xref: /openbmc/linux/fs/xfs/xfs_log_recover.c (revision baf2c002)
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 = 1 << ffs(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 	kmem_free(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 	kmem_free(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 	kmem_free(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 	kmem_free(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 	kmem_free(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 	kmem_free(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 			xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1181 					log->l_curr_cycle, after_umount_blk);
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 	atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1216 	xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1217 					BBTOB(log->l_curr_block));
1218 	xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1219 					BBTOB(log->l_curr_block));
1220 }
1221 
1222 /*
1223  * Find the sync block number or the tail of the log.
1224  *
1225  * This will be the block number of the last record to have its
1226  * associated buffers synced to disk.  Every log record header has
1227  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
1228  * to get a sync block number.  The only concern is to figure out which
1229  * log record header to believe.
1230  *
1231  * The following algorithm uses the log record header with the largest
1232  * lsn.  The entire log record does not need to be valid.  We only care
1233  * that the header is valid.
1234  *
1235  * We could speed up search by using current head_blk buffer, but it is not
1236  * available.
1237  */
1238 STATIC int
1239 xlog_find_tail(
1240 	struct xlog		*log,
1241 	xfs_daddr_t		*head_blk,
1242 	xfs_daddr_t		*tail_blk)
1243 {
1244 	xlog_rec_header_t	*rhead;
1245 	char			*offset = NULL;
1246 	char			*buffer;
1247 	int			error;
1248 	xfs_daddr_t		rhead_blk;
1249 	xfs_lsn_t		tail_lsn;
1250 	bool			wrapped = false;
1251 	bool			clean = false;
1252 
1253 	/*
1254 	 * Find previous log record
1255 	 */
1256 	if ((error = xlog_find_head(log, head_blk)))
1257 		return error;
1258 	ASSERT(*head_blk < INT_MAX);
1259 
1260 	buffer = xlog_alloc_buffer(log, 1);
1261 	if (!buffer)
1262 		return -ENOMEM;
1263 	if (*head_blk == 0) {				/* special case */
1264 		error = xlog_bread(log, 0, 1, buffer, &offset);
1265 		if (error)
1266 			goto done;
1267 
1268 		if (xlog_get_cycle(offset) == 0) {
1269 			*tail_blk = 0;
1270 			/* leave all other log inited values alone */
1271 			goto done;
1272 		}
1273 	}
1274 
1275 	/*
1276 	 * Search backwards through the log looking for the log record header
1277 	 * block. This wraps all the way back around to the head so something is
1278 	 * seriously wrong if we can't find it.
1279 	 */
1280 	error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, buffer,
1281 				      &rhead_blk, &rhead, &wrapped);
1282 	if (error < 0)
1283 		goto done;
1284 	if (!error) {
1285 		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1286 		error = -EFSCORRUPTED;
1287 		goto done;
1288 	}
1289 	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1290 
1291 	/*
1292 	 * Set the log state based on the current head record.
1293 	 */
1294 	xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1295 	tail_lsn = atomic64_read(&log->l_tail_lsn);
1296 
1297 	/*
1298 	 * Look for an unmount record at the head of the log. This sets the log
1299 	 * state to determine whether recovery is necessary.
1300 	 */
1301 	error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1302 				       rhead_blk, buffer, &clean);
1303 	if (error)
1304 		goto done;
1305 
1306 	/*
1307 	 * Verify the log head if the log is not clean (e.g., we have anything
1308 	 * but an unmount record at the head). This uses CRC verification to
1309 	 * detect and trim torn writes. If discovered, CRC failures are
1310 	 * considered torn writes and the log head is trimmed accordingly.
1311 	 *
1312 	 * Note that we can only run CRC verification when the log is dirty
1313 	 * because there's no guarantee that the log data behind an unmount
1314 	 * record is compatible with the current architecture.
1315 	 */
1316 	if (!clean) {
1317 		xfs_daddr_t	orig_head = *head_blk;
1318 
1319 		error = xlog_verify_head(log, head_blk, tail_blk, buffer,
1320 					 &rhead_blk, &rhead, &wrapped);
1321 		if (error)
1322 			goto done;
1323 
1324 		/* update in-core state again if the head changed */
1325 		if (*head_blk != orig_head) {
1326 			xlog_set_state(log, *head_blk, rhead, rhead_blk,
1327 				       wrapped);
1328 			tail_lsn = atomic64_read(&log->l_tail_lsn);
1329 			error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1330 						       rhead, rhead_blk, buffer,
1331 						       &clean);
1332 			if (error)
1333 				goto done;
1334 		}
1335 	}
1336 
1337 	/*
1338 	 * Note that the unmount was clean. If the unmount was not clean, we
1339 	 * need to know this to rebuild the superblock counters from the perag
1340 	 * headers if we have a filesystem using non-persistent counters.
1341 	 */
1342 	if (clean)
1343 		set_bit(XFS_OPSTATE_CLEAN, &log->l_mp->m_opstate);
1344 
1345 	/*
1346 	 * Make sure that there are no blocks in front of the head
1347 	 * with the same cycle number as the head.  This can happen
1348 	 * because we allow multiple outstanding log writes concurrently,
1349 	 * and the later writes might make it out before earlier ones.
1350 	 *
1351 	 * We use the lsn from before modifying it so that we'll never
1352 	 * overwrite the unmount record after a clean unmount.
1353 	 *
1354 	 * Do this only if we are going to recover the filesystem
1355 	 *
1356 	 * NOTE: This used to say "if (!readonly)"
1357 	 * However on Linux, we can & do recover a read-only filesystem.
1358 	 * We only skip recovery if NORECOVERY is specified on mount,
1359 	 * in which case we would not be here.
1360 	 *
1361 	 * But... if the -device- itself is readonly, just skip this.
1362 	 * We can't recover this device anyway, so it won't matter.
1363 	 */
1364 	if (!xfs_readonly_buftarg(log->l_targ))
1365 		error = xlog_clear_stale_blocks(log, tail_lsn);
1366 
1367 done:
1368 	kmem_free(buffer);
1369 
1370 	if (error)
1371 		xfs_warn(log->l_mp, "failed to locate log tail");
1372 	return error;
1373 }
1374 
1375 /*
1376  * Is the log zeroed at all?
1377  *
1378  * The last binary search should be changed to perform an X block read
1379  * once X becomes small enough.  You can then search linearly through
1380  * the X blocks.  This will cut down on the number of reads we need to do.
1381  *
1382  * If the log is partially zeroed, this routine will pass back the blkno
1383  * of the first block with cycle number 0.  It won't have a complete LR
1384  * preceding it.
1385  *
1386  * Return:
1387  *	0  => the log is completely written to
1388  *	1 => use *blk_no as the first block of the log
1389  *	<0 => error has occurred
1390  */
1391 STATIC int
1392 xlog_find_zeroed(
1393 	struct xlog	*log,
1394 	xfs_daddr_t	*blk_no)
1395 {
1396 	char		*buffer;
1397 	char		*offset;
1398 	uint	        first_cycle, last_cycle;
1399 	xfs_daddr_t	new_blk, last_blk, start_blk;
1400 	xfs_daddr_t     num_scan_bblks;
1401 	int	        error, log_bbnum = log->l_logBBsize;
1402 
1403 	*blk_no = 0;
1404 
1405 	/* check totally zeroed log */
1406 	buffer = xlog_alloc_buffer(log, 1);
1407 	if (!buffer)
1408 		return -ENOMEM;
1409 	error = xlog_bread(log, 0, 1, buffer, &offset);
1410 	if (error)
1411 		goto out_free_buffer;
1412 
1413 	first_cycle = xlog_get_cycle(offset);
1414 	if (first_cycle == 0) {		/* completely zeroed log */
1415 		*blk_no = 0;
1416 		kmem_free(buffer);
1417 		return 1;
1418 	}
1419 
1420 	/* check partially zeroed log */
1421 	error = xlog_bread(log, log_bbnum-1, 1, buffer, &offset);
1422 	if (error)
1423 		goto out_free_buffer;
1424 
1425 	last_cycle = xlog_get_cycle(offset);
1426 	if (last_cycle != 0) {		/* log completely written to */
1427 		kmem_free(buffer);
1428 		return 0;
1429 	}
1430 
1431 	/* we have a partially zeroed log */
1432 	last_blk = log_bbnum-1;
1433 	error = xlog_find_cycle_start(log, buffer, 0, &last_blk, 0);
1434 	if (error)
1435 		goto out_free_buffer;
1436 
1437 	/*
1438 	 * Validate the answer.  Because there is no way to guarantee that
1439 	 * the entire log is made up of log records which are the same size,
1440 	 * we scan over the defined maximum blocks.  At this point, the maximum
1441 	 * is not chosen to mean anything special.   XXXmiken
1442 	 */
1443 	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1444 	ASSERT(num_scan_bblks <= INT_MAX);
1445 
1446 	if (last_blk < num_scan_bblks)
1447 		num_scan_bblks = last_blk;
1448 	start_blk = last_blk - num_scan_bblks;
1449 
1450 	/*
1451 	 * We search for any instances of cycle number 0 that occur before
1452 	 * our current estimate of the head.  What we're trying to detect is
1453 	 *        1 ... | 0 | 1 | 0...
1454 	 *                       ^ binary search ends here
1455 	 */
1456 	if ((error = xlog_find_verify_cycle(log, start_blk,
1457 					 (int)num_scan_bblks, 0, &new_blk)))
1458 		goto out_free_buffer;
1459 	if (new_blk != -1)
1460 		last_blk = new_blk;
1461 
1462 	/*
1463 	 * Potentially backup over partial log record write.  We don't need
1464 	 * to search the end of the log because we know it is zero.
1465 	 */
1466 	error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1467 	if (error == 1)
1468 		error = -EIO;
1469 	if (error)
1470 		goto out_free_buffer;
1471 
1472 	*blk_no = last_blk;
1473 out_free_buffer:
1474 	kmem_free(buffer);
1475 	if (error)
1476 		return error;
1477 	return 1;
1478 }
1479 
1480 /*
1481  * These are simple subroutines used by xlog_clear_stale_blocks() below
1482  * to initialize a buffer full of empty log record headers and write
1483  * them into the log.
1484  */
1485 STATIC void
1486 xlog_add_record(
1487 	struct xlog		*log,
1488 	char			*buf,
1489 	int			cycle,
1490 	int			block,
1491 	int			tail_cycle,
1492 	int			tail_block)
1493 {
1494 	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1495 
1496 	memset(buf, 0, BBSIZE);
1497 	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1498 	recp->h_cycle = cpu_to_be32(cycle);
1499 	recp->h_version = cpu_to_be32(
1500 			xfs_has_logv2(log->l_mp) ? 2 : 1);
1501 	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1502 	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1503 	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1504 	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1505 }
1506 
1507 STATIC int
1508 xlog_write_log_records(
1509 	struct xlog	*log,
1510 	int		cycle,
1511 	int		start_block,
1512 	int		blocks,
1513 	int		tail_cycle,
1514 	int		tail_block)
1515 {
1516 	char		*offset;
1517 	char		*buffer;
1518 	int		balign, ealign;
1519 	int		sectbb = log->l_sectBBsize;
1520 	int		end_block = start_block + blocks;
1521 	int		bufblks;
1522 	int		error = 0;
1523 	int		i, j = 0;
1524 
1525 	/*
1526 	 * Greedily allocate a buffer big enough to handle the full
1527 	 * range of basic blocks to be written.  If that fails, try
1528 	 * a smaller size.  We need to be able to write at least a
1529 	 * log sector, or we're out of luck.
1530 	 */
1531 	bufblks = 1 << ffs(blocks);
1532 	while (bufblks > log->l_logBBsize)
1533 		bufblks >>= 1;
1534 	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
1535 		bufblks >>= 1;
1536 		if (bufblks < sectbb)
1537 			return -ENOMEM;
1538 	}
1539 
1540 	/* We may need to do a read at the start to fill in part of
1541 	 * the buffer in the starting sector not covered by the first
1542 	 * write below.
1543 	 */
1544 	balign = round_down(start_block, sectbb);
1545 	if (balign != start_block) {
1546 		error = xlog_bread_noalign(log, start_block, 1, buffer);
1547 		if (error)
1548 			goto out_free_buffer;
1549 
1550 		j = start_block - balign;
1551 	}
1552 
1553 	for (i = start_block; i < end_block; i += bufblks) {
1554 		int		bcount, endcount;
1555 
1556 		bcount = min(bufblks, end_block - start_block);
1557 		endcount = bcount - j;
1558 
1559 		/* We may need to do a read at the end to fill in part of
1560 		 * the buffer in the final sector not covered by the write.
1561 		 * If this is the same sector as the above read, skip it.
1562 		 */
1563 		ealign = round_down(end_block, sectbb);
1564 		if (j == 0 && (start_block + endcount > ealign)) {
1565 			error = xlog_bread_noalign(log, ealign, sectbb,
1566 					buffer + BBTOB(ealign - start_block));
1567 			if (error)
1568 				break;
1569 
1570 		}
1571 
1572 		offset = buffer + xlog_align(log, start_block);
1573 		for (; j < endcount; j++) {
1574 			xlog_add_record(log, offset, cycle, i+j,
1575 					tail_cycle, tail_block);
1576 			offset += BBSIZE;
1577 		}
1578 		error = xlog_bwrite(log, start_block, endcount, buffer);
1579 		if (error)
1580 			break;
1581 		start_block += endcount;
1582 		j = 0;
1583 	}
1584 
1585 out_free_buffer:
1586 	kmem_free(buffer);
1587 	return error;
1588 }
1589 
1590 /*
1591  * This routine is called to blow away any incomplete log writes out
1592  * in front of the log head.  We do this so that we won't become confused
1593  * if we come up, write only a little bit more, and then crash again.
1594  * If we leave the partial log records out there, this situation could
1595  * cause us to think those partial writes are valid blocks since they
1596  * have the current cycle number.  We get rid of them by overwriting them
1597  * with empty log records with the old cycle number rather than the
1598  * current one.
1599  *
1600  * The tail lsn is passed in rather than taken from
1601  * the log so that we will not write over the unmount record after a
1602  * clean unmount in a 512 block log.  Doing so would leave the log without
1603  * any valid log records in it until a new one was written.  If we crashed
1604  * during that time we would not be able to recover.
1605  */
1606 STATIC int
1607 xlog_clear_stale_blocks(
1608 	struct xlog	*log,
1609 	xfs_lsn_t	tail_lsn)
1610 {
1611 	int		tail_cycle, head_cycle;
1612 	int		tail_block, head_block;
1613 	int		tail_distance, max_distance;
1614 	int		distance;
1615 	int		error;
1616 
1617 	tail_cycle = CYCLE_LSN(tail_lsn);
1618 	tail_block = BLOCK_LSN(tail_lsn);
1619 	head_cycle = log->l_curr_cycle;
1620 	head_block = log->l_curr_block;
1621 
1622 	/*
1623 	 * Figure out the distance between the new head of the log
1624 	 * and the tail.  We want to write over any blocks beyond the
1625 	 * head that we may have written just before the crash, but
1626 	 * we don't want to overwrite the tail of the log.
1627 	 */
1628 	if (head_cycle == tail_cycle) {
1629 		/*
1630 		 * The tail is behind the head in the physical log,
1631 		 * so the distance from the head to the tail is the
1632 		 * distance from the head to the end of the log plus
1633 		 * the distance from the beginning of the log to the
1634 		 * tail.
1635 		 */
1636 		if (XFS_IS_CORRUPT(log->l_mp,
1637 				   head_block < tail_block ||
1638 				   head_block >= log->l_logBBsize))
1639 			return -EFSCORRUPTED;
1640 		tail_distance = tail_block + (log->l_logBBsize - head_block);
1641 	} else {
1642 		/*
1643 		 * The head is behind the tail in the physical log,
1644 		 * so the distance from the head to the tail is just
1645 		 * the tail block minus the head block.
1646 		 */
1647 		if (XFS_IS_CORRUPT(log->l_mp,
1648 				   head_block >= tail_block ||
1649 				   head_cycle != tail_cycle + 1))
1650 			return -EFSCORRUPTED;
1651 		tail_distance = tail_block - head_block;
1652 	}
1653 
1654 	/*
1655 	 * If the head is right up against the tail, we can't clear
1656 	 * anything.
1657 	 */
1658 	if (tail_distance <= 0) {
1659 		ASSERT(tail_distance == 0);
1660 		return 0;
1661 	}
1662 
1663 	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1664 	/*
1665 	 * Take the smaller of the maximum amount of outstanding I/O
1666 	 * we could have and the distance to the tail to clear out.
1667 	 * We take the smaller so that we don't overwrite the tail and
1668 	 * we don't waste all day writing from the head to the tail
1669 	 * for no reason.
1670 	 */
1671 	max_distance = min(max_distance, tail_distance);
1672 
1673 	if ((head_block + max_distance) <= log->l_logBBsize) {
1674 		/*
1675 		 * We can stomp all the blocks we need to without
1676 		 * wrapping around the end of the log.  Just do it
1677 		 * in a single write.  Use the cycle number of the
1678 		 * current cycle minus one so that the log will look like:
1679 		 *     n ... | n - 1 ...
1680 		 */
1681 		error = xlog_write_log_records(log, (head_cycle - 1),
1682 				head_block, max_distance, tail_cycle,
1683 				tail_block);
1684 		if (error)
1685 			return error;
1686 	} else {
1687 		/*
1688 		 * We need to wrap around the end of the physical log in
1689 		 * order to clear all the blocks.  Do it in two separate
1690 		 * I/Os.  The first write should be from the head to the
1691 		 * end of the physical log, and it should use the current
1692 		 * cycle number minus one just like above.
1693 		 */
1694 		distance = log->l_logBBsize - head_block;
1695 		error = xlog_write_log_records(log, (head_cycle - 1),
1696 				head_block, distance, tail_cycle,
1697 				tail_block);
1698 
1699 		if (error)
1700 			return error;
1701 
1702 		/*
1703 		 * Now write the blocks at the start of the physical log.
1704 		 * This writes the remainder of the blocks we want to clear.
1705 		 * It uses the current cycle number since we're now on the
1706 		 * same cycle as the head so that we get:
1707 		 *    n ... n ... | n - 1 ...
1708 		 *    ^^^^^ blocks we're writing
1709 		 */
1710 		distance = max_distance - (log->l_logBBsize - head_block);
1711 		error = xlog_write_log_records(log, head_cycle, 0, distance,
1712 				tail_cycle, tail_block);
1713 		if (error)
1714 			return error;
1715 	}
1716 
1717 	return 0;
1718 }
1719 
1720 /*
1721  * Release the recovered intent item in the AIL that matches the given intent
1722  * type and intent id.
1723  */
1724 void
1725 xlog_recover_release_intent(
1726 	struct xlog		*log,
1727 	unsigned short		intent_type,
1728 	uint64_t		intent_id)
1729 {
1730 	struct xfs_ail_cursor	cur;
1731 	struct xfs_log_item	*lip;
1732 	struct xfs_ail		*ailp = log->l_ailp;
1733 
1734 	spin_lock(&ailp->ail_lock);
1735 	for (lip = xfs_trans_ail_cursor_first(ailp, &cur, 0); lip != NULL;
1736 	     lip = xfs_trans_ail_cursor_next(ailp, &cur)) {
1737 		if (lip->li_type != intent_type)
1738 			continue;
1739 		if (!lip->li_ops->iop_match(lip, intent_id))
1740 			continue;
1741 
1742 		spin_unlock(&ailp->ail_lock);
1743 		lip->li_ops->iop_release(lip);
1744 		spin_lock(&ailp->ail_lock);
1745 		break;
1746 	}
1747 
1748 	xfs_trans_ail_cursor_done(&cur);
1749 	spin_unlock(&ailp->ail_lock);
1750 }
1751 
1752 int
1753 xlog_recover_iget(
1754 	struct xfs_mount	*mp,
1755 	xfs_ino_t		ino,
1756 	struct xfs_inode	**ipp)
1757 {
1758 	int			error;
1759 
1760 	error = xfs_iget(mp, NULL, ino, 0, 0, ipp);
1761 	if (error)
1762 		return error;
1763 
1764 	error = xfs_qm_dqattach(*ipp);
1765 	if (error) {
1766 		xfs_irele(*ipp);
1767 		return error;
1768 	}
1769 
1770 	if (VFS_I(*ipp)->i_nlink == 0)
1771 		xfs_iflags_set(*ipp, XFS_IRECOVERY);
1772 
1773 	return 0;
1774 }
1775 
1776 /******************************************************************************
1777  *
1778  *		Log recover routines
1779  *
1780  ******************************************************************************
1781  */
1782 static const struct xlog_recover_item_ops *xlog_recover_item_ops[] = {
1783 	&xlog_buf_item_ops,
1784 	&xlog_inode_item_ops,
1785 	&xlog_dquot_item_ops,
1786 	&xlog_quotaoff_item_ops,
1787 	&xlog_icreate_item_ops,
1788 	&xlog_efi_item_ops,
1789 	&xlog_efd_item_ops,
1790 	&xlog_rui_item_ops,
1791 	&xlog_rud_item_ops,
1792 	&xlog_cui_item_ops,
1793 	&xlog_cud_item_ops,
1794 	&xlog_bui_item_ops,
1795 	&xlog_bud_item_ops,
1796 	&xlog_attri_item_ops,
1797 	&xlog_attrd_item_ops,
1798 };
1799 
1800 static const struct xlog_recover_item_ops *
1801 xlog_find_item_ops(
1802 	struct xlog_recover_item		*item)
1803 {
1804 	unsigned int				i;
1805 
1806 	for (i = 0; i < ARRAY_SIZE(xlog_recover_item_ops); i++)
1807 		if (ITEM_TYPE(item) == xlog_recover_item_ops[i]->item_type)
1808 			return xlog_recover_item_ops[i];
1809 
1810 	return NULL;
1811 }
1812 
1813 /*
1814  * Sort the log items in the transaction.
1815  *
1816  * The ordering constraints are defined by the inode allocation and unlink
1817  * behaviour. The rules are:
1818  *
1819  *	1. Every item is only logged once in a given transaction. Hence it
1820  *	   represents the last logged state of the item. Hence ordering is
1821  *	   dependent on the order in which operations need to be performed so
1822  *	   required initial conditions are always met.
1823  *
1824  *	2. Cancelled buffers are recorded in pass 1 in a separate table and
1825  *	   there's nothing to replay from them so we can simply cull them
1826  *	   from the transaction. However, we can't do that until after we've
1827  *	   replayed all the other items because they may be dependent on the
1828  *	   cancelled buffer and replaying the cancelled buffer can remove it
1829  *	   form the cancelled buffer table. Hence they have tobe done last.
1830  *
1831  *	3. Inode allocation buffers must be replayed before inode items that
1832  *	   read the buffer and replay changes into it. For filesystems using the
1833  *	   ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1834  *	   treated the same as inode allocation buffers as they create and
1835  *	   initialise the buffers directly.
1836  *
1837  *	4. Inode unlink buffers must be replayed after inode items are replayed.
1838  *	   This ensures that inodes are completely flushed to the inode buffer
1839  *	   in a "free" state before we remove the unlinked inode list pointer.
1840  *
1841  * Hence the ordering needs to be inode allocation buffers first, inode items
1842  * second, inode unlink buffers third and cancelled buffers last.
1843  *
1844  * But there's a problem with that - we can't tell an inode allocation buffer
1845  * apart from a regular buffer, so we can't separate them. We can, however,
1846  * tell an inode unlink buffer from the others, and so we can separate them out
1847  * from all the other buffers and move them to last.
1848  *
1849  * Hence, 4 lists, in order from head to tail:
1850  *	- buffer_list for all buffers except cancelled/inode unlink buffers
1851  *	- item_list for all non-buffer items
1852  *	- inode_buffer_list for inode unlink buffers
1853  *	- cancel_list for the cancelled buffers
1854  *
1855  * Note that we add objects to the tail of the lists so that first-to-last
1856  * ordering is preserved within the lists. Adding objects to the head of the
1857  * list means when we traverse from the head we walk them in last-to-first
1858  * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1859  * but for all other items there may be specific ordering that we need to
1860  * preserve.
1861  */
1862 STATIC int
1863 xlog_recover_reorder_trans(
1864 	struct xlog		*log,
1865 	struct xlog_recover	*trans,
1866 	int			pass)
1867 {
1868 	struct xlog_recover_item *item, *n;
1869 	int			error = 0;
1870 	LIST_HEAD(sort_list);
1871 	LIST_HEAD(cancel_list);
1872 	LIST_HEAD(buffer_list);
1873 	LIST_HEAD(inode_buffer_list);
1874 	LIST_HEAD(item_list);
1875 
1876 	list_splice_init(&trans->r_itemq, &sort_list);
1877 	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1878 		enum xlog_recover_reorder	fate = XLOG_REORDER_ITEM_LIST;
1879 
1880 		item->ri_ops = xlog_find_item_ops(item);
1881 		if (!item->ri_ops) {
1882 			xfs_warn(log->l_mp,
1883 				"%s: unrecognized type of log operation (%d)",
1884 				__func__, ITEM_TYPE(item));
1885 			ASSERT(0);
1886 			/*
1887 			 * return the remaining items back to the transaction
1888 			 * item list so they can be freed in caller.
1889 			 */
1890 			if (!list_empty(&sort_list))
1891 				list_splice_init(&sort_list, &trans->r_itemq);
1892 			error = -EFSCORRUPTED;
1893 			break;
1894 		}
1895 
1896 		if (item->ri_ops->reorder)
1897 			fate = item->ri_ops->reorder(item);
1898 
1899 		switch (fate) {
1900 		case XLOG_REORDER_BUFFER_LIST:
1901 			list_move_tail(&item->ri_list, &buffer_list);
1902 			break;
1903 		case XLOG_REORDER_CANCEL_LIST:
1904 			trace_xfs_log_recover_item_reorder_head(log,
1905 					trans, item, pass);
1906 			list_move(&item->ri_list, &cancel_list);
1907 			break;
1908 		case XLOG_REORDER_INODE_BUFFER_LIST:
1909 			list_move(&item->ri_list, &inode_buffer_list);
1910 			break;
1911 		case XLOG_REORDER_ITEM_LIST:
1912 			trace_xfs_log_recover_item_reorder_tail(log,
1913 							trans, item, pass);
1914 			list_move_tail(&item->ri_list, &item_list);
1915 			break;
1916 		}
1917 	}
1918 
1919 	ASSERT(list_empty(&sort_list));
1920 	if (!list_empty(&buffer_list))
1921 		list_splice(&buffer_list, &trans->r_itemq);
1922 	if (!list_empty(&item_list))
1923 		list_splice_tail(&item_list, &trans->r_itemq);
1924 	if (!list_empty(&inode_buffer_list))
1925 		list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1926 	if (!list_empty(&cancel_list))
1927 		list_splice_tail(&cancel_list, &trans->r_itemq);
1928 	return error;
1929 }
1930 
1931 void
1932 xlog_buf_readahead(
1933 	struct xlog		*log,
1934 	xfs_daddr_t		blkno,
1935 	uint			len,
1936 	const struct xfs_buf_ops *ops)
1937 {
1938 	if (!xlog_is_buffer_cancelled(log, blkno, len))
1939 		xfs_buf_readahead(log->l_mp->m_ddev_targp, blkno, len, ops);
1940 }
1941 
1942 STATIC int
1943 xlog_recover_items_pass2(
1944 	struct xlog                     *log,
1945 	struct xlog_recover             *trans,
1946 	struct list_head                *buffer_list,
1947 	struct list_head                *item_list)
1948 {
1949 	struct xlog_recover_item	*item;
1950 	int				error = 0;
1951 
1952 	list_for_each_entry(item, item_list, ri_list) {
1953 		trace_xfs_log_recover_item_recover(log, trans, item,
1954 				XLOG_RECOVER_PASS2);
1955 
1956 		if (item->ri_ops->commit_pass2)
1957 			error = item->ri_ops->commit_pass2(log, buffer_list,
1958 					item, trans->r_lsn);
1959 		if (error)
1960 			return error;
1961 	}
1962 
1963 	return error;
1964 }
1965 
1966 /*
1967  * Perform the transaction.
1968  *
1969  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
1970  * EFIs and EFDs get queued up by adding entries into the AIL for them.
1971  */
1972 STATIC int
1973 xlog_recover_commit_trans(
1974 	struct xlog		*log,
1975 	struct xlog_recover	*trans,
1976 	int			pass,
1977 	struct list_head	*buffer_list)
1978 {
1979 	int				error = 0;
1980 	int				items_queued = 0;
1981 	struct xlog_recover_item	*item;
1982 	struct xlog_recover_item	*next;
1983 	LIST_HEAD			(ra_list);
1984 	LIST_HEAD			(done_list);
1985 
1986 	#define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
1987 
1988 	hlist_del_init(&trans->r_list);
1989 
1990 	error = xlog_recover_reorder_trans(log, trans, pass);
1991 	if (error)
1992 		return error;
1993 
1994 	list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
1995 		trace_xfs_log_recover_item_recover(log, trans, item, pass);
1996 
1997 		switch (pass) {
1998 		case XLOG_RECOVER_PASS1:
1999 			if (item->ri_ops->commit_pass1)
2000 				error = item->ri_ops->commit_pass1(log, item);
2001 			break;
2002 		case XLOG_RECOVER_PASS2:
2003 			if (item->ri_ops->ra_pass2)
2004 				item->ri_ops->ra_pass2(log, item);
2005 			list_move_tail(&item->ri_list, &ra_list);
2006 			items_queued++;
2007 			if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
2008 				error = xlog_recover_items_pass2(log, trans,
2009 						buffer_list, &ra_list);
2010 				list_splice_tail_init(&ra_list, &done_list);
2011 				items_queued = 0;
2012 			}
2013 
2014 			break;
2015 		default:
2016 			ASSERT(0);
2017 		}
2018 
2019 		if (error)
2020 			goto out;
2021 	}
2022 
2023 out:
2024 	if (!list_empty(&ra_list)) {
2025 		if (!error)
2026 			error = xlog_recover_items_pass2(log, trans,
2027 					buffer_list, &ra_list);
2028 		list_splice_tail_init(&ra_list, &done_list);
2029 	}
2030 
2031 	if (!list_empty(&done_list))
2032 		list_splice_init(&done_list, &trans->r_itemq);
2033 
2034 	return error;
2035 }
2036 
2037 STATIC void
2038 xlog_recover_add_item(
2039 	struct list_head	*head)
2040 {
2041 	struct xlog_recover_item *item;
2042 
2043 	item = kmem_zalloc(sizeof(struct xlog_recover_item), 0);
2044 	INIT_LIST_HEAD(&item->ri_list);
2045 	list_add_tail(&item->ri_list, head);
2046 }
2047 
2048 STATIC int
2049 xlog_recover_add_to_cont_trans(
2050 	struct xlog		*log,
2051 	struct xlog_recover	*trans,
2052 	char			*dp,
2053 	int			len)
2054 {
2055 	struct xlog_recover_item *item;
2056 	char			*ptr, *old_ptr;
2057 	int			old_len;
2058 
2059 	/*
2060 	 * If the transaction is empty, the header was split across this and the
2061 	 * previous record. Copy the rest of the header.
2062 	 */
2063 	if (list_empty(&trans->r_itemq)) {
2064 		ASSERT(len <= sizeof(struct xfs_trans_header));
2065 		if (len > sizeof(struct xfs_trans_header)) {
2066 			xfs_warn(log->l_mp, "%s: bad header length", __func__);
2067 			return -EFSCORRUPTED;
2068 		}
2069 
2070 		xlog_recover_add_item(&trans->r_itemq);
2071 		ptr = (char *)&trans->r_theader +
2072 				sizeof(struct xfs_trans_header) - len;
2073 		memcpy(ptr, dp, len);
2074 		return 0;
2075 	}
2076 
2077 	/* take the tail entry */
2078 	item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2079 			  ri_list);
2080 
2081 	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
2082 	old_len = item->ri_buf[item->ri_cnt-1].i_len;
2083 
2084 	ptr = kvrealloc(old_ptr, old_len, len + old_len, GFP_KERNEL);
2085 	if (!ptr)
2086 		return -ENOMEM;
2087 	memcpy(&ptr[old_len], dp, len);
2088 	item->ri_buf[item->ri_cnt-1].i_len += len;
2089 	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
2090 	trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
2091 	return 0;
2092 }
2093 
2094 /*
2095  * The next region to add is the start of a new region.  It could be
2096  * a whole region or it could be the first part of a new region.  Because
2097  * of this, the assumption here is that the type and size fields of all
2098  * format structures fit into the first 32 bits of the structure.
2099  *
2100  * This works because all regions must be 32 bit aligned.  Therefore, we
2101  * either have both fields or we have neither field.  In the case we have
2102  * neither field, the data part of the region is zero length.  We only have
2103  * a log_op_header and can throw away the header since a new one will appear
2104  * later.  If we have at least 4 bytes, then we can determine how many regions
2105  * will appear in the current log item.
2106  */
2107 STATIC int
2108 xlog_recover_add_to_trans(
2109 	struct xlog		*log,
2110 	struct xlog_recover	*trans,
2111 	char			*dp,
2112 	int			len)
2113 {
2114 	struct xfs_inode_log_format	*in_f;			/* any will do */
2115 	struct xlog_recover_item *item;
2116 	char			*ptr;
2117 
2118 	if (!len)
2119 		return 0;
2120 	if (list_empty(&trans->r_itemq)) {
2121 		/* we need to catch log corruptions here */
2122 		if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
2123 			xfs_warn(log->l_mp, "%s: bad header magic number",
2124 				__func__);
2125 			ASSERT(0);
2126 			return -EFSCORRUPTED;
2127 		}
2128 
2129 		if (len > sizeof(struct xfs_trans_header)) {
2130 			xfs_warn(log->l_mp, "%s: bad header length", __func__);
2131 			ASSERT(0);
2132 			return -EFSCORRUPTED;
2133 		}
2134 
2135 		/*
2136 		 * The transaction header can be arbitrarily split across op
2137 		 * records. If we don't have the whole thing here, copy what we
2138 		 * do have and handle the rest in the next record.
2139 		 */
2140 		if (len == sizeof(struct xfs_trans_header))
2141 			xlog_recover_add_item(&trans->r_itemq);
2142 		memcpy(&trans->r_theader, dp, len);
2143 		return 0;
2144 	}
2145 
2146 	ptr = kmem_alloc(len, 0);
2147 	memcpy(ptr, dp, len);
2148 	in_f = (struct xfs_inode_log_format *)ptr;
2149 
2150 	/* take the tail entry */
2151 	item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2152 			  ri_list);
2153 	if (item->ri_total != 0 &&
2154 	     item->ri_total == item->ri_cnt) {
2155 		/* tail item is in use, get a new one */
2156 		xlog_recover_add_item(&trans->r_itemq);
2157 		item = list_entry(trans->r_itemq.prev,
2158 					struct xlog_recover_item, ri_list);
2159 	}
2160 
2161 	if (item->ri_total == 0) {		/* first region to be added */
2162 		if (in_f->ilf_size == 0 ||
2163 		    in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
2164 			xfs_warn(log->l_mp,
2165 		"bad number of regions (%d) in inode log format",
2166 				  in_f->ilf_size);
2167 			ASSERT(0);
2168 			kmem_free(ptr);
2169 			return -EFSCORRUPTED;
2170 		}
2171 
2172 		item->ri_total = in_f->ilf_size;
2173 		item->ri_buf =
2174 			kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
2175 				    0);
2176 	}
2177 
2178 	if (item->ri_total <= item->ri_cnt) {
2179 		xfs_warn(log->l_mp,
2180 	"log item region count (%d) overflowed size (%d)",
2181 				item->ri_cnt, item->ri_total);
2182 		ASSERT(0);
2183 		kmem_free(ptr);
2184 		return -EFSCORRUPTED;
2185 	}
2186 
2187 	/* Description region is ri_buf[0] */
2188 	item->ri_buf[item->ri_cnt].i_addr = ptr;
2189 	item->ri_buf[item->ri_cnt].i_len  = len;
2190 	item->ri_cnt++;
2191 	trace_xfs_log_recover_item_add(log, trans, item, 0);
2192 	return 0;
2193 }
2194 
2195 /*
2196  * Free up any resources allocated by the transaction
2197  *
2198  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2199  */
2200 STATIC void
2201 xlog_recover_free_trans(
2202 	struct xlog_recover	*trans)
2203 {
2204 	struct xlog_recover_item *item, *n;
2205 	int			i;
2206 
2207 	hlist_del_init(&trans->r_list);
2208 
2209 	list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2210 		/* Free the regions in the item. */
2211 		list_del(&item->ri_list);
2212 		for (i = 0; i < item->ri_cnt; i++)
2213 			kmem_free(item->ri_buf[i].i_addr);
2214 		/* Free the item itself */
2215 		kmem_free(item->ri_buf);
2216 		kmem_free(item);
2217 	}
2218 	/* Free the transaction recover structure */
2219 	kmem_free(trans);
2220 }
2221 
2222 /*
2223  * On error or completion, trans is freed.
2224  */
2225 STATIC int
2226 xlog_recovery_process_trans(
2227 	struct xlog		*log,
2228 	struct xlog_recover	*trans,
2229 	char			*dp,
2230 	unsigned int		len,
2231 	unsigned int		flags,
2232 	int			pass,
2233 	struct list_head	*buffer_list)
2234 {
2235 	int			error = 0;
2236 	bool			freeit = false;
2237 
2238 	/* mask off ophdr transaction container flags */
2239 	flags &= ~XLOG_END_TRANS;
2240 	if (flags & XLOG_WAS_CONT_TRANS)
2241 		flags &= ~XLOG_CONTINUE_TRANS;
2242 
2243 	/*
2244 	 * Callees must not free the trans structure. We'll decide if we need to
2245 	 * free it or not based on the operation being done and it's result.
2246 	 */
2247 	switch (flags) {
2248 	/* expected flag values */
2249 	case 0:
2250 	case XLOG_CONTINUE_TRANS:
2251 		error = xlog_recover_add_to_trans(log, trans, dp, len);
2252 		break;
2253 	case XLOG_WAS_CONT_TRANS:
2254 		error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
2255 		break;
2256 	case XLOG_COMMIT_TRANS:
2257 		error = xlog_recover_commit_trans(log, trans, pass,
2258 						  buffer_list);
2259 		/* success or fail, we are now done with this transaction. */
2260 		freeit = true;
2261 		break;
2262 
2263 	/* unexpected flag values */
2264 	case XLOG_UNMOUNT_TRANS:
2265 		/* just skip trans */
2266 		xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
2267 		freeit = true;
2268 		break;
2269 	case XLOG_START_TRANS:
2270 	default:
2271 		xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
2272 		ASSERT(0);
2273 		error = -EFSCORRUPTED;
2274 		break;
2275 	}
2276 	if (error || freeit)
2277 		xlog_recover_free_trans(trans);
2278 	return error;
2279 }
2280 
2281 /*
2282  * Lookup the transaction recovery structure associated with the ID in the
2283  * current ophdr. If the transaction doesn't exist and the start flag is set in
2284  * the ophdr, then allocate a new transaction for future ID matches to find.
2285  * Either way, return what we found during the lookup - an existing transaction
2286  * or nothing.
2287  */
2288 STATIC struct xlog_recover *
2289 xlog_recover_ophdr_to_trans(
2290 	struct hlist_head	rhash[],
2291 	struct xlog_rec_header	*rhead,
2292 	struct xlog_op_header	*ohead)
2293 {
2294 	struct xlog_recover	*trans;
2295 	xlog_tid_t		tid;
2296 	struct hlist_head	*rhp;
2297 
2298 	tid = be32_to_cpu(ohead->oh_tid);
2299 	rhp = &rhash[XLOG_RHASH(tid)];
2300 	hlist_for_each_entry(trans, rhp, r_list) {
2301 		if (trans->r_log_tid == tid)
2302 			return trans;
2303 	}
2304 
2305 	/*
2306 	 * skip over non-start transaction headers - we could be
2307 	 * processing slack space before the next transaction starts
2308 	 */
2309 	if (!(ohead->oh_flags & XLOG_START_TRANS))
2310 		return NULL;
2311 
2312 	ASSERT(be32_to_cpu(ohead->oh_len) == 0);
2313 
2314 	/*
2315 	 * This is a new transaction so allocate a new recovery container to
2316 	 * hold the recovery ops that will follow.
2317 	 */
2318 	trans = kmem_zalloc(sizeof(struct xlog_recover), 0);
2319 	trans->r_log_tid = tid;
2320 	trans->r_lsn = be64_to_cpu(rhead->h_lsn);
2321 	INIT_LIST_HEAD(&trans->r_itemq);
2322 	INIT_HLIST_NODE(&trans->r_list);
2323 	hlist_add_head(&trans->r_list, rhp);
2324 
2325 	/*
2326 	 * Nothing more to do for this ophdr. Items to be added to this new
2327 	 * transaction will be in subsequent ophdr containers.
2328 	 */
2329 	return NULL;
2330 }
2331 
2332 STATIC int
2333 xlog_recover_process_ophdr(
2334 	struct xlog		*log,
2335 	struct hlist_head	rhash[],
2336 	struct xlog_rec_header	*rhead,
2337 	struct xlog_op_header	*ohead,
2338 	char			*dp,
2339 	char			*end,
2340 	int			pass,
2341 	struct list_head	*buffer_list)
2342 {
2343 	struct xlog_recover	*trans;
2344 	unsigned int		len;
2345 	int			error;
2346 
2347 	/* Do we understand who wrote this op? */
2348 	if (ohead->oh_clientid != XFS_TRANSACTION &&
2349 	    ohead->oh_clientid != XFS_LOG) {
2350 		xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
2351 			__func__, ohead->oh_clientid);
2352 		ASSERT(0);
2353 		return -EFSCORRUPTED;
2354 	}
2355 
2356 	/*
2357 	 * Check the ophdr contains all the data it is supposed to contain.
2358 	 */
2359 	len = be32_to_cpu(ohead->oh_len);
2360 	if (dp + len > end) {
2361 		xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
2362 		WARN_ON(1);
2363 		return -EFSCORRUPTED;
2364 	}
2365 
2366 	trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
2367 	if (!trans) {
2368 		/* nothing to do, so skip over this ophdr */
2369 		return 0;
2370 	}
2371 
2372 	/*
2373 	 * The recovered buffer queue is drained only once we know that all
2374 	 * recovery items for the current LSN have been processed. This is
2375 	 * required because:
2376 	 *
2377 	 * - Buffer write submission updates the metadata LSN of the buffer.
2378 	 * - Log recovery skips items with a metadata LSN >= the current LSN of
2379 	 *   the recovery item.
2380 	 * - Separate recovery items against the same metadata buffer can share
2381 	 *   a current LSN. I.e., consider that the LSN of a recovery item is
2382 	 *   defined as the starting LSN of the first record in which its
2383 	 *   transaction appears, that a record can hold multiple transactions,
2384 	 *   and/or that a transaction can span multiple records.
2385 	 *
2386 	 * In other words, we are allowed to submit a buffer from log recovery
2387 	 * once per current LSN. Otherwise, we may incorrectly skip recovery
2388 	 * items and cause corruption.
2389 	 *
2390 	 * We don't know up front whether buffers are updated multiple times per
2391 	 * LSN. Therefore, track the current LSN of each commit log record as it
2392 	 * is processed and drain the queue when it changes. Use commit records
2393 	 * because they are ordered correctly by the logging code.
2394 	 */
2395 	if (log->l_recovery_lsn != trans->r_lsn &&
2396 	    ohead->oh_flags & XLOG_COMMIT_TRANS) {
2397 		error = xfs_buf_delwri_submit(buffer_list);
2398 		if (error)
2399 			return error;
2400 		log->l_recovery_lsn = trans->r_lsn;
2401 	}
2402 
2403 	return xlog_recovery_process_trans(log, trans, dp, len,
2404 					   ohead->oh_flags, pass, buffer_list);
2405 }
2406 
2407 /*
2408  * There are two valid states of the r_state field.  0 indicates that the
2409  * transaction structure is in a normal state.  We have either seen the
2410  * start of the transaction or the last operation we added was not a partial
2411  * operation.  If the last operation we added to the transaction was a
2412  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2413  *
2414  * NOTE: skip LRs with 0 data length.
2415  */
2416 STATIC int
2417 xlog_recover_process_data(
2418 	struct xlog		*log,
2419 	struct hlist_head	rhash[],
2420 	struct xlog_rec_header	*rhead,
2421 	char			*dp,
2422 	int			pass,
2423 	struct list_head	*buffer_list)
2424 {
2425 	struct xlog_op_header	*ohead;
2426 	char			*end;
2427 	int			num_logops;
2428 	int			error;
2429 
2430 	end = dp + be32_to_cpu(rhead->h_len);
2431 	num_logops = be32_to_cpu(rhead->h_num_logops);
2432 
2433 	/* check the log format matches our own - else we can't recover */
2434 	if (xlog_header_check_recover(log->l_mp, rhead))
2435 		return -EIO;
2436 
2437 	trace_xfs_log_recover_record(log, rhead, pass);
2438 	while ((dp < end) && num_logops) {
2439 
2440 		ohead = (struct xlog_op_header *)dp;
2441 		dp += sizeof(*ohead);
2442 		ASSERT(dp <= end);
2443 
2444 		/* errors will abort recovery */
2445 		error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
2446 						   dp, end, pass, buffer_list);
2447 		if (error)
2448 			return error;
2449 
2450 		dp += be32_to_cpu(ohead->oh_len);
2451 		num_logops--;
2452 	}
2453 	return 0;
2454 }
2455 
2456 /* Take all the collected deferred ops and finish them in order. */
2457 static int
2458 xlog_finish_defer_ops(
2459 	struct xfs_mount	*mp,
2460 	struct list_head	*capture_list)
2461 {
2462 	struct xfs_defer_capture *dfc, *next;
2463 	struct xfs_trans	*tp;
2464 	int			error = 0;
2465 
2466 	list_for_each_entry_safe(dfc, next, capture_list, dfc_list) {
2467 		struct xfs_trans_res	resv;
2468 		struct xfs_defer_resources dres;
2469 
2470 		/*
2471 		 * Create a new transaction reservation from the captured
2472 		 * information.  Set logcount to 1 to force the new transaction
2473 		 * to regrant every roll so that we can make forward progress
2474 		 * in recovery no matter how full the log might be.
2475 		 */
2476 		resv.tr_logres = dfc->dfc_logres;
2477 		resv.tr_logcount = 1;
2478 		resv.tr_logflags = XFS_TRANS_PERM_LOG_RES;
2479 
2480 		error = xfs_trans_alloc(mp, &resv, dfc->dfc_blkres,
2481 				dfc->dfc_rtxres, XFS_TRANS_RESERVE, &tp);
2482 		if (error) {
2483 			xlog_force_shutdown(mp->m_log, SHUTDOWN_LOG_IO_ERROR);
2484 			return error;
2485 		}
2486 
2487 		/*
2488 		 * Transfer to this new transaction all the dfops we captured
2489 		 * from recovering a single intent item.
2490 		 */
2491 		list_del_init(&dfc->dfc_list);
2492 		xfs_defer_ops_continue(dfc, tp, &dres);
2493 		error = xfs_trans_commit(tp);
2494 		xfs_defer_resources_rele(&dres);
2495 		if (error)
2496 			return error;
2497 	}
2498 
2499 	ASSERT(list_empty(capture_list));
2500 	return 0;
2501 }
2502 
2503 /* Release all the captured defer ops and capture structures in this list. */
2504 static void
2505 xlog_abort_defer_ops(
2506 	struct xfs_mount		*mp,
2507 	struct list_head		*capture_list)
2508 {
2509 	struct xfs_defer_capture	*dfc;
2510 	struct xfs_defer_capture	*next;
2511 
2512 	list_for_each_entry_safe(dfc, next, capture_list, dfc_list) {
2513 		list_del_init(&dfc->dfc_list);
2514 		xfs_defer_ops_capture_free(mp, dfc);
2515 	}
2516 }
2517 
2518 /*
2519  * When this is called, all of the log intent items which did not have
2520  * corresponding log done items should be in the AIL.  What we do now is update
2521  * the data structures associated with each one.
2522  *
2523  * Since we process the log intent items in normal transactions, they will be
2524  * removed at some point after the commit.  This prevents us from just walking
2525  * down the list processing each one.  We'll use a flag in the intent item to
2526  * skip those that we've already processed and use the AIL iteration mechanism's
2527  * generation count to try to speed this up at least a bit.
2528  *
2529  * When we start, we know that the intents are the only things in the AIL. As we
2530  * process them, however, other items are added to the AIL. Hence we know we
2531  * have started recovery on all the pending intents when we find an non-intent
2532  * item in the AIL.
2533  */
2534 STATIC int
2535 xlog_recover_process_intents(
2536 	struct xlog		*log)
2537 {
2538 	LIST_HEAD(capture_list);
2539 	struct xfs_ail_cursor	cur;
2540 	struct xfs_log_item	*lip;
2541 	struct xfs_ail		*ailp;
2542 	int			error = 0;
2543 #if defined(DEBUG) || defined(XFS_WARN)
2544 	xfs_lsn_t		last_lsn;
2545 #endif
2546 
2547 	ailp = log->l_ailp;
2548 	spin_lock(&ailp->ail_lock);
2549 #if defined(DEBUG) || defined(XFS_WARN)
2550 	last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
2551 #endif
2552 	for (lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2553 	     lip != NULL;
2554 	     lip = xfs_trans_ail_cursor_next(ailp, &cur)) {
2555 		if (!xlog_item_is_intent(lip))
2556 			break;
2557 
2558 		/*
2559 		 * We should never see a redo item with a LSN higher than
2560 		 * the last transaction we found in the log at the start
2561 		 * of recovery.
2562 		 */
2563 		ASSERT(XFS_LSN_CMP(last_lsn, lip->li_lsn) >= 0);
2564 
2565 		/*
2566 		 * NOTE: If your intent processing routine can create more
2567 		 * deferred ops, you /must/ attach them to the capture list in
2568 		 * the recover routine or else those subsequent intents will be
2569 		 * replayed in the wrong order!
2570 		 */
2571 		spin_unlock(&ailp->ail_lock);
2572 		error = lip->li_ops->iop_recover(lip, &capture_list);
2573 		spin_lock(&ailp->ail_lock);
2574 		if (error) {
2575 			trace_xlog_intent_recovery_failed(log->l_mp, error,
2576 					lip->li_ops->iop_recover);
2577 			break;
2578 		}
2579 	}
2580 
2581 	xfs_trans_ail_cursor_done(&cur);
2582 	spin_unlock(&ailp->ail_lock);
2583 	if (error)
2584 		goto err;
2585 
2586 	error = xlog_finish_defer_ops(log->l_mp, &capture_list);
2587 	if (error)
2588 		goto err;
2589 
2590 	return 0;
2591 err:
2592 	xlog_abort_defer_ops(log->l_mp, &capture_list);
2593 	return error;
2594 }
2595 
2596 /*
2597  * A cancel occurs when the mount has failed and we're bailing out.  Release all
2598  * pending log intent items that we haven't started recovery on so they don't
2599  * pin the AIL.
2600  */
2601 STATIC void
2602 xlog_recover_cancel_intents(
2603 	struct xlog		*log)
2604 {
2605 	struct xfs_log_item	*lip;
2606 	struct xfs_ail_cursor	cur;
2607 	struct xfs_ail		*ailp;
2608 
2609 	ailp = log->l_ailp;
2610 	spin_lock(&ailp->ail_lock);
2611 	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2612 	while (lip != NULL) {
2613 		if (!xlog_item_is_intent(lip))
2614 			break;
2615 
2616 		spin_unlock(&ailp->ail_lock);
2617 		lip->li_ops->iop_release(lip);
2618 		spin_lock(&ailp->ail_lock);
2619 		lip = xfs_trans_ail_cursor_next(ailp, &cur);
2620 	}
2621 
2622 	xfs_trans_ail_cursor_done(&cur);
2623 	spin_unlock(&ailp->ail_lock);
2624 }
2625 
2626 /*
2627  * This routine performs a transaction to null out a bad inode pointer
2628  * in an agi unlinked inode hash bucket.
2629  */
2630 STATIC void
2631 xlog_recover_clear_agi_bucket(
2632 	struct xfs_perag	*pag,
2633 	int			bucket)
2634 {
2635 	struct xfs_mount	*mp = pag->pag_mount;
2636 	struct xfs_trans	*tp;
2637 	struct xfs_agi		*agi;
2638 	struct xfs_buf		*agibp;
2639 	int			offset;
2640 	int			error;
2641 
2642 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
2643 	if (error)
2644 		goto out_error;
2645 
2646 	error = xfs_read_agi(pag, tp, &agibp);
2647 	if (error)
2648 		goto out_abort;
2649 
2650 	agi = agibp->b_addr;
2651 	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
2652 	offset = offsetof(xfs_agi_t, agi_unlinked) +
2653 		 (sizeof(xfs_agino_t) * bucket);
2654 	xfs_trans_log_buf(tp, agibp, offset,
2655 			  (offset + sizeof(xfs_agino_t) - 1));
2656 
2657 	error = xfs_trans_commit(tp);
2658 	if (error)
2659 		goto out_error;
2660 	return;
2661 
2662 out_abort:
2663 	xfs_trans_cancel(tp);
2664 out_error:
2665 	xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__,
2666 			pag->pag_agno);
2667 	return;
2668 }
2669 
2670 static int
2671 xlog_recover_iunlink_bucket(
2672 	struct xfs_perag	*pag,
2673 	struct xfs_agi		*agi,
2674 	int			bucket)
2675 {
2676 	struct xfs_mount	*mp = pag->pag_mount;
2677 	struct xfs_inode	*prev_ip = NULL;
2678 	struct xfs_inode	*ip;
2679 	xfs_agino_t		prev_agino, agino;
2680 	int			error = 0;
2681 
2682 	agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2683 	while (agino != NULLAGINO) {
2684 		error = xfs_iget(mp, NULL,
2685 				XFS_AGINO_TO_INO(mp, pag->pag_agno, agino),
2686 				0, 0, &ip);
2687 		if (error)
2688 			break;
2689 
2690 		ASSERT(VFS_I(ip)->i_nlink == 0);
2691 		ASSERT(VFS_I(ip)->i_mode != 0);
2692 		xfs_iflags_clear(ip, XFS_IRECOVERY);
2693 		agino = ip->i_next_unlinked;
2694 
2695 		if (prev_ip) {
2696 			ip->i_prev_unlinked = prev_agino;
2697 			xfs_irele(prev_ip);
2698 
2699 			/*
2700 			 * Ensure the inode is removed from the unlinked list
2701 			 * before we continue so that it won't race with
2702 			 * building the in-memory list here. This could be
2703 			 * serialised with the agibp lock, but that just
2704 			 * serialises via lockstepping and it's much simpler
2705 			 * just to flush the inodegc queue and wait for it to
2706 			 * complete.
2707 			 */
2708 			xfs_inodegc_flush(mp);
2709 		}
2710 
2711 		prev_agino = agino;
2712 		prev_ip = ip;
2713 	}
2714 
2715 	if (prev_ip) {
2716 		ip->i_prev_unlinked = prev_agino;
2717 		xfs_irele(prev_ip);
2718 	}
2719 	xfs_inodegc_flush(mp);
2720 	return error;
2721 }
2722 
2723 /*
2724  * Recover AGI unlinked lists
2725  *
2726  * This is called during recovery to process any inodes which we unlinked but
2727  * not freed when the system crashed.  These inodes will be on the lists in the
2728  * AGI blocks. What we do here is scan all the AGIs and fully truncate and free
2729  * any inodes found on the lists. Each inode is removed from the lists when it
2730  * has been fully truncated and is freed. The freeing of the inode and its
2731  * removal from the list must be atomic.
2732  *
2733  * If everything we touch in the agi processing loop is already in memory, this
2734  * loop can hold the cpu for a long time. It runs without lock contention,
2735  * memory allocation contention, the need wait for IO, etc, and so will run
2736  * until we either run out of inodes to process, run low on memory or we run out
2737  * of log space.
2738  *
2739  * This behaviour is bad for latency on single CPU and non-preemptible kernels,
2740  * and can prevent other filesystem work (such as CIL pushes) from running. This
2741  * can lead to deadlocks if the recovery process runs out of log reservation
2742  * space. Hence we need to yield the CPU when there is other kernel work
2743  * scheduled on this CPU to ensure other scheduled work can run without undue
2744  * latency.
2745  */
2746 static void
2747 xlog_recover_iunlink_ag(
2748 	struct xfs_perag	*pag)
2749 {
2750 	struct xfs_agi		*agi;
2751 	struct xfs_buf		*agibp;
2752 	int			bucket;
2753 	int			error;
2754 
2755 	error = xfs_read_agi(pag, NULL, &agibp);
2756 	if (error) {
2757 		/*
2758 		 * AGI is b0rked. Don't process it.
2759 		 *
2760 		 * We should probably mark the filesystem as corrupt after we've
2761 		 * recovered all the ag's we can....
2762 		 */
2763 		return;
2764 	}
2765 
2766 	/*
2767 	 * Unlock the buffer so that it can be acquired in the normal course of
2768 	 * the transaction to truncate and free each inode.  Because we are not
2769 	 * racing with anyone else here for the AGI buffer, we don't even need
2770 	 * to hold it locked to read the initial unlinked bucket entries out of
2771 	 * the buffer. We keep buffer reference though, so that it stays pinned
2772 	 * in memory while we need the buffer.
2773 	 */
2774 	agi = agibp->b_addr;
2775 	xfs_buf_unlock(agibp);
2776 
2777 	for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
2778 		error = xlog_recover_iunlink_bucket(pag, agi, bucket);
2779 		if (error) {
2780 			/*
2781 			 * Bucket is unrecoverable, so only a repair scan can
2782 			 * free the remaining unlinked inodes. Just empty the
2783 			 * bucket and remaining inodes on it unreferenced and
2784 			 * unfreeable.
2785 			 */
2786 			xfs_inodegc_flush(pag->pag_mount);
2787 			xlog_recover_clear_agi_bucket(pag, bucket);
2788 		}
2789 	}
2790 
2791 	xfs_buf_rele(agibp);
2792 }
2793 
2794 static void
2795 xlog_recover_process_iunlinks(
2796 	struct xlog	*log)
2797 {
2798 	struct xfs_perag	*pag;
2799 	xfs_agnumber_t		agno;
2800 
2801 	for_each_perag(log->l_mp, agno, pag)
2802 		xlog_recover_iunlink_ag(pag);
2803 
2804 	/*
2805 	 * Flush the pending unlinked inodes to ensure that the inactivations
2806 	 * are fully completed on disk and the incore inodes can be reclaimed
2807 	 * before we signal that recovery is complete.
2808 	 */
2809 	xfs_inodegc_flush(log->l_mp);
2810 }
2811 
2812 STATIC void
2813 xlog_unpack_data(
2814 	struct xlog_rec_header	*rhead,
2815 	char			*dp,
2816 	struct xlog		*log)
2817 {
2818 	int			i, j, k;
2819 
2820 	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
2821 		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
2822 		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
2823 		dp += BBSIZE;
2824 	}
2825 
2826 	if (xfs_has_logv2(log->l_mp)) {
2827 		xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
2828 		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
2829 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2830 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2831 			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
2832 			dp += BBSIZE;
2833 		}
2834 	}
2835 }
2836 
2837 /*
2838  * CRC check, unpack and process a log record.
2839  */
2840 STATIC int
2841 xlog_recover_process(
2842 	struct xlog		*log,
2843 	struct hlist_head	rhash[],
2844 	struct xlog_rec_header	*rhead,
2845 	char			*dp,
2846 	int			pass,
2847 	struct list_head	*buffer_list)
2848 {
2849 	__le32			old_crc = rhead->h_crc;
2850 	__le32			crc;
2851 
2852 	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
2853 
2854 	/*
2855 	 * Nothing else to do if this is a CRC verification pass. Just return
2856 	 * if this a record with a non-zero crc. Unfortunately, mkfs always
2857 	 * sets old_crc to 0 so we must consider this valid even on v5 supers.
2858 	 * Otherwise, return EFSBADCRC on failure so the callers up the stack
2859 	 * know precisely what failed.
2860 	 */
2861 	if (pass == XLOG_RECOVER_CRCPASS) {
2862 		if (old_crc && crc != old_crc)
2863 			return -EFSBADCRC;
2864 		return 0;
2865 	}
2866 
2867 	/*
2868 	 * We're in the normal recovery path. Issue a warning if and only if the
2869 	 * CRC in the header is non-zero. This is an advisory warning and the
2870 	 * zero CRC check prevents warnings from being emitted when upgrading
2871 	 * the kernel from one that does not add CRCs by default.
2872 	 */
2873 	if (crc != old_crc) {
2874 		if (old_crc || xfs_has_crc(log->l_mp)) {
2875 			xfs_alert(log->l_mp,
2876 		"log record CRC mismatch: found 0x%x, expected 0x%x.",
2877 					le32_to_cpu(old_crc),
2878 					le32_to_cpu(crc));
2879 			xfs_hex_dump(dp, 32);
2880 		}
2881 
2882 		/*
2883 		 * If the filesystem is CRC enabled, this mismatch becomes a
2884 		 * fatal log corruption failure.
2885 		 */
2886 		if (xfs_has_crc(log->l_mp)) {
2887 			XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, log->l_mp);
2888 			return -EFSCORRUPTED;
2889 		}
2890 	}
2891 
2892 	xlog_unpack_data(rhead, dp, log);
2893 
2894 	return xlog_recover_process_data(log, rhash, rhead, dp, pass,
2895 					 buffer_list);
2896 }
2897 
2898 STATIC int
2899 xlog_valid_rec_header(
2900 	struct xlog		*log,
2901 	struct xlog_rec_header	*rhead,
2902 	xfs_daddr_t		blkno,
2903 	int			bufsize)
2904 {
2905 	int			hlen;
2906 
2907 	if (XFS_IS_CORRUPT(log->l_mp,
2908 			   rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM)))
2909 		return -EFSCORRUPTED;
2910 	if (XFS_IS_CORRUPT(log->l_mp,
2911 			   (!rhead->h_version ||
2912 			   (be32_to_cpu(rhead->h_version) &
2913 			    (~XLOG_VERSION_OKBITS))))) {
2914 		xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
2915 			__func__, be32_to_cpu(rhead->h_version));
2916 		return -EFSCORRUPTED;
2917 	}
2918 
2919 	/*
2920 	 * LR body must have data (or it wouldn't have been written)
2921 	 * and h_len must not be greater than LR buffer size.
2922 	 */
2923 	hlen = be32_to_cpu(rhead->h_len);
2924 	if (XFS_IS_CORRUPT(log->l_mp, hlen <= 0 || hlen > bufsize))
2925 		return -EFSCORRUPTED;
2926 
2927 	if (XFS_IS_CORRUPT(log->l_mp,
2928 			   blkno > log->l_logBBsize || blkno > INT_MAX))
2929 		return -EFSCORRUPTED;
2930 	return 0;
2931 }
2932 
2933 /*
2934  * Read the log from tail to head and process the log records found.
2935  * Handle the two cases where the tail and head are in the same cycle
2936  * and where the active portion of the log wraps around the end of
2937  * the physical log separately.  The pass parameter is passed through
2938  * to the routines called to process the data and is not looked at
2939  * here.
2940  */
2941 STATIC int
2942 xlog_do_recovery_pass(
2943 	struct xlog		*log,
2944 	xfs_daddr_t		head_blk,
2945 	xfs_daddr_t		tail_blk,
2946 	int			pass,
2947 	xfs_daddr_t		*first_bad)	/* out: first bad log rec */
2948 {
2949 	xlog_rec_header_t	*rhead;
2950 	xfs_daddr_t		blk_no, rblk_no;
2951 	xfs_daddr_t		rhead_blk;
2952 	char			*offset;
2953 	char			*hbp, *dbp;
2954 	int			error = 0, h_size, h_len;
2955 	int			error2 = 0;
2956 	int			bblks, split_bblks;
2957 	int			hblks, split_hblks, wrapped_hblks;
2958 	int			i;
2959 	struct hlist_head	rhash[XLOG_RHASH_SIZE];
2960 	LIST_HEAD		(buffer_list);
2961 
2962 	ASSERT(head_blk != tail_blk);
2963 	blk_no = rhead_blk = tail_blk;
2964 
2965 	for (i = 0; i < XLOG_RHASH_SIZE; i++)
2966 		INIT_HLIST_HEAD(&rhash[i]);
2967 
2968 	/*
2969 	 * Read the header of the tail block and get the iclog buffer size from
2970 	 * h_size.  Use this to tell how many sectors make up the log header.
2971 	 */
2972 	if (xfs_has_logv2(log->l_mp)) {
2973 		/*
2974 		 * When using variable length iclogs, read first sector of
2975 		 * iclog header and extract the header size from it.  Get a
2976 		 * new hbp that is the correct size.
2977 		 */
2978 		hbp = xlog_alloc_buffer(log, 1);
2979 		if (!hbp)
2980 			return -ENOMEM;
2981 
2982 		error = xlog_bread(log, tail_blk, 1, hbp, &offset);
2983 		if (error)
2984 			goto bread_err1;
2985 
2986 		rhead = (xlog_rec_header_t *)offset;
2987 
2988 		/*
2989 		 * xfsprogs has a bug where record length is based on lsunit but
2990 		 * h_size (iclog size) is hardcoded to 32k. Now that we
2991 		 * unconditionally CRC verify the unmount record, this means the
2992 		 * log buffer can be too small for the record and cause an
2993 		 * overrun.
2994 		 *
2995 		 * Detect this condition here. Use lsunit for the buffer size as
2996 		 * long as this looks like the mkfs case. Otherwise, return an
2997 		 * error to avoid a buffer overrun.
2998 		 */
2999 		h_size = be32_to_cpu(rhead->h_size);
3000 		h_len = be32_to_cpu(rhead->h_len);
3001 		if (h_len > h_size && h_len <= log->l_mp->m_logbsize &&
3002 		    rhead->h_num_logops == cpu_to_be32(1)) {
3003 			xfs_warn(log->l_mp,
3004 		"invalid iclog size (%d bytes), using lsunit (%d bytes)",
3005 				 h_size, log->l_mp->m_logbsize);
3006 			h_size = log->l_mp->m_logbsize;
3007 		}
3008 
3009 		error = xlog_valid_rec_header(log, rhead, tail_blk, h_size);
3010 		if (error)
3011 			goto bread_err1;
3012 
3013 		hblks = xlog_logrec_hblks(log, rhead);
3014 		if (hblks != 1) {
3015 			kmem_free(hbp);
3016 			hbp = xlog_alloc_buffer(log, hblks);
3017 		}
3018 	} else {
3019 		ASSERT(log->l_sectBBsize == 1);
3020 		hblks = 1;
3021 		hbp = xlog_alloc_buffer(log, 1);
3022 		h_size = XLOG_BIG_RECORD_BSIZE;
3023 	}
3024 
3025 	if (!hbp)
3026 		return -ENOMEM;
3027 	dbp = xlog_alloc_buffer(log, BTOBB(h_size));
3028 	if (!dbp) {
3029 		kmem_free(hbp);
3030 		return -ENOMEM;
3031 	}
3032 
3033 	memset(rhash, 0, sizeof(rhash));
3034 	if (tail_blk > head_blk) {
3035 		/*
3036 		 * Perform recovery around the end of the physical log.
3037 		 * When the head is not on the same cycle number as the tail,
3038 		 * we can't do a sequential recovery.
3039 		 */
3040 		while (blk_no < log->l_logBBsize) {
3041 			/*
3042 			 * Check for header wrapping around physical end-of-log
3043 			 */
3044 			offset = hbp;
3045 			split_hblks = 0;
3046 			wrapped_hblks = 0;
3047 			if (blk_no + hblks <= log->l_logBBsize) {
3048 				/* Read header in one read */
3049 				error = xlog_bread(log, blk_no, hblks, hbp,
3050 						   &offset);
3051 				if (error)
3052 					goto bread_err2;
3053 			} else {
3054 				/* This LR is split across physical log end */
3055 				if (blk_no != log->l_logBBsize) {
3056 					/* some data before physical log end */
3057 					ASSERT(blk_no <= INT_MAX);
3058 					split_hblks = log->l_logBBsize - (int)blk_no;
3059 					ASSERT(split_hblks > 0);
3060 					error = xlog_bread(log, blk_no,
3061 							   split_hblks, hbp,
3062 							   &offset);
3063 					if (error)
3064 						goto bread_err2;
3065 				}
3066 
3067 				/*
3068 				 * Note: this black magic still works with
3069 				 * large sector sizes (non-512) only because:
3070 				 * - we increased the buffer size originally
3071 				 *   by 1 sector giving us enough extra space
3072 				 *   for the second read;
3073 				 * - the log start is guaranteed to be sector
3074 				 *   aligned;
3075 				 * - we read the log end (LR header start)
3076 				 *   _first_, then the log start (LR header end)
3077 				 *   - order is important.
3078 				 */
3079 				wrapped_hblks = hblks - split_hblks;
3080 				error = xlog_bread_noalign(log, 0,
3081 						wrapped_hblks,
3082 						offset + BBTOB(split_hblks));
3083 				if (error)
3084 					goto bread_err2;
3085 			}
3086 			rhead = (xlog_rec_header_t *)offset;
3087 			error = xlog_valid_rec_header(log, rhead,
3088 					split_hblks ? blk_no : 0, h_size);
3089 			if (error)
3090 				goto bread_err2;
3091 
3092 			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3093 			blk_no += hblks;
3094 
3095 			/*
3096 			 * Read the log record data in multiple reads if it
3097 			 * wraps around the end of the log. Note that if the
3098 			 * header already wrapped, blk_no could point past the
3099 			 * end of the log. The record data is contiguous in
3100 			 * that case.
3101 			 */
3102 			if (blk_no + bblks <= log->l_logBBsize ||
3103 			    blk_no >= log->l_logBBsize) {
3104 				rblk_no = xlog_wrap_logbno(log, blk_no);
3105 				error = xlog_bread(log, rblk_no, bblks, dbp,
3106 						   &offset);
3107 				if (error)
3108 					goto bread_err2;
3109 			} else {
3110 				/* This log record is split across the
3111 				 * physical end of log */
3112 				offset = dbp;
3113 				split_bblks = 0;
3114 				if (blk_no != log->l_logBBsize) {
3115 					/* some data is before the physical
3116 					 * end of log */
3117 					ASSERT(!wrapped_hblks);
3118 					ASSERT(blk_no <= INT_MAX);
3119 					split_bblks =
3120 						log->l_logBBsize - (int)blk_no;
3121 					ASSERT(split_bblks > 0);
3122 					error = xlog_bread(log, blk_no,
3123 							split_bblks, dbp,
3124 							&offset);
3125 					if (error)
3126 						goto bread_err2;
3127 				}
3128 
3129 				/*
3130 				 * Note: this black magic still works with
3131 				 * large sector sizes (non-512) only because:
3132 				 * - we increased the buffer size originally
3133 				 *   by 1 sector giving us enough extra space
3134 				 *   for the second read;
3135 				 * - the log start is guaranteed to be sector
3136 				 *   aligned;
3137 				 * - we read the log end (LR header start)
3138 				 *   _first_, then the log start (LR header end)
3139 				 *   - order is important.
3140 				 */
3141 				error = xlog_bread_noalign(log, 0,
3142 						bblks - split_bblks,
3143 						offset + BBTOB(split_bblks));
3144 				if (error)
3145 					goto bread_err2;
3146 			}
3147 
3148 			error = xlog_recover_process(log, rhash, rhead, offset,
3149 						     pass, &buffer_list);
3150 			if (error)
3151 				goto bread_err2;
3152 
3153 			blk_no += bblks;
3154 			rhead_blk = blk_no;
3155 		}
3156 
3157 		ASSERT(blk_no >= log->l_logBBsize);
3158 		blk_no -= log->l_logBBsize;
3159 		rhead_blk = blk_no;
3160 	}
3161 
3162 	/* read first part of physical log */
3163 	while (blk_no < head_blk) {
3164 		error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3165 		if (error)
3166 			goto bread_err2;
3167 
3168 		rhead = (xlog_rec_header_t *)offset;
3169 		error = xlog_valid_rec_header(log, rhead, blk_no, h_size);
3170 		if (error)
3171 			goto bread_err2;
3172 
3173 		/* blocks in data section */
3174 		bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3175 		error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3176 				   &offset);
3177 		if (error)
3178 			goto bread_err2;
3179 
3180 		error = xlog_recover_process(log, rhash, rhead, offset, pass,
3181 					     &buffer_list);
3182 		if (error)
3183 			goto bread_err2;
3184 
3185 		blk_no += bblks + hblks;
3186 		rhead_blk = blk_no;
3187 	}
3188 
3189  bread_err2:
3190 	kmem_free(dbp);
3191  bread_err1:
3192 	kmem_free(hbp);
3193 
3194 	/*
3195 	 * Submit buffers that have been added from the last record processed,
3196 	 * regardless of error status.
3197 	 */
3198 	if (!list_empty(&buffer_list))
3199 		error2 = xfs_buf_delwri_submit(&buffer_list);
3200 
3201 	if (error && first_bad)
3202 		*first_bad = rhead_blk;
3203 
3204 	/*
3205 	 * Transactions are freed at commit time but transactions without commit
3206 	 * records on disk are never committed. Free any that may be left in the
3207 	 * hash table.
3208 	 */
3209 	for (i = 0; i < XLOG_RHASH_SIZE; i++) {
3210 		struct hlist_node	*tmp;
3211 		struct xlog_recover	*trans;
3212 
3213 		hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
3214 			xlog_recover_free_trans(trans);
3215 	}
3216 
3217 	return error ? error : error2;
3218 }
3219 
3220 /*
3221  * Do the recovery of the log.  We actually do this in two phases.
3222  * The two passes are necessary in order to implement the function
3223  * of cancelling a record written into the log.  The first pass
3224  * determines those things which have been cancelled, and the
3225  * second pass replays log items normally except for those which
3226  * have been cancelled.  The handling of the replay and cancellations
3227  * takes place in the log item type specific routines.
3228  *
3229  * The table of items which have cancel records in the log is allocated
3230  * and freed at this level, since only here do we know when all of
3231  * the log recovery has been completed.
3232  */
3233 STATIC int
3234 xlog_do_log_recovery(
3235 	struct xlog	*log,
3236 	xfs_daddr_t	head_blk,
3237 	xfs_daddr_t	tail_blk)
3238 {
3239 	int		error;
3240 
3241 	ASSERT(head_blk != tail_blk);
3242 
3243 	/*
3244 	 * First do a pass to find all of the cancelled buf log items.
3245 	 * Store them in the buf_cancel_table for use in the second pass.
3246 	 */
3247 	error = xlog_alloc_buf_cancel_table(log);
3248 	if (error)
3249 		return error;
3250 
3251 	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3252 				      XLOG_RECOVER_PASS1, NULL);
3253 	if (error != 0)
3254 		goto out_cancel;
3255 
3256 	/*
3257 	 * Then do a second pass to actually recover the items in the log.
3258 	 * When it is complete free the table of buf cancel items.
3259 	 */
3260 	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3261 				      XLOG_RECOVER_PASS2, NULL);
3262 	if (!error)
3263 		xlog_check_buf_cancel_table(log);
3264 out_cancel:
3265 	xlog_free_buf_cancel_table(log);
3266 	return error;
3267 }
3268 
3269 /*
3270  * Do the actual recovery
3271  */
3272 STATIC int
3273 xlog_do_recover(
3274 	struct xlog		*log,
3275 	xfs_daddr_t		head_blk,
3276 	xfs_daddr_t		tail_blk)
3277 {
3278 	struct xfs_mount	*mp = log->l_mp;
3279 	struct xfs_buf		*bp = mp->m_sb_bp;
3280 	struct xfs_sb		*sbp = &mp->m_sb;
3281 	int			error;
3282 
3283 	trace_xfs_log_recover(log, head_blk, tail_blk);
3284 
3285 	/*
3286 	 * First replay the images in the log.
3287 	 */
3288 	error = xlog_do_log_recovery(log, head_blk, tail_blk);
3289 	if (error)
3290 		return error;
3291 
3292 	if (xlog_is_shutdown(log))
3293 		return -EIO;
3294 
3295 	/*
3296 	 * We now update the tail_lsn since much of the recovery has completed
3297 	 * and there may be space available to use.  If there were no extent
3298 	 * or iunlinks, we can free up the entire log and set the tail_lsn to
3299 	 * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3300 	 * lsn of the last known good LR on disk.  If there are extent frees
3301 	 * or iunlinks they will have some entries in the AIL; so we look at
3302 	 * the AIL to determine how to set the tail_lsn.
3303 	 */
3304 	xlog_assign_tail_lsn(mp);
3305 
3306 	/*
3307 	 * Now that we've finished replaying all buffer and inode updates,
3308 	 * re-read the superblock and reverify it.
3309 	 */
3310 	xfs_buf_lock(bp);
3311 	xfs_buf_hold(bp);
3312 	error = _xfs_buf_read(bp, XBF_READ);
3313 	if (error) {
3314 		if (!xlog_is_shutdown(log)) {
3315 			xfs_buf_ioerror_alert(bp, __this_address);
3316 			ASSERT(0);
3317 		}
3318 		xfs_buf_relse(bp);
3319 		return error;
3320 	}
3321 
3322 	/* Convert superblock from on-disk format */
3323 	xfs_sb_from_disk(sbp, bp->b_addr);
3324 	xfs_buf_relse(bp);
3325 
3326 	/* re-initialise in-core superblock and geometry structures */
3327 	mp->m_features |= xfs_sb_version_to_features(sbp);
3328 	xfs_reinit_percpu_counters(mp);
3329 	error = xfs_initialize_perag(mp, sbp->sb_agcount, sbp->sb_dblocks,
3330 			&mp->m_maxagi);
3331 	if (error) {
3332 		xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
3333 		return error;
3334 	}
3335 	mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
3336 
3337 	/* Normal transactions can now occur */
3338 	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
3339 	return 0;
3340 }
3341 
3342 /*
3343  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3344  *
3345  * Return error or zero.
3346  */
3347 int
3348 xlog_recover(
3349 	struct xlog	*log)
3350 {
3351 	xfs_daddr_t	head_blk, tail_blk;
3352 	int		error;
3353 
3354 	/* find the tail of the log */
3355 	error = xlog_find_tail(log, &head_blk, &tail_blk);
3356 	if (error)
3357 		return error;
3358 
3359 	/*
3360 	 * The superblock was read before the log was available and thus the LSN
3361 	 * could not be verified. Check the superblock LSN against the current
3362 	 * LSN now that it's known.
3363 	 */
3364 	if (xfs_has_crc(log->l_mp) &&
3365 	    !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
3366 		return -EINVAL;
3367 
3368 	if (tail_blk != head_blk) {
3369 		/* There used to be a comment here:
3370 		 *
3371 		 * disallow recovery on read-only mounts.  note -- mount
3372 		 * checks for ENOSPC and turns it into an intelligent
3373 		 * error message.
3374 		 * ...but this is no longer true.  Now, unless you specify
3375 		 * NORECOVERY (in which case this function would never be
3376 		 * called), we just go ahead and recover.  We do this all
3377 		 * under the vfs layer, so we can get away with it unless
3378 		 * the device itself is read-only, in which case we fail.
3379 		 */
3380 		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3381 			return error;
3382 		}
3383 
3384 		/*
3385 		 * Version 5 superblock log feature mask validation. We know the
3386 		 * log is dirty so check if there are any unknown log features
3387 		 * in what we need to recover. If there are unknown features
3388 		 * (e.g. unsupported transactions, then simply reject the
3389 		 * attempt at recovery before touching anything.
3390 		 */
3391 		if (xfs_sb_is_v5(&log->l_mp->m_sb) &&
3392 		    xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
3393 					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
3394 			xfs_warn(log->l_mp,
3395 "Superblock has unknown incompatible log features (0x%x) enabled.",
3396 				(log->l_mp->m_sb.sb_features_log_incompat &
3397 					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
3398 			xfs_warn(log->l_mp,
3399 "The log can not be fully and/or safely recovered by this kernel.");
3400 			xfs_warn(log->l_mp,
3401 "Please recover the log on a kernel that supports the unknown features.");
3402 			return -EINVAL;
3403 		}
3404 
3405 		/*
3406 		 * Delay log recovery if the debug hook is set. This is debug
3407 		 * instrumentation to coordinate simulation of I/O failures with
3408 		 * log recovery.
3409 		 */
3410 		if (xfs_globals.log_recovery_delay) {
3411 			xfs_notice(log->l_mp,
3412 				"Delaying log recovery for %d seconds.",
3413 				xfs_globals.log_recovery_delay);
3414 			msleep(xfs_globals.log_recovery_delay * 1000);
3415 		}
3416 
3417 		xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
3418 				log->l_mp->m_logname ? log->l_mp->m_logname
3419 						     : "internal");
3420 
3421 		error = xlog_do_recover(log, head_blk, tail_blk);
3422 		set_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
3423 	}
3424 	return error;
3425 }
3426 
3427 /*
3428  * In the first part of recovery we replay inodes and buffers and build up the
3429  * list of intents which need to be processed. Here we process the intents and
3430  * clean up the on disk unlinked inode lists. This is separated from the first
3431  * part of recovery so that the root and real-time bitmap inodes can be read in
3432  * from disk in between the two stages.  This is necessary so that we can free
3433  * space in the real-time portion of the file system.
3434  */
3435 int
3436 xlog_recover_finish(
3437 	struct xlog	*log)
3438 {
3439 	int	error;
3440 
3441 	error = xlog_recover_process_intents(log);
3442 	if (error) {
3443 		/*
3444 		 * Cancel all the unprocessed intent items now so that we don't
3445 		 * leave them pinned in the AIL.  This can cause the AIL to
3446 		 * livelock on the pinned item if anyone tries to push the AIL
3447 		 * (inode reclaim does this) before we get around to
3448 		 * xfs_log_mount_cancel.
3449 		 */
3450 		xlog_recover_cancel_intents(log);
3451 		xfs_alert(log->l_mp, "Failed to recover intents");
3452 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
3453 		return error;
3454 	}
3455 
3456 	/*
3457 	 * Sync the log to get all the intents out of the AIL.  This isn't
3458 	 * absolutely necessary, but it helps in case the unlink transactions
3459 	 * would have problems pushing the intents out of the way.
3460 	 */
3461 	xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3462 
3463 	/*
3464 	 * Now that we've recovered the log and all the intents, we can clear
3465 	 * the log incompat feature bits in the superblock because there's no
3466 	 * longer anything to protect.  We rely on the AIL push to write out the
3467 	 * updated superblock after everything else.
3468 	 */
3469 	if (xfs_clear_incompat_log_features(log->l_mp)) {
3470 		error = xfs_sync_sb(log->l_mp, false);
3471 		if (error < 0) {
3472 			xfs_alert(log->l_mp,
3473 	"Failed to clear log incompat features on recovery");
3474 			return error;
3475 		}
3476 	}
3477 
3478 	xlog_recover_process_iunlinks(log);
3479 
3480 	/*
3481 	 * Recover any CoW staging blocks that are still referenced by the
3482 	 * ondisk refcount metadata.  During mount there cannot be any live
3483 	 * staging extents as we have not permitted any user modifications.
3484 	 * Therefore, it is safe to free them all right now, even on a
3485 	 * read-only mount.
3486 	 */
3487 	error = xfs_reflink_recover_cow(log->l_mp);
3488 	if (error) {
3489 		xfs_alert(log->l_mp,
3490 	"Failed to recover leftover CoW staging extents, err %d.",
3491 				error);
3492 		/*
3493 		 * If we get an error here, make sure the log is shut down
3494 		 * but return zero so that any log items committed since the
3495 		 * end of intents processing can be pushed through the CIL
3496 		 * and AIL.
3497 		 */
3498 		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
3499 	}
3500 
3501 	return 0;
3502 }
3503 
3504 void
3505 xlog_recover_cancel(
3506 	struct xlog	*log)
3507 {
3508 	if (xlog_recovery_needed(log))
3509 		xlog_recover_cancel_intents(log);
3510 }
3511 
3512