of regions (%d) in inode log format",
3925 in_f->ilf_size);
3926 ASSERT(0);
3927 kmem_free(ptr);
3928 return -EIO;
3929 }
3930
3931 item->ri_total = in_f->ilf_size;
3932 item->ri_buf =
3933 kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
3934 KM_SLEEP);
3935 }
3936 ASSERT(item->ri_total > item->ri_cnt);
3937 /* Description region is ri_buf[0] */
3938 item->ri_buf[item->ri_cnt].i_addr = ptr;
3939 item->ri_buf[item->ri_cnt].i_len = len;
3940 item->ri_cnt++;
3941 trace_xfs_log_recover_item_add(log, trans, item, 0);
3942 return 0;
3943}
3944
3945/*
3946 * Free up any resources allocated by the transaction
3947 *
3948 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
3949 */
3950STATIC void
3951xlog_recover_free_trans(
3952 struct xlog_recover *trans)
3953{
3954 xlog_recover_item_t *item, *n;
3955 int i;
3956
3957 list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
3958 /* Free the regions in the item. */
3959 list_del(&item->ri_list);
3960 for (i = 0; i < item->ri_cnt; i++)
3961 kmem_free(item->ri_buf[i].i_addr);
3962 /* Free the item itself */
3963 kmem_free(item->ri_buf);
3964 kmem_free(item);
3965 }
3966 /* Free the transaction recover structure */
3967 kmem_free(trans);
3968}
3969
3970/*
3971 * On error or completion, trans is freed.
3972 */
3973STATIC int
3974xlog_recovery_process_trans(
3975 struct xlog *log,
3976 struct xlog_recover *trans,
3977 char *dp,
3978 unsigned int len,
3979 unsigned int flags,
3980 int pass)
3981{
3982 int error = 0;
3983 bool freeit = false;
3984
3985 /* mask off ophdr transaction container flags */
3986 flags &= ~XLOG_END_TRANS;
3987 if (flags & XLOG_WAS_CONT_TRANS)
3988 flags &= ~XLOG_CONTINUE_TRANS;
3989
3990 /*
3991 * Callees must not free the trans structure. We'll decide if we need to
3992 * free it or not based on the operation being done and it's result.
3993 */
3994 switch (flags) {
3995 /* expected flag values */
3996 case 0:
3997 case XLOG_CONTINUE_TRANS:
3998 error = xlog_recover_add_to_trans(log, trans, dp, len);
3999 break;
4000 case XLOG_WAS_CONT_TRANS:
4001 error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
4002 break;
4003 case XLOG_COMMIT_TRANS:
4004 error = xlog_recover_commit_trans(log, trans, pass);
4005 /* success or fail, we are now done with this transaction. */
4006 freeit = true;
4007 break;
4008
4009 /* unexpected flag values */
4010 case XLOG_UNMOUNT_TRANS:
4011 /* just skip trans */
4012 xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
4013 freeit = true;
4014 break;
4015 case XLOG_START_TRANS:
4016 default:
4017 xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
4018 ASSERT(0);
4019 error = -EIO;
4020 break;
4021 }
4022 if (error || freeit)
4023 xlog_recover_free_trans(trans);
4024 return error;
4025}
4026
4027/*
4028 * Lookup the transaction recovery structure associated with the ID in the
4029 * current ophdr. If the transaction doesn't exist and the start flag is set in
4030 * the ophdr, then allocate a new transaction for future ID matches to find.
4031 * Either way, return what we found during the lookup - an existing transaction
4032 * or nothing.
4033 */
4034STATIC struct xlog_recover *
4035xlog_recover_ophdr_to_trans(
4036 struct hlist_head rhash[],
4037 struct xlog_rec_header *rhead,
4038 struct xlog_op_header *ohead)
4039{
4040 struct xlog_recover *trans;
4041 xlog_tid_t tid;
4042 struct hlist_head *rhp;
4043
4044 tid = be32_to_cpu(ohead->oh_tid);
4045 rhp = &rhash[XLOG_RHASH(tid)];
4046 hlist_for_each_entry(trans, rhp, r_list) {
4047 if (trans->r_log_tid == tid)
4048 return trans;
4049 }
4050
4051 /*
4052 * skip over non-start transaction headers - we could be
4053 * processing slack space before the next transaction starts
4054 */
4055 if (!(ohead->oh_flags & XLOG_START_TRANS))
4056 return NULL;
4057
4058 ASSERT(be32_to_cpu(ohead->oh_len) == 0);
4059
4060 /*
4061 * This is a new transaction so allocate a new recovery container to
4062 * hold the recovery ops that will follow.
4063 */
4064 trans = kmem_zalloc(sizeof(struct xlog_recover), KM_SLEEP);
4065 trans->r_log_tid = tid;
4066 trans->r_lsn = be64_to_cpu(rhead->h_lsn);
4067 INIT_LIST_HEAD(&trans->r_itemq);
4068 INIT_HLIST_NODE(&trans->r_list);
4069 hlist_add_head(&trans->r_list, rhp);
4070
4071 /*
4072 * Nothing more to do for this ophdr. Items to be added to this new
4073 * transaction will be in subsequent ophdr containers.
4074 */
4075 return NULL;
4076}
4077
4078STATIC int
4079xlog_recover_process_ophdr(
4080 struct xlog *log,
4081 struct hlist_head rhash[],
4082 struct xlog_rec_header *rhead,
4083 struct xlog_op_header *ohead,
4084 char *dp,
4085 char *end,
4086 int pass)
4087{
4088 struct xlog_recover *trans;
4089 unsigned int len;
4090
4091 /* Do we understand who wrote this op? */
4092 if (ohead->oh_clientid != XFS_TRANSACTION &&
4093 ohead->oh_clientid != XFS_LOG) {
4094 xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
4095 __func__, ohead->oh_clientid);
4096 ASSERT(0);
4097 return -EIO;
4098 }
4099
4100 /*
4101 * Check the ophdr contains all the data it is supposed to contain.
4102 */
4103 len = be32_to_cpu(ohead->oh_len);
4104 if (dp + len > end) {
4105 xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
4106 WARN_ON(1);
4107 return -EIO;
4108 }
4109
4110 trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
4111 if (!trans) {
4112 /* nothing to do, so skip over this ophdr */
4113 return 0;
4114 }
4115
4116 return xlog_recovery_process_trans(log, trans, dp, len,
4117 ohead->oh_flags, pass);
4118}
4119
4120/*
4121 * There are two valid states of the r_state field. 0 indicates that the
4122 * transaction structure is in a normal state. We have either seen the
4123 * start of the transaction or the last operation we added was not a partial
4124 * operation. If the last operation we added to the transaction was a
4125 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
4126 *
4127 * NOTE: skip LRs with 0 data length.
4128 */
4129STATIC int
4130xlog_recover_process_data(
4131 struct xlog *log,
4132 struct hlist_head rhash[],
4133 struct xlog_rec_header *rhead,
4134 char *dp,
4135 int pass)
4136{
4137 struct xlog_op_header *ohead;
4138 char *end;
4139 int num_logops;
4140 int error;
4141
4142 end = dp + be32_to_cpu(rhead->h_len);
4143 num_logops = be32_to_cpu(rhead->h_num_logops);
4144
4145 /* check the log format matches our own - else we can't recover */
4146 if (xlog_header_check_recover(log->l_mp, rhead))
4147 return -EIO;
4148
4149 while ((dp < end) && num_logops) {
4150
4151 ohead = (struct xlog_op_header *)dp;
4152 dp += sizeof(*ohead);
4153 ASSERT(dp <= end);
4154
4155 /* errors will abort recovery */
4156 error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
4157 dp, end, pass);
4158 if (error)
4159 return error;
4160
4161 dp += be32_to_cpu(ohead->oh_len);
4162 num_logops--;
4163 }
4164 return 0;
4165}
4166
4167/*
4168 * Process an extent free intent item that was recovered from
4169 * the log. We need to free the extents that it describes.
4170 */
4171STATIC int
4172xlog_recover_process_efi(
4173 xfs_mount_t *mp,
4174 xfs_efi_log_item_t *efip)
4175{
4176 xfs_efd_log_item_t *efdp;
4177 xfs_trans_t *tp;
4178 int i;
4179 int error = 0;
4180 xfs_extent_t *extp;
4181 xfs_fsblock_t startblock_fsb;
4182
4183 ASSERT(!test_bit(XFS_EFI_RECOVERED, &efip->efi_flags));
4184
4185 /*
4186 * First check the validity of the extents described by the
4187 * EFI. If any are bad, then assume that all are bad and
4188 * just toss the EFI.
4189 */
4190 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
4191 extp = &(efip->efi_format.efi_extents[i]);
4192 startblock_fsb = XFS_BB_TO_FSB(mp,
4193 XFS_FSB_TO_DADDR(mp, extp->ext_start));
4194 if ((startblock_fsb == 0) ||
4195 (extp->ext_len == 0) ||
4196 (startblock_fsb >= mp->m_sb.sb_dblocks) ||
4197 (extp->ext_len >= mp->m_sb.sb_agblocks)) {
4198 /*
4199 * This will pull the EFI from the AIL and
4200 * free the memory associated with it.
4201 */
4202 set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
4203 xfs_efi_release(efip);
4204 return -EIO;
4205 }
4206 }
4207
4208 tp = xfs_trans_alloc(mp, 0);
4209 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_itruncate, 0, 0);
4210 if (error)
4211 goto abort_error;
4212 efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
4213
4214 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
4215 extp = &(efip->efi_format.efi_extents[i]);
4216 error = xfs_trans_free_extent(tp, efdp, extp->ext_start,
4217 extp->ext_len);
4218 if (error)
4219 goto abort_error;
4220
4221 }
4222
4223 set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
4224 error = xfs_trans_commit(tp);
4225 return error;
4226
4227abort_error:
4228 xfs_trans_cancel(tp);
4229 return error;
4230}
4231
4232/*
4233 * When this is called, all of the EFIs which did not have
4234 * corresponding EFDs should be in the AIL. What we do now
4235 * is free the extents associated with each one.
4236 *
4237 * Since we process the EFIs in normal transactions, they
4238 * will be removed at some point after the commit. This prevents
4239 * us from just walking down the list processing each one.
4240 * We'll use a flag in the EFI to skip those that we've already
4241 * processed and use the AIL iteration mechanism's generation
4242 * count to try to speed this up at least a bit.
4243 *
4244 * When we start, we know that the EFIs are the only things in
4245 * the AIL. As we process them, however, other items are added
4246 * to the AIL. Since everything added to the AIL must come after
4247 * everything already in the AIL, we stop processing as soon as
4248 * we see something other than an EFI in the AIL.
4249 */
4250STATIC int
4251xlog_recover_process_efis(
4252 struct xlog *log)
4253{
4254 struct xfs_log_item *lip;
4255 struct xfs_efi_log_item *efip;
4256 int error = 0;
4257 struct xfs_ail_cursor cur;
4258 struct xfs_ail *ailp;
4259
4260 ailp = log->l_ailp;
4261 spin_lock(&ailp->xa_lock);
4262 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4263 while (lip != NULL) {
4264 /*
4265 * We're done when we see something other than an EFI.
4266 * There should be no EFIs left in the AIL now.
4267 */
4268 if (lip->li_type != XFS_LI_EFI) {
4269#ifdef DEBUG
4270 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4271 ASSERT(lip->li_type != XFS_LI_EFI);
4272#endif
4273 break;
4274 }
4275
4276 /*
4277 * Skip EFIs that we've already processed.
4278 */
4279 efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4280 if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags)) {
4281 lip = xfs_trans_ail_cursor_next(ailp, &cur);
4282 continue;
4283 }
4284
4285 spin_unlock(&ailp->xa_lock);
4286 error = xlog_recover_process_efi(log->l_mp, efip);
4287 spin_lock(&ailp->xa_lock);
4288 if (error)
4289 goto out;
4290 lip = xfs_trans_ail_cursor_next(ailp, &cur);
4291 }
4292out:
4293 xfs_trans_ail_cursor_done(&cur);
4294 spin_unlock(&ailp->xa_lock);
4295 return error;
4296}
4297
4298/*
4299 * A cancel occurs when the mount has failed and we're bailing out. Release all
4300 * pending EFIs so they don't pin the AIL.
4301 */
4302STATIC int
4303xlog_recover_cancel_efis(
4304 struct xlog *log)
4305{
4306 struct xfs_log_item *lip;
4307 struct xfs_efi_log_item *efip;
4308 int error = 0;
4309 struct xfs_ail_cursor cur;
4310 struct xfs_ail *ailp;
4311
4312 ailp = log->l_ailp;
4313 spin_lock(&ailp->xa_lock);
4314 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4315 while (lip != NULL) {
4316 /*
4317 * We're done when we see something other than an EFI.
4318 * There should be no EFIs left in the AIL now.
4319 */
4320 if (lip->li_type != XFS_LI_EFI) {
4321#ifdef DEBUG
4322 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4323 ASSERT(lip->li_type != XFS_LI_EFI);
4324#endif
4325 break;
4326 }
4327
4328 efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4329
4330 spin_unlock(&ailp->xa_lock);
4331 xfs_efi_release(efip);
4332 spin_lock(&ailp->xa_lock);
4333
4334 lip = xfs_trans_ail_cursor_next(ailp, &cur);
4335 }
4336
4337 xfs_trans_ail_cursor_done(&cur);
4338 spin_unlock(&ailp->xa_lock);
4339 return error;
4340}
4341
4342/*
4343 * This routine performs a transaction to null out a bad inode pointer
4344 * in an agi unlinked inode hash bucket.
4345 */
4346STATIC void
4347xlog_recover_clear_agi_bucket(
4348 xfs_mount_t *mp,
4349 xfs_agnumber_t agno,
4350 int bucket)
4351{
4352 xfs_trans_t *tp;
4353 xfs_agi_t *agi;
4354 xfs_buf_t *agibp;
4355 int offset;
4356 int error;
4357
4358 tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
4359 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_clearagi, 0, 0);
4360 if (error)
4361 goto out_abort;
4362
4363 error = xfs_read_agi(mp, tp, agno, &agibp);
4364 if (error)
4365 goto out_abort;
4366
4367 agi = XFS_BUF_TO_AGI(agibp);
4368 agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
4369 offset = offsetof(xfs_agi_t, agi_unlinked) +
4370 (sizeof(xfs_agino_t) * bucket);
4371 xfs_trans_log_buf(tp, agibp, offset,
4372 (offset + sizeof(xfs_agino_t) - 1));
4373
4374 error = xfs_trans_commit(tp);
4375 if (error)
4376 goto out_error;
4377 return;
4378
4379out_abort:
4380 xfs_trans_cancel(tp);
4381out_error:
4382 xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
4383 return;
4384}
4385
4386STATIC xfs_agino_t
4387xlog_recover_process_one_iunlink(
4388 struct xfs_mount *mp,
4389 xfs_agnumber_t agno,
4390 xfs_agino_t agino,
4391 int bucket)
4392{
4393 struct xfs_buf *ibp;
4394 struct xfs_dinode *dip;
4395 struct xfs_inode *ip;
4396 xfs_ino_t ino;
4397 int error;
4398
4399 ino = XFS_AGINO_TO_INO(mp, agno, agino);
4400 error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
4401 if (error)
4402 goto fail;
4403
4404 /*
4405 * Get the on disk inode to find the next inode in the bucket.
4406 */
4407 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0, 0);
4408 if (error)
4409 goto fail_iput;
4410
4411 ASSERT(VFS_I(ip)->i_nlink == 0);
4412 ASSERT(VFS_I(ip)->i_mode != 0);
4413
4414 /* setup for the next pass */
4415 agino = be32_to_cpu(dip->di_next_unlinked);
4416 xfs_buf_relse(ibp);
4417
4418 /*
4419 * Prevent any DMAPI event from being sent when the reference on
4420 * the inode is dropped.
4421 */
4422 ip->i_d.di_dmevmask = 0;
4423
4424 IRELE(ip);
4425 return agino;
4426
4427 fail_iput:
4428 IRELE(ip);
4429 fail:
4430 /*
4431 * We can't read in the inode this bucket points to, or this inode
4432 * is messed up. Just ditch this bucket of inodes. We will lose
4433 * some inodes and space, but at least we won't hang.
4434 *
4435 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
4436 * clear the inode pointer in the bucket.
4437 */
4438 xlog_recover_clear_agi_bucket(mp, agno, bucket);
4439 return NULLAGINO;
4440}
4441
4442/*
4443 * xlog_iunlink_recover
4444 *
4445 * This is called during recovery to process any inodes which
4446 * we unlinked but not freed when the system crashed. These
4447 * inodes will be on the lists in the AGI blocks. What we do
4448 * here is scan all the AGIs and fully truncate and free any
4449 * inodes found on the lists. Each inode is removed from the
4450 * lists when it has been fully truncated and is freed. The
4451 * freeing of the inode and its removal from the list must be
4452 * atomic.
4453 */
4454STATIC void
4455xlog_recover_process_iunlinks(
4456 struct xlog *log)
4457{
4458 xfs_mount_t *mp;
4459 xfs_agnumber_t agno;
4460 xfs_agi_t *agi;
4461 xfs_buf_t *agibp;
4462 xfs_agino_t agino;
4463 int bucket;
4464 int error;
4465 uint mp_dmevmask;
4466
4467 mp = log->l_mp;
4468
4469 /*
4470 * Prevent any DMAPI event from being sent while in this function.
4471 */
4472 mp_dmevmask = mp->m_dmevmask;
4473 mp->m_dmevmask = 0;
4474
4475 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4476 /*
4477 * Find the agi for this ag.
4478 */
4479 error = xfs_read_agi(mp, NULL, agno, &agibp);
4480 if (error) {
4481 /*
4482 * AGI is b0rked. Don't process it.
4483 *
4484 * We should probably mark the filesystem as corrupt
4485 * after we've recovered all the ag's we can....
4486 */
4487 continue;
4488 }
4489 /*
4490 * Unlock the buffer so that it can be acquired in the normal
4491 * course of the transaction to truncate and free each inode.
4492 * Because we are not racing with anyone else here for the AGI
4493 * buffer, we don't even need to hold it locked to read the
4494 * initial unlinked bucket entries out of the buffer. We keep
4495 * buffer reference though, so that it stays pinned in memory
4496 * while we need the buffer.
4497 */
4498 agi = XFS_BUF_TO_AGI(agibp);
4499 xfs_buf_unlock(agibp);
4500
4501 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
4502 agino = be32_to_cpu(agi->agi_unlinked[bucket]);
4503 while (agino != NULLAGINO) {
4504 agino = xlog_recover_process_one_iunlink(mp,
4505 agno, agino, bucket);
4506 }
4507 }
4508 xfs_buf_rele(agibp);
4509 }
4510
4511 mp->m_dmevmask = mp_dmevmask;
4512}
4513
4514STATIC int
4515xlog_unpack_data(
4516 struct xlog_rec_header *rhead,
4517 char *dp,
4518 struct xlog *log)
4519{
4520 int i, j, k;
4521
4522 for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
4523 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
4524 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
4525 dp += BBSIZE;
4526 }
4527
4528 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
4529 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
4530 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
4531 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
4532 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
4533 *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
4534 dp += BBSIZE;
4535 }
4536 }
4537
4538 return 0;
4539}
4540
4541/*
4542 * CRC check, unpack and process a log record.
4543 */
4544STATIC int
4545xlog_recover_process(
4546 struct xlog *log,
4547 struct hlist_head rhash[],
4548 struct xlog_rec_header *rhead,
4549 char *dp,
4550 int pass)
4551{
4552 int error;
4553 __le32 crc;
4554
4555 crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
4556
4557 /*
4558 * Nothing else to do if this is a CRC verification pass. Just return
4559 * if this a record with a non-zero crc. Unfortunately, mkfs always
4560 * sets h_crc to 0 so we must consider this valid even on v5 supers.
4561 * Otherwise, return EFSBADCRC on failure so the callers up the stack
4562 * know precisely what failed.
4563 */
4564 if (pass == XLOG_RECOVER_CRCPASS) {
4565 if (rhead->h_crc && crc != rhead->h_crc)
4566 return -EFSBADCRC;
4567 return 0;
4568 }
4569
4570 /*
4571 * We're in the normal recovery path. Issue a warning if and only if the
4572 * CRC in the header is non-zero. This is an advisory warning and the
4573 * zero CRC check prevents warnings from being emitted when upgrading
4574 * the kernel from one that does not add CRCs by default.
4575 */
4576 if (crc != rhead->h_crc) {
4577 if (rhead->h_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
4578 xfs_alert(log->l_mp,
4579 "log record CRC mismatch: found 0x%x, expected 0x%x.",
4580 le32_to_cpu(rhead->h_crc),
4581 le32_to_cpu(crc));
4582 xfs_hex_dump(dp, 32);
4583 }
4584
4585 /*
4586 * If the filesystem is CRC enabled, this mismatch becomes a
4587 * fatal log corruption failure.
4588 */
4589 if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
4590 return -EFSCORRUPTED;
4591 }
4592
4593 error = xlog_unpack_data(rhead, dp, log);
4594 if (error)
4595 return error;
4596
4597 return xlog_recover_process_data(log, rhash, rhead, dp, pass);
4598}
4599
4600STATIC int
4601xlog_valid_rec_header(
4602 struct xlog *log,
4603 struct xlog_rec_header *rhead,
4604 xfs_daddr_t blkno)
4605{
4606 int hlen;
4607
4608 if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
4609 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
4610 XFS_ERRLEVEL_LOW, log->l_mp);
4611 return -EFSCORRUPTED;
4612 }
4613 if (unlikely(
4614 (!rhead->h_version ||
4615 (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
4616 xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
4617 __func__, be32_to_cpu(rhead->h_version));
4618 return -EIO;
4619 }
4620
4621 /* LR body must have data or it wouldn't have been written */
4622 hlen = be32_to_cpu(rhead->h_len);
4623 if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
4624 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
4625 XFS_ERRLEVEL_LOW, log->l_mp);
4626 return -EFSCORRUPTED;
4627 }
4628 if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
4629 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
4630 XFS_ERRLEVEL_LOW, log->l_mp);
4631 return -EFSCORRUPTED;
4632 }
4633 return 0;
4634}
4635
4636/*
4637 * Read the log from tail to head and process the log records found.
4638 * Handle the two cases where the tail and head are in the same cycle
4639 * and where the active portion of the log wraps around the end of
4640 * the physical log separately. The pass parameter is passed through
4641 * to the routines called to process the data and is not looked at
4642 * here.
4643 */
4644STATIC int
4645xlog_do_recovery_pass(
4646 struct xlog *log,
4647 xfs_daddr_t head_blk,
4648 xfs_daddr_t tail_blk,
4649 int pass,
4650 xfs_daddr_t *first_bad) /* out: first bad log rec */
4651{
4652 xlog_rec_header_t *rhead;
4653 xfs_daddr_t blk_no;
4654 xfs_daddr_t rhead_blk;
4655 char *offset;
4656 xfs_buf_t *hbp, *dbp;
4657 int error = 0, h_size, h_len;
4658 int bblks, split_bblks;
4659 int hblks, split_hblks, wrapped_hblks;
4660 struct hlist_head rhash[XLOG_RHASH_SIZE];
4661
4662 ASSERT(head_blk != tail_blk);
4663 rhead_blk = 0;
4664
4665 /*
4666 * Read the header of the tail block and get the iclog buffer size from
4667 * h_size. Use this to tell how many sectors make up the log header.
4668 */
4669 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
4670 /*
4671 * When using variable length iclogs, read first sector of
4672 * iclog header and extract the header size from it. Get a
4673 * new hbp that is the correct size.
4674 */
4675 hbp = xlog_get_bp(log, 1);
4676 if (!hbp)
4677 return -ENOMEM;
4678
4679 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
4680 if (error)
4681 goto bread_err1;
4682
4683 rhead = (xlog_rec_header_t *)offset;
4684 error = xlog_valid_rec_header(log, rhead, tail_blk);
4685 if (error)
4686 goto bread_err1;
4687
4688 /*
4689 * xfsprogs has a bug where record length is based on lsunit but
4690 * h_size (iclog size) is hardcoded to 32k. Now that we
4691 * unconditionally CRC verify the unmount record, this means the
4692 * log buffer can be too small for the record and cause an
4693 * overrun.
4694 *
4695 * Detect this condition here. Use lsunit for the buffer size as
4696 * long as this looks like the mkfs case. Otherwise, return an
4697 * error to avoid a buffer overrun.
4698 */
4699 h_size = be32_to_cpu(rhead->h_size);
4700 h_len = be32_to_cpu(rhead->h_len);
4701 if (h_len > h_size) {
4702 if (h_len <= log->l_mp->m_logbsize &&
4703 be32_to_cpu(rhead->h_num_logops) == 1) {
4704 xfs_warn(log->l_mp,
4705 "invalid iclog size (%d bytes), using lsunit (%d bytes)",
4706 h_size, log->l_mp->m_logbsize);
4707 h_size = log->l_mp->m_logbsize;
4708 } else
4709 return -EFSCORRUPTED;
4710 }
4711
4712 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
4713 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
4714 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
4715 if (h_size % XLOG_HEADER_CYCLE_SIZE)
4716 hblks++;
4717 xlog_put_bp(hbp);
4718 hbp = xlog_get_bp(log, hblks);
4719 } else {
4720 hblks = 1;
4721 }
4722 } else {
4723 ASSERT(log->l_sectBBsize == 1);
4724 hblks = 1;
4725 hbp = xlog_get_bp(log, 1);
4726 h_size = XLOG_BIG_RECORD_BSIZE;
4727 }
4728
4729 if (!hbp)
4730 return -ENOMEM;
4731 dbp = xlog_get_bp(log, BTOBB(h_size));
4732 if (!dbp) {
4733 xlog_put_bp(hbp);
4734 return -ENOMEM;
4735 }
4736
4737 memset(rhash, 0, sizeof(rhash));
4738 blk_no = rhead_blk = tail_blk;
4739 if (tail_blk > head_blk) {
4740 /*
4741 * Perform recovery around the end of the physical log.
4742 * When the head is not on the same cycle number as the tail,
4743 * we can't do a sequential recovery.
4744 */
4745 while (blk_no < log->l_logBBsize) {
4746 /*
4747 * Check for header wrapping around physical end-of-log
4748 */
4749 offset = hbp->b_addr;
4750 split_hblks = 0;
4751 wrapped_hblks = 0;
4752 if (blk_no + hblks <= log->l_logBBsize) {
4753 /* Read header in one read */
4754 error = xlog_bread(log, blk_no, hblks, hbp,
4755 &offset);
4756 if (error)
4757 goto bread_err2;
4758 } else {
4759 /* This LR is split across physical log end */
4760 if (blk_no != log->l_logBBsize) {
4761 /* some data before physical log end */
4762 ASSERT(blk_no <= INT_MAX);
4763 split_hblks = log->l_logBBsize - (int)blk_no;
4764 ASSERT(split_hblks > 0);
4765 error = xlog_bread(log, blk_no,
4766 split_hblks, hbp,
4767 &offset);
4768 if (error)
4769 goto bread_err2;
4770 }
4771
4772 /*
4773 * Note: this black magic still works with
4774 * large sector sizes (non-512) only because:
4775 * - we increased the buffer size originally
4776 * by 1 sector giving us enough extra space
4777 * for the second read;
4778 * - the log start is guaranteed to be sector
4779 * aligned;
4780 * - we read the log end (LR header start)
4781 * _first_, then the log start (LR header end)
4782 * - order is important.
4783 */
4784 wrapped_hblks = hblks - split_hblks;
4785 error = xlog_bread_offset(log, 0,
4786 wrapped_hblks, hbp,
4787 offset + BBTOB(split_hblks));
4788 if (error)
4789 goto bread_err2;
4790 }
4791 rhead = (xlog_rec_header_t *)offset;
4792 error = xlog_valid_rec_header(log, rhead,
4793 split_hblks ? blk_no : 0);
4794 if (error)
4795 goto bread_err2;
4796
4797 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
4798 blk_no += hblks;
4799
4800 /* Read in data for log record */
4801 if (blk_no + bblks <= log->l_logBBsize) {
4802 error = xlog_bread(log, blk_no, bblks, dbp,
4803 &offset);
4804 if (error)
4805 goto bread_err2;
4806 } else {
4807 /* This log record is split across the
4808 * physical end of log */
4809 offset = dbp->b_addr;
4810 split_bblks = 0;
4811 if (blk_no != log->l_logBBsize) {
4812 /* some data is before the physical
4813 * end of log */
4814 ASSERT(!wrapped_hblks);
4815 ASSERT(blk_no <= INT_MAX);
4816 split_bblks =
4817 log->l_logBBsize - (int)blk_no;
4818 ASSERT(split_bblks > 0);
4819 error = xlog_bread(log, blk_no,
4820 split_bblks, dbp,
4821 &offset);
4822 if (error)
4823 goto bread_err2;
4824 }
4825
4826 /*
4827 * Note: this black magic still works with
4828 * large sector sizes (non-512) only because:
4829 * - we increased the buffer size originally
4830 * by 1 sector giving us enough extra space
4831 * for the second read;
4832 * - the log start is guaranteed to be sector
4833 * aligned;
4834 * - we read the log end (LR header start)
4835 * _first_, then the log start (LR header end)
4836 * - order is important.
4837 */
4838 error = xlog_bread_offset(log, 0,
4839 bblks - split_bblks, dbp,
4840 offset + BBTOB(split_bblks));
4841 if (error)
4842 goto bread_err2;
4843 }
4844
4845 error = xlog_recover_process(log, rhash, rhead, offset,
4846 pass);
4847 if (error)
4848 goto bread_err2;
4849
4850 blk_no += bblks;
4851 rhead_blk = blk_no;
4852 }
4853
4854 ASSERT(blk_no >= log->l_logBBsize);
4855 blk_no -= log->l_logBBsize;
4856 rhead_blk = blk_no;
4857 }
4858
4859 /* read first part of physical log */
4860 while (blk_no < head_blk) {
4861 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
4862 if (error)
4863 goto bread_err2;
4864
4865 rhead = (xlog_rec_header_t *)offset;
4866 error = xlog_valid_rec_header(log, rhead, blk_no);
4867 if (error)
4868 goto bread_err2;
4869
4870 /* blocks in data section */
4871 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
4872 error = xlog_bread(log, blk_no+hblks, bblks, dbp,
4873 &offset);
4874 if (error)
4875 goto bread_err2;
4876
4877 error = xlog_recover_process(log, rhash, rhead, offset, pass);
4878 if (error)
4879 goto bread_err2;
4880
4881 blk_no += bblks + hblks;
4882 rhead_blk = blk_no;
4883 }
4884
4885 bread_err2:
4886 xlog_put_bp(dbp);
4887 bread_err1:
4888 xlog_put_bp(hbp);
4889
4890 if (error && first_bad)
4891 *first_bad = rhead_blk;
4892
4893 return error;
4894}
4895
4896/*
4897 * Do the recovery of the log. We actually do this in two phases.
4898 * The two passes are necessary in order to implement the function
4899 * of cancelling a record written into the log. The first pass
4900 * determines those things which have been cancelled, and the
4901 * second pass replays log items normally except for those which
4902 * have been cancelled. The handling of the replay and cancellations
4903 * takes place in the log item type specific routines.
4904 *
4905 * The table of items which have cancel records in the log is allocated
4906 * and freed at this level, since only here do we know when all of
4907 * the log recovery has been completed.
4908 */
4909STATIC int
4910xlog_do_log_recovery(
4911 struct xlog *log,
4912 xfs_daddr_t head_blk,
4913 xfs_daddr_t tail_blk)
4914{
4915 int error, i;
4916
4917 ASSERT(head_blk != tail_blk);
4918
4919 /*
4920 * First do a pass to find all of the cancelled buf log items.
4921 * Store them in the buf_cancel_table for use in the second pass.
4922 */
4923 log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
4924 sizeof(struct list_head),
4925 KM_SLEEP);
4926 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
4927 INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
4928
4929 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
4930 XLOG_RECOVER_PASS1, NULL);
4931 if (error != 0) {
4932 kmem_free(log->l_buf_cancel_table);
4933 log->l_buf_cancel_table = NULL;
4934 return error;
4935 }
4936 /*
4937 * Then do a second pass to actually recover the items in the log.
4938 * When it is complete free the table of buf cancel items.
4939 */
4940 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
4941 XLOG_RECOVER_PASS2, NULL);
4942#ifdef DEBUG
4943 if (!error) {
4944 int i;
4945
4946 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
4947 ASSERT(list_empty(&log->l_buf_cancel_table[i]));
4948 }
4949#endif /* DEBUG */
4950
4951 kmem_free(log->l_buf_cancel_table);
4952 log->l_buf_cancel_table = NULL;
4953
4954 return error;
4955}
4956
4957/*
4958 * Do the actual recovery
4959 */
4960STATIC int
4961xlog_do_recover(
4962 struct xlog *log,
4963 xfs_daddr_t head_blk,
4964 xfs_daddr_t tail_blk)
4965{
4966 struct xfs_mount *mp = log->l_mp;
4967 int error;
4968 xfs_buf_t *bp;
4969 xfs_sb_t *sbp;
4970
4971 /*
4972 * First replay the images in the log.
4973 */
4974 error = xlog_do_log_recovery(log, head_blk, tail_blk);
4975 if (error)
4976 return error;
4977
4978 /*
4979 * If IO errors happened during recovery, bail out.
4980 */
4981 if (XFS_FORCED_SHUTDOWN(mp)) {
4982 return -EIO;
4983 }
4984
4985 /*
4986 * We now update the tail_lsn since much of the recovery has completed
4987 * and there may be space available to use. If there were no extent
4988 * or iunlinks, we can free up the entire log and set the tail_lsn to
4989 * be the last_sync_lsn. This was set in xlog_find_tail to be the
4990 * lsn of the last known good LR on disk. If there are extent frees
4991 * or iunlinks they will have some entries in the AIL; so we look at
4992 * the AIL to determine how to set the tail_lsn.
4993 */
4994 xlog_assign_tail_lsn(mp);
4995
4996 /*
4997 * Now that we've finished replaying all buffer and inode
4998 * updates, re-read in the superblock and reverify it.
4999 */
5000 bp = xfs_getsb(mp, 0);
5001 bp->b_flags &= ~(XBF_DONE | XBF_ASYNC);
5002 ASSERT(!(bp->b_flags & XBF_WRITE));
5003 bp->b_flags |= XBF_READ;
5004 bp->b_ops = &xfs_sb_buf_ops;
5005
5006 error = xfs_buf_submit_wait(bp);
5007 if (error) {
5008 if (!XFS_FORCED_SHUTDOWN(mp)) {
5009 xfs_buf_ioerror_alert(bp, __func__);
5010 ASSERT(0);
5011 }
5012 xfs_buf_relse(bp);
5013 return error;
5014 }
5015
5016 /* Convert superblock from on-disk format */
5017 sbp = &mp->m_sb;
5018 xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
5019 xfs_buf_relse(bp);
5020
5021 /* re-initialise in-core superblock and geometry structures */
5022 xfs_reinit_percpu_counters(mp);
5023 error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
5024 if (error) {
5025 xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
5026 return error;
5027 }
5028
5029 xlog_recover_check_summary(log);
5030
5031 /* Normal transactions can now occur */
5032 log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
5033 return 0;
5034}
5035
5036/*
5037 * Perform recovery and re-initialize some log variables in xlog_find_tail.
5038 *
5039 * Return error or zero.
5040 */
5041int
5042xlog_recover(
5043 struct xlog *log)
5044{
5045 xfs_daddr_t head_blk, tail_blk;
5046 int error;
5047
5048 /* find the tail of the log */
5049 error = xlog_find_tail(log, &head_blk, &tail_blk);
5050 if (error)
5051 return error;
5052
5053 /*
5054 * The superblock was read before the log was available and thus the LSN
5055 * could not be verified. Check the superblock LSN against the current
5056 * LSN now that it's known.
5057 */
5058 if (xfs_sb_version_hascrc(&log->l_mp->m_sb) &&
5059 !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
5060 return -EINVAL;
5061
5062 if (tail_blk != head_blk) {
5063 /* There used to be a comment here:
5064 *
5065 * disallow recovery on read-only mounts. note -- mount
5066 * checks for ENOSPC and turns it into an intelligent
5067 * error message.
5068 * ...but this is no longer true. Now, unless you specify
5069 * NORECOVERY (in which case this function would never be
5070 * called), we just go ahead and recover. We do this all
5071 * under the vfs layer, so we can get away with it unless
5072 * the device itself is read-only, in which case we fail.
5073 */
5074 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
5075 return error;
5076 }
5077
5078 /*
5079 * Version 5 superblock log feature mask validation. We know the
5080 * log is dirty so check if there are any unknown log features
5081 * in what we need to recover. If there are unknown features
5082 * (e.g. unsupported transactions, then simply reject the
5083 * attempt at recovery before touching anything.
5084 */
5085 if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
5086 xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
5087 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
5088 xfs_warn(log->l_mp,
5089"Superblock has unknown incompatible log features (0x%x) enabled.",
5090 (log->l_mp->m_sb.sb_features_log_incompat &
5091 XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
5092 xfs_warn(log->l_mp,
5093"The log can not be fully and/or safely recovered by this kernel.");
5094 xfs_warn(log->l_mp,
5095"Please recover the log on a kernel that supports the unknown features.");
5096 return -EINVAL;
5097 }
5098
5099 /*
5100 * Delay log recovery if the debug hook is set. This is debug
5101 * instrumention to coordinate simulation of I/O failures with
5102 * log recovery.
5103 */
5104 if (xfs_globals.log_recovery_delay) {
5105 xfs_notice(log->l_mp,
5106 "Delaying log recovery for %d seconds.",
5107 xfs_globals.log_recovery_delay);
5108 msleep(xfs_globals.log_recovery_delay * 1000);
5109 }
5110
5111 xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
5112 log->l_mp->m_logname ? log->l_mp->m_logname
5113 : "internal");
5114
5115 error = xlog_do_recover(log, head_blk, tail_blk);
5116 log->l_flags |= XLOG_RECOVERY_NEEDED;
5117 }
5118 return error;
5119}
5120
5121/*
5122 * In the first part of recovery we replay inodes and buffers and build
5123 * up the list of extent free items which need to be processed. Here
5124 * we process the extent free items and clean up the on disk unlinked
5125 * inode lists. This is separated from the first part of recovery so
5126 * that the root and real-time bitmap inodes can be read in from disk in
5127 * between the two stages. This is necessary so that we can free space
5128 * in the real-time portion of the file system.
5129 */
5130int
5131xlog_recover_finish(
5132 struct xlog *log)
5133{
5134 /*
5135 * Now we're ready to do the transactions needed for the
5136 * rest of recovery. Start with completing all the extent
5137 * free intent records and then process the unlinked inode
5138 * lists. At this point, we essentially run in normal mode
5139 * except that we're still performing recovery actions
5140 * rather than accepting new requests.
5141 */
5142 if (log->l_flags & XLOG_RECOVERY_NEEDED) {
5143 int error;
5144 error = xlog_recover_process_efis(log);
5145 if (error) {
5146 xfs_alert(log->l_mp, "Failed to recover EFIs");
5147 return error;
5148 }
5149 /*
5150 * Sync the log to get all the EFIs out of the AIL.
5151 * This isn't absolutely necessary, but it helps in
5152 * case the unlink transactions would have problems
5153 * pushing the EFIs out of the way.
5154 */
5155 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
5156
5157 xlog_recover_process_iunlinks(log);
5158
5159 xlog_recover_check_summary(log);
5160
5161 xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
5162 log->l_mp->m_logname ? log->l_mp->m_logname
5163 : "internal");
5164 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
5165 } else {
5166 xfs_info(log->l_mp, "Ending clean mount");
5167 }
5168 return 0;
5169}
5170
5171int
5172xlog_recover_cancel(
5173 struct xlog *log)
5174{
5175 int error = 0;
5176
5177 if (log->l_flags & XLOG_RECOVERY_NEEDED)
5178 error = xlog_recover_cancel_efis(log);
5179
5180 return error;
5181}
5182
5183#if defined(DEBUG)
5184/*
5185 * Read all of the agf and agi counters and check that they
5186 * are consistent with the superblock counters.
5187 */
5188void
5189xlog_recover_check_summary(
5190 struct xlog *log)
5191{
5192 xfs_mount_t *mp;
5193 xfs_agf_t *agfp;
5194 xfs_buf_t *agfbp;
5195 xfs_buf_t *agibp;
5196 xfs_agnumber_t agno;
5197 __uint64_t freeblks;
5198 __uint64_t itotal;
5199 __uint64_t ifree;
5200 int error;
5201
5202 mp = log->l_mp;
5203
5204 freeblks = 0LL;
5205 itotal = 0LL;
5206 ifree = 0LL;
5207 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5208 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
5209 if (error) {
5210 xfs_alert(mp, "%s agf read failed agno %d error %d",
5211 __func__, agno, error);
5212 } else {
5213 agfp = XFS_BUF_TO_AGF(agfbp);
5214 freeblks += be32_to_cpu(agfp->agf_freeblks) +
5215 be32_to_cpu(agfp->agf_flcount);
5216 xfs_buf_relse(agfbp);
5217 }
5218
5219 error = xfs_read_agi(mp, NULL, agno, &agibp);
5220 if (error) {
5221 xfs_alert(mp, "%s agi read failed agno %d error %d",
5222 __func__, agno, error);
5223 } else {
5224 struct xfs_agi *agi = XFS_BUF_TO_AGI(agibp);
5225
5226 itotal += be32_to_cpu(agi->agi_count);
5227 ifree += be32_to_cpu(agi->agi_freecount);
5228 xfs_buf_relse(agibp);
5229 }
5230 }
5231}
5232#endif /* DEBUG */