xref: /linux/fs/xfs/xfs_log_recover.c (revision 8751b21ad9dc33f31dff20297dcae2063cbbcfc9)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4  * All Rights Reserved.
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_bit.h"
13 #include "xfs_sb.h"
14 #include "xfs_mount.h"
15 #include "xfs_defer.h"
16 #include "xfs_inode.h"
17 #include "xfs_trans.h"
18 #include "xfs_log.h"
19 #include "xfs_log_priv.h"
20 #include "xfs_log_recover.h"
21 #include "xfs_trans_priv.h"
22 #include "xfs_alloc.h"
23 #include "xfs_ialloc.h"
24 #include "xfs_trace.h"
25 #include "xfs_icache.h"
26 #include "xfs_error.h"
27 #include "xfs_buf_item.h"
28 #include "xfs_ag.h"
29 #include "xfs_quota.h"
30 #include "xfs_reflink.h"
31 
32 #define BLK_AVG(blk1, blk2)	((blk1+blk2) >> 1)
33 
34 STATIC int
35 xlog_find_zeroed(
36 	struct xlog	*,
37 	xfs_daddr_t	*);
38 STATIC int
39 xlog_clear_stale_blocks(
40 	struct xlog	*,
41 	xfs_lsn_t);
42 STATIC int
43 xlog_do_recovery_pass(
44         struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
45 
46 /*
47  * Sector aligned buffer routines for buffer create/read/write/access
48  */
49 
50 /*
51  * Verify the log-relative block number and length in basic blocks are valid for
52  * an operation involving the given XFS log buffer. Returns true if the fields
53  * are valid, false otherwise.
54  */
55 static inline bool
56 xlog_verify_bno(
57 	struct xlog	*log,
58 	xfs_daddr_t	blk_no,
59 	int		bbcount)
60 {
61 	if (blk_no < 0 || blk_no >= log->l_logBBsize)
62 		return false;
63 	if (bbcount <= 0 || (blk_no + bbcount) > log->l_logBBsize)
64 		return false;
65 	return true;
66 }
67 
68 /*
69  * Allocate a buffer to hold log data.  The buffer needs to be able to map to
70  * a range of nbblks basic blocks at any valid offset within the log.
71  */
72 static char *
73 xlog_alloc_buffer(
74 	struct xlog	*log,
75 	int		nbblks)
76 {
77 	/*
78 	 * Pass log block 0 since we don't have an addr yet, buffer will be
79 	 * verified on read.
80 	 */
81 	if (XFS_IS_CORRUPT(log->l_mp, !xlog_verify_bno(log, 0, nbblks))) {
82 		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
83 			nbblks);
84 		return NULL;
85 	}
86 
87 	/*
88 	 * We do log I/O in units of log sectors (a power-of-2 multiple of the
89 	 * basic block size), so we round up the requested size to accommodate
90 	 * the basic blocks required for complete log sectors.
91 	 *
92 	 * In addition, the buffer may be used for a non-sector-aligned block
93 	 * offset, in which case an I/O of the requested size could extend
94 	 * beyond the end of the buffer.  If the requested size is only 1 basic
95 	 * block it will never straddle a sector boundary, so this won't be an
96 	 * issue.  Nor will this be a problem if the log I/O is done in basic
97 	 * blocks (sector size 1).  But otherwise we extend the buffer by one
98 	 * extra log sector to ensure there's space to accommodate this
99 	 * possibility.
100 	 */
101 	if (nbblks > 1 && log->l_sectBBsize > 1)
102 		nbblks += log->l_sectBBsize;
103 	nbblks = round_up(nbblks, log->l_sectBBsize);
104 	return kvzalloc(BBTOB(nbblks), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
105 }
106 
107 /*
108  * Return the address of the start of the given block number's data
109  * in a log buffer.  The buffer covers a log sector-aligned region.
110  */
111 static inline unsigned int
112 xlog_align(
113 	struct xlog	*log,
114 	xfs_daddr_t	blk_no)
115 {
116 	return BBTOB(blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1));
117 }
118 
119 static int
120 xlog_do_io(
121 	struct xlog		*log,
122 	xfs_daddr_t		blk_no,
123 	unsigned int		nbblks,
124 	char			*data,
125 	enum req_op		op)
126 {
127 	int			error;
128 
129 	if (XFS_IS_CORRUPT(log->l_mp, !xlog_verify_bno(log, blk_no, nbblks))) {
130 		xfs_warn(log->l_mp,
131 			 "Invalid log block/length (0x%llx, 0x%x) for buffer",
132 			 blk_no, nbblks);
133 		return -EFSCORRUPTED;
134 	}
135 
136 	blk_no = round_down(blk_no, log->l_sectBBsize);
137 	nbblks = round_up(nbblks, log->l_sectBBsize);
138 	ASSERT(nbblks > 0);
139 
140 	error = xfs_rw_bdev(log->l_targ->bt_bdev, log->l_logBBstart + blk_no,
141 			BBTOB(nbblks), data, op);
142 	if (error && !xlog_is_shutdown(log)) {
143 		xfs_alert(log->l_mp,
144 			  "log recovery %s I/O error at daddr 0x%llx len %d error %d",
145 			  op == REQ_OP_WRITE ? "write" : "read",
146 			  blk_no, nbblks, error);
147 	}
148 	return error;
149 }
150 
151 STATIC int
152 xlog_bread_noalign(
153 	struct xlog	*log,
154 	xfs_daddr_t	blk_no,
155 	int		nbblks,
156 	char		*data)
157 {
158 	return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
159 }
160 
161 STATIC int
162 xlog_bread(
163 	struct xlog	*log,
164 	xfs_daddr_t	blk_no,
165 	int		nbblks,
166 	char		*data,
167 	char		**offset)
168 {
169 	int		error;
170 
171 	error = xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
172 	if (!error)
173 		*offset = data + xlog_align(log, blk_no);
174 	return error;
175 }
176 
177 STATIC int
178 xlog_bwrite(
179 	struct xlog	*log,
180 	xfs_daddr_t	blk_no,
181 	int		nbblks,
182 	char		*data)
183 {
184 	return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_WRITE);
185 }
186 
187 #ifdef DEBUG
188 /*
189  * dump debug superblock and log record information
190  */
191 STATIC void
192 xlog_header_check_dump(
193 	xfs_mount_t		*mp,
194 	xlog_rec_header_t	*head)
195 {
196 	xfs_debug(mp, "%s:  SB : uuid = %pU, fmt = %d",
197 		__func__, &mp->m_sb.sb_uuid, XLOG_FMT);
198 	xfs_debug(mp, "    log : uuid = %pU, fmt = %d",
199 		&head->h_fs_uuid, be32_to_cpu(head->h_fmt));
200 }
201 #else
202 #define xlog_header_check_dump(mp, head)
203 #endif
204 
205 /*
206  * check log record header for recovery
207  */
208 STATIC int
209 xlog_header_check_recover(
210 	xfs_mount_t		*mp,
211 	xlog_rec_header_t	*head)
212 {
213 	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
214 
215 	/*
216 	 * IRIX doesn't write the h_fmt field and leaves it zeroed
217 	 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
218 	 * a dirty log created in IRIX.
219 	 */
220 	if (XFS_IS_CORRUPT(mp, head->h_fmt != cpu_to_be32(XLOG_FMT))) {
221 		xfs_warn(mp,
222 	"dirty log written in incompatible format - can't recover");
223 		xlog_header_check_dump(mp, head);
224 		return -EFSCORRUPTED;
225 	}
226 	if (XFS_IS_CORRUPT(mp, !uuid_equal(&mp->m_sb.sb_uuid,
227 					   &head->h_fs_uuid))) {
228 		xfs_warn(mp,
229 	"dirty log entry has mismatched uuid - can't recover");
230 		xlog_header_check_dump(mp, head);
231 		return -EFSCORRUPTED;
232 	}
233 	return 0;
234 }
235 
236 /*
237  * read the head block of the log and check the header
238  */
239 STATIC int
240 xlog_header_check_mount(
241 	xfs_mount_t		*mp,
242 	xlog_rec_header_t	*head)
243 {
244 	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
245 
246 	if (uuid_is_null(&head->h_fs_uuid)) {
247 		/*
248 		 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
249 		 * h_fs_uuid is null, we assume this log was last mounted
250 		 * by IRIX and continue.
251 		 */
252 		xfs_warn(mp, "null uuid in log - IRIX style log");
253 	} else if (XFS_IS_CORRUPT(mp, !uuid_equal(&mp->m_sb.sb_uuid,
254 						  &head->h_fs_uuid))) {
255 		xfs_warn(mp, "log has mismatched uuid - can't recover");
256 		xlog_header_check_dump(mp, head);
257 		return -EFSCORRUPTED;
258 	}
259 	return 0;
260 }
261 
262 /*
263  * This routine finds (to an approximation) the first block in the physical
264  * log which contains the given cycle.  It uses a binary search algorithm.
265  * Note that the algorithm can not be perfect because the disk will not
266  * necessarily be perfect.
267  */
268 STATIC int
269 xlog_find_cycle_start(
270 	struct xlog	*log,
271 	char		*buffer,
272 	xfs_daddr_t	first_blk,
273 	xfs_daddr_t	*last_blk,
274 	uint		cycle)
275 {
276 	char		*offset;
277 	xfs_daddr_t	mid_blk;
278 	xfs_daddr_t	end_blk;
279 	uint		mid_cycle;
280 	int		error;
281 
282 	end_blk = *last_blk;
283 	mid_blk = BLK_AVG(first_blk, end_blk);
284 	while (mid_blk != first_blk && mid_blk != end_blk) {
285 		error = xlog_bread(log, mid_blk, 1, buffer, &offset);
286 		if (error)
287 			return error;
288 		mid_cycle = xlog_get_cycle(offset);
289 		if (mid_cycle == cycle)
290 			end_blk = mid_blk;   /* last_half_cycle == mid_cycle */
291 		else
292 			first_blk = mid_blk; /* first_half_cycle == mid_cycle */
293 		mid_blk = BLK_AVG(first_blk, end_blk);
294 	}
295 	ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
296 	       (mid_blk == end_blk && mid_blk-1 == first_blk));
297 
298 	*last_blk = end_blk;
299 
300 	return 0;
301 }
302 
303 /*
304  * Check that a range of blocks does not contain stop_on_cycle_no.
305  * Fill in *new_blk with the block offset where such a block is
306  * found, or with -1 (an invalid block number) if there is no such
307  * block in the range.  The scan needs to occur from front to back
308  * and the pointer into the region must be updated since a later
309  * routine will need to perform another test.
310  */
311 STATIC int
312 xlog_find_verify_cycle(
313 	struct xlog	*log,
314 	xfs_daddr_t	start_blk,
315 	int		nbblks,
316 	uint		stop_on_cycle_no,
317 	xfs_daddr_t	*new_blk)
318 {
319 	xfs_daddr_t	i, j;
320 	uint		cycle;
321 	char		*buffer;
322 	xfs_daddr_t	bufblks;
323 	char		*buf = NULL;
324 	int		error = 0;
325 
326 	/*
327 	 * Greedily allocate a buffer big enough to handle the full
328 	 * range of basic blocks we'll be examining.  If that fails,
329 	 * try a smaller size.  We need to be able to read at least
330 	 * a log sector, or we're out of luck.
331 	 */
332 	bufblks = roundup_pow_of_two(nbblks);
333 	while (bufblks > log->l_logBBsize)
334 		bufblks >>= 1;
335 	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
336 		bufblks >>= 1;
337 		if (bufblks < log->l_sectBBsize)
338 			return -ENOMEM;
339 	}
340 
341 	for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
342 		int	bcount;
343 
344 		bcount = min(bufblks, (start_blk + nbblks - i));
345 
346 		error = xlog_bread(log, i, bcount, buffer, &buf);
347 		if (error)
348 			goto out;
349 
350 		for (j = 0; j < bcount; j++) {
351 			cycle = xlog_get_cycle(buf);
352 			if (cycle == stop_on_cycle_no) {
353 				*new_blk = i+j;
354 				goto out;
355 			}
356 
357 			buf += BBSIZE;
358 		}
359 	}
360 
361 	*new_blk = -1;
362 
363 out:
364 	kvfree(buffer);
365 	return error;
366 }
367 
368 static inline int
369 xlog_logrec_hblks(struct xlog *log, struct xlog_rec_header *rh)
370 {
371 	if (xfs_has_logv2(log->l_mp)) {
372 		int	h_size = be32_to_cpu(rh->h_size);
373 
374 		if ((be32_to_cpu(rh->h_version) & XLOG_VERSION_2) &&
375 		    h_size > XLOG_HEADER_CYCLE_SIZE)
376 			return DIV_ROUND_UP(h_size, XLOG_HEADER_CYCLE_SIZE);
377 	}
378 	return 1;
379 }
380 
381 /*
382  * Potentially backup over partial log record write.
383  *
384  * In the typical case, last_blk is the number of the block directly after
385  * a good log record.  Therefore, we subtract one to get the block number
386  * of the last block in the given buffer.  extra_bblks contains the number
387  * of blocks we would have read on a previous read.  This happens when the
388  * last log record is split over the end of the physical log.
389  *
390  * extra_bblks is the number of blocks potentially verified on a previous
391  * call to this routine.
392  */
393 STATIC int
394 xlog_find_verify_log_record(
395 	struct xlog		*log,
396 	xfs_daddr_t		start_blk,
397 	xfs_daddr_t		*last_blk,
398 	int			extra_bblks)
399 {
400 	xfs_daddr_t		i;
401 	char			*buffer;
402 	char			*offset = NULL;
403 	xlog_rec_header_t	*head = NULL;
404 	int			error = 0;
405 	int			smallmem = 0;
406 	int			num_blks = *last_blk - start_blk;
407 	int			xhdrs;
408 
409 	ASSERT(start_blk != 0 || *last_blk != start_blk);
410 
411 	buffer = xlog_alloc_buffer(log, num_blks);
412 	if (!buffer) {
413 		buffer = xlog_alloc_buffer(log, 1);
414 		if (!buffer)
415 			return -ENOMEM;
416 		smallmem = 1;
417 	} else {
418 		error = xlog_bread(log, start_blk, num_blks, buffer, &offset);
419 		if (error)
420 			goto out;
421 		offset += ((num_blks - 1) << BBSHIFT);
422 	}
423 
424 	for (i = (*last_blk) - 1; i >= 0; i--) {
425 		if (i < start_blk) {
426 			/* valid log record not found */
427 			xfs_warn(log->l_mp,
428 		"Log inconsistent (didn't find previous header)");
429 			ASSERT(0);
430 			error = -EFSCORRUPTED;
431 			goto out;
432 		}
433 
434 		if (smallmem) {
435 			error = xlog_bread(log, i, 1, buffer, &offset);
436 			if (error)
437 				goto out;
438 		}
439 
440 		head = (xlog_rec_header_t *)offset;
441 
442 		if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
443 			break;
444 
445 		if (!smallmem)
446 			offset -= BBSIZE;
447 	}
448 
449 	/*
450 	 * We hit the beginning of the physical log & still no header.  Return
451 	 * to caller.  If caller can handle a return of -1, then this routine
452 	 * will be called again for the end of the physical log.
453 	 */
454 	if (i == -1) {
455 		error = 1;
456 		goto out;
457 	}
458 
459 	/*
460 	 * We have the final block of the good log (the first block
461 	 * of the log record _before_ the head. So we check the uuid.
462 	 */
463 	if ((error = xlog_header_check_mount(log->l_mp, head)))
464 		goto out;
465 
466 	/*
467 	 * We may have found a log record header before we expected one.
468 	 * last_blk will be the 1st block # with a given cycle #.  We may end
469 	 * up reading an entire log record.  In this case, we don't want to
470 	 * reset last_blk.  Only when last_blk points in the middle of a log
471 	 * record do we update last_blk.
472 	 */
473 	xhdrs = xlog_logrec_hblks(log, head);
474 
475 	if (*last_blk - i + extra_bblks !=
476 	    BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
477 		*last_blk = i;
478 
479 out:
480 	kvfree(buffer);
481 	return error;
482 }
483 
484 /*
485  * Head is defined to be the point of the log where the next log write
486  * could go.  This means that incomplete LR writes at the end are
487  * eliminated when calculating the head.  We aren't guaranteed that previous
488  * LR have complete transactions.  We only know that a cycle number of
489  * current cycle number -1 won't be present in the log if we start writing
490  * from our current block number.
491  *
492  * last_blk contains the block number of the first block with a given
493  * cycle number.
494  *
495  * Return: zero if normal, non-zero if error.
496  */
497 STATIC int
498 xlog_find_head(
499 	struct xlog	*log,
500 	xfs_daddr_t	*return_head_blk)
501 {
502 	char		*buffer;
503 	char		*offset;
504 	xfs_daddr_t	new_blk, first_blk, start_blk, last_blk, head_blk;
505 	int		num_scan_bblks;
506 	uint		first_half_cycle, last_half_cycle;
507 	uint		stop_on_cycle;
508 	int		error, log_bbnum = log->l_logBBsize;
509 
510 	/* Is the end of the log device zeroed? */
511 	error = xlog_find_zeroed(log, &first_blk);
512 	if (error < 0) {
513 		xfs_warn(log->l_mp, "empty log check failed");
514 		return error;
515 	}
516 	if (error == 1) {
517 		*return_head_blk = first_blk;
518 
519 		/* Is the whole lot zeroed? */
520 		if (!first_blk) {
521 			/* Linux XFS shouldn't generate totally zeroed logs -
522 			 * mkfs etc write a dummy unmount record to a fresh
523 			 * log so we can store the uuid in there
524 			 */
525 			xfs_warn(log->l_mp, "totally zeroed log");
526 		}
527 
528 		return 0;
529 	}
530 
531 	first_blk = 0;			/* get cycle # of 1st block */
532 	buffer = xlog_alloc_buffer(log, 1);
533 	if (!buffer)
534 		return -ENOMEM;
535 
536 	error = xlog_bread(log, 0, 1, buffer, &offset);
537 	if (error)
538 		goto out_free_buffer;
539 
540 	first_half_cycle = xlog_get_cycle(offset);
541 
542 	last_blk = head_blk = log_bbnum - 1;	/* get cycle # of last block */
543 	error = xlog_bread(log, last_blk, 1, buffer, &offset);
544 	if (error)
545 		goto out_free_buffer;
546 
547 	last_half_cycle = xlog_get_cycle(offset);
548 	ASSERT(last_half_cycle != 0);
549 
550 	/*
551 	 * If the 1st half cycle number is equal to the last half cycle number,
552 	 * then the entire log is stamped with the same cycle number.  In this
553 	 * case, head_blk can't be set to zero (which makes sense).  The below
554 	 * math doesn't work out properly with head_blk equal to zero.  Instead,
555 	 * we set it to log_bbnum which is an invalid block number, but this
556 	 * value makes the math correct.  If head_blk doesn't changed through
557 	 * all the tests below, *head_blk is set to zero at the very end rather
558 	 * than log_bbnum.  In a sense, log_bbnum and zero are the same block
559 	 * in a circular file.
560 	 */
561 	if (first_half_cycle == last_half_cycle) {
562 		/*
563 		 * In this case we believe that the entire log should have
564 		 * cycle number last_half_cycle.  We need to scan backwards
565 		 * from the end verifying that there are no holes still
566 		 * containing last_half_cycle - 1.  If we find such a hole,
567 		 * then the start of that hole will be the new head.  The
568 		 * simple case looks like
569 		 *        x | x ... | x - 1 | x
570 		 * Another case that fits this picture would be
571 		 *        x | x + 1 | x ... | x
572 		 * In this case the head really is somewhere at the end of the
573 		 * log, as one of the latest writes at the beginning was
574 		 * incomplete.
575 		 * One more case is
576 		 *        x | x + 1 | x ... | x - 1 | x
577 		 * This is really the combination of the above two cases, and
578 		 * the head has to end up at the start of the x-1 hole at the
579 		 * end of the log.
580 		 *
581 		 * In the 256k log case, we will read from the beginning to the
582 		 * end of the log and search for cycle numbers equal to x-1.
583 		 * We don't worry about the x+1 blocks that we encounter,
584 		 * because we know that they cannot be the head since the log
585 		 * started with x.
586 		 */
587 		head_blk = log_bbnum;
588 		stop_on_cycle = last_half_cycle - 1;
589 	} else {
590 		/*
591 		 * In this case we want to find the first block with cycle
592 		 * number matching last_half_cycle.  We expect the log to be
593 		 * some variation on
594 		 *        x + 1 ... | x ... | x
595 		 * The first block with cycle number x (last_half_cycle) will
596 		 * be where the new head belongs.  First we do a binary search
597 		 * for the first occurrence of last_half_cycle.  The binary
598 		 * search may not be totally accurate, so then we scan back
599 		 * from there looking for occurrences of last_half_cycle before
600 		 * us.  If that backwards scan wraps around the beginning of
601 		 * the log, then we look for occurrences of last_half_cycle - 1
602 		 * at the end of the log.  The cases we're looking for look
603 		 * like
604 		 *                               v binary search stopped here
605 		 *        x + 1 ... | x | x + 1 | x ... | x
606 		 *                   ^ but we want to locate this spot
607 		 * or
608 		 *        <---------> less than scan distance
609 		 *        x + 1 ... | x ... | x - 1 | x
610 		 *                           ^ we want to locate this spot
611 		 */
612 		stop_on_cycle = last_half_cycle;
613 		error = xlog_find_cycle_start(log, buffer, first_blk, &head_blk,
614 				last_half_cycle);
615 		if (error)
616 			goto out_free_buffer;
617 	}
618 
619 	/*
620 	 * Now validate the answer.  Scan back some number of maximum possible
621 	 * blocks and make sure each one has the expected cycle number.  The
622 	 * maximum is determined by the total possible amount of buffering
623 	 * in the in-core log.  The following number can be made tighter if
624 	 * we actually look at the block size of the filesystem.
625 	 */
626 	num_scan_bblks = min_t(int, log_bbnum, XLOG_TOTAL_REC_SHIFT(log));
627 	if (head_blk >= num_scan_bblks) {
628 		/*
629 		 * We are guaranteed that the entire check can be performed
630 		 * in one buffer.
631 		 */
632 		start_blk = head_blk - num_scan_bblks;
633 		if ((error = xlog_find_verify_cycle(log,
634 						start_blk, num_scan_bblks,
635 						stop_on_cycle, &new_blk)))
636 			goto out_free_buffer;
637 		if (new_blk != -1)
638 			head_blk = new_blk;
639 	} else {		/* need to read 2 parts of log */
640 		/*
641 		 * We are going to scan backwards in the log in two parts.
642 		 * First we scan the physical end of the log.  In this part
643 		 * of the log, we are looking for blocks with cycle number
644 		 * last_half_cycle - 1.
645 		 * If we find one, then we know that the log starts there, as
646 		 * we've found a hole that didn't get written in going around
647 		 * the end of the physical log.  The simple case for this is
648 		 *        x + 1 ... | x ... | x - 1 | x
649 		 *        <---------> less than scan distance
650 		 * If all of the blocks at the end of the log have cycle number
651 		 * last_half_cycle, then we check the blocks at the start of
652 		 * the log looking for occurrences of last_half_cycle.  If we
653 		 * find one, then our current estimate for the location of the
654 		 * first occurrence of last_half_cycle is wrong and we move
655 		 * back to the hole we've found.  This case looks like
656 		 *        x + 1 ... | x | x + 1 | x ...
657 		 *                               ^ binary search stopped here
658 		 * Another case we need to handle that only occurs in 256k
659 		 * logs is
660 		 *        x + 1 ... | x ... | x+1 | x ...
661 		 *                   ^ binary search stops here
662 		 * In a 256k log, the scan at the end of the log will see the
663 		 * x + 1 blocks.  We need to skip past those since that is
664 		 * certainly not the head of the log.  By searching for
665 		 * last_half_cycle-1 we accomplish that.
666 		 */
667 		ASSERT(head_blk <= INT_MAX &&
668 			(xfs_daddr_t) num_scan_bblks >= head_blk);
669 		start_blk = log_bbnum - (num_scan_bblks - head_blk);
670 		if ((error = xlog_find_verify_cycle(log, start_blk,
671 					num_scan_bblks - (int)head_blk,
672 					(stop_on_cycle - 1), &new_blk)))
673 			goto out_free_buffer;
674 		if (new_blk != -1) {
675 			head_blk = new_blk;
676 			goto validate_head;
677 		}
678 
679 		/*
680 		 * Scan beginning of log now.  The last part of the physical
681 		 * log is good.  This scan needs to verify that it doesn't find
682 		 * the last_half_cycle.
683 		 */
684 		start_blk = 0;
685 		ASSERT(head_blk <= INT_MAX);
686 		if ((error = xlog_find_verify_cycle(log,
687 					start_blk, (int)head_blk,
688 					stop_on_cycle, &new_blk)))
689 			goto out_free_buffer;
690 		if (new_blk != -1)
691 			head_blk = new_blk;
692 	}
693 
694 validate_head:
695 	/*
696 	 * Now we need to make sure head_blk is not pointing to a block in
697 	 * the middle of a log record.
698 	 */
699 	num_scan_bblks = XLOG_REC_SHIFT(log);
700 	if (head_blk >= num_scan_bblks) {
701 		start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
702 
703 		/* start ptr at last block ptr before head_blk */
704 		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
705 		if (error == 1)
706 			error = -EIO;
707 		if (error)
708 			goto out_free_buffer;
709 	} else {
710 		start_blk = 0;
711 		ASSERT(head_blk <= INT_MAX);
712 		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
713 		if (error < 0)
714 			goto out_free_buffer;
715 		if (error == 1) {
716 			/* We hit the beginning of the log during our search */
717 			start_blk = log_bbnum - (num_scan_bblks - head_blk);
718 			new_blk = log_bbnum;
719 			ASSERT(start_blk <= INT_MAX &&
720 				(xfs_daddr_t) log_bbnum-start_blk >= 0);
721 			ASSERT(head_blk <= INT_MAX);
722 			error = xlog_find_verify_log_record(log, start_blk,
723 							&new_blk, (int)head_blk);
724 			if (error == 1)
725 				error = -EIO;
726 			if (error)
727 				goto out_free_buffer;
728 			if (new_blk != log_bbnum)
729 				head_blk = new_blk;
730 		} else if (error)
731 			goto out_free_buffer;
732 	}
733 
734 	kvfree(buffer);
735 	if (head_blk == log_bbnum)
736 		*return_head_blk = 0;
737 	else
738 		*return_head_blk = head_blk;
739 	/*
740 	 * When returning here, we have a good block number.  Bad block
741 	 * means that during a previous crash, we didn't have a clean break
742 	 * from cycle number N to cycle number N-1.  In this case, we need
743 	 * to find the first block with cycle number N-1.
744 	 */
745 	return 0;
746 
747 out_free_buffer:
748 	kvfree(buffer);
749 	if (error)
750 		xfs_warn(log->l_mp, "failed to find log head");
751 	return error;
752 }
753 
754 /*
755  * Seek backwards in the log for log record headers.
756  *
757  * Given a starting log block, walk backwards until we find the provided number
758  * of records or hit the provided tail block. The return value is the number of
759  * records encountered or a negative error code. The log block and buffer
760  * pointer of the last record seen are returned in rblk and rhead respectively.
761  */
762 STATIC int
763 xlog_rseek_logrec_hdr(
764 	struct xlog		*log,
765 	xfs_daddr_t		head_blk,
766 	xfs_daddr_t		tail_blk,
767 	int			count,
768 	char			*buffer,
769 	xfs_daddr_t		*rblk,
770 	struct xlog_rec_header	**rhead,
771 	bool			*wrapped)
772 {
773 	int			i;
774 	int			error;
775 	int			found = 0;
776 	char			*offset = NULL;
777 	xfs_daddr_t		end_blk;
778 
779 	*wrapped = false;
780 
781 	/*
782 	 * Walk backwards from the head block until we hit the tail or the first
783 	 * block in the log.
784 	 */
785 	end_blk = head_blk > tail_blk ? tail_blk : 0;
786 	for (i = (int) head_blk - 1; i >= end_blk; i--) {
787 		error = xlog_bread(log, i, 1, buffer, &offset);
788 		if (error)
789 			goto out_error;
790 
791 		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
792 			*rblk = i;
793 			*rhead = (struct xlog_rec_header *) offset;
794 			if (++found == count)
795 				break;
796 		}
797 	}
798 
799 	/*
800 	 * If we haven't hit the tail block or the log record header count,
801 	 * start looking again from the end of the physical log. Note that
802 	 * callers can pass head == tail if the tail is not yet known.
803 	 */
804 	if (tail_blk >= head_blk && found != count) {
805 		for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
806 			error = xlog_bread(log, i, 1, buffer, &offset);
807 			if (error)
808 				goto out_error;
809 
810 			if (*(__be32 *)offset ==
811 			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
812 				*wrapped = true;
813 				*rblk = i;
814 				*rhead = (struct xlog_rec_header *) offset;
815 				if (++found == count)
816 					break;
817 			}
818 		}
819 	}
820 
821 	return found;
822 
823 out_error:
824 	return error;
825 }
826 
827 /*
828  * Seek forward in the log for log record headers.
829  *
830  * Given head and tail blocks, walk forward from the tail block until we find
831  * the provided number of records or hit the head block. The return value is the
832  * number of records encountered or a negative error code. The log block and
833  * buffer pointer of the last record seen are returned in rblk and rhead
834  * respectively.
835  */
836 STATIC int
837 xlog_seek_logrec_hdr(
838 	struct xlog		*log,
839 	xfs_daddr_t		head_blk,
840 	xfs_daddr_t		tail_blk,
841 	int			count,
842 	char			*buffer,
843 	xfs_daddr_t		*rblk,
844 	struct xlog_rec_header	**rhead,
845 	bool			*wrapped)
846 {
847 	int			i;
848 	int			error;
849 	int			found = 0;
850 	char			*offset = NULL;
851 	xfs_daddr_t		end_blk;
852 
853 	*wrapped = false;
854 
855 	/*
856 	 * Walk forward from the tail block until we hit the head or the last
857 	 * block in the log.
858 	 */
859 	end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
860 	for (i = (int) tail_blk; i <= end_blk; i++) {
861 		error = xlog_bread(log, i, 1, buffer, &offset);
862 		if (error)
863 			goto out_error;
864 
865 		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
866 			*rblk = i;
867 			*rhead = (struct xlog_rec_header *) offset;
868 			if (++found == count)
869 				break;
870 		}
871 	}
872 
873 	/*
874 	 * If we haven't hit the head block or the log record header count,
875 	 * start looking again from the start of the physical log.
876 	 */
877 	if (tail_blk > head_blk && found != count) {
878 		for (i = 0; i < (int) head_blk; i++) {
879 			error = xlog_bread(log, i, 1, buffer, &offset);
880 			if (error)
881 				goto out_error;
882 
883 			if (*(__be32 *)offset ==
884 			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
885 				*wrapped = true;
886 				*rblk = i;
887 				*rhead = (struct xlog_rec_header *) offset;
888 				if (++found == count)
889 					break;
890 			}
891 		}
892 	}
893 
894 	return found;
895 
896 out_error:
897 	return error;
898 }
899 
900 /*
901  * Calculate distance from head to tail (i.e., unused space in the log).
902  */
903 static inline int
904 xlog_tail_distance(
905 	struct xlog	*log,
906 	xfs_daddr_t	head_blk,
907 	xfs_daddr_t	tail_blk)
908 {
909 	if (head_blk < tail_blk)
910 		return tail_blk - head_blk;
911 
912 	return tail_blk + (log->l_logBBsize - head_blk);
913 }
914 
915 /*
916  * Verify the log tail. This is particularly important when torn or incomplete
917  * writes have been detected near the front of the log and the head has been
918  * walked back accordingly.
919  *
920  * We also have to handle the case where the tail was pinned and the head
921  * blocked behind the tail right before a crash. If the tail had been pushed
922  * immediately prior to the crash and the subsequent checkpoint was only
923  * partially written, it's possible it overwrote the last referenced tail in the
924  * log with garbage. This is not a coherency problem because the tail must have
925  * been pushed before it can be overwritten, but appears as log corruption to
926  * recovery because we have no way to know the tail was updated if the
927  * subsequent checkpoint didn't write successfully.
928  *
929  * Therefore, CRC check the log from tail to head. If a failure occurs and the
930  * offending record is within max iclog bufs from the head, walk the tail
931  * forward and retry until a valid tail is found or corruption is detected out
932  * of the range of a possible overwrite.
933  */
934 STATIC int
935 xlog_verify_tail(
936 	struct xlog		*log,
937 	xfs_daddr_t		head_blk,
938 	xfs_daddr_t		*tail_blk,
939 	int			hsize)
940 {
941 	struct xlog_rec_header	*thead;
942 	char			*buffer;
943 	xfs_daddr_t		first_bad;
944 	int			error = 0;
945 	bool			wrapped;
946 	xfs_daddr_t		tmp_tail;
947 	xfs_daddr_t		orig_tail = *tail_blk;
948 
949 	buffer = xlog_alloc_buffer(log, 1);
950 	if (!buffer)
951 		return -ENOMEM;
952 
953 	/*
954 	 * Make sure the tail points to a record (returns positive count on
955 	 * success).
956 	 */
957 	error = xlog_seek_logrec_hdr(log, head_blk, *tail_blk, 1, buffer,
958 			&tmp_tail, &thead, &wrapped);
959 	if (error < 0)
960 		goto out;
961 	if (*tail_blk != tmp_tail)
962 		*tail_blk = tmp_tail;
963 
964 	/*
965 	 * Run a CRC check from the tail to the head. We can't just check
966 	 * MAX_ICLOGS records past the tail because the tail may point to stale
967 	 * blocks cleared during the search for the head/tail. These blocks are
968 	 * overwritten with zero-length records and thus record count is not a
969 	 * reliable indicator of the iclog state before a crash.
970 	 */
971 	first_bad = 0;
972 	error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
973 				      XLOG_RECOVER_CRCPASS, &first_bad);
974 	while ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
975 		int	tail_distance;
976 
977 		/*
978 		 * Is corruption within range of the head? If so, retry from
979 		 * the next record. Otherwise return an error.
980 		 */
981 		tail_distance = xlog_tail_distance(log, head_blk, first_bad);
982 		if (tail_distance > BTOBB(XLOG_MAX_ICLOGS * hsize))
983 			break;
984 
985 		/* skip to the next record; returns positive count on success */
986 		error = xlog_seek_logrec_hdr(log, head_blk, first_bad, 2,
987 				buffer, &tmp_tail, &thead, &wrapped);
988 		if (error < 0)
989 			goto out;
990 
991 		*tail_blk = tmp_tail;
992 		first_bad = 0;
993 		error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
994 					      XLOG_RECOVER_CRCPASS, &first_bad);
995 	}
996 
997 	if (!error && *tail_blk != orig_tail)
998 		xfs_warn(log->l_mp,
999 		"Tail block (0x%llx) overwrite detected. Updated to 0x%llx",
1000 			 orig_tail, *tail_blk);
1001 out:
1002 	kvfree(buffer);
1003 	return error;
1004 }
1005 
1006 /*
1007  * Detect and trim torn writes from the head of the log.
1008  *
1009  * Storage without sector atomicity guarantees can result in torn writes in the
1010  * log in the event of a crash. Our only means to detect this scenario is via
1011  * CRC verification. While we can't always be certain that CRC verification
1012  * failure is due to a torn write vs. an unrelated corruption, we do know that
1013  * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1014  * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1015  * the log and treat failures in this range as torn writes as a matter of
1016  * policy. In the event of CRC failure, the head is walked back to the last good
1017  * record in the log and the tail is updated from that record and verified.
1018  */
1019 STATIC int
1020 xlog_verify_head(
1021 	struct xlog		*log,
1022 	xfs_daddr_t		*head_blk,	/* in/out: unverified head */
1023 	xfs_daddr_t		*tail_blk,	/* out: tail block */
1024 	char			*buffer,
1025 	xfs_daddr_t		*rhead_blk,	/* start blk of last record */
1026 	struct xlog_rec_header	**rhead,	/* ptr to last record */
1027 	bool			*wrapped)	/* last rec. wraps phys. log */
1028 {
1029 	struct xlog_rec_header	*tmp_rhead;
1030 	char			*tmp_buffer;
1031 	xfs_daddr_t		first_bad;
1032 	xfs_daddr_t		tmp_rhead_blk;
1033 	int			found;
1034 	int			error;
1035 	bool			tmp_wrapped;
1036 
1037 	/*
1038 	 * Check the head of the log for torn writes. Search backwards from the
1039 	 * head until we hit the tail or the maximum number of log record I/Os
1040 	 * that could have been in flight at one time. Use a temporary buffer so
1041 	 * we don't trash the rhead/buffer pointers from the caller.
1042 	 */
1043 	tmp_buffer = xlog_alloc_buffer(log, 1);
1044 	if (!tmp_buffer)
1045 		return -ENOMEM;
1046 	error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1047 				      XLOG_MAX_ICLOGS, tmp_buffer,
1048 				      &tmp_rhead_blk, &tmp_rhead, &tmp_wrapped);
1049 	kvfree(tmp_buffer);
1050 	if (error < 0)
1051 		return error;
1052 
1053 	/*
1054 	 * Now run a CRC verification pass over the records starting at the
1055 	 * block found above to the current head. If a CRC failure occurs, the
1056 	 * log block of the first bad record is saved in first_bad.
1057 	 */
1058 	error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1059 				      XLOG_RECOVER_CRCPASS, &first_bad);
1060 	if ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1061 		/*
1062 		 * We've hit a potential torn write. Reset the error and warn
1063 		 * about it.
1064 		 */
1065 		error = 0;
1066 		xfs_warn(log->l_mp,
1067 "Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1068 			 first_bad, *head_blk);
1069 
1070 		/*
1071 		 * Get the header block and buffer pointer for the last good
1072 		 * record before the bad record.
1073 		 *
1074 		 * Note that xlog_find_tail() clears the blocks at the new head
1075 		 * (i.e., the records with invalid CRC) if the cycle number
1076 		 * matches the current cycle.
1077 		 */
1078 		found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1,
1079 				buffer, rhead_blk, rhead, wrapped);
1080 		if (found < 0)
1081 			return found;
1082 		if (found == 0)		/* XXX: right thing to do here? */
1083 			return -EIO;
1084 
1085 		/*
1086 		 * Reset the head block to the starting block of the first bad
1087 		 * log record and set the tail block based on the last good
1088 		 * record.
1089 		 *
1090 		 * Bail out if the updated head/tail match as this indicates
1091 		 * possible corruption outside of the acceptable
1092 		 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1093 		 */
1094 		*head_blk = first_bad;
1095 		*tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1096 		if (*head_blk == *tail_blk) {
1097 			ASSERT(0);
1098 			return 0;
1099 		}
1100 	}
1101 	if (error)
1102 		return error;
1103 
1104 	return xlog_verify_tail(log, *head_blk, tail_blk,
1105 				be32_to_cpu((*rhead)->h_size));
1106 }
1107 
1108 /*
1109  * We need to make sure we handle log wrapping properly, so we can't use the
1110  * calculated logbno directly. Make sure it wraps to the correct bno inside the
1111  * log.
1112  *
1113  * The log is limited to 32 bit sizes, so we use the appropriate modulus
1114  * operation here and cast it back to a 64 bit daddr on return.
1115  */
1116 static inline xfs_daddr_t
1117 xlog_wrap_logbno(
1118 	struct xlog		*log,
1119 	xfs_daddr_t		bno)
1120 {
1121 	int			mod;
1122 
1123 	div_s64_rem(bno, log->l_logBBsize, &mod);
1124 	return mod;
1125 }
1126 
1127 /*
1128  * Check whether the head of the log points to an unmount record. In other
1129  * words, determine whether the log is clean. If so, update the in-core state
1130  * appropriately.
1131  */
1132 static int
1133 xlog_check_unmount_rec(
1134 	struct xlog		*log,
1135 	xfs_daddr_t		*head_blk,
1136 	xfs_daddr_t		*tail_blk,
1137 	struct xlog_rec_header	*rhead,
1138 	xfs_daddr_t		rhead_blk,
1139 	char			*buffer,
1140 	bool			*clean)
1141 {
1142 	struct xlog_op_header	*op_head;
1143 	xfs_daddr_t		umount_data_blk;
1144 	xfs_daddr_t		after_umount_blk;
1145 	int			hblks;
1146 	int			error;
1147 	char			*offset;
1148 
1149 	*clean = false;
1150 
1151 	/*
1152 	 * Look for unmount record. If we find it, then we know there was a
1153 	 * clean unmount. Since 'i' could be the last block in the physical
1154 	 * log, we convert to a log block before comparing to the head_blk.
1155 	 *
1156 	 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1157 	 * below. We won't want to clear the unmount record if there is one, so
1158 	 * we pass the lsn of the unmount record rather than the block after it.
1159 	 */
1160 	hblks = xlog_logrec_hblks(log, rhead);
1161 	after_umount_blk = xlog_wrap_logbno(log,
1162 			rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len)));
1163 
1164 	if (*head_blk == after_umount_blk &&
1165 	    be32_to_cpu(rhead->h_num_logops) == 1) {
1166 		umount_data_blk = xlog_wrap_logbno(log, rhead_blk + hblks);
1167 		error = xlog_bread(log, umount_data_blk, 1, buffer, &offset);
1168 		if (error)
1169 			return error;
1170 
1171 		op_head = (struct xlog_op_header *)offset;
1172 		if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1173 			/*
1174 			 * Set tail and last sync so that newly written log
1175 			 * records will point recovery to after the current
1176 			 * unmount record.
1177 			 */
1178 			xlog_assign_atomic_lsn(&log->l_tail_lsn,
1179 					log->l_curr_cycle, after_umount_blk);
1180 			log->l_ailp->ail_head_lsn =
1181 					atomic64_read(&log->l_tail_lsn);
1182 			*tail_blk = after_umount_blk;
1183 
1184 			*clean = true;
1185 		}
1186 	}
1187 
1188 	return 0;
1189 }
1190 
1191 static void
1192 xlog_set_state(
1193 	struct xlog		*log,
1194 	xfs_daddr_t		head_blk,
1195 	struct xlog_rec_header	*rhead,
1196 	xfs_daddr_t		rhead_blk,
1197 	bool			bump_cycle)
1198 {
1199 	/*
1200 	 * Reset log values according to the state of the log when we
1201 	 * crashed.  In the case where head_blk == 0, we bump curr_cycle
1202 	 * one because the next write starts a new cycle rather than
1203 	 * continuing the cycle of the last good log record.  At this
1204 	 * point we have guaranteed that all partial log records have been
1205 	 * accounted for.  Therefore, we know that the last good log record
1206 	 * written was complete and ended exactly on the end boundary
1207 	 * of the physical log.
1208 	 */
1209 	log->l_prev_block = rhead_blk;
1210 	log->l_curr_block = (int)head_blk;
1211 	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1212 	if (bump_cycle)
1213 		log->l_curr_cycle++;
1214 	atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1215 	log->l_ailp->ail_head_lsn = be64_to_cpu(rhead->h_lsn);
1216 }
1217 
1218 /*
1219  * Find the sync block number or the tail of the log.
1220  *
1221  * This will be the block number of the last record to have its
1222  * associated buffers synced to disk.  Every log record header has
1223  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
1224  * to get a sync block number.  The only concern is to figure out which
1225  * log record header to believe.
1226  *
1227  * The following algorithm uses the log record header with the largest
1228  * lsn.  The entire log record does not need to be valid.  We only care
1229  * that the header is valid.
1230  *
1231  * We could speed up search by using current head_blk buffer, but it is not
1232  * available.
1233  */
1234 STATIC int
1235 xlog_find_tail(
1236 	struct xlog		*log,
1237 	xfs_daddr_t		*head_blk,
1238 	xfs_daddr_t		*tail_blk)
1239 {
1240 	xlog_rec_header_t	*rhead;
1241 	char			*offset = NULL;
1242 	char			*buffer;
1243 	int			error;
1244 	xfs_daddr_t		rhead_blk;
1245 	xfs_lsn_t		tail_lsn;
1246 	bool			wrapped = false;
1247 	bool			clean = false;
1248 
1249 	/*
1250 	 * Find previous log record
1251 	 */
1252 	if ((error = xlog_find_head(log, head_blk)))
1253 		return error;
1254 	ASSERT(*head_blk < INT_MAX);
1255 
1256 	buffer = xlog_alloc_buffer(log, 1);
1257 	if (!buffer)
1258 		return -ENOMEM;
1259 	if (*head_blk == 0) {				/* special case */
1260 		error = xlog_bread(log, 0, 1, buffer, &offset);
1261 		if (error)
1262 			goto done;
1263 
1264 		if (xlog_get_cycle(offset) == 0) {
1265 			*tail_blk = 0;
1266 			/* leave all other log inited values alone */
1267 			goto done;
1268 		}
1269 	}
1270 
1271 	/*
1272 	 * Search backwards through the log looking for the log record header
1273 	 * block. This wraps all the way back around to the head so something is
1274 	 * seriously wrong if we can't find it.
1275 	 */
1276 	error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, buffer,
1277 				      &rhead_blk, &rhead, &wrapped);
1278 	if (error < 0)
1279 		goto done;
1280 	if (!error) {
1281 		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1282 		error = -EFSCORRUPTED;
1283 		goto done;
1284 	}
1285 	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1286 
1287 	/*
1288 	 * Set the log state based on the current head record.
1289 	 */
1290 	xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1291 	tail_lsn = atomic64_read(&log->l_tail_lsn);
1292 
1293 	/*
1294 	 * Look for an unmount record at the head of the log. This sets the log
1295 	 * state to determine whether recovery is necessary.
1296 	 */
1297 	error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1298 				       rhead_blk, buffer, &clean);
1299 	if (error)
1300 		goto done;
1301 
1302 	/*
1303 	 * Verify the log head if the log is not clean (e.g., we have anything
1304 	 * but an unmount record at the head). This uses CRC verification to
1305 	 * detect and trim torn writes. If discovered, CRC failures are
1306 	 * considered torn writes and the log head is trimmed accordingly.
1307 	 *
1308 	 * Note that we can only run CRC verification when the log is dirty
1309 	 * because there's no guarantee that the log data behind an unmount
1310 	 * record is compatible with the current architecture.
1311 	 */
1312 	if (!clean) {
1313 		xfs_daddr_t	orig_head = *head_blk;
1314 
1315 		error = xlog_verify_head(log, head_blk, tail_blk, buffer,
1316 					 &rhead_blk, &rhead, &wrapped);
1317 		if (error)
1318 			goto done;
1319 
1320 		/* update in-core state again if the head changed */
1321 		if (*head_blk != orig_head) {
1322 			xlog_set_state(log, *head_blk, rhead, rhead_blk,
1323 				       wrapped);
1324 			tail_lsn = atomic64_read(&log->l_tail_lsn);
1325 			error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1326 						       rhead, rhead_blk, buffer,
1327 						       &clean);
1328 			if (error)
1329 				goto done;
1330 		}
1331 	}
1332 
1333 	/*
1334 	 * Note that the unmount was clean. If the unmount was not clean, we
1335 	 * need to know this to rebuild the superblock counters from the perag
1336 	 * headers if we have a filesystem using non-persistent counters.
1337 	 */
1338 	if (clean)
1339 		xfs_set_clean(log->l_mp);
1340 
1341 	/*
1342 	 * Make sure that there are no blocks in front of the head
1343 	 * with the same cycle number as the head.  This can happen
1344 	 * because we allow multiple outstanding log writes concurrently,
1345 	 * and the later writes might make it out before earlier ones.
1346 	 *
1347 	 * We use the lsn from before modifying it so that we'll never
1348 	 * overwrite the unmount record after a clean unmount.
1349 	 *
1350 	 * Do this only if we are going to recover the filesystem
1351 	 *
1352 	 * NOTE: This used to say "if (!readonly)"
1353 	 * However on Linux, we can & do recover a read-only filesystem.
1354 	 * We only skip recovery if NORECOVERY is specified on mount,
1355 	 * in which case we would not be here.
1356 	 *
1357 	 * But... if the -device- itself is readonly, just skip this.
1358 	 * We can't recover this device anyway, so it won't matter.
1359 	 */
1360 	if (!xfs_readonly_buftarg(log->l_targ))
1361 		error = xlog_clear_stale_blocks(log, tail_lsn);
1362 
1363 done:
1364 	kvfree(buffer);
1365 
1366 	if (error)
1367 		xfs_warn(log->l_mp, "failed to locate log tail");
1368 	return error;
1369 }
1370 
1371 /*
1372  * Is the log zeroed at all?
1373  *
1374  * The last binary search should be changed to perform an X block read
1375  * once X becomes small enough.  You can then search linearly through
1376  * the X blocks.  This will cut down on the number of reads we need to do.
1377  *
1378  * If the log is partially zeroed, this routine will pass back the blkno
1379  * of the first block with cycle number 0.  It won't have a complete LR
1380  * preceding it.
1381  *
1382  * Return:
1383  *	0  => the log is completely written to
1384  *	1 => use *blk_no as the first block of the log
1385  *	<0 => error has occurred
1386  */
1387 STATIC int
1388 xlog_find_zeroed(
1389 	struct xlog	*log,
1390 	xfs_daddr_t	*blk_no)
1391 {
1392 	char		*buffer;
1393 	char		*offset;
1394 	uint	        first_cycle, last_cycle;
1395 	xfs_daddr_t	new_blk, last_blk, start_blk;
1396 	xfs_daddr_t     num_scan_bblks;
1397 	int	        error, log_bbnum = log->l_logBBsize;
1398 	int		ret = 1;
1399 
1400 	*blk_no = 0;
1401 
1402 	/* check totally zeroed log */
1403 	buffer = xlog_alloc_buffer(log, 1);
1404 	if (!buffer)
1405 		return -ENOMEM;
1406 	error = xlog_bread(log, 0, 1, buffer, &offset);
1407 	if (error)
1408 		goto out_free_buffer;
1409 
1410 	first_cycle = xlog_get_cycle(offset);
1411 	if (first_cycle == 0) {		/* completely zeroed log */
1412 		*blk_no = 0;
1413 		goto out_free_buffer;
1414 	}
1415 
1416 	/* check partially zeroed log */
1417 	error = xlog_bread(log, log_bbnum-1, 1, buffer, &offset);
1418 	if (error)
1419 		goto out_free_buffer;
1420 
1421 	last_cycle = xlog_get_cycle(offset);
1422 	if (last_cycle != 0) {		/* log completely written to */
1423 		ret = 0;
1424 		goto out_free_buffer;
1425 	}
1426 
1427 	/* we have a partially zeroed log */
1428 	last_blk = log_bbnum-1;
1429 	error = xlog_find_cycle_start(log, buffer, 0, &last_blk, 0);
1430 	if (error)
1431 		goto out_free_buffer;
1432 
1433 	/*
1434 	 * Validate the answer.  Because there is no way to guarantee that
1435 	 * the entire log is made up of log records which are the same size,
1436 	 * we scan over the defined maximum blocks.  At this point, the maximum
1437 	 * is not chosen to mean anything special.   XXXmiken
1438 	 */
1439 	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1440 	ASSERT(num_scan_bblks <= INT_MAX);
1441 
1442 	if (last_blk < num_scan_bblks)
1443 		num_scan_bblks = last_blk;
1444 	start_blk = last_blk - num_scan_bblks;
1445 
1446 	/*
1447 	 * We search for any instances of cycle number 0 that occur before
1448 	 * our current estimate of the head.  What we're trying to detect is
1449 	 *        1 ... | 0 | 1 | 0...
1450 	 *                       ^ binary search ends here
1451 	 */
1452 	if ((error = xlog_find_verify_cycle(log, start_blk,
1453 					 (int)num_scan_bblks, 0, &new_blk)))
1454 		goto out_free_buffer;
1455 	if (new_blk != -1)
1456 		last_blk = new_blk;
1457 
1458 	/*
1459 	 * Potentially backup over partial log record write.  We don't need
1460 	 * to search the end of the log because we know it is zero.
1461 	 */
1462 	error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1463 	if (error == 1)
1464 		error = -EIO;
1465 	if (error)
1466 		goto out_free_buffer;
1467 
1468 	*blk_no = last_blk;
1469 out_free_buffer:
1470 	kvfree(buffer);
1471 	if (error)
1472 		return error;
1473 	return ret;
1474 }
1475 
1476 /*
1477  * These are simple subroutines used by xlog_clear_stale_blocks() below
1478  * to initialize a buffer full of empty log record headers and write
1479  * them into the log.
1480  */
1481 STATIC void
1482 xlog_add_record(
1483 	struct xlog		*log,
1484 	char			*buf,
1485 	int			cycle,
1486 	int			block,
1487 	int			tail_cycle,
1488 	int			tail_block)
1489 {
1490 	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1491 
1492 	memset(buf, 0, BBSIZE);
1493 	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1494 	recp->h_cycle = cpu_to_be32(cycle);
1495 	recp->h_version = cpu_to_be32(
1496 			xfs_has_logv2(log->l_mp) ? 2 : 1);
1497 	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1498 	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1499 	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1500 	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1501 }
1502 
1503 STATIC int
1504 xlog_write_log_records(
1505 	struct xlog	*log,
1506 	int		cycle,
1507 	int		start_block,
1508 	int		blocks,
1509 	int		tail_cycle,
1510 	int		tail_block)
1511 {
1512 	char		*offset;
1513 	char		*buffer;
1514 	int		balign, ealign;
1515 	int		sectbb = log->l_sectBBsize;
1516 	int		end_block = start_block + blocks;
1517 	int		bufblks;
1518 	int		error = 0;
1519 	int		i, j = 0;
1520 
1521 	/*
1522 	 * Greedily allocate a buffer big enough to handle the full
1523 	 * range of basic blocks to be written.  If that fails, try
1524 	 * a smaller size.  We need to be able to write at least a
1525 	 * log sector, or we're out of luck.
1526 	 */
1527 	bufblks = roundup_pow_of_two(blocks);
1528 	while (bufblks > log->l_logBBsize)
1529 		bufblks >>= 1;
1530 	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
1531 		bufblks >>= 1;
1532 		if (bufblks < sectbb)
1533 			return -ENOMEM;
1534 	}
1535 
1536 	/* We may need to do a read at the start to fill in part of
1537 	 * the buffer in the starting sector not covered by the first
1538 	 * write below.
1539 	 */
1540 	balign = round_down(start_block, sectbb);
1541 	if (balign != start_block) {
1542 		error = xlog_bread_noalign(log, start_block, 1, buffer);
1543 		if (error)
1544 			goto out_free_buffer;
1545 
1546 		j = start_block - balign;
1547 	}
1548 
1549 	for (i = start_block; i < end_block; i += bufblks) {
1550 		int		bcount, endcount;
1551 
1552 		bcount = min(bufblks, end_block - start_block);
1553 		endcount = bcount - j;
1554 
1555 		/* We may need to do a read at the end to fill in part of
1556 		 * the buffer in the final sector not covered by the write.
1557 		 * If this is the same sector as the above read, skip it.
1558 		 */
1559 		ealign = round_down(end_block, sectbb);
1560 		if (j == 0 && (start_block + endcount > ealign)) {
1561 			error = xlog_bread_noalign(log, ealign, sectbb,
1562 					buffer + BBTOB(ealign - start_block));
1563 			if (error)
1564 				break;
1565 
1566 		}
1567 
1568 		offset = buffer + xlog_align(log, start_block);
1569 		for (; j < endcount; j++) {
1570 			xlog_add_record(log, offset, cycle, i+j,
1571 					tail_cycle, tail_block);
1572 			offset += BBSIZE;
1573 		}
1574 		error = xlog_bwrite(log, start_block, endcount, buffer);
1575 		if (error)
1576 			break;
1577 		start_block += endcount;
1578 		j = 0;
1579 	}
1580 
1581 out_free_buffer:
1582 	kvfree(buffer);
1583 	return error;
1584 }
1585 
1586 /*
1587  * This routine is called to blow away any incomplete log writes out
1588  * in front of the log head.  We do this so that we won't become confused
1589  * if we come up, write only a little bit more, and then crash again.
1590  * If we leave the partial log records out there, this situation could
1591  * cause us to think those partial writes are valid blocks since they
1592  * have the current cycle number.  We get rid of them by overwriting them
1593  * with empty log records with the old cycle number rather than the
1594  * current one.
1595  *
1596  * The tail lsn is passed in rather than taken from
1597  * the log so that we will not write over the unmount record after a
1598  * clean unmount in a 512 block log.  Doing so would leave the log without
1599  * any valid log records in it until a new one was written.  If we crashed
1600  * during that time we would not be able to recover.
1601  */
1602 STATIC int
1603 xlog_clear_stale_blocks(
1604 	struct xlog	*log,
1605 	xfs_lsn_t	tail_lsn)
1606 {
1607 	int		tail_cycle, head_cycle;
1608 	int		tail_block, head_block;
1609 	int		tail_distance, max_distance;
1610 	int		distance;
1611 	int		error;
1612 
1613 	tail_cycle = CYCLE_LSN(tail_lsn);
1614 	tail_block = BLOCK_LSN(tail_lsn);
1615 	head_cycle = log->l_curr_cycle;
1616 	head_block = log->l_curr_block;
1617 
1618 	/*
1619 	 * Figure out the distance between the new head of the log
1620 	 * and the tail.  We want to write over any blocks beyond the
1621 	 * head that we may have written just before the crash, but
1622 	 * we don't want to overwrite the tail of the log.
1623 	 */
1624 	if (head_cycle == tail_cycle) {
1625 		/*
1626 		 * The tail is behind the head in the physical log,
1627 		 * so the distance from the head to the tail is the
1628 		 * distance from the head to the end of the log plus
1629 		 * the distance from the beginning of the log to the
1630 		 * tail.
1631 		 */
1632 		if (XFS_IS_CORRUPT(log->l_mp,
1633 				   head_block < tail_block ||
1634 				   head_block >= log->l_logBBsize))
1635 			return -EFSCORRUPTED;
1636 		tail_distance = tail_block + (log->l_logBBsize - head_block);
1637 	} else {
1638 		/*
1639 		 * The head is behind the tail in the physical log,
1640 		 * so the distance from the head to the tail is just
1641 		 * the tail block minus the head block.
1642 		 */
1643 		if (XFS_IS_CORRUPT(log->l_mp,
1644 				   head_block >= tail_block ||
1645 				   head_cycle != tail_cycle + 1))
1646 			return -EFSCORRUPTED;
1647 		tail_distance = tail_block - head_block;
1648 	}
1649 
1650 	/*
1651 	 * If the head is right up against the tail, we can't clear
1652 	 * anything.
1653 	 */
1654 	if (tail_distance <= 0) {
1655 		ASSERT(tail_distance == 0);
1656 		return 0;
1657 	}
1658 
1659 	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1660 	/*
1661 	 * Take the smaller of the maximum amount of outstanding I/O
1662 	 * we could have and the distance to the tail to clear out.
1663 	 * We take the smaller so that we don't overwrite the tail and
1664 	 * we don't waste all day writing from the head to the tail
1665 	 * for no reason.
1666 	 */
1667 	max_distance = min(max_distance, tail_distance);
1668 
1669 	if ((head_block + max_distance) <= log->l_logBBsize) {
1670 		/*
1671 		 * We can stomp all the blocks we need to without
1672 		 * wrapping around the end of the log.  Just do it
1673 		 * in a single write.  Use the cycle number of the
1674 		 * current cycle minus one so that the log will look like:
1675 		 *     n ... | n - 1 ...
1676 		 */
1677 		error = xlog_write_log_records(log, (head_cycle - 1),
1678 				head_block, max_distance, tail_cycle,
1679 				tail_block);
1680 		if (error)
1681 			return error;
1682 	} else {
1683 		/*
1684 		 * We need to wrap around the end of the physical log in
1685 		 * order to clear all the blocks.  Do it in two separate
1686 		 * I/Os.  The first write should be from the head to the
1687 		 * end of the physical log, and it should use the current
1688 		 * cycle number minus one just like above.
1689 		 */
1690 		distance = log->l_logBBsize - head_block;
1691 		error = xlog_write_log_records(log, (head_cycle - 1),
1692 				head_block, distance, tail_cycle,
1693 				tail_block);
1694 
1695 		if (error)
1696 			return error;
1697 
1698 		/*
1699 		 * Now write the blocks at the start of the physical log.
1700 		 * This writes the remainder of the blocks we want to clear.
1701 		 * It uses the current cycle number since we're now on the
1702 		 * same cycle as the head so that we get:
1703 		 *    n ... n ... | n - 1 ...
1704 		 *    ^^^^^ blocks we're writing
1705 		 */
1706 		distance = max_distance - (log->l_logBBsize - head_block);
1707 		error = xlog_write_log_records(log, head_cycle, 0, distance,
1708 				tail_cycle, tail_block);
1709 		if (error)
1710 			return error;
1711 	}
1712 
1713 	return 0;
1714 }
1715 
1716 /*
1717  * Release the recovered intent item in the AIL that matches the given intent
1718  * type and intent id.
1719  */
1720 void
1721 xlog_recover_release_intent(
1722 	struct xlog			*log,
1723 	unsigned short			intent_type,
1724 	uint64_t			intent_id)
1725 {
1726 	struct xfs_defer_pending	*dfp, *n;
1727 
1728 	list_for_each_entry_safe(dfp, n, &log->r_dfops, dfp_list) {
1729 		struct xfs_log_item	*lip = dfp->dfp_intent;
1730 
1731 		if (lip->li_type != intent_type)
1732 			continue;
1733 		if (!lip->li_ops->iop_match(lip, intent_id))
1734 			continue;
1735 
1736 		ASSERT(xlog_item_is_intent(lip));
1737 
1738 		xfs_defer_cancel_recovery(log->l_mp, dfp);
1739 	}
1740 }
1741 
1742 int
1743 xlog_recover_iget(
1744 	struct xfs_mount	*mp,
1745 	xfs_ino_t		ino,
1746 	struct xfs_inode	**ipp)
1747 {
1748 	int			error;
1749 
1750 	error = xfs_iget(mp, NULL, ino, 0, 0, ipp);
1751 	if (error)
1752 		return error;
1753 
1754 	error = xfs_qm_dqattach(*ipp);
1755 	if (error) {
1756 		xfs_irele(*ipp);
1757 		return error;
1758 	}
1759 
1760 	if (VFS_I(*ipp)->i_nlink == 0)
1761 		xfs_iflags_set(*ipp, XFS_IRECOVERY);
1762 
1763 	return 0;
1764 }
1765 
1766 /*
1767  * Get an inode so that we can recover a log operation.
1768  *
1769  * Log intent items that target inodes effectively contain a file handle.
1770  * Check that the generation number matches the intent item like we do for
1771  * other file handles.  Log intent items defined after this validation weakness
1772  * was identified must use this function.
1773  */
1774 int
1775 xlog_recover_iget_handle(
1776 	struct xfs_mount	*mp,
1777 	xfs_ino_t		ino,
1778 	uint32_t		gen,
1779 	struct xfs_inode	**ipp)
1780 {
1781 	struct xfs_inode	*ip;
1782 	int			error;
1783 
1784 	error = xlog_recover_iget(mp, ino, &ip);
1785 	if (error)
1786 		return error;
1787 
1788 	if (VFS_I(ip)->i_generation != gen) {
1789 		xfs_irele(ip);
1790 		return -EFSCORRUPTED;
1791 	}
1792 
1793 	*ipp = ip;
1794 	return 0;
1795 }
1796 
1797 /******************************************************************************
1798  *
1799  *		Log recover routines
1800  *
1801  ******************************************************************************
1802  */
1803 static const struct xlog_recover_item_ops *xlog_recover_item_ops[] = {
1804 	&xlog_buf_item_ops,
1805 	&xlog_inode_item_ops,
1806 	&xlog_dquot_item_ops,
1807 	&xlog_quotaoff_item_ops,
1808 	&xlog_icreate_item_ops,
1809 	&xlog_efi_item_ops,
1810 	&xlog_efd_item_ops,
1811 	&xlog_rui_item_ops,
1812 	&xlog_rud_item_ops,
1813 	&xlog_cui_item_ops,
1814 	&xlog_cud_item_ops,
1815 	&xlog_bui_item_ops,
1816 	&xlog_bud_item_ops,
1817 	&xlog_attri_item_ops,
1818 	&xlog_attrd_item_ops,
1819 	&xlog_xmi_item_ops,
1820 	&xlog_xmd_item_ops,
1821 };
1822 
1823 static const struct xlog_recover_item_ops *
1824 xlog_find_item_ops(
1825 	struct xlog_recover_item		*item)
1826 {
1827 	unsigned int				i;
1828 
1829 	for (i = 0; i < ARRAY_SIZE(xlog_recover_item_ops); i++)
1830 		if (ITEM_TYPE(item) == xlog_recover_item_ops[i]->item_type)
1831 			return xlog_recover_item_ops[i];
1832 
1833 	return NULL;
1834 }
1835 
1836 /*
1837  * Sort the log items in the transaction.
1838  *
1839  * The ordering constraints are defined by the inode allocation and unlink
1840  * behaviour. The rules are:
1841  *
1842  *	1. Every item is only logged once in a given transaction. Hence it
1843  *	   represents the last logged state of the item. Hence ordering is
1844  *	   dependent on the order in which operations need to be performed so
1845  *	   required initial conditions are always met.
1846  *
1847  *	2. Cancelled buffers are recorded in pass 1 in a separate table and
1848  *	   there's nothing to replay from them so we can simply cull them
1849  *	   from the transaction. However, we can't do that until after we've
1850  *	   replayed all the other items because they may be dependent on the
1851  *	   cancelled buffer and replaying the cancelled buffer can remove it
1852  *	   form the cancelled buffer table. Hence they have tobe done last.
1853  *
1854  *	3. Inode allocation buffers must be replayed before inode items that
1855  *	   read the buffer and replay changes into it. For filesystems using the
1856  *	   ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1857  *	   treated the same as inode allocation buffers as they create and
1858  *	   initialise the buffers directly.
1859  *
1860  *	4. Inode unlink buffers must be replayed after inode items are replayed.
1861  *	   This ensures that inodes are completely flushed to the inode buffer
1862  *	   in a "free" state before we remove the unlinked inode list pointer.
1863  *
1864  * Hence the ordering needs to be inode allocation buffers first, inode items
1865  * second, inode unlink buffers third and cancelled buffers last.
1866  *
1867  * But there's a problem with that - we can't tell an inode allocation buffer
1868  * apart from a regular buffer, so we can't separate them. We can, however,
1869  * tell an inode unlink buffer from the others, and so we can separate them out
1870  * from all the other buffers and move them to last.
1871  *
1872  * Hence, 4 lists, in order from head to tail:
1873  *	- buffer_list for all buffers except cancelled/inode unlink buffers
1874  *	- item_list for all non-buffer items
1875  *	- inode_buffer_list for inode unlink buffers
1876  *	- cancel_list for the cancelled buffers
1877  *
1878  * Note that we add objects to the tail of the lists so that first-to-last
1879  * ordering is preserved within the lists. Adding objects to the head of the
1880  * list means when we traverse from the head we walk them in last-to-first
1881  * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1882  * but for all other items there may be specific ordering that we need to
1883  * preserve.
1884  */
1885 STATIC int
1886 xlog_recover_reorder_trans(
1887 	struct xlog		*log,
1888 	struct xlog_recover	*trans,
1889 	int			pass)
1890 {
1891 	struct xlog_recover_item *item, *n;
1892 	int			error = 0;
1893 	LIST_HEAD(sort_list);
1894 	LIST_HEAD(cancel_list);
1895 	LIST_HEAD(buffer_list);
1896 	LIST_HEAD(inode_buffer_list);
1897 	LIST_HEAD(item_list);
1898 
1899 	list_splice_init(&trans->r_itemq, &sort_list);
1900 	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1901 		enum xlog_recover_reorder	fate = XLOG_REORDER_ITEM_LIST;
1902 
1903 		item->ri_ops = xlog_find_item_ops(item);
1904 		if (!item->ri_ops) {
1905 			xfs_warn(log->l_mp,
1906 				"%s: unrecognized type of log operation (%d)",
1907 				__func__, ITEM_TYPE(item));
1908 			ASSERT(0);
1909 			/*
1910 			 * return the remaining items back to the transaction
1911 			 * item list so they can be freed in caller.
1912 			 */
1913 			if (!list_empty(&sort_list))
1914 				list_splice_init(&sort_list, &trans->r_itemq);
1915 			error = -EFSCORRUPTED;
1916 			break;
1917 		}
1918 
1919 		if (item->ri_ops->reorder)
1920 			fate = item->ri_ops->reorder(item);
1921 
1922 		switch (fate) {
1923 		case XLOG_REORDER_BUFFER_LIST:
1924 			list_move_tail(&item->ri_list, &buffer_list);
1925 			break;
1926 		case XLOG_REORDER_CANCEL_LIST:
1927 			trace_xfs_log_recover_item_reorder_head(log,
1928 					trans, item, pass);
1929 			list_move(&item->ri_list, &cancel_list);
1930 			break;
1931 		case XLOG_REORDER_INODE_BUFFER_LIST:
1932 			list_move(&item->ri_list, &inode_buffer_list);
1933 			break;
1934 		case XLOG_REORDER_ITEM_LIST:
1935 			trace_xfs_log_recover_item_reorder_tail(log,
1936 							trans, item, pass);
1937 			list_move_tail(&item->ri_list, &item_list);
1938 			break;
1939 		}
1940 	}
1941 
1942 	ASSERT(list_empty(&sort_list));
1943 	if (!list_empty(&buffer_list))
1944 		list_splice(&buffer_list, &trans->r_itemq);
1945 	if (!list_empty(&item_list))
1946 		list_splice_tail(&item_list, &trans->r_itemq);
1947 	if (!list_empty(&inode_buffer_list))
1948 		list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1949 	if (!list_empty(&cancel_list))
1950 		list_splice_tail(&cancel_list, &trans->r_itemq);
1951 	return error;
1952 }
1953 
1954 void
1955 xlog_buf_readahead(
1956 	struct xlog		*log,
1957 	xfs_daddr_t		blkno,
1958 	uint			len,
1959 	const struct xfs_buf_ops *ops)
1960 {
1961 	if (!xlog_is_buffer_cancelled(log, blkno, len))
1962 		xfs_buf_readahead(log->l_mp->m_ddev_targp, blkno, len, ops);
1963 }
1964 
1965 /*
1966  * Create a deferred work structure for resuming and tracking the progress of a
1967  * log intent item that was found during recovery.
1968  */
1969 void
1970 xlog_recover_intent_item(
1971 	struct xlog			*log,
1972 	struct xfs_log_item		*lip,
1973 	xfs_lsn_t			lsn,
1974 	const struct xfs_defer_op_type	*ops)
1975 {
1976 	ASSERT(xlog_item_is_intent(lip));
1977 
1978 	xfs_defer_start_recovery(lip, &log->r_dfops, ops);
1979 
1980 	/*
1981 	 * Insert the intent into the AIL directly and drop one reference so
1982 	 * that finishing or canceling the work will drop the other.
1983 	 */
1984 	xfs_trans_ail_insert(log->l_ailp, lip, lsn);
1985 	lip->li_ops->iop_unpin(lip, 0);
1986 }
1987 
1988 STATIC int
1989 xlog_recover_items_pass2(
1990 	struct xlog                     *log,
1991 	struct xlog_recover             *trans,
1992 	struct list_head                *buffer_list,
1993 	struct list_head                *item_list)
1994 {
1995 	struct xlog_recover_item	*item;
1996 	int				error = 0;
1997 
1998 	list_for_each_entry(item, item_list, ri_list) {
1999 		trace_xfs_log_recover_item_recover(log, trans, item,
2000 				XLOG_RECOVER_PASS2);
2001 
2002 		if (item->ri_ops->commit_pass2)
2003 			error = item->ri_ops->commit_pass2(log, buffer_list,
2004 					item, trans->r_lsn);
2005 		if (error)
2006 			return error;
2007 	}
2008 
2009 	return error;
2010 }
2011 
2012 /*
2013  * Perform the transaction.
2014  *
2015  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2016  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2017  */
2018 STATIC int
2019 xlog_recover_commit_trans(
2020 	struct xlog		*log,
2021 	struct xlog_recover	*trans,
2022 	int			pass,
2023 	struct list_head	*buffer_list)
2024 {
2025 	int				error = 0;
2026 	int				items_queued = 0;
2027 	struct xlog_recover_item	*item;
2028 	struct xlog_recover_item	*next;
2029 	LIST_HEAD			(ra_list);
2030 	LIST_HEAD			(done_list);
2031 
2032 	#define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
2033 
2034 	hlist_del_init(&trans->r_list);
2035 
2036 	error = xlog_recover_reorder_trans(log, trans, pass);
2037 	if (error)
2038 		return error;
2039 
2040 	list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
2041 		trace_xfs_log_recover_item_recover(log, trans, item, pass);
2042 
2043 		switch (pass) {
2044 		case XLOG_RECOVER_PASS1:
2045 			if (item->ri_ops->commit_pass1)
2046 				error = item->ri_ops->commit_pass1(log, item);
2047 			break;
2048 		case XLOG_RECOVER_PASS2:
2049 			if (item->ri_ops->ra_pass2)
2050 				item->ri_ops->ra_pass2(log, item);
2051 			list_move_tail(&item->ri_list, &ra_list);
2052 			items_queued++;
2053 			if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
2054 				error = xlog_recover_items_pass2(log, trans,
2055 						buffer_list, &ra_list);
2056 				list_splice_tail_init(&ra_list, &done_list);
2057 				items_queued = 0;
2058 			}
2059 
2060 			break;
2061 		default:
2062 			ASSERT(0);
2063 		}
2064 
2065 		if (error)
2066 			goto out;
2067 	}
2068 
2069 out:
2070 	if (!list_empty(&ra_list)) {
2071 		if (!error)
2072 			error = xlog_recover_items_pass2(log, trans,
2073 					buffer_list, &ra_list);
2074 		list_splice_tail_init(&ra_list, &done_list);
2075 	}
2076 
2077 	if (!list_empty(&done_list))
2078 		list_splice_init(&done_list, &trans->r_itemq);
2079 
2080 	return error;
2081 }
2082 
2083 STATIC void
2084 xlog_recover_add_item(
2085 	struct list_head	*head)
2086 {
2087 	struct xlog_recover_item *item;
2088 
2089 	item = kzalloc(sizeof(struct xlog_recover_item),
2090 			GFP_KERNEL | __GFP_NOFAIL);
2091 	INIT_LIST_HEAD(&item->ri_list);
2092 	list_add_tail(&item->ri_list, head);
2093 }
2094 
2095 STATIC int
2096 xlog_recover_add_to_cont_trans(
2097 	struct xlog		*log,
2098 	struct xlog_recover	*trans,
2099 	char			*dp,
2100 	int			len)
2101 {
2102 	struct xlog_recover_item *item;
2103 	char			*ptr, *old_ptr;
2104 	int			old_len;
2105 
2106 	/*
2107 	 * If the transaction is empty, the header was split across this and the
2108 	 * previous record. Copy the rest of the header.
2109 	 */
2110 	if (list_empty(&trans->r_itemq)) {
2111 		ASSERT(len <= sizeof(struct xfs_trans_header));
2112 		if (len > sizeof(struct xfs_trans_header)) {
2113 			xfs_warn(log->l_mp, "%s: bad header length", __func__);
2114 			return -EFSCORRUPTED;
2115 		}
2116 
2117 		xlog_recover_add_item(&trans->r_itemq);
2118 		ptr = (char *)&trans->r_theader +
2119 				sizeof(struct xfs_trans_header) - len;
2120 		memcpy(ptr, dp, len);
2121 		return 0;
2122 	}
2123 
2124 	/* take the tail entry */
2125 	item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2126 			  ri_list);
2127 
2128 	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
2129 	old_len = item->ri_buf[item->ri_cnt-1].i_len;
2130 
2131 	ptr = kvrealloc(old_ptr, old_len, len + old_len, GFP_KERNEL);
2132 	if (!ptr)
2133 		return -ENOMEM;
2134 	memcpy(&ptr[old_len], dp, len);
2135 	item->ri_buf[item->ri_cnt-1].i_len += len;
2136 	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
2137 	trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
2138 	return 0;
2139 }
2140 
2141 /*
2142  * The next region to add is the start of a new region.  It could be
2143  * a whole region or it could be the first part of a new region.  Because
2144  * of this, the assumption here is that the type and size fields of all
2145  * format structures fit into the first 32 bits of the structure.
2146  *
2147  * This works because all regions must be 32 bit aligned.  Therefore, we
2148  * either have both fields or we have neither field.  In the case we have
2149  * neither field, the data part of the region is zero length.  We only have
2150  * a log_op_header and can throw away the header since a new one will appear
2151  * later.  If we have at least 4 bytes, then we can determine how many regions
2152  * will appear in the current log item.
2153  */
2154 STATIC int
2155 xlog_recover_add_to_trans(
2156 	struct xlog		*log,
2157 	struct xlog_recover	*trans,
2158 	char			*dp,
2159 	int			len)
2160 {
2161 	struct xfs_inode_log_format	*in_f;			/* any will do */
2162 	struct xlog_recover_item *item;
2163 	char			*ptr;
2164 
2165 	if (!len)
2166 		return 0;
2167 	if (list_empty(&trans->r_itemq)) {
2168 		/* we need to catch log corruptions here */
2169 		if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
2170 			xfs_warn(log->l_mp, "%s: bad header magic number",
2171 				__func__);
2172 			ASSERT(0);
2173 			return -EFSCORRUPTED;
2174 		}
2175 
2176 		if (len > sizeof(struct xfs_trans_header)) {
2177 			xfs_warn(log->l_mp, "%s: bad header length", __func__);
2178 			ASSERT(0);
2179 			return -EFSCORRUPTED;
2180 		}
2181 
2182 		/*
2183 		 * The transaction header can be arbitrarily split across op
2184 		 * records. If we don't have the whole thing here, copy what we
2185 		 * do have and handle the rest in the next record.
2186 		 */
2187 		if (len == sizeof(struct xfs_trans_header))
2188 			xlog_recover_add_item(&trans->r_itemq);
2189 		memcpy(&trans->r_theader, dp, len);
2190 		return 0;
2191 	}
2192 
2193 	ptr = xlog_kvmalloc(len);
2194 	memcpy(ptr, dp, len);
2195 	in_f = (struct xfs_inode_log_format *)ptr;
2196 
2197 	/* take the tail entry */
2198 	item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2199 			  ri_list);
2200 	if (item->ri_total != 0 &&
2201 	     item->ri_total == item->ri_cnt) {
2202 		/* tail item is in use, get a new one */
2203 		xlog_recover_add_item(&trans->r_itemq);
2204 		item = list_entry(trans->r_itemq.prev,
2205 					struct xlog_recover_item, ri_list);
2206 	}
2207 
2208 	if (item->ri_total == 0) {		/* first region to be added */
2209 		if (in_f->ilf_size == 0 ||
2210 		    in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
2211 			xfs_warn(log->l_mp,
2212 		"bad number of regions (%d) in inode log format",
2213 				  in_f->ilf_size);
2214 			ASSERT(0);
2215 			kvfree(ptr);
2216 			return -EFSCORRUPTED;
2217 		}
2218 
2219 		item->ri_total = in_f->ilf_size;
2220 		item->ri_buf = kzalloc(item->ri_total * sizeof(xfs_log_iovec_t),
2221 				GFP_KERNEL | __GFP_NOFAIL);
2222 	}
2223 
2224 	if (item->ri_total <= item->ri_cnt) {
2225 		xfs_warn(log->l_mp,
2226 	"log item region count (%d) overflowed size (%d)",
2227 				item->ri_cnt, item->ri_total);
2228 		ASSERT(0);
2229 		kvfree(ptr);
2230 		return -EFSCORRUPTED;
2231 	}
2232 
2233 	/* Description region is ri_buf[0] */
2234 	item->ri_buf[item->ri_cnt].i_addr = ptr;
2235 	item->ri_buf[item->ri_cnt].i_len  = len;
2236 	item->ri_cnt++;
2237 	trace_xfs_log_recover_item_add(log, trans, item, 0);
2238 	return 0;
2239 }
2240 
2241 /*
2242  * Free up any resources allocated by the transaction
2243  *
2244  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2245  */
2246 STATIC void
2247 xlog_recover_free_trans(
2248 	struct xlog_recover	*trans)
2249 {
2250 	struct xlog_recover_item *item, *n;
2251 	int			i;
2252 
2253 	hlist_del_init(&trans->r_list);
2254 
2255 	list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2256 		/* Free the regions in the item. */
2257 		list_del(&item->ri_list);
2258 		for (i = 0; i < item->ri_cnt; i++)
2259 			kvfree(item->ri_buf[i].i_addr);
2260 		/* Free the item itself */
2261 		kfree(item->ri_buf);
2262 		kfree(item);
2263 	}
2264 	/* Free the transaction recover structure */
2265 	kfree(trans);
2266 }
2267 
2268 /*
2269  * On error or completion, trans is freed.
2270  */
2271 STATIC int
2272 xlog_recovery_process_trans(
2273 	struct xlog		*log,
2274 	struct xlog_recover	*trans,
2275 	char			*dp,
2276 	unsigned int		len,
2277 	unsigned int		flags,
2278 	int			pass,
2279 	struct list_head	*buffer_list)
2280 {
2281 	int			error = 0;
2282 	bool			freeit = false;
2283 
2284 	/* mask off ophdr transaction container flags */
2285 	flags &= ~XLOG_END_TRANS;
2286 	if (flags & XLOG_WAS_CONT_TRANS)
2287 		flags &= ~XLOG_CONTINUE_TRANS;
2288 
2289 	/*
2290 	 * Callees must not free the trans structure. We'll decide if we need to
2291 	 * free it or not based on the operation being done and it's result.
2292 	 */
2293 	switch (flags) {
2294 	/* expected flag values */
2295 	case 0:
2296 	case XLOG_CONTINUE_TRANS:
2297 		error = xlog_recover_add_to_trans(log, trans, dp, len);
2298 		break;
2299 	case XLOG_WAS_CONT_TRANS:
2300 		error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
2301 		break;
2302 	case XLOG_COMMIT_TRANS:
2303 		error = xlog_recover_commit_trans(log, trans, pass,
2304 						  buffer_list);
2305 		/* success or fail, we are now done with this transaction. */
2306 		freeit = true;
2307 		break;
2308 
2309 	/* unexpected flag values */
2310 	case XLOG_UNMOUNT_TRANS:
2311 		/* just skip trans */
2312 		xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
2313 		freeit = true;
2314 		break;
2315 	case XLOG_START_TRANS:
2316 	default:
2317 		xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
2318 		ASSERT(0);
2319 		error = -EFSCORRUPTED;
2320 		break;
2321 	}
2322 	if (error || freeit)
2323 		xlog_recover_free_trans(trans);
2324 	return error;
2325 }
2326 
2327 /*
2328  * Lookup the transaction recovery structure associated with the ID in the
2329  * current ophdr. If the transaction doesn't exist and the start flag is set in
2330  * the ophdr, then allocate a new transaction for future ID matches to find.
2331  * Either way, return what we found during the lookup - an existing transaction
2332  * or nothing.
2333  */
2334 STATIC struct xlog_recover *
2335 xlog_recover_ophdr_to_trans(
2336 	struct hlist_head	rhash[],
2337 	struct xlog_rec_header	*rhead,
2338 	struct xlog_op_header	*ohead)
2339 {
2340 	struct xlog_recover	*trans;
2341 	xlog_tid_t		tid;
2342 	struct hlist_head	*rhp;
2343 
2344 	tid = be32_to_cpu(ohead->oh_tid);
2345 	rhp = &rhash[XLOG_RHASH(tid)];
2346 	hlist_for_each_entry(trans, rhp, r_list) {
2347 		if (trans->r_log_tid == tid)
2348 			return trans;
2349 	}
2350 
2351 	/*
2352 	 * skip over non-start transaction headers - we could be
2353 	 * processing slack space before the next transaction starts
2354 	 */
2355 	if (!(ohead->oh_flags & XLOG_START_TRANS))
2356 		return NULL;
2357 
2358 	ASSERT(be32_to_cpu(ohead->oh_len) == 0);
2359 
2360 	/*
2361 	 * This is a new transaction so allocate a new recovery container to
2362 	 * hold the recovery ops that will follow.
2363 	 */
2364 	trans = kzalloc(sizeof(struct xlog_recover), GFP_KERNEL | __GFP_NOFAIL);
2365 	trans->r_log_tid = tid;
2366 	trans->r_lsn = be64_to_cpu(rhead->h_lsn);
2367 	INIT_LIST_HEAD(&trans->r_itemq);
2368 	INIT_HLIST_NODE(&trans->r_list);
2369 	hlist_add_head(&trans->r_list, rhp);
2370 
2371 	/*
2372 	 * Nothing more to do for this ophdr. Items to be added to this new
2373 	 * transaction will be in subsequent ophdr containers.
2374 	 */
2375 	return NULL;
2376 }
2377 
2378 STATIC int
2379 xlog_recover_process_ophdr(
2380 	struct xlog		*log,
2381 	struct hlist_head	rhash[],
2382 	struct xlog_rec_header	*rhead,
2383 	struct xlog_op_header	*ohead,
2384 	char			*dp,
2385 	char			*end,
2386 	int			pass,
2387 	struct list_head	*buffer_list)
2388 {
2389 	struct xlog_recover	*trans;
2390 	unsigned int		len;
2391 	int			error;
2392 
2393 	/* Do we understand who wrote this op? */
2394 	if (ohead->oh_clientid != XFS_TRANSACTION &&
2395 	    ohead->oh_clientid != XFS_LOG) {
2396 		xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
2397 			__func__, ohead->oh_clientid);
2398 		ASSERT(0);
2399 		return -EFSCORRUPTED;
2400 	}
2401 
2402 	/*
2403 	 * Check the ophdr contains all the data it is supposed to contain.
2404 	 */
2405 	len = be32_to_cpu(ohead->oh_len);
2406 	if (dp + len > end) {
2407 		xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
2408 		WARN_ON(1);
2409 		return -EFSCORRUPTED;
2410 	}
2411 
2412 	trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
2413 	if (!trans) {
2414 		/* nothing to do, so skip over this ophdr */
2415 		return 0;
2416 	}
2417 
2418 	/*
2419 	 * The recovered buffer queue is drained only once we know that all
2420 	 * recovery items for the current LSN have been processed. This is
2421 	 * required because:
2422 	 *
2423 	 * - Buffer write submission updates the metadata LSN of the buffer.
2424 	 * - Log recovery skips items with a metadata LSN >= the current LSN of
2425 	 *   the recovery item.
2426 	 * - Separate recovery items against the same metadata buffer can share
2427 	 *   a current LSN. I.e., consider that the LSN of a recovery item is
2428 	 *   defined as the starting LSN of the first record in which its
2429 	 *   transaction appears, that a record can hold multiple transactions,
2430 	 *   and/or that a transaction can span multiple records.
2431 	 *
2432 	 * In other words, we are allowed to submit a buffer from log recovery
2433 	 * once per current LSN. Otherwise, we may incorrectly skip recovery
2434 	 * items and cause corruption.
2435 	 *
2436 	 * We don't know up front whether buffers are updated multiple times per
2437 	 * LSN. Therefore, track the current LSN of each commit log record as it
2438 	 * is processed and drain the queue when it changes. Use commit records
2439 	 * because they are ordered correctly by the logging code.
2440 	 */
2441 	if (log->l_recovery_lsn != trans->r_lsn &&
2442 	    ohead->oh_flags & XLOG_COMMIT_TRANS) {
2443 		error = xfs_buf_delwri_submit(buffer_list);
2444 		if (error)
2445 			return error;
2446 		log->l_recovery_lsn = trans->r_lsn;
2447 	}
2448 
2449 	return xlog_recovery_process_trans(log, trans, dp, len,
2450 					   ohead->oh_flags, pass, buffer_list);
2451 }
2452 
2453 /*
2454  * There are two valid states of the r_state field.  0 indicates that the
2455  * transaction structure is in a normal state.  We have either seen the
2456  * start of the transaction or the last operation we added was not a partial
2457  * operation.  If the last operation we added to the transaction was a
2458  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2459  *
2460  * NOTE: skip LRs with 0 data length.
2461  */
2462 STATIC int
2463 xlog_recover_process_data(
2464 	struct xlog		*log,
2465 	struct hlist_head	rhash[],
2466 	struct xlog_rec_header	*rhead,
2467 	char			*dp,
2468 	int			pass,
2469 	struct list_head	*buffer_list)
2470 {
2471 	struct xlog_op_header	*ohead;
2472 	char			*end;
2473 	int			num_logops;
2474 	int			error;
2475 
2476 	end = dp + be32_to_cpu(rhead->h_len);
2477 	num_logops = be32_to_cpu(rhead->h_num_logops);
2478 
2479 	/* check the log format matches our own - else we can't recover */
2480 	if (xlog_header_check_recover(log->l_mp, rhead))
2481 		return -EIO;
2482 
2483 	trace_xfs_log_recover_record(log, rhead, pass);
2484 	while ((dp < end) && num_logops) {
2485 
2486 		ohead = (struct xlog_op_header *)dp;
2487 		dp += sizeof(*ohead);
2488 		if (dp > end) {
2489 			xfs_warn(log->l_mp, "%s: op header overrun", __func__);
2490 			return -EFSCORRUPTED;
2491 		}
2492 
2493 		/* errors will abort recovery */
2494 		error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
2495 						   dp, end, pass, buffer_list);
2496 		if (error)
2497 			return error;
2498 
2499 		dp += be32_to_cpu(ohead->oh_len);
2500 		num_logops--;
2501 	}
2502 	return 0;
2503 }
2504 
2505 /* Take all the collected deferred ops and finish them in order. */
2506 static int
2507 xlog_finish_defer_ops(
2508 	struct xfs_mount	*mp,
2509 	struct list_head	*capture_list)
2510 {
2511 	struct xfs_defer_capture *dfc, *next;
2512 	struct xfs_trans	*tp;
2513 	int			error = 0;
2514 
2515 	list_for_each_entry_safe(dfc, next, capture_list, dfc_list) {
2516 		struct xfs_trans_res	resv;
2517 		struct xfs_defer_resources dres;
2518 
2519 		/*
2520 		 * Create a new transaction reservation from the captured
2521 		 * information.  Set logcount to 1 to force the new transaction
2522 		 * to regrant every roll so that we can make forward progress
2523 		 * in recovery no matter how full the log might be.
2524 		 */
2525 		resv.tr_logres = dfc->dfc_logres;
2526 		resv.tr_logcount = 1;
2527 		resv.tr_logflags = XFS_TRANS_PERM_LOG_RES;
2528 
2529 		error = xfs_trans_alloc(mp, &resv, dfc->dfc_blkres,
2530 				dfc->dfc_rtxres, XFS_TRANS_RESERVE, &tp);
2531 		if (error) {
2532 			xlog_force_shutdown(mp->m_log, SHUTDOWN_LOG_IO_ERROR);
2533 			return error;
2534 		}
2535 
2536 		/*
2537 		 * Transfer to this new transaction all the dfops we captured
2538 		 * from recovering a single intent item.
2539 		 */
2540 		list_del_init(&dfc->dfc_list);
2541 		xfs_defer_ops_continue(dfc, tp, &dres);
2542 		error = xfs_trans_commit(tp);
2543 		xfs_defer_resources_rele(&dres);
2544 		if (error)
2545 			return error;
2546 	}
2547 
2548 	ASSERT(list_empty(capture_list));
2549 	return 0;
2550 }
2551 
2552 /* Release all the captured defer ops and capture structures in this list. */
2553 static void
2554 xlog_abort_defer_ops(
2555 	struct xfs_mount		*mp,
2556 	struct list_head		*capture_list)
2557 {
2558 	struct xfs_defer_capture	*dfc;
2559 	struct xfs_defer_capture	*next;
2560 
2561 	list_for_each_entry_safe(dfc, next, capture_list, dfc_list) {
2562 		list_del_init(&dfc->dfc_list);
2563 		xfs_defer_ops_capture_abort(mp, dfc);
2564 	}
2565 }
2566 
2567 /*
2568  * When this is called, all of the log intent items which did not have
2569  * corresponding log done items should be in the AIL.  What we do now is update
2570  * the data structures associated with each one.
2571  *
2572  * Since we process the log intent items in normal transactions, they will be
2573  * removed at some point after the commit.  This prevents us from just walking
2574  * down the list processing each one.  We'll use a flag in the intent item to
2575  * skip those that we've already processed and use the AIL iteration mechanism's
2576  * generation count to try to speed this up at least a bit.
2577  *
2578  * When we start, we know that the intents are the only things in the AIL. As we
2579  * process them, however, other items are added to the AIL. Hence we know we
2580  * have started recovery on all the pending intents when we find an non-intent
2581  * item in the AIL.
2582  */
2583 STATIC int
2584 xlog_recover_process_intents(
2585 	struct xlog			*log)
2586 {
2587 	LIST_HEAD(capture_list);
2588 	struct xfs_defer_pending	*dfp, *n;
2589 	int				error = 0;
2590 #if defined(DEBUG) || defined(XFS_WARN)
2591 	xfs_lsn_t			last_lsn;
2592 
2593 	last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
2594 #endif
2595 
2596 	list_for_each_entry_safe(dfp, n, &log->r_dfops, dfp_list) {
2597 		ASSERT(xlog_item_is_intent(dfp->dfp_intent));
2598 
2599 		/*
2600 		 * We should never see a redo item with a LSN higher than
2601 		 * the last transaction we found in the log at the start
2602 		 * of recovery.
2603 		 */
2604 		ASSERT(XFS_LSN_CMP(last_lsn, dfp->dfp_intent->li_lsn) >= 0);
2605 
2606 		/*
2607 		 * NOTE: If your intent processing routine can create more
2608 		 * deferred ops, you /must/ attach them to the capture list in
2609 		 * the recover routine or else those subsequent intents will be
2610 		 * replayed in the wrong order!
2611 		 *
2612 		 * The recovery function can free the log item, so we must not
2613 		 * access dfp->dfp_intent after it returns.  It must dispose of
2614 		 * @dfp if it returns 0.
2615 		 */
2616 		error = xfs_defer_finish_recovery(log->l_mp, dfp,
2617 				&capture_list);
2618 		if (error)
2619 			break;
2620 	}
2621 	if (error)
2622 		goto err;
2623 
2624 	error = xlog_finish_defer_ops(log->l_mp, &capture_list);
2625 	if (error)
2626 		goto err;
2627 
2628 	return 0;
2629 err:
2630 	xlog_abort_defer_ops(log->l_mp, &capture_list);
2631 	return error;
2632 }
2633 
2634 /*
2635  * A cancel occurs when the mount has failed and we're bailing out.  Release all
2636  * pending log intent items that we haven't started recovery on so they don't
2637  * pin the AIL.
2638  */
2639 STATIC void
2640 xlog_recover_cancel_intents(
2641 	struct xlog			*log)
2642 {
2643 	struct xfs_defer_pending	*dfp, *n;
2644 
2645 	list_for_each_entry_safe(dfp, n, &log->r_dfops, dfp_list) {
2646 		ASSERT(xlog_item_is_intent(dfp->dfp_intent));
2647 
2648 		xfs_defer_cancel_recovery(log->l_mp, dfp);
2649 	}
2650 }
2651 
2652 /*
2653  * Transfer ownership of the recovered pending work to the recovery transaction
2654  * and try to finish the work.  If there is more work to be done, the dfp will
2655  * remain attached to the transaction.  If not, the dfp is freed.
2656  */
2657 int
2658 xlog_recover_finish_intent(
2659 	struct xfs_trans		*tp,
2660 	struct xfs_defer_pending	*dfp)
2661 {
2662 	int				error;
2663 
2664 	list_move(&dfp->dfp_list, &tp->t_dfops);
2665 	error = xfs_defer_finish_one(tp, dfp);
2666 	if (error == -EAGAIN)
2667 		return 0;
2668 	return error;
2669 }
2670 
2671 /*
2672  * This routine performs a transaction to null out a bad inode pointer
2673  * in an agi unlinked inode hash bucket.
2674  */
2675 STATIC void
2676 xlog_recover_clear_agi_bucket(
2677 	struct xfs_perag	*pag,
2678 	int			bucket)
2679 {
2680 	struct xfs_mount	*mp = pag->pag_mount;
2681 	struct xfs_trans	*tp;
2682 	struct xfs_agi		*agi;
2683 	struct xfs_buf		*agibp;
2684 	int			offset;
2685 	int			error;
2686 
2687 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
2688 	if (error)
2689 		goto out_error;
2690 
2691 	error = xfs_read_agi(pag, tp, 0, &agibp);
2692 	if (error)
2693 		goto out_abort;
2694 
2695 	agi = agibp->b_addr;
2696 	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
2697 	offset = offsetof(xfs_agi_t, agi_unlinked) +
2698 		 (sizeof(xfs_agino_t) * bucket);
2699 	xfs_trans_log_buf(tp, agibp, offset,
2700 			  (offset + sizeof(xfs_agino_t) - 1));
2701 
2702 	error = xfs_trans_commit(tp);
2703 	if (error)
2704 		goto out_error;
2705 	return;
2706 
2707 out_abort:
2708 	xfs_trans_cancel(tp);
2709 out_error:
2710 	xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__,
2711 			pag->pag_agno);
2712 	return;
2713 }
2714 
2715 static int
2716 xlog_recover_iunlink_bucket(
2717 	struct xfs_perag	*pag,
2718 	struct xfs_agi		*agi,
2719 	int			bucket)
2720 {
2721 	struct xfs_mount	*mp = pag->pag_mount;
2722 	struct xfs_inode	*prev_ip = NULL;
2723 	struct xfs_inode	*ip;
2724 	xfs_agino_t		prev_agino, agino;
2725 	int			error = 0;
2726 
2727 	agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2728 	while (agino != NULLAGINO) {
2729 		error = xfs_iget(mp, NULL,
2730 				XFS_AGINO_TO_INO(mp, pag->pag_agno, agino),
2731 				0, 0, &ip);
2732 		if (error)
2733 			break;
2734 
2735 		ASSERT(VFS_I(ip)->i_nlink == 0);
2736 		ASSERT(VFS_I(ip)->i_mode != 0);
2737 		xfs_iflags_clear(ip, XFS_IRECOVERY);
2738 		agino = ip->i_next_unlinked;
2739 
2740 		if (prev_ip) {
2741 			ip->i_prev_unlinked = prev_agino;
2742 			xfs_irele(prev_ip);
2743 
2744 			/*
2745 			 * Ensure the inode is removed from the unlinked list
2746 			 * before we continue so that it won't race with
2747 			 * building the in-memory list here. This could be
2748 			 * serialised with the agibp lock, but that just
2749 			 * serialises via lockstepping and it's much simpler
2750 			 * just to flush the inodegc queue and wait for it to
2751 			 * complete.
2752 			 */
2753 			error = xfs_inodegc_flush(mp);
2754 			if (error)
2755 				break;
2756 		}
2757 
2758 		prev_agino = agino;
2759 		prev_ip = ip;
2760 	}
2761 
2762 	if (prev_ip) {
2763 		int	error2;
2764 
2765 		ip->i_prev_unlinked = prev_agino;
2766 		xfs_irele(prev_ip);
2767 
2768 		error2 = xfs_inodegc_flush(mp);
2769 		if (error2 && !error)
2770 			return error2;
2771 	}
2772 	return error;
2773 }
2774 
2775 /*
2776  * Recover AGI unlinked lists
2777  *
2778  * This is called during recovery to process any inodes which we unlinked but
2779  * not freed when the system crashed.  These inodes will be on the lists in the
2780  * AGI blocks. What we do here is scan all the AGIs and fully truncate and free
2781  * any inodes found on the lists. Each inode is removed from the lists when it
2782  * has been fully truncated and is freed. The freeing of the inode and its
2783  * removal from the list must be atomic.
2784  *
2785  * If everything we touch in the agi processing loop is already in memory, this
2786  * loop can hold the cpu for a long time. It runs without lock contention,
2787  * memory allocation contention, the need wait for IO, etc, and so will run
2788  * until we either run out of inodes to process, run low on memory or we run out
2789  * of log space.
2790  *
2791  * This behaviour is bad for latency on single CPU and non-preemptible kernels,
2792  * and can prevent other filesystem work (such as CIL pushes) from running. This
2793  * can lead to deadlocks if the recovery process runs out of log reservation
2794  * space. Hence we need to yield the CPU when there is other kernel work
2795  * scheduled on this CPU to ensure other scheduled work can run without undue
2796  * latency.
2797  */
2798 static void
2799 xlog_recover_iunlink_ag(
2800 	struct xfs_perag	*pag)
2801 {
2802 	struct xfs_agi		*agi;
2803 	struct xfs_buf		*agibp;
2804 	int			bucket;
2805 	int			error;
2806 
2807 	error = xfs_read_agi(pag, NULL, 0, &agibp);
2808 	if (error) {
2809 		/*
2810 		 * AGI is b0rked. Don't process it.
2811 		 *
2812 		 * We should probably mark the filesystem as corrupt after we've
2813 		 * recovered all the ag's we can....
2814 		 */
2815 		return;
2816 	}
2817 
2818 	/*
2819 	 * Unlock the buffer so that it can be acquired in the normal course of
2820 	 * the transaction to truncate and free each inode.  Because we are not
2821 	 * racing with anyone else here for the AGI buffer, we don't even need
2822 	 * to hold it locked to read the initial unlinked bucket entries out of
2823 	 * the buffer. We keep buffer reference though, so that it stays pinned
2824 	 * in memory while we need the buffer.
2825 	 */
2826 	agi = agibp->b_addr;
2827 	xfs_buf_unlock(agibp);
2828 
2829 	for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
2830 		error = xlog_recover_iunlink_bucket(pag, agi, bucket);
2831 		if (error) {
2832 			/*
2833 			 * Bucket is unrecoverable, so only a repair scan can
2834 			 * free the remaining unlinked inodes. Just empty the
2835 			 * bucket and remaining inodes on it unreferenced and
2836 			 * unfreeable.
2837 			 */
2838 			xlog_recover_clear_agi_bucket(pag, bucket);
2839 		}
2840 	}
2841 
2842 	xfs_buf_rele(agibp);
2843 }
2844 
2845 static void
2846 xlog_recover_process_iunlinks(
2847 	struct xlog	*log)
2848 {
2849 	struct xfs_perag	*pag;
2850 	xfs_agnumber_t		agno;
2851 
2852 	for_each_perag(log->l_mp, agno, pag)
2853 		xlog_recover_iunlink_ag(pag);
2854 }
2855 
2856 STATIC void
2857 xlog_unpack_data(
2858 	struct xlog_rec_header	*rhead,
2859 	char			*dp,
2860 	struct xlog		*log)
2861 {
2862 	int			i, j, k;
2863 
2864 	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
2865 		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
2866 		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
2867 		dp += BBSIZE;
2868 	}
2869 
2870 	if (xfs_has_logv2(log->l_mp)) {
2871 		xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
2872 		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
2873 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2874 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2875 			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
2876 			dp += BBSIZE;
2877 		}
2878 	}
2879 }
2880 
2881 /*
2882  * CRC check, unpack and process a log record.
2883  */
2884 STATIC int
2885 xlog_recover_process(
2886 	struct xlog		*log,
2887 	struct hlist_head	rhash[],
2888 	struct xlog_rec_header	*rhead,
2889 	char			*dp,
2890 	int			pass,
2891 	struct list_head	*buffer_list)
2892 {
2893 	__le32			old_crc = rhead->h_crc;
2894 	__le32			crc;
2895 
2896 	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
2897 
2898 	/*
2899 	 * Nothing else to do if this is a CRC verification pass. Just return
2900 	 * if this a record with a non-zero crc. Unfortunately, mkfs always
2901 	 * sets old_crc to 0 so we must consider this valid even on v5 supers.
2902 	 * Otherwise, return EFSBADCRC on failure so the callers up the stack
2903 	 * know precisely what failed.
2904 	 */
2905 	if (pass == XLOG_RECOVER_CRCPASS) {
2906 		if (old_crc && crc != old_crc)
2907 			return -EFSBADCRC;
2908 		return 0;
2909 	}
2910 
2911 	/*
2912 	 * We're in the normal recovery path. Issue a warning if and only if the
2913 	 * CRC in the header is non-zero. This is an advisory warning and the
2914 	 * zero CRC check prevents warnings from being emitted when upgrading
2915 	 * the kernel from one that does not add CRCs by default.
2916 	 */
2917 	if (crc != old_crc) {
2918 		if (old_crc || xfs_has_crc(log->l_mp)) {
2919 			xfs_alert(log->l_mp,
2920 		"log record CRC mismatch: found 0x%x, expected 0x%x.",
2921 					le32_to_cpu(old_crc),
2922 					le32_to_cpu(crc));
2923 			xfs_hex_dump(dp, 32);
2924 		}
2925 
2926 		/*
2927 		 * If the filesystem is CRC enabled, this mismatch becomes a
2928 		 * fatal log corruption failure.
2929 		 */
2930 		if (xfs_has_crc(log->l_mp)) {
2931 			XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, log->l_mp);
2932 			return -EFSCORRUPTED;
2933 		}
2934 	}
2935 
2936 	xlog_unpack_data(rhead, dp, log);
2937 
2938 	return xlog_recover_process_data(log, rhash, rhead, dp, pass,
2939 					 buffer_list);
2940 }
2941 
2942 STATIC int
2943 xlog_valid_rec_header(
2944 	struct xlog		*log,
2945 	struct xlog_rec_header	*rhead,
2946 	xfs_daddr_t		blkno,
2947 	int			bufsize)
2948 {
2949 	int			hlen;
2950 
2951 	if (XFS_IS_CORRUPT(log->l_mp,
2952 			   rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM)))
2953 		return -EFSCORRUPTED;
2954 	if (XFS_IS_CORRUPT(log->l_mp,
2955 			   (!rhead->h_version ||
2956 			   (be32_to_cpu(rhead->h_version) &
2957 			    (~XLOG_VERSION_OKBITS))))) {
2958 		xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
2959 			__func__, be32_to_cpu(rhead->h_version));
2960 		return -EFSCORRUPTED;
2961 	}
2962 
2963 	/*
2964 	 * LR body must have data (or it wouldn't have been written)
2965 	 * and h_len must not be greater than LR buffer size.
2966 	 */
2967 	hlen = be32_to_cpu(rhead->h_len);
2968 	if (XFS_IS_CORRUPT(log->l_mp, hlen <= 0 || hlen > bufsize))
2969 		return -EFSCORRUPTED;
2970 
2971 	if (XFS_IS_CORRUPT(log->l_mp,
2972 			   blkno > log->l_logBBsize || blkno > INT_MAX))
2973 		return -EFSCORRUPTED;
2974 	return 0;
2975 }
2976 
2977 /*
2978  * Read the log from tail to head and process the log records found.
2979  * Handle the two cases where the tail and head are in the same cycle
2980  * and where the active portion of the log wraps around the end of
2981  * the physical log separately.  The pass parameter is passed through
2982  * to the routines called to process the data and is not looked at
2983  * here.
2984  */
2985 STATIC int
2986 xlog_do_recovery_pass(
2987 	struct xlog		*log,
2988 	xfs_daddr_t		head_blk,
2989 	xfs_daddr_t		tail_blk,
2990 	int			pass,
2991 	xfs_daddr_t		*first_bad)	/* out: first bad log rec */
2992 {
2993 	xlog_rec_header_t	*rhead;
2994 	xfs_daddr_t		blk_no, rblk_no;
2995 	xfs_daddr_t		rhead_blk;
2996 	char			*offset;
2997 	char			*hbp, *dbp;
2998 	int			error = 0, h_size, h_len;
2999 	int			error2 = 0;
3000 	int			bblks, split_bblks;
3001 	int			hblks = 1, split_hblks, wrapped_hblks;
3002 	int			i;
3003 	struct hlist_head	rhash[XLOG_RHASH_SIZE];
3004 	LIST_HEAD		(buffer_list);
3005 
3006 	ASSERT(head_blk != tail_blk);
3007 	blk_no = rhead_blk = tail_blk;
3008 
3009 	for (i = 0; i < XLOG_RHASH_SIZE; i++)
3010 		INIT_HLIST_HEAD(&rhash[i]);
3011 
3012 	hbp = xlog_alloc_buffer(log, hblks);
3013 	if (!hbp)
3014 		return -ENOMEM;
3015 
3016 	/*
3017 	 * Read the header of the tail block and get the iclog buffer size from
3018 	 * h_size.  Use this to tell how many sectors make up the log header.
3019 	 */
3020 	if (xfs_has_logv2(log->l_mp)) {
3021 		/*
3022 		 * When using variable length iclogs, read first sector of
3023 		 * iclog header and extract the header size from it.  Get a
3024 		 * new hbp that is the correct size.
3025 		 */
3026 		error = xlog_bread(log, tail_blk, 1, hbp, &offset);
3027 		if (error)
3028 			goto bread_err1;
3029 
3030 		rhead = (xlog_rec_header_t *)offset;
3031 
3032 		/*
3033 		 * xfsprogs has a bug where record length is based on lsunit but
3034 		 * h_size (iclog size) is hardcoded to 32k. Now that we
3035 		 * unconditionally CRC verify the unmount record, this means the
3036 		 * log buffer can be too small for the record and cause an
3037 		 * overrun.
3038 		 *
3039 		 * Detect this condition here. Use lsunit for the buffer size as
3040 		 * long as this looks like the mkfs case. Otherwise, return an
3041 		 * error to avoid a buffer overrun.
3042 		 */
3043 		h_size = be32_to_cpu(rhead->h_size);
3044 		h_len = be32_to_cpu(rhead->h_len);
3045 		if (h_len > h_size && h_len <= log->l_mp->m_logbsize &&
3046 		    rhead->h_num_logops == cpu_to_be32(1)) {
3047 			xfs_warn(log->l_mp,
3048 		"invalid iclog size (%d bytes), using lsunit (%d bytes)",
3049 				 h_size, log->l_mp->m_logbsize);
3050 			h_size = log->l_mp->m_logbsize;
3051 		}
3052 
3053 		error = xlog_valid_rec_header(log, rhead, tail_blk, h_size);
3054 		if (error)
3055 			goto bread_err1;
3056 
3057 		/*
3058 		 * This open codes xlog_logrec_hblks so that we can reuse the
3059 		 * fixed up h_size value calculated above.  Without that we'd
3060 		 * still allocate the buffer based on the incorrect on-disk
3061 		 * size.
3062 		 */
3063 		if (h_size > XLOG_HEADER_CYCLE_SIZE &&
3064 		    (rhead->h_version & cpu_to_be32(XLOG_VERSION_2))) {
3065 			hblks = DIV_ROUND_UP(h_size, XLOG_HEADER_CYCLE_SIZE);
3066 			if (hblks > 1) {
3067 				kvfree(hbp);
3068 				hbp = xlog_alloc_buffer(log, hblks);
3069 				if (!hbp)
3070 					return -ENOMEM;
3071 			}
3072 		}
3073 	} else {
3074 		ASSERT(log->l_sectBBsize == 1);
3075 		h_size = XLOG_BIG_RECORD_BSIZE;
3076 	}
3077 
3078 	dbp = xlog_alloc_buffer(log, BTOBB(h_size));
3079 	if (!dbp) {
3080 		kvfree(hbp);
3081 		return -ENOMEM;
3082 	}
3083 
3084 	memset(rhash, 0, sizeof(rhash));
3085 	if (tail_blk > head_blk) {
3086 		/*
3087 		 * Perform recovery around the end of the physical log.
3088 		 * When the head is not on the same cycle number as the tail,
3089 		 * we can't do a sequential recovery.
3090 		 */
3091 		while (blk_no < log->l_logBBsize) {
3092 			/*
3093 			 * Check for header wrapping around physical end-of-log
3094 			 */
3095 			offset = hbp;
3096 			split_hblks = 0;
3097 			wrapped_hblks = 0;
3098 			if (blk_no + hblks <= log->l_logBBsize) {
3099 				/* Read header in one read */
3100 				error = xlog_bread(log, blk_no, hblks, hbp,
3101 						   &offset);
3102 				if (error)
3103 					goto bread_err2;
3104 			} else {
3105 				/* This LR is split across physical log end */
3106 				if (blk_no != log->l_logBBsize) {
3107 					/* some data before physical log end */
3108 					ASSERT(blk_no <= INT_MAX);
3109 					split_hblks = log->l_logBBsize - (int)blk_no;
3110 					ASSERT(split_hblks > 0);
3111 					error = xlog_bread(log, blk_no,
3112 							   split_hblks, hbp,
3113 							   &offset);
3114 					if (error)
3115 						goto bread_err2;
3116 				}
3117 
3118 				/*
3119 				 * Note: this black magic still works with
3120 				 * large sector sizes (non-512) only because:
3121 				 * - we increased the buffer size originally
3122 				 *   by 1 sector giving us enough extra space
3123 				 *   for the second read;
3124 				 * - the log start is guaranteed to be sector
3125 				 *   aligned;
3126 				 * - we read the log end (LR header start)
3127 				 *   _first_, then the log start (LR header end)
3128 				 *   - order is important.
3129 				 */
3130 				wrapped_hblks = hblks - split_hblks;
3131 				error = xlog_bread_noalign(log, 0,
3132 						wrapped_hblks,
3133 						offset + BBTOB(split_hblks));
3134 				if (error)
3135 					goto bread_err2;
3136 			}
3137 			rhead = (xlog_rec_header_t *)offset;
3138 			error = xlog_valid_rec_header(log, rhead,
3139 					split_hblks ? blk_no : 0, h_size);
3140 			if (error)
3141 				goto bread_err2;
3142 
3143 			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3144 			blk_no += hblks;
3145 
3146 			/*
3147 			 * Read the log record data in multiple reads if it
3148 			 * wraps around the end of the log. Note that if the
3149 			 * header already wrapped, blk_no could point past the
3150 			 * end of the log. The record data is contiguous in
3151 			 * that case.
3152 			 */
3153 			if (blk_no + bblks <= log->l_logBBsize ||
3154 			    blk_no >= log->l_logBBsize) {
3155 				rblk_no = xlog_wrap_logbno(log, blk_no);
3156 				error = xlog_bread(log, rblk_no, bblks, dbp,
3157 						   &offset);
3158 				if (error)
3159 					goto bread_err2;
3160 			} else {
3161 				/* This log record is split across the
3162 				 * physical end of log */
3163 				offset = dbp;
3164 				split_bblks = 0;
3165 				if (blk_no != log->l_logBBsize) {
3166 					/* some data is before the physical
3167 					 * end of log */
3168 					ASSERT(!wrapped_hblks);
3169 					ASSERT(blk_no <= INT_MAX);
3170 					split_bblks =
3171 						log->l_logBBsize - (int)blk_no;
3172 					ASSERT(split_bblks > 0);
3173 					error = xlog_bread(log, blk_no,
3174 							split_bblks, dbp,
3175 							&offset);
3176 					if (error)
3177 						goto bread_err2;
3178 				}
3179 
3180 				/*
3181 				 * Note: this black magic still works with
3182 				 * large sector sizes (non-512) only because:
3183 				 * - we increased the buffer size originally
3184 				 *   by 1 sector giving us enough extra space
3185 				 *   for the second read;
3186 				 * - the log start is guaranteed to be sector
3187 				 *   aligned;
3188 				 * - we read the log end (LR header start)
3189 				 *   _first_, then the log start (LR header end)
3190 				 *   - order is important.
3191 				 */
3192 				error = xlog_bread_noalign(log, 0,
3193 						bblks - split_bblks,
3194 						offset + BBTOB(split_bblks));
3195 				if (error)
3196 					goto bread_err2;
3197 			}
3198 
3199 			error = xlog_recover_process(log, rhash, rhead, offset,
3200 						     pass, &buffer_list);
3201 			if (error)
3202 				goto bread_err2;
3203 
3204 			blk_no += bblks;
3205 			rhead_blk = blk_no;
3206 		}
3207 
3208 		ASSERT(blk_no >= log->l_logBBsize);
3209 		blk_no -= log->l_logBBsize;
3210 		rhead_blk = blk_no;
3211 	}
3212 
3213 	/* read first part of physical log */
3214 	while (blk_no < head_blk) {
3215 		error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3216 		if (error)
3217 			goto bread_err2;
3218 
3219 		rhead = (xlog_rec_header_t *)offset;
3220 		error = xlog_valid_rec_header(log, rhead, blk_no, h_size);
3221 		if (error)
3222 			goto bread_err2;
3223 
3224 		/* blocks in data section */
3225 		bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3226 		error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3227 				   &offset);
3228 		if (error)
3229 			goto bread_err2;
3230 
3231 		error = xlog_recover_process(log, rhash, rhead, offset, pass,
3232 					     &buffer_list);
3233 		if (error)
3234 			goto bread_err2;
3235 
3236 		blk_no += bblks + hblks;
3237 		rhead_blk = blk_no;
3238 	}
3239 
3240  bread_err2:
3241 	kvfree(dbp);
3242  bread_err1:
3243 	kvfree(hbp);
3244 
3245 	/*
3246 	 * Submit buffers that have been dirtied by the last record recovered.
3247 	 */
3248 	if (!list_empty(&buffer_list)) {
3249 		if (error) {
3250 			/*
3251 			 * If there has been an item recovery error then we
3252 			 * cannot allow partial checkpoint writeback to
3253 			 * occur.  We might have multiple checkpoints with the
3254 			 * same start LSN in this buffer list, and partial
3255 			 * writeback of a checkpoint in this situation can
3256 			 * prevent future recovery of all the changes in the
3257 			 * checkpoints at this start LSN.
3258 			 *
3259 			 * Note: Shutting down the filesystem will result in the
3260 			 * delwri submission marking all the buffers stale,
3261 			 * completing them and cleaning up _XBF_LOGRECOVERY
3262 			 * state without doing any IO.
3263 			 */
3264 			xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
3265 		}
3266 		error2 = xfs_buf_delwri_submit(&buffer_list);
3267 	}
3268 
3269 	if (error && first_bad)
3270 		*first_bad = rhead_blk;
3271 
3272 	/*
3273 	 * Transactions are freed at commit time but transactions without commit
3274 	 * records on disk are never committed. Free any that may be left in the
3275 	 * hash table.
3276 	 */
3277 	for (i = 0; i < XLOG_RHASH_SIZE; i++) {
3278 		struct hlist_node	*tmp;
3279 		struct xlog_recover	*trans;
3280 
3281 		hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
3282 			xlog_recover_free_trans(trans);
3283 	}
3284 
3285 	return error ? error : error2;
3286 }
3287 
3288 /*
3289  * Do the recovery of the log.  We actually do this in two phases.
3290  * The two passes are necessary in order to implement the function
3291  * of cancelling a record written into the log.  The first pass
3292  * determines those things which have been cancelled, and the
3293  * second pass replays log items normally except for those which
3294  * have been cancelled.  The handling of the replay and cancellations
3295  * takes place in the log item type specific routines.
3296  *
3297  * The table of items which have cancel records in the log is allocated
3298  * and freed at this level, since only here do we know when all of
3299  * the log recovery has been completed.
3300  */
3301 STATIC int
3302 xlog_do_log_recovery(
3303 	struct xlog	*log,
3304 	xfs_daddr_t	head_blk,
3305 	xfs_daddr_t	tail_blk)
3306 {
3307 	int		error;
3308 
3309 	ASSERT(head_blk != tail_blk);
3310 
3311 	/*
3312 	 * First do a pass to find all of the cancelled buf log items.
3313 	 * Store them in the buf_cancel_table for use in the second pass.
3314 	 */
3315 	error = xlog_alloc_buf_cancel_table(log);
3316 	if (error)
3317 		return error;
3318 
3319 	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3320 				      XLOG_RECOVER_PASS1, NULL);
3321 	if (error != 0)
3322 		goto out_cancel;
3323 
3324 	/*
3325 	 * Then do a second pass to actually recover the items in the log.
3326 	 * When it is complete free the table of buf cancel items.
3327 	 */
3328 	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3329 				      XLOG_RECOVER_PASS2, NULL);
3330 	if (!error)
3331 		xlog_check_buf_cancel_table(log);
3332 out_cancel:
3333 	xlog_free_buf_cancel_table(log);
3334 	return error;
3335 }
3336 
3337 /*
3338  * Do the actual recovery
3339  */
3340 STATIC int
3341 xlog_do_recover(
3342 	struct xlog		*log,
3343 	xfs_daddr_t		head_blk,
3344 	xfs_daddr_t		tail_blk)
3345 {
3346 	struct xfs_mount	*mp = log->l_mp;
3347 	struct xfs_buf		*bp = mp->m_sb_bp;
3348 	struct xfs_sb		*sbp = &mp->m_sb;
3349 	int			error;
3350 
3351 	trace_xfs_log_recover(log, head_blk, tail_blk);
3352 
3353 	/*
3354 	 * First replay the images in the log.
3355 	 */
3356 	error = xlog_do_log_recovery(log, head_blk, tail_blk);
3357 	if (error)
3358 		return error;
3359 
3360 	if (xlog_is_shutdown(log))
3361 		return -EIO;
3362 
3363 	/*
3364 	 * We now update the tail_lsn since much of the recovery has completed
3365 	 * and there may be space available to use.  If there were no extent or
3366 	 * iunlinks, we can free up the entire log.  This was set in
3367 	 * xlog_find_tail to be the lsn of the last known good LR on disk.  If
3368 	 * there are extent frees or iunlinks they will have some entries in the
3369 	 * AIL; so we look at the AIL to determine how to set the tail_lsn.
3370 	 */
3371 	xfs_ail_assign_tail_lsn(log->l_ailp);
3372 
3373 	/*
3374 	 * Now that we've finished replaying all buffer and inode updates,
3375 	 * re-read the superblock and reverify it.
3376 	 */
3377 	xfs_buf_lock(bp);
3378 	xfs_buf_hold(bp);
3379 	error = _xfs_buf_read(bp, XBF_READ);
3380 	if (error) {
3381 		if (!xlog_is_shutdown(log)) {
3382 			xfs_buf_ioerror_alert(bp, __this_address);
3383 			ASSERT(0);
3384 		}
3385 		xfs_buf_relse(bp);
3386 		return error;
3387 	}
3388 
3389 	/* Convert superblock from on-disk format */
3390 	xfs_sb_from_disk(sbp, bp->b_addr);
3391 	xfs_buf_relse(bp);
3392 
3393 	/* re-initialise in-core superblock and geometry structures */
3394 	mp->m_features |= xfs_sb_version_to_features(sbp);
3395 	xfs_reinit_percpu_counters(mp);
3396 	error = xfs_initialize_perag(mp, sbp->sb_agcount, sbp->sb_dblocks,
3397 			&mp->m_maxagi);
3398 	if (error) {
3399 		xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
3400 		return error;
3401 	}
3402 	mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
3403 
3404 	/* Normal transactions can now occur */
3405 	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
3406 	return 0;
3407 }
3408 
3409 /*
3410  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3411  *
3412  * Return error or zero.
3413  */
3414 int
3415 xlog_recover(
3416 	struct xlog	*log)
3417 {
3418 	xfs_daddr_t	head_blk, tail_blk;
3419 	int		error;
3420 
3421 	/* find the tail of the log */
3422 	error = xlog_find_tail(log, &head_blk, &tail_blk);
3423 	if (error)
3424 		return error;
3425 
3426 	/*
3427 	 * The superblock was read before the log was available and thus the LSN
3428 	 * could not be verified. Check the superblock LSN against the current
3429 	 * LSN now that it's known.
3430 	 */
3431 	if (xfs_has_crc(log->l_mp) &&
3432 	    !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
3433 		return -EINVAL;
3434 
3435 	if (tail_blk != head_blk) {
3436 		/* There used to be a comment here:
3437 		 *
3438 		 * disallow recovery on read-only mounts.  note -- mount
3439 		 * checks for ENOSPC and turns it into an intelligent
3440 		 * error message.
3441 		 * ...but this is no longer true.  Now, unless you specify
3442 		 * NORECOVERY (in which case this function would never be
3443 		 * called), we just go ahead and recover.  We do this all
3444 		 * under the vfs layer, so we can get away with it unless
3445 		 * the device itself is read-only, in which case we fail.
3446 		 */
3447 		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3448 			return error;
3449 		}
3450 
3451 		/*
3452 		 * Version 5 superblock log feature mask validation. We know the
3453 		 * log is dirty so check if there are any unknown log features
3454 		 * in what we need to recover. If there are unknown features
3455 		 * (e.g. unsupported transactions, then simply reject the
3456 		 * attempt at recovery before touching anything.
3457 		 */
3458 		if (xfs_sb_is_v5(&log->l_mp->m_sb) &&
3459 		    xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
3460 					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
3461 			xfs_warn(log->l_mp,
3462 "Superblock has unknown incompatible log features (0x%x) enabled.",
3463 				(log->l_mp->m_sb.sb_features_log_incompat &
3464 					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
3465 			xfs_warn(log->l_mp,
3466 "The log can not be fully and/or safely recovered by this kernel.");
3467 			xfs_warn(log->l_mp,
3468 "Please recover the log on a kernel that supports the unknown features.");
3469 			return -EINVAL;
3470 		}
3471 
3472 		/*
3473 		 * Delay log recovery if the debug hook is set. This is debug
3474 		 * instrumentation to coordinate simulation of I/O failures with
3475 		 * log recovery.
3476 		 */
3477 		if (xfs_globals.log_recovery_delay) {
3478 			xfs_notice(log->l_mp,
3479 				"Delaying log recovery for %d seconds.",
3480 				xfs_globals.log_recovery_delay);
3481 			msleep(xfs_globals.log_recovery_delay * 1000);
3482 		}
3483 
3484 		xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
3485 				log->l_mp->m_logname ? log->l_mp->m_logname
3486 						     : "internal");
3487 
3488 		error = xlog_do_recover(log, head_blk, tail_blk);
3489 		set_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
3490 	}
3491 	return error;
3492 }
3493 
3494 /*
3495  * In the first part of recovery we replay inodes and buffers and build up the
3496  * list of intents which need to be processed. Here we process the intents and
3497  * clean up the on disk unlinked inode lists. This is separated from the first
3498  * part of recovery so that the root and real-time bitmap inodes can be read in
3499  * from disk in between the two stages.  This is necessary so that we can free
3500  * space in the real-time portion of the file system.
3501  *
3502  * We run this whole process under GFP_NOFS allocation context. We do a
3503  * combination of non-transactional and transactional work, yet we really don't
3504  * want to recurse into the filesystem from direct reclaim during any of this
3505  * processing. This allows all the recovery code run here not to care about the
3506  * memory allocation context it is running in.
3507  */
3508 int
3509 xlog_recover_finish(
3510 	struct xlog	*log)
3511 {
3512 	unsigned int	nofs_flags = memalloc_nofs_save();
3513 	int		error;
3514 
3515 	error = xlog_recover_process_intents(log);
3516 	if (error) {
3517 		/*
3518 		 * Cancel all the unprocessed intent items now so that we don't
3519 		 * leave them pinned in the AIL.  This can cause the AIL to
3520 		 * livelock on the pinned item if anyone tries to push the AIL
3521 		 * (inode reclaim does this) before we get around to
3522 		 * xfs_log_mount_cancel.
3523 		 */
3524 		xlog_recover_cancel_intents(log);
3525 		xfs_alert(log->l_mp, "Failed to recover intents");
3526 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
3527 		goto out_error;
3528 	}
3529 
3530 	/*
3531 	 * Sync the log to get all the intents out of the AIL.  This isn't
3532 	 * absolutely necessary, but it helps in case the unlink transactions
3533 	 * would have problems pushing the intents out of the way.
3534 	 */
3535 	xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3536 
3537 	xlog_recover_process_iunlinks(log);
3538 
3539 	/*
3540 	 * Recover any CoW staging blocks that are still referenced by the
3541 	 * ondisk refcount metadata.  During mount there cannot be any live
3542 	 * staging extents as we have not permitted any user modifications.
3543 	 * Therefore, it is safe to free them all right now, even on a
3544 	 * read-only mount.
3545 	 */
3546 	error = xfs_reflink_recover_cow(log->l_mp);
3547 	if (error) {
3548 		xfs_alert(log->l_mp,
3549 	"Failed to recover leftover CoW staging extents, err %d.",
3550 				error);
3551 		/*
3552 		 * If we get an error here, make sure the log is shut down
3553 		 * but return zero so that any log items committed since the
3554 		 * end of intents processing can be pushed through the CIL
3555 		 * and AIL.
3556 		 */
3557 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
3558 		error = 0;
3559 		goto out_error;
3560 	}
3561 
3562 out_error:
3563 	memalloc_nofs_restore(nofs_flags);
3564 	return error;
3565 }
3566 
3567 void
3568 xlog_recover_cancel(
3569 	struct xlog	*log)
3570 {
3571 	if (xlog_recovery_needed(log))
3572 		xlog_recover_cancel_intents(log);
3573 }
3574 
3575