xref: /openbmc/linux/fs/xfs/xfs_log_recover.c (revision 9ac8d3fb)
1 /*
2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_dir2_sf.h"
35 #include "xfs_attr_sf.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
38 #include "xfs_inode_item.h"
39 #include "xfs_imap.h"
40 #include "xfs_alloc.h"
41 #include "xfs_ialloc.h"
42 #include "xfs_log_priv.h"
43 #include "xfs_buf_item.h"
44 #include "xfs_log_recover.h"
45 #include "xfs_extfree_item.h"
46 #include "xfs_trans_priv.h"
47 #include "xfs_quota.h"
48 #include "xfs_rw.h"
49 #include "xfs_utils.h"
50 
51 STATIC int	xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
52 STATIC int	xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
53 STATIC void	xlog_recover_insert_item_backq(xlog_recover_item_t **q,
54 					       xlog_recover_item_t *item);
55 #if defined(DEBUG)
56 STATIC void	xlog_recover_check_summary(xlog_t *);
57 STATIC void	xlog_recover_check_ail(xfs_mount_t *, xfs_log_item_t *, int);
58 #else
59 #define	xlog_recover_check_summary(log)
60 #define	xlog_recover_check_ail(mp, lip, gen)
61 #endif
62 
63 
64 /*
65  * Sector aligned buffer routines for buffer create/read/write/access
66  */
67 
68 #define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs)	\
69 	( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
70 	((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
71 #define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno)	((bno) & ~(log)->l_sectbb_mask)
72 
73 xfs_buf_t *
74 xlog_get_bp(
75 	xlog_t		*log,
76 	int		num_bblks)
77 {
78 	ASSERT(num_bblks > 0);
79 
80 	if (log->l_sectbb_log) {
81 		if (num_bblks > 1)
82 			num_bblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
83 		num_bblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, num_bblks);
84 	}
85 	return xfs_buf_get_noaddr(BBTOB(num_bblks), log->l_mp->m_logdev_targp);
86 }
87 
88 void
89 xlog_put_bp(
90 	xfs_buf_t	*bp)
91 {
92 	xfs_buf_free(bp);
93 }
94 
95 
96 /*
97  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
98  */
99 int
100 xlog_bread(
101 	xlog_t		*log,
102 	xfs_daddr_t	blk_no,
103 	int		nbblks,
104 	xfs_buf_t	*bp)
105 {
106 	int		error;
107 
108 	if (log->l_sectbb_log) {
109 		blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
110 		nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
111 	}
112 
113 	ASSERT(nbblks > 0);
114 	ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
115 	ASSERT(bp);
116 
117 	XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
118 	XFS_BUF_READ(bp);
119 	XFS_BUF_BUSY(bp);
120 	XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
121 	XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
122 
123 	xfsbdstrat(log->l_mp, bp);
124 	error = xfs_iowait(bp);
125 	if (error)
126 		xfs_ioerror_alert("xlog_bread", log->l_mp,
127 				  bp, XFS_BUF_ADDR(bp));
128 	return error;
129 }
130 
131 /*
132  * Write out the buffer at the given block for the given number of blocks.
133  * The buffer is kept locked across the write and is returned locked.
134  * This can only be used for synchronous log writes.
135  */
136 STATIC int
137 xlog_bwrite(
138 	xlog_t		*log,
139 	xfs_daddr_t	blk_no,
140 	int		nbblks,
141 	xfs_buf_t	*bp)
142 {
143 	int		error;
144 
145 	if (log->l_sectbb_log) {
146 		blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
147 		nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
148 	}
149 
150 	ASSERT(nbblks > 0);
151 	ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
152 
153 	XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
154 	XFS_BUF_ZEROFLAGS(bp);
155 	XFS_BUF_BUSY(bp);
156 	XFS_BUF_HOLD(bp);
157 	XFS_BUF_PSEMA(bp, PRIBIO);
158 	XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
159 	XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
160 
161 	if ((error = xfs_bwrite(log->l_mp, bp)))
162 		xfs_ioerror_alert("xlog_bwrite", log->l_mp,
163 				  bp, XFS_BUF_ADDR(bp));
164 	return error;
165 }
166 
167 STATIC xfs_caddr_t
168 xlog_align(
169 	xlog_t		*log,
170 	xfs_daddr_t	blk_no,
171 	int		nbblks,
172 	xfs_buf_t	*bp)
173 {
174 	xfs_caddr_t	ptr;
175 
176 	if (!log->l_sectbb_log)
177 		return XFS_BUF_PTR(bp);
178 
179 	ptr = XFS_BUF_PTR(bp) + BBTOB((int)blk_no & log->l_sectbb_mask);
180 	ASSERT(XFS_BUF_SIZE(bp) >=
181 		BBTOB(nbblks + (blk_no & log->l_sectbb_mask)));
182 	return ptr;
183 }
184 
185 #ifdef DEBUG
186 /*
187  * dump debug superblock and log record information
188  */
189 STATIC void
190 xlog_header_check_dump(
191 	xfs_mount_t		*mp,
192 	xlog_rec_header_t	*head)
193 {
194 	int			b;
195 
196 	cmn_err(CE_DEBUG, "%s:  SB : uuid = ", __func__);
197 	for (b = 0; b < 16; b++)
198 		cmn_err(CE_DEBUG, "%02x", ((uchar_t *)&mp->m_sb.sb_uuid)[b]);
199 	cmn_err(CE_DEBUG, ", fmt = %d\n", XLOG_FMT);
200 	cmn_err(CE_DEBUG, "    log : uuid = ");
201 	for (b = 0; b < 16; b++)
202 		cmn_err(CE_DEBUG, "%02x",((uchar_t *)&head->h_fs_uuid)[b]);
203 	cmn_err(CE_DEBUG, ", fmt = %d\n", be32_to_cpu(head->h_fmt));
204 }
205 #else
206 #define xlog_header_check_dump(mp, head)
207 #endif
208 
209 /*
210  * check log record header for recovery
211  */
212 STATIC int
213 xlog_header_check_recover(
214 	xfs_mount_t		*mp,
215 	xlog_rec_header_t	*head)
216 {
217 	ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
218 
219 	/*
220 	 * IRIX doesn't write the h_fmt field and leaves it zeroed
221 	 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
222 	 * a dirty log created in IRIX.
223 	 */
224 	if (unlikely(be32_to_cpu(head->h_fmt) != XLOG_FMT)) {
225 		xlog_warn(
226 	"XFS: dirty log written in incompatible format - can't recover");
227 		xlog_header_check_dump(mp, head);
228 		XFS_ERROR_REPORT("xlog_header_check_recover(1)",
229 				 XFS_ERRLEVEL_HIGH, mp);
230 		return XFS_ERROR(EFSCORRUPTED);
231 	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
232 		xlog_warn(
233 	"XFS: dirty log entry has mismatched uuid - can't recover");
234 		xlog_header_check_dump(mp, head);
235 		XFS_ERROR_REPORT("xlog_header_check_recover(2)",
236 				 XFS_ERRLEVEL_HIGH, mp);
237 		return XFS_ERROR(EFSCORRUPTED);
238 	}
239 	return 0;
240 }
241 
242 /*
243  * read the head block of the log and check the header
244  */
245 STATIC int
246 xlog_header_check_mount(
247 	xfs_mount_t		*mp,
248 	xlog_rec_header_t	*head)
249 {
250 	ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
251 
252 	if (uuid_is_nil(&head->h_fs_uuid)) {
253 		/*
254 		 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
255 		 * h_fs_uuid is nil, we assume this log was last mounted
256 		 * by IRIX and continue.
257 		 */
258 		xlog_warn("XFS: nil uuid in log - IRIX style log");
259 	} else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
260 		xlog_warn("XFS: log has mismatched uuid - can't recover");
261 		xlog_header_check_dump(mp, head);
262 		XFS_ERROR_REPORT("xlog_header_check_mount",
263 				 XFS_ERRLEVEL_HIGH, mp);
264 		return XFS_ERROR(EFSCORRUPTED);
265 	}
266 	return 0;
267 }
268 
269 STATIC void
270 xlog_recover_iodone(
271 	struct xfs_buf	*bp)
272 {
273 	xfs_mount_t	*mp;
274 
275 	ASSERT(XFS_BUF_FSPRIVATE(bp, void *));
276 
277 	if (XFS_BUF_GETERROR(bp)) {
278 		/*
279 		 * We're not going to bother about retrying
280 		 * this during recovery. One strike!
281 		 */
282 		mp = XFS_BUF_FSPRIVATE(bp, xfs_mount_t *);
283 		xfs_ioerror_alert("xlog_recover_iodone",
284 				  mp, bp, XFS_BUF_ADDR(bp));
285 		xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
286 	}
287 	XFS_BUF_SET_FSPRIVATE(bp, NULL);
288 	XFS_BUF_CLR_IODONE_FUNC(bp);
289 	xfs_biodone(bp);
290 }
291 
292 /*
293  * This routine finds (to an approximation) the first block in the physical
294  * log which contains the given cycle.  It uses a binary search algorithm.
295  * Note that the algorithm can not be perfect because the disk will not
296  * necessarily be perfect.
297  */
298 STATIC int
299 xlog_find_cycle_start(
300 	xlog_t		*log,
301 	xfs_buf_t	*bp,
302 	xfs_daddr_t	first_blk,
303 	xfs_daddr_t	*last_blk,
304 	uint		cycle)
305 {
306 	xfs_caddr_t	offset;
307 	xfs_daddr_t	mid_blk;
308 	uint		mid_cycle;
309 	int		error;
310 
311 	mid_blk = BLK_AVG(first_blk, *last_blk);
312 	while (mid_blk != first_blk && mid_blk != *last_blk) {
313 		if ((error = xlog_bread(log, mid_blk, 1, bp)))
314 			return error;
315 		offset = xlog_align(log, mid_blk, 1, bp);
316 		mid_cycle = xlog_get_cycle(offset);
317 		if (mid_cycle == cycle) {
318 			*last_blk = mid_blk;
319 			/* last_half_cycle == mid_cycle */
320 		} else {
321 			first_blk = mid_blk;
322 			/* first_half_cycle == mid_cycle */
323 		}
324 		mid_blk = BLK_AVG(first_blk, *last_blk);
325 	}
326 	ASSERT((mid_blk == first_blk && mid_blk+1 == *last_blk) ||
327 	       (mid_blk == *last_blk && mid_blk-1 == first_blk));
328 
329 	return 0;
330 }
331 
332 /*
333  * Check that the range of blocks does not contain the cycle number
334  * given.  The scan needs to occur from front to back and the ptr into the
335  * region must be updated since a later routine will need to perform another
336  * test.  If the region is completely good, we end up returning the same
337  * last block number.
338  *
339  * Set blkno to -1 if we encounter no errors.  This is an invalid block number
340  * since we don't ever expect logs to get this large.
341  */
342 STATIC int
343 xlog_find_verify_cycle(
344 	xlog_t		*log,
345 	xfs_daddr_t	start_blk,
346 	int		nbblks,
347 	uint		stop_on_cycle_no,
348 	xfs_daddr_t	*new_blk)
349 {
350 	xfs_daddr_t	i, j;
351 	uint		cycle;
352 	xfs_buf_t	*bp;
353 	xfs_daddr_t	bufblks;
354 	xfs_caddr_t	buf = NULL;
355 	int		error = 0;
356 
357 	bufblks = 1 << ffs(nbblks);
358 
359 	while (!(bp = xlog_get_bp(log, bufblks))) {
360 		/* can't get enough memory to do everything in one big buffer */
361 		bufblks >>= 1;
362 		if (bufblks <= log->l_sectbb_log)
363 			return ENOMEM;
364 	}
365 
366 	for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
367 		int	bcount;
368 
369 		bcount = min(bufblks, (start_blk + nbblks - i));
370 
371 		if ((error = xlog_bread(log, i, bcount, bp)))
372 			goto out;
373 
374 		buf = xlog_align(log, i, bcount, bp);
375 		for (j = 0; j < bcount; j++) {
376 			cycle = xlog_get_cycle(buf);
377 			if (cycle == stop_on_cycle_no) {
378 				*new_blk = i+j;
379 				goto out;
380 			}
381 
382 			buf += BBSIZE;
383 		}
384 	}
385 
386 	*new_blk = -1;
387 
388 out:
389 	xlog_put_bp(bp);
390 	return error;
391 }
392 
393 /*
394  * Potentially backup over partial log record write.
395  *
396  * In the typical case, last_blk is the number of the block directly after
397  * a good log record.  Therefore, we subtract one to get the block number
398  * of the last block in the given buffer.  extra_bblks contains the number
399  * of blocks we would have read on a previous read.  This happens when the
400  * last log record is split over the end of the physical log.
401  *
402  * extra_bblks is the number of blocks potentially verified on a previous
403  * call to this routine.
404  */
405 STATIC int
406 xlog_find_verify_log_record(
407 	xlog_t			*log,
408 	xfs_daddr_t		start_blk,
409 	xfs_daddr_t		*last_blk,
410 	int			extra_bblks)
411 {
412 	xfs_daddr_t		i;
413 	xfs_buf_t		*bp;
414 	xfs_caddr_t		offset = NULL;
415 	xlog_rec_header_t	*head = NULL;
416 	int			error = 0;
417 	int			smallmem = 0;
418 	int			num_blks = *last_blk - start_blk;
419 	int			xhdrs;
420 
421 	ASSERT(start_blk != 0 || *last_blk != start_blk);
422 
423 	if (!(bp = xlog_get_bp(log, num_blks))) {
424 		if (!(bp = xlog_get_bp(log, 1)))
425 			return ENOMEM;
426 		smallmem = 1;
427 	} else {
428 		if ((error = xlog_bread(log, start_blk, num_blks, bp)))
429 			goto out;
430 		offset = xlog_align(log, start_blk, num_blks, bp);
431 		offset += ((num_blks - 1) << BBSHIFT);
432 	}
433 
434 	for (i = (*last_blk) - 1; i >= 0; i--) {
435 		if (i < start_blk) {
436 			/* valid log record not found */
437 			xlog_warn(
438 		"XFS: Log inconsistent (didn't find previous header)");
439 			ASSERT(0);
440 			error = XFS_ERROR(EIO);
441 			goto out;
442 		}
443 
444 		if (smallmem) {
445 			if ((error = xlog_bread(log, i, 1, bp)))
446 				goto out;
447 			offset = xlog_align(log, i, 1, bp);
448 		}
449 
450 		head = (xlog_rec_header_t *)offset;
451 
452 		if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(head->h_magicno))
453 			break;
454 
455 		if (!smallmem)
456 			offset -= BBSIZE;
457 	}
458 
459 	/*
460 	 * We hit the beginning of the physical log & still no header.  Return
461 	 * to caller.  If caller can handle a return of -1, then this routine
462 	 * will be called again for the end of the physical log.
463 	 */
464 	if (i == -1) {
465 		error = -1;
466 		goto out;
467 	}
468 
469 	/*
470 	 * We have the final block of the good log (the first block
471 	 * of the log record _before_ the head. So we check the uuid.
472 	 */
473 	if ((error = xlog_header_check_mount(log->l_mp, head)))
474 		goto out;
475 
476 	/*
477 	 * We may have found a log record header before we expected one.
478 	 * last_blk will be the 1st block # with a given cycle #.  We may end
479 	 * up reading an entire log record.  In this case, we don't want to
480 	 * reset last_blk.  Only when last_blk points in the middle of a log
481 	 * record do we update last_blk.
482 	 */
483 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
484 		uint	h_size = be32_to_cpu(head->h_size);
485 
486 		xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
487 		if (h_size % XLOG_HEADER_CYCLE_SIZE)
488 			xhdrs++;
489 	} else {
490 		xhdrs = 1;
491 	}
492 
493 	if (*last_blk - i + extra_bblks !=
494 	    BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
495 		*last_blk = i;
496 
497 out:
498 	xlog_put_bp(bp);
499 	return error;
500 }
501 
502 /*
503  * Head is defined to be the point of the log where the next log write
504  * write could go.  This means that incomplete LR writes at the end are
505  * eliminated when calculating the head.  We aren't guaranteed that previous
506  * LR have complete transactions.  We only know that a cycle number of
507  * current cycle number -1 won't be present in the log if we start writing
508  * from our current block number.
509  *
510  * last_blk contains the block number of the first block with a given
511  * cycle number.
512  *
513  * Return: zero if normal, non-zero if error.
514  */
515 STATIC int
516 xlog_find_head(
517 	xlog_t 		*log,
518 	xfs_daddr_t	*return_head_blk)
519 {
520 	xfs_buf_t	*bp;
521 	xfs_caddr_t	offset;
522 	xfs_daddr_t	new_blk, first_blk, start_blk, last_blk, head_blk;
523 	int		num_scan_bblks;
524 	uint		first_half_cycle, last_half_cycle;
525 	uint		stop_on_cycle;
526 	int		error, log_bbnum = log->l_logBBsize;
527 
528 	/* Is the end of the log device zeroed? */
529 	if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
530 		*return_head_blk = first_blk;
531 
532 		/* Is the whole lot zeroed? */
533 		if (!first_blk) {
534 			/* Linux XFS shouldn't generate totally zeroed logs -
535 			 * mkfs etc write a dummy unmount record to a fresh
536 			 * log so we can store the uuid in there
537 			 */
538 			xlog_warn("XFS: totally zeroed log");
539 		}
540 
541 		return 0;
542 	} else if (error) {
543 		xlog_warn("XFS: empty log check failed");
544 		return error;
545 	}
546 
547 	first_blk = 0;			/* get cycle # of 1st block */
548 	bp = xlog_get_bp(log, 1);
549 	if (!bp)
550 		return ENOMEM;
551 	if ((error = xlog_bread(log, 0, 1, bp)))
552 		goto bp_err;
553 	offset = xlog_align(log, 0, 1, bp);
554 	first_half_cycle = xlog_get_cycle(offset);
555 
556 	last_blk = head_blk = log_bbnum - 1;	/* get cycle # of last block */
557 	if ((error = xlog_bread(log, last_blk, 1, bp)))
558 		goto bp_err;
559 	offset = xlog_align(log, last_blk, 1, bp);
560 	last_half_cycle = xlog_get_cycle(offset);
561 	ASSERT(last_half_cycle != 0);
562 
563 	/*
564 	 * If the 1st half cycle number is equal to the last half cycle number,
565 	 * then the entire log is stamped with the same cycle number.  In this
566 	 * case, head_blk can't be set to zero (which makes sense).  The below
567 	 * math doesn't work out properly with head_blk equal to zero.  Instead,
568 	 * we set it to log_bbnum which is an invalid block number, but this
569 	 * value makes the math correct.  If head_blk doesn't changed through
570 	 * all the tests below, *head_blk is set to zero at the very end rather
571 	 * than log_bbnum.  In a sense, log_bbnum and zero are the same block
572 	 * in a circular file.
573 	 */
574 	if (first_half_cycle == last_half_cycle) {
575 		/*
576 		 * In this case we believe that the entire log should have
577 		 * cycle number last_half_cycle.  We need to scan backwards
578 		 * from the end verifying that there are no holes still
579 		 * containing last_half_cycle - 1.  If we find such a hole,
580 		 * then the start of that hole will be the new head.  The
581 		 * simple case looks like
582 		 *        x | x ... | x - 1 | x
583 		 * Another case that fits this picture would be
584 		 *        x | x + 1 | x ... | x
585 		 * In this case the head really is somewhere at the end of the
586 		 * log, as one of the latest writes at the beginning was
587 		 * incomplete.
588 		 * One more case is
589 		 *        x | x + 1 | x ... | x - 1 | x
590 		 * This is really the combination of the above two cases, and
591 		 * the head has to end up at the start of the x-1 hole at the
592 		 * end of the log.
593 		 *
594 		 * In the 256k log case, we will read from the beginning to the
595 		 * end of the log and search for cycle numbers equal to x-1.
596 		 * We don't worry about the x+1 blocks that we encounter,
597 		 * because we know that they cannot be the head since the log
598 		 * started with x.
599 		 */
600 		head_blk = log_bbnum;
601 		stop_on_cycle = last_half_cycle - 1;
602 	} else {
603 		/*
604 		 * In this case we want to find the first block with cycle
605 		 * number matching last_half_cycle.  We expect the log to be
606 		 * some variation on
607 		 *        x + 1 ... | x ...
608 		 * The first block with cycle number x (last_half_cycle) will
609 		 * be where the new head belongs.  First we do a binary search
610 		 * for the first occurrence of last_half_cycle.  The binary
611 		 * search may not be totally accurate, so then we scan back
612 		 * from there looking for occurrences of last_half_cycle before
613 		 * us.  If that backwards scan wraps around the beginning of
614 		 * the log, then we look for occurrences of last_half_cycle - 1
615 		 * at the end of the log.  The cases we're looking for look
616 		 * like
617 		 *        x + 1 ... | x | x + 1 | x ...
618 		 *                               ^ binary search stopped here
619 		 * or
620 		 *        x + 1 ... | x ... | x - 1 | x
621 		 *        <---------> less than scan distance
622 		 */
623 		stop_on_cycle = last_half_cycle;
624 		if ((error = xlog_find_cycle_start(log, bp, first_blk,
625 						&head_blk, last_half_cycle)))
626 			goto bp_err;
627 	}
628 
629 	/*
630 	 * Now validate the answer.  Scan back some number of maximum possible
631 	 * blocks and make sure each one has the expected cycle number.  The
632 	 * maximum is determined by the total possible amount of buffering
633 	 * in the in-core log.  The following number can be made tighter if
634 	 * we actually look at the block size of the filesystem.
635 	 */
636 	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
637 	if (head_blk >= num_scan_bblks) {
638 		/*
639 		 * We are guaranteed that the entire check can be performed
640 		 * in one buffer.
641 		 */
642 		start_blk = head_blk - num_scan_bblks;
643 		if ((error = xlog_find_verify_cycle(log,
644 						start_blk, num_scan_bblks,
645 						stop_on_cycle, &new_blk)))
646 			goto bp_err;
647 		if (new_blk != -1)
648 			head_blk = new_blk;
649 	} else {		/* need to read 2 parts of log */
650 		/*
651 		 * We are going to scan backwards in the log in two parts.
652 		 * First we scan the physical end of the log.  In this part
653 		 * of the log, we are looking for blocks with cycle number
654 		 * last_half_cycle - 1.
655 		 * If we find one, then we know that the log starts there, as
656 		 * we've found a hole that didn't get written in going around
657 		 * the end of the physical log.  The simple case for this is
658 		 *        x + 1 ... | x ... | x - 1 | x
659 		 *        <---------> less than scan distance
660 		 * If all of the blocks at the end of the log have cycle number
661 		 * last_half_cycle, then we check the blocks at the start of
662 		 * the log looking for occurrences of last_half_cycle.  If we
663 		 * find one, then our current estimate for the location of the
664 		 * first occurrence of last_half_cycle is wrong and we move
665 		 * back to the hole we've found.  This case looks like
666 		 *        x + 1 ... | x | x + 1 | x ...
667 		 *                               ^ binary search stopped here
668 		 * Another case we need to handle that only occurs in 256k
669 		 * logs is
670 		 *        x + 1 ... | x ... | x+1 | x ...
671 		 *                   ^ binary search stops here
672 		 * In a 256k log, the scan at the end of the log will see the
673 		 * x + 1 blocks.  We need to skip past those since that is
674 		 * certainly not the head of the log.  By searching for
675 		 * last_half_cycle-1 we accomplish that.
676 		 */
677 		start_blk = log_bbnum - num_scan_bblks + head_blk;
678 		ASSERT(head_blk <= INT_MAX &&
679 			(xfs_daddr_t) num_scan_bblks - head_blk >= 0);
680 		if ((error = xlog_find_verify_cycle(log, start_blk,
681 					num_scan_bblks - (int)head_blk,
682 					(stop_on_cycle - 1), &new_blk)))
683 			goto bp_err;
684 		if (new_blk != -1) {
685 			head_blk = new_blk;
686 			goto bad_blk;
687 		}
688 
689 		/*
690 		 * Scan beginning of log now.  The last part of the physical
691 		 * log is good.  This scan needs to verify that it doesn't find
692 		 * the last_half_cycle.
693 		 */
694 		start_blk = 0;
695 		ASSERT(head_blk <= INT_MAX);
696 		if ((error = xlog_find_verify_cycle(log,
697 					start_blk, (int)head_blk,
698 					stop_on_cycle, &new_blk)))
699 			goto bp_err;
700 		if (new_blk != -1)
701 			head_blk = new_blk;
702 	}
703 
704  bad_blk:
705 	/*
706 	 * Now we need to make sure head_blk is not pointing to a block in
707 	 * the middle of a log record.
708 	 */
709 	num_scan_bblks = XLOG_REC_SHIFT(log);
710 	if (head_blk >= num_scan_bblks) {
711 		start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
712 
713 		/* start ptr at last block ptr before head_blk */
714 		if ((error = xlog_find_verify_log_record(log, start_blk,
715 							&head_blk, 0)) == -1) {
716 			error = XFS_ERROR(EIO);
717 			goto bp_err;
718 		} else if (error)
719 			goto bp_err;
720 	} else {
721 		start_blk = 0;
722 		ASSERT(head_blk <= INT_MAX);
723 		if ((error = xlog_find_verify_log_record(log, start_blk,
724 							&head_blk, 0)) == -1) {
725 			/* We hit the beginning of the log during our search */
726 			start_blk = log_bbnum - num_scan_bblks + head_blk;
727 			new_blk = log_bbnum;
728 			ASSERT(start_blk <= INT_MAX &&
729 				(xfs_daddr_t) log_bbnum-start_blk >= 0);
730 			ASSERT(head_blk <= INT_MAX);
731 			if ((error = xlog_find_verify_log_record(log,
732 							start_blk, &new_blk,
733 							(int)head_blk)) == -1) {
734 				error = XFS_ERROR(EIO);
735 				goto bp_err;
736 			} else if (error)
737 				goto bp_err;
738 			if (new_blk != log_bbnum)
739 				head_blk = new_blk;
740 		} else if (error)
741 			goto bp_err;
742 	}
743 
744 	xlog_put_bp(bp);
745 	if (head_blk == log_bbnum)
746 		*return_head_blk = 0;
747 	else
748 		*return_head_blk = head_blk;
749 	/*
750 	 * When returning here, we have a good block number.  Bad block
751 	 * means that during a previous crash, we didn't have a clean break
752 	 * from cycle number N to cycle number N-1.  In this case, we need
753 	 * to find the first block with cycle number N-1.
754 	 */
755 	return 0;
756 
757  bp_err:
758 	xlog_put_bp(bp);
759 
760 	if (error)
761 	    xlog_warn("XFS: failed to find log head");
762 	return error;
763 }
764 
765 /*
766  * Find the sync block number or the tail of the log.
767  *
768  * This will be the block number of the last record to have its
769  * associated buffers synced to disk.  Every log record header has
770  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
771  * to get a sync block number.  The only concern is to figure out which
772  * log record header to believe.
773  *
774  * The following algorithm uses the log record header with the largest
775  * lsn.  The entire log record does not need to be valid.  We only care
776  * that the header is valid.
777  *
778  * We could speed up search by using current head_blk buffer, but it is not
779  * available.
780  */
781 int
782 xlog_find_tail(
783 	xlog_t			*log,
784 	xfs_daddr_t		*head_blk,
785 	xfs_daddr_t		*tail_blk)
786 {
787 	xlog_rec_header_t	*rhead;
788 	xlog_op_header_t	*op_head;
789 	xfs_caddr_t		offset = NULL;
790 	xfs_buf_t		*bp;
791 	int			error, i, found;
792 	xfs_daddr_t		umount_data_blk;
793 	xfs_daddr_t		after_umount_blk;
794 	xfs_lsn_t		tail_lsn;
795 	int			hblks;
796 
797 	found = 0;
798 
799 	/*
800 	 * Find previous log record
801 	 */
802 	if ((error = xlog_find_head(log, head_blk)))
803 		return error;
804 
805 	bp = xlog_get_bp(log, 1);
806 	if (!bp)
807 		return ENOMEM;
808 	if (*head_blk == 0) {				/* special case */
809 		if ((error = xlog_bread(log, 0, 1, bp)))
810 			goto bread_err;
811 		offset = xlog_align(log, 0, 1, bp);
812 		if (xlog_get_cycle(offset) == 0) {
813 			*tail_blk = 0;
814 			/* leave all other log inited values alone */
815 			goto exit;
816 		}
817 	}
818 
819 	/*
820 	 * Search backwards looking for log record header block
821 	 */
822 	ASSERT(*head_blk < INT_MAX);
823 	for (i = (int)(*head_blk) - 1; i >= 0; i--) {
824 		if ((error = xlog_bread(log, i, 1, bp)))
825 			goto bread_err;
826 		offset = xlog_align(log, i, 1, bp);
827 		if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(*(__be32 *)offset)) {
828 			found = 1;
829 			break;
830 		}
831 	}
832 	/*
833 	 * If we haven't found the log record header block, start looking
834 	 * again from the end of the physical log.  XXXmiken: There should be
835 	 * a check here to make sure we didn't search more than N blocks in
836 	 * the previous code.
837 	 */
838 	if (!found) {
839 		for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
840 			if ((error = xlog_bread(log, i, 1, bp)))
841 				goto bread_err;
842 			offset = xlog_align(log, i, 1, bp);
843 			if (XLOG_HEADER_MAGIC_NUM ==
844 			    be32_to_cpu(*(__be32 *)offset)) {
845 				found = 2;
846 				break;
847 			}
848 		}
849 	}
850 	if (!found) {
851 		xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
852 		ASSERT(0);
853 		return XFS_ERROR(EIO);
854 	}
855 
856 	/* find blk_no of tail of log */
857 	rhead = (xlog_rec_header_t *)offset;
858 	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
859 
860 	/*
861 	 * Reset log values according to the state of the log when we
862 	 * crashed.  In the case where head_blk == 0, we bump curr_cycle
863 	 * one because the next write starts a new cycle rather than
864 	 * continuing the cycle of the last good log record.  At this
865 	 * point we have guaranteed that all partial log records have been
866 	 * accounted for.  Therefore, we know that the last good log record
867 	 * written was complete and ended exactly on the end boundary
868 	 * of the physical log.
869 	 */
870 	log->l_prev_block = i;
871 	log->l_curr_block = (int)*head_blk;
872 	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
873 	if (found == 2)
874 		log->l_curr_cycle++;
875 	log->l_tail_lsn = be64_to_cpu(rhead->h_tail_lsn);
876 	log->l_last_sync_lsn = be64_to_cpu(rhead->h_lsn);
877 	log->l_grant_reserve_cycle = log->l_curr_cycle;
878 	log->l_grant_reserve_bytes = BBTOB(log->l_curr_block);
879 	log->l_grant_write_cycle = log->l_curr_cycle;
880 	log->l_grant_write_bytes = BBTOB(log->l_curr_block);
881 
882 	/*
883 	 * Look for unmount record.  If we find it, then we know there
884 	 * was a clean unmount.  Since 'i' could be the last block in
885 	 * the physical log, we convert to a log block before comparing
886 	 * to the head_blk.
887 	 *
888 	 * Save the current tail lsn to use to pass to
889 	 * xlog_clear_stale_blocks() below.  We won't want to clear the
890 	 * unmount record if there is one, so we pass the lsn of the
891 	 * unmount record rather than the block after it.
892 	 */
893 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
894 		int	h_size = be32_to_cpu(rhead->h_size);
895 		int	h_version = be32_to_cpu(rhead->h_version);
896 
897 		if ((h_version & XLOG_VERSION_2) &&
898 		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
899 			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
900 			if (h_size % XLOG_HEADER_CYCLE_SIZE)
901 				hblks++;
902 		} else {
903 			hblks = 1;
904 		}
905 	} else {
906 		hblks = 1;
907 	}
908 	after_umount_blk = (i + hblks + (int)
909 		BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
910 	tail_lsn = log->l_tail_lsn;
911 	if (*head_blk == after_umount_blk &&
912 	    be32_to_cpu(rhead->h_num_logops) == 1) {
913 		umount_data_blk = (i + hblks) % log->l_logBBsize;
914 		if ((error = xlog_bread(log, umount_data_blk, 1, bp))) {
915 			goto bread_err;
916 		}
917 		offset = xlog_align(log, umount_data_blk, 1, bp);
918 		op_head = (xlog_op_header_t *)offset;
919 		if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
920 			/*
921 			 * Set tail and last sync so that newly written
922 			 * log records will point recovery to after the
923 			 * current unmount record.
924 			 */
925 			log->l_tail_lsn =
926 				xlog_assign_lsn(log->l_curr_cycle,
927 						after_umount_blk);
928 			log->l_last_sync_lsn =
929 				xlog_assign_lsn(log->l_curr_cycle,
930 						after_umount_blk);
931 			*tail_blk = after_umount_blk;
932 
933 			/*
934 			 * Note that the unmount was clean. If the unmount
935 			 * was not clean, we need to know this to rebuild the
936 			 * superblock counters from the perag headers if we
937 			 * have a filesystem using non-persistent counters.
938 			 */
939 			log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
940 		}
941 	}
942 
943 	/*
944 	 * Make sure that there are no blocks in front of the head
945 	 * with the same cycle number as the head.  This can happen
946 	 * because we allow multiple outstanding log writes concurrently,
947 	 * and the later writes might make it out before earlier ones.
948 	 *
949 	 * We use the lsn from before modifying it so that we'll never
950 	 * overwrite the unmount record after a clean unmount.
951 	 *
952 	 * Do this only if we are going to recover the filesystem
953 	 *
954 	 * NOTE: This used to say "if (!readonly)"
955 	 * However on Linux, we can & do recover a read-only filesystem.
956 	 * We only skip recovery if NORECOVERY is specified on mount,
957 	 * in which case we would not be here.
958 	 *
959 	 * But... if the -device- itself is readonly, just skip this.
960 	 * We can't recover this device anyway, so it won't matter.
961 	 */
962 	if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
963 		error = xlog_clear_stale_blocks(log, tail_lsn);
964 	}
965 
966 bread_err:
967 exit:
968 	xlog_put_bp(bp);
969 
970 	if (error)
971 		xlog_warn("XFS: failed to locate log tail");
972 	return error;
973 }
974 
975 /*
976  * Is the log zeroed at all?
977  *
978  * The last binary search should be changed to perform an X block read
979  * once X becomes small enough.  You can then search linearly through
980  * the X blocks.  This will cut down on the number of reads we need to do.
981  *
982  * If the log is partially zeroed, this routine will pass back the blkno
983  * of the first block with cycle number 0.  It won't have a complete LR
984  * preceding it.
985  *
986  * Return:
987  *	0  => the log is completely written to
988  *	-1 => use *blk_no as the first block of the log
989  *	>0 => error has occurred
990  */
991 STATIC int
992 xlog_find_zeroed(
993 	xlog_t		*log,
994 	xfs_daddr_t	*blk_no)
995 {
996 	xfs_buf_t	*bp;
997 	xfs_caddr_t	offset;
998 	uint	        first_cycle, last_cycle;
999 	xfs_daddr_t	new_blk, last_blk, start_blk;
1000 	xfs_daddr_t     num_scan_bblks;
1001 	int	        error, log_bbnum = log->l_logBBsize;
1002 
1003 	*blk_no = 0;
1004 
1005 	/* check totally zeroed log */
1006 	bp = xlog_get_bp(log, 1);
1007 	if (!bp)
1008 		return ENOMEM;
1009 	if ((error = xlog_bread(log, 0, 1, bp)))
1010 		goto bp_err;
1011 	offset = xlog_align(log, 0, 1, bp);
1012 	first_cycle = xlog_get_cycle(offset);
1013 	if (first_cycle == 0) {		/* completely zeroed log */
1014 		*blk_no = 0;
1015 		xlog_put_bp(bp);
1016 		return -1;
1017 	}
1018 
1019 	/* check partially zeroed log */
1020 	if ((error = xlog_bread(log, log_bbnum-1, 1, bp)))
1021 		goto bp_err;
1022 	offset = xlog_align(log, log_bbnum-1, 1, bp);
1023 	last_cycle = xlog_get_cycle(offset);
1024 	if (last_cycle != 0) {		/* log completely written to */
1025 		xlog_put_bp(bp);
1026 		return 0;
1027 	} else if (first_cycle != 1) {
1028 		/*
1029 		 * If the cycle of the last block is zero, the cycle of
1030 		 * the first block must be 1. If it's not, maybe we're
1031 		 * not looking at a log... Bail out.
1032 		 */
1033 		xlog_warn("XFS: Log inconsistent or not a log (last==0, first!=1)");
1034 		return XFS_ERROR(EINVAL);
1035 	}
1036 
1037 	/* we have a partially zeroed log */
1038 	last_blk = log_bbnum-1;
1039 	if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1040 		goto bp_err;
1041 
1042 	/*
1043 	 * Validate the answer.  Because there is no way to guarantee that
1044 	 * the entire log is made up of log records which are the same size,
1045 	 * we scan over the defined maximum blocks.  At this point, the maximum
1046 	 * is not chosen to mean anything special.   XXXmiken
1047 	 */
1048 	num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1049 	ASSERT(num_scan_bblks <= INT_MAX);
1050 
1051 	if (last_blk < num_scan_bblks)
1052 		num_scan_bblks = last_blk;
1053 	start_blk = last_blk - num_scan_bblks;
1054 
1055 	/*
1056 	 * We search for any instances of cycle number 0 that occur before
1057 	 * our current estimate of the head.  What we're trying to detect is
1058 	 *        1 ... | 0 | 1 | 0...
1059 	 *                       ^ binary search ends here
1060 	 */
1061 	if ((error = xlog_find_verify_cycle(log, start_blk,
1062 					 (int)num_scan_bblks, 0, &new_blk)))
1063 		goto bp_err;
1064 	if (new_blk != -1)
1065 		last_blk = new_blk;
1066 
1067 	/*
1068 	 * Potentially backup over partial log record write.  We don't need
1069 	 * to search the end of the log because we know it is zero.
1070 	 */
1071 	if ((error = xlog_find_verify_log_record(log, start_blk,
1072 				&last_blk, 0)) == -1) {
1073 	    error = XFS_ERROR(EIO);
1074 	    goto bp_err;
1075 	} else if (error)
1076 	    goto bp_err;
1077 
1078 	*blk_no = last_blk;
1079 bp_err:
1080 	xlog_put_bp(bp);
1081 	if (error)
1082 		return error;
1083 	return -1;
1084 }
1085 
1086 /*
1087  * These are simple subroutines used by xlog_clear_stale_blocks() below
1088  * to initialize a buffer full of empty log record headers and write
1089  * them into the log.
1090  */
1091 STATIC void
1092 xlog_add_record(
1093 	xlog_t			*log,
1094 	xfs_caddr_t		buf,
1095 	int			cycle,
1096 	int			block,
1097 	int			tail_cycle,
1098 	int			tail_block)
1099 {
1100 	xlog_rec_header_t	*recp = (xlog_rec_header_t *)buf;
1101 
1102 	memset(buf, 0, BBSIZE);
1103 	recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1104 	recp->h_cycle = cpu_to_be32(cycle);
1105 	recp->h_version = cpu_to_be32(
1106 			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1107 	recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1108 	recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1109 	recp->h_fmt = cpu_to_be32(XLOG_FMT);
1110 	memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1111 }
1112 
1113 STATIC int
1114 xlog_write_log_records(
1115 	xlog_t		*log,
1116 	int		cycle,
1117 	int		start_block,
1118 	int		blocks,
1119 	int		tail_cycle,
1120 	int		tail_block)
1121 {
1122 	xfs_caddr_t	offset;
1123 	xfs_buf_t	*bp;
1124 	int		balign, ealign;
1125 	int		sectbb = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
1126 	int		end_block = start_block + blocks;
1127 	int		bufblks;
1128 	int		error = 0;
1129 	int		i, j = 0;
1130 
1131 	bufblks = 1 << ffs(blocks);
1132 	while (!(bp = xlog_get_bp(log, bufblks))) {
1133 		bufblks >>= 1;
1134 		if (bufblks <= log->l_sectbb_log)
1135 			return ENOMEM;
1136 	}
1137 
1138 	/* We may need to do a read at the start to fill in part of
1139 	 * the buffer in the starting sector not covered by the first
1140 	 * write below.
1141 	 */
1142 	balign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, start_block);
1143 	if (balign != start_block) {
1144 		if ((error = xlog_bread(log, start_block, 1, bp))) {
1145 			xlog_put_bp(bp);
1146 			return error;
1147 		}
1148 		j = start_block - balign;
1149 	}
1150 
1151 	for (i = start_block; i < end_block; i += bufblks) {
1152 		int		bcount, endcount;
1153 
1154 		bcount = min(bufblks, end_block - start_block);
1155 		endcount = bcount - j;
1156 
1157 		/* We may need to do a read at the end to fill in part of
1158 		 * the buffer in the final sector not covered by the write.
1159 		 * If this is the same sector as the above read, skip it.
1160 		 */
1161 		ealign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, end_block);
1162 		if (j == 0 && (start_block + endcount > ealign)) {
1163 			offset = XFS_BUF_PTR(bp);
1164 			balign = BBTOB(ealign - start_block);
1165 			error = XFS_BUF_SET_PTR(bp, offset + balign,
1166 						BBTOB(sectbb));
1167 			if (!error)
1168 				error = xlog_bread(log, ealign, sectbb, bp);
1169 			if (!error)
1170 				error = XFS_BUF_SET_PTR(bp, offset, bufblks);
1171 			if (error)
1172 				break;
1173 		}
1174 
1175 		offset = xlog_align(log, start_block, endcount, bp);
1176 		for (; j < endcount; j++) {
1177 			xlog_add_record(log, offset, cycle, i+j,
1178 					tail_cycle, tail_block);
1179 			offset += BBSIZE;
1180 		}
1181 		error = xlog_bwrite(log, start_block, endcount, bp);
1182 		if (error)
1183 			break;
1184 		start_block += endcount;
1185 		j = 0;
1186 	}
1187 	xlog_put_bp(bp);
1188 	return error;
1189 }
1190 
1191 /*
1192  * This routine is called to blow away any incomplete log writes out
1193  * in front of the log head.  We do this so that we won't become confused
1194  * if we come up, write only a little bit more, and then crash again.
1195  * If we leave the partial log records out there, this situation could
1196  * cause us to think those partial writes are valid blocks since they
1197  * have the current cycle number.  We get rid of them by overwriting them
1198  * with empty log records with the old cycle number rather than the
1199  * current one.
1200  *
1201  * The tail lsn is passed in rather than taken from
1202  * the log so that we will not write over the unmount record after a
1203  * clean unmount in a 512 block log.  Doing so would leave the log without
1204  * any valid log records in it until a new one was written.  If we crashed
1205  * during that time we would not be able to recover.
1206  */
1207 STATIC int
1208 xlog_clear_stale_blocks(
1209 	xlog_t		*log,
1210 	xfs_lsn_t	tail_lsn)
1211 {
1212 	int		tail_cycle, head_cycle;
1213 	int		tail_block, head_block;
1214 	int		tail_distance, max_distance;
1215 	int		distance;
1216 	int		error;
1217 
1218 	tail_cycle = CYCLE_LSN(tail_lsn);
1219 	tail_block = BLOCK_LSN(tail_lsn);
1220 	head_cycle = log->l_curr_cycle;
1221 	head_block = log->l_curr_block;
1222 
1223 	/*
1224 	 * Figure out the distance between the new head of the log
1225 	 * and the tail.  We want to write over any blocks beyond the
1226 	 * head that we may have written just before the crash, but
1227 	 * we don't want to overwrite the tail of the log.
1228 	 */
1229 	if (head_cycle == tail_cycle) {
1230 		/*
1231 		 * The tail is behind the head in the physical log,
1232 		 * so the distance from the head to the tail is the
1233 		 * distance from the head to the end of the log plus
1234 		 * the distance from the beginning of the log to the
1235 		 * tail.
1236 		 */
1237 		if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1238 			XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1239 					 XFS_ERRLEVEL_LOW, log->l_mp);
1240 			return XFS_ERROR(EFSCORRUPTED);
1241 		}
1242 		tail_distance = tail_block + (log->l_logBBsize - head_block);
1243 	} else {
1244 		/*
1245 		 * The head is behind the tail in the physical log,
1246 		 * so the distance from the head to the tail is just
1247 		 * the tail block minus the head block.
1248 		 */
1249 		if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1250 			XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1251 					 XFS_ERRLEVEL_LOW, log->l_mp);
1252 			return XFS_ERROR(EFSCORRUPTED);
1253 		}
1254 		tail_distance = tail_block - head_block;
1255 	}
1256 
1257 	/*
1258 	 * If the head is right up against the tail, we can't clear
1259 	 * anything.
1260 	 */
1261 	if (tail_distance <= 0) {
1262 		ASSERT(tail_distance == 0);
1263 		return 0;
1264 	}
1265 
1266 	max_distance = XLOG_TOTAL_REC_SHIFT(log);
1267 	/*
1268 	 * Take the smaller of the maximum amount of outstanding I/O
1269 	 * we could have and the distance to the tail to clear out.
1270 	 * We take the smaller so that we don't overwrite the tail and
1271 	 * we don't waste all day writing from the head to the tail
1272 	 * for no reason.
1273 	 */
1274 	max_distance = MIN(max_distance, tail_distance);
1275 
1276 	if ((head_block + max_distance) <= log->l_logBBsize) {
1277 		/*
1278 		 * We can stomp all the blocks we need to without
1279 		 * wrapping around the end of the log.  Just do it
1280 		 * in a single write.  Use the cycle number of the
1281 		 * current cycle minus one so that the log will look like:
1282 		 *     n ... | n - 1 ...
1283 		 */
1284 		error = xlog_write_log_records(log, (head_cycle - 1),
1285 				head_block, max_distance, tail_cycle,
1286 				tail_block);
1287 		if (error)
1288 			return error;
1289 	} else {
1290 		/*
1291 		 * We need to wrap around the end of the physical log in
1292 		 * order to clear all the blocks.  Do it in two separate
1293 		 * I/Os.  The first write should be from the head to the
1294 		 * end of the physical log, and it should use the current
1295 		 * cycle number minus one just like above.
1296 		 */
1297 		distance = log->l_logBBsize - head_block;
1298 		error = xlog_write_log_records(log, (head_cycle - 1),
1299 				head_block, distance, tail_cycle,
1300 				tail_block);
1301 
1302 		if (error)
1303 			return error;
1304 
1305 		/*
1306 		 * Now write the blocks at the start of the physical log.
1307 		 * This writes the remainder of the blocks we want to clear.
1308 		 * It uses the current cycle number since we're now on the
1309 		 * same cycle as the head so that we get:
1310 		 *    n ... n ... | n - 1 ...
1311 		 *    ^^^^^ blocks we're writing
1312 		 */
1313 		distance = max_distance - (log->l_logBBsize - head_block);
1314 		error = xlog_write_log_records(log, head_cycle, 0, distance,
1315 				tail_cycle, tail_block);
1316 		if (error)
1317 			return error;
1318 	}
1319 
1320 	return 0;
1321 }
1322 
1323 /******************************************************************************
1324  *
1325  *		Log recover routines
1326  *
1327  ******************************************************************************
1328  */
1329 
1330 STATIC xlog_recover_t *
1331 xlog_recover_find_tid(
1332 	xlog_recover_t		*q,
1333 	xlog_tid_t		tid)
1334 {
1335 	xlog_recover_t		*p = q;
1336 
1337 	while (p != NULL) {
1338 		if (p->r_log_tid == tid)
1339 		    break;
1340 		p = p->r_next;
1341 	}
1342 	return p;
1343 }
1344 
1345 STATIC void
1346 xlog_recover_put_hashq(
1347 	xlog_recover_t		**q,
1348 	xlog_recover_t		*trans)
1349 {
1350 	trans->r_next = *q;
1351 	*q = trans;
1352 }
1353 
1354 STATIC void
1355 xlog_recover_add_item(
1356 	xlog_recover_item_t	**itemq)
1357 {
1358 	xlog_recover_item_t	*item;
1359 
1360 	item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1361 	xlog_recover_insert_item_backq(itemq, item);
1362 }
1363 
1364 STATIC int
1365 xlog_recover_add_to_cont_trans(
1366 	xlog_recover_t		*trans,
1367 	xfs_caddr_t		dp,
1368 	int			len)
1369 {
1370 	xlog_recover_item_t	*item;
1371 	xfs_caddr_t		ptr, old_ptr;
1372 	int			old_len;
1373 
1374 	item = trans->r_itemq;
1375 	if (item == NULL) {
1376 		/* finish copying rest of trans header */
1377 		xlog_recover_add_item(&trans->r_itemq);
1378 		ptr = (xfs_caddr_t) &trans->r_theader +
1379 				sizeof(xfs_trans_header_t) - len;
1380 		memcpy(ptr, dp, len); /* d, s, l */
1381 		return 0;
1382 	}
1383 	item = item->ri_prev;
1384 
1385 	old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1386 	old_len = item->ri_buf[item->ri_cnt-1].i_len;
1387 
1388 	ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1389 	memcpy(&ptr[old_len], dp, len); /* d, s, l */
1390 	item->ri_buf[item->ri_cnt-1].i_len += len;
1391 	item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1392 	return 0;
1393 }
1394 
1395 /*
1396  * The next region to add is the start of a new region.  It could be
1397  * a whole region or it could be the first part of a new region.  Because
1398  * of this, the assumption here is that the type and size fields of all
1399  * format structures fit into the first 32 bits of the structure.
1400  *
1401  * This works because all regions must be 32 bit aligned.  Therefore, we
1402  * either have both fields or we have neither field.  In the case we have
1403  * neither field, the data part of the region is zero length.  We only have
1404  * a log_op_header and can throw away the header since a new one will appear
1405  * later.  If we have at least 4 bytes, then we can determine how many regions
1406  * will appear in the current log item.
1407  */
1408 STATIC int
1409 xlog_recover_add_to_trans(
1410 	xlog_recover_t		*trans,
1411 	xfs_caddr_t		dp,
1412 	int			len)
1413 {
1414 	xfs_inode_log_format_t	*in_f;			/* any will do */
1415 	xlog_recover_item_t	*item;
1416 	xfs_caddr_t		ptr;
1417 
1418 	if (!len)
1419 		return 0;
1420 	item = trans->r_itemq;
1421 	if (item == NULL) {
1422 		ASSERT(*(uint *)dp == XFS_TRANS_HEADER_MAGIC);
1423 		if (len == sizeof(xfs_trans_header_t))
1424 			xlog_recover_add_item(&trans->r_itemq);
1425 		memcpy(&trans->r_theader, dp, len); /* d, s, l */
1426 		return 0;
1427 	}
1428 
1429 	ptr = kmem_alloc(len, KM_SLEEP);
1430 	memcpy(ptr, dp, len);
1431 	in_f = (xfs_inode_log_format_t *)ptr;
1432 
1433 	if (item->ri_prev->ri_total != 0 &&
1434 	     item->ri_prev->ri_total == item->ri_prev->ri_cnt) {
1435 		xlog_recover_add_item(&trans->r_itemq);
1436 	}
1437 	item = trans->r_itemq;
1438 	item = item->ri_prev;
1439 
1440 	if (item->ri_total == 0) {		/* first region to be added */
1441 		item->ri_total	= in_f->ilf_size;
1442 		ASSERT(item->ri_total <= XLOG_MAX_REGIONS_IN_ITEM);
1443 		item->ri_buf = kmem_zalloc((item->ri_total *
1444 					    sizeof(xfs_log_iovec_t)), KM_SLEEP);
1445 	}
1446 	ASSERT(item->ri_total > item->ri_cnt);
1447 	/* Description region is ri_buf[0] */
1448 	item->ri_buf[item->ri_cnt].i_addr = ptr;
1449 	item->ri_buf[item->ri_cnt].i_len  = len;
1450 	item->ri_cnt++;
1451 	return 0;
1452 }
1453 
1454 STATIC void
1455 xlog_recover_new_tid(
1456 	xlog_recover_t		**q,
1457 	xlog_tid_t		tid,
1458 	xfs_lsn_t		lsn)
1459 {
1460 	xlog_recover_t		*trans;
1461 
1462 	trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1463 	trans->r_log_tid   = tid;
1464 	trans->r_lsn	   = lsn;
1465 	xlog_recover_put_hashq(q, trans);
1466 }
1467 
1468 STATIC int
1469 xlog_recover_unlink_tid(
1470 	xlog_recover_t		**q,
1471 	xlog_recover_t		*trans)
1472 {
1473 	xlog_recover_t		*tp;
1474 	int			found = 0;
1475 
1476 	ASSERT(trans != NULL);
1477 	if (trans == *q) {
1478 		*q = (*q)->r_next;
1479 	} else {
1480 		tp = *q;
1481 		while (tp) {
1482 			if (tp->r_next == trans) {
1483 				found = 1;
1484 				break;
1485 			}
1486 			tp = tp->r_next;
1487 		}
1488 		if (!found) {
1489 			xlog_warn(
1490 			     "XFS: xlog_recover_unlink_tid: trans not found");
1491 			ASSERT(0);
1492 			return XFS_ERROR(EIO);
1493 		}
1494 		tp->r_next = tp->r_next->r_next;
1495 	}
1496 	return 0;
1497 }
1498 
1499 STATIC void
1500 xlog_recover_insert_item_backq(
1501 	xlog_recover_item_t	**q,
1502 	xlog_recover_item_t	*item)
1503 {
1504 	if (*q == NULL) {
1505 		item->ri_prev = item->ri_next = item;
1506 		*q = item;
1507 	} else {
1508 		item->ri_next		= *q;
1509 		item->ri_prev		= (*q)->ri_prev;
1510 		(*q)->ri_prev		= item;
1511 		item->ri_prev->ri_next	= item;
1512 	}
1513 }
1514 
1515 STATIC void
1516 xlog_recover_insert_item_frontq(
1517 	xlog_recover_item_t	**q,
1518 	xlog_recover_item_t	*item)
1519 {
1520 	xlog_recover_insert_item_backq(q, item);
1521 	*q = item;
1522 }
1523 
1524 STATIC int
1525 xlog_recover_reorder_trans(
1526 	xlog_recover_t		*trans)
1527 {
1528 	xlog_recover_item_t	*first_item, *itemq, *itemq_next;
1529 	xfs_buf_log_format_t	*buf_f;
1530 	ushort			flags = 0;
1531 
1532 	first_item = itemq = trans->r_itemq;
1533 	trans->r_itemq = NULL;
1534 	do {
1535 		itemq_next = itemq->ri_next;
1536 		buf_f = (xfs_buf_log_format_t *)itemq->ri_buf[0].i_addr;
1537 
1538 		switch (ITEM_TYPE(itemq)) {
1539 		case XFS_LI_BUF:
1540 			flags = buf_f->blf_flags;
1541 			if (!(flags & XFS_BLI_CANCEL)) {
1542 				xlog_recover_insert_item_frontq(&trans->r_itemq,
1543 								itemq);
1544 				break;
1545 			}
1546 		case XFS_LI_INODE:
1547 		case XFS_LI_DQUOT:
1548 		case XFS_LI_QUOTAOFF:
1549 		case XFS_LI_EFD:
1550 		case XFS_LI_EFI:
1551 			xlog_recover_insert_item_backq(&trans->r_itemq, itemq);
1552 			break;
1553 		default:
1554 			xlog_warn(
1555 	"XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1556 			ASSERT(0);
1557 			return XFS_ERROR(EIO);
1558 		}
1559 		itemq = itemq_next;
1560 	} while (first_item != itemq);
1561 	return 0;
1562 }
1563 
1564 /*
1565  * Build up the table of buf cancel records so that we don't replay
1566  * cancelled data in the second pass.  For buffer records that are
1567  * not cancel records, there is nothing to do here so we just return.
1568  *
1569  * If we get a cancel record which is already in the table, this indicates
1570  * that the buffer was cancelled multiple times.  In order to ensure
1571  * that during pass 2 we keep the record in the table until we reach its
1572  * last occurrence in the log, we keep a reference count in the cancel
1573  * record in the table to tell us how many times we expect to see this
1574  * record during the second pass.
1575  */
1576 STATIC void
1577 xlog_recover_do_buffer_pass1(
1578 	xlog_t			*log,
1579 	xfs_buf_log_format_t	*buf_f)
1580 {
1581 	xfs_buf_cancel_t	*bcp;
1582 	xfs_buf_cancel_t	*nextp;
1583 	xfs_buf_cancel_t	*prevp;
1584 	xfs_buf_cancel_t	**bucket;
1585 	xfs_daddr_t		blkno = 0;
1586 	uint			len = 0;
1587 	ushort			flags = 0;
1588 
1589 	switch (buf_f->blf_type) {
1590 	case XFS_LI_BUF:
1591 		blkno = buf_f->blf_blkno;
1592 		len = buf_f->blf_len;
1593 		flags = buf_f->blf_flags;
1594 		break;
1595 	}
1596 
1597 	/*
1598 	 * If this isn't a cancel buffer item, then just return.
1599 	 */
1600 	if (!(flags & XFS_BLI_CANCEL))
1601 		return;
1602 
1603 	/*
1604 	 * Insert an xfs_buf_cancel record into the hash table of
1605 	 * them.  If there is already an identical record, bump
1606 	 * its reference count.
1607 	 */
1608 	bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1609 					  XLOG_BC_TABLE_SIZE];
1610 	/*
1611 	 * If the hash bucket is empty then just insert a new record into
1612 	 * the bucket.
1613 	 */
1614 	if (*bucket == NULL) {
1615 		bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1616 						     KM_SLEEP);
1617 		bcp->bc_blkno = blkno;
1618 		bcp->bc_len = len;
1619 		bcp->bc_refcount = 1;
1620 		bcp->bc_next = NULL;
1621 		*bucket = bcp;
1622 		return;
1623 	}
1624 
1625 	/*
1626 	 * The hash bucket is not empty, so search for duplicates of our
1627 	 * record.  If we find one them just bump its refcount.  If not
1628 	 * then add us at the end of the list.
1629 	 */
1630 	prevp = NULL;
1631 	nextp = *bucket;
1632 	while (nextp != NULL) {
1633 		if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1634 			nextp->bc_refcount++;
1635 			return;
1636 		}
1637 		prevp = nextp;
1638 		nextp = nextp->bc_next;
1639 	}
1640 	ASSERT(prevp != NULL);
1641 	bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1642 					     KM_SLEEP);
1643 	bcp->bc_blkno = blkno;
1644 	bcp->bc_len = len;
1645 	bcp->bc_refcount = 1;
1646 	bcp->bc_next = NULL;
1647 	prevp->bc_next = bcp;
1648 }
1649 
1650 /*
1651  * Check to see whether the buffer being recovered has a corresponding
1652  * entry in the buffer cancel record table.  If it does then return 1
1653  * so that it will be cancelled, otherwise return 0.  If the buffer is
1654  * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1655  * the refcount on the entry in the table and remove it from the table
1656  * if this is the last reference.
1657  *
1658  * We remove the cancel record from the table when we encounter its
1659  * last occurrence in the log so that if the same buffer is re-used
1660  * again after its last cancellation we actually replay the changes
1661  * made at that point.
1662  */
1663 STATIC int
1664 xlog_check_buffer_cancelled(
1665 	xlog_t			*log,
1666 	xfs_daddr_t		blkno,
1667 	uint			len,
1668 	ushort			flags)
1669 {
1670 	xfs_buf_cancel_t	*bcp;
1671 	xfs_buf_cancel_t	*prevp;
1672 	xfs_buf_cancel_t	**bucket;
1673 
1674 	if (log->l_buf_cancel_table == NULL) {
1675 		/*
1676 		 * There is nothing in the table built in pass one,
1677 		 * so this buffer must not be cancelled.
1678 		 */
1679 		ASSERT(!(flags & XFS_BLI_CANCEL));
1680 		return 0;
1681 	}
1682 
1683 	bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1684 					  XLOG_BC_TABLE_SIZE];
1685 	bcp = *bucket;
1686 	if (bcp == NULL) {
1687 		/*
1688 		 * There is no corresponding entry in the table built
1689 		 * in pass one, so this buffer has not been cancelled.
1690 		 */
1691 		ASSERT(!(flags & XFS_BLI_CANCEL));
1692 		return 0;
1693 	}
1694 
1695 	/*
1696 	 * Search for an entry in the buffer cancel table that
1697 	 * matches our buffer.
1698 	 */
1699 	prevp = NULL;
1700 	while (bcp != NULL) {
1701 		if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1702 			/*
1703 			 * We've go a match, so return 1 so that the
1704 			 * recovery of this buffer is cancelled.
1705 			 * If this buffer is actually a buffer cancel
1706 			 * log item, then decrement the refcount on the
1707 			 * one in the table and remove it if this is the
1708 			 * last reference.
1709 			 */
1710 			if (flags & XFS_BLI_CANCEL) {
1711 				bcp->bc_refcount--;
1712 				if (bcp->bc_refcount == 0) {
1713 					if (prevp == NULL) {
1714 						*bucket = bcp->bc_next;
1715 					} else {
1716 						prevp->bc_next = bcp->bc_next;
1717 					}
1718 					kmem_free(bcp);
1719 				}
1720 			}
1721 			return 1;
1722 		}
1723 		prevp = bcp;
1724 		bcp = bcp->bc_next;
1725 	}
1726 	/*
1727 	 * We didn't find a corresponding entry in the table, so
1728 	 * return 0 so that the buffer is NOT cancelled.
1729 	 */
1730 	ASSERT(!(flags & XFS_BLI_CANCEL));
1731 	return 0;
1732 }
1733 
1734 STATIC int
1735 xlog_recover_do_buffer_pass2(
1736 	xlog_t			*log,
1737 	xfs_buf_log_format_t	*buf_f)
1738 {
1739 	xfs_daddr_t		blkno = 0;
1740 	ushort			flags = 0;
1741 	uint			len = 0;
1742 
1743 	switch (buf_f->blf_type) {
1744 	case XFS_LI_BUF:
1745 		blkno = buf_f->blf_blkno;
1746 		flags = buf_f->blf_flags;
1747 		len = buf_f->blf_len;
1748 		break;
1749 	}
1750 
1751 	return xlog_check_buffer_cancelled(log, blkno, len, flags);
1752 }
1753 
1754 /*
1755  * Perform recovery for a buffer full of inodes.  In these buffers,
1756  * the only data which should be recovered is that which corresponds
1757  * to the di_next_unlinked pointers in the on disk inode structures.
1758  * The rest of the data for the inodes is always logged through the
1759  * inodes themselves rather than the inode buffer and is recovered
1760  * in xlog_recover_do_inode_trans().
1761  *
1762  * The only time when buffers full of inodes are fully recovered is
1763  * when the buffer is full of newly allocated inodes.  In this case
1764  * the buffer will not be marked as an inode buffer and so will be
1765  * sent to xlog_recover_do_reg_buffer() below during recovery.
1766  */
1767 STATIC int
1768 xlog_recover_do_inode_buffer(
1769 	xfs_mount_t		*mp,
1770 	xlog_recover_item_t	*item,
1771 	xfs_buf_t		*bp,
1772 	xfs_buf_log_format_t	*buf_f)
1773 {
1774 	int			i;
1775 	int			item_index;
1776 	int			bit;
1777 	int			nbits;
1778 	int			reg_buf_offset;
1779 	int			reg_buf_bytes;
1780 	int			next_unlinked_offset;
1781 	int			inodes_per_buf;
1782 	xfs_agino_t		*logged_nextp;
1783 	xfs_agino_t		*buffer_nextp;
1784 	unsigned int		*data_map = NULL;
1785 	unsigned int		map_size = 0;
1786 
1787 	switch (buf_f->blf_type) {
1788 	case XFS_LI_BUF:
1789 		data_map = buf_f->blf_data_map;
1790 		map_size = buf_f->blf_map_size;
1791 		break;
1792 	}
1793 	/*
1794 	 * Set the variables corresponding to the current region to
1795 	 * 0 so that we'll initialize them on the first pass through
1796 	 * the loop.
1797 	 */
1798 	reg_buf_offset = 0;
1799 	reg_buf_bytes = 0;
1800 	bit = 0;
1801 	nbits = 0;
1802 	item_index = 0;
1803 	inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1804 	for (i = 0; i < inodes_per_buf; i++) {
1805 		next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1806 			offsetof(xfs_dinode_t, di_next_unlinked);
1807 
1808 		while (next_unlinked_offset >=
1809 		       (reg_buf_offset + reg_buf_bytes)) {
1810 			/*
1811 			 * The next di_next_unlinked field is beyond
1812 			 * the current logged region.  Find the next
1813 			 * logged region that contains or is beyond
1814 			 * the current di_next_unlinked field.
1815 			 */
1816 			bit += nbits;
1817 			bit = xfs_next_bit(data_map, map_size, bit);
1818 
1819 			/*
1820 			 * If there are no more logged regions in the
1821 			 * buffer, then we're done.
1822 			 */
1823 			if (bit == -1) {
1824 				return 0;
1825 			}
1826 
1827 			nbits = xfs_contig_bits(data_map, map_size,
1828 							 bit);
1829 			ASSERT(nbits > 0);
1830 			reg_buf_offset = bit << XFS_BLI_SHIFT;
1831 			reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1832 			item_index++;
1833 		}
1834 
1835 		/*
1836 		 * If the current logged region starts after the current
1837 		 * di_next_unlinked field, then move on to the next
1838 		 * di_next_unlinked field.
1839 		 */
1840 		if (next_unlinked_offset < reg_buf_offset) {
1841 			continue;
1842 		}
1843 
1844 		ASSERT(item->ri_buf[item_index].i_addr != NULL);
1845 		ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1846 		ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1847 
1848 		/*
1849 		 * The current logged region contains a copy of the
1850 		 * current di_next_unlinked field.  Extract its value
1851 		 * and copy it to the buffer copy.
1852 		 */
1853 		logged_nextp = (xfs_agino_t *)
1854 			       ((char *)(item->ri_buf[item_index].i_addr) +
1855 				(next_unlinked_offset - reg_buf_offset));
1856 		if (unlikely(*logged_nextp == 0)) {
1857 			xfs_fs_cmn_err(CE_ALERT, mp,
1858 				"bad inode buffer log record (ptr = 0x%p, bp = 0x%p).  XFS trying to replay bad (0) inode di_next_unlinked field",
1859 				item, bp);
1860 			XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1861 					 XFS_ERRLEVEL_LOW, mp);
1862 			return XFS_ERROR(EFSCORRUPTED);
1863 		}
1864 
1865 		buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1866 					      next_unlinked_offset);
1867 		*buffer_nextp = *logged_nextp;
1868 	}
1869 
1870 	return 0;
1871 }
1872 
1873 /*
1874  * Perform a 'normal' buffer recovery.  Each logged region of the
1875  * buffer should be copied over the corresponding region in the
1876  * given buffer.  The bitmap in the buf log format structure indicates
1877  * where to place the logged data.
1878  */
1879 /*ARGSUSED*/
1880 STATIC void
1881 xlog_recover_do_reg_buffer(
1882 	xlog_recover_item_t	*item,
1883 	xfs_buf_t		*bp,
1884 	xfs_buf_log_format_t	*buf_f)
1885 {
1886 	int			i;
1887 	int			bit;
1888 	int			nbits;
1889 	unsigned int		*data_map = NULL;
1890 	unsigned int		map_size = 0;
1891 	int                     error;
1892 
1893 	switch (buf_f->blf_type) {
1894 	case XFS_LI_BUF:
1895 		data_map = buf_f->blf_data_map;
1896 		map_size = buf_f->blf_map_size;
1897 		break;
1898 	}
1899 	bit = 0;
1900 	i = 1;  /* 0 is the buf format structure */
1901 	while (1) {
1902 		bit = xfs_next_bit(data_map, map_size, bit);
1903 		if (bit == -1)
1904 			break;
1905 		nbits = xfs_contig_bits(data_map, map_size, bit);
1906 		ASSERT(nbits > 0);
1907 		ASSERT(item->ri_buf[i].i_addr != NULL);
1908 		ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1909 		ASSERT(XFS_BUF_COUNT(bp) >=
1910 		       ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1911 
1912 		/*
1913 		 * Do a sanity check if this is a dquot buffer. Just checking
1914 		 * the first dquot in the buffer should do. XXXThis is
1915 		 * probably a good thing to do for other buf types also.
1916 		 */
1917 		error = 0;
1918 		if (buf_f->blf_flags &
1919 		   (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1920 			error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1921 					       item->ri_buf[i].i_addr,
1922 					       -1, 0, XFS_QMOPT_DOWARN,
1923 					       "dquot_buf_recover");
1924 		}
1925 		if (!error)
1926 			memcpy(xfs_buf_offset(bp,
1927 				(uint)bit << XFS_BLI_SHIFT),	/* dest */
1928 				item->ri_buf[i].i_addr,		/* source */
1929 				nbits<<XFS_BLI_SHIFT);		/* length */
1930 		i++;
1931 		bit += nbits;
1932 	}
1933 
1934 	/* Shouldn't be any more regions */
1935 	ASSERT(i == item->ri_total);
1936 }
1937 
1938 /*
1939  * Do some primitive error checking on ondisk dquot data structures.
1940  */
1941 int
1942 xfs_qm_dqcheck(
1943 	xfs_disk_dquot_t *ddq,
1944 	xfs_dqid_t	 id,
1945 	uint		 type,	  /* used only when IO_dorepair is true */
1946 	uint		 flags,
1947 	char		 *str)
1948 {
1949 	xfs_dqblk_t	 *d = (xfs_dqblk_t *)ddq;
1950 	int		errs = 0;
1951 
1952 	/*
1953 	 * We can encounter an uninitialized dquot buffer for 2 reasons:
1954 	 * 1. If we crash while deleting the quotainode(s), and those blks got
1955 	 *    used for user data. This is because we take the path of regular
1956 	 *    file deletion; however, the size field of quotainodes is never
1957 	 *    updated, so all the tricks that we play in itruncate_finish
1958 	 *    don't quite matter.
1959 	 *
1960 	 * 2. We don't play the quota buffers when there's a quotaoff logitem.
1961 	 *    But the allocation will be replayed so we'll end up with an
1962 	 *    uninitialized quota block.
1963 	 *
1964 	 * This is all fine; things are still consistent, and we haven't lost
1965 	 * any quota information. Just don't complain about bad dquot blks.
1966 	 */
1967 	if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
1968 		if (flags & XFS_QMOPT_DOWARN)
1969 			cmn_err(CE_ALERT,
1970 			"%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1971 			str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
1972 		errs++;
1973 	}
1974 	if (ddq->d_version != XFS_DQUOT_VERSION) {
1975 		if (flags & XFS_QMOPT_DOWARN)
1976 			cmn_err(CE_ALERT,
1977 			"%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1978 			str, id, ddq->d_version, XFS_DQUOT_VERSION);
1979 		errs++;
1980 	}
1981 
1982 	if (ddq->d_flags != XFS_DQ_USER &&
1983 	    ddq->d_flags != XFS_DQ_PROJ &&
1984 	    ddq->d_flags != XFS_DQ_GROUP) {
1985 		if (flags & XFS_QMOPT_DOWARN)
1986 			cmn_err(CE_ALERT,
1987 			"%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1988 			str, id, ddq->d_flags);
1989 		errs++;
1990 	}
1991 
1992 	if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
1993 		if (flags & XFS_QMOPT_DOWARN)
1994 			cmn_err(CE_ALERT,
1995 			"%s : ondisk-dquot 0x%p, ID mismatch: "
1996 			"0x%x expected, found id 0x%x",
1997 			str, ddq, id, be32_to_cpu(ddq->d_id));
1998 		errs++;
1999 	}
2000 
2001 	if (!errs && ddq->d_id) {
2002 		if (ddq->d_blk_softlimit &&
2003 		    be64_to_cpu(ddq->d_bcount) >=
2004 				be64_to_cpu(ddq->d_blk_softlimit)) {
2005 			if (!ddq->d_btimer) {
2006 				if (flags & XFS_QMOPT_DOWARN)
2007 					cmn_err(CE_ALERT,
2008 					"%s : Dquot ID 0x%x (0x%p) "
2009 					"BLK TIMER NOT STARTED",
2010 					str, (int)be32_to_cpu(ddq->d_id), ddq);
2011 				errs++;
2012 			}
2013 		}
2014 		if (ddq->d_ino_softlimit &&
2015 		    be64_to_cpu(ddq->d_icount) >=
2016 				be64_to_cpu(ddq->d_ino_softlimit)) {
2017 			if (!ddq->d_itimer) {
2018 				if (flags & XFS_QMOPT_DOWARN)
2019 					cmn_err(CE_ALERT,
2020 					"%s : Dquot ID 0x%x (0x%p) "
2021 					"INODE TIMER NOT STARTED",
2022 					str, (int)be32_to_cpu(ddq->d_id), ddq);
2023 				errs++;
2024 			}
2025 		}
2026 		if (ddq->d_rtb_softlimit &&
2027 		    be64_to_cpu(ddq->d_rtbcount) >=
2028 				be64_to_cpu(ddq->d_rtb_softlimit)) {
2029 			if (!ddq->d_rtbtimer) {
2030 				if (flags & XFS_QMOPT_DOWARN)
2031 					cmn_err(CE_ALERT,
2032 					"%s : Dquot ID 0x%x (0x%p) "
2033 					"RTBLK TIMER NOT STARTED",
2034 					str, (int)be32_to_cpu(ddq->d_id), ddq);
2035 				errs++;
2036 			}
2037 		}
2038 	}
2039 
2040 	if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2041 		return errs;
2042 
2043 	if (flags & XFS_QMOPT_DOWARN)
2044 		cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2045 
2046 	/*
2047 	 * Typically, a repair is only requested by quotacheck.
2048 	 */
2049 	ASSERT(id != -1);
2050 	ASSERT(flags & XFS_QMOPT_DQREPAIR);
2051 	memset(d, 0, sizeof(xfs_dqblk_t));
2052 
2053 	d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2054 	d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2055 	d->dd_diskdq.d_flags = type;
2056 	d->dd_diskdq.d_id = cpu_to_be32(id);
2057 
2058 	return errs;
2059 }
2060 
2061 /*
2062  * Perform a dquot buffer recovery.
2063  * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2064  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2065  * Else, treat it as a regular buffer and do recovery.
2066  */
2067 STATIC void
2068 xlog_recover_do_dquot_buffer(
2069 	xfs_mount_t		*mp,
2070 	xlog_t			*log,
2071 	xlog_recover_item_t	*item,
2072 	xfs_buf_t		*bp,
2073 	xfs_buf_log_format_t	*buf_f)
2074 {
2075 	uint			type;
2076 
2077 	/*
2078 	 * Filesystems are required to send in quota flags at mount time.
2079 	 */
2080 	if (mp->m_qflags == 0) {
2081 		return;
2082 	}
2083 
2084 	type = 0;
2085 	if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2086 		type |= XFS_DQ_USER;
2087 	if (buf_f->blf_flags & XFS_BLI_PDQUOT_BUF)
2088 		type |= XFS_DQ_PROJ;
2089 	if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2090 		type |= XFS_DQ_GROUP;
2091 	/*
2092 	 * This type of quotas was turned off, so ignore this buffer
2093 	 */
2094 	if (log->l_quotaoffs_flag & type)
2095 		return;
2096 
2097 	xlog_recover_do_reg_buffer(item, bp, buf_f);
2098 }
2099 
2100 /*
2101  * This routine replays a modification made to a buffer at runtime.
2102  * There are actually two types of buffer, regular and inode, which
2103  * are handled differently.  Inode buffers are handled differently
2104  * in that we only recover a specific set of data from them, namely
2105  * the inode di_next_unlinked fields.  This is because all other inode
2106  * data is actually logged via inode records and any data we replay
2107  * here which overlaps that may be stale.
2108  *
2109  * When meta-data buffers are freed at run time we log a buffer item
2110  * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2111  * of the buffer in the log should not be replayed at recovery time.
2112  * This is so that if the blocks covered by the buffer are reused for
2113  * file data before we crash we don't end up replaying old, freed
2114  * meta-data into a user's file.
2115  *
2116  * To handle the cancellation of buffer log items, we make two passes
2117  * over the log during recovery.  During the first we build a table of
2118  * those buffers which have been cancelled, and during the second we
2119  * only replay those buffers which do not have corresponding cancel
2120  * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2121  * for more details on the implementation of the table of cancel records.
2122  */
2123 STATIC int
2124 xlog_recover_do_buffer_trans(
2125 	xlog_t			*log,
2126 	xlog_recover_item_t	*item,
2127 	int			pass)
2128 {
2129 	xfs_buf_log_format_t	*buf_f;
2130 	xfs_mount_t		*mp;
2131 	xfs_buf_t		*bp;
2132 	int			error;
2133 	int			cancel;
2134 	xfs_daddr_t		blkno;
2135 	int			len;
2136 	ushort			flags;
2137 
2138 	buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2139 
2140 	if (pass == XLOG_RECOVER_PASS1) {
2141 		/*
2142 		 * In this pass we're only looking for buf items
2143 		 * with the XFS_BLI_CANCEL bit set.
2144 		 */
2145 		xlog_recover_do_buffer_pass1(log, buf_f);
2146 		return 0;
2147 	} else {
2148 		/*
2149 		 * In this pass we want to recover all the buffers
2150 		 * which have not been cancelled and are not
2151 		 * cancellation buffers themselves.  The routine
2152 		 * we call here will tell us whether or not to
2153 		 * continue with the replay of this buffer.
2154 		 */
2155 		cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2156 		if (cancel) {
2157 			return 0;
2158 		}
2159 	}
2160 	switch (buf_f->blf_type) {
2161 	case XFS_LI_BUF:
2162 		blkno = buf_f->blf_blkno;
2163 		len = buf_f->blf_len;
2164 		flags = buf_f->blf_flags;
2165 		break;
2166 	default:
2167 		xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2168 			"xfs_log_recover: unknown buffer type 0x%x, logdev %s",
2169 			buf_f->blf_type, log->l_mp->m_logname ?
2170 			log->l_mp->m_logname : "internal");
2171 		XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2172 				 XFS_ERRLEVEL_LOW, log->l_mp);
2173 		return XFS_ERROR(EFSCORRUPTED);
2174 	}
2175 
2176 	mp = log->l_mp;
2177 	if (flags & XFS_BLI_INODE_BUF) {
2178 		bp = xfs_buf_read_flags(mp->m_ddev_targp, blkno, len,
2179 								XFS_BUF_LOCK);
2180 	} else {
2181 		bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, 0);
2182 	}
2183 	if (XFS_BUF_ISERROR(bp)) {
2184 		xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2185 				  bp, blkno);
2186 		error = XFS_BUF_GETERROR(bp);
2187 		xfs_buf_relse(bp);
2188 		return error;
2189 	}
2190 
2191 	error = 0;
2192 	if (flags & XFS_BLI_INODE_BUF) {
2193 		error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2194 	} else if (flags &
2195 		  (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
2196 		xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2197 	} else {
2198 		xlog_recover_do_reg_buffer(item, bp, buf_f);
2199 	}
2200 	if (error)
2201 		return XFS_ERROR(error);
2202 
2203 	/*
2204 	 * Perform delayed write on the buffer.  Asynchronous writes will be
2205 	 * slower when taking into account all the buffers to be flushed.
2206 	 *
2207 	 * Also make sure that only inode buffers with good sizes stay in
2208 	 * the buffer cache.  The kernel moves inodes in buffers of 1 block
2209 	 * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2210 	 * buffers in the log can be a different size if the log was generated
2211 	 * by an older kernel using unclustered inode buffers or a newer kernel
2212 	 * running with a different inode cluster size.  Regardless, if the
2213 	 * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2214 	 * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2215 	 * the buffer out of the buffer cache so that the buffer won't
2216 	 * overlap with future reads of those inodes.
2217 	 */
2218 	if (XFS_DINODE_MAGIC ==
2219 	    be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2220 	    (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2221 			(__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2222 		XFS_BUF_STALE(bp);
2223 		error = xfs_bwrite(mp, bp);
2224 	} else {
2225 		ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2226 		       XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2227 		XFS_BUF_SET_FSPRIVATE(bp, mp);
2228 		XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2229 		xfs_bdwrite(mp, bp);
2230 	}
2231 
2232 	return (error);
2233 }
2234 
2235 STATIC int
2236 xlog_recover_do_inode_trans(
2237 	xlog_t			*log,
2238 	xlog_recover_item_t	*item,
2239 	int			pass)
2240 {
2241 	xfs_inode_log_format_t	*in_f;
2242 	xfs_mount_t		*mp;
2243 	xfs_buf_t		*bp;
2244 	xfs_imap_t		imap;
2245 	xfs_dinode_t		*dip;
2246 	xfs_ino_t		ino;
2247 	int			len;
2248 	xfs_caddr_t		src;
2249 	xfs_caddr_t		dest;
2250 	int			error;
2251 	int			attr_index;
2252 	uint			fields;
2253 	xfs_icdinode_t		*dicp;
2254 	int			need_free = 0;
2255 
2256 	if (pass == XLOG_RECOVER_PASS1) {
2257 		return 0;
2258 	}
2259 
2260 	if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2261 		in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2262 	} else {
2263 		in_f = (xfs_inode_log_format_t *)kmem_alloc(
2264 			sizeof(xfs_inode_log_format_t), KM_SLEEP);
2265 		need_free = 1;
2266 		error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2267 		if (error)
2268 			goto error;
2269 	}
2270 	ino = in_f->ilf_ino;
2271 	mp = log->l_mp;
2272 	if (ITEM_TYPE(item) == XFS_LI_INODE) {
2273 		imap.im_blkno = (xfs_daddr_t)in_f->ilf_blkno;
2274 		imap.im_len = in_f->ilf_len;
2275 		imap.im_boffset = in_f->ilf_boffset;
2276 	} else {
2277 		/*
2278 		 * It's an old inode format record.  We don't know where
2279 		 * its cluster is located on disk, and we can't allow
2280 		 * xfs_imap() to figure it out because the inode btrees
2281 		 * are not ready to be used.  Therefore do not pass the
2282 		 * XFS_IMAP_LOOKUP flag to xfs_imap().  This will give
2283 		 * us only the single block in which the inode lives
2284 		 * rather than its cluster, so we must make sure to
2285 		 * invalidate the buffer when we write it out below.
2286 		 */
2287 		imap.im_blkno = 0;
2288 		error = xfs_imap(log->l_mp, NULL, ino, &imap, 0);
2289 		if (error)
2290 			goto error;
2291 	}
2292 
2293 	/*
2294 	 * Inode buffers can be freed, look out for it,
2295 	 * and do not replay the inode.
2296 	 */
2297 	if (xlog_check_buffer_cancelled(log, imap.im_blkno, imap.im_len, 0)) {
2298 		error = 0;
2299 		goto error;
2300 	}
2301 
2302 	bp = xfs_buf_read_flags(mp->m_ddev_targp, imap.im_blkno, imap.im_len,
2303 								XFS_BUF_LOCK);
2304 	if (XFS_BUF_ISERROR(bp)) {
2305 		xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2306 				  bp, imap.im_blkno);
2307 		error = XFS_BUF_GETERROR(bp);
2308 		xfs_buf_relse(bp);
2309 		goto error;
2310 	}
2311 	error = 0;
2312 	ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2313 	dip = (xfs_dinode_t *)xfs_buf_offset(bp, imap.im_boffset);
2314 
2315 	/*
2316 	 * Make sure the place we're flushing out to really looks
2317 	 * like an inode!
2318 	 */
2319 	if (unlikely(be16_to_cpu(dip->di_core.di_magic) != XFS_DINODE_MAGIC)) {
2320 		xfs_buf_relse(bp);
2321 		xfs_fs_cmn_err(CE_ALERT, mp,
2322 			"xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2323 			dip, bp, ino);
2324 		XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2325 				 XFS_ERRLEVEL_LOW, mp);
2326 		error = EFSCORRUPTED;
2327 		goto error;
2328 	}
2329 	dicp = (xfs_icdinode_t *)(item->ri_buf[1].i_addr);
2330 	if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2331 		xfs_buf_relse(bp);
2332 		xfs_fs_cmn_err(CE_ALERT, mp,
2333 			"xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2334 			item, ino);
2335 		XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2336 				 XFS_ERRLEVEL_LOW, mp);
2337 		error = EFSCORRUPTED;
2338 		goto error;
2339 	}
2340 
2341 	/* Skip replay when the on disk inode is newer than the log one */
2342 	if (dicp->di_flushiter < be16_to_cpu(dip->di_core.di_flushiter)) {
2343 		/*
2344 		 * Deal with the wrap case, DI_MAX_FLUSH is less
2345 		 * than smaller numbers
2346 		 */
2347 		if (be16_to_cpu(dip->di_core.di_flushiter) == DI_MAX_FLUSH &&
2348 		    dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2349 			/* do nothing */
2350 		} else {
2351 			xfs_buf_relse(bp);
2352 			error = 0;
2353 			goto error;
2354 		}
2355 	}
2356 	/* Take the opportunity to reset the flush iteration count */
2357 	dicp->di_flushiter = 0;
2358 
2359 	if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2360 		if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2361 		    (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2362 			XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2363 					 XFS_ERRLEVEL_LOW, mp, dicp);
2364 			xfs_buf_relse(bp);
2365 			xfs_fs_cmn_err(CE_ALERT, mp,
2366 				"xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2367 				item, dip, bp, ino);
2368 			error = EFSCORRUPTED;
2369 			goto error;
2370 		}
2371 	} else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2372 		if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2373 		    (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2374 		    (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2375 			XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2376 					     XFS_ERRLEVEL_LOW, mp, dicp);
2377 			xfs_buf_relse(bp);
2378 			xfs_fs_cmn_err(CE_ALERT, mp,
2379 				"xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2380 				item, dip, bp, ino);
2381 			error = EFSCORRUPTED;
2382 			goto error;
2383 		}
2384 	}
2385 	if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2386 		XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2387 				     XFS_ERRLEVEL_LOW, mp, dicp);
2388 		xfs_buf_relse(bp);
2389 		xfs_fs_cmn_err(CE_ALERT, mp,
2390 			"xfs_inode_recover: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2391 			item, dip, bp, ino,
2392 			dicp->di_nextents + dicp->di_anextents,
2393 			dicp->di_nblocks);
2394 		error = EFSCORRUPTED;
2395 		goto error;
2396 	}
2397 	if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2398 		XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2399 				     XFS_ERRLEVEL_LOW, mp, dicp);
2400 		xfs_buf_relse(bp);
2401 		xfs_fs_cmn_err(CE_ALERT, mp,
2402 			"xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2403 			item, dip, bp, ino, dicp->di_forkoff);
2404 		error = EFSCORRUPTED;
2405 		goto error;
2406 	}
2407 	if (unlikely(item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t))) {
2408 		XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
2409 				     XFS_ERRLEVEL_LOW, mp, dicp);
2410 		xfs_buf_relse(bp);
2411 		xfs_fs_cmn_err(CE_ALERT, mp,
2412 			"xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2413 			item->ri_buf[1].i_len, item);
2414 		error = EFSCORRUPTED;
2415 		goto error;
2416 	}
2417 
2418 	/* The core is in in-core format */
2419 	xfs_dinode_to_disk(&dip->di_core,
2420 		(xfs_icdinode_t *)item->ri_buf[1].i_addr);
2421 
2422 	/* the rest is in on-disk format */
2423 	if (item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t)) {
2424 		memcpy((xfs_caddr_t) dip + sizeof(xfs_dinode_core_t),
2425 			item->ri_buf[1].i_addr + sizeof(xfs_dinode_core_t),
2426 			item->ri_buf[1].i_len  - sizeof(xfs_dinode_core_t));
2427 	}
2428 
2429 	fields = in_f->ilf_fields;
2430 	switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2431 	case XFS_ILOG_DEV:
2432 		dip->di_u.di_dev = cpu_to_be32(in_f->ilf_u.ilfu_rdev);
2433 		break;
2434 	case XFS_ILOG_UUID:
2435 		dip->di_u.di_muuid = in_f->ilf_u.ilfu_uuid;
2436 		break;
2437 	}
2438 
2439 	if (in_f->ilf_size == 2)
2440 		goto write_inode_buffer;
2441 	len = item->ri_buf[2].i_len;
2442 	src = item->ri_buf[2].i_addr;
2443 	ASSERT(in_f->ilf_size <= 4);
2444 	ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2445 	ASSERT(!(fields & XFS_ILOG_DFORK) ||
2446 	       (len == in_f->ilf_dsize));
2447 
2448 	switch (fields & XFS_ILOG_DFORK) {
2449 	case XFS_ILOG_DDATA:
2450 	case XFS_ILOG_DEXT:
2451 		memcpy(&dip->di_u, src, len);
2452 		break;
2453 
2454 	case XFS_ILOG_DBROOT:
2455 		xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2456 				 &(dip->di_u.di_bmbt),
2457 				 XFS_DFORK_DSIZE(dip, mp));
2458 		break;
2459 
2460 	default:
2461 		/*
2462 		 * There are no data fork flags set.
2463 		 */
2464 		ASSERT((fields & XFS_ILOG_DFORK) == 0);
2465 		break;
2466 	}
2467 
2468 	/*
2469 	 * If we logged any attribute data, recover it.  There may or
2470 	 * may not have been any other non-core data logged in this
2471 	 * transaction.
2472 	 */
2473 	if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2474 		if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2475 			attr_index = 3;
2476 		} else {
2477 			attr_index = 2;
2478 		}
2479 		len = item->ri_buf[attr_index].i_len;
2480 		src = item->ri_buf[attr_index].i_addr;
2481 		ASSERT(len == in_f->ilf_asize);
2482 
2483 		switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2484 		case XFS_ILOG_ADATA:
2485 		case XFS_ILOG_AEXT:
2486 			dest = XFS_DFORK_APTR(dip);
2487 			ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2488 			memcpy(dest, src, len);
2489 			break;
2490 
2491 		case XFS_ILOG_ABROOT:
2492 			dest = XFS_DFORK_APTR(dip);
2493 			xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2494 					 (xfs_bmdr_block_t*)dest,
2495 					 XFS_DFORK_ASIZE(dip, mp));
2496 			break;
2497 
2498 		default:
2499 			xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2500 			ASSERT(0);
2501 			xfs_buf_relse(bp);
2502 			error = EIO;
2503 			goto error;
2504 		}
2505 	}
2506 
2507 write_inode_buffer:
2508 	if (ITEM_TYPE(item) == XFS_LI_INODE) {
2509 		ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2510 		       XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2511 		XFS_BUF_SET_FSPRIVATE(bp, mp);
2512 		XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2513 		xfs_bdwrite(mp, bp);
2514 	} else {
2515 		XFS_BUF_STALE(bp);
2516 		error = xfs_bwrite(mp, bp);
2517 	}
2518 
2519 error:
2520 	if (need_free)
2521 		kmem_free(in_f);
2522 	return XFS_ERROR(error);
2523 }
2524 
2525 /*
2526  * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2527  * structure, so that we know not to do any dquot item or dquot buffer recovery,
2528  * of that type.
2529  */
2530 STATIC int
2531 xlog_recover_do_quotaoff_trans(
2532 	xlog_t			*log,
2533 	xlog_recover_item_t	*item,
2534 	int			pass)
2535 {
2536 	xfs_qoff_logformat_t	*qoff_f;
2537 
2538 	if (pass == XLOG_RECOVER_PASS2) {
2539 		return (0);
2540 	}
2541 
2542 	qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2543 	ASSERT(qoff_f);
2544 
2545 	/*
2546 	 * The logitem format's flag tells us if this was user quotaoff,
2547 	 * group/project quotaoff or both.
2548 	 */
2549 	if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2550 		log->l_quotaoffs_flag |= XFS_DQ_USER;
2551 	if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2552 		log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2553 	if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2554 		log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2555 
2556 	return (0);
2557 }
2558 
2559 /*
2560  * Recover a dquot record
2561  */
2562 STATIC int
2563 xlog_recover_do_dquot_trans(
2564 	xlog_t			*log,
2565 	xlog_recover_item_t	*item,
2566 	int			pass)
2567 {
2568 	xfs_mount_t		*mp;
2569 	xfs_buf_t		*bp;
2570 	struct xfs_disk_dquot	*ddq, *recddq;
2571 	int			error;
2572 	xfs_dq_logformat_t	*dq_f;
2573 	uint			type;
2574 
2575 	if (pass == XLOG_RECOVER_PASS1) {
2576 		return 0;
2577 	}
2578 	mp = log->l_mp;
2579 
2580 	/*
2581 	 * Filesystems are required to send in quota flags at mount time.
2582 	 */
2583 	if (mp->m_qflags == 0)
2584 		return (0);
2585 
2586 	recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2587 	ASSERT(recddq);
2588 	/*
2589 	 * This type of quotas was turned off, so ignore this record.
2590 	 */
2591 	type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2592 	ASSERT(type);
2593 	if (log->l_quotaoffs_flag & type)
2594 		return (0);
2595 
2596 	/*
2597 	 * At this point we know that quota was _not_ turned off.
2598 	 * Since the mount flags are not indicating to us otherwise, this
2599 	 * must mean that quota is on, and the dquot needs to be replayed.
2600 	 * Remember that we may not have fully recovered the superblock yet,
2601 	 * so we can't do the usual trick of looking at the SB quota bits.
2602 	 *
2603 	 * The other possibility, of course, is that the quota subsystem was
2604 	 * removed since the last mount - ENOSYS.
2605 	 */
2606 	dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2607 	ASSERT(dq_f);
2608 	if ((error = xfs_qm_dqcheck(recddq,
2609 			   dq_f->qlf_id,
2610 			   0, XFS_QMOPT_DOWARN,
2611 			   "xlog_recover_do_dquot_trans (log copy)"))) {
2612 		return XFS_ERROR(EIO);
2613 	}
2614 	ASSERT(dq_f->qlf_len == 1);
2615 
2616 	error = xfs_read_buf(mp, mp->m_ddev_targp,
2617 			     dq_f->qlf_blkno,
2618 			     XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2619 			     0, &bp);
2620 	if (error) {
2621 		xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2622 				  bp, dq_f->qlf_blkno);
2623 		return error;
2624 	}
2625 	ASSERT(bp);
2626 	ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2627 
2628 	/*
2629 	 * At least the magic num portion should be on disk because this
2630 	 * was among a chunk of dquots created earlier, and we did some
2631 	 * minimal initialization then.
2632 	 */
2633 	if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2634 			   "xlog_recover_do_dquot_trans")) {
2635 		xfs_buf_relse(bp);
2636 		return XFS_ERROR(EIO);
2637 	}
2638 
2639 	memcpy(ddq, recddq, item->ri_buf[1].i_len);
2640 
2641 	ASSERT(dq_f->qlf_size == 2);
2642 	ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2643 	       XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2644 	XFS_BUF_SET_FSPRIVATE(bp, mp);
2645 	XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2646 	xfs_bdwrite(mp, bp);
2647 
2648 	return (0);
2649 }
2650 
2651 /*
2652  * This routine is called to create an in-core extent free intent
2653  * item from the efi format structure which was logged on disk.
2654  * It allocates an in-core efi, copies the extents from the format
2655  * structure into it, and adds the efi to the AIL with the given
2656  * LSN.
2657  */
2658 STATIC int
2659 xlog_recover_do_efi_trans(
2660 	xlog_t			*log,
2661 	xlog_recover_item_t	*item,
2662 	xfs_lsn_t		lsn,
2663 	int			pass)
2664 {
2665 	int			error;
2666 	xfs_mount_t		*mp;
2667 	xfs_efi_log_item_t	*efip;
2668 	xfs_efi_log_format_t	*efi_formatp;
2669 
2670 	if (pass == XLOG_RECOVER_PASS1) {
2671 		return 0;
2672 	}
2673 
2674 	efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2675 
2676 	mp = log->l_mp;
2677 	efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2678 	if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2679 					 &(efip->efi_format)))) {
2680 		xfs_efi_item_free(efip);
2681 		return error;
2682 	}
2683 	efip->efi_next_extent = efi_formatp->efi_nextents;
2684 	efip->efi_flags |= XFS_EFI_COMMITTED;
2685 
2686 	spin_lock(&mp->m_ail_lock);
2687 	/*
2688 	 * xfs_trans_update_ail() drops the AIL lock.
2689 	 */
2690 	xfs_trans_update_ail(mp, (xfs_log_item_t *)efip, lsn);
2691 	return 0;
2692 }
2693 
2694 
2695 /*
2696  * This routine is called when an efd format structure is found in
2697  * a committed transaction in the log.  It's purpose is to cancel
2698  * the corresponding efi if it was still in the log.  To do this
2699  * it searches the AIL for the efi with an id equal to that in the
2700  * efd format structure.  If we find it, we remove the efi from the
2701  * AIL and free it.
2702  */
2703 STATIC void
2704 xlog_recover_do_efd_trans(
2705 	xlog_t			*log,
2706 	xlog_recover_item_t	*item,
2707 	int			pass)
2708 {
2709 	xfs_mount_t		*mp;
2710 	xfs_efd_log_format_t	*efd_formatp;
2711 	xfs_efi_log_item_t	*efip = NULL;
2712 	xfs_log_item_t		*lip;
2713 	int			gen;
2714 	__uint64_t		efi_id;
2715 
2716 	if (pass == XLOG_RECOVER_PASS1) {
2717 		return;
2718 	}
2719 
2720 	efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2721 	ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2722 		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2723 	       (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2724 		((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2725 	efi_id = efd_formatp->efd_efi_id;
2726 
2727 	/*
2728 	 * Search for the efi with the id in the efd format structure
2729 	 * in the AIL.
2730 	 */
2731 	mp = log->l_mp;
2732 	spin_lock(&mp->m_ail_lock);
2733 	lip = xfs_trans_first_ail(mp, &gen);
2734 	while (lip != NULL) {
2735 		if (lip->li_type == XFS_LI_EFI) {
2736 			efip = (xfs_efi_log_item_t *)lip;
2737 			if (efip->efi_format.efi_id == efi_id) {
2738 				/*
2739 				 * xfs_trans_delete_ail() drops the
2740 				 * AIL lock.
2741 				 */
2742 				xfs_trans_delete_ail(mp, lip);
2743 				xfs_efi_item_free(efip);
2744 				return;
2745 			}
2746 		}
2747 		lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
2748 	}
2749 	spin_unlock(&mp->m_ail_lock);
2750 }
2751 
2752 /*
2753  * Perform the transaction
2754  *
2755  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2756  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2757  */
2758 STATIC int
2759 xlog_recover_do_trans(
2760 	xlog_t			*log,
2761 	xlog_recover_t		*trans,
2762 	int			pass)
2763 {
2764 	int			error = 0;
2765 	xlog_recover_item_t	*item, *first_item;
2766 
2767 	if ((error = xlog_recover_reorder_trans(trans)))
2768 		return error;
2769 	first_item = item = trans->r_itemq;
2770 	do {
2771 		/*
2772 		 * we don't need to worry about the block number being
2773 		 * truncated in > 1 TB buffers because in user-land,
2774 		 * we're now n32 or 64-bit so xfs_daddr_t is 64-bits so
2775 		 * the blknos will get through the user-mode buffer
2776 		 * cache properly.  The only bad case is o32 kernels
2777 		 * where xfs_daddr_t is 32-bits but mount will warn us
2778 		 * off a > 1 TB filesystem before we get here.
2779 		 */
2780 		if ((ITEM_TYPE(item) == XFS_LI_BUF)) {
2781 			if  ((error = xlog_recover_do_buffer_trans(log, item,
2782 								 pass)))
2783 				break;
2784 		} else if ((ITEM_TYPE(item) == XFS_LI_INODE)) {
2785 			if ((error = xlog_recover_do_inode_trans(log, item,
2786 								pass)))
2787 				break;
2788 		} else if (ITEM_TYPE(item) == XFS_LI_EFI) {
2789 			if ((error = xlog_recover_do_efi_trans(log, item, trans->r_lsn,
2790 						  pass)))
2791 				break;
2792 		} else if (ITEM_TYPE(item) == XFS_LI_EFD) {
2793 			xlog_recover_do_efd_trans(log, item, pass);
2794 		} else if (ITEM_TYPE(item) == XFS_LI_DQUOT) {
2795 			if ((error = xlog_recover_do_dquot_trans(log, item,
2796 								   pass)))
2797 					break;
2798 		} else if ((ITEM_TYPE(item) == XFS_LI_QUOTAOFF)) {
2799 			if ((error = xlog_recover_do_quotaoff_trans(log, item,
2800 								   pass)))
2801 					break;
2802 		} else {
2803 			xlog_warn("XFS: xlog_recover_do_trans");
2804 			ASSERT(0);
2805 			error = XFS_ERROR(EIO);
2806 			break;
2807 		}
2808 		item = item->ri_next;
2809 	} while (first_item != item);
2810 
2811 	return error;
2812 }
2813 
2814 /*
2815  * Free up any resources allocated by the transaction
2816  *
2817  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2818  */
2819 STATIC void
2820 xlog_recover_free_trans(
2821 	xlog_recover_t		*trans)
2822 {
2823 	xlog_recover_item_t	*first_item, *item, *free_item;
2824 	int			i;
2825 
2826 	item = first_item = trans->r_itemq;
2827 	do {
2828 		free_item = item;
2829 		item = item->ri_next;
2830 		 /* Free the regions in the item. */
2831 		for (i = 0; i < free_item->ri_cnt; i++) {
2832 			kmem_free(free_item->ri_buf[i].i_addr);
2833 		}
2834 		/* Free the item itself */
2835 		kmem_free(free_item->ri_buf);
2836 		kmem_free(free_item);
2837 	} while (first_item != item);
2838 	/* Free the transaction recover structure */
2839 	kmem_free(trans);
2840 }
2841 
2842 STATIC int
2843 xlog_recover_commit_trans(
2844 	xlog_t			*log,
2845 	xlog_recover_t		**q,
2846 	xlog_recover_t		*trans,
2847 	int			pass)
2848 {
2849 	int			error;
2850 
2851 	if ((error = xlog_recover_unlink_tid(q, trans)))
2852 		return error;
2853 	if ((error = xlog_recover_do_trans(log, trans, pass)))
2854 		return error;
2855 	xlog_recover_free_trans(trans);			/* no error */
2856 	return 0;
2857 }
2858 
2859 STATIC int
2860 xlog_recover_unmount_trans(
2861 	xlog_recover_t		*trans)
2862 {
2863 	/* Do nothing now */
2864 	xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2865 	return 0;
2866 }
2867 
2868 /*
2869  * There are two valid states of the r_state field.  0 indicates that the
2870  * transaction structure is in a normal state.  We have either seen the
2871  * start of the transaction or the last operation we added was not a partial
2872  * operation.  If the last operation we added to the transaction was a
2873  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2874  *
2875  * NOTE: skip LRs with 0 data length.
2876  */
2877 STATIC int
2878 xlog_recover_process_data(
2879 	xlog_t			*log,
2880 	xlog_recover_t		*rhash[],
2881 	xlog_rec_header_t	*rhead,
2882 	xfs_caddr_t		dp,
2883 	int			pass)
2884 {
2885 	xfs_caddr_t		lp;
2886 	int			num_logops;
2887 	xlog_op_header_t	*ohead;
2888 	xlog_recover_t		*trans;
2889 	xlog_tid_t		tid;
2890 	int			error;
2891 	unsigned long		hash;
2892 	uint			flags;
2893 
2894 	lp = dp + be32_to_cpu(rhead->h_len);
2895 	num_logops = be32_to_cpu(rhead->h_num_logops);
2896 
2897 	/* check the log format matches our own - else we can't recover */
2898 	if (xlog_header_check_recover(log->l_mp, rhead))
2899 		return (XFS_ERROR(EIO));
2900 
2901 	while ((dp < lp) && num_logops) {
2902 		ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2903 		ohead = (xlog_op_header_t *)dp;
2904 		dp += sizeof(xlog_op_header_t);
2905 		if (ohead->oh_clientid != XFS_TRANSACTION &&
2906 		    ohead->oh_clientid != XFS_LOG) {
2907 			xlog_warn(
2908 		"XFS: xlog_recover_process_data: bad clientid");
2909 			ASSERT(0);
2910 			return (XFS_ERROR(EIO));
2911 		}
2912 		tid = be32_to_cpu(ohead->oh_tid);
2913 		hash = XLOG_RHASH(tid);
2914 		trans = xlog_recover_find_tid(rhash[hash], tid);
2915 		if (trans == NULL) {		   /* not found; add new tid */
2916 			if (ohead->oh_flags & XLOG_START_TRANS)
2917 				xlog_recover_new_tid(&rhash[hash], tid,
2918 					be64_to_cpu(rhead->h_lsn));
2919 		} else {
2920 			if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2921 				xlog_warn(
2922 			"XFS: xlog_recover_process_data: bad length");
2923 				WARN_ON(1);
2924 				return (XFS_ERROR(EIO));
2925 			}
2926 			flags = ohead->oh_flags & ~XLOG_END_TRANS;
2927 			if (flags & XLOG_WAS_CONT_TRANS)
2928 				flags &= ~XLOG_CONTINUE_TRANS;
2929 			switch (flags) {
2930 			case XLOG_COMMIT_TRANS:
2931 				error = xlog_recover_commit_trans(log,
2932 						&rhash[hash], trans, pass);
2933 				break;
2934 			case XLOG_UNMOUNT_TRANS:
2935 				error = xlog_recover_unmount_trans(trans);
2936 				break;
2937 			case XLOG_WAS_CONT_TRANS:
2938 				error = xlog_recover_add_to_cont_trans(trans,
2939 						dp, be32_to_cpu(ohead->oh_len));
2940 				break;
2941 			case XLOG_START_TRANS:
2942 				xlog_warn(
2943 			"XFS: xlog_recover_process_data: bad transaction");
2944 				ASSERT(0);
2945 				error = XFS_ERROR(EIO);
2946 				break;
2947 			case 0:
2948 			case XLOG_CONTINUE_TRANS:
2949 				error = xlog_recover_add_to_trans(trans,
2950 						dp, be32_to_cpu(ohead->oh_len));
2951 				break;
2952 			default:
2953 				xlog_warn(
2954 			"XFS: xlog_recover_process_data: bad flag");
2955 				ASSERT(0);
2956 				error = XFS_ERROR(EIO);
2957 				break;
2958 			}
2959 			if (error)
2960 				return error;
2961 		}
2962 		dp += be32_to_cpu(ohead->oh_len);
2963 		num_logops--;
2964 	}
2965 	return 0;
2966 }
2967 
2968 /*
2969  * Process an extent free intent item that was recovered from
2970  * the log.  We need to free the extents that it describes.
2971  */
2972 STATIC int
2973 xlog_recover_process_efi(
2974 	xfs_mount_t		*mp,
2975 	xfs_efi_log_item_t	*efip)
2976 {
2977 	xfs_efd_log_item_t	*efdp;
2978 	xfs_trans_t		*tp;
2979 	int			i;
2980 	int			error = 0;
2981 	xfs_extent_t		*extp;
2982 	xfs_fsblock_t		startblock_fsb;
2983 
2984 	ASSERT(!(efip->efi_flags & XFS_EFI_RECOVERED));
2985 
2986 	/*
2987 	 * First check the validity of the extents described by the
2988 	 * EFI.  If any are bad, then assume that all are bad and
2989 	 * just toss the EFI.
2990 	 */
2991 	for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2992 		extp = &(efip->efi_format.efi_extents[i]);
2993 		startblock_fsb = XFS_BB_TO_FSB(mp,
2994 				   XFS_FSB_TO_DADDR(mp, extp->ext_start));
2995 		if ((startblock_fsb == 0) ||
2996 		    (extp->ext_len == 0) ||
2997 		    (startblock_fsb >= mp->m_sb.sb_dblocks) ||
2998 		    (extp->ext_len >= mp->m_sb.sb_agblocks)) {
2999 			/*
3000 			 * This will pull the EFI from the AIL and
3001 			 * free the memory associated with it.
3002 			 */
3003 			xfs_efi_release(efip, efip->efi_format.efi_nextents);
3004 			return XFS_ERROR(EIO);
3005 		}
3006 	}
3007 
3008 	tp = xfs_trans_alloc(mp, 0);
3009 	error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3010 	if (error)
3011 		goto abort_error;
3012 	efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3013 
3014 	for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3015 		extp = &(efip->efi_format.efi_extents[i]);
3016 		error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3017 		if (error)
3018 			goto abort_error;
3019 		xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3020 					 extp->ext_len);
3021 	}
3022 
3023 	efip->efi_flags |= XFS_EFI_RECOVERED;
3024 	error = xfs_trans_commit(tp, 0);
3025 	return error;
3026 
3027 abort_error:
3028 	xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3029 	return error;
3030 }
3031 
3032 /*
3033  * Verify that once we've encountered something other than an EFI
3034  * in the AIL that there are no more EFIs in the AIL.
3035  */
3036 #if defined(DEBUG)
3037 STATIC void
3038 xlog_recover_check_ail(
3039 	xfs_mount_t		*mp,
3040 	xfs_log_item_t		*lip,
3041 	int			gen)
3042 {
3043 	int			orig_gen = gen;
3044 
3045 	do {
3046 		ASSERT(lip->li_type != XFS_LI_EFI);
3047 		lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3048 		/*
3049 		 * The check will be bogus if we restart from the
3050 		 * beginning of the AIL, so ASSERT that we don't.
3051 		 * We never should since we're holding the AIL lock
3052 		 * the entire time.
3053 		 */
3054 		ASSERT(gen == orig_gen);
3055 	} while (lip != NULL);
3056 }
3057 #endif	/* DEBUG */
3058 
3059 /*
3060  * When this is called, all of the EFIs which did not have
3061  * corresponding EFDs should be in the AIL.  What we do now
3062  * is free the extents associated with each one.
3063  *
3064  * Since we process the EFIs in normal transactions, they
3065  * will be removed at some point after the commit.  This prevents
3066  * us from just walking down the list processing each one.
3067  * We'll use a flag in the EFI to skip those that we've already
3068  * processed and use the AIL iteration mechanism's generation
3069  * count to try to speed this up at least a bit.
3070  *
3071  * When we start, we know that the EFIs are the only things in
3072  * the AIL.  As we process them, however, other items are added
3073  * to the AIL.  Since everything added to the AIL must come after
3074  * everything already in the AIL, we stop processing as soon as
3075  * we see something other than an EFI in the AIL.
3076  */
3077 STATIC int
3078 xlog_recover_process_efis(
3079 	xlog_t			*log)
3080 {
3081 	xfs_log_item_t		*lip;
3082 	xfs_efi_log_item_t	*efip;
3083 	int			gen;
3084 	xfs_mount_t		*mp;
3085 	int			error = 0;
3086 
3087 	mp = log->l_mp;
3088 	spin_lock(&mp->m_ail_lock);
3089 
3090 	lip = xfs_trans_first_ail(mp, &gen);
3091 	while (lip != NULL) {
3092 		/*
3093 		 * We're done when we see something other than an EFI.
3094 		 */
3095 		if (lip->li_type != XFS_LI_EFI) {
3096 			xlog_recover_check_ail(mp, lip, gen);
3097 			break;
3098 		}
3099 
3100 		/*
3101 		 * Skip EFIs that we've already processed.
3102 		 */
3103 		efip = (xfs_efi_log_item_t *)lip;
3104 		if (efip->efi_flags & XFS_EFI_RECOVERED) {
3105 			lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3106 			continue;
3107 		}
3108 
3109 		spin_unlock(&mp->m_ail_lock);
3110 		error = xlog_recover_process_efi(mp, efip);
3111 		if (error)
3112 			return error;
3113 		spin_lock(&mp->m_ail_lock);
3114 		lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3115 	}
3116 	spin_unlock(&mp->m_ail_lock);
3117 	return error;
3118 }
3119 
3120 /*
3121  * This routine performs a transaction to null out a bad inode pointer
3122  * in an agi unlinked inode hash bucket.
3123  */
3124 STATIC void
3125 xlog_recover_clear_agi_bucket(
3126 	xfs_mount_t	*mp,
3127 	xfs_agnumber_t	agno,
3128 	int		bucket)
3129 {
3130 	xfs_trans_t	*tp;
3131 	xfs_agi_t	*agi;
3132 	xfs_buf_t	*agibp;
3133 	int		offset;
3134 	int		error;
3135 
3136 	tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3137 	error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp), 0, 0, 0);
3138 	if (!error)
3139 		error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
3140 				   XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3141 				   XFS_FSS_TO_BB(mp, 1), 0, &agibp);
3142 	if (error)
3143 		goto out_abort;
3144 
3145 	error = EINVAL;
3146 	agi = XFS_BUF_TO_AGI(agibp);
3147 	if (be32_to_cpu(agi->agi_magicnum) != XFS_AGI_MAGIC)
3148 		goto out_abort;
3149 
3150 	agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3151 	offset = offsetof(xfs_agi_t, agi_unlinked) +
3152 		 (sizeof(xfs_agino_t) * bucket);
3153 	xfs_trans_log_buf(tp, agibp, offset,
3154 			  (offset + sizeof(xfs_agino_t) - 1));
3155 
3156 	error = xfs_trans_commit(tp, 0);
3157 	if (error)
3158 		goto out_error;
3159 	return;
3160 
3161 out_abort:
3162 	xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3163 out_error:
3164 	xfs_fs_cmn_err(CE_WARN, mp, "xlog_recover_clear_agi_bucket: "
3165 			"failed to clear agi %d. Continuing.", agno);
3166 	return;
3167 }
3168 
3169 /*
3170  * xlog_iunlink_recover
3171  *
3172  * This is called during recovery to process any inodes which
3173  * we unlinked but not freed when the system crashed.  These
3174  * inodes will be on the lists in the AGI blocks.  What we do
3175  * here is scan all the AGIs and fully truncate and free any
3176  * inodes found on the lists.  Each inode is removed from the
3177  * lists when it has been fully truncated and is freed.  The
3178  * freeing of the inode and its removal from the list must be
3179  * atomic.
3180  */
3181 void
3182 xlog_recover_process_iunlinks(
3183 	xlog_t		*log)
3184 {
3185 	xfs_mount_t	*mp;
3186 	xfs_agnumber_t	agno;
3187 	xfs_agi_t	*agi;
3188 	xfs_buf_t	*agibp;
3189 	xfs_buf_t	*ibp;
3190 	xfs_dinode_t	*dip;
3191 	xfs_inode_t	*ip;
3192 	xfs_agino_t	agino;
3193 	xfs_ino_t	ino;
3194 	int		bucket;
3195 	int		error;
3196 	uint		mp_dmevmask;
3197 
3198 	mp = log->l_mp;
3199 
3200 	/*
3201 	 * Prevent any DMAPI event from being sent while in this function.
3202 	 */
3203 	mp_dmevmask = mp->m_dmevmask;
3204 	mp->m_dmevmask = 0;
3205 
3206 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3207 		/*
3208 		 * Find the agi for this ag.
3209 		 */
3210 		agibp = xfs_buf_read(mp->m_ddev_targp,
3211 				XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3212 				XFS_FSS_TO_BB(mp, 1), 0);
3213 		if (XFS_BUF_ISERROR(agibp)) {
3214 			xfs_ioerror_alert("xlog_recover_process_iunlinks(#1)",
3215 				log->l_mp, agibp,
3216 				XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)));
3217 		}
3218 		agi = XFS_BUF_TO_AGI(agibp);
3219 		ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agi->agi_magicnum));
3220 
3221 		for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3222 
3223 			agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3224 			while (agino != NULLAGINO) {
3225 
3226 				/*
3227 				 * Release the agi buffer so that it can
3228 				 * be acquired in the normal course of the
3229 				 * transaction to truncate and free the inode.
3230 				 */
3231 				xfs_buf_relse(agibp);
3232 
3233 				ino = XFS_AGINO_TO_INO(mp, agno, agino);
3234 				error = xfs_iget(mp, NULL, ino, 0, 0, &ip, 0);
3235 				ASSERT(error || (ip != NULL));
3236 
3237 				if (!error) {
3238 					/*
3239 					 * Get the on disk inode to find the
3240 					 * next inode in the bucket.
3241 					 */
3242 					error = xfs_itobp(mp, NULL, ip, &dip,
3243 							&ibp, 0, 0,
3244 							XFS_BUF_LOCK);
3245 					ASSERT(error || (dip != NULL));
3246 				}
3247 
3248 				if (!error) {
3249 					ASSERT(ip->i_d.di_nlink == 0);
3250 
3251 					/* setup for the next pass */
3252 					agino = be32_to_cpu(
3253 							dip->di_next_unlinked);
3254 					xfs_buf_relse(ibp);
3255 					/*
3256 					 * Prevent any DMAPI event from
3257 					 * being sent when the
3258 					 * reference on the inode is
3259 					 * dropped.
3260 					 */
3261 					ip->i_d.di_dmevmask = 0;
3262 
3263 					/*
3264 					 * If this is a new inode, handle
3265 					 * it specially.  Otherwise,
3266 					 * just drop our reference to the
3267 					 * inode.  If there are no
3268 					 * other references, this will
3269 					 * send the inode to
3270 					 * xfs_inactive() which will
3271 					 * truncate the file and free
3272 					 * the inode.
3273 					 */
3274 					if (ip->i_d.di_mode == 0)
3275 						xfs_iput_new(ip, 0);
3276 					else
3277 						IRELE(ip);
3278 				} else {
3279 					/*
3280 					 * We can't read in the inode
3281 					 * this bucket points to, or
3282 					 * this inode is messed up.  Just
3283 					 * ditch this bucket of inodes.  We
3284 					 * will lose some inodes and space,
3285 					 * but at least we won't hang.  Call
3286 					 * xlog_recover_clear_agi_bucket()
3287 					 * to perform a transaction to clear
3288 					 * the inode pointer in the bucket.
3289 					 */
3290 					xlog_recover_clear_agi_bucket(mp, agno,
3291 							bucket);
3292 
3293 					agino = NULLAGINO;
3294 				}
3295 
3296 				/*
3297 				 * Reacquire the agibuffer and continue around
3298 				 * the loop.
3299 				 */
3300 				agibp = xfs_buf_read(mp->m_ddev_targp,
3301 						XFS_AG_DADDR(mp, agno,
3302 							XFS_AGI_DADDR(mp)),
3303 						XFS_FSS_TO_BB(mp, 1), 0);
3304 				if (XFS_BUF_ISERROR(agibp)) {
3305 					xfs_ioerror_alert(
3306 				"xlog_recover_process_iunlinks(#2)",
3307 						log->l_mp, agibp,
3308 						XFS_AG_DADDR(mp, agno,
3309 							XFS_AGI_DADDR(mp)));
3310 				}
3311 				agi = XFS_BUF_TO_AGI(agibp);
3312 				ASSERT(XFS_AGI_MAGIC == be32_to_cpu(
3313 					agi->agi_magicnum));
3314 			}
3315 		}
3316 
3317 		/*
3318 		 * Release the buffer for the current agi so we can
3319 		 * go on to the next one.
3320 		 */
3321 		xfs_buf_relse(agibp);
3322 	}
3323 
3324 	mp->m_dmevmask = mp_dmevmask;
3325 }
3326 
3327 
3328 #ifdef DEBUG
3329 STATIC void
3330 xlog_pack_data_checksum(
3331 	xlog_t		*log,
3332 	xlog_in_core_t	*iclog,
3333 	int		size)
3334 {
3335 	int		i;
3336 	__be32		*up;
3337 	uint		chksum = 0;
3338 
3339 	up = (__be32 *)iclog->ic_datap;
3340 	/* divide length by 4 to get # words */
3341 	for (i = 0; i < (size >> 2); i++) {
3342 		chksum ^= be32_to_cpu(*up);
3343 		up++;
3344 	}
3345 	iclog->ic_header.h_chksum = cpu_to_be32(chksum);
3346 }
3347 #else
3348 #define xlog_pack_data_checksum(log, iclog, size)
3349 #endif
3350 
3351 /*
3352  * Stamp cycle number in every block
3353  */
3354 void
3355 xlog_pack_data(
3356 	xlog_t			*log,
3357 	xlog_in_core_t		*iclog,
3358 	int			roundoff)
3359 {
3360 	int			i, j, k;
3361 	int			size = iclog->ic_offset + roundoff;
3362 	__be32			cycle_lsn;
3363 	xfs_caddr_t		dp;
3364 	xlog_in_core_2_t	*xhdr;
3365 
3366 	xlog_pack_data_checksum(log, iclog, size);
3367 
3368 	cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3369 
3370 	dp = iclog->ic_datap;
3371 	for (i = 0; i < BTOBB(size) &&
3372 		i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3373 		iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
3374 		*(__be32 *)dp = cycle_lsn;
3375 		dp += BBSIZE;
3376 	}
3377 
3378 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3379 		xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3380 		for ( ; i < BTOBB(size); i++) {
3381 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3382 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3383 			xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
3384 			*(__be32 *)dp = cycle_lsn;
3385 			dp += BBSIZE;
3386 		}
3387 
3388 		for (i = 1; i < log->l_iclog_heads; i++) {
3389 			xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3390 		}
3391 	}
3392 }
3393 
3394 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
3395 STATIC void
3396 xlog_unpack_data_checksum(
3397 	xlog_rec_header_t	*rhead,
3398 	xfs_caddr_t		dp,
3399 	xlog_t			*log)
3400 {
3401 	__be32			*up = (__be32 *)dp;
3402 	uint			chksum = 0;
3403 	int			i;
3404 
3405 	/* divide length by 4 to get # words */
3406 	for (i=0; i < be32_to_cpu(rhead->h_len) >> 2; i++) {
3407 		chksum ^= be32_to_cpu(*up);
3408 		up++;
3409 	}
3410 	if (chksum != be32_to_cpu(rhead->h_chksum)) {
3411 	    if (rhead->h_chksum ||
3412 		((log->l_flags & XLOG_CHKSUM_MISMATCH) == 0)) {
3413 		    cmn_err(CE_DEBUG,
3414 			"XFS: LogR chksum mismatch: was (0x%x) is (0x%x)\n",
3415 			    be32_to_cpu(rhead->h_chksum), chksum);
3416 		    cmn_err(CE_DEBUG,
3417 "XFS: Disregard message if filesystem was created with non-DEBUG kernel");
3418 		    if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3419 			    cmn_err(CE_DEBUG,
3420 				"XFS: LogR this is a LogV2 filesystem\n");
3421 		    }
3422 		    log->l_flags |= XLOG_CHKSUM_MISMATCH;
3423 	    }
3424 	}
3425 }
3426 #else
3427 #define xlog_unpack_data_checksum(rhead, dp, log)
3428 #endif
3429 
3430 STATIC void
3431 xlog_unpack_data(
3432 	xlog_rec_header_t	*rhead,
3433 	xfs_caddr_t		dp,
3434 	xlog_t			*log)
3435 {
3436 	int			i, j, k;
3437 	xlog_in_core_2_t	*xhdr;
3438 
3439 	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3440 		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3441 		*(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3442 		dp += BBSIZE;
3443 	}
3444 
3445 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3446 		xhdr = (xlog_in_core_2_t *)rhead;
3447 		for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3448 			j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3449 			k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3450 			*(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3451 			dp += BBSIZE;
3452 		}
3453 	}
3454 
3455 	xlog_unpack_data_checksum(rhead, dp, log);
3456 }
3457 
3458 STATIC int
3459 xlog_valid_rec_header(
3460 	xlog_t			*log,
3461 	xlog_rec_header_t	*rhead,
3462 	xfs_daddr_t		blkno)
3463 {
3464 	int			hlen;
3465 
3466 	if (unlikely(be32_to_cpu(rhead->h_magicno) != XLOG_HEADER_MAGIC_NUM)) {
3467 		XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3468 				XFS_ERRLEVEL_LOW, log->l_mp);
3469 		return XFS_ERROR(EFSCORRUPTED);
3470 	}
3471 	if (unlikely(
3472 	    (!rhead->h_version ||
3473 	    (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3474 		xlog_warn("XFS: %s: unrecognised log version (%d).",
3475 			__func__, be32_to_cpu(rhead->h_version));
3476 		return XFS_ERROR(EIO);
3477 	}
3478 
3479 	/* LR body must have data or it wouldn't have been written */
3480 	hlen = be32_to_cpu(rhead->h_len);
3481 	if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3482 		XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3483 				XFS_ERRLEVEL_LOW, log->l_mp);
3484 		return XFS_ERROR(EFSCORRUPTED);
3485 	}
3486 	if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3487 		XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3488 				XFS_ERRLEVEL_LOW, log->l_mp);
3489 		return XFS_ERROR(EFSCORRUPTED);
3490 	}
3491 	return 0;
3492 }
3493 
3494 /*
3495  * Read the log from tail to head and process the log records found.
3496  * Handle the two cases where the tail and head are in the same cycle
3497  * and where the active portion of the log wraps around the end of
3498  * the physical log separately.  The pass parameter is passed through
3499  * to the routines called to process the data and is not looked at
3500  * here.
3501  */
3502 STATIC int
3503 xlog_do_recovery_pass(
3504 	xlog_t			*log,
3505 	xfs_daddr_t		head_blk,
3506 	xfs_daddr_t		tail_blk,
3507 	int			pass)
3508 {
3509 	xlog_rec_header_t	*rhead;
3510 	xfs_daddr_t		blk_no;
3511 	xfs_caddr_t		bufaddr, offset;
3512 	xfs_buf_t		*hbp, *dbp;
3513 	int			error = 0, h_size;
3514 	int			bblks, split_bblks;
3515 	int			hblks, split_hblks, wrapped_hblks;
3516 	xlog_recover_t		*rhash[XLOG_RHASH_SIZE];
3517 
3518 	ASSERT(head_blk != tail_blk);
3519 
3520 	/*
3521 	 * Read the header of the tail block and get the iclog buffer size from
3522 	 * h_size.  Use this to tell how many sectors make up the log header.
3523 	 */
3524 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3525 		/*
3526 		 * When using variable length iclogs, read first sector of
3527 		 * iclog header and extract the header size from it.  Get a
3528 		 * new hbp that is the correct size.
3529 		 */
3530 		hbp = xlog_get_bp(log, 1);
3531 		if (!hbp)
3532 			return ENOMEM;
3533 		if ((error = xlog_bread(log, tail_blk, 1, hbp)))
3534 			goto bread_err1;
3535 		offset = xlog_align(log, tail_blk, 1, hbp);
3536 		rhead = (xlog_rec_header_t *)offset;
3537 		error = xlog_valid_rec_header(log, rhead, tail_blk);
3538 		if (error)
3539 			goto bread_err1;
3540 		h_size = be32_to_cpu(rhead->h_size);
3541 		if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3542 		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3543 			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3544 			if (h_size % XLOG_HEADER_CYCLE_SIZE)
3545 				hblks++;
3546 			xlog_put_bp(hbp);
3547 			hbp = xlog_get_bp(log, hblks);
3548 		} else {
3549 			hblks = 1;
3550 		}
3551 	} else {
3552 		ASSERT(log->l_sectbb_log == 0);
3553 		hblks = 1;
3554 		hbp = xlog_get_bp(log, 1);
3555 		h_size = XLOG_BIG_RECORD_BSIZE;
3556 	}
3557 
3558 	if (!hbp)
3559 		return ENOMEM;
3560 	dbp = xlog_get_bp(log, BTOBB(h_size));
3561 	if (!dbp) {
3562 		xlog_put_bp(hbp);
3563 		return ENOMEM;
3564 	}
3565 
3566 	memset(rhash, 0, sizeof(rhash));
3567 	if (tail_blk <= head_blk) {
3568 		for (blk_no = tail_blk; blk_no < head_blk; ) {
3569 			if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3570 				goto bread_err2;
3571 			offset = xlog_align(log, blk_no, hblks, hbp);
3572 			rhead = (xlog_rec_header_t *)offset;
3573 			error = xlog_valid_rec_header(log, rhead, blk_no);
3574 			if (error)
3575 				goto bread_err2;
3576 
3577 			/* blocks in data section */
3578 			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3579 			error = xlog_bread(log, blk_no + hblks, bblks, dbp);
3580 			if (error)
3581 				goto bread_err2;
3582 			offset = xlog_align(log, blk_no + hblks, bblks, dbp);
3583 			xlog_unpack_data(rhead, offset, log);
3584 			if ((error = xlog_recover_process_data(log,
3585 						rhash, rhead, offset, pass)))
3586 				goto bread_err2;
3587 			blk_no += bblks + hblks;
3588 		}
3589 	} else {
3590 		/*
3591 		 * Perform recovery around the end of the physical log.
3592 		 * When the head is not on the same cycle number as the tail,
3593 		 * we can't do a sequential recovery as above.
3594 		 */
3595 		blk_no = tail_blk;
3596 		while (blk_no < log->l_logBBsize) {
3597 			/*
3598 			 * Check for header wrapping around physical end-of-log
3599 			 */
3600 			offset = NULL;
3601 			split_hblks = 0;
3602 			wrapped_hblks = 0;
3603 			if (blk_no + hblks <= log->l_logBBsize) {
3604 				/* Read header in one read */
3605 				error = xlog_bread(log, blk_no, hblks, hbp);
3606 				if (error)
3607 					goto bread_err2;
3608 				offset = xlog_align(log, blk_no, hblks, hbp);
3609 			} else {
3610 				/* This LR is split across physical log end */
3611 				if (blk_no != log->l_logBBsize) {
3612 					/* some data before physical log end */
3613 					ASSERT(blk_no <= INT_MAX);
3614 					split_hblks = log->l_logBBsize - (int)blk_no;
3615 					ASSERT(split_hblks > 0);
3616 					if ((error = xlog_bread(log, blk_no,
3617 							split_hblks, hbp)))
3618 						goto bread_err2;
3619 					offset = xlog_align(log, blk_no,
3620 							split_hblks, hbp);
3621 				}
3622 				/*
3623 				 * Note: this black magic still works with
3624 				 * large sector sizes (non-512) only because:
3625 				 * - we increased the buffer size originally
3626 				 *   by 1 sector giving us enough extra space
3627 				 *   for the second read;
3628 				 * - the log start is guaranteed to be sector
3629 				 *   aligned;
3630 				 * - we read the log end (LR header start)
3631 				 *   _first_, then the log start (LR header end)
3632 				 *   - order is important.
3633 				 */
3634 				wrapped_hblks = hblks - split_hblks;
3635 				bufaddr = XFS_BUF_PTR(hbp);
3636 				error = XFS_BUF_SET_PTR(hbp,
3637 						bufaddr + BBTOB(split_hblks),
3638 						BBTOB(hblks - split_hblks));
3639 				if (!error)
3640 					error = xlog_bread(log, 0,
3641 							wrapped_hblks, hbp);
3642 				if (!error)
3643 					error = XFS_BUF_SET_PTR(hbp, bufaddr,
3644 							BBTOB(hblks));
3645 				if (error)
3646 					goto bread_err2;
3647 				if (!offset)
3648 					offset = xlog_align(log, 0,
3649 							wrapped_hblks, hbp);
3650 			}
3651 			rhead = (xlog_rec_header_t *)offset;
3652 			error = xlog_valid_rec_header(log, rhead,
3653 						split_hblks ? blk_no : 0);
3654 			if (error)
3655 				goto bread_err2;
3656 
3657 			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3658 			blk_no += hblks;
3659 
3660 			/* Read in data for log record */
3661 			if (blk_no + bblks <= log->l_logBBsize) {
3662 				error = xlog_bread(log, blk_no, bblks, dbp);
3663 				if (error)
3664 					goto bread_err2;
3665 				offset = xlog_align(log, blk_no, bblks, dbp);
3666 			} else {
3667 				/* This log record is split across the
3668 				 * physical end of log */
3669 				offset = NULL;
3670 				split_bblks = 0;
3671 				if (blk_no != log->l_logBBsize) {
3672 					/* some data is before the physical
3673 					 * end of log */
3674 					ASSERT(!wrapped_hblks);
3675 					ASSERT(blk_no <= INT_MAX);
3676 					split_bblks =
3677 						log->l_logBBsize - (int)blk_no;
3678 					ASSERT(split_bblks > 0);
3679 					if ((error = xlog_bread(log, blk_no,
3680 							split_bblks, dbp)))
3681 						goto bread_err2;
3682 					offset = xlog_align(log, blk_no,
3683 							split_bblks, dbp);
3684 				}
3685 				/*
3686 				 * Note: this black magic still works with
3687 				 * large sector sizes (non-512) only because:
3688 				 * - we increased the buffer size originally
3689 				 *   by 1 sector giving us enough extra space
3690 				 *   for the second read;
3691 				 * - the log start is guaranteed to be sector
3692 				 *   aligned;
3693 				 * - we read the log end (LR header start)
3694 				 *   _first_, then the log start (LR header end)
3695 				 *   - order is important.
3696 				 */
3697 				bufaddr = XFS_BUF_PTR(dbp);
3698 				error = XFS_BUF_SET_PTR(dbp,
3699 						bufaddr + BBTOB(split_bblks),
3700 						BBTOB(bblks - split_bblks));
3701 				if (!error)
3702 					error = xlog_bread(log, wrapped_hblks,
3703 							bblks - split_bblks,
3704 							dbp);
3705 				if (!error)
3706 					error = XFS_BUF_SET_PTR(dbp, bufaddr,
3707 							h_size);
3708 				if (error)
3709 					goto bread_err2;
3710 				if (!offset)
3711 					offset = xlog_align(log, wrapped_hblks,
3712 						bblks - split_bblks, dbp);
3713 			}
3714 			xlog_unpack_data(rhead, offset, log);
3715 			if ((error = xlog_recover_process_data(log, rhash,
3716 							rhead, offset, pass)))
3717 				goto bread_err2;
3718 			blk_no += bblks;
3719 		}
3720 
3721 		ASSERT(blk_no >= log->l_logBBsize);
3722 		blk_no -= log->l_logBBsize;
3723 
3724 		/* read first part of physical log */
3725 		while (blk_no < head_blk) {
3726 			if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3727 				goto bread_err2;
3728 			offset = xlog_align(log, blk_no, hblks, hbp);
3729 			rhead = (xlog_rec_header_t *)offset;
3730 			error = xlog_valid_rec_header(log, rhead, blk_no);
3731 			if (error)
3732 				goto bread_err2;
3733 			bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3734 			if ((error = xlog_bread(log, blk_no+hblks, bblks, dbp)))
3735 				goto bread_err2;
3736 			offset = xlog_align(log, blk_no+hblks, bblks, dbp);
3737 			xlog_unpack_data(rhead, offset, log);
3738 			if ((error = xlog_recover_process_data(log, rhash,
3739 							rhead, offset, pass)))
3740 				goto bread_err2;
3741 			blk_no += bblks + hblks;
3742 		}
3743 	}
3744 
3745  bread_err2:
3746 	xlog_put_bp(dbp);
3747  bread_err1:
3748 	xlog_put_bp(hbp);
3749 	return error;
3750 }
3751 
3752 /*
3753  * Do the recovery of the log.  We actually do this in two phases.
3754  * The two passes are necessary in order to implement the function
3755  * of cancelling a record written into the log.  The first pass
3756  * determines those things which have been cancelled, and the
3757  * second pass replays log items normally except for those which
3758  * have been cancelled.  The handling of the replay and cancellations
3759  * takes place in the log item type specific routines.
3760  *
3761  * The table of items which have cancel records in the log is allocated
3762  * and freed at this level, since only here do we know when all of
3763  * the log recovery has been completed.
3764  */
3765 STATIC int
3766 xlog_do_log_recovery(
3767 	xlog_t		*log,
3768 	xfs_daddr_t	head_blk,
3769 	xfs_daddr_t	tail_blk)
3770 {
3771 	int		error;
3772 
3773 	ASSERT(head_blk != tail_blk);
3774 
3775 	/*
3776 	 * First do a pass to find all of the cancelled buf log items.
3777 	 * Store them in the buf_cancel_table for use in the second pass.
3778 	 */
3779 	log->l_buf_cancel_table =
3780 		(xfs_buf_cancel_t **)kmem_zalloc(XLOG_BC_TABLE_SIZE *
3781 						 sizeof(xfs_buf_cancel_t*),
3782 						 KM_SLEEP);
3783 	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3784 				      XLOG_RECOVER_PASS1);
3785 	if (error != 0) {
3786 		kmem_free(log->l_buf_cancel_table);
3787 		log->l_buf_cancel_table = NULL;
3788 		return error;
3789 	}
3790 	/*
3791 	 * Then do a second pass to actually recover the items in the log.
3792 	 * When it is complete free the table of buf cancel items.
3793 	 */
3794 	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3795 				      XLOG_RECOVER_PASS2);
3796 #ifdef DEBUG
3797 	if (!error) {
3798 		int	i;
3799 
3800 		for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3801 			ASSERT(log->l_buf_cancel_table[i] == NULL);
3802 	}
3803 #endif	/* DEBUG */
3804 
3805 	kmem_free(log->l_buf_cancel_table);
3806 	log->l_buf_cancel_table = NULL;
3807 
3808 	return error;
3809 }
3810 
3811 /*
3812  * Do the actual recovery
3813  */
3814 STATIC int
3815 xlog_do_recover(
3816 	xlog_t		*log,
3817 	xfs_daddr_t	head_blk,
3818 	xfs_daddr_t	tail_blk)
3819 {
3820 	int		error;
3821 	xfs_buf_t	*bp;
3822 	xfs_sb_t	*sbp;
3823 
3824 	/*
3825 	 * First replay the images in the log.
3826 	 */
3827 	error = xlog_do_log_recovery(log, head_blk, tail_blk);
3828 	if (error) {
3829 		return error;
3830 	}
3831 
3832 	XFS_bflush(log->l_mp->m_ddev_targp);
3833 
3834 	/*
3835 	 * If IO errors happened during recovery, bail out.
3836 	 */
3837 	if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3838 		return (EIO);
3839 	}
3840 
3841 	/*
3842 	 * We now update the tail_lsn since much of the recovery has completed
3843 	 * and there may be space available to use.  If there were no extent
3844 	 * or iunlinks, we can free up the entire log and set the tail_lsn to
3845 	 * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3846 	 * lsn of the last known good LR on disk.  If there are extent frees
3847 	 * or iunlinks they will have some entries in the AIL; so we look at
3848 	 * the AIL to determine how to set the tail_lsn.
3849 	 */
3850 	xlog_assign_tail_lsn(log->l_mp);
3851 
3852 	/*
3853 	 * Now that we've finished replaying all buffer and inode
3854 	 * updates, re-read in the superblock.
3855 	 */
3856 	bp = xfs_getsb(log->l_mp, 0);
3857 	XFS_BUF_UNDONE(bp);
3858 	ASSERT(!(XFS_BUF_ISWRITE(bp)));
3859 	ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
3860 	XFS_BUF_READ(bp);
3861 	XFS_BUF_UNASYNC(bp);
3862 	xfsbdstrat(log->l_mp, bp);
3863 	error = xfs_iowait(bp);
3864 	if (error) {
3865 		xfs_ioerror_alert("xlog_do_recover",
3866 				  log->l_mp, bp, XFS_BUF_ADDR(bp));
3867 		ASSERT(0);
3868 		xfs_buf_relse(bp);
3869 		return error;
3870 	}
3871 
3872 	/* Convert superblock from on-disk format */
3873 	sbp = &log->l_mp->m_sb;
3874 	xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3875 	ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3876 	ASSERT(xfs_sb_good_version(sbp));
3877 	xfs_buf_relse(bp);
3878 
3879 	/* We've re-read the superblock so re-initialize per-cpu counters */
3880 	xfs_icsb_reinit_counters(log->l_mp);
3881 
3882 	xlog_recover_check_summary(log);
3883 
3884 	/* Normal transactions can now occur */
3885 	log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3886 	return 0;
3887 }
3888 
3889 /*
3890  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3891  *
3892  * Return error or zero.
3893  */
3894 int
3895 xlog_recover(
3896 	xlog_t		*log)
3897 {
3898 	xfs_daddr_t	head_blk, tail_blk;
3899 	int		error;
3900 
3901 	/* find the tail of the log */
3902 	if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3903 		return error;
3904 
3905 	if (tail_blk != head_blk) {
3906 		/* There used to be a comment here:
3907 		 *
3908 		 * disallow recovery on read-only mounts.  note -- mount
3909 		 * checks for ENOSPC and turns it into an intelligent
3910 		 * error message.
3911 		 * ...but this is no longer true.  Now, unless you specify
3912 		 * NORECOVERY (in which case this function would never be
3913 		 * called), we just go ahead and recover.  We do this all
3914 		 * under the vfs layer, so we can get away with it unless
3915 		 * the device itself is read-only, in which case we fail.
3916 		 */
3917 		if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3918 			return error;
3919 		}
3920 
3921 		cmn_err(CE_NOTE,
3922 			"Starting XFS recovery on filesystem: %s (logdev: %s)",
3923 			log->l_mp->m_fsname, log->l_mp->m_logname ?
3924 			log->l_mp->m_logname : "internal");
3925 
3926 		error = xlog_do_recover(log, head_blk, tail_blk);
3927 		log->l_flags |= XLOG_RECOVERY_NEEDED;
3928 	}
3929 	return error;
3930 }
3931 
3932 /*
3933  * In the first part of recovery we replay inodes and buffers and build
3934  * up the list of extent free items which need to be processed.  Here
3935  * we process the extent free items and clean up the on disk unlinked
3936  * inode lists.  This is separated from the first part of recovery so
3937  * that the root and real-time bitmap inodes can be read in from disk in
3938  * between the two stages.  This is necessary so that we can free space
3939  * in the real-time portion of the file system.
3940  */
3941 int
3942 xlog_recover_finish(
3943 	xlog_t		*log)
3944 {
3945 	/*
3946 	 * Now we're ready to do the transactions needed for the
3947 	 * rest of recovery.  Start with completing all the extent
3948 	 * free intent records and then process the unlinked inode
3949 	 * lists.  At this point, we essentially run in normal mode
3950 	 * except that we're still performing recovery actions
3951 	 * rather than accepting new requests.
3952 	 */
3953 	if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3954 		int	error;
3955 		error = xlog_recover_process_efis(log);
3956 		if (error) {
3957 			cmn_err(CE_ALERT,
3958 				"Failed to recover EFIs on filesystem: %s",
3959 				log->l_mp->m_fsname);
3960 			return error;
3961 		}
3962 		/*
3963 		 * Sync the log to get all the EFIs out of the AIL.
3964 		 * This isn't absolutely necessary, but it helps in
3965 		 * case the unlink transactions would have problems
3966 		 * pushing the EFIs out of the way.
3967 		 */
3968 		xfs_log_force(log->l_mp, (xfs_lsn_t)0,
3969 			      (XFS_LOG_FORCE | XFS_LOG_SYNC));
3970 
3971 		xlog_recover_process_iunlinks(log);
3972 
3973 		xlog_recover_check_summary(log);
3974 
3975 		cmn_err(CE_NOTE,
3976 			"Ending XFS recovery on filesystem: %s (logdev: %s)",
3977 			log->l_mp->m_fsname, log->l_mp->m_logname ?
3978 			log->l_mp->m_logname : "internal");
3979 		log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3980 	} else {
3981 		cmn_err(CE_DEBUG,
3982 			"!Ending clean XFS mount for filesystem: %s\n",
3983 			log->l_mp->m_fsname);
3984 	}
3985 	return 0;
3986 }
3987 
3988 
3989 #if defined(DEBUG)
3990 /*
3991  * Read all of the agf and agi counters and check that they
3992  * are consistent with the superblock counters.
3993  */
3994 void
3995 xlog_recover_check_summary(
3996 	xlog_t		*log)
3997 {
3998 	xfs_mount_t	*mp;
3999 	xfs_agf_t	*agfp;
4000 	xfs_agi_t	*agip;
4001 	xfs_buf_t	*agfbp;
4002 	xfs_buf_t	*agibp;
4003 	xfs_daddr_t	agfdaddr;
4004 	xfs_daddr_t	agidaddr;
4005 	xfs_buf_t	*sbbp;
4006 #ifdef XFS_LOUD_RECOVERY
4007 	xfs_sb_t	*sbp;
4008 #endif
4009 	xfs_agnumber_t	agno;
4010 	__uint64_t	freeblks;
4011 	__uint64_t	itotal;
4012 	__uint64_t	ifree;
4013 
4014 	mp = log->l_mp;
4015 
4016 	freeblks = 0LL;
4017 	itotal = 0LL;
4018 	ifree = 0LL;
4019 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4020 		agfdaddr = XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp));
4021 		agfbp = xfs_buf_read(mp->m_ddev_targp, agfdaddr,
4022 				XFS_FSS_TO_BB(mp, 1), 0);
4023 		if (XFS_BUF_ISERROR(agfbp)) {
4024 			xfs_ioerror_alert("xlog_recover_check_summary(agf)",
4025 						mp, agfbp, agfdaddr);
4026 		}
4027 		agfp = XFS_BUF_TO_AGF(agfbp);
4028 		ASSERT(XFS_AGF_MAGIC == be32_to_cpu(agfp->agf_magicnum));
4029 		ASSERT(XFS_AGF_GOOD_VERSION(be32_to_cpu(agfp->agf_versionnum)));
4030 		ASSERT(be32_to_cpu(agfp->agf_seqno) == agno);
4031 
4032 		freeblks += be32_to_cpu(agfp->agf_freeblks) +
4033 			    be32_to_cpu(agfp->agf_flcount);
4034 		xfs_buf_relse(agfbp);
4035 
4036 		agidaddr = XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp));
4037 		agibp = xfs_buf_read(mp->m_ddev_targp, agidaddr,
4038 				XFS_FSS_TO_BB(mp, 1), 0);
4039 		if (XFS_BUF_ISERROR(agibp)) {
4040 			xfs_ioerror_alert("xlog_recover_check_summary(agi)",
4041 					  mp, agibp, agidaddr);
4042 		}
4043 		agip = XFS_BUF_TO_AGI(agibp);
4044 		ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agip->agi_magicnum));
4045 		ASSERT(XFS_AGI_GOOD_VERSION(be32_to_cpu(agip->agi_versionnum)));
4046 		ASSERT(be32_to_cpu(agip->agi_seqno) == agno);
4047 
4048 		itotal += be32_to_cpu(agip->agi_count);
4049 		ifree += be32_to_cpu(agip->agi_freecount);
4050 		xfs_buf_relse(agibp);
4051 	}
4052 
4053 	sbbp = xfs_getsb(mp, 0);
4054 #ifdef XFS_LOUD_RECOVERY
4055 	sbp = &mp->m_sb;
4056 	xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(sbbp));
4057 	cmn_err(CE_NOTE,
4058 		"xlog_recover_check_summary: sb_icount %Lu itotal %Lu",
4059 		sbp->sb_icount, itotal);
4060 	cmn_err(CE_NOTE,
4061 		"xlog_recover_check_summary: sb_ifree %Lu itotal %Lu",
4062 		sbp->sb_ifree, ifree);
4063 	cmn_err(CE_NOTE,
4064 		"xlog_recover_check_summary: sb_fdblocks %Lu freeblks %Lu",
4065 		sbp->sb_fdblocks, freeblks);
4066 #if 0
4067 	/*
4068 	 * This is turned off until I account for the allocation
4069 	 * btree blocks which live in free space.
4070 	 */
4071 	ASSERT(sbp->sb_icount == itotal);
4072 	ASSERT(sbp->sb_ifree == ifree);
4073 	ASSERT(sbp->sb_fdblocks == freeblks);
4074 #endif
4075 #endif
4076 	xfs_buf_relse(sbbp);
4077 }
4078 #endif /* DEBUG */
4079