xfs_log_recover.c revision d6837c1a
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
31
32#define BLK_AVG(blk1, blk2)	((blk1+blk2) >> 1)
33
34STATIC int
35xlog_find_zeroed(
36	struct xlog	*,
37	xfs_daddr_t	*);
38STATIC int
39xlog_clear_stale_blocks(
40	struct xlog	*,
41	xfs_lsn_t);
42#if defined(DEBUG)
43STATIC void
44xlog_recover_check_summary(
45	struct xlog *);
46#else
47#define	xlog_recover_check_summary(log)
48#endif
49STATIC int
50xlog_do_recovery_pass(
51        struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
52
53/*
54 * Sector aligned buffer routines for buffer create/read/write/access
55 */
56
57/*
58 * Verify the log-relative block number and length in basic blocks are valid for
59 * an operation involving the given XFS log buffer. Returns true if the fields
60 * are valid, false otherwise.
61 */
62static inline bool
63xlog_verify_bno(
64	struct xlog	*log,
65	xfs_daddr_t	blk_no,
66	int		bbcount)
67{
68	if (blk_no < 0 || blk_no >= log->l_logBBsize)
69		return false;
70	if (bbcount <= 0 || (blk_no + bbcount) > log->l_logBBsize)
71		return false;
72	return true;
73}
74
75/*
76 * Allocate a buffer to hold log data.  The buffer needs to be able to map to
77 * a range of nbblks basic blocks at any valid offset within the log.
78 */
79static char *
80xlog_alloc_buffer(
81	struct xlog	*log,
82	int		nbblks)
83{
84	/*
85	 * Pass log block 0 since we don't have an addr yet, buffer will be
86	 * verified on read.
87	 */
88	if (XFS_IS_CORRUPT(log->l_mp, !xlog_verify_bno(log, 0, nbblks))) {
89		xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
90			nbblks);
91		return NULL;
92	}
93
94	/*
95	 * We do log I/O in units of log sectors (a power-of-2 multiple of the
96	 * basic block size), so we round up the requested size to accommodate
97	 * the basic blocks required for complete log sectors.
98	 *
99	 * In addition, the buffer may be used for a non-sector-aligned block
100	 * offset, in which case an I/O of the requested size could extend
101	 * beyond the end of the buffer.  If the requested size is only 1 basic
102	 * block it will never straddle a sector boundary, so this won't be an
103	 * issue.  Nor will this be a problem if the log I/O is done in basic
104	 * blocks (sector size 1).  But otherwise we extend the buffer by one
105	 * extra log sector to ensure there's space to accommodate this
106	 * possibility.
107	 */
108	if (nbblks > 1 && log->l_sectBBsize > 1)
109		nbblks += log->l_sectBBsize;
110	nbblks = round_up(nbblks, log->l_sectBBsize);
111	return kvzalloc(BBTOB(nbblks), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
112}
113
114/*
115 * Return the address of the start of the given block number's data
116 * in a log buffer.  The buffer covers a log sector-aligned region.
117 */
118static inline unsigned int
119xlog_align(
120	struct xlog	*log,
121	xfs_daddr_t	blk_no)
122{
123	return BBTOB(blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1));
124}
125
126static int
127xlog_do_io(
128	struct xlog		*log,
129	xfs_daddr_t		blk_no,
130	unsigned int		nbblks,
131	char			*data,
132	unsigned int		op)
133{
134	int			error;
135
136	if (XFS_IS_CORRUPT(log->l_mp, !xlog_verify_bno(log, blk_no, nbblks))) {
137		xfs_warn(log->l_mp,
138			 "Invalid log block/length (0x%llx, 0x%x) for buffer",
139			 blk_no, nbblks);
140		return -EFSCORRUPTED;
141	}
142
143	blk_no = round_down(blk_no, log->l_sectBBsize);
144	nbblks = round_up(nbblks, log->l_sectBBsize);
145	ASSERT(nbblks > 0);
146
147	error = xfs_rw_bdev(log->l_targ->bt_bdev, log->l_logBBstart + blk_no,
148			BBTOB(nbblks), data, op);
149	if (error && !xlog_is_shutdown(log)) {
150		xfs_alert(log->l_mp,
151			  "log recovery %s I/O error at daddr 0x%llx len %d error %d",
152			  op == REQ_OP_WRITE ? "write" : "read",
153			  blk_no, nbblks, error);
154	}
155	return error;
156}
157
158STATIC int
159xlog_bread_noalign(
160	struct xlog	*log,
161	xfs_daddr_t	blk_no,
162	int		nbblks,
163	char		*data)
164{
165	return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
166}
167
168STATIC int
169xlog_bread(
170	struct xlog	*log,
171	xfs_daddr_t	blk_no,
172	int		nbblks,
173	char		*data,
174	char		**offset)
175{
176	int		error;
177
178	error = xlog_do_io(log, blk_no, nbblks, data, REQ_OP_READ);
179	if (!error)
180		*offset = data + xlog_align(log, blk_no);
181	return error;
182}
183
184STATIC int
185xlog_bwrite(
186	struct xlog	*log,
187	xfs_daddr_t	blk_no,
188	int		nbblks,
189	char		*data)
190{
191	return xlog_do_io(log, blk_no, nbblks, data, REQ_OP_WRITE);
192}
193
194#ifdef DEBUG
195/*
196 * dump debug superblock and log record information
197 */
198STATIC void
199xlog_header_check_dump(
200	xfs_mount_t		*mp,
201	xlog_rec_header_t	*head)
202{
203	xfs_debug(mp, "%s:  SB : uuid = %pU, fmt = %d",
204		__func__, &mp->m_sb.sb_uuid, XLOG_FMT);
205	xfs_debug(mp, "    log : uuid = %pU, fmt = %d",
206		&head->h_fs_uuid, be32_to_cpu(head->h_fmt));
207}
208#else
209#define xlog_header_check_dump(mp, head)
210#endif
211
212/*
213 * check log record header for recovery
214 */
215STATIC int
216xlog_header_check_recover(
217	xfs_mount_t		*mp,
218	xlog_rec_header_t	*head)
219{
220	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
221
222	/*
223	 * IRIX doesn't write the h_fmt field and leaves it zeroed
224	 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
225	 * a dirty log created in IRIX.
226	 */
227	if (XFS_IS_CORRUPT(mp, head->h_fmt != cpu_to_be32(XLOG_FMT))) {
228		xfs_warn(mp,
229	"dirty log written in incompatible format - can't recover");
230		xlog_header_check_dump(mp, head);
231		return -EFSCORRUPTED;
232	}
233	if (XFS_IS_CORRUPT(mp, !uuid_equal(&mp->m_sb.sb_uuid,
234					   &head->h_fs_uuid))) {
235		xfs_warn(mp,
236	"dirty log entry has mismatched uuid - can't recover");
237		xlog_header_check_dump(mp, head);
238		return -EFSCORRUPTED;
239	}
240	return 0;
241}
242
243/*
244 * read the head block of the log and check the header
245 */
246STATIC int
247xlog_header_check_mount(
248	xfs_mount_t		*mp,
249	xlog_rec_header_t	*head)
250{
251	ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
252
253	if (uuid_is_null(&head->h_fs_uuid)) {
254		/*
255		 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
256		 * h_fs_uuid is null, we assume this log was last mounted
257		 * by IRIX and continue.
258		 */
259		xfs_warn(mp, "null uuid in log - IRIX style log");
260	} else if (XFS_IS_CORRUPT(mp, !uuid_equal(&mp->m_sb.sb_uuid,
261						  &head->h_fs_uuid))) {
262		xfs_warn(mp, "log has mismatched uuid - can't recover");
263		xlog_header_check_dump(mp, head);
264		return -EFSCORRUPTED;
265	}
266	return 0;
267}
268
269/*
270 * This routine finds (to an approximation) the first block in the physical
271 * log which contains the given cycle.  It uses a binary search algorithm.
272 * Note that the algorithm can not be perfect because the disk will not
273 * necessarily be perfect.
274 */
275STATIC int
276xlog_find_cycle_start(
277	struct xlog	*log,
278	char		*buffer,
279	xfs_daddr_t	first_blk,
280	xfs_daddr_t	*last_blk,
281	uint		cycle)
282{
283	char		*offset;
284	xfs_daddr_t	mid_blk;
285	xfs_daddr_t	end_blk;
286	uint		mid_cycle;
287	int		error;
288
289	end_blk = *last_blk;
290	mid_blk = BLK_AVG(first_blk, end_blk);
291	while (mid_blk != first_blk && mid_blk != end_blk) {
292		error = xlog_bread(log, mid_blk, 1, buffer, &offset);
293		if (error)
294			return error;
295		mid_cycle = xlog_get_cycle(offset);
296		if (mid_cycle == cycle)
297			end_blk = mid_blk;   /* last_half_cycle == mid_cycle */
298		else
299			first_blk = mid_blk; /* first_half_cycle == mid_cycle */
300		mid_blk = BLK_AVG(first_blk, end_blk);
301	}
302	ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
303	       (mid_blk == end_blk && mid_blk-1 == first_blk));
304
305	*last_blk = end_blk;
306
307	return 0;
308}
309
310/*
311 * Check that a range of blocks does not contain stop_on_cycle_no.
312 * Fill in *new_blk with the block offset where such a block is
313 * found, or with -1 (an invalid block number) if there is no such
314 * block in the range.  The scan needs to occur from front to back
315 * and the pointer into the region must be updated since a later
316 * routine will need to perform another test.
317 */
318STATIC int
319xlog_find_verify_cycle(
320	struct xlog	*log,
321	xfs_daddr_t	start_blk,
322	int		nbblks,
323	uint		stop_on_cycle_no,
324	xfs_daddr_t	*new_blk)
325{
326	xfs_daddr_t	i, j;
327	uint		cycle;
328	char		*buffer;
329	xfs_daddr_t	bufblks;
330	char		*buf = NULL;
331	int		error = 0;
332
333	/*
334	 * Greedily allocate a buffer big enough to handle the full
335	 * range of basic blocks we'll be examining.  If that fails,
336	 * try a smaller size.  We need to be able to read at least
337	 * a log sector, or we're out of luck.
338	 */
339	bufblks = 1 << ffs(nbblks);
340	while (bufblks > log->l_logBBsize)
341		bufblks >>= 1;
342	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
343		bufblks >>= 1;
344		if (bufblks < log->l_sectBBsize)
345			return -ENOMEM;
346	}
347
348	for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
349		int	bcount;
350
351		bcount = min(bufblks, (start_blk + nbblks - i));
352
353		error = xlog_bread(log, i, bcount, buffer, &buf);
354		if (error)
355			goto out;
356
357		for (j = 0; j < bcount; j++) {
358			cycle = xlog_get_cycle(buf);
359			if (cycle == stop_on_cycle_no) {
360				*new_blk = i+j;
361				goto out;
362			}
363
364			buf += BBSIZE;
365		}
366	}
367
368	*new_blk = -1;
369
370out:
371	kmem_free(buffer);
372	return error;
373}
374
375static inline int
376xlog_logrec_hblks(struct xlog *log, struct xlog_rec_header *rh)
377{
378	if (xfs_has_logv2(log->l_mp)) {
379		int	h_size = be32_to_cpu(rh->h_size);
380
381		if ((be32_to_cpu(rh->h_version) & XLOG_VERSION_2) &&
382		    h_size > XLOG_HEADER_CYCLE_SIZE)
383			return DIV_ROUND_UP(h_size, XLOG_HEADER_CYCLE_SIZE);
384	}
385	return 1;
386}
387
388/*
389 * Potentially backup over partial log record write.
390 *
391 * In the typical case, last_blk is the number of the block directly after
392 * a good log record.  Therefore, we subtract one to get the block number
393 * of the last block in the given buffer.  extra_bblks contains the number
394 * of blocks we would have read on a previous read.  This happens when the
395 * last log record is split over the end of the physical log.
396 *
397 * extra_bblks is the number of blocks potentially verified on a previous
398 * call to this routine.
399 */
400STATIC int
401xlog_find_verify_log_record(
402	struct xlog		*log,
403	xfs_daddr_t		start_blk,
404	xfs_daddr_t		*last_blk,
405	int			extra_bblks)
406{
407	xfs_daddr_t		i;
408	char			*buffer;
409	char			*offset = NULL;
410	xlog_rec_header_t	*head = NULL;
411	int			error = 0;
412	int			smallmem = 0;
413	int			num_blks = *last_blk - start_blk;
414	int			xhdrs;
415
416	ASSERT(start_blk != 0 || *last_blk != start_blk);
417
418	buffer = xlog_alloc_buffer(log, num_blks);
419	if (!buffer) {
420		buffer = xlog_alloc_buffer(log, 1);
421		if (!buffer)
422			return -ENOMEM;
423		smallmem = 1;
424	} else {
425		error = xlog_bread(log, start_blk, num_blks, buffer, &offset);
426		if (error)
427			goto out;
428		offset += ((num_blks - 1) << BBSHIFT);
429	}
430
431	for (i = (*last_blk) - 1; i >= 0; i--) {
432		if (i < start_blk) {
433			/* valid log record not found */
434			xfs_warn(log->l_mp,
435		"Log inconsistent (didn't find previous header)");
436			ASSERT(0);
437			error = -EFSCORRUPTED;
438			goto out;
439		}
440
441		if (smallmem) {
442			error = xlog_bread(log, i, 1, buffer, &offset);
443			if (error)
444				goto out;
445		}
446
447		head = (xlog_rec_header_t *)offset;
448
449		if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
450			break;
451
452		if (!smallmem)
453			offset -= BBSIZE;
454	}
455
456	/*
457	 * We hit the beginning of the physical log & still no header.  Return
458	 * to caller.  If caller can handle a return of -1, then this routine
459	 * will be called again for the end of the physical log.
460	 */
461	if (i == -1) {
462		error = 1;
463		goto out;
464	}
465
466	/*
467	 * We have the final block of the good log (the first block
468	 * of the log record _before_ the head. So we check the uuid.
469	 */
470	if ((error = xlog_header_check_mount(log->l_mp, head)))
471		goto out;
472
473	/*
474	 * We may have found a log record header before we expected one.
475	 * last_blk will be the 1st block # with a given cycle #.  We may end
476	 * up reading an entire log record.  In this case, we don't want to
477	 * reset last_blk.  Only when last_blk points in the middle of a log
478	 * record do we update last_blk.
479	 */
480	xhdrs = xlog_logrec_hblks(log, head);
481
482	if (*last_blk - i + extra_bblks !=
483	    BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
484		*last_blk = i;
485
486out:
487	kmem_free(buffer);
488	return error;
489}
490
491/*
492 * Head is defined to be the point of the log where the next log write
493 * could go.  This means that incomplete LR writes at the end are
494 * eliminated when calculating the head.  We aren't guaranteed that previous
495 * LR have complete transactions.  We only know that a cycle number of
496 * current cycle number -1 won't be present in the log if we start writing
497 * from our current block number.
498 *
499 * last_blk contains the block number of the first block with a given
500 * cycle number.
501 *
502 * Return: zero if normal, non-zero if error.
503 */
504STATIC int
505xlog_find_head(
506	struct xlog	*log,
507	xfs_daddr_t	*return_head_blk)
508{
509	char		*buffer;
510	char		*offset;
511	xfs_daddr_t	new_blk, first_blk, start_blk, last_blk, head_blk;
512	int		num_scan_bblks;
513	uint		first_half_cycle, last_half_cycle;
514	uint		stop_on_cycle;
515	int		error, log_bbnum = log->l_logBBsize;
516
517	/* Is the end of the log device zeroed? */
518	error = xlog_find_zeroed(log, &first_blk);
519	if (error < 0) {
520		xfs_warn(log->l_mp, "empty log check failed");
521		return error;
522	}
523	if (error == 1) {
524		*return_head_blk = first_blk;
525
526		/* Is the whole lot zeroed? */
527		if (!first_blk) {
528			/* Linux XFS shouldn't generate totally zeroed logs -
529			 * mkfs etc write a dummy unmount record to a fresh
530			 * log so we can store the uuid in there
531			 */
532			xfs_warn(log->l_mp, "totally zeroed log");
533		}
534
535		return 0;
536	}
537
538	first_blk = 0;			/* get cycle # of 1st block */
539	buffer = xlog_alloc_buffer(log, 1);
540	if (!buffer)
541		return -ENOMEM;
542
543	error = xlog_bread(log, 0, 1, buffer, &offset);
544	if (error)
545		goto out_free_buffer;
546
547	first_half_cycle = xlog_get_cycle(offset);
548
549	last_blk = head_blk = log_bbnum - 1;	/* get cycle # of last block */
550	error = xlog_bread(log, last_blk, 1, buffer, &offset);
551	if (error)
552		goto out_free_buffer;
553
554	last_half_cycle = xlog_get_cycle(offset);
555	ASSERT(last_half_cycle != 0);
556
557	/*
558	 * If the 1st half cycle number is equal to the last half cycle number,
559	 * then the entire log is stamped with the same cycle number.  In this
560	 * case, head_blk can't be set to zero (which makes sense).  The below
561	 * math doesn't work out properly with head_blk equal to zero.  Instead,
562	 * we set it to log_bbnum which is an invalid block number, but this
563	 * value makes the math correct.  If head_blk doesn't changed through
564	 * all the tests below, *head_blk is set to zero at the very end rather
565	 * than log_bbnum.  In a sense, log_bbnum and zero are the same block
566	 * in a circular file.
567	 */
568	if (first_half_cycle == last_half_cycle) {
569		/*
570		 * In this case we believe that the entire log should have
571		 * cycle number last_half_cycle.  We need to scan backwards
572		 * from the end verifying that there are no holes still
573		 * containing last_half_cycle - 1.  If we find such a hole,
574		 * then the start of that hole will be the new head.  The
575		 * simple case looks like
576		 *        x | x ... | x - 1 | x
577		 * Another case that fits this picture would be
578		 *        x | x + 1 | x ... | x
579		 * In this case the head really is somewhere at the end of the
580		 * log, as one of the latest writes at the beginning was
581		 * incomplete.
582		 * One more case is
583		 *        x | x + 1 | x ... | x - 1 | x
584		 * This is really the combination of the above two cases, and
585		 * the head has to end up at the start of the x-1 hole at the
586		 * end of the log.
587		 *
588		 * In the 256k log case, we will read from the beginning to the
589		 * end of the log and search for cycle numbers equal to x-1.
590		 * We don't worry about the x+1 blocks that we encounter,
591		 * because we know that they cannot be the head since the log
592		 * started with x.
593		 */
594		head_blk = log_bbnum;
595		stop_on_cycle = last_half_cycle - 1;
596	} else {
597		/*
598		 * In this case we want to find the first block with cycle
599		 * number matching last_half_cycle.  We expect the log to be
600		 * some variation on
601		 *        x + 1 ... | x ... | x
602		 * The first block with cycle number x (last_half_cycle) will
603		 * be where the new head belongs.  First we do a binary search
604		 * for the first occurrence of last_half_cycle.  The binary
605		 * search may not be totally accurate, so then we scan back
606		 * from there looking for occurrences of last_half_cycle before
607		 * us.  If that backwards scan wraps around the beginning of
608		 * the log, then we look for occurrences of last_half_cycle - 1
609		 * at the end of the log.  The cases we're looking for look
610		 * like
611		 *                               v binary search stopped here
612		 *        x + 1 ... | x | x + 1 | x ... | x
613		 *                   ^ but we want to locate this spot
614		 * or
615		 *        <---------> less than scan distance
616		 *        x + 1 ... | x ... | x - 1 | x
617		 *                           ^ we want to locate this spot
618		 */
619		stop_on_cycle = last_half_cycle;
620		error = xlog_find_cycle_start(log, buffer, first_blk, &head_blk,
621				last_half_cycle);
622		if (error)
623			goto out_free_buffer;
624	}
625
626	/*
627	 * Now validate the answer.  Scan back some number of maximum possible
628	 * blocks and make sure each one has the expected cycle number.  The
629	 * maximum is determined by the total possible amount of buffering
630	 * in the in-core log.  The following number can be made tighter if
631	 * we actually look at the block size of the filesystem.
632	 */
633	num_scan_bblks = min_t(int, log_bbnum, XLOG_TOTAL_REC_SHIFT(log));
634	if (head_blk >= num_scan_bblks) {
635		/*
636		 * We are guaranteed that the entire check can be performed
637		 * in one buffer.
638		 */
639		start_blk = head_blk - num_scan_bblks;
640		if ((error = xlog_find_verify_cycle(log,
641						start_blk, num_scan_bblks,
642						stop_on_cycle, &new_blk)))
643			goto out_free_buffer;
644		if (new_blk != -1)
645			head_blk = new_blk;
646	} else {		/* need to read 2 parts of log */
647		/*
648		 * We are going to scan backwards in the log in two parts.
649		 * First we scan the physical end of the log.  In this part
650		 * of the log, we are looking for blocks with cycle number
651		 * last_half_cycle - 1.
652		 * If we find one, then we know that the log starts there, as
653		 * we've found a hole that didn't get written in going around
654		 * the end of the physical log.  The simple case for this is
655		 *        x + 1 ... | x ... | x - 1 | x
656		 *        <---------> less than scan distance
657		 * If all of the blocks at the end of the log have cycle number
658		 * last_half_cycle, then we check the blocks at the start of
659		 * the log looking for occurrences of last_half_cycle.  If we
660		 * find one, then our current estimate for the location of the
661		 * first occurrence of last_half_cycle is wrong and we move
662		 * back to the hole we've found.  This case looks like
663		 *        x + 1 ... | x | x + 1 | x ...
664		 *                               ^ binary search stopped here
665		 * Another case we need to handle that only occurs in 256k
666		 * logs is
667		 *        x + 1 ... | x ... | x+1 | x ...
668		 *                   ^ binary search stops here
669		 * In a 256k log, the scan at the end of the log will see the
670		 * x + 1 blocks.  We need to skip past those since that is
671		 * certainly not the head of the log.  By searching for
672		 * last_half_cycle-1 we accomplish that.
673		 */
674		ASSERT(head_blk <= INT_MAX &&
675			(xfs_daddr_t) num_scan_bblks >= head_blk);
676		start_blk = log_bbnum - (num_scan_bblks - head_blk);
677		if ((error = xlog_find_verify_cycle(log, start_blk,
678					num_scan_bblks - (int)head_blk,
679					(stop_on_cycle - 1), &new_blk)))
680			goto out_free_buffer;
681		if (new_blk != -1) {
682			head_blk = new_blk;
683			goto validate_head;
684		}
685
686		/*
687		 * Scan beginning of log now.  The last part of the physical
688		 * log is good.  This scan needs to verify that it doesn't find
689		 * the last_half_cycle.
690		 */
691		start_blk = 0;
692		ASSERT(head_blk <= INT_MAX);
693		if ((error = xlog_find_verify_cycle(log,
694					start_blk, (int)head_blk,
695					stop_on_cycle, &new_blk)))
696			goto out_free_buffer;
697		if (new_blk != -1)
698			head_blk = new_blk;
699	}
700
701validate_head:
702	/*
703	 * Now we need to make sure head_blk is not pointing to a block in
704	 * the middle of a log record.
705	 */
706	num_scan_bblks = XLOG_REC_SHIFT(log);
707	if (head_blk >= num_scan_bblks) {
708		start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
709
710		/* start ptr at last block ptr before head_blk */
711		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
712		if (error == 1)
713			error = -EIO;
714		if (error)
715			goto out_free_buffer;
716	} else {
717		start_blk = 0;
718		ASSERT(head_blk <= INT_MAX);
719		error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
720		if (error < 0)
721			goto out_free_buffer;
722		if (error == 1) {
723			/* We hit the beginning of the log during our search */
724			start_blk = log_bbnum - (num_scan_bblks - head_blk);
725			new_blk = log_bbnum;
726			ASSERT(start_blk <= INT_MAX &&
727				(xfs_daddr_t) log_bbnum-start_blk >= 0);
728			ASSERT(head_blk <= INT_MAX);
729			error = xlog_find_verify_log_record(log, start_blk,
730							&new_blk, (int)head_blk);
731			if (error == 1)
732				error = -EIO;
733			if (error)
734				goto out_free_buffer;
735			if (new_blk != log_bbnum)
736				head_blk = new_blk;
737		} else if (error)
738			goto out_free_buffer;
739	}
740
741	kmem_free(buffer);
742	if (head_blk == log_bbnum)
743		*return_head_blk = 0;
744	else
745		*return_head_blk = head_blk;
746	/*
747	 * When returning here, we have a good block number.  Bad block
748	 * means that during a previous crash, we didn't have a clean break
749	 * from cycle number N to cycle number N-1.  In this case, we need
750	 * to find the first block with cycle number N-1.
751	 */
752	return 0;
753
754out_free_buffer:
755	kmem_free(buffer);
756	if (error)
757		xfs_warn(log->l_mp, "failed to find log head");
758	return error;
759}
760
761/*
762 * Seek backwards in the log for log record headers.
763 *
764 * Given a starting log block, walk backwards until we find the provided number
765 * of records or hit the provided tail block. The return value is the number of
766 * records encountered or a negative error code. The log block and buffer
767 * pointer of the last record seen are returned in rblk and rhead respectively.
768 */
769STATIC int
770xlog_rseek_logrec_hdr(
771	struct xlog		*log,
772	xfs_daddr_t		head_blk,
773	xfs_daddr_t		tail_blk,
774	int			count,
775	char			*buffer,
776	xfs_daddr_t		*rblk,
777	struct xlog_rec_header	**rhead,
778	bool			*wrapped)
779{
780	int			i;
781	int			error;
782	int			found = 0;
783	char			*offset = NULL;
784	xfs_daddr_t		end_blk;
785
786	*wrapped = false;
787
788	/*
789	 * Walk backwards from the head block until we hit the tail or the first
790	 * block in the log.
791	 */
792	end_blk = head_blk > tail_blk ? tail_blk : 0;
793	for (i = (int) head_blk - 1; i >= end_blk; i--) {
794		error = xlog_bread(log, i, 1, buffer, &offset);
795		if (error)
796			goto out_error;
797
798		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
799			*rblk = i;
800			*rhead = (struct xlog_rec_header *) offset;
801			if (++found == count)
802				break;
803		}
804	}
805
806	/*
807	 * If we haven't hit the tail block or the log record header count,
808	 * start looking again from the end of the physical log. Note that
809	 * callers can pass head == tail if the tail is not yet known.
810	 */
811	if (tail_blk >= head_blk && found != count) {
812		for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
813			error = xlog_bread(log, i, 1, buffer, &offset);
814			if (error)
815				goto out_error;
816
817			if (*(__be32 *)offset ==
818			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
819				*wrapped = true;
820				*rblk = i;
821				*rhead = (struct xlog_rec_header *) offset;
822				if (++found == count)
823					break;
824			}
825		}
826	}
827
828	return found;
829
830out_error:
831	return error;
832}
833
834/*
835 * Seek forward in the log for log record headers.
836 *
837 * Given head and tail blocks, walk forward from the tail block until we find
838 * the provided number of records or hit the head block. The return value is the
839 * number of records encountered or a negative error code. The log block and
840 * buffer pointer of the last record seen are returned in rblk and rhead
841 * respectively.
842 */
843STATIC int
844xlog_seek_logrec_hdr(
845	struct xlog		*log,
846	xfs_daddr_t		head_blk,
847	xfs_daddr_t		tail_blk,
848	int			count,
849	char			*buffer,
850	xfs_daddr_t		*rblk,
851	struct xlog_rec_header	**rhead,
852	bool			*wrapped)
853{
854	int			i;
855	int			error;
856	int			found = 0;
857	char			*offset = NULL;
858	xfs_daddr_t		end_blk;
859
860	*wrapped = false;
861
862	/*
863	 * Walk forward from the tail block until we hit the head or the last
864	 * block in the log.
865	 */
866	end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
867	for (i = (int) tail_blk; i <= end_blk; i++) {
868		error = xlog_bread(log, i, 1, buffer, &offset);
869		if (error)
870			goto out_error;
871
872		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
873			*rblk = i;
874			*rhead = (struct xlog_rec_header *) offset;
875			if (++found == count)
876				break;
877		}
878	}
879
880	/*
881	 * If we haven't hit the head block or the log record header count,
882	 * start looking again from the start of the physical log.
883	 */
884	if (tail_blk > head_blk && found != count) {
885		for (i = 0; i < (int) head_blk; i++) {
886			error = xlog_bread(log, i, 1, buffer, &offset);
887			if (error)
888				goto out_error;
889
890			if (*(__be32 *)offset ==
891			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
892				*wrapped = true;
893				*rblk = i;
894				*rhead = (struct xlog_rec_header *) offset;
895				if (++found == count)
896					break;
897			}
898		}
899	}
900
901	return found;
902
903out_error:
904	return error;
905}
906
907/*
908 * Calculate distance from head to tail (i.e., unused space in the log).
909 */
910static inline int
911xlog_tail_distance(
912	struct xlog	*log,
913	xfs_daddr_t	head_blk,
914	xfs_daddr_t	tail_blk)
915{
916	if (head_blk < tail_blk)
917		return tail_blk - head_blk;
918
919	return tail_blk + (log->l_logBBsize - head_blk);
920}
921
922/*
923 * Verify the log tail. This is particularly important when torn or incomplete
924 * writes have been detected near the front of the log and the head has been
925 * walked back accordingly.
926 *
927 * We also have to handle the case where the tail was pinned and the head
928 * blocked behind the tail right before a crash. If the tail had been pushed
929 * immediately prior to the crash and the subsequent checkpoint was only
930 * partially written, it's possible it overwrote the last referenced tail in the
931 * log with garbage. This is not a coherency problem because the tail must have
932 * been pushed before it can be overwritten, but appears as log corruption to
933 * recovery because we have no way to know the tail was updated if the
934 * subsequent checkpoint didn't write successfully.
935 *
936 * Therefore, CRC check the log from tail to head. If a failure occurs and the
937 * offending record is within max iclog bufs from the head, walk the tail
938 * forward and retry until a valid tail is found or corruption is detected out
939 * of the range of a possible overwrite.
940 */
941STATIC int
942xlog_verify_tail(
943	struct xlog		*log,
944	xfs_daddr_t		head_blk,
945	xfs_daddr_t		*tail_blk,
946	int			hsize)
947{
948	struct xlog_rec_header	*thead;
949	char			*buffer;
950	xfs_daddr_t		first_bad;
951	int			error = 0;
952	bool			wrapped;
953	xfs_daddr_t		tmp_tail;
954	xfs_daddr_t		orig_tail = *tail_blk;
955
956	buffer = xlog_alloc_buffer(log, 1);
957	if (!buffer)
958		return -ENOMEM;
959
960	/*
961	 * Make sure the tail points to a record (returns positive count on
962	 * success).
963	 */
964	error = xlog_seek_logrec_hdr(log, head_blk, *tail_blk, 1, buffer,
965			&tmp_tail, &thead, &wrapped);
966	if (error < 0)
967		goto out;
968	if (*tail_blk != tmp_tail)
969		*tail_blk = tmp_tail;
970
971	/*
972	 * Run a CRC check from the tail to the head. We can't just check
973	 * MAX_ICLOGS records past the tail because the tail may point to stale
974	 * blocks cleared during the search for the head/tail. These blocks are
975	 * overwritten with zero-length records and thus record count is not a
976	 * reliable indicator of the iclog state before a crash.
977	 */
978	first_bad = 0;
979	error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
980				      XLOG_RECOVER_CRCPASS, &first_bad);
981	while ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
982		int	tail_distance;
983
984		/*
985		 * Is corruption within range of the head? If so, retry from
986		 * the next record. Otherwise return an error.
987		 */
988		tail_distance = xlog_tail_distance(log, head_blk, first_bad);
989		if (tail_distance > BTOBB(XLOG_MAX_ICLOGS * hsize))
990			break;
991
992		/* skip to the next record; returns positive count on success */
993		error = xlog_seek_logrec_hdr(log, head_blk, first_bad, 2,
994				buffer, &tmp_tail, &thead, &wrapped);
995		if (error < 0)
996			goto out;
997
998		*tail_blk = tmp_tail;
999		first_bad = 0;
1000		error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1001					      XLOG_RECOVER_CRCPASS, &first_bad);
1002	}
1003
1004	if (!error && *tail_blk != orig_tail)
1005		xfs_warn(log->l_mp,
1006		"Tail block (0x%llx) overwrite detected. Updated to 0x%llx",
1007			 orig_tail, *tail_blk);
1008out:
1009	kmem_free(buffer);
1010	return error;
1011}
1012
1013/*
1014 * Detect and trim torn writes from the head of the log.
1015 *
1016 * Storage without sector atomicity guarantees can result in torn writes in the
1017 * log in the event of a crash. Our only means to detect this scenario is via
1018 * CRC verification. While we can't always be certain that CRC verification
1019 * failure is due to a torn write vs. an unrelated corruption, we do know that
1020 * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1021 * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1022 * the log and treat failures in this range as torn writes as a matter of
1023 * policy. In the event of CRC failure, the head is walked back to the last good
1024 * record in the log and the tail is updated from that record and verified.
1025 */
1026STATIC int
1027xlog_verify_head(
1028	struct xlog		*log,
1029	xfs_daddr_t		*head_blk,	/* in/out: unverified head */
1030	xfs_daddr_t		*tail_blk,	/* out: tail block */
1031	char			*buffer,
1032	xfs_daddr_t		*rhead_blk,	/* start blk of last record */
1033	struct xlog_rec_header	**rhead,	/* ptr to last record */
1034	bool			*wrapped)	/* last rec. wraps phys. log */
1035{
1036	struct xlog_rec_header	*tmp_rhead;
1037	char			*tmp_buffer;
1038	xfs_daddr_t		first_bad;
1039	xfs_daddr_t		tmp_rhead_blk;
1040	int			found;
1041	int			error;
1042	bool			tmp_wrapped;
1043
1044	/*
1045	 * Check the head of the log for torn writes. Search backwards from the
1046	 * head until we hit the tail or the maximum number of log record I/Os
1047	 * that could have been in flight at one time. Use a temporary buffer so
1048	 * we don't trash the rhead/buffer pointers from the caller.
1049	 */
1050	tmp_buffer = xlog_alloc_buffer(log, 1);
1051	if (!tmp_buffer)
1052		return -ENOMEM;
1053	error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1054				      XLOG_MAX_ICLOGS, tmp_buffer,
1055				      &tmp_rhead_blk, &tmp_rhead, &tmp_wrapped);
1056	kmem_free(tmp_buffer);
1057	if (error < 0)
1058		return error;
1059
1060	/*
1061	 * Now run a CRC verification pass over the records starting at the
1062	 * block found above to the current head. If a CRC failure occurs, the
1063	 * log block of the first bad record is saved in first_bad.
1064	 */
1065	error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1066				      XLOG_RECOVER_CRCPASS, &first_bad);
1067	if ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1068		/*
1069		 * We've hit a potential torn write. Reset the error and warn
1070		 * about it.
1071		 */
1072		error = 0;
1073		xfs_warn(log->l_mp,
1074"Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1075			 first_bad, *head_blk);
1076
1077		/*
1078		 * Get the header block and buffer pointer for the last good
1079		 * record before the bad record.
1080		 *
1081		 * Note that xlog_find_tail() clears the blocks at the new head
1082		 * (i.e., the records with invalid CRC) if the cycle number
1083		 * matches the current cycle.
1084		 */
1085		found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1,
1086				buffer, rhead_blk, rhead, wrapped);
1087		if (found < 0)
1088			return found;
1089		if (found == 0)		/* XXX: right thing to do here? */
1090			return -EIO;
1091
1092		/*
1093		 * Reset the head block to the starting block of the first bad
1094		 * log record and set the tail block based on the last good
1095		 * record.
1096		 *
1097		 * Bail out if the updated head/tail match as this indicates
1098		 * possible corruption outside of the acceptable
1099		 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1100		 */
1101		*head_blk = first_bad;
1102		*tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1103		if (*head_blk == *tail_blk) {
1104			ASSERT(0);
1105			return 0;
1106		}
1107	}
1108	if (error)
1109		return error;
1110
1111	return xlog_verify_tail(log, *head_blk, tail_blk,
1112				be32_to_cpu((*rhead)->h_size));
1113}
1114
1115/*
1116 * We need to make sure we handle log wrapping properly, so we can't use the
1117 * calculated logbno directly. Make sure it wraps to the correct bno inside the
1118 * log.
1119 *
1120 * The log is limited to 32 bit sizes, so we use the appropriate modulus
1121 * operation here and cast it back to a 64 bit daddr on return.
1122 */
1123static inline xfs_daddr_t
1124xlog_wrap_logbno(
1125	struct xlog		*log,
1126	xfs_daddr_t		bno)
1127{
1128	int			mod;
1129
1130	div_s64_rem(bno, log->l_logBBsize, &mod);
1131	return mod;
1132}
1133
1134/*
1135 * Check whether the head of the log points to an unmount record. In other
1136 * words, determine whether the log is clean. If so, update the in-core state
1137 * appropriately.
1138 */
1139static int
1140xlog_check_unmount_rec(
1141	struct xlog		*log,
1142	xfs_daddr_t		*head_blk,
1143	xfs_daddr_t		*tail_blk,
1144	struct xlog_rec_header	*rhead,
1145	xfs_daddr_t		rhead_blk,
1146	char			*buffer,
1147	bool			*clean)
1148{
1149	struct xlog_op_header	*op_head;
1150	xfs_daddr_t		umount_data_blk;
1151	xfs_daddr_t		after_umount_blk;
1152	int			hblks;
1153	int			error;
1154	char			*offset;
1155
1156	*clean = false;
1157
1158	/*
1159	 * Look for unmount record. If we find it, then we know there was a
1160	 * clean unmount. Since 'i' could be the last block in the physical
1161	 * log, we convert to a log block before comparing to the head_blk.
1162	 *
1163	 * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1164	 * below. We won't want to clear the unmount record if there is one, so
1165	 * we pass the lsn of the unmount record rather than the block after it.
1166	 */
1167	hblks = xlog_logrec_hblks(log, rhead);
1168	after_umount_blk = xlog_wrap_logbno(log,
1169			rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len)));
1170
1171	if (*head_blk == after_umount_blk &&
1172	    be32_to_cpu(rhead->h_num_logops) == 1) {
1173		umount_data_blk = xlog_wrap_logbno(log, rhead_blk + hblks);
1174		error = xlog_bread(log, umount_data_blk, 1, buffer, &offset);
1175		if (error)
1176			return error;
1177
1178		op_head = (struct xlog_op_header *)offset;
1179		if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1180			/*
1181			 * Set tail and last sync so that newly written log
1182			 * records will point recovery to after the current
1183			 * unmount record.
1184			 */
1185			xlog_assign_atomic_lsn(&log->l_tail_lsn,
1186					log->l_curr_cycle, after_umount_blk);
1187			xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1188					log->l_curr_cycle, after_umount_blk);
1189			*tail_blk = after_umount_blk;
1190
1191			*clean = true;
1192		}
1193	}
1194
1195	return 0;
1196}
1197
1198static void
1199xlog_set_state(
1200	struct xlog		*log,
1201	xfs_daddr_t		head_blk,
1202	struct xlog_rec_header	*rhead,
1203	xfs_daddr_t		rhead_blk,
1204	bool			bump_cycle)
1205{
1206	/*
1207	 * Reset log values according to the state of the log when we
1208	 * crashed.  In the case where head_blk == 0, we bump curr_cycle
1209	 * one because the next write starts a new cycle rather than
1210	 * continuing the cycle of the last good log record.  At this
1211	 * point we have guaranteed that all partial log records have been
1212	 * accounted for.  Therefore, we know that the last good log record
1213	 * written was complete and ended exactly on the end boundary
1214	 * of the physical log.
1215	 */
1216	log->l_prev_block = rhead_blk;
1217	log->l_curr_block = (int)head_blk;
1218	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1219	if (bump_cycle)
1220		log->l_curr_cycle++;
1221	atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1222	atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1223	xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1224					BBTOB(log->l_curr_block));
1225	xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1226					BBTOB(log->l_curr_block));
1227}
1228
1229/*
1230 * Find the sync block number or the tail of the log.
1231 *
1232 * This will be the block number of the last record to have its
1233 * associated buffers synced to disk.  Every log record header has
1234 * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
1235 * to get a sync block number.  The only concern is to figure out which
1236 * log record header to believe.
1237 *
1238 * The following algorithm uses the log record header with the largest
1239 * lsn.  The entire log record does not need to be valid.  We only care
1240 * that the header is valid.
1241 *
1242 * We could speed up search by using current head_blk buffer, but it is not
1243 * available.
1244 */
1245STATIC int
1246xlog_find_tail(
1247	struct xlog		*log,
1248	xfs_daddr_t		*head_blk,
1249	xfs_daddr_t		*tail_blk)
1250{
1251	xlog_rec_header_t	*rhead;
1252	char			*offset = NULL;
1253	char			*buffer;
1254	int			error;
1255	xfs_daddr_t		rhead_blk;
1256	xfs_lsn_t		tail_lsn;
1257	bool			wrapped = false;
1258	bool			clean = false;
1259
1260	/*
1261	 * Find previous log record
1262	 */
1263	if ((error = xlog_find_head(log, head_blk)))
1264		return error;
1265	ASSERT(*head_blk < INT_MAX);
1266
1267	buffer = xlog_alloc_buffer(log, 1);
1268	if (!buffer)
1269		return -ENOMEM;
1270	if (*head_blk == 0) {				/* special case */
1271		error = xlog_bread(log, 0, 1, buffer, &offset);
1272		if (error)
1273			goto done;
1274
1275		if (xlog_get_cycle(offset) == 0) {
1276			*tail_blk = 0;
1277			/* leave all other log inited values alone */
1278			goto done;
1279		}
1280	}
1281
1282	/*
1283	 * Search backwards through the log looking for the log record header
1284	 * block. This wraps all the way back around to the head so something is
1285	 * seriously wrong if we can't find it.
1286	 */
1287	error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, buffer,
1288				      &rhead_blk, &rhead, &wrapped);
1289	if (error < 0)
1290		goto done;
1291	if (!error) {
1292		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1293		error = -EFSCORRUPTED;
1294		goto done;
1295	}
1296	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1297
1298	/*
1299	 * Set the log state based on the current head record.
1300	 */
1301	xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1302	tail_lsn = atomic64_read(&log->l_tail_lsn);
1303
1304	/*
1305	 * Look for an unmount record at the head of the log. This sets the log
1306	 * state to determine whether recovery is necessary.
1307	 */
1308	error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1309				       rhead_blk, buffer, &clean);
1310	if (error)
1311		goto done;
1312
1313	/*
1314	 * Verify the log head if the log is not clean (e.g., we have anything
1315	 * but an unmount record at the head). This uses CRC verification to
1316	 * detect and trim torn writes. If discovered, CRC failures are
1317	 * considered torn writes and the log head is trimmed accordingly.
1318	 *
1319	 * Note that we can only run CRC verification when the log is dirty
1320	 * because there's no guarantee that the log data behind an unmount
1321	 * record is compatible with the current architecture.
1322	 */
1323	if (!clean) {
1324		xfs_daddr_t	orig_head = *head_blk;
1325
1326		error = xlog_verify_head(log, head_blk, tail_blk, buffer,
1327					 &rhead_blk, &rhead, &wrapped);
1328		if (error)
1329			goto done;
1330
1331		/* update in-core state again if the head changed */
1332		if (*head_blk != orig_head) {
1333			xlog_set_state(log, *head_blk, rhead, rhead_blk,
1334				       wrapped);
1335			tail_lsn = atomic64_read(&log->l_tail_lsn);
1336			error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1337						       rhead, rhead_blk, buffer,
1338						       &clean);
1339			if (error)
1340				goto done;
1341		}
1342	}
1343
1344	/*
1345	 * Note that the unmount was clean. If the unmount was not clean, we
1346	 * need to know this to rebuild the superblock counters from the perag
1347	 * headers if we have a filesystem using non-persistent counters.
1348	 */
1349	if (clean)
1350		set_bit(XFS_OPSTATE_CLEAN, &log->l_mp->m_opstate);
1351
1352	/*
1353	 * Make sure that there are no blocks in front of the head
1354	 * with the same cycle number as the head.  This can happen
1355	 * because we allow multiple outstanding log writes concurrently,
1356	 * and the later writes might make it out before earlier ones.
1357	 *
1358	 * We use the lsn from before modifying it so that we'll never
1359	 * overwrite the unmount record after a clean unmount.
1360	 *
1361	 * Do this only if we are going to recover the filesystem
1362	 *
1363	 * NOTE: This used to say "if (!readonly)"
1364	 * However on Linux, we can & do recover a read-only filesystem.
1365	 * We only skip recovery if NORECOVERY is specified on mount,
1366	 * in which case we would not be here.
1367	 *
1368	 * But... if the -device- itself is readonly, just skip this.
1369	 * We can't recover this device anyway, so it won't matter.
1370	 */
1371	if (!xfs_readonly_buftarg(log->l_targ))
1372		error = xlog_clear_stale_blocks(log, tail_lsn);
1373
1374done:
1375	kmem_free(buffer);
1376
1377	if (error)
1378		xfs_warn(log->l_mp, "failed to locate log tail");
1379	return error;
1380}
1381
1382/*
1383 * Is the log zeroed at all?
1384 *
1385 * The last binary search should be changed to perform an X block read
1386 * once X becomes small enough.  You can then search linearly through
1387 * the X blocks.  This will cut down on the number of reads we need to do.
1388 *
1389 * If the log is partially zeroed, this routine will pass back the blkno
1390 * of the first block with cycle number 0.  It won't have a complete LR
1391 * preceding it.
1392 *
1393 * Return:
1394 *	0  => the log is completely written to
1395 *	1 => use *blk_no as the first block of the log
1396 *	<0 => error has occurred
1397 */
1398STATIC int
1399xlog_find_zeroed(
1400	struct xlog	*log,
1401	xfs_daddr_t	*blk_no)
1402{
1403	char		*buffer;
1404	char		*offset;
1405	uint	        first_cycle, last_cycle;
1406	xfs_daddr_t	new_blk, last_blk, start_blk;
1407	xfs_daddr_t     num_scan_bblks;
1408	int	        error, log_bbnum = log->l_logBBsize;
1409
1410	*blk_no = 0;
1411
1412	/* check totally zeroed log */
1413	buffer = xlog_alloc_buffer(log, 1);
1414	if (!buffer)
1415		return -ENOMEM;
1416	error = xlog_bread(log, 0, 1, buffer, &offset);
1417	if (error)
1418		goto out_free_buffer;
1419
1420	first_cycle = xlog_get_cycle(offset);
1421	if (first_cycle == 0) {		/* completely zeroed log */
1422		*blk_no = 0;
1423		kmem_free(buffer);
1424		return 1;
1425	}
1426
1427	/* check partially zeroed log */
1428	error = xlog_bread(log, log_bbnum-1, 1, buffer, &offset);
1429	if (error)
1430		goto out_free_buffer;
1431
1432	last_cycle = xlog_get_cycle(offset);
1433	if (last_cycle != 0) {		/* log completely written to */
1434		kmem_free(buffer);
1435		return 0;
1436	}
1437
1438	/* we have a partially zeroed log */
1439	last_blk = log_bbnum-1;
1440	error = xlog_find_cycle_start(log, buffer, 0, &last_blk, 0);
1441	if (error)
1442		goto out_free_buffer;
1443
1444	/*
1445	 * Validate the answer.  Because there is no way to guarantee that
1446	 * the entire log is made up of log records which are the same size,
1447	 * we scan over the defined maximum blocks.  At this point, the maximum
1448	 * is not chosen to mean anything special.   XXXmiken
1449	 */
1450	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1451	ASSERT(num_scan_bblks <= INT_MAX);
1452
1453	if (last_blk < num_scan_bblks)
1454		num_scan_bblks = last_blk;
1455	start_blk = last_blk - num_scan_bblks;
1456
1457	/*
1458	 * We search for any instances of cycle number 0 that occur before
1459	 * our current estimate of the head.  What we're trying to detect is
1460	 *        1 ... | 0 | 1 | 0...
1461	 *                       ^ binary search ends here
1462	 */
1463	if ((error = xlog_find_verify_cycle(log, start_blk,
1464					 (int)num_scan_bblks, 0, &new_blk)))
1465		goto out_free_buffer;
1466	if (new_blk != -1)
1467		last_blk = new_blk;
1468
1469	/*
1470	 * Potentially backup over partial log record write.  We don't need
1471	 * to search the end of the log because we know it is zero.
1472	 */
1473	error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1474	if (error == 1)
1475		error = -EIO;
1476	if (error)
1477		goto out_free_buffer;
1478
1479	*blk_no = last_blk;
1480out_free_buffer:
1481	kmem_free(buffer);
1482	if (error)
1483		return error;
1484	return 1;
1485}
1486
1487/*
1488 * These are simple subroutines used by xlog_clear_stale_blocks() below
1489 * to initialize a buffer full of empty log record headers and write
1490 * them into the log.
1491 */
1492STATIC void
1493xlog_add_record(
1494	struct xlog		*log,
1495	char			*buf,
1496	int			cycle,
1497	int			block,
1498	int			tail_cycle,
1499	int			tail_block)
1500{
1501	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1502
1503	memset(buf, 0, BBSIZE);
1504	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1505	recp->h_cycle = cpu_to_be32(cycle);
1506	recp->h_version = cpu_to_be32(
1507			xfs_has_logv2(log->l_mp) ? 2 : 1);
1508	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1509	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1510	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1511	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1512}
1513
1514STATIC int
1515xlog_write_log_records(
1516	struct xlog	*log,
1517	int		cycle,
1518	int		start_block,
1519	int		blocks,
1520	int		tail_cycle,
1521	int		tail_block)
1522{
1523	char		*offset;
1524	char		*buffer;
1525	int		balign, ealign;
1526	int		sectbb = log->l_sectBBsize;
1527	int		end_block = start_block + blocks;
1528	int		bufblks;
1529	int		error = 0;
1530	int		i, j = 0;
1531
1532	/*
1533	 * Greedily allocate a buffer big enough to handle the full
1534	 * range of basic blocks to be written.  If that fails, try
1535	 * a smaller size.  We need to be able to write at least a
1536	 * log sector, or we're out of luck.
1537	 */
1538	bufblks = 1 << ffs(blocks);
1539	while (bufblks > log->l_logBBsize)
1540		bufblks >>= 1;
1541	while (!(buffer = xlog_alloc_buffer(log, bufblks))) {
1542		bufblks >>= 1;
1543		if (bufblks < sectbb)
1544			return -ENOMEM;
1545	}
1546
1547	/* We may need to do a read at the start to fill in part of
1548	 * the buffer in the starting sector not covered by the first
1549	 * write below.
1550	 */
1551	balign = round_down(start_block, sectbb);
1552	if (balign != start_block) {
1553		error = xlog_bread_noalign(log, start_block, 1, buffer);
1554		if (error)
1555			goto out_free_buffer;
1556
1557		j = start_block - balign;
1558	}
1559
1560	for (i = start_block; i < end_block; i += bufblks) {
1561		int		bcount, endcount;
1562
1563		bcount = min(bufblks, end_block - start_block);
1564		endcount = bcount - j;
1565
1566		/* We may need to do a read at the end to fill in part of
1567		 * the buffer in the final sector not covered by the write.
1568		 * If this is the same sector as the above read, skip it.
1569		 */
1570		ealign = round_down(end_block, sectbb);
1571		if (j == 0 && (start_block + endcount > ealign)) {
1572			error = xlog_bread_noalign(log, ealign, sectbb,
1573					buffer + BBTOB(ealign - start_block));
1574			if (error)
1575				break;
1576
1577		}
1578
1579		offset = buffer + xlog_align(log, start_block);
1580		for (; j < endcount; j++) {
1581			xlog_add_record(log, offset, cycle, i+j,
1582					tail_cycle, tail_block);
1583			offset += BBSIZE;
1584		}
1585		error = xlog_bwrite(log, start_block, endcount, buffer);
1586		if (error)
1587			break;
1588		start_block += endcount;
1589		j = 0;
1590	}
1591
1592out_free_buffer:
1593	kmem_free(buffer);
1594	return error;
1595}
1596
1597/*
1598 * This routine is called to blow away any incomplete log writes out
1599 * in front of the log head.  We do this so that we won't become confused
1600 * if we come up, write only a little bit more, and then crash again.
1601 * If we leave the partial log records out there, this situation could
1602 * cause us to think those partial writes are valid blocks since they
1603 * have the current cycle number.  We get rid of them by overwriting them
1604 * with empty log records with the old cycle number rather than the
1605 * current one.
1606 *
1607 * The tail lsn is passed in rather than taken from
1608 * the log so that we will not write over the unmount record after a
1609 * clean unmount in a 512 block log.  Doing so would leave the log without
1610 * any valid log records in it until a new one was written.  If we crashed
1611 * during that time we would not be able to recover.
1612 */
1613STATIC int
1614xlog_clear_stale_blocks(
1615	struct xlog	*log,
1616	xfs_lsn_t	tail_lsn)
1617{
1618	int		tail_cycle, head_cycle;
1619	int		tail_block, head_block;
1620	int		tail_distance, max_distance;
1621	int		distance;
1622	int		error;
1623
1624	tail_cycle = CYCLE_LSN(tail_lsn);
1625	tail_block = BLOCK_LSN(tail_lsn);
1626	head_cycle = log->l_curr_cycle;
1627	head_block = log->l_curr_block;
1628
1629	/*
1630	 * Figure out the distance between the new head of the log
1631	 * and the tail.  We want to write over any blocks beyond the
1632	 * head that we may have written just before the crash, but
1633	 * we don't want to overwrite the tail of the log.
1634	 */
1635	if (head_cycle == tail_cycle) {
1636		/*
1637		 * The tail is behind the head in the physical log,
1638		 * so the distance from the head to the tail is the
1639		 * distance from the head to the end of the log plus
1640		 * the distance from the beginning of the log to the
1641		 * tail.
1642		 */
1643		if (XFS_IS_CORRUPT(log->l_mp,
1644				   head_block < tail_block ||
1645				   head_block >= log->l_logBBsize))
1646			return -EFSCORRUPTED;
1647		tail_distance = tail_block + (log->l_logBBsize - head_block);
1648	} else {
1649		/*
1650		 * The head is behind the tail in the physical log,
1651		 * so the distance from the head to the tail is just
1652		 * the tail block minus the head block.
1653		 */
1654		if (XFS_IS_CORRUPT(log->l_mp,
1655				   head_block >= tail_block ||
1656				   head_cycle != tail_cycle + 1))
1657			return -EFSCORRUPTED;
1658		tail_distance = tail_block - head_block;
1659	}
1660
1661	/*
1662	 * If the head is right up against the tail, we can't clear
1663	 * anything.
1664	 */
1665	if (tail_distance <= 0) {
1666		ASSERT(tail_distance == 0);
1667		return 0;
1668	}
1669
1670	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1671	/*
1672	 * Take the smaller of the maximum amount of outstanding I/O
1673	 * we could have and the distance to the tail to clear out.
1674	 * We take the smaller so that we don't overwrite the tail and
1675	 * we don't waste all day writing from the head to the tail
1676	 * for no reason.
1677	 */
1678	max_distance = min(max_distance, tail_distance);
1679
1680	if ((head_block + max_distance) <= log->l_logBBsize) {
1681		/*
1682		 * We can stomp all the blocks we need to without
1683		 * wrapping around the end of the log.  Just do it
1684		 * in a single write.  Use the cycle number of the
1685		 * current cycle minus one so that the log will look like:
1686		 *     n ... | n - 1 ...
1687		 */
1688		error = xlog_write_log_records(log, (head_cycle - 1),
1689				head_block, max_distance, tail_cycle,
1690				tail_block);
1691		if (error)
1692			return error;
1693	} else {
1694		/*
1695		 * We need to wrap around the end of the physical log in
1696		 * order to clear all the blocks.  Do it in two separate
1697		 * I/Os.  The first write should be from the head to the
1698		 * end of the physical log, and it should use the current
1699		 * cycle number minus one just like above.
1700		 */
1701		distance = log->l_logBBsize - head_block;
1702		error = xlog_write_log_records(log, (head_cycle - 1),
1703				head_block, distance, tail_cycle,
1704				tail_block);
1705
1706		if (error)
1707			return error;
1708
1709		/*
1710		 * Now write the blocks at the start of the physical log.
1711		 * This writes the remainder of the blocks we want to clear.
1712		 * It uses the current cycle number since we're now on the
1713		 * same cycle as the head so that we get:
1714		 *    n ... n ... | n - 1 ...
1715		 *    ^^^^^ blocks we're writing
1716		 */
1717		distance = max_distance - (log->l_logBBsize - head_block);
1718		error = xlog_write_log_records(log, head_cycle, 0, distance,
1719				tail_cycle, tail_block);
1720		if (error)
1721			return error;
1722	}
1723
1724	return 0;
1725}
1726
1727/*
1728 * Release the recovered intent item in the AIL that matches the given intent
1729 * type and intent id.
1730 */
1731void
1732xlog_recover_release_intent(
1733	struct xlog		*log,
1734	unsigned short		intent_type,
1735	uint64_t		intent_id)
1736{
1737	struct xfs_ail_cursor	cur;
1738	struct xfs_log_item	*lip;
1739	struct xfs_ail		*ailp = log->l_ailp;
1740
1741	spin_lock(&ailp->ail_lock);
1742	for (lip = xfs_trans_ail_cursor_first(ailp, &cur, 0); lip != NULL;
1743	     lip = xfs_trans_ail_cursor_next(ailp, &cur)) {
1744		if (lip->li_type != intent_type)
1745			continue;
1746		if (!lip->li_ops->iop_match(lip, intent_id))
1747			continue;
1748
1749		spin_unlock(&ailp->ail_lock);
1750		lip->li_ops->iop_release(lip);
1751		spin_lock(&ailp->ail_lock);
1752		break;
1753	}
1754
1755	xfs_trans_ail_cursor_done(&cur);
1756	spin_unlock(&ailp->ail_lock);
1757}
1758
1759int
1760xlog_recover_iget(
1761	struct xfs_mount	*mp,
1762	xfs_ino_t		ino,
1763	struct xfs_inode	**ipp)
1764{
1765	int			error;
1766
1767	error = xfs_iget(mp, NULL, ino, 0, 0, ipp);
1768	if (error)
1769		return error;
1770
1771	error = xfs_qm_dqattach(*ipp);
1772	if (error) {
1773		xfs_irele(*ipp);
1774		return error;
1775	}
1776
1777	if (VFS_I(*ipp)->i_nlink == 0)
1778		xfs_iflags_set(*ipp, XFS_IRECOVERY);
1779
1780	return 0;
1781}
1782
1783/******************************************************************************
1784 *
1785 *		Log recover routines
1786 *
1787 ******************************************************************************
1788 */
1789static const struct xlog_recover_item_ops *xlog_recover_item_ops[] = {
1790	&xlog_buf_item_ops,
1791	&xlog_inode_item_ops,
1792	&xlog_dquot_item_ops,
1793	&xlog_quotaoff_item_ops,
1794	&xlog_icreate_item_ops,
1795	&xlog_efi_item_ops,
1796	&xlog_efd_item_ops,
1797	&xlog_rui_item_ops,
1798	&xlog_rud_item_ops,
1799	&xlog_cui_item_ops,
1800	&xlog_cud_item_ops,
1801	&xlog_bui_item_ops,
1802	&xlog_bud_item_ops,
1803};
1804
1805static const struct xlog_recover_item_ops *
1806xlog_find_item_ops(
1807	struct xlog_recover_item		*item)
1808{
1809	unsigned int				i;
1810
1811	for (i = 0; i < ARRAY_SIZE(xlog_recover_item_ops); i++)
1812		if (ITEM_TYPE(item) == xlog_recover_item_ops[i]->item_type)
1813			return xlog_recover_item_ops[i];
1814
1815	return NULL;
1816}
1817
1818/*
1819 * Sort the log items in the transaction.
1820 *
1821 * The ordering constraints are defined by the inode allocation and unlink
1822 * behaviour. The rules are:
1823 *
1824 *	1. Every item is only logged once in a given transaction. Hence it
1825 *	   represents the last logged state of the item. Hence ordering is
1826 *	   dependent on the order in which operations need to be performed so
1827 *	   required initial conditions are always met.
1828 *
1829 *	2. Cancelled buffers are recorded in pass 1 in a separate table and
1830 *	   there's nothing to replay from them so we can simply cull them
1831 *	   from the transaction. However, we can't do that until after we've
1832 *	   replayed all the other items because they may be dependent on the
1833 *	   cancelled buffer and replaying the cancelled buffer can remove it
1834 *	   form the cancelled buffer table. Hence they have tobe done last.
1835 *
1836 *	3. Inode allocation buffers must be replayed before inode items that
1837 *	   read the buffer and replay changes into it. For filesystems using the
1838 *	   ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1839 *	   treated the same as inode allocation buffers as they create and
1840 *	   initialise the buffers directly.
1841 *
1842 *	4. Inode unlink buffers must be replayed after inode items are replayed.
1843 *	   This ensures that inodes are completely flushed to the inode buffer
1844 *	   in a "free" state before we remove the unlinked inode list pointer.
1845 *
1846 * Hence the ordering needs to be inode allocation buffers first, inode items
1847 * second, inode unlink buffers third and cancelled buffers last.
1848 *
1849 * But there's a problem with that - we can't tell an inode allocation buffer
1850 * apart from a regular buffer, so we can't separate them. We can, however,
1851 * tell an inode unlink buffer from the others, and so we can separate them out
1852 * from all the other buffers and move them to last.
1853 *
1854 * Hence, 4 lists, in order from head to tail:
1855 *	- buffer_list for all buffers except cancelled/inode unlink buffers
1856 *	- item_list for all non-buffer items
1857 *	- inode_buffer_list for inode unlink buffers
1858 *	- cancel_list for the cancelled buffers
1859 *
1860 * Note that we add objects to the tail of the lists so that first-to-last
1861 * ordering is preserved within the lists. Adding objects to the head of the
1862 * list means when we traverse from the head we walk them in last-to-first
1863 * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1864 * but for all other items there may be specific ordering that we need to
1865 * preserve.
1866 */
1867STATIC int
1868xlog_recover_reorder_trans(
1869	struct xlog		*log,
1870	struct xlog_recover	*trans,
1871	int			pass)
1872{
1873	struct xlog_recover_item *item, *n;
1874	int			error = 0;
1875	LIST_HEAD(sort_list);
1876	LIST_HEAD(cancel_list);
1877	LIST_HEAD(buffer_list);
1878	LIST_HEAD(inode_buffer_list);
1879	LIST_HEAD(item_list);
1880
1881	list_splice_init(&trans->r_itemq, &sort_list);
1882	list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1883		enum xlog_recover_reorder	fate = XLOG_REORDER_ITEM_LIST;
1884
1885		item->ri_ops = xlog_find_item_ops(item);
1886		if (!item->ri_ops) {
1887			xfs_warn(log->l_mp,
1888				"%s: unrecognized type of log operation (%d)",
1889				__func__, ITEM_TYPE(item));
1890			ASSERT(0);
1891			/*
1892			 * return the remaining items back to the transaction
1893			 * item list so they can be freed in caller.
1894			 */
1895			if (!list_empty(&sort_list))
1896				list_splice_init(&sort_list, &trans->r_itemq);
1897			error = -EFSCORRUPTED;
1898			break;
1899		}
1900
1901		if (item->ri_ops->reorder)
1902			fate = item->ri_ops->reorder(item);
1903
1904		switch (fate) {
1905		case XLOG_REORDER_BUFFER_LIST:
1906			list_move_tail(&item->ri_list, &buffer_list);
1907			break;
1908		case XLOG_REORDER_CANCEL_LIST:
1909			trace_xfs_log_recover_item_reorder_head(log,
1910					trans, item, pass);
1911			list_move(&item->ri_list, &cancel_list);
1912			break;
1913		case XLOG_REORDER_INODE_BUFFER_LIST:
1914			list_move(&item->ri_list, &inode_buffer_list);
1915			break;
1916		case XLOG_REORDER_ITEM_LIST:
1917			trace_xfs_log_recover_item_reorder_tail(log,
1918							trans, item, pass);
1919			list_move_tail(&item->ri_list, &item_list);
1920			break;
1921		}
1922	}
1923
1924	ASSERT(list_empty(&sort_list));
1925	if (!list_empty(&buffer_list))
1926		list_splice(&buffer_list, &trans->r_itemq);
1927	if (!list_empty(&item_list))
1928		list_splice_tail(&item_list, &trans->r_itemq);
1929	if (!list_empty(&inode_buffer_list))
1930		list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1931	if (!list_empty(&cancel_list))
1932		list_splice_tail(&cancel_list, &trans->r_itemq);
1933	return error;
1934}
1935
1936void
1937xlog_buf_readahead(
1938	struct xlog		*log,
1939	xfs_daddr_t		blkno,
1940	uint			len,
1941	const struct xfs_buf_ops *ops)
1942{
1943	if (!xlog_is_buffer_cancelled(log, blkno, len))
1944		xfs_buf_readahead(log->l_mp->m_ddev_targp, blkno, len, ops);
1945}
1946
1947STATIC int
1948xlog_recover_items_pass2(
1949	struct xlog                     *log,
1950	struct xlog_recover             *trans,
1951	struct list_head                *buffer_list,
1952	struct list_head                *item_list)
1953{
1954	struct xlog_recover_item	*item;
1955	int				error = 0;
1956
1957	list_for_each_entry(item, item_list, ri_list) {
1958		trace_xfs_log_recover_item_recover(log, trans, item,
1959				XLOG_RECOVER_PASS2);
1960
1961		if (item->ri_ops->commit_pass2)
1962			error = item->ri_ops->commit_pass2(log, buffer_list,
1963					item, trans->r_lsn);
1964		if (error)
1965			return error;
1966	}
1967
1968	return error;
1969}
1970
1971/*
1972 * Perform the transaction.
1973 *
1974 * If the transaction modifies a buffer or inode, do it now.  Otherwise,
1975 * EFIs and EFDs get queued up by adding entries into the AIL for them.
1976 */
1977STATIC int
1978xlog_recover_commit_trans(
1979	struct xlog		*log,
1980	struct xlog_recover	*trans,
1981	int			pass,
1982	struct list_head	*buffer_list)
1983{
1984	int				error = 0;
1985	int				items_queued = 0;
1986	struct xlog_recover_item	*item;
1987	struct xlog_recover_item	*next;
1988	LIST_HEAD			(ra_list);
1989	LIST_HEAD			(done_list);
1990
1991	#define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
1992
1993	hlist_del_init(&trans->r_list);
1994
1995	error = xlog_recover_reorder_trans(log, trans, pass);
1996	if (error)
1997		return error;
1998
1999	list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
2000		trace_xfs_log_recover_item_recover(log, trans, item, pass);
2001
2002		switch (pass) {
2003		case XLOG_RECOVER_PASS1:
2004			if (item->ri_ops->commit_pass1)
2005				error = item->ri_ops->commit_pass1(log, item);
2006			break;
2007		case XLOG_RECOVER_PASS2:
2008			if (item->ri_ops->ra_pass2)
2009				item->ri_ops->ra_pass2(log, item);
2010			list_move_tail(&item->ri_list, &ra_list);
2011			items_queued++;
2012			if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
2013				error = xlog_recover_items_pass2(log, trans,
2014						buffer_list, &ra_list);
2015				list_splice_tail_init(&ra_list, &done_list);
2016				items_queued = 0;
2017			}
2018
2019			break;
2020		default:
2021			ASSERT(0);
2022		}
2023
2024		if (error)
2025			goto out;
2026	}
2027
2028out:
2029	if (!list_empty(&ra_list)) {
2030		if (!error)
2031			error = xlog_recover_items_pass2(log, trans,
2032					buffer_list, &ra_list);
2033		list_splice_tail_init(&ra_list, &done_list);
2034	}
2035
2036	if (!list_empty(&done_list))
2037		list_splice_init(&done_list, &trans->r_itemq);
2038
2039	return error;
2040}
2041
2042STATIC void
2043xlog_recover_add_item(
2044	struct list_head	*head)
2045{
2046	struct xlog_recover_item *item;
2047
2048	item = kmem_zalloc(sizeof(struct xlog_recover_item), 0);
2049	INIT_LIST_HEAD(&item->ri_list);
2050	list_add_tail(&item->ri_list, head);
2051}
2052
2053STATIC int
2054xlog_recover_add_to_cont_trans(
2055	struct xlog		*log,
2056	struct xlog_recover	*trans,
2057	char			*dp,
2058	int			len)
2059{
2060	struct xlog_recover_item *item;
2061	char			*ptr, *old_ptr;
2062	int			old_len;
2063
2064	/*
2065	 * If the transaction is empty, the header was split across this and the
2066	 * previous record. Copy the rest of the header.
2067	 */
2068	if (list_empty(&trans->r_itemq)) {
2069		ASSERT(len <= sizeof(struct xfs_trans_header));
2070		if (len > sizeof(struct xfs_trans_header)) {
2071			xfs_warn(log->l_mp, "%s: bad header length", __func__);
2072			return -EFSCORRUPTED;
2073		}
2074
2075		xlog_recover_add_item(&trans->r_itemq);
2076		ptr = (char *)&trans->r_theader +
2077				sizeof(struct xfs_trans_header) - len;
2078		memcpy(ptr, dp, len);
2079		return 0;
2080	}
2081
2082	/* take the tail entry */
2083	item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2084			  ri_list);
2085
2086	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
2087	old_len = item->ri_buf[item->ri_cnt-1].i_len;
2088
2089	ptr = kvrealloc(old_ptr, old_len, len + old_len, GFP_KERNEL);
2090	if (!ptr)
2091		return -ENOMEM;
2092	memcpy(&ptr[old_len], dp, len);
2093	item->ri_buf[item->ri_cnt-1].i_len += len;
2094	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
2095	trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
2096	return 0;
2097}
2098
2099/*
2100 * The next region to add is the start of a new region.  It could be
2101 * a whole region or it could be the first part of a new region.  Because
2102 * of this, the assumption here is that the type and size fields of all
2103 * format structures fit into the first 32 bits of the structure.
2104 *
2105 * This works because all regions must be 32 bit aligned.  Therefore, we
2106 * either have both fields or we have neither field.  In the case we have
2107 * neither field, the data part of the region is zero length.  We only have
2108 * a log_op_header and can throw away the header since a new one will appear
2109 * later.  If we have at least 4 bytes, then we can determine how many regions
2110 * will appear in the current log item.
2111 */
2112STATIC int
2113xlog_recover_add_to_trans(
2114	struct xlog		*log,
2115	struct xlog_recover	*trans,
2116	char			*dp,
2117	int			len)
2118{
2119	struct xfs_inode_log_format	*in_f;			/* any will do */
2120	struct xlog_recover_item *item;
2121	char			*ptr;
2122
2123	if (!len)
2124		return 0;
2125	if (list_empty(&trans->r_itemq)) {
2126		/* we need to catch log corruptions here */
2127		if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
2128			xfs_warn(log->l_mp, "%s: bad header magic number",
2129				__func__);
2130			ASSERT(0);
2131			return -EFSCORRUPTED;
2132		}
2133
2134		if (len > sizeof(struct xfs_trans_header)) {
2135			xfs_warn(log->l_mp, "%s: bad header length", __func__);
2136			ASSERT(0);
2137			return -EFSCORRUPTED;
2138		}
2139
2140		/*
2141		 * The transaction header can be arbitrarily split across op
2142		 * records. If we don't have the whole thing here, copy what we
2143		 * do have and handle the rest in the next record.
2144		 */
2145		if (len == sizeof(struct xfs_trans_header))
2146			xlog_recover_add_item(&trans->r_itemq);
2147		memcpy(&trans->r_theader, dp, len);
2148		return 0;
2149	}
2150
2151	ptr = kmem_alloc(len, 0);
2152	memcpy(ptr, dp, len);
2153	in_f = (struct xfs_inode_log_format *)ptr;
2154
2155	/* take the tail entry */
2156	item = list_entry(trans->r_itemq.prev, struct xlog_recover_item,
2157			  ri_list);
2158	if (item->ri_total != 0 &&
2159	     item->ri_total == item->ri_cnt) {
2160		/* tail item is in use, get a new one */
2161		xlog_recover_add_item(&trans->r_itemq);
2162		item = list_entry(trans->r_itemq.prev,
2163					struct xlog_recover_item, ri_list);
2164	}
2165
2166	if (item->ri_total == 0) {		/* first region to be added */
2167		if (in_f->ilf_size == 0 ||
2168		    in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
2169			xfs_warn(log->l_mp,
2170		"bad number of regions (%d) in inode log format",
2171				  in_f->ilf_size);
2172			ASSERT(0);
2173			kmem_free(ptr);
2174			return -EFSCORRUPTED;
2175		}
2176
2177		item->ri_total = in_f->ilf_size;
2178		item->ri_buf =
2179			kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
2180				    0);
2181	}
2182
2183	if (item->ri_total <= item->ri_cnt) {
2184		xfs_warn(log->l_mp,
2185	"log item region count (%d) overflowed size (%d)",
2186				item->ri_cnt, item->ri_total);
2187		ASSERT(0);
2188		kmem_free(ptr);
2189		return -EFSCORRUPTED;
2190	}
2191
2192	/* Description region is ri_buf[0] */
2193	item->ri_buf[item->ri_cnt].i_addr = ptr;
2194	item->ri_buf[item->ri_cnt].i_len  = len;
2195	item->ri_cnt++;
2196	trace_xfs_log_recover_item_add(log, trans, item, 0);
2197	return 0;
2198}
2199
2200/*
2201 * Free up any resources allocated by the transaction
2202 *
2203 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2204 */
2205STATIC void
2206xlog_recover_free_trans(
2207	struct xlog_recover	*trans)
2208{
2209	struct xlog_recover_item *item, *n;
2210	int			i;
2211
2212	hlist_del_init(&trans->r_list);
2213
2214	list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2215		/* Free the regions in the item. */
2216		list_del(&item->ri_list);
2217		for (i = 0; i < item->ri_cnt; i++)
2218			kmem_free(item->ri_buf[i].i_addr);
2219		/* Free the item itself */
2220		kmem_free(item->ri_buf);
2221		kmem_free(item);
2222	}
2223	/* Free the transaction recover structure */
2224	kmem_free(trans);
2225}
2226
2227/*
2228 * On error or completion, trans is freed.
2229 */
2230STATIC int
2231xlog_recovery_process_trans(
2232	struct xlog		*log,
2233	struct xlog_recover	*trans,
2234	char			*dp,
2235	unsigned int		len,
2236	unsigned int		flags,
2237	int			pass,
2238	struct list_head	*buffer_list)
2239{
2240	int			error = 0;
2241	bool			freeit = false;
2242
2243	/* mask off ophdr transaction container flags */
2244	flags &= ~XLOG_END_TRANS;
2245	if (flags & XLOG_WAS_CONT_TRANS)
2246		flags &= ~XLOG_CONTINUE_TRANS;
2247
2248	/*
2249	 * Callees must not free the trans structure. We'll decide if we need to
2250	 * free it or not based on the operation being done and it's result.
2251	 */
2252	switch (flags) {
2253	/* expected flag values */
2254	case 0:
2255	case XLOG_CONTINUE_TRANS:
2256		error = xlog_recover_add_to_trans(log, trans, dp, len);
2257		break;
2258	case XLOG_WAS_CONT_TRANS:
2259		error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
2260		break;
2261	case XLOG_COMMIT_TRANS:
2262		error = xlog_recover_commit_trans(log, trans, pass,
2263						  buffer_list);
2264		/* success or fail, we are now done with this transaction. */
2265		freeit = true;
2266		break;
2267
2268	/* unexpected flag values */
2269	case XLOG_UNMOUNT_TRANS:
2270		/* just skip trans */
2271		xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
2272		freeit = true;
2273		break;
2274	case XLOG_START_TRANS:
2275	default:
2276		xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
2277		ASSERT(0);
2278		error = -EFSCORRUPTED;
2279		break;
2280	}
2281	if (error || freeit)
2282		xlog_recover_free_trans(trans);
2283	return error;
2284}
2285
2286/*
2287 * Lookup the transaction recovery structure associated with the ID in the
2288 * current ophdr. If the transaction doesn't exist and the start flag is set in
2289 * the ophdr, then allocate a new transaction for future ID matches to find.
2290 * Either way, return what we found during the lookup - an existing transaction
2291 * or nothing.
2292 */
2293STATIC struct xlog_recover *
2294xlog_recover_ophdr_to_trans(
2295	struct hlist_head	rhash[],
2296	struct xlog_rec_header	*rhead,
2297	struct xlog_op_header	*ohead)
2298{
2299	struct xlog_recover	*trans;
2300	xlog_tid_t		tid;
2301	struct hlist_head	*rhp;
2302
2303	tid = be32_to_cpu(ohead->oh_tid);
2304	rhp = &rhash[XLOG_RHASH(tid)];
2305	hlist_for_each_entry(trans, rhp, r_list) {
2306		if (trans->r_log_tid == tid)
2307			return trans;
2308	}
2309
2310	/*
2311	 * skip over non-start transaction headers - we could be
2312	 * processing slack space before the next transaction starts
2313	 */
2314	if (!(ohead->oh_flags & XLOG_START_TRANS))
2315		return NULL;
2316
2317	ASSERT(be32_to_cpu(ohead->oh_len) == 0);
2318
2319	/*
2320	 * This is a new transaction so allocate a new recovery container to
2321	 * hold the recovery ops that will follow.
2322	 */
2323	trans = kmem_zalloc(sizeof(struct xlog_recover), 0);
2324	trans->r_log_tid = tid;
2325	trans->r_lsn = be64_to_cpu(rhead->h_lsn);
2326	INIT_LIST_HEAD(&trans->r_itemq);
2327	INIT_HLIST_NODE(&trans->r_list);
2328	hlist_add_head(&trans->r_list, rhp);
2329
2330	/*
2331	 * Nothing more to do for this ophdr. Items to be added to this new
2332	 * transaction will be in subsequent ophdr containers.
2333	 */
2334	return NULL;
2335}
2336
2337STATIC int
2338xlog_recover_process_ophdr(
2339	struct xlog		*log,
2340	struct hlist_head	rhash[],
2341	struct xlog_rec_header	*rhead,
2342	struct xlog_op_header	*ohead,
2343	char			*dp,
2344	char			*end,
2345	int			pass,
2346	struct list_head	*buffer_list)
2347{
2348	struct xlog_recover	*trans;
2349	unsigned int		len;
2350	int			error;
2351
2352	/* Do we understand who wrote this op? */
2353	if (ohead->oh_clientid != XFS_TRANSACTION &&
2354	    ohead->oh_clientid != XFS_LOG) {
2355		xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
2356			__func__, ohead->oh_clientid);
2357		ASSERT(0);
2358		return -EFSCORRUPTED;
2359	}
2360
2361	/*
2362	 * Check the ophdr contains all the data it is supposed to contain.
2363	 */
2364	len = be32_to_cpu(ohead->oh_len);
2365	if (dp + len > end) {
2366		xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
2367		WARN_ON(1);
2368		return -EFSCORRUPTED;
2369	}
2370
2371	trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
2372	if (!trans) {
2373		/* nothing to do, so skip over this ophdr */
2374		return 0;
2375	}
2376
2377	/*
2378	 * The recovered buffer queue is drained only once we know that all
2379	 * recovery items for the current LSN have been processed. This is
2380	 * required because:
2381	 *
2382	 * - Buffer write submission updates the metadata LSN of the buffer.
2383	 * - Log recovery skips items with a metadata LSN >= the current LSN of
2384	 *   the recovery item.
2385	 * - Separate recovery items against the same metadata buffer can share
2386	 *   a current LSN. I.e., consider that the LSN of a recovery item is
2387	 *   defined as the starting LSN of the first record in which its
2388	 *   transaction appears, that a record can hold multiple transactions,
2389	 *   and/or that a transaction can span multiple records.
2390	 *
2391	 * In other words, we are allowed to submit a buffer from log recovery
2392	 * once per current LSN. Otherwise, we may incorrectly skip recovery
2393	 * items and cause corruption.
2394	 *
2395	 * We don't know up front whether buffers are updated multiple times per
2396	 * LSN. Therefore, track the current LSN of each commit log record as it
2397	 * is processed and drain the queue when it changes. Use commit records
2398	 * because they are ordered correctly by the logging code.
2399	 */
2400	if (log->l_recovery_lsn != trans->r_lsn &&
2401	    ohead->oh_flags & XLOG_COMMIT_TRANS) {
2402		error = xfs_buf_delwri_submit(buffer_list);
2403		if (error)
2404			return error;
2405		log->l_recovery_lsn = trans->r_lsn;
2406	}
2407
2408	return xlog_recovery_process_trans(log, trans, dp, len,
2409					   ohead->oh_flags, pass, buffer_list);
2410}
2411
2412/*
2413 * There are two valid states of the r_state field.  0 indicates that the
2414 * transaction structure is in a normal state.  We have either seen the
2415 * start of the transaction or the last operation we added was not a partial
2416 * operation.  If the last operation we added to the transaction was a
2417 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2418 *
2419 * NOTE: skip LRs with 0 data length.
2420 */
2421STATIC int
2422xlog_recover_process_data(
2423	struct xlog		*log,
2424	struct hlist_head	rhash[],
2425	struct xlog_rec_header	*rhead,
2426	char			*dp,
2427	int			pass,
2428	struct list_head	*buffer_list)
2429{
2430	struct xlog_op_header	*ohead;
2431	char			*end;
2432	int			num_logops;
2433	int			error;
2434
2435	end = dp + be32_to_cpu(rhead->h_len);
2436	num_logops = be32_to_cpu(rhead->h_num_logops);
2437
2438	/* check the log format matches our own - else we can't recover */
2439	if (xlog_header_check_recover(log->l_mp, rhead))
2440		return -EIO;
2441
2442	trace_xfs_log_recover_record(log, rhead, pass);
2443	while ((dp < end) && num_logops) {
2444
2445		ohead = (struct xlog_op_header *)dp;
2446		dp += sizeof(*ohead);
2447		ASSERT(dp <= end);
2448
2449		/* errors will abort recovery */
2450		error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
2451						   dp, end, pass, buffer_list);
2452		if (error)
2453			return error;
2454
2455		dp += be32_to_cpu(ohead->oh_len);
2456		num_logops--;
2457	}
2458	return 0;
2459}
2460
2461/* Take all the collected deferred ops and finish them in order. */
2462static int
2463xlog_finish_defer_ops(
2464	struct xfs_mount	*mp,
2465	struct list_head	*capture_list)
2466{
2467	struct xfs_defer_capture *dfc, *next;
2468	struct xfs_trans	*tp;
2469	struct xfs_inode	*ip;
2470	int			error = 0;
2471
2472	list_for_each_entry_safe(dfc, next, capture_list, dfc_list) {
2473		struct xfs_trans_res	resv;
2474
2475		/*
2476		 * Create a new transaction reservation from the captured
2477		 * information.  Set logcount to 1 to force the new transaction
2478		 * to regrant every roll so that we can make forward progress
2479		 * in recovery no matter how full the log might be.
2480		 */
2481		resv.tr_logres = dfc->dfc_logres;
2482		resv.tr_logcount = 1;
2483		resv.tr_logflags = XFS_TRANS_PERM_LOG_RES;
2484
2485		error = xfs_trans_alloc(mp, &resv, dfc->dfc_blkres,
2486				dfc->dfc_rtxres, XFS_TRANS_RESERVE, &tp);
2487		if (error) {
2488			xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2489			return error;
2490		}
2491
2492		/*
2493		 * Transfer to this new transaction all the dfops we captured
2494		 * from recovering a single intent item.
2495		 */
2496		list_del_init(&dfc->dfc_list);
2497		xfs_defer_ops_continue(dfc, tp, &ip);
2498
2499		error = xfs_trans_commit(tp);
2500		if (ip) {
2501			xfs_iunlock(ip, XFS_ILOCK_EXCL);
2502			xfs_irele(ip);
2503		}
2504		if (error)
2505			return error;
2506	}
2507
2508	ASSERT(list_empty(capture_list));
2509	return 0;
2510}
2511
2512/* Release all the captured defer ops and capture structures in this list. */
2513static void
2514xlog_abort_defer_ops(
2515	struct xfs_mount		*mp,
2516	struct list_head		*capture_list)
2517{
2518	struct xfs_defer_capture	*dfc;
2519	struct xfs_defer_capture	*next;
2520
2521	list_for_each_entry_safe(dfc, next, capture_list, dfc_list) {
2522		list_del_init(&dfc->dfc_list);
2523		xfs_defer_ops_release(mp, dfc);
2524	}
2525}
2526/*
2527 * When this is called, all of the log intent items which did not have
2528 * corresponding log done items should be in the AIL.  What we do now
2529 * is update the data structures associated with each one.
2530 *
2531 * Since we process the log intent items in normal transactions, they
2532 * will be removed at some point after the commit.  This prevents us
2533 * from just walking down the list processing each one.  We'll use a
2534 * flag in the intent item to skip those that we've already processed
2535 * and use the AIL iteration mechanism's generation count to try to
2536 * speed this up at least a bit.
2537 *
2538 * When we start, we know that the intents are the only things in the
2539 * AIL.  As we process them, however, other items are added to the
2540 * AIL.
2541 */
2542STATIC int
2543xlog_recover_process_intents(
2544	struct xlog		*log)
2545{
2546	LIST_HEAD(capture_list);
2547	struct xfs_ail_cursor	cur;
2548	struct xfs_log_item	*lip;
2549	struct xfs_ail		*ailp;
2550	int			error = 0;
2551#if defined(DEBUG) || defined(XFS_WARN)
2552	xfs_lsn_t		last_lsn;
2553#endif
2554
2555	ailp = log->l_ailp;
2556	spin_lock(&ailp->ail_lock);
2557#if defined(DEBUG) || defined(XFS_WARN)
2558	last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
2559#endif
2560	for (lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2561	     lip != NULL;
2562	     lip = xfs_trans_ail_cursor_next(ailp, &cur)) {
2563		/*
2564		 * We're done when we see something other than an intent.
2565		 * There should be no intents left in the AIL now.
2566		 */
2567		if (!xlog_item_is_intent(lip)) {
2568#ifdef DEBUG
2569			for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2570				ASSERT(!xlog_item_is_intent(lip));
2571#endif
2572			break;
2573		}
2574
2575		/*
2576		 * We should never see a redo item with a LSN higher than
2577		 * the last transaction we found in the log at the start
2578		 * of recovery.
2579		 */
2580		ASSERT(XFS_LSN_CMP(last_lsn, lip->li_lsn) >= 0);
2581
2582		/*
2583		 * NOTE: If your intent processing routine can create more
2584		 * deferred ops, you /must/ attach them to the capture list in
2585		 * the recover routine or else those subsequent intents will be
2586		 * replayed in the wrong order!
2587		 */
2588		spin_unlock(&ailp->ail_lock);
2589		error = lip->li_ops->iop_recover(lip, &capture_list);
2590		spin_lock(&ailp->ail_lock);
2591		if (error) {
2592			trace_xlog_intent_recovery_failed(log->l_mp, error,
2593					lip->li_ops->iop_recover);
2594			break;
2595		}
2596	}
2597
2598	xfs_trans_ail_cursor_done(&cur);
2599	spin_unlock(&ailp->ail_lock);
2600	if (error)
2601		goto err;
2602
2603	error = xlog_finish_defer_ops(log->l_mp, &capture_list);
2604	if (error)
2605		goto err;
2606
2607	return 0;
2608err:
2609	xlog_abort_defer_ops(log->l_mp, &capture_list);
2610	return error;
2611}
2612
2613/*
2614 * A cancel occurs when the mount has failed and we're bailing out.
2615 * Release all pending log intent items so they don't pin the AIL.
2616 */
2617STATIC void
2618xlog_recover_cancel_intents(
2619	struct xlog		*log)
2620{
2621	struct xfs_log_item	*lip;
2622	struct xfs_ail_cursor	cur;
2623	struct xfs_ail		*ailp;
2624
2625	ailp = log->l_ailp;
2626	spin_lock(&ailp->ail_lock);
2627	lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2628	while (lip != NULL) {
2629		/*
2630		 * We're done when we see something other than an intent.
2631		 * There should be no intents left in the AIL now.
2632		 */
2633		if (!xlog_item_is_intent(lip)) {
2634#ifdef DEBUG
2635			for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2636				ASSERT(!xlog_item_is_intent(lip));
2637#endif
2638			break;
2639		}
2640
2641		spin_unlock(&ailp->ail_lock);
2642		lip->li_ops->iop_release(lip);
2643		spin_lock(&ailp->ail_lock);
2644		lip = xfs_trans_ail_cursor_next(ailp, &cur);
2645	}
2646
2647	xfs_trans_ail_cursor_done(&cur);
2648	spin_unlock(&ailp->ail_lock);
2649}
2650
2651/*
2652 * This routine performs a transaction to null out a bad inode pointer
2653 * in an agi unlinked inode hash bucket.
2654 */
2655STATIC void
2656xlog_recover_clear_agi_bucket(
2657	xfs_mount_t	*mp,
2658	xfs_agnumber_t	agno,
2659	int		bucket)
2660{
2661	xfs_trans_t	*tp;
2662	xfs_agi_t	*agi;
2663	struct xfs_buf	*agibp;
2664	int		offset;
2665	int		error;
2666
2667	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
2668	if (error)
2669		goto out_error;
2670
2671	error = xfs_read_agi(mp, tp, agno, &agibp);
2672	if (error)
2673		goto out_abort;
2674
2675	agi = agibp->b_addr;
2676	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
2677	offset = offsetof(xfs_agi_t, agi_unlinked) +
2678		 (sizeof(xfs_agino_t) * bucket);
2679	xfs_trans_log_buf(tp, agibp, offset,
2680			  (offset + sizeof(xfs_agino_t) - 1));
2681
2682	error = xfs_trans_commit(tp);
2683	if (error)
2684		goto out_error;
2685	return;
2686
2687out_abort:
2688	xfs_trans_cancel(tp);
2689out_error:
2690	xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
2691	return;
2692}
2693
2694STATIC xfs_agino_t
2695xlog_recover_process_one_iunlink(
2696	struct xfs_mount		*mp,
2697	xfs_agnumber_t			agno,
2698	xfs_agino_t			agino,
2699	int				bucket)
2700{
2701	struct xfs_buf			*ibp;
2702	struct xfs_dinode		*dip;
2703	struct xfs_inode		*ip;
2704	xfs_ino_t			ino;
2705	int				error;
2706
2707	ino = XFS_AGINO_TO_INO(mp, agno, agino);
2708	error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
2709	if (error)
2710		goto fail;
2711
2712	/*
2713	 * Get the on disk inode to find the next inode in the bucket.
2714	 */
2715	error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &ibp);
2716	if (error)
2717		goto fail_iput;
2718	dip = xfs_buf_offset(ibp, ip->i_imap.im_boffset);
2719
2720	xfs_iflags_clear(ip, XFS_IRECOVERY);
2721	ASSERT(VFS_I(ip)->i_nlink == 0);
2722	ASSERT(VFS_I(ip)->i_mode != 0);
2723
2724	/* setup for the next pass */
2725	agino = be32_to_cpu(dip->di_next_unlinked);
2726	xfs_buf_relse(ibp);
2727
2728	xfs_irele(ip);
2729	return agino;
2730
2731 fail_iput:
2732	xfs_irele(ip);
2733 fail:
2734	/*
2735	 * We can't read in the inode this bucket points to, or this inode
2736	 * is messed up.  Just ditch this bucket of inodes.  We will lose
2737	 * some inodes and space, but at least we won't hang.
2738	 *
2739	 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
2740	 * clear the inode pointer in the bucket.
2741	 */
2742	xlog_recover_clear_agi_bucket(mp, agno, bucket);
2743	return NULLAGINO;
2744}
2745
2746/*
2747 * Recover AGI unlinked lists
2748 *
2749 * This is called during recovery to process any inodes which we unlinked but
2750 * not freed when the system crashed.  These inodes will be on the lists in the
2751 * AGI blocks. What we do here is scan all the AGIs and fully truncate and free
2752 * any inodes found on the lists. Each inode is removed from the lists when it
2753 * has been fully truncated and is freed. The freeing of the inode and its
2754 * removal from the list must be atomic.
2755 *
2756 * If everything we touch in the agi processing loop is already in memory, this
2757 * loop can hold the cpu for a long time. It runs without lock contention,
2758 * memory allocation contention, the need wait for IO, etc, and so will run
2759 * until we either run out of inodes to process, run low on memory or we run out
2760 * of log space.
2761 *
2762 * This behaviour is bad for latency on single CPU and non-preemptible kernels,
2763 * and can prevent other filesystem work (such as CIL pushes) from running. This
2764 * can lead to deadlocks if the recovery process runs out of log reservation
2765 * space. Hence we need to yield the CPU when there is other kernel work
2766 * scheduled on this CPU to ensure other scheduled work can run without undue
2767 * latency.
2768 */
2769STATIC void
2770xlog_recover_process_iunlinks(
2771	struct xlog	*log)
2772{
2773	struct xfs_mount	*mp = log->l_mp;
2774	struct xfs_perag	*pag;
2775	xfs_agnumber_t		agno;
2776	struct xfs_agi		*agi;
2777	struct xfs_buf		*agibp;
2778	xfs_agino_t		agino;
2779	int			bucket;
2780	int			error;
2781
2782	for_each_perag(mp, agno, pag) {
2783		error = xfs_read_agi(mp, NULL, pag->pag_agno, &agibp);
2784		if (error) {
2785			/*
2786			 * AGI is b0rked. Don't process it.
2787			 *
2788			 * We should probably mark the filesystem as corrupt
2789			 * after we've recovered all the ag's we can....
2790			 */
2791			continue;
2792		}
2793		/*
2794		 * Unlock the buffer so that it can be acquired in the normal
2795		 * course of the transaction to truncate and free each inode.
2796		 * Because we are not racing with anyone else here for the AGI
2797		 * buffer, we don't even need to hold it locked to read the
2798		 * initial unlinked bucket entries out of the buffer. We keep
2799		 * buffer reference though, so that it stays pinned in memory
2800		 * while we need the buffer.
2801		 */
2802		agi = agibp->b_addr;
2803		xfs_buf_unlock(agibp);
2804
2805		for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
2806			agino = be32_to_cpu(agi->agi_unlinked[bucket]);
2807			while (agino != NULLAGINO) {
2808				agino = xlog_recover_process_one_iunlink(mp,
2809						pag->pag_agno, agino, bucket);
2810				cond_resched();
2811			}
2812		}
2813		xfs_buf_rele(agibp);
2814	}
2815
2816	/*
2817	 * Flush the pending unlinked inodes to ensure that the inactivations
2818	 * are fully completed on disk and the incore inodes can be reclaimed
2819	 * before we signal that recovery is complete.
2820	 */
2821	xfs_inodegc_flush(mp);
2822}
2823
2824STATIC void
2825xlog_unpack_data(
2826	struct xlog_rec_header	*rhead,
2827	char			*dp,
2828	struct xlog		*log)
2829{
2830	int			i, j, k;
2831
2832	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
2833		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
2834		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
2835		dp += BBSIZE;
2836	}
2837
2838	if (xfs_has_logv2(log->l_mp)) {
2839		xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
2840		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
2841			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2842			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
2843			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
2844			dp += BBSIZE;
2845		}
2846	}
2847}
2848
2849/*
2850 * CRC check, unpack and process a log record.
2851 */
2852STATIC int
2853xlog_recover_process(
2854	struct xlog		*log,
2855	struct hlist_head	rhash[],
2856	struct xlog_rec_header	*rhead,
2857	char			*dp,
2858	int			pass,
2859	struct list_head	*buffer_list)
2860{
2861	__le32			old_crc = rhead->h_crc;
2862	__le32			crc;
2863
2864	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
2865
2866	/*
2867	 * Nothing else to do if this is a CRC verification pass. Just return
2868	 * if this a record with a non-zero crc. Unfortunately, mkfs always
2869	 * sets old_crc to 0 so we must consider this valid even on v5 supers.
2870	 * Otherwise, return EFSBADCRC on failure so the callers up the stack
2871	 * know precisely what failed.
2872	 */
2873	if (pass == XLOG_RECOVER_CRCPASS) {
2874		if (old_crc && crc != old_crc)
2875			return -EFSBADCRC;
2876		return 0;
2877	}
2878
2879	/*
2880	 * We're in the normal recovery path. Issue a warning if and only if the
2881	 * CRC in the header is non-zero. This is an advisory warning and the
2882	 * zero CRC check prevents warnings from being emitted when upgrading
2883	 * the kernel from one that does not add CRCs by default.
2884	 */
2885	if (crc != old_crc) {
2886		if (old_crc || xfs_has_crc(log->l_mp)) {
2887			xfs_alert(log->l_mp,
2888		"log record CRC mismatch: found 0x%x, expected 0x%x.",
2889					le32_to_cpu(old_crc),
2890					le32_to_cpu(crc));
2891			xfs_hex_dump(dp, 32);
2892		}
2893
2894		/*
2895		 * If the filesystem is CRC enabled, this mismatch becomes a
2896		 * fatal log corruption failure.
2897		 */
2898		if (xfs_has_crc(log->l_mp)) {
2899			XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_LOW, log->l_mp);
2900			return -EFSCORRUPTED;
2901		}
2902	}
2903
2904	xlog_unpack_data(rhead, dp, log);
2905
2906	return xlog_recover_process_data(log, rhash, rhead, dp, pass,
2907					 buffer_list);
2908}
2909
2910STATIC int
2911xlog_valid_rec_header(
2912	struct xlog		*log,
2913	struct xlog_rec_header	*rhead,
2914	xfs_daddr_t		blkno,
2915	int			bufsize)
2916{
2917	int			hlen;
2918
2919	if (XFS_IS_CORRUPT(log->l_mp,
2920			   rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM)))
2921		return -EFSCORRUPTED;
2922	if (XFS_IS_CORRUPT(log->l_mp,
2923			   (!rhead->h_version ||
2924			   (be32_to_cpu(rhead->h_version) &
2925			    (~XLOG_VERSION_OKBITS))))) {
2926		xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
2927			__func__, be32_to_cpu(rhead->h_version));
2928		return -EFSCORRUPTED;
2929	}
2930
2931	/*
2932	 * LR body must have data (or it wouldn't have been written)
2933	 * and h_len must not be greater than LR buffer size.
2934	 */
2935	hlen = be32_to_cpu(rhead->h_len);
2936	if (XFS_IS_CORRUPT(log->l_mp, hlen <= 0 || hlen > bufsize))
2937		return -EFSCORRUPTED;
2938
2939	if (XFS_IS_CORRUPT(log->l_mp,
2940			   blkno > log->l_logBBsize || blkno > INT_MAX))
2941		return -EFSCORRUPTED;
2942	return 0;
2943}
2944
2945/*
2946 * Read the log from tail to head and process the log records found.
2947 * Handle the two cases where the tail and head are in the same cycle
2948 * and where the active portion of the log wraps around the end of
2949 * the physical log separately.  The pass parameter is passed through
2950 * to the routines called to process the data and is not looked at
2951 * here.
2952 */
2953STATIC int
2954xlog_do_recovery_pass(
2955	struct xlog		*log,
2956	xfs_daddr_t		head_blk,
2957	xfs_daddr_t		tail_blk,
2958	int			pass,
2959	xfs_daddr_t		*first_bad)	/* out: first bad log rec */
2960{
2961	xlog_rec_header_t	*rhead;
2962	xfs_daddr_t		blk_no, rblk_no;
2963	xfs_daddr_t		rhead_blk;
2964	char			*offset;
2965	char			*hbp, *dbp;
2966	int			error = 0, h_size, h_len;
2967	int			error2 = 0;
2968	int			bblks, split_bblks;
2969	int			hblks, split_hblks, wrapped_hblks;
2970	int			i;
2971	struct hlist_head	rhash[XLOG_RHASH_SIZE];
2972	LIST_HEAD		(buffer_list);
2973
2974	ASSERT(head_blk != tail_blk);
2975	blk_no = rhead_blk = tail_blk;
2976
2977	for (i = 0; i < XLOG_RHASH_SIZE; i++)
2978		INIT_HLIST_HEAD(&rhash[i]);
2979
2980	/*
2981	 * Read the header of the tail block and get the iclog buffer size from
2982	 * h_size.  Use this to tell how many sectors make up the log header.
2983	 */
2984	if (xfs_has_logv2(log->l_mp)) {
2985		/*
2986		 * When using variable length iclogs, read first sector of
2987		 * iclog header and extract the header size from it.  Get a
2988		 * new hbp that is the correct size.
2989		 */
2990		hbp = xlog_alloc_buffer(log, 1);
2991		if (!hbp)
2992			return -ENOMEM;
2993
2994		error = xlog_bread(log, tail_blk, 1, hbp, &offset);
2995		if (error)
2996			goto bread_err1;
2997
2998		rhead = (xlog_rec_header_t *)offset;
2999
3000		/*
3001		 * xfsprogs has a bug where record length is based on lsunit but
3002		 * h_size (iclog size) is hardcoded to 32k. Now that we
3003		 * unconditionally CRC verify the unmount record, this means the
3004		 * log buffer can be too small for the record and cause an
3005		 * overrun.
3006		 *
3007		 * Detect this condition here. Use lsunit for the buffer size as
3008		 * long as this looks like the mkfs case. Otherwise, return an
3009		 * error to avoid a buffer overrun.
3010		 */
3011		h_size = be32_to_cpu(rhead->h_size);
3012		h_len = be32_to_cpu(rhead->h_len);
3013		if (h_len > h_size && h_len <= log->l_mp->m_logbsize &&
3014		    rhead->h_num_logops == cpu_to_be32(1)) {
3015			xfs_warn(log->l_mp,
3016		"invalid iclog size (%d bytes), using lsunit (%d bytes)",
3017				 h_size, log->l_mp->m_logbsize);
3018			h_size = log->l_mp->m_logbsize;
3019		}
3020
3021		error = xlog_valid_rec_header(log, rhead, tail_blk, h_size);
3022		if (error)
3023			goto bread_err1;
3024
3025		hblks = xlog_logrec_hblks(log, rhead);
3026		if (hblks != 1) {
3027			kmem_free(hbp);
3028			hbp = xlog_alloc_buffer(log, hblks);
3029		}
3030	} else {
3031		ASSERT(log->l_sectBBsize == 1);
3032		hblks = 1;
3033		hbp = xlog_alloc_buffer(log, 1);
3034		h_size = XLOG_BIG_RECORD_BSIZE;
3035	}
3036
3037	if (!hbp)
3038		return -ENOMEM;
3039	dbp = xlog_alloc_buffer(log, BTOBB(h_size));
3040	if (!dbp) {
3041		kmem_free(hbp);
3042		return -ENOMEM;
3043	}
3044
3045	memset(rhash, 0, sizeof(rhash));
3046	if (tail_blk > head_blk) {
3047		/*
3048		 * Perform recovery around the end of the physical log.
3049		 * When the head is not on the same cycle number as the tail,
3050		 * we can't do a sequential recovery.
3051		 */
3052		while (blk_no < log->l_logBBsize) {
3053			/*
3054			 * Check for header wrapping around physical end-of-log
3055			 */
3056			offset = hbp;
3057			split_hblks = 0;
3058			wrapped_hblks = 0;
3059			if (blk_no + hblks <= log->l_logBBsize) {
3060				/* Read header in one read */
3061				error = xlog_bread(log, blk_no, hblks, hbp,
3062						   &offset);
3063				if (error)
3064					goto bread_err2;
3065			} else {
3066				/* This LR is split across physical log end */
3067				if (blk_no != log->l_logBBsize) {
3068					/* some data before physical log end */
3069					ASSERT(blk_no <= INT_MAX);
3070					split_hblks = log->l_logBBsize - (int)blk_no;
3071					ASSERT(split_hblks > 0);
3072					error = xlog_bread(log, blk_no,
3073							   split_hblks, hbp,
3074							   &offset);
3075					if (error)
3076						goto bread_err2;
3077				}
3078
3079				/*
3080				 * Note: this black magic still works with
3081				 * large sector sizes (non-512) only because:
3082				 * - we increased the buffer size originally
3083				 *   by 1 sector giving us enough extra space
3084				 *   for the second read;
3085				 * - the log start is guaranteed to be sector
3086				 *   aligned;
3087				 * - we read the log end (LR header start)
3088				 *   _first_, then the log start (LR header end)
3089				 *   - order is important.
3090				 */
3091				wrapped_hblks = hblks - split_hblks;
3092				error = xlog_bread_noalign(log, 0,
3093						wrapped_hblks,
3094						offset + BBTOB(split_hblks));
3095				if (error)
3096					goto bread_err2;
3097			}
3098			rhead = (xlog_rec_header_t *)offset;
3099			error = xlog_valid_rec_header(log, rhead,
3100					split_hblks ? blk_no : 0, h_size);
3101			if (error)
3102				goto bread_err2;
3103
3104			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3105			blk_no += hblks;
3106
3107			/*
3108			 * Read the log record data in multiple reads if it
3109			 * wraps around the end of the log. Note that if the
3110			 * header already wrapped, blk_no could point past the
3111			 * end of the log. The record data is contiguous in
3112			 * that case.
3113			 */
3114			if (blk_no + bblks <= log->l_logBBsize ||
3115			    blk_no >= log->l_logBBsize) {
3116				rblk_no = xlog_wrap_logbno(log, blk_no);
3117				error = xlog_bread(log, rblk_no, bblks, dbp,
3118						   &offset);
3119				if (error)
3120					goto bread_err2;
3121			} else {
3122				/* This log record is split across the
3123				 * physical end of log */
3124				offset = dbp;
3125				split_bblks = 0;
3126				if (blk_no != log->l_logBBsize) {
3127					/* some data is before the physical
3128					 * end of log */
3129					ASSERT(!wrapped_hblks);
3130					ASSERT(blk_no <= INT_MAX);
3131					split_bblks =
3132						log->l_logBBsize - (int)blk_no;
3133					ASSERT(split_bblks > 0);
3134					error = xlog_bread(log, blk_no,
3135							split_bblks, dbp,
3136							&offset);
3137					if (error)
3138						goto bread_err2;
3139				}
3140
3141				/*
3142				 * Note: this black magic still works with
3143				 * large sector sizes (non-512) only because:
3144				 * - we increased the buffer size originally
3145				 *   by 1 sector giving us enough extra space
3146				 *   for the second read;
3147				 * - the log start is guaranteed to be sector
3148				 *   aligned;
3149				 * - we read the log end (LR header start)
3150				 *   _first_, then the log start (LR header end)
3151				 *   - order is important.
3152				 */
3153				error = xlog_bread_noalign(log, 0,
3154						bblks - split_bblks,
3155						offset + BBTOB(split_bblks));
3156				if (error)
3157					goto bread_err2;
3158			}
3159
3160			error = xlog_recover_process(log, rhash, rhead, offset,
3161						     pass, &buffer_list);
3162			if (error)
3163				goto bread_err2;
3164
3165			blk_no += bblks;
3166			rhead_blk = blk_no;
3167		}
3168
3169		ASSERT(blk_no >= log->l_logBBsize);
3170		blk_no -= log->l_logBBsize;
3171		rhead_blk = blk_no;
3172	}
3173
3174	/* read first part of physical log */
3175	while (blk_no < head_blk) {
3176		error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3177		if (error)
3178			goto bread_err2;
3179
3180		rhead = (xlog_rec_header_t *)offset;
3181		error = xlog_valid_rec_header(log, rhead, blk_no, h_size);
3182		if (error)
3183			goto bread_err2;
3184
3185		/* blocks in data section */
3186		bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3187		error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3188				   &offset);
3189		if (error)
3190			goto bread_err2;
3191
3192		error = xlog_recover_process(log, rhash, rhead, offset, pass,
3193					     &buffer_list);
3194		if (error)
3195			goto bread_err2;
3196
3197		blk_no += bblks + hblks;
3198		rhead_blk = blk_no;
3199	}
3200
3201 bread_err2:
3202	kmem_free(dbp);
3203 bread_err1:
3204	kmem_free(hbp);
3205
3206	/*
3207	 * Submit buffers that have been added from the last record processed,
3208	 * regardless of error status.
3209	 */
3210	if (!list_empty(&buffer_list))
3211		error2 = xfs_buf_delwri_submit(&buffer_list);
3212
3213	if (error && first_bad)
3214		*first_bad = rhead_blk;
3215
3216	/*
3217	 * Transactions are freed at commit time but transactions without commit
3218	 * records on disk are never committed. Free any that may be left in the
3219	 * hash table.
3220	 */
3221	for (i = 0; i < XLOG_RHASH_SIZE; i++) {
3222		struct hlist_node	*tmp;
3223		struct xlog_recover	*trans;
3224
3225		hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
3226			xlog_recover_free_trans(trans);
3227	}
3228
3229	return error ? error : error2;
3230}
3231
3232/*
3233 * Do the recovery of the log.  We actually do this in two phases.
3234 * The two passes are necessary in order to implement the function
3235 * of cancelling a record written into the log.  The first pass
3236 * determines those things which have been cancelled, and the
3237 * second pass replays log items normally except for those which
3238 * have been cancelled.  The handling of the replay and cancellations
3239 * takes place in the log item type specific routines.
3240 *
3241 * The table of items which have cancel records in the log is allocated
3242 * and freed at this level, since only here do we know when all of
3243 * the log recovery has been completed.
3244 */
3245STATIC int
3246xlog_do_log_recovery(
3247	struct xlog	*log,
3248	xfs_daddr_t	head_blk,
3249	xfs_daddr_t	tail_blk)
3250{
3251	int		error, i;
3252
3253	ASSERT(head_blk != tail_blk);
3254
3255	/*
3256	 * First do a pass to find all of the cancelled buf log items.
3257	 * Store them in the buf_cancel_table for use in the second pass.
3258	 */
3259	log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
3260						 sizeof(struct list_head),
3261						 0);
3262	for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3263		INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
3264
3265	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3266				      XLOG_RECOVER_PASS1, NULL);
3267	if (error != 0) {
3268		kmem_free(log->l_buf_cancel_table);
3269		log->l_buf_cancel_table = NULL;
3270		return error;
3271	}
3272	/*
3273	 * Then do a second pass to actually recover the items in the log.
3274	 * When it is complete free the table of buf cancel items.
3275	 */
3276	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3277				      XLOG_RECOVER_PASS2, NULL);
3278#ifdef DEBUG
3279	if (!error) {
3280		int	i;
3281
3282		for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3283			ASSERT(list_empty(&log->l_buf_cancel_table[i]));
3284	}
3285#endif	/* DEBUG */
3286
3287	kmem_free(log->l_buf_cancel_table);
3288	log->l_buf_cancel_table = NULL;
3289
3290	return error;
3291}
3292
3293/*
3294 * Do the actual recovery
3295 */
3296STATIC int
3297xlog_do_recover(
3298	struct xlog		*log,
3299	xfs_daddr_t		head_blk,
3300	xfs_daddr_t		tail_blk)
3301{
3302	struct xfs_mount	*mp = log->l_mp;
3303	struct xfs_buf		*bp = mp->m_sb_bp;
3304	struct xfs_sb		*sbp = &mp->m_sb;
3305	int			error;
3306
3307	trace_xfs_log_recover(log, head_blk, tail_blk);
3308
3309	/*
3310	 * First replay the images in the log.
3311	 */
3312	error = xlog_do_log_recovery(log, head_blk, tail_blk);
3313	if (error)
3314		return error;
3315
3316	if (xlog_is_shutdown(log))
3317		return -EIO;
3318
3319	/*
3320	 * We now update the tail_lsn since much of the recovery has completed
3321	 * and there may be space available to use.  If there were no extent
3322	 * or iunlinks, we can free up the entire log and set the tail_lsn to
3323	 * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3324	 * lsn of the last known good LR on disk.  If there are extent frees
3325	 * or iunlinks they will have some entries in the AIL; so we look at
3326	 * the AIL to determine how to set the tail_lsn.
3327	 */
3328	xlog_assign_tail_lsn(mp);
3329
3330	/*
3331	 * Now that we've finished replaying all buffer and inode updates,
3332	 * re-read the superblock and reverify it.
3333	 */
3334	xfs_buf_lock(bp);
3335	xfs_buf_hold(bp);
3336	error = _xfs_buf_read(bp, XBF_READ);
3337	if (error) {
3338		if (!xlog_is_shutdown(log)) {
3339			xfs_buf_ioerror_alert(bp, __this_address);
3340			ASSERT(0);
3341		}
3342		xfs_buf_relse(bp);
3343		return error;
3344	}
3345
3346	/* Convert superblock from on-disk format */
3347	xfs_sb_from_disk(sbp, bp->b_addr);
3348	xfs_buf_relse(bp);
3349
3350	/* re-initialise in-core superblock and geometry structures */
3351	mp->m_features |= xfs_sb_version_to_features(sbp);
3352	xfs_reinit_percpu_counters(mp);
3353	error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
3354	if (error) {
3355		xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
3356		return error;
3357	}
3358	mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
3359
3360	xlog_recover_check_summary(log);
3361
3362	/* Normal transactions can now occur */
3363	clear_bit(XLOG_ACTIVE_RECOVERY, &log->l_opstate);
3364	return 0;
3365}
3366
3367/*
3368 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3369 *
3370 * Return error or zero.
3371 */
3372int
3373xlog_recover(
3374	struct xlog	*log)
3375{
3376	xfs_daddr_t	head_blk, tail_blk;
3377	int		error;
3378
3379	/* find the tail of the log */
3380	error = xlog_find_tail(log, &head_blk, &tail_blk);
3381	if (error)
3382		return error;
3383
3384	/*
3385	 * The superblock was read before the log was available and thus the LSN
3386	 * could not be verified. Check the superblock LSN against the current
3387	 * LSN now that it's known.
3388	 */
3389	if (xfs_has_crc(log->l_mp) &&
3390	    !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
3391		return -EINVAL;
3392
3393	if (tail_blk != head_blk) {
3394		/* There used to be a comment here:
3395		 *
3396		 * disallow recovery on read-only mounts.  note -- mount
3397		 * checks for ENOSPC and turns it into an intelligent
3398		 * error message.
3399		 * ...but this is no longer true.  Now, unless you specify
3400		 * NORECOVERY (in which case this function would never be
3401		 * called), we just go ahead and recover.  We do this all
3402		 * under the vfs layer, so we can get away with it unless
3403		 * the device itself is read-only, in which case we fail.
3404		 */
3405		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3406			return error;
3407		}
3408
3409		/*
3410		 * Version 5 superblock log feature mask validation. We know the
3411		 * log is dirty so check if there are any unknown log features
3412		 * in what we need to recover. If there are unknown features
3413		 * (e.g. unsupported transactions, then simply reject the
3414		 * attempt at recovery before touching anything.
3415		 */
3416		if (xfs_sb_is_v5(&log->l_mp->m_sb) &&
3417		    xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
3418					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
3419			xfs_warn(log->l_mp,
3420"Superblock has unknown incompatible log features (0x%x) enabled.",
3421				(log->l_mp->m_sb.sb_features_log_incompat &
3422					XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
3423			xfs_warn(log->l_mp,
3424"The log can not be fully and/or safely recovered by this kernel.");
3425			xfs_warn(log->l_mp,
3426"Please recover the log on a kernel that supports the unknown features.");
3427			return -EINVAL;
3428		}
3429
3430		/*
3431		 * Delay log recovery if the debug hook is set. This is debug
3432		 * instrumentation to coordinate simulation of I/O failures with
3433		 * log recovery.
3434		 */
3435		if (xfs_globals.log_recovery_delay) {
3436			xfs_notice(log->l_mp,
3437				"Delaying log recovery for %d seconds.",
3438				xfs_globals.log_recovery_delay);
3439			msleep(xfs_globals.log_recovery_delay * 1000);
3440		}
3441
3442		xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
3443				log->l_mp->m_logname ? log->l_mp->m_logname
3444						     : "internal");
3445
3446		error = xlog_do_recover(log, head_blk, tail_blk);
3447		set_bit(XLOG_RECOVERY_NEEDED, &log->l_opstate);
3448	}
3449	return error;
3450}
3451
3452/*
3453 * In the first part of recovery we replay inodes and buffers and build up the
3454 * list of intents which need to be processed. Here we process the intents and
3455 * clean up the on disk unlinked inode lists. This is separated from the first
3456 * part of recovery so that the root and real-time bitmap inodes can be read in
3457 * from disk in between the two stages.  This is necessary so that we can free
3458 * space in the real-time portion of the file system.
3459 */
3460int
3461xlog_recover_finish(
3462	struct xlog	*log)
3463{
3464	int	error;
3465
3466	error = xlog_recover_process_intents(log);
3467	if (error) {
3468		/*
3469		 * Cancel all the unprocessed intent items now so that we don't
3470		 * leave them pinned in the AIL.  This can cause the AIL to
3471		 * livelock on the pinned item if anyone tries to push the AIL
3472		 * (inode reclaim does this) before we get around to
3473		 * xfs_log_mount_cancel.
3474		 */
3475		xlog_recover_cancel_intents(log);
3476		xfs_alert(log->l_mp, "Failed to recover intents");
3477		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
3478		return error;
3479	}
3480
3481	/*
3482	 * Sync the log to get all the intents out of the AIL.  This isn't
3483	 * absolutely necessary, but it helps in case the unlink transactions
3484	 * would have problems pushing the intents out of the way.
3485	 */
3486	xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3487
3488	/*
3489	 * Now that we've recovered the log and all the intents, we can clear
3490	 * the log incompat feature bits in the superblock because there's no
3491	 * longer anything to protect.  We rely on the AIL push to write out the
3492	 * updated superblock after everything else.
3493	 */
3494	if (xfs_clear_incompat_log_features(log->l_mp)) {
3495		error = xfs_sync_sb(log->l_mp, false);
3496		if (error < 0) {
3497			xfs_alert(log->l_mp,
3498	"Failed to clear log incompat features on recovery");
3499			return error;
3500		}
3501	}
3502
3503	xlog_recover_process_iunlinks(log);
3504	xlog_recover_check_summary(log);
3505	return 0;
3506}
3507
3508void
3509xlog_recover_cancel(
3510	struct xlog	*log)
3511{
3512	if (xlog_recovery_needed(log))
3513		xlog_recover_cancel_intents(log);
3514}
3515
3516#if defined(DEBUG)
3517/*
3518 * Read all of the agf and agi counters and check that they
3519 * are consistent with the superblock counters.
3520 */
3521STATIC void
3522xlog_recover_check_summary(
3523	struct xlog		*log)
3524{
3525	struct xfs_mount	*mp = log->l_mp;
3526	struct xfs_perag	*pag;
3527	struct xfs_buf		*agfbp;
3528	struct xfs_buf		*agibp;
3529	xfs_agnumber_t		agno;
3530	uint64_t		freeblks;
3531	uint64_t		itotal;
3532	uint64_t		ifree;
3533	int			error;
3534
3535	mp = log->l_mp;
3536
3537	freeblks = 0LL;
3538	itotal = 0LL;
3539	ifree = 0LL;
3540	for_each_perag(mp, agno, pag) {
3541		error = xfs_read_agf(mp, NULL, pag->pag_agno, 0, &agfbp);
3542		if (error) {
3543			xfs_alert(mp, "%s agf read failed agno %d error %d",
3544						__func__, pag->pag_agno, error);
3545		} else {
3546			struct xfs_agf	*agfp = agfbp->b_addr;
3547
3548			freeblks += be32_to_cpu(agfp->agf_freeblks) +
3549				    be32_to_cpu(agfp->agf_flcount);
3550			xfs_buf_relse(agfbp);
3551		}
3552
3553		error = xfs_read_agi(mp, NULL, pag->pag_agno, &agibp);
3554		if (error) {
3555			xfs_alert(mp, "%s agi read failed agno %d error %d",
3556						__func__, pag->pag_agno, error);
3557		} else {
3558			struct xfs_agi	*agi = agibp->b_addr;
3559
3560			itotal += be32_to_cpu(agi->agi_count);
3561			ifree += be32_to_cpu(agi->agi_freecount);
3562			xfs_buf_relse(agibp);
3563		}
3564	}
3565}
3566#endif /* DEBUG */
3567