xref: /openbmc/linux/fs/xfs/xfs_log.c (revision 55b66332d0921146a914d5d75a7b870a65dc4938)
1 /*
2  * Copyright (c) 2000-2005 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_log_priv.h"
32 #include "xfs_buf_item.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_log_recover.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_dir2_sf.h"
39 #include "xfs_attr_sf.h"
40 #include "xfs_dinode.h"
41 #include "xfs_inode.h"
42 #include "xfs_rw.h"
43 #include "xfs_trace.h"
44 
45 kmem_zone_t	*xfs_log_ticket_zone;
46 
47 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
48 	{ (ptr) += (bytes); \
49 	  (len) -= (bytes); \
50 	  (off) += (bytes);}
51 
52 /* Local miscellaneous function prototypes */
53 STATIC int	 xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
54 				    xlog_in_core_t **, xfs_lsn_t *);
55 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t	*mp,
56 				xfs_buftarg_t	*log_target,
57 				xfs_daddr_t	blk_offset,
58 				int		num_bblks);
59 STATIC int	 xlog_space_left(xlog_t *log, int cycle, int bytes);
60 STATIC int	 xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
61 STATIC void	 xlog_dealloc_log(xlog_t *log);
62 STATIC int	 xlog_write(struct log *log, struct xfs_log_vec *log_vector,
63 			    struct xlog_ticket *tic, xfs_lsn_t *start_lsn,
64 			    xlog_in_core_t **commit_iclog, uint flags);
65 
66 /* local state machine functions */
67 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
68 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
69 STATIC int  xlog_state_get_iclog_space(xlog_t		*log,
70 				       int		len,
71 				       xlog_in_core_t	**iclog,
72 				       xlog_ticket_t	*ticket,
73 				       int		*continued_write,
74 				       int		*logoffsetp);
75 STATIC int  xlog_state_release_iclog(xlog_t		*log,
76 				     xlog_in_core_t	*iclog);
77 STATIC void xlog_state_switch_iclogs(xlog_t		*log,
78 				     xlog_in_core_t *iclog,
79 				     int		eventual_size);
80 STATIC void xlog_state_want_sync(xlog_t	*log, xlog_in_core_t *iclog);
81 
82 /* local functions to manipulate grant head */
83 STATIC int  xlog_grant_log_space(xlog_t		*log,
84 				 xlog_ticket_t	*xtic);
85 STATIC void xlog_grant_push_ail(xfs_mount_t	*mp,
86 				int		need_bytes);
87 STATIC void xlog_regrant_reserve_log_space(xlog_t	 *log,
88 					   xlog_ticket_t *ticket);
89 STATIC int xlog_regrant_write_log_space(xlog_t		*log,
90 					 xlog_ticket_t  *ticket);
91 STATIC void xlog_ungrant_log_space(xlog_t	 *log,
92 				   xlog_ticket_t *ticket);
93 
94 
95 /* local ticket functions */
96 STATIC xlog_ticket_t	*xlog_ticket_alloc(xlog_t *log,
97 					 int	unit_bytes,
98 					 int	count,
99 					 char	clientid,
100 					 uint	flags);
101 
102 #if defined(DEBUG)
103 STATIC void	xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
104 STATIC void	xlog_verify_grant_head(xlog_t *log, int equals);
105 STATIC void	xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
106 				  int count, boolean_t syncing);
107 STATIC void	xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
108 				     xfs_lsn_t tail_lsn);
109 #else
110 #define xlog_verify_dest_ptr(a,b)
111 #define xlog_verify_grant_head(a,b)
112 #define xlog_verify_iclog(a,b,c,d)
113 #define xlog_verify_tail_lsn(a,b,c)
114 #endif
115 
116 STATIC int	xlog_iclogs_empty(xlog_t *log);
117 
118 
119 static void
120 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
121 {
122 	if (*qp) {
123 		tic->t_next	    = (*qp);
124 		tic->t_prev	    = (*qp)->t_prev;
125 		(*qp)->t_prev->t_next = tic;
126 		(*qp)->t_prev	    = tic;
127 	} else {
128 		tic->t_prev = tic->t_next = tic;
129 		*qp = tic;
130 	}
131 
132 	tic->t_flags |= XLOG_TIC_IN_Q;
133 }
134 
135 static void
136 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
137 {
138 	if (tic == tic->t_next) {
139 		*qp = NULL;
140 	} else {
141 		*qp = tic->t_next;
142 		tic->t_next->t_prev = tic->t_prev;
143 		tic->t_prev->t_next = tic->t_next;
144 	}
145 
146 	tic->t_next = tic->t_prev = NULL;
147 	tic->t_flags &= ~XLOG_TIC_IN_Q;
148 }
149 
150 static void
151 xlog_grant_sub_space(struct log *log, int bytes)
152 {
153 	log->l_grant_write_bytes -= bytes;
154 	if (log->l_grant_write_bytes < 0) {
155 		log->l_grant_write_bytes += log->l_logsize;
156 		log->l_grant_write_cycle--;
157 	}
158 
159 	log->l_grant_reserve_bytes -= bytes;
160 	if ((log)->l_grant_reserve_bytes < 0) {
161 		log->l_grant_reserve_bytes += log->l_logsize;
162 		log->l_grant_reserve_cycle--;
163 	}
164 
165 }
166 
167 static void
168 xlog_grant_add_space_write(struct log *log, int bytes)
169 {
170 	int tmp = log->l_logsize - log->l_grant_write_bytes;
171 	if (tmp > bytes)
172 		log->l_grant_write_bytes += bytes;
173 	else {
174 		log->l_grant_write_cycle++;
175 		log->l_grant_write_bytes = bytes - tmp;
176 	}
177 }
178 
179 static void
180 xlog_grant_add_space_reserve(struct log *log, int bytes)
181 {
182 	int tmp = log->l_logsize - log->l_grant_reserve_bytes;
183 	if (tmp > bytes)
184 		log->l_grant_reserve_bytes += bytes;
185 	else {
186 		log->l_grant_reserve_cycle++;
187 		log->l_grant_reserve_bytes = bytes - tmp;
188 	}
189 }
190 
191 static inline void
192 xlog_grant_add_space(struct log *log, int bytes)
193 {
194 	xlog_grant_add_space_write(log, bytes);
195 	xlog_grant_add_space_reserve(log, bytes);
196 }
197 
198 static void
199 xlog_tic_reset_res(xlog_ticket_t *tic)
200 {
201 	tic->t_res_num = 0;
202 	tic->t_res_arr_sum = 0;
203 	tic->t_res_num_ophdrs = 0;
204 }
205 
206 static void
207 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
208 {
209 	if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
210 		/* add to overflow and start again */
211 		tic->t_res_o_flow += tic->t_res_arr_sum;
212 		tic->t_res_num = 0;
213 		tic->t_res_arr_sum = 0;
214 	}
215 
216 	tic->t_res_arr[tic->t_res_num].r_len = len;
217 	tic->t_res_arr[tic->t_res_num].r_type = type;
218 	tic->t_res_arr_sum += len;
219 	tic->t_res_num++;
220 }
221 
222 /*
223  * NOTES:
224  *
225  *	1. currblock field gets updated at startup and after in-core logs
226  *		marked as with WANT_SYNC.
227  */
228 
229 /*
230  * This routine is called when a user of a log manager ticket is done with
231  * the reservation.  If the ticket was ever used, then a commit record for
232  * the associated transaction is written out as a log operation header with
233  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
234  * a given ticket.  If the ticket was one with a permanent reservation, then
235  * a few operations are done differently.  Permanent reservation tickets by
236  * default don't release the reservation.  They just commit the current
237  * transaction with the belief that the reservation is still needed.  A flag
238  * must be passed in before permanent reservations are actually released.
239  * When these type of tickets are not released, they need to be set into
240  * the inited state again.  By doing this, a start record will be written
241  * out when the next write occurs.
242  */
243 xfs_lsn_t
244 xfs_log_done(
245 	struct xfs_mount	*mp,
246 	struct xlog_ticket	*ticket,
247 	struct xlog_in_core	**iclog,
248 	uint			flags)
249 {
250 	struct log		*log = mp->m_log;
251 	xfs_lsn_t		lsn = 0;
252 
253 	if (XLOG_FORCED_SHUTDOWN(log) ||
254 	    /*
255 	     * If nothing was ever written, don't write out commit record.
256 	     * If we get an error, just continue and give back the log ticket.
257 	     */
258 	    (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
259 	     (xlog_commit_record(log, ticket, iclog, &lsn)))) {
260 		lsn = (xfs_lsn_t) -1;
261 		if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
262 			flags |= XFS_LOG_REL_PERM_RESERV;
263 		}
264 	}
265 
266 
267 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
268 	    (flags & XFS_LOG_REL_PERM_RESERV)) {
269 		trace_xfs_log_done_nonperm(log, ticket);
270 
271 		/*
272 		 * Release ticket if not permanent reservation or a specific
273 		 * request has been made to release a permanent reservation.
274 		 */
275 		xlog_ungrant_log_space(log, ticket);
276 		xfs_log_ticket_put(ticket);
277 	} else {
278 		trace_xfs_log_done_perm(log, ticket);
279 
280 		xlog_regrant_reserve_log_space(log, ticket);
281 		/* If this ticket was a permanent reservation and we aren't
282 		 * trying to release it, reset the inited flags; so next time
283 		 * we write, a start record will be written out.
284 		 */
285 		ticket->t_flags |= XLOG_TIC_INITED;
286 	}
287 
288 	return lsn;
289 }
290 
291 /*
292  * Attaches a new iclog I/O completion callback routine during
293  * transaction commit.  If the log is in error state, a non-zero
294  * return code is handed back and the caller is responsible for
295  * executing the callback at an appropriate time.
296  */
297 int
298 xfs_log_notify(
299 	struct xfs_mount	*mp,
300 	struct xlog_in_core	*iclog,
301 	xfs_log_callback_t	*cb)
302 {
303 	int	abortflg;
304 
305 	spin_lock(&iclog->ic_callback_lock);
306 	abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
307 	if (!abortflg) {
308 		ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
309 			      (iclog->ic_state == XLOG_STATE_WANT_SYNC));
310 		cb->cb_next = NULL;
311 		*(iclog->ic_callback_tail) = cb;
312 		iclog->ic_callback_tail = &(cb->cb_next);
313 	}
314 	spin_unlock(&iclog->ic_callback_lock);
315 	return abortflg;
316 }
317 
318 int
319 xfs_log_release_iclog(
320 	struct xfs_mount	*mp,
321 	struct xlog_in_core	*iclog)
322 {
323 	if (xlog_state_release_iclog(mp->m_log, iclog)) {
324 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
325 		return EIO;
326 	}
327 
328 	return 0;
329 }
330 
331 /*
332  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
333  *	to the reservation.
334  *  2. Potentially, push buffers at tail of log to disk.
335  *
336  * Each reservation is going to reserve extra space for a log record header.
337  * When writes happen to the on-disk log, we don't subtract the length of the
338  * log record header from any reservation.  By wasting space in each
339  * reservation, we prevent over allocation problems.
340  */
341 int
342 xfs_log_reserve(
343 	struct xfs_mount	*mp,
344 	int		 	unit_bytes,
345 	int		 	cnt,
346 	struct xlog_ticket	**ticket,
347 	__uint8_t	 	client,
348 	uint		 	flags,
349 	uint		 	t_type)
350 {
351 	struct log		*log = mp->m_log;
352 	struct xlog_ticket	*internal_ticket;
353 	int			retval = 0;
354 
355 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
356 	ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
357 
358 	if (XLOG_FORCED_SHUTDOWN(log))
359 		return XFS_ERROR(EIO);
360 
361 	XFS_STATS_INC(xs_try_logspace);
362 
363 
364 	if (*ticket != NULL) {
365 		ASSERT(flags & XFS_LOG_PERM_RESERV);
366 		internal_ticket = *ticket;
367 
368 		trace_xfs_log_reserve(log, internal_ticket);
369 
370 		xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
371 		retval = xlog_regrant_write_log_space(log, internal_ticket);
372 	} else {
373 		/* may sleep if need to allocate more tickets */
374 		internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
375 						  client, flags);
376 		if (!internal_ticket)
377 			return XFS_ERROR(ENOMEM);
378 		internal_ticket->t_trans_type = t_type;
379 		*ticket = internal_ticket;
380 
381 		trace_xfs_log_reserve(log, internal_ticket);
382 
383 		xlog_grant_push_ail(mp,
384 				    (internal_ticket->t_unit_res *
385 				     internal_ticket->t_cnt));
386 		retval = xlog_grant_log_space(log, internal_ticket);
387 	}
388 
389 	return retval;
390 }	/* xfs_log_reserve */
391 
392 
393 /*
394  * Mount a log filesystem
395  *
396  * mp		- ubiquitous xfs mount point structure
397  * log_target	- buftarg of on-disk log device
398  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
399  * num_bblocks	- Number of BBSIZE blocks in on-disk log
400  *
401  * Return error or zero.
402  */
403 int
404 xfs_log_mount(
405 	xfs_mount_t	*mp,
406 	xfs_buftarg_t	*log_target,
407 	xfs_daddr_t	blk_offset,
408 	int		num_bblks)
409 {
410 	int		error;
411 
412 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
413 		cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
414 	else {
415 		cmn_err(CE_NOTE,
416 			"!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
417 			mp->m_fsname);
418 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
419 	}
420 
421 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
422 	if (IS_ERR(mp->m_log)) {
423 		error = -PTR_ERR(mp->m_log);
424 		goto out;
425 	}
426 
427 	/*
428 	 * Initialize the AIL now we have a log.
429 	 */
430 	error = xfs_trans_ail_init(mp);
431 	if (error) {
432 		cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
433 		goto out_free_log;
434 	}
435 	mp->m_log->l_ailp = mp->m_ail;
436 
437 	/*
438 	 * skip log recovery on a norecovery mount.  pretend it all
439 	 * just worked.
440 	 */
441 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
442 		int	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
443 
444 		if (readonly)
445 			mp->m_flags &= ~XFS_MOUNT_RDONLY;
446 
447 		error = xlog_recover(mp->m_log);
448 
449 		if (readonly)
450 			mp->m_flags |= XFS_MOUNT_RDONLY;
451 		if (error) {
452 			cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
453 			goto out_destroy_ail;
454 		}
455 	}
456 
457 	/* Normal transactions can now occur */
458 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
459 
460 	return 0;
461 
462 out_destroy_ail:
463 	xfs_trans_ail_destroy(mp);
464 out_free_log:
465 	xlog_dealloc_log(mp->m_log);
466 out:
467 	return error;
468 }
469 
470 /*
471  * Finish the recovery of the file system.  This is separate from
472  * the xfs_log_mount() call, because it depends on the code in
473  * xfs_mountfs() to read in the root and real-time bitmap inodes
474  * between calling xfs_log_mount() and here.
475  *
476  * mp		- ubiquitous xfs mount point structure
477  */
478 int
479 xfs_log_mount_finish(xfs_mount_t *mp)
480 {
481 	int	error;
482 
483 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
484 		error = xlog_recover_finish(mp->m_log);
485 	else {
486 		error = 0;
487 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
488 	}
489 
490 	return error;
491 }
492 
493 /*
494  * Final log writes as part of unmount.
495  *
496  * Mark the filesystem clean as unmount happens.  Note that during relocation
497  * this routine needs to be executed as part of source-bag while the
498  * deallocation must not be done until source-end.
499  */
500 
501 /*
502  * Unmount record used to have a string "Unmount filesystem--" in the
503  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
504  * We just write the magic number now since that particular field isn't
505  * currently architecture converted and "nUmount" is a bit foo.
506  * As far as I know, there weren't any dependencies on the old behaviour.
507  */
508 
509 int
510 xfs_log_unmount_write(xfs_mount_t *mp)
511 {
512 	xlog_t		 *log = mp->m_log;
513 	xlog_in_core_t	 *iclog;
514 #ifdef DEBUG
515 	xlog_in_core_t	 *first_iclog;
516 #endif
517 	xlog_ticket_t	*tic = NULL;
518 	xfs_lsn_t	 lsn;
519 	int		 error;
520 
521 	/*
522 	 * Don't write out unmount record on read-only mounts.
523 	 * Or, if we are doing a forced umount (typically because of IO errors).
524 	 */
525 	if (mp->m_flags & XFS_MOUNT_RDONLY)
526 		return 0;
527 
528 	error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
529 	ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
530 
531 #ifdef DEBUG
532 	first_iclog = iclog = log->l_iclog;
533 	do {
534 		if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
535 			ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
536 			ASSERT(iclog->ic_offset == 0);
537 		}
538 		iclog = iclog->ic_next;
539 	} while (iclog != first_iclog);
540 #endif
541 	if (! (XLOG_FORCED_SHUTDOWN(log))) {
542 		error = xfs_log_reserve(mp, 600, 1, &tic,
543 					XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
544 		if (!error) {
545 			/* the data section must be 32 bit size aligned */
546 			struct {
547 			    __uint16_t magic;
548 			    __uint16_t pad1;
549 			    __uint32_t pad2; /* may as well make it 64 bits */
550 			} magic = {
551 				.magic = XLOG_UNMOUNT_TYPE,
552 			};
553 			struct xfs_log_iovec reg = {
554 				.i_addr = (void *)&magic,
555 				.i_len = sizeof(magic),
556 				.i_type = XLOG_REG_TYPE_UNMOUNT,
557 			};
558 			struct xfs_log_vec vec = {
559 				.lv_niovecs = 1,
560 				.lv_iovecp = &reg,
561 			};
562 
563 			/* remove inited flag */
564 			tic->t_flags = 0;
565 			error = xlog_write(log, &vec, tic, &lsn,
566 					   NULL, XLOG_UNMOUNT_TRANS);
567 			/*
568 			 * At this point, we're umounting anyway,
569 			 * so there's no point in transitioning log state
570 			 * to IOERROR. Just continue...
571 			 */
572 		}
573 
574 		if (error) {
575 			xfs_fs_cmn_err(CE_ALERT, mp,
576 				"xfs_log_unmount: unmount record failed");
577 		}
578 
579 
580 		spin_lock(&log->l_icloglock);
581 		iclog = log->l_iclog;
582 		atomic_inc(&iclog->ic_refcnt);
583 		xlog_state_want_sync(log, iclog);
584 		spin_unlock(&log->l_icloglock);
585 		error = xlog_state_release_iclog(log, iclog);
586 
587 		spin_lock(&log->l_icloglock);
588 		if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
589 		      iclog->ic_state == XLOG_STATE_DIRTY)) {
590 			if (!XLOG_FORCED_SHUTDOWN(log)) {
591 				sv_wait(&iclog->ic_force_wait, PMEM,
592 					&log->l_icloglock, s);
593 			} else {
594 				spin_unlock(&log->l_icloglock);
595 			}
596 		} else {
597 			spin_unlock(&log->l_icloglock);
598 		}
599 		if (tic) {
600 			trace_xfs_log_umount_write(log, tic);
601 			xlog_ungrant_log_space(log, tic);
602 			xfs_log_ticket_put(tic);
603 		}
604 	} else {
605 		/*
606 		 * We're already in forced_shutdown mode, couldn't
607 		 * even attempt to write out the unmount transaction.
608 		 *
609 		 * Go through the motions of sync'ing and releasing
610 		 * the iclog, even though no I/O will actually happen,
611 		 * we need to wait for other log I/Os that may already
612 		 * be in progress.  Do this as a separate section of
613 		 * code so we'll know if we ever get stuck here that
614 		 * we're in this odd situation of trying to unmount
615 		 * a file system that went into forced_shutdown as
616 		 * the result of an unmount..
617 		 */
618 		spin_lock(&log->l_icloglock);
619 		iclog = log->l_iclog;
620 		atomic_inc(&iclog->ic_refcnt);
621 
622 		xlog_state_want_sync(log, iclog);
623 		spin_unlock(&log->l_icloglock);
624 		error =  xlog_state_release_iclog(log, iclog);
625 
626 		spin_lock(&log->l_icloglock);
627 
628 		if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
629 			|| iclog->ic_state == XLOG_STATE_DIRTY
630 			|| iclog->ic_state == XLOG_STATE_IOERROR) ) {
631 
632 				sv_wait(&iclog->ic_force_wait, PMEM,
633 					&log->l_icloglock, s);
634 		} else {
635 			spin_unlock(&log->l_icloglock);
636 		}
637 	}
638 
639 	return error;
640 }	/* xfs_log_unmount_write */
641 
642 /*
643  * Deallocate log structures for unmount/relocation.
644  *
645  * We need to stop the aild from running before we destroy
646  * and deallocate the log as the aild references the log.
647  */
648 void
649 xfs_log_unmount(xfs_mount_t *mp)
650 {
651 	xfs_trans_ail_destroy(mp);
652 	xlog_dealloc_log(mp->m_log);
653 }
654 
655 void
656 xfs_log_item_init(
657 	struct xfs_mount	*mp,
658 	struct xfs_log_item	*item,
659 	int			type,
660 	struct xfs_item_ops	*ops)
661 {
662 	item->li_mountp = mp;
663 	item->li_ailp = mp->m_ail;
664 	item->li_type = type;
665 	item->li_ops = ops;
666 }
667 
668 /*
669  * Write region vectors to log.  The write happens using the space reservation
670  * of the ticket (tic).  It is not a requirement that all writes for a given
671  * transaction occur with one call to xfs_log_write(). However, it is important
672  * to note that the transaction reservation code makes an assumption about the
673  * number of log headers a transaction requires that may be violated if you
674  * don't pass all the transaction vectors in one call....
675  */
676 int
677 xfs_log_write(
678 	struct xfs_mount	*mp,
679 	struct xfs_log_iovec	reg[],
680 	int			nentries,
681 	struct xlog_ticket	*tic,
682 	xfs_lsn_t		*start_lsn)
683 {
684 	struct log		*log = mp->m_log;
685 	int			error;
686 	struct xfs_log_vec	vec = {
687 		.lv_niovecs = nentries,
688 		.lv_iovecp = reg,
689 	};
690 
691 	if (XLOG_FORCED_SHUTDOWN(log))
692 		return XFS_ERROR(EIO);
693 
694 	error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
695 	if (error)
696 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
697 	return error;
698 }
699 
700 void
701 xfs_log_move_tail(xfs_mount_t	*mp,
702 		  xfs_lsn_t	tail_lsn)
703 {
704 	xlog_ticket_t	*tic;
705 	xlog_t		*log = mp->m_log;
706 	int		need_bytes, free_bytes, cycle, bytes;
707 
708 	if (XLOG_FORCED_SHUTDOWN(log))
709 		return;
710 
711 	if (tail_lsn == 0) {
712 		/* needed since sync_lsn is 64 bits */
713 		spin_lock(&log->l_icloglock);
714 		tail_lsn = log->l_last_sync_lsn;
715 		spin_unlock(&log->l_icloglock);
716 	}
717 
718 	spin_lock(&log->l_grant_lock);
719 
720 	/* Also an invalid lsn.  1 implies that we aren't passing in a valid
721 	 * tail_lsn.
722 	 */
723 	if (tail_lsn != 1) {
724 		log->l_tail_lsn = tail_lsn;
725 	}
726 
727 	if ((tic = log->l_write_headq)) {
728 #ifdef DEBUG
729 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
730 			panic("Recovery problem");
731 #endif
732 		cycle = log->l_grant_write_cycle;
733 		bytes = log->l_grant_write_bytes;
734 		free_bytes = xlog_space_left(log, cycle, bytes);
735 		do {
736 			ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
737 
738 			if (free_bytes < tic->t_unit_res && tail_lsn != 1)
739 				break;
740 			tail_lsn = 0;
741 			free_bytes -= tic->t_unit_res;
742 			sv_signal(&tic->t_wait);
743 			tic = tic->t_next;
744 		} while (tic != log->l_write_headq);
745 	}
746 	if ((tic = log->l_reserve_headq)) {
747 #ifdef DEBUG
748 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
749 			panic("Recovery problem");
750 #endif
751 		cycle = log->l_grant_reserve_cycle;
752 		bytes = log->l_grant_reserve_bytes;
753 		free_bytes = xlog_space_left(log, cycle, bytes);
754 		do {
755 			if (tic->t_flags & XLOG_TIC_PERM_RESERV)
756 				need_bytes = tic->t_unit_res*tic->t_cnt;
757 			else
758 				need_bytes = tic->t_unit_res;
759 			if (free_bytes < need_bytes && tail_lsn != 1)
760 				break;
761 			tail_lsn = 0;
762 			free_bytes -= need_bytes;
763 			sv_signal(&tic->t_wait);
764 			tic = tic->t_next;
765 		} while (tic != log->l_reserve_headq);
766 	}
767 	spin_unlock(&log->l_grant_lock);
768 }	/* xfs_log_move_tail */
769 
770 /*
771  * Determine if we have a transaction that has gone to disk
772  * that needs to be covered. To begin the transition to the idle state
773  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
774  * If we are then in a state where covering is needed, the caller is informed
775  * that dummy transactions are required to move the log into the idle state.
776  *
777  * Because this is called as part of the sync process, we should also indicate
778  * that dummy transactions should be issued in anything but the covered or
779  * idle states. This ensures that the log tail is accurately reflected in
780  * the log at the end of the sync, hence if a crash occurrs avoids replay
781  * of transactions where the metadata is already on disk.
782  */
783 int
784 xfs_log_need_covered(xfs_mount_t *mp)
785 {
786 	int		needed = 0;
787 	xlog_t		*log = mp->m_log;
788 
789 	if (!xfs_fs_writable(mp))
790 		return 0;
791 
792 	spin_lock(&log->l_icloglock);
793 	switch (log->l_covered_state) {
794 	case XLOG_STATE_COVER_DONE:
795 	case XLOG_STATE_COVER_DONE2:
796 	case XLOG_STATE_COVER_IDLE:
797 		break;
798 	case XLOG_STATE_COVER_NEED:
799 	case XLOG_STATE_COVER_NEED2:
800 		if (!xfs_trans_ail_tail(log->l_ailp) &&
801 		    xlog_iclogs_empty(log)) {
802 			if (log->l_covered_state == XLOG_STATE_COVER_NEED)
803 				log->l_covered_state = XLOG_STATE_COVER_DONE;
804 			else
805 				log->l_covered_state = XLOG_STATE_COVER_DONE2;
806 		}
807 		/* FALLTHRU */
808 	default:
809 		needed = 1;
810 		break;
811 	}
812 	spin_unlock(&log->l_icloglock);
813 	return needed;
814 }
815 
816 /******************************************************************************
817  *
818  *	local routines
819  *
820  ******************************************************************************
821  */
822 
823 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
824  * The log manager must keep track of the last LR which was committed
825  * to disk.  The lsn of this LR will become the new tail_lsn whenever
826  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
827  * the situation where stuff could be written into the log but nothing
828  * was ever in the AIL when asked.  Eventually, we panic since the
829  * tail hits the head.
830  *
831  * We may be holding the log iclog lock upon entering this routine.
832  */
833 xfs_lsn_t
834 xlog_assign_tail_lsn(xfs_mount_t *mp)
835 {
836 	xfs_lsn_t tail_lsn;
837 	xlog_t	  *log = mp->m_log;
838 
839 	tail_lsn = xfs_trans_ail_tail(mp->m_ail);
840 	spin_lock(&log->l_grant_lock);
841 	if (tail_lsn != 0) {
842 		log->l_tail_lsn = tail_lsn;
843 	} else {
844 		tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
845 	}
846 	spin_unlock(&log->l_grant_lock);
847 
848 	return tail_lsn;
849 }	/* xlog_assign_tail_lsn */
850 
851 
852 /*
853  * Return the space in the log between the tail and the head.  The head
854  * is passed in the cycle/bytes formal parms.  In the special case where
855  * the reserve head has wrapped passed the tail, this calculation is no
856  * longer valid.  In this case, just return 0 which means there is no space
857  * in the log.  This works for all places where this function is called
858  * with the reserve head.  Of course, if the write head were to ever
859  * wrap the tail, we should blow up.  Rather than catch this case here,
860  * we depend on other ASSERTions in other parts of the code.   XXXmiken
861  *
862  * This code also handles the case where the reservation head is behind
863  * the tail.  The details of this case are described below, but the end
864  * result is that we return the size of the log as the amount of space left.
865  */
866 STATIC int
867 xlog_space_left(xlog_t *log, int cycle, int bytes)
868 {
869 	int free_bytes;
870 	int tail_bytes;
871 	int tail_cycle;
872 
873 	tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
874 	tail_cycle = CYCLE_LSN(log->l_tail_lsn);
875 	if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
876 		free_bytes = log->l_logsize - (bytes - tail_bytes);
877 	} else if ((tail_cycle + 1) < cycle) {
878 		return 0;
879 	} else if (tail_cycle < cycle) {
880 		ASSERT(tail_cycle == (cycle - 1));
881 		free_bytes = tail_bytes - bytes;
882 	} else {
883 		/*
884 		 * The reservation head is behind the tail.
885 		 * In this case we just want to return the size of the
886 		 * log as the amount of space left.
887 		 */
888 		xfs_fs_cmn_err(CE_ALERT, log->l_mp,
889 			"xlog_space_left: head behind tail\n"
890 			"  tail_cycle = %d, tail_bytes = %d\n"
891 			"  GH   cycle = %d, GH   bytes = %d",
892 			tail_cycle, tail_bytes, cycle, bytes);
893 		ASSERT(0);
894 		free_bytes = log->l_logsize;
895 	}
896 	return free_bytes;
897 }	/* xlog_space_left */
898 
899 
900 /*
901  * Log function which is called when an io completes.
902  *
903  * The log manager needs its own routine, in order to control what
904  * happens with the buffer after the write completes.
905  */
906 void
907 xlog_iodone(xfs_buf_t *bp)
908 {
909 	xlog_in_core_t	*iclog;
910 	xlog_t		*l;
911 	int		aborted;
912 
913 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
914 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
915 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
916 	aborted = 0;
917 	l = iclog->ic_log;
918 
919 	/*
920 	 * If the _XFS_BARRIER_FAILED flag was set by a lower
921 	 * layer, it means the underlying device no longer supports
922 	 * barrier I/O. Warn loudly and turn off barriers.
923 	 */
924 	if (bp->b_flags & _XFS_BARRIER_FAILED) {
925 		bp->b_flags &= ~_XFS_BARRIER_FAILED;
926 		l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
927 		xfs_fs_cmn_err(CE_WARN, l->l_mp,
928 				"xlog_iodone: Barriers are no longer supported"
929 				" by device. Disabling barriers\n");
930 	}
931 
932 	/*
933 	 * Race to shutdown the filesystem if we see an error.
934 	 */
935 	if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
936 			XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
937 		xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
938 		XFS_BUF_STALE(bp);
939 		xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
940 		/*
941 		 * This flag will be propagated to the trans-committed
942 		 * callback routines to let them know that the log-commit
943 		 * didn't succeed.
944 		 */
945 		aborted = XFS_LI_ABORTED;
946 	} else if (iclog->ic_state & XLOG_STATE_IOERROR) {
947 		aborted = XFS_LI_ABORTED;
948 	}
949 
950 	/* log I/O is always issued ASYNC */
951 	ASSERT(XFS_BUF_ISASYNC(bp));
952 	xlog_state_done_syncing(iclog, aborted);
953 	/*
954 	 * do not reference the buffer (bp) here as we could race
955 	 * with it being freed after writing the unmount record to the
956 	 * log.
957 	 */
958 
959 }	/* xlog_iodone */
960 
961 /*
962  * Return size of each in-core log record buffer.
963  *
964  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
965  *
966  * If the filesystem blocksize is too large, we may need to choose a
967  * larger size since the directory code currently logs entire blocks.
968  */
969 
970 STATIC void
971 xlog_get_iclog_buffer_size(xfs_mount_t	*mp,
972 			   xlog_t	*log)
973 {
974 	int size;
975 	int xhdrs;
976 
977 	if (mp->m_logbufs <= 0)
978 		log->l_iclog_bufs = XLOG_MAX_ICLOGS;
979 	else
980 		log->l_iclog_bufs = mp->m_logbufs;
981 
982 	/*
983 	 * Buffer size passed in from mount system call.
984 	 */
985 	if (mp->m_logbsize > 0) {
986 		size = log->l_iclog_size = mp->m_logbsize;
987 		log->l_iclog_size_log = 0;
988 		while (size != 1) {
989 			log->l_iclog_size_log++;
990 			size >>= 1;
991 		}
992 
993 		if (xfs_sb_version_haslogv2(&mp->m_sb)) {
994 			/* # headers = size / 32k
995 			 * one header holds cycles from 32k of data
996 			 */
997 
998 			xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
999 			if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1000 				xhdrs++;
1001 			log->l_iclog_hsize = xhdrs << BBSHIFT;
1002 			log->l_iclog_heads = xhdrs;
1003 		} else {
1004 			ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1005 			log->l_iclog_hsize = BBSIZE;
1006 			log->l_iclog_heads = 1;
1007 		}
1008 		goto done;
1009 	}
1010 
1011 	/* All machines use 32kB buffers by default. */
1012 	log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1013 	log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1014 
1015 	/* the default log size is 16k or 32k which is one header sector */
1016 	log->l_iclog_hsize = BBSIZE;
1017 	log->l_iclog_heads = 1;
1018 
1019 done:
1020 	/* are we being asked to make the sizes selected above visible? */
1021 	if (mp->m_logbufs == 0)
1022 		mp->m_logbufs = log->l_iclog_bufs;
1023 	if (mp->m_logbsize == 0)
1024 		mp->m_logbsize = log->l_iclog_size;
1025 }	/* xlog_get_iclog_buffer_size */
1026 
1027 
1028 /*
1029  * This routine initializes some of the log structure for a given mount point.
1030  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1031  * some other stuff may be filled in too.
1032  */
1033 STATIC xlog_t *
1034 xlog_alloc_log(xfs_mount_t	*mp,
1035 	       xfs_buftarg_t	*log_target,
1036 	       xfs_daddr_t	blk_offset,
1037 	       int		num_bblks)
1038 {
1039 	xlog_t			*log;
1040 	xlog_rec_header_t	*head;
1041 	xlog_in_core_t		**iclogp;
1042 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1043 	xfs_buf_t		*bp;
1044 	int			i;
1045 	int			iclogsize;
1046 	int			error = ENOMEM;
1047 
1048 	log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1049 	if (!log) {
1050 		xlog_warn("XFS: Log allocation failed: No memory!");
1051 		goto out;
1052 	}
1053 
1054 	log->l_mp	   = mp;
1055 	log->l_targ	   = log_target;
1056 	log->l_logsize     = BBTOB(num_bblks);
1057 	log->l_logBBstart  = blk_offset;
1058 	log->l_logBBsize   = num_bblks;
1059 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1060 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1061 
1062 	log->l_prev_block  = -1;
1063 	log->l_tail_lsn	   = xlog_assign_lsn(1, 0);
1064 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1065 	log->l_last_sync_lsn = log->l_tail_lsn;
1066 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1067 	log->l_grant_reserve_cycle = 1;
1068 	log->l_grant_write_cycle = 1;
1069 
1070 	error = EFSCORRUPTED;
1071 	if (xfs_sb_version_hassector(&mp->m_sb)) {
1072 		log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1073 		if (log->l_sectbb_log < 0 ||
1074 		    log->l_sectbb_log > mp->m_sectbb_log) {
1075 			xlog_warn("XFS: Log sector size (0x%x) out of range.",
1076 						log->l_sectbb_log);
1077 			goto out_free_log;
1078 		}
1079 
1080 		/* for larger sector sizes, must have v2 or external log */
1081 		if (log->l_sectbb_log != 0 &&
1082 		    (log->l_logBBstart != 0 &&
1083 		     !xfs_sb_version_haslogv2(&mp->m_sb))) {
1084 			xlog_warn("XFS: log sector size (0x%x) invalid "
1085 				  "for configuration.", log->l_sectbb_log);
1086 			goto out_free_log;
1087 		}
1088 		if (mp->m_sb.sb_logsectlog < BBSHIFT) {
1089 			xlog_warn("XFS: Log sector log (0x%x) too small.",
1090 						mp->m_sb.sb_logsectlog);
1091 			goto out_free_log;
1092 		}
1093 	}
1094 	log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1095 
1096 	xlog_get_iclog_buffer_size(mp, log);
1097 
1098 	error = ENOMEM;
1099 	bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1100 	if (!bp)
1101 		goto out_free_log;
1102 	XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1103 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1104 	ASSERT(XFS_BUF_ISBUSY(bp));
1105 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1106 	log->l_xbuf = bp;
1107 
1108 	spin_lock_init(&log->l_icloglock);
1109 	spin_lock_init(&log->l_grant_lock);
1110 	sv_init(&log->l_flush_wait, 0, "flush_wait");
1111 
1112 	/* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1113 	ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1114 
1115 	iclogp = &log->l_iclog;
1116 	/*
1117 	 * The amount of memory to allocate for the iclog structure is
1118 	 * rather funky due to the way the structure is defined.  It is
1119 	 * done this way so that we can use different sizes for machines
1120 	 * with different amounts of memory.  See the definition of
1121 	 * xlog_in_core_t in xfs_log_priv.h for details.
1122 	 */
1123 	iclogsize = log->l_iclog_size;
1124 	ASSERT(log->l_iclog_size >= 4096);
1125 	for (i=0; i < log->l_iclog_bufs; i++) {
1126 		*iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1127 		if (!*iclogp)
1128 			goto out_free_iclog;
1129 
1130 		iclog = *iclogp;
1131 		iclog->ic_prev = prev_iclog;
1132 		prev_iclog = iclog;
1133 
1134 		bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1135 		if (!bp)
1136 			goto out_free_iclog;
1137 		if (!XFS_BUF_CPSEMA(bp))
1138 			ASSERT(0);
1139 		XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1140 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1141 		iclog->ic_bp = bp;
1142 		iclog->ic_data = bp->b_addr;
1143 #ifdef DEBUG
1144 		log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1145 #endif
1146 		head = &iclog->ic_header;
1147 		memset(head, 0, sizeof(xlog_rec_header_t));
1148 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1149 		head->h_version = cpu_to_be32(
1150 			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1151 		head->h_size = cpu_to_be32(log->l_iclog_size);
1152 		/* new fields */
1153 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1154 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1155 
1156 		iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1157 		iclog->ic_state = XLOG_STATE_ACTIVE;
1158 		iclog->ic_log = log;
1159 		atomic_set(&iclog->ic_refcnt, 0);
1160 		spin_lock_init(&iclog->ic_callback_lock);
1161 		iclog->ic_callback_tail = &(iclog->ic_callback);
1162 		iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1163 
1164 		ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1165 		ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1166 		sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1167 		sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1168 
1169 		iclogp = &iclog->ic_next;
1170 	}
1171 	*iclogp = log->l_iclog;			/* complete ring */
1172 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1173 
1174 	return log;
1175 
1176 out_free_iclog:
1177 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1178 		prev_iclog = iclog->ic_next;
1179 		if (iclog->ic_bp) {
1180 			sv_destroy(&iclog->ic_force_wait);
1181 			sv_destroy(&iclog->ic_write_wait);
1182 			xfs_buf_free(iclog->ic_bp);
1183 		}
1184 		kmem_free(iclog);
1185 	}
1186 	spinlock_destroy(&log->l_icloglock);
1187 	spinlock_destroy(&log->l_grant_lock);
1188 	xfs_buf_free(log->l_xbuf);
1189 out_free_log:
1190 	kmem_free(log);
1191 out:
1192 	return ERR_PTR(-error);
1193 }	/* xlog_alloc_log */
1194 
1195 
1196 /*
1197  * Write out the commit record of a transaction associated with the given
1198  * ticket.  Return the lsn of the commit record.
1199  */
1200 STATIC int
1201 xlog_commit_record(
1202 	struct log		*log,
1203 	struct xlog_ticket	*ticket,
1204 	struct xlog_in_core	**iclog,
1205 	xfs_lsn_t		*commitlsnp)
1206 {
1207 	struct xfs_mount *mp = log->l_mp;
1208 	int	error;
1209 	struct xfs_log_iovec reg = {
1210 		.i_addr = NULL,
1211 		.i_len = 0,
1212 		.i_type = XLOG_REG_TYPE_COMMIT,
1213 	};
1214 	struct xfs_log_vec vec = {
1215 		.lv_niovecs = 1,
1216 		.lv_iovecp = &reg,
1217 	};
1218 
1219 	ASSERT_ALWAYS(iclog);
1220 	error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1221 					XLOG_COMMIT_TRANS);
1222 	if (error)
1223 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1224 	return error;
1225 }
1226 
1227 /*
1228  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1229  * log space.  This code pushes on the lsn which would supposedly free up
1230  * the 25% which we want to leave free.  We may need to adopt a policy which
1231  * pushes on an lsn which is further along in the log once we reach the high
1232  * water mark.  In this manner, we would be creating a low water mark.
1233  */
1234 STATIC void
1235 xlog_grant_push_ail(xfs_mount_t	*mp,
1236 		    int		need_bytes)
1237 {
1238     xlog_t	*log = mp->m_log;	/* pointer to the log */
1239     xfs_lsn_t	tail_lsn;		/* lsn of the log tail */
1240     xfs_lsn_t	threshold_lsn = 0;	/* lsn we'd like to be at */
1241     int		free_blocks;		/* free blocks left to write to */
1242     int		free_bytes;		/* free bytes left to write to */
1243     int		threshold_block;	/* block in lsn we'd like to be at */
1244     int		threshold_cycle;	/* lsn cycle we'd like to be at */
1245     int		free_threshold;
1246 
1247     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1248 
1249     spin_lock(&log->l_grant_lock);
1250     free_bytes = xlog_space_left(log,
1251 				 log->l_grant_reserve_cycle,
1252 				 log->l_grant_reserve_bytes);
1253     tail_lsn = log->l_tail_lsn;
1254     free_blocks = BTOBBT(free_bytes);
1255 
1256     /*
1257      * Set the threshold for the minimum number of free blocks in the
1258      * log to the maximum of what the caller needs, one quarter of the
1259      * log, and 256 blocks.
1260      */
1261     free_threshold = BTOBB(need_bytes);
1262     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1263     free_threshold = MAX(free_threshold, 256);
1264     if (free_blocks < free_threshold) {
1265 	threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1266 	threshold_cycle = CYCLE_LSN(tail_lsn);
1267 	if (threshold_block >= log->l_logBBsize) {
1268 	    threshold_block -= log->l_logBBsize;
1269 	    threshold_cycle += 1;
1270 	}
1271 	threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1272 
1273 	/* Don't pass in an lsn greater than the lsn of the last
1274 	 * log record known to be on disk.
1275 	 */
1276 	if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1277 	    threshold_lsn = log->l_last_sync_lsn;
1278     }
1279     spin_unlock(&log->l_grant_lock);
1280 
1281     /*
1282      * Get the transaction layer to kick the dirty buffers out to
1283      * disk asynchronously. No point in trying to do this if
1284      * the filesystem is shutting down.
1285      */
1286     if (threshold_lsn &&
1287 	!XLOG_FORCED_SHUTDOWN(log))
1288 	    xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1289 }	/* xlog_grant_push_ail */
1290 
1291 /*
1292  * The bdstrat callback function for log bufs. This gives us a central
1293  * place to trap bufs in case we get hit by a log I/O error and need to
1294  * shutdown. Actually, in practice, even when we didn't get a log error,
1295  * we transition the iclogs to IOERROR state *after* flushing all existing
1296  * iclogs to disk. This is because we don't want anymore new transactions to be
1297  * started or completed afterwards.
1298  */
1299 STATIC int
1300 xlog_bdstrat(
1301 	struct xfs_buf		*bp)
1302 {
1303 	struct xlog_in_core	*iclog;
1304 
1305 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1306 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
1307 		XFS_BUF_ERROR(bp, EIO);
1308 		XFS_BUF_STALE(bp);
1309 		xfs_biodone(bp);
1310 		/*
1311 		 * It would seem logical to return EIO here, but we rely on
1312 		 * the log state machine to propagate I/O errors instead of
1313 		 * doing it here.
1314 		 */
1315 		return 0;
1316 	}
1317 
1318 	bp->b_flags |= _XBF_RUN_QUEUES;
1319 	xfs_buf_iorequest(bp);
1320 	return 0;
1321 }
1322 
1323 /*
1324  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1325  * fashion.  Previously, we should have moved the current iclog
1326  * ptr in the log to point to the next available iclog.  This allows further
1327  * write to continue while this code syncs out an iclog ready to go.
1328  * Before an in-core log can be written out, the data section must be scanned
1329  * to save away the 1st word of each BBSIZE block into the header.  We replace
1330  * it with the current cycle count.  Each BBSIZE block is tagged with the
1331  * cycle count because there in an implicit assumption that drives will
1332  * guarantee that entire 512 byte blocks get written at once.  In other words,
1333  * we can't have part of a 512 byte block written and part not written.  By
1334  * tagging each block, we will know which blocks are valid when recovering
1335  * after an unclean shutdown.
1336  *
1337  * This routine is single threaded on the iclog.  No other thread can be in
1338  * this routine with the same iclog.  Changing contents of iclog can there-
1339  * fore be done without grabbing the state machine lock.  Updating the global
1340  * log will require grabbing the lock though.
1341  *
1342  * The entire log manager uses a logical block numbering scheme.  Only
1343  * log_sync (and then only bwrite()) know about the fact that the log may
1344  * not start with block zero on a given device.  The log block start offset
1345  * is added immediately before calling bwrite().
1346  */
1347 
1348 STATIC int
1349 xlog_sync(xlog_t		*log,
1350 	  xlog_in_core_t	*iclog)
1351 {
1352 	xfs_caddr_t	dptr;		/* pointer to byte sized element */
1353 	xfs_buf_t	*bp;
1354 	int		i;
1355 	uint		count;		/* byte count of bwrite */
1356 	uint		count_init;	/* initial count before roundup */
1357 	int		roundoff;       /* roundoff to BB or stripe */
1358 	int		split = 0;	/* split write into two regions */
1359 	int		error;
1360 	int		v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1361 
1362 	XFS_STATS_INC(xs_log_writes);
1363 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1364 
1365 	/* Add for LR header */
1366 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1367 
1368 	/* Round out the log write size */
1369 	if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1370 		/* we have a v2 stripe unit to use */
1371 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1372 	} else {
1373 		count = BBTOB(BTOBB(count_init));
1374 	}
1375 	roundoff = count - count_init;
1376 	ASSERT(roundoff >= 0);
1377 	ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1378                 roundoff < log->l_mp->m_sb.sb_logsunit)
1379 		||
1380 		(log->l_mp->m_sb.sb_logsunit <= 1 &&
1381 		 roundoff < BBTOB(1)));
1382 
1383 	/* move grant heads by roundoff in sync */
1384 	spin_lock(&log->l_grant_lock);
1385 	xlog_grant_add_space(log, roundoff);
1386 	spin_unlock(&log->l_grant_lock);
1387 
1388 	/* put cycle number in every block */
1389 	xlog_pack_data(log, iclog, roundoff);
1390 
1391 	/* real byte length */
1392 	if (v2) {
1393 		iclog->ic_header.h_len =
1394 			cpu_to_be32(iclog->ic_offset + roundoff);
1395 	} else {
1396 		iclog->ic_header.h_len =
1397 			cpu_to_be32(iclog->ic_offset);
1398 	}
1399 
1400 	bp = iclog->ic_bp;
1401 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1402 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1403 	XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1404 
1405 	XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1406 
1407 	/* Do we need to split this write into 2 parts? */
1408 	if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1409 		split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1410 		count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1411 		iclog->ic_bwritecnt = 2;	/* split into 2 writes */
1412 	} else {
1413 		iclog->ic_bwritecnt = 1;
1414 	}
1415 	XFS_BUF_SET_COUNT(bp, count);
1416 	XFS_BUF_SET_FSPRIVATE(bp, iclog);	/* save for later */
1417 	XFS_BUF_ZEROFLAGS(bp);
1418 	XFS_BUF_BUSY(bp);
1419 	XFS_BUF_ASYNC(bp);
1420 	bp->b_flags |= XBF_LOG_BUFFER;
1421 	/*
1422 	 * Do an ordered write for the log block.
1423 	 * Its unnecessary to flush the first split block in the log wrap case.
1424 	 */
1425 	if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1426 		XFS_BUF_ORDERED(bp);
1427 
1428 	ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1429 	ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1430 
1431 	xlog_verify_iclog(log, iclog, count, B_TRUE);
1432 
1433 	/* account for log which doesn't start at block #0 */
1434 	XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1435 	/*
1436 	 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1437 	 * is shutting down.
1438 	 */
1439 	XFS_BUF_WRITE(bp);
1440 
1441 	if ((error = xlog_bdstrat(bp))) {
1442 		xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1443 				  XFS_BUF_ADDR(bp));
1444 		return error;
1445 	}
1446 	if (split) {
1447 		bp = iclog->ic_log->l_xbuf;
1448 		ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1449 							(unsigned long)1);
1450 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1451 		XFS_BUF_SET_ADDR(bp, 0);	     /* logical 0 */
1452 		XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1453 					    (__psint_t)count), split);
1454 		XFS_BUF_SET_FSPRIVATE(bp, iclog);
1455 		XFS_BUF_ZEROFLAGS(bp);
1456 		XFS_BUF_BUSY(bp);
1457 		XFS_BUF_ASYNC(bp);
1458 		bp->b_flags |= XBF_LOG_BUFFER;
1459 		if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1460 			XFS_BUF_ORDERED(bp);
1461 		dptr = XFS_BUF_PTR(bp);
1462 		/*
1463 		 * Bump the cycle numbers at the start of each block
1464 		 * since this part of the buffer is at the start of
1465 		 * a new cycle.  Watch out for the header magic number
1466 		 * case, though.
1467 		 */
1468 		for (i = 0; i < split; i += BBSIZE) {
1469 			be32_add_cpu((__be32 *)dptr, 1);
1470 			if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1471 				be32_add_cpu((__be32 *)dptr, 1);
1472 			dptr += BBSIZE;
1473 		}
1474 
1475 		ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1476 		ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1477 
1478 		/* account for internal log which doesn't start at block #0 */
1479 		XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1480 		XFS_BUF_WRITE(bp);
1481 		if ((error = xlog_bdstrat(bp))) {
1482 			xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1483 					  bp, XFS_BUF_ADDR(bp));
1484 			return error;
1485 		}
1486 	}
1487 	return 0;
1488 }	/* xlog_sync */
1489 
1490 
1491 /*
1492  * Deallocate a log structure
1493  */
1494 STATIC void
1495 xlog_dealloc_log(xlog_t *log)
1496 {
1497 	xlog_in_core_t	*iclog, *next_iclog;
1498 	int		i;
1499 
1500 	iclog = log->l_iclog;
1501 	for (i=0; i<log->l_iclog_bufs; i++) {
1502 		sv_destroy(&iclog->ic_force_wait);
1503 		sv_destroy(&iclog->ic_write_wait);
1504 		xfs_buf_free(iclog->ic_bp);
1505 		next_iclog = iclog->ic_next;
1506 		kmem_free(iclog);
1507 		iclog = next_iclog;
1508 	}
1509 	spinlock_destroy(&log->l_icloglock);
1510 	spinlock_destroy(&log->l_grant_lock);
1511 
1512 	xfs_buf_free(log->l_xbuf);
1513 	log->l_mp->m_log = NULL;
1514 	kmem_free(log);
1515 }	/* xlog_dealloc_log */
1516 
1517 /*
1518  * Update counters atomically now that memcpy is done.
1519  */
1520 /* ARGSUSED */
1521 static inline void
1522 xlog_state_finish_copy(xlog_t		*log,
1523 		       xlog_in_core_t	*iclog,
1524 		       int		record_cnt,
1525 		       int		copy_bytes)
1526 {
1527 	spin_lock(&log->l_icloglock);
1528 
1529 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1530 	iclog->ic_offset += copy_bytes;
1531 
1532 	spin_unlock(&log->l_icloglock);
1533 }	/* xlog_state_finish_copy */
1534 
1535 
1536 
1537 
1538 /*
1539  * print out info relating to regions written which consume
1540  * the reservation
1541  */
1542 STATIC void
1543 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1544 {
1545 	uint i;
1546 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1547 
1548 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
1549 	static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1550 	    "bformat",
1551 	    "bchunk",
1552 	    "efi_format",
1553 	    "efd_format",
1554 	    "iformat",
1555 	    "icore",
1556 	    "iext",
1557 	    "ibroot",
1558 	    "ilocal",
1559 	    "iattr_ext",
1560 	    "iattr_broot",
1561 	    "iattr_local",
1562 	    "qformat",
1563 	    "dquot",
1564 	    "quotaoff",
1565 	    "LR header",
1566 	    "unmount",
1567 	    "commit",
1568 	    "trans header"
1569 	};
1570 	static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1571 	    "SETATTR_NOT_SIZE",
1572 	    "SETATTR_SIZE",
1573 	    "INACTIVE",
1574 	    "CREATE",
1575 	    "CREATE_TRUNC",
1576 	    "TRUNCATE_FILE",
1577 	    "REMOVE",
1578 	    "LINK",
1579 	    "RENAME",
1580 	    "MKDIR",
1581 	    "RMDIR",
1582 	    "SYMLINK",
1583 	    "SET_DMATTRS",
1584 	    "GROWFS",
1585 	    "STRAT_WRITE",
1586 	    "DIOSTRAT",
1587 	    "WRITE_SYNC",
1588 	    "WRITEID",
1589 	    "ADDAFORK",
1590 	    "ATTRINVAL",
1591 	    "ATRUNCATE",
1592 	    "ATTR_SET",
1593 	    "ATTR_RM",
1594 	    "ATTR_FLAG",
1595 	    "CLEAR_AGI_BUCKET",
1596 	    "QM_SBCHANGE",
1597 	    "DUMMY1",
1598 	    "DUMMY2",
1599 	    "QM_QUOTAOFF",
1600 	    "QM_DQALLOC",
1601 	    "QM_SETQLIM",
1602 	    "QM_DQCLUSTER",
1603 	    "QM_QINOCREATE",
1604 	    "QM_QUOTAOFF_END",
1605 	    "SB_UNIT",
1606 	    "FSYNC_TS",
1607 	    "GROWFSRT_ALLOC",
1608 	    "GROWFSRT_ZERO",
1609 	    "GROWFSRT_FREE",
1610 	    "SWAPEXT"
1611 	};
1612 
1613 	xfs_fs_cmn_err(CE_WARN, mp,
1614 			"xfs_log_write: reservation summary:\n"
1615 			"  trans type  = %s (%u)\n"
1616 			"  unit res    = %d bytes\n"
1617 			"  current res = %d bytes\n"
1618 			"  total reg   = %u bytes (o/flow = %u bytes)\n"
1619 			"  ophdrs      = %u (ophdr space = %u bytes)\n"
1620 			"  ophdr + reg = %u bytes\n"
1621 			"  num regions = %u\n",
1622 			((ticket->t_trans_type <= 0 ||
1623 			  ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1624 			  "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1625 			ticket->t_trans_type,
1626 			ticket->t_unit_res,
1627 			ticket->t_curr_res,
1628 			ticket->t_res_arr_sum, ticket->t_res_o_flow,
1629 			ticket->t_res_num_ophdrs, ophdr_spc,
1630 			ticket->t_res_arr_sum +
1631 			ticket->t_res_o_flow + ophdr_spc,
1632 			ticket->t_res_num);
1633 
1634 	for (i = 0; i < ticket->t_res_num; i++) {
1635 		uint r_type = ticket->t_res_arr[i].r_type;
1636 		cmn_err(CE_WARN,
1637 			    "region[%u]: %s - %u bytes\n",
1638 			    i,
1639 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1640 			    "bad-rtype" : res_type_str[r_type-1]),
1641 			    ticket->t_res_arr[i].r_len);
1642 	}
1643 }
1644 
1645 /*
1646  * Calculate the potential space needed by the log vector.  Each region gets
1647  * its own xlog_op_header_t and may need to be double word aligned.
1648  */
1649 static int
1650 xlog_write_calc_vec_length(
1651 	struct xlog_ticket	*ticket,
1652 	struct xfs_log_vec	*log_vector)
1653 {
1654 	struct xfs_log_vec	*lv;
1655 	int			headers = 0;
1656 	int			len = 0;
1657 	int			i;
1658 
1659 	/* acct for start rec of xact */
1660 	if (ticket->t_flags & XLOG_TIC_INITED)
1661 		headers++;
1662 
1663 	for (lv = log_vector; lv; lv = lv->lv_next) {
1664 		headers += lv->lv_niovecs;
1665 
1666 		for (i = 0; i < lv->lv_niovecs; i++) {
1667 			struct xfs_log_iovec	*vecp = &lv->lv_iovecp[i];
1668 
1669 			len += vecp->i_len;
1670 			xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1671 		}
1672 	}
1673 
1674 	ticket->t_res_num_ophdrs += headers;
1675 	len += headers * sizeof(struct xlog_op_header);
1676 
1677 	return len;
1678 }
1679 
1680 /*
1681  * If first write for transaction, insert start record  We can't be trying to
1682  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1683  */
1684 static int
1685 xlog_write_start_rec(
1686 	__psint_t		ptr,
1687 	struct xlog_ticket	*ticket)
1688 {
1689 	struct xlog_op_header	*ophdr = (struct xlog_op_header *)ptr;
1690 
1691 	if (!(ticket->t_flags & XLOG_TIC_INITED))
1692 		return 0;
1693 
1694 	ophdr->oh_tid	= cpu_to_be32(ticket->t_tid);
1695 	ophdr->oh_clientid = ticket->t_clientid;
1696 	ophdr->oh_len = 0;
1697 	ophdr->oh_flags = XLOG_START_TRANS;
1698 	ophdr->oh_res2 = 0;
1699 
1700 	ticket->t_flags &= ~XLOG_TIC_INITED;
1701 
1702 	return sizeof(struct xlog_op_header);
1703 }
1704 
1705 static xlog_op_header_t *
1706 xlog_write_setup_ophdr(
1707 	struct log		*log,
1708 	__psint_t		ptr,
1709 	struct xlog_ticket	*ticket,
1710 	uint			flags)
1711 {
1712 	struct xlog_op_header	*ophdr = (struct xlog_op_header *)ptr;
1713 
1714 	ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1715 	ophdr->oh_clientid = ticket->t_clientid;
1716 	ophdr->oh_res2 = 0;
1717 
1718 	/* are we copying a commit or unmount record? */
1719 	ophdr->oh_flags = flags;
1720 
1721 	/*
1722 	 * We've seen logs corrupted with bad transaction client ids.  This
1723 	 * makes sure that XFS doesn't generate them on.  Turn this into an EIO
1724 	 * and shut down the filesystem.
1725 	 */
1726 	switch (ophdr->oh_clientid)  {
1727 	case XFS_TRANSACTION:
1728 	case XFS_VOLUME:
1729 	case XFS_LOG:
1730 		break;
1731 	default:
1732 		xfs_fs_cmn_err(CE_WARN, log->l_mp,
1733 			"Bad XFS transaction clientid 0x%x in ticket 0x%p",
1734 			ophdr->oh_clientid, ticket);
1735 		return NULL;
1736 	}
1737 
1738 	return ophdr;
1739 }
1740 
1741 /*
1742  * Set up the parameters of the region copy into the log. This has
1743  * to handle region write split across multiple log buffers - this
1744  * state is kept external to this function so that this code can
1745  * can be written in an obvious, self documenting manner.
1746  */
1747 static int
1748 xlog_write_setup_copy(
1749 	struct xlog_ticket	*ticket,
1750 	struct xlog_op_header	*ophdr,
1751 	int			space_available,
1752 	int			space_required,
1753 	int			*copy_off,
1754 	int			*copy_len,
1755 	int			*last_was_partial_copy,
1756 	int			*bytes_consumed)
1757 {
1758 	int			still_to_copy;
1759 
1760 	still_to_copy = space_required - *bytes_consumed;
1761 	*copy_off = *bytes_consumed;
1762 
1763 	if (still_to_copy <= space_available) {
1764 		/* write of region completes here */
1765 		*copy_len = still_to_copy;
1766 		ophdr->oh_len = cpu_to_be32(*copy_len);
1767 		if (*last_was_partial_copy)
1768 			ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1769 		*last_was_partial_copy = 0;
1770 		*bytes_consumed = 0;
1771 		return 0;
1772 	}
1773 
1774 	/* partial write of region, needs extra log op header reservation */
1775 	*copy_len = space_available;
1776 	ophdr->oh_len = cpu_to_be32(*copy_len);
1777 	ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1778 	if (*last_was_partial_copy)
1779 		ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1780 	*bytes_consumed += *copy_len;
1781 	(*last_was_partial_copy)++;
1782 
1783 	/* account for new log op header */
1784 	ticket->t_curr_res -= sizeof(struct xlog_op_header);
1785 	ticket->t_res_num_ophdrs++;
1786 
1787 	return sizeof(struct xlog_op_header);
1788 }
1789 
1790 static int
1791 xlog_write_copy_finish(
1792 	struct log		*log,
1793 	struct xlog_in_core	*iclog,
1794 	uint			flags,
1795 	int			*record_cnt,
1796 	int			*data_cnt,
1797 	int			*partial_copy,
1798 	int			*partial_copy_len,
1799 	int			log_offset,
1800 	struct xlog_in_core	**commit_iclog)
1801 {
1802 	if (*partial_copy) {
1803 		/*
1804 		 * This iclog has already been marked WANT_SYNC by
1805 		 * xlog_state_get_iclog_space.
1806 		 */
1807 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1808 		*record_cnt = 0;
1809 		*data_cnt = 0;
1810 		return xlog_state_release_iclog(log, iclog);
1811 	}
1812 
1813 	*partial_copy = 0;
1814 	*partial_copy_len = 0;
1815 
1816 	if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1817 		/* no more space in this iclog - push it. */
1818 		xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1819 		*record_cnt = 0;
1820 		*data_cnt = 0;
1821 
1822 		spin_lock(&log->l_icloglock);
1823 		xlog_state_want_sync(log, iclog);
1824 		spin_unlock(&log->l_icloglock);
1825 
1826 		if (!commit_iclog)
1827 			return xlog_state_release_iclog(log, iclog);
1828 		ASSERT(flags & XLOG_COMMIT_TRANS);
1829 		*commit_iclog = iclog;
1830 	}
1831 
1832 	return 0;
1833 }
1834 
1835 /*
1836  * Write some region out to in-core log
1837  *
1838  * This will be called when writing externally provided regions or when
1839  * writing out a commit record for a given transaction.
1840  *
1841  * General algorithm:
1842  *	1. Find total length of this write.  This may include adding to the
1843  *		lengths passed in.
1844  *	2. Check whether we violate the tickets reservation.
1845  *	3. While writing to this iclog
1846  *	    A. Reserve as much space in this iclog as can get
1847  *	    B. If this is first write, save away start lsn
1848  *	    C. While writing this region:
1849  *		1. If first write of transaction, write start record
1850  *		2. Write log operation header (header per region)
1851  *		3. Find out if we can fit entire region into this iclog
1852  *		4. Potentially, verify destination memcpy ptr
1853  *		5. Memcpy (partial) region
1854  *		6. If partial copy, release iclog; otherwise, continue
1855  *			copying more regions into current iclog
1856  *	4. Mark want sync bit (in simulation mode)
1857  *	5. Release iclog for potential flush to on-disk log.
1858  *
1859  * ERRORS:
1860  * 1.	Panic if reservation is overrun.  This should never happen since
1861  *	reservation amounts are generated internal to the filesystem.
1862  * NOTES:
1863  * 1. Tickets are single threaded data structures.
1864  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1865  *	syncing routine.  When a single log_write region needs to span
1866  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1867  *	on all log operation writes which don't contain the end of the
1868  *	region.  The XLOG_END_TRANS bit is used for the in-core log
1869  *	operation which contains the end of the continued log_write region.
1870  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1871  *	we don't really know exactly how much space will be used.  As a result,
1872  *	we don't update ic_offset until the end when we know exactly how many
1873  *	bytes have been written out.
1874  */
1875 STATIC int
1876 xlog_write(
1877 	struct log		*log,
1878 	struct xfs_log_vec	*log_vector,
1879 	struct xlog_ticket	*ticket,
1880 	xfs_lsn_t		*start_lsn,
1881 	struct xlog_in_core	**commit_iclog,
1882 	uint			flags)
1883 {
1884 	struct xlog_in_core	*iclog = NULL;
1885 	struct xfs_log_iovec	*vecp;
1886 	struct xfs_log_vec	*lv;
1887 	int			len;
1888 	int			index;
1889 	int			partial_copy = 0;
1890 	int			partial_copy_len = 0;
1891 	int			contwr = 0;
1892 	int			record_cnt = 0;
1893 	int			data_cnt = 0;
1894 	int			error;
1895 
1896 	*start_lsn = 0;
1897 
1898 	len = xlog_write_calc_vec_length(ticket, log_vector);
1899 	if (ticket->t_curr_res < len) {
1900 		xlog_print_tic_res(log->l_mp, ticket);
1901 #ifdef DEBUG
1902 		xlog_panic(
1903 	"xfs_log_write: reservation ran out. Need to up reservation");
1904 #else
1905 		/* Customer configurable panic */
1906 		xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, log->l_mp,
1907 	"xfs_log_write: reservation ran out. Need to up reservation");
1908 
1909 		/* If we did not panic, shutdown the filesystem */
1910 		xfs_force_shutdown(log->l_mp, SHUTDOWN_CORRUPT_INCORE);
1911 #endif
1912 	}
1913 
1914 	ticket->t_curr_res -= len;
1915 
1916 	index = 0;
1917 	lv = log_vector;
1918 	vecp = lv->lv_iovecp;
1919 	while (lv && index < lv->lv_niovecs) {
1920 		__psint_t	ptr;
1921 		int		log_offset;
1922 
1923 		error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1924 						   &contwr, &log_offset);
1925 		if (error)
1926 			return error;
1927 
1928 		ASSERT(log_offset <= iclog->ic_size - 1);
1929 		ptr = (__psint_t)((char *)iclog->ic_datap + log_offset);
1930 
1931 		/* start_lsn is the first lsn written to. That's all we need. */
1932 		if (!*start_lsn)
1933 			*start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1934 
1935 		/*
1936 		 * This loop writes out as many regions as can fit in the amount
1937 		 * of space which was allocated by xlog_state_get_iclog_space().
1938 		 */
1939 		while (lv && index < lv->lv_niovecs) {
1940 			struct xfs_log_iovec	*reg = &vecp[index];
1941 			struct xlog_op_header	*ophdr;
1942 			int			start_rec_copy;
1943 			int			copy_len;
1944 			int			copy_off;
1945 
1946 			ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1947 			ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1948 
1949 			start_rec_copy = xlog_write_start_rec(ptr, ticket);
1950 			if (start_rec_copy) {
1951 				record_cnt++;
1952 				xlog_write_adv_cnt(ptr, len, log_offset,
1953 						   start_rec_copy);
1954 			}
1955 
1956 			ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1957 			if (!ophdr)
1958 				return XFS_ERROR(EIO);
1959 
1960 			xlog_write_adv_cnt(ptr, len, log_offset,
1961 					   sizeof(struct xlog_op_header));
1962 
1963 			len += xlog_write_setup_copy(ticket, ophdr,
1964 						     iclog->ic_size-log_offset,
1965 						     reg->i_len,
1966 						     &copy_off, &copy_len,
1967 						     &partial_copy,
1968 						     &partial_copy_len);
1969 			xlog_verify_dest_ptr(log, ptr);
1970 
1971 			/* copy region */
1972 			ASSERT(copy_len >= 0);
1973 			memcpy((xfs_caddr_t)ptr, reg->i_addr + copy_off,
1974 			       copy_len);
1975 			xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1976 
1977 			copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1978 			record_cnt++;
1979 			data_cnt += contwr ? copy_len : 0;
1980 
1981 			error = xlog_write_copy_finish(log, iclog, flags,
1982 						       &record_cnt, &data_cnt,
1983 						       &partial_copy,
1984 						       &partial_copy_len,
1985 						       log_offset,
1986 						       commit_iclog);
1987 			if (error)
1988 				return error;
1989 
1990 			/*
1991 			 * if we had a partial copy, we need to get more iclog
1992 			 * space but we don't want to increment the region
1993 			 * index because there is still more is this region to
1994 			 * write.
1995 			 *
1996 			 * If we completed writing this region, and we flushed
1997 			 * the iclog (indicated by resetting of the record
1998 			 * count), then we also need to get more log space. If
1999 			 * this was the last record, though, we are done and
2000 			 * can just return.
2001 			 */
2002 			if (partial_copy)
2003 				break;
2004 
2005 			if (++index == lv->lv_niovecs) {
2006 				lv = lv->lv_next;
2007 				index = 0;
2008 				if (lv)
2009 					vecp = lv->lv_iovecp;
2010 			}
2011 			if (record_cnt == 0) {
2012 				if (!lv)
2013 					return 0;
2014 				break;
2015 			}
2016 		}
2017 	}
2018 
2019 	ASSERT(len == 0);
2020 
2021 	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2022 	if (!commit_iclog)
2023 		return xlog_state_release_iclog(log, iclog);
2024 
2025 	ASSERT(flags & XLOG_COMMIT_TRANS);
2026 	*commit_iclog = iclog;
2027 	return 0;
2028 }
2029 
2030 
2031 /*****************************************************************************
2032  *
2033  *		State Machine functions
2034  *
2035  *****************************************************************************
2036  */
2037 
2038 /* Clean iclogs starting from the head.  This ordering must be
2039  * maintained, so an iclog doesn't become ACTIVE beyond one that
2040  * is SYNCING.  This is also required to maintain the notion that we use
2041  * a ordered wait queue to hold off would be writers to the log when every
2042  * iclog is trying to sync to disk.
2043  *
2044  * State Change: DIRTY -> ACTIVE
2045  */
2046 STATIC void
2047 xlog_state_clean_log(xlog_t *log)
2048 {
2049 	xlog_in_core_t	*iclog;
2050 	int changed = 0;
2051 
2052 	iclog = log->l_iclog;
2053 	do {
2054 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
2055 			iclog->ic_state	= XLOG_STATE_ACTIVE;
2056 			iclog->ic_offset       = 0;
2057 			ASSERT(iclog->ic_callback == NULL);
2058 			/*
2059 			 * If the number of ops in this iclog indicate it just
2060 			 * contains the dummy transaction, we can
2061 			 * change state into IDLE (the second time around).
2062 			 * Otherwise we should change the state into
2063 			 * NEED a dummy.
2064 			 * We don't need to cover the dummy.
2065 			 */
2066 			if (!changed &&
2067 			   (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2068 			   		XLOG_COVER_OPS)) {
2069 				changed = 1;
2070 			} else {
2071 				/*
2072 				 * We have two dirty iclogs so start over
2073 				 * This could also be num of ops indicates
2074 				 * this is not the dummy going out.
2075 				 */
2076 				changed = 2;
2077 			}
2078 			iclog->ic_header.h_num_logops = 0;
2079 			memset(iclog->ic_header.h_cycle_data, 0,
2080 			      sizeof(iclog->ic_header.h_cycle_data));
2081 			iclog->ic_header.h_lsn = 0;
2082 		} else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2083 			/* do nothing */;
2084 		else
2085 			break;	/* stop cleaning */
2086 		iclog = iclog->ic_next;
2087 	} while (iclog != log->l_iclog);
2088 
2089 	/* log is locked when we are called */
2090 	/*
2091 	 * Change state for the dummy log recording.
2092 	 * We usually go to NEED. But we go to NEED2 if the changed indicates
2093 	 * we are done writing the dummy record.
2094 	 * If we are done with the second dummy recored (DONE2), then
2095 	 * we go to IDLE.
2096 	 */
2097 	if (changed) {
2098 		switch (log->l_covered_state) {
2099 		case XLOG_STATE_COVER_IDLE:
2100 		case XLOG_STATE_COVER_NEED:
2101 		case XLOG_STATE_COVER_NEED2:
2102 			log->l_covered_state = XLOG_STATE_COVER_NEED;
2103 			break;
2104 
2105 		case XLOG_STATE_COVER_DONE:
2106 			if (changed == 1)
2107 				log->l_covered_state = XLOG_STATE_COVER_NEED2;
2108 			else
2109 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2110 			break;
2111 
2112 		case XLOG_STATE_COVER_DONE2:
2113 			if (changed == 1)
2114 				log->l_covered_state = XLOG_STATE_COVER_IDLE;
2115 			else
2116 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2117 			break;
2118 
2119 		default:
2120 			ASSERT(0);
2121 		}
2122 	}
2123 }	/* xlog_state_clean_log */
2124 
2125 STATIC xfs_lsn_t
2126 xlog_get_lowest_lsn(
2127 	xlog_t		*log)
2128 {
2129 	xlog_in_core_t  *lsn_log;
2130 	xfs_lsn_t	lowest_lsn, lsn;
2131 
2132 	lsn_log = log->l_iclog;
2133 	lowest_lsn = 0;
2134 	do {
2135 	    if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2136 		lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2137 		if ((lsn && !lowest_lsn) ||
2138 		    (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2139 			lowest_lsn = lsn;
2140 		}
2141 	    }
2142 	    lsn_log = lsn_log->ic_next;
2143 	} while (lsn_log != log->l_iclog);
2144 	return lowest_lsn;
2145 }
2146 
2147 
2148 STATIC void
2149 xlog_state_do_callback(
2150 	xlog_t		*log,
2151 	int		aborted,
2152 	xlog_in_core_t	*ciclog)
2153 {
2154 	xlog_in_core_t	   *iclog;
2155 	xlog_in_core_t	   *first_iclog;	/* used to know when we've
2156 						 * processed all iclogs once */
2157 	xfs_log_callback_t *cb, *cb_next;
2158 	int		   flushcnt = 0;
2159 	xfs_lsn_t	   lowest_lsn;
2160 	int		   ioerrors;	/* counter: iclogs with errors */
2161 	int		   loopdidcallbacks; /* flag: inner loop did callbacks*/
2162 	int		   funcdidcallbacks; /* flag: function did callbacks */
2163 	int		   repeats;	/* for issuing console warnings if
2164 					 * looping too many times */
2165 	int		   wake = 0;
2166 
2167 	spin_lock(&log->l_icloglock);
2168 	first_iclog = iclog = log->l_iclog;
2169 	ioerrors = 0;
2170 	funcdidcallbacks = 0;
2171 	repeats = 0;
2172 
2173 	do {
2174 		/*
2175 		 * Scan all iclogs starting with the one pointed to by the
2176 		 * log.  Reset this starting point each time the log is
2177 		 * unlocked (during callbacks).
2178 		 *
2179 		 * Keep looping through iclogs until one full pass is made
2180 		 * without running any callbacks.
2181 		 */
2182 		first_iclog = log->l_iclog;
2183 		iclog = log->l_iclog;
2184 		loopdidcallbacks = 0;
2185 		repeats++;
2186 
2187 		do {
2188 
2189 			/* skip all iclogs in the ACTIVE & DIRTY states */
2190 			if (iclog->ic_state &
2191 			    (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2192 				iclog = iclog->ic_next;
2193 				continue;
2194 			}
2195 
2196 			/*
2197 			 * Between marking a filesystem SHUTDOWN and stopping
2198 			 * the log, we do flush all iclogs to disk (if there
2199 			 * wasn't a log I/O error). So, we do want things to
2200 			 * go smoothly in case of just a SHUTDOWN  w/o a
2201 			 * LOG_IO_ERROR.
2202 			 */
2203 			if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2204 				/*
2205 				 * Can only perform callbacks in order.  Since
2206 				 * this iclog is not in the DONE_SYNC/
2207 				 * DO_CALLBACK state, we skip the rest and
2208 				 * just try to clean up.  If we set our iclog
2209 				 * to DO_CALLBACK, we will not process it when
2210 				 * we retry since a previous iclog is in the
2211 				 * CALLBACK and the state cannot change since
2212 				 * we are holding the l_icloglock.
2213 				 */
2214 				if (!(iclog->ic_state &
2215 					(XLOG_STATE_DONE_SYNC |
2216 						 XLOG_STATE_DO_CALLBACK))) {
2217 					if (ciclog && (ciclog->ic_state ==
2218 							XLOG_STATE_DONE_SYNC)) {
2219 						ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2220 					}
2221 					break;
2222 				}
2223 				/*
2224 				 * We now have an iclog that is in either the
2225 				 * DO_CALLBACK or DONE_SYNC states. The other
2226 				 * states (WANT_SYNC, SYNCING, or CALLBACK were
2227 				 * caught by the above if and are going to
2228 				 * clean (i.e. we aren't doing their callbacks)
2229 				 * see the above if.
2230 				 */
2231 
2232 				/*
2233 				 * We will do one more check here to see if we
2234 				 * have chased our tail around.
2235 				 */
2236 
2237 				lowest_lsn = xlog_get_lowest_lsn(log);
2238 				if (lowest_lsn &&
2239 				    XFS_LSN_CMP(lowest_lsn,
2240 				    		be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2241 					iclog = iclog->ic_next;
2242 					continue; /* Leave this iclog for
2243 						   * another thread */
2244 				}
2245 
2246 				iclog->ic_state = XLOG_STATE_CALLBACK;
2247 
2248 				spin_unlock(&log->l_icloglock);
2249 
2250 				/* l_last_sync_lsn field protected by
2251 				 * l_grant_lock. Don't worry about iclog's lsn.
2252 				 * No one else can be here except us.
2253 				 */
2254 				spin_lock(&log->l_grant_lock);
2255 				ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2256 				       be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2257 				log->l_last_sync_lsn =
2258 					be64_to_cpu(iclog->ic_header.h_lsn);
2259 				spin_unlock(&log->l_grant_lock);
2260 
2261 			} else {
2262 				spin_unlock(&log->l_icloglock);
2263 				ioerrors++;
2264 			}
2265 
2266 			/*
2267 			 * Keep processing entries in the callback list until
2268 			 * we come around and it is empty.  We need to
2269 			 * atomically see that the list is empty and change the
2270 			 * state to DIRTY so that we don't miss any more
2271 			 * callbacks being added.
2272 			 */
2273 			spin_lock(&iclog->ic_callback_lock);
2274 			cb = iclog->ic_callback;
2275 			while (cb) {
2276 				iclog->ic_callback_tail = &(iclog->ic_callback);
2277 				iclog->ic_callback = NULL;
2278 				spin_unlock(&iclog->ic_callback_lock);
2279 
2280 				/* perform callbacks in the order given */
2281 				for (; cb; cb = cb_next) {
2282 					cb_next = cb->cb_next;
2283 					cb->cb_func(cb->cb_arg, aborted);
2284 				}
2285 				spin_lock(&iclog->ic_callback_lock);
2286 				cb = iclog->ic_callback;
2287 			}
2288 
2289 			loopdidcallbacks++;
2290 			funcdidcallbacks++;
2291 
2292 			spin_lock(&log->l_icloglock);
2293 			ASSERT(iclog->ic_callback == NULL);
2294 			spin_unlock(&iclog->ic_callback_lock);
2295 			if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2296 				iclog->ic_state = XLOG_STATE_DIRTY;
2297 
2298 			/*
2299 			 * Transition from DIRTY to ACTIVE if applicable.
2300 			 * NOP if STATE_IOERROR.
2301 			 */
2302 			xlog_state_clean_log(log);
2303 
2304 			/* wake up threads waiting in xfs_log_force() */
2305 			sv_broadcast(&iclog->ic_force_wait);
2306 
2307 			iclog = iclog->ic_next;
2308 		} while (first_iclog != iclog);
2309 
2310 		if (repeats > 5000) {
2311 			flushcnt += repeats;
2312 			repeats = 0;
2313 			xfs_fs_cmn_err(CE_WARN, log->l_mp,
2314 				"%s: possible infinite loop (%d iterations)",
2315 				__func__, flushcnt);
2316 		}
2317 	} while (!ioerrors && loopdidcallbacks);
2318 
2319 	/*
2320 	 * make one last gasp attempt to see if iclogs are being left in
2321 	 * limbo..
2322 	 */
2323 #ifdef DEBUG
2324 	if (funcdidcallbacks) {
2325 		first_iclog = iclog = log->l_iclog;
2326 		do {
2327 			ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2328 			/*
2329 			 * Terminate the loop if iclogs are found in states
2330 			 * which will cause other threads to clean up iclogs.
2331 			 *
2332 			 * SYNCING - i/o completion will go through logs
2333 			 * DONE_SYNC - interrupt thread should be waiting for
2334 			 *              l_icloglock
2335 			 * IOERROR - give up hope all ye who enter here
2336 			 */
2337 			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2338 			    iclog->ic_state == XLOG_STATE_SYNCING ||
2339 			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2340 			    iclog->ic_state == XLOG_STATE_IOERROR )
2341 				break;
2342 			iclog = iclog->ic_next;
2343 		} while (first_iclog != iclog);
2344 	}
2345 #endif
2346 
2347 	if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2348 		wake = 1;
2349 	spin_unlock(&log->l_icloglock);
2350 
2351 	if (wake)
2352 		sv_broadcast(&log->l_flush_wait);
2353 }
2354 
2355 
2356 /*
2357  * Finish transitioning this iclog to the dirty state.
2358  *
2359  * Make sure that we completely execute this routine only when this is
2360  * the last call to the iclog.  There is a good chance that iclog flushes,
2361  * when we reach the end of the physical log, get turned into 2 separate
2362  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2363  * routine.  By using the reference count bwritecnt, we guarantee that only
2364  * the second completion goes through.
2365  *
2366  * Callbacks could take time, so they are done outside the scope of the
2367  * global state machine log lock.
2368  */
2369 STATIC void
2370 xlog_state_done_syncing(
2371 	xlog_in_core_t	*iclog,
2372 	int		aborted)
2373 {
2374 	xlog_t		   *log = iclog->ic_log;
2375 
2376 	spin_lock(&log->l_icloglock);
2377 
2378 	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2379 	       iclog->ic_state == XLOG_STATE_IOERROR);
2380 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2381 	ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2382 
2383 
2384 	/*
2385 	 * If we got an error, either on the first buffer, or in the case of
2386 	 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2387 	 * and none should ever be attempted to be written to disk
2388 	 * again.
2389 	 */
2390 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
2391 		if (--iclog->ic_bwritecnt == 1) {
2392 			spin_unlock(&log->l_icloglock);
2393 			return;
2394 		}
2395 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2396 	}
2397 
2398 	/*
2399 	 * Someone could be sleeping prior to writing out the next
2400 	 * iclog buffer, we wake them all, one will get to do the
2401 	 * I/O, the others get to wait for the result.
2402 	 */
2403 	sv_broadcast(&iclog->ic_write_wait);
2404 	spin_unlock(&log->l_icloglock);
2405 	xlog_state_do_callback(log, aborted, iclog);	/* also cleans log */
2406 }	/* xlog_state_done_syncing */
2407 
2408 
2409 /*
2410  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2411  * sleep.  We wait on the flush queue on the head iclog as that should be
2412  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2413  * we will wait here and all new writes will sleep until a sync completes.
2414  *
2415  * The in-core logs are used in a circular fashion. They are not used
2416  * out-of-order even when an iclog past the head is free.
2417  *
2418  * return:
2419  *	* log_offset where xlog_write() can start writing into the in-core
2420  *		log's data space.
2421  *	* in-core log pointer to which xlog_write() should write.
2422  *	* boolean indicating this is a continued write to an in-core log.
2423  *		If this is the last write, then the in-core log's offset field
2424  *		needs to be incremented, depending on the amount of data which
2425  *		is copied.
2426  */
2427 STATIC int
2428 xlog_state_get_iclog_space(xlog_t	  *log,
2429 			   int		  len,
2430 			   xlog_in_core_t **iclogp,
2431 			   xlog_ticket_t  *ticket,
2432 			   int		  *continued_write,
2433 			   int		  *logoffsetp)
2434 {
2435 	int		  log_offset;
2436 	xlog_rec_header_t *head;
2437 	xlog_in_core_t	  *iclog;
2438 	int		  error;
2439 
2440 restart:
2441 	spin_lock(&log->l_icloglock);
2442 	if (XLOG_FORCED_SHUTDOWN(log)) {
2443 		spin_unlock(&log->l_icloglock);
2444 		return XFS_ERROR(EIO);
2445 	}
2446 
2447 	iclog = log->l_iclog;
2448 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2449 		XFS_STATS_INC(xs_log_noiclogs);
2450 
2451 		/* Wait for log writes to have flushed */
2452 		sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2453 		goto restart;
2454 	}
2455 
2456 	head = &iclog->ic_header;
2457 
2458 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2459 	log_offset = iclog->ic_offset;
2460 
2461 	/* On the 1st write to an iclog, figure out lsn.  This works
2462 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2463 	 * committing to.  If the offset is set, that's how many blocks
2464 	 * must be written.
2465 	 */
2466 	if (log_offset == 0) {
2467 		ticket->t_curr_res -= log->l_iclog_hsize;
2468 		xlog_tic_add_region(ticket,
2469 				    log->l_iclog_hsize,
2470 				    XLOG_REG_TYPE_LRHEADER);
2471 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2472 		head->h_lsn = cpu_to_be64(
2473 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2474 		ASSERT(log->l_curr_block >= 0);
2475 	}
2476 
2477 	/* If there is enough room to write everything, then do it.  Otherwise,
2478 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2479 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2480 	 * until you know exactly how many bytes get copied.  Therefore, wait
2481 	 * until later to update ic_offset.
2482 	 *
2483 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2484 	 * can fit into remaining data section.
2485 	 */
2486 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2487 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2488 
2489 		/*
2490 		 * If I'm the only one writing to this iclog, sync it to disk.
2491 		 * We need to do an atomic compare and decrement here to avoid
2492 		 * racing with concurrent atomic_dec_and_lock() calls in
2493 		 * xlog_state_release_iclog() when there is more than one
2494 		 * reference to the iclog.
2495 		 */
2496 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2497 			/* we are the only one */
2498 			spin_unlock(&log->l_icloglock);
2499 			error = xlog_state_release_iclog(log, iclog);
2500 			if (error)
2501 				return error;
2502 		} else {
2503 			spin_unlock(&log->l_icloglock);
2504 		}
2505 		goto restart;
2506 	}
2507 
2508 	/* Do we have enough room to write the full amount in the remainder
2509 	 * of this iclog?  Or must we continue a write on the next iclog and
2510 	 * mark this iclog as completely taken?  In the case where we switch
2511 	 * iclogs (to mark it taken), this particular iclog will release/sync
2512 	 * to disk in xlog_write().
2513 	 */
2514 	if (len <= iclog->ic_size - iclog->ic_offset) {
2515 		*continued_write = 0;
2516 		iclog->ic_offset += len;
2517 	} else {
2518 		*continued_write = 1;
2519 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2520 	}
2521 	*iclogp = iclog;
2522 
2523 	ASSERT(iclog->ic_offset <= iclog->ic_size);
2524 	spin_unlock(&log->l_icloglock);
2525 
2526 	*logoffsetp = log_offset;
2527 	return 0;
2528 }	/* xlog_state_get_iclog_space */
2529 
2530 /*
2531  * Atomically get the log space required for a log ticket.
2532  *
2533  * Once a ticket gets put onto the reserveq, it will only return after
2534  * the needed reservation is satisfied.
2535  */
2536 STATIC int
2537 xlog_grant_log_space(xlog_t	   *log,
2538 		     xlog_ticket_t *tic)
2539 {
2540 	int		 free_bytes;
2541 	int		 need_bytes;
2542 #ifdef DEBUG
2543 	xfs_lsn_t	 tail_lsn;
2544 #endif
2545 
2546 
2547 #ifdef DEBUG
2548 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2549 		panic("grant Recovery problem");
2550 #endif
2551 
2552 	/* Is there space or do we need to sleep? */
2553 	spin_lock(&log->l_grant_lock);
2554 
2555 	trace_xfs_log_grant_enter(log, tic);
2556 
2557 	/* something is already sleeping; insert new transaction at end */
2558 	if (log->l_reserve_headq) {
2559 		xlog_ins_ticketq(&log->l_reserve_headq, tic);
2560 
2561 		trace_xfs_log_grant_sleep1(log, tic);
2562 
2563 		/*
2564 		 * Gotta check this before going to sleep, while we're
2565 		 * holding the grant lock.
2566 		 */
2567 		if (XLOG_FORCED_SHUTDOWN(log))
2568 			goto error_return;
2569 
2570 		XFS_STATS_INC(xs_sleep_logspace);
2571 		sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2572 		/*
2573 		 * If we got an error, and the filesystem is shutting down,
2574 		 * we'll catch it down below. So just continue...
2575 		 */
2576 		trace_xfs_log_grant_wake1(log, tic);
2577 		spin_lock(&log->l_grant_lock);
2578 	}
2579 	if (tic->t_flags & XFS_LOG_PERM_RESERV)
2580 		need_bytes = tic->t_unit_res*tic->t_ocnt;
2581 	else
2582 		need_bytes = tic->t_unit_res;
2583 
2584 redo:
2585 	if (XLOG_FORCED_SHUTDOWN(log))
2586 		goto error_return;
2587 
2588 	free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2589 				     log->l_grant_reserve_bytes);
2590 	if (free_bytes < need_bytes) {
2591 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2592 			xlog_ins_ticketq(&log->l_reserve_headq, tic);
2593 
2594 		trace_xfs_log_grant_sleep2(log, tic);
2595 
2596 		spin_unlock(&log->l_grant_lock);
2597 		xlog_grant_push_ail(log->l_mp, need_bytes);
2598 		spin_lock(&log->l_grant_lock);
2599 
2600 		XFS_STATS_INC(xs_sleep_logspace);
2601 		sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2602 
2603 		spin_lock(&log->l_grant_lock);
2604 		if (XLOG_FORCED_SHUTDOWN(log))
2605 			goto error_return;
2606 
2607 		trace_xfs_log_grant_wake2(log, tic);
2608 
2609 		goto redo;
2610 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2611 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2612 
2613 	/* we've got enough space */
2614 	xlog_grant_add_space(log, need_bytes);
2615 #ifdef DEBUG
2616 	tail_lsn = log->l_tail_lsn;
2617 	/*
2618 	 * Check to make sure the grant write head didn't just over lap the
2619 	 * tail.  If the cycles are the same, we can't be overlapping.
2620 	 * Otherwise, make sure that the cycles differ by exactly one and
2621 	 * check the byte count.
2622 	 */
2623 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2624 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2625 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2626 	}
2627 #endif
2628 	trace_xfs_log_grant_exit(log, tic);
2629 	xlog_verify_grant_head(log, 1);
2630 	spin_unlock(&log->l_grant_lock);
2631 	return 0;
2632 
2633  error_return:
2634 	if (tic->t_flags & XLOG_TIC_IN_Q)
2635 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2636 
2637 	trace_xfs_log_grant_error(log, tic);
2638 
2639 	/*
2640 	 * If we are failing, make sure the ticket doesn't have any
2641 	 * current reservations. We don't want to add this back when
2642 	 * the ticket/transaction gets cancelled.
2643 	 */
2644 	tic->t_curr_res = 0;
2645 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2646 	spin_unlock(&log->l_grant_lock);
2647 	return XFS_ERROR(EIO);
2648 }	/* xlog_grant_log_space */
2649 
2650 
2651 /*
2652  * Replenish the byte reservation required by moving the grant write head.
2653  *
2654  *
2655  */
2656 STATIC int
2657 xlog_regrant_write_log_space(xlog_t	   *log,
2658 			     xlog_ticket_t *tic)
2659 {
2660 	int		free_bytes, need_bytes;
2661 	xlog_ticket_t	*ntic;
2662 #ifdef DEBUG
2663 	xfs_lsn_t	tail_lsn;
2664 #endif
2665 
2666 	tic->t_curr_res = tic->t_unit_res;
2667 	xlog_tic_reset_res(tic);
2668 
2669 	if (tic->t_cnt > 0)
2670 		return 0;
2671 
2672 #ifdef DEBUG
2673 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2674 		panic("regrant Recovery problem");
2675 #endif
2676 
2677 	spin_lock(&log->l_grant_lock);
2678 
2679 	trace_xfs_log_regrant_write_enter(log, tic);
2680 
2681 	if (XLOG_FORCED_SHUTDOWN(log))
2682 		goto error_return;
2683 
2684 	/* If there are other waiters on the queue then give them a
2685 	 * chance at logspace before us. Wake up the first waiters,
2686 	 * if we do not wake up all the waiters then go to sleep waiting
2687 	 * for more free space, otherwise try to get some space for
2688 	 * this transaction.
2689 	 */
2690 	need_bytes = tic->t_unit_res;
2691 	if ((ntic = log->l_write_headq)) {
2692 		free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2693 					     log->l_grant_write_bytes);
2694 		do {
2695 			ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2696 
2697 			if (free_bytes < ntic->t_unit_res)
2698 				break;
2699 			free_bytes -= ntic->t_unit_res;
2700 			sv_signal(&ntic->t_wait);
2701 			ntic = ntic->t_next;
2702 		} while (ntic != log->l_write_headq);
2703 
2704 		if (ntic != log->l_write_headq) {
2705 			if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2706 				xlog_ins_ticketq(&log->l_write_headq, tic);
2707 
2708 			trace_xfs_log_regrant_write_sleep1(log, tic);
2709 
2710 			spin_unlock(&log->l_grant_lock);
2711 			xlog_grant_push_ail(log->l_mp, need_bytes);
2712 			spin_lock(&log->l_grant_lock);
2713 
2714 			XFS_STATS_INC(xs_sleep_logspace);
2715 			sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2716 				&log->l_grant_lock, s);
2717 
2718 			/* If we're shutting down, this tic is already
2719 			 * off the queue */
2720 			spin_lock(&log->l_grant_lock);
2721 			if (XLOG_FORCED_SHUTDOWN(log))
2722 				goto error_return;
2723 
2724 			trace_xfs_log_regrant_write_wake1(log, tic);
2725 		}
2726 	}
2727 
2728 redo:
2729 	if (XLOG_FORCED_SHUTDOWN(log))
2730 		goto error_return;
2731 
2732 	free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2733 				     log->l_grant_write_bytes);
2734 	if (free_bytes < need_bytes) {
2735 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2736 			xlog_ins_ticketq(&log->l_write_headq, tic);
2737 		spin_unlock(&log->l_grant_lock);
2738 		xlog_grant_push_ail(log->l_mp, need_bytes);
2739 		spin_lock(&log->l_grant_lock);
2740 
2741 		XFS_STATS_INC(xs_sleep_logspace);
2742 		trace_xfs_log_regrant_write_sleep2(log, tic);
2743 
2744 		sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2745 
2746 		/* If we're shutting down, this tic is already off the queue */
2747 		spin_lock(&log->l_grant_lock);
2748 		if (XLOG_FORCED_SHUTDOWN(log))
2749 			goto error_return;
2750 
2751 		trace_xfs_log_regrant_write_wake2(log, tic);
2752 		goto redo;
2753 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2754 		xlog_del_ticketq(&log->l_write_headq, tic);
2755 
2756 	/* we've got enough space */
2757 	xlog_grant_add_space_write(log, need_bytes);
2758 #ifdef DEBUG
2759 	tail_lsn = log->l_tail_lsn;
2760 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2761 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2762 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2763 	}
2764 #endif
2765 
2766 	trace_xfs_log_regrant_write_exit(log, tic);
2767 
2768 	xlog_verify_grant_head(log, 1);
2769 	spin_unlock(&log->l_grant_lock);
2770 	return 0;
2771 
2772 
2773  error_return:
2774 	if (tic->t_flags & XLOG_TIC_IN_Q)
2775 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2776 
2777 	trace_xfs_log_regrant_write_error(log, tic);
2778 
2779 	/*
2780 	 * If we are failing, make sure the ticket doesn't have any
2781 	 * current reservations. We don't want to add this back when
2782 	 * the ticket/transaction gets cancelled.
2783 	 */
2784 	tic->t_curr_res = 0;
2785 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2786 	spin_unlock(&log->l_grant_lock);
2787 	return XFS_ERROR(EIO);
2788 }	/* xlog_regrant_write_log_space */
2789 
2790 
2791 /* The first cnt-1 times through here we don't need to
2792  * move the grant write head because the permanent
2793  * reservation has reserved cnt times the unit amount.
2794  * Release part of current permanent unit reservation and
2795  * reset current reservation to be one units worth.  Also
2796  * move grant reservation head forward.
2797  */
2798 STATIC void
2799 xlog_regrant_reserve_log_space(xlog_t	     *log,
2800 			       xlog_ticket_t *ticket)
2801 {
2802 	trace_xfs_log_regrant_reserve_enter(log, ticket);
2803 
2804 	if (ticket->t_cnt > 0)
2805 		ticket->t_cnt--;
2806 
2807 	spin_lock(&log->l_grant_lock);
2808 	xlog_grant_sub_space(log, ticket->t_curr_res);
2809 	ticket->t_curr_res = ticket->t_unit_res;
2810 	xlog_tic_reset_res(ticket);
2811 
2812 	trace_xfs_log_regrant_reserve_sub(log, ticket);
2813 
2814 	xlog_verify_grant_head(log, 1);
2815 
2816 	/* just return if we still have some of the pre-reserved space */
2817 	if (ticket->t_cnt > 0) {
2818 		spin_unlock(&log->l_grant_lock);
2819 		return;
2820 	}
2821 
2822 	xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2823 
2824 	trace_xfs_log_regrant_reserve_exit(log, ticket);
2825 
2826 	xlog_verify_grant_head(log, 0);
2827 	spin_unlock(&log->l_grant_lock);
2828 	ticket->t_curr_res = ticket->t_unit_res;
2829 	xlog_tic_reset_res(ticket);
2830 }	/* xlog_regrant_reserve_log_space */
2831 
2832 
2833 /*
2834  * Give back the space left from a reservation.
2835  *
2836  * All the information we need to make a correct determination of space left
2837  * is present.  For non-permanent reservations, things are quite easy.  The
2838  * count should have been decremented to zero.  We only need to deal with the
2839  * space remaining in the current reservation part of the ticket.  If the
2840  * ticket contains a permanent reservation, there may be left over space which
2841  * needs to be released.  A count of N means that N-1 refills of the current
2842  * reservation can be done before we need to ask for more space.  The first
2843  * one goes to fill up the first current reservation.  Once we run out of
2844  * space, the count will stay at zero and the only space remaining will be
2845  * in the current reservation field.
2846  */
2847 STATIC void
2848 xlog_ungrant_log_space(xlog_t	     *log,
2849 		       xlog_ticket_t *ticket)
2850 {
2851 	if (ticket->t_cnt > 0)
2852 		ticket->t_cnt--;
2853 
2854 	spin_lock(&log->l_grant_lock);
2855 	trace_xfs_log_ungrant_enter(log, ticket);
2856 
2857 	xlog_grant_sub_space(log, ticket->t_curr_res);
2858 
2859 	trace_xfs_log_ungrant_sub(log, ticket);
2860 
2861 	/* If this is a permanent reservation ticket, we may be able to free
2862 	 * up more space based on the remaining count.
2863 	 */
2864 	if (ticket->t_cnt > 0) {
2865 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2866 		xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2867 	}
2868 
2869 	trace_xfs_log_ungrant_exit(log, ticket);
2870 
2871 	xlog_verify_grant_head(log, 1);
2872 	spin_unlock(&log->l_grant_lock);
2873 	xfs_log_move_tail(log->l_mp, 1);
2874 }	/* xlog_ungrant_log_space */
2875 
2876 
2877 /*
2878  * Flush iclog to disk if this is the last reference to the given iclog and
2879  * the WANT_SYNC bit is set.
2880  *
2881  * When this function is entered, the iclog is not necessarily in the
2882  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2883  *
2884  *
2885  */
2886 STATIC int
2887 xlog_state_release_iclog(
2888 	xlog_t		*log,
2889 	xlog_in_core_t	*iclog)
2890 {
2891 	int		sync = 0;	/* do we sync? */
2892 
2893 	if (iclog->ic_state & XLOG_STATE_IOERROR)
2894 		return XFS_ERROR(EIO);
2895 
2896 	ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2897 	if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2898 		return 0;
2899 
2900 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2901 		spin_unlock(&log->l_icloglock);
2902 		return XFS_ERROR(EIO);
2903 	}
2904 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2905 	       iclog->ic_state == XLOG_STATE_WANT_SYNC);
2906 
2907 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2908 		/* update tail before writing to iclog */
2909 		xlog_assign_tail_lsn(log->l_mp);
2910 		sync++;
2911 		iclog->ic_state = XLOG_STATE_SYNCING;
2912 		iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2913 		xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2914 		/* cycle incremented when incrementing curr_block */
2915 	}
2916 	spin_unlock(&log->l_icloglock);
2917 
2918 	/*
2919 	 * We let the log lock go, so it's possible that we hit a log I/O
2920 	 * error or some other SHUTDOWN condition that marks the iclog
2921 	 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2922 	 * this iclog has consistent data, so we ignore IOERROR
2923 	 * flags after this point.
2924 	 */
2925 	if (sync)
2926 		return xlog_sync(log, iclog);
2927 	return 0;
2928 }	/* xlog_state_release_iclog */
2929 
2930 
2931 /*
2932  * This routine will mark the current iclog in the ring as WANT_SYNC
2933  * and move the current iclog pointer to the next iclog in the ring.
2934  * When this routine is called from xlog_state_get_iclog_space(), the
2935  * exact size of the iclog has not yet been determined.  All we know is
2936  * that every data block.  We have run out of space in this log record.
2937  */
2938 STATIC void
2939 xlog_state_switch_iclogs(xlog_t		*log,
2940 			 xlog_in_core_t *iclog,
2941 			 int		eventual_size)
2942 {
2943 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2944 	if (!eventual_size)
2945 		eventual_size = iclog->ic_offset;
2946 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2947 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2948 	log->l_prev_block = log->l_curr_block;
2949 	log->l_prev_cycle = log->l_curr_cycle;
2950 
2951 	/* roll log?: ic_offset changed later */
2952 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2953 
2954 	/* Round up to next log-sunit */
2955 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2956 	    log->l_mp->m_sb.sb_logsunit > 1) {
2957 		__uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2958 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2959 	}
2960 
2961 	if (log->l_curr_block >= log->l_logBBsize) {
2962 		log->l_curr_cycle++;
2963 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2964 			log->l_curr_cycle++;
2965 		log->l_curr_block -= log->l_logBBsize;
2966 		ASSERT(log->l_curr_block >= 0);
2967 	}
2968 	ASSERT(iclog == log->l_iclog);
2969 	log->l_iclog = iclog->ic_next;
2970 }	/* xlog_state_switch_iclogs */
2971 
2972 /*
2973  * Write out all data in the in-core log as of this exact moment in time.
2974  *
2975  * Data may be written to the in-core log during this call.  However,
2976  * we don't guarantee this data will be written out.  A change from past
2977  * implementation means this routine will *not* write out zero length LRs.
2978  *
2979  * Basically, we try and perform an intelligent scan of the in-core logs.
2980  * If we determine there is no flushable data, we just return.  There is no
2981  * flushable data if:
2982  *
2983  *	1. the current iclog is active and has no data; the previous iclog
2984  *		is in the active or dirty state.
2985  *	2. the current iclog is drity, and the previous iclog is in the
2986  *		active or dirty state.
2987  *
2988  * We may sleep if:
2989  *
2990  *	1. the current iclog is not in the active nor dirty state.
2991  *	2. the current iclog dirty, and the previous iclog is not in the
2992  *		active nor dirty state.
2993  *	3. the current iclog is active, and there is another thread writing
2994  *		to this particular iclog.
2995  *	4. a) the current iclog is active and has no other writers
2996  *	   b) when we return from flushing out this iclog, it is still
2997  *		not in the active nor dirty state.
2998  */
2999 int
3000 _xfs_log_force(
3001 	struct xfs_mount	*mp,
3002 	uint			flags,
3003 	int			*log_flushed)
3004 {
3005 	struct log		*log = mp->m_log;
3006 	struct xlog_in_core	*iclog;
3007 	xfs_lsn_t		lsn;
3008 
3009 	XFS_STATS_INC(xs_log_force);
3010 
3011 	spin_lock(&log->l_icloglock);
3012 
3013 	iclog = log->l_iclog;
3014 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
3015 		spin_unlock(&log->l_icloglock);
3016 		return XFS_ERROR(EIO);
3017 	}
3018 
3019 	/* If the head iclog is not active nor dirty, we just attach
3020 	 * ourselves to the head and go to sleep.
3021 	 */
3022 	if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3023 	    iclog->ic_state == XLOG_STATE_DIRTY) {
3024 		/*
3025 		 * If the head is dirty or (active and empty), then
3026 		 * we need to look at the previous iclog.  If the previous
3027 		 * iclog is active or dirty we are done.  There is nothing
3028 		 * to sync out.  Otherwise, we attach ourselves to the
3029 		 * previous iclog and go to sleep.
3030 		 */
3031 		if (iclog->ic_state == XLOG_STATE_DIRTY ||
3032 		    (atomic_read(&iclog->ic_refcnt) == 0
3033 		     && iclog->ic_offset == 0)) {
3034 			iclog = iclog->ic_prev;
3035 			if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3036 			    iclog->ic_state == XLOG_STATE_DIRTY)
3037 				goto no_sleep;
3038 			else
3039 				goto maybe_sleep;
3040 		} else {
3041 			if (atomic_read(&iclog->ic_refcnt) == 0) {
3042 				/* We are the only one with access to this
3043 				 * iclog.  Flush it out now.  There should
3044 				 * be a roundoff of zero to show that someone
3045 				 * has already taken care of the roundoff from
3046 				 * the previous sync.
3047 				 */
3048 				atomic_inc(&iclog->ic_refcnt);
3049 				lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3050 				xlog_state_switch_iclogs(log, iclog, 0);
3051 				spin_unlock(&log->l_icloglock);
3052 
3053 				if (xlog_state_release_iclog(log, iclog))
3054 					return XFS_ERROR(EIO);
3055 
3056 				if (log_flushed)
3057 					*log_flushed = 1;
3058 				spin_lock(&log->l_icloglock);
3059 				if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3060 				    iclog->ic_state != XLOG_STATE_DIRTY)
3061 					goto maybe_sleep;
3062 				else
3063 					goto no_sleep;
3064 			} else {
3065 				/* Someone else is writing to this iclog.
3066 				 * Use its call to flush out the data.  However,
3067 				 * the other thread may not force out this LR,
3068 				 * so we mark it WANT_SYNC.
3069 				 */
3070 				xlog_state_switch_iclogs(log, iclog, 0);
3071 				goto maybe_sleep;
3072 			}
3073 		}
3074 	}
3075 
3076 	/* By the time we come around again, the iclog could've been filled
3077 	 * which would give it another lsn.  If we have a new lsn, just
3078 	 * return because the relevant data has been flushed.
3079 	 */
3080 maybe_sleep:
3081 	if (flags & XFS_LOG_SYNC) {
3082 		/*
3083 		 * We must check if we're shutting down here, before
3084 		 * we wait, while we're holding the l_icloglock.
3085 		 * Then we check again after waking up, in case our
3086 		 * sleep was disturbed by a bad news.
3087 		 */
3088 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
3089 			spin_unlock(&log->l_icloglock);
3090 			return XFS_ERROR(EIO);
3091 		}
3092 		XFS_STATS_INC(xs_log_force_sleep);
3093 		sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
3094 		/*
3095 		 * No need to grab the log lock here since we're
3096 		 * only deciding whether or not to return EIO
3097 		 * and the memory read should be atomic.
3098 		 */
3099 		if (iclog->ic_state & XLOG_STATE_IOERROR)
3100 			return XFS_ERROR(EIO);
3101 		if (log_flushed)
3102 			*log_flushed = 1;
3103 	} else {
3104 
3105 no_sleep:
3106 		spin_unlock(&log->l_icloglock);
3107 	}
3108 	return 0;
3109 }
3110 
3111 /*
3112  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3113  * about errors or whether the log was flushed or not. This is the normal
3114  * interface to use when trying to unpin items or move the log forward.
3115  */
3116 void
3117 xfs_log_force(
3118 	xfs_mount_t	*mp,
3119 	uint		flags)
3120 {
3121 	int	error;
3122 
3123 	error = _xfs_log_force(mp, flags, NULL);
3124 	if (error) {
3125 		xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3126 			"error %d returned.", error);
3127 	}
3128 }
3129 
3130 /*
3131  * Force the in-core log to disk for a specific LSN.
3132  *
3133  * Find in-core log with lsn.
3134  *	If it is in the DIRTY state, just return.
3135  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3136  *		state and go to sleep or return.
3137  *	If it is in any other state, go to sleep or return.
3138  *
3139  * Synchronous forces are implemented with a signal variable. All callers
3140  * to force a given lsn to disk will wait on a the sv attached to the
3141  * specific in-core log.  When given in-core log finally completes its
3142  * write to disk, that thread will wake up all threads waiting on the
3143  * sv.
3144  */
3145 int
3146 _xfs_log_force_lsn(
3147 	struct xfs_mount	*mp,
3148 	xfs_lsn_t		lsn,
3149 	uint			flags,
3150 	int			*log_flushed)
3151 {
3152 	struct log		*log = mp->m_log;
3153 	struct xlog_in_core	*iclog;
3154 	int			already_slept = 0;
3155 
3156 	ASSERT(lsn != 0);
3157 
3158 	XFS_STATS_INC(xs_log_force);
3159 
3160 try_again:
3161 	spin_lock(&log->l_icloglock);
3162 	iclog = log->l_iclog;
3163 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
3164 		spin_unlock(&log->l_icloglock);
3165 		return XFS_ERROR(EIO);
3166 	}
3167 
3168 	do {
3169 		if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3170 			iclog = iclog->ic_next;
3171 			continue;
3172 		}
3173 
3174 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
3175 			spin_unlock(&log->l_icloglock);
3176 			return 0;
3177 		}
3178 
3179 		if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3180 			/*
3181 			 * We sleep here if we haven't already slept (e.g.
3182 			 * this is the first time we've looked at the correct
3183 			 * iclog buf) and the buffer before us is going to
3184 			 * be sync'ed. The reason for this is that if we
3185 			 * are doing sync transactions here, by waiting for
3186 			 * the previous I/O to complete, we can allow a few
3187 			 * more transactions into this iclog before we close
3188 			 * it down.
3189 			 *
3190 			 * Otherwise, we mark the buffer WANT_SYNC, and bump
3191 			 * up the refcnt so we can release the log (which
3192 			 * drops the ref count).  The state switch keeps new
3193 			 * transaction commits from using this buffer.  When
3194 			 * the current commits finish writing into the buffer,
3195 			 * the refcount will drop to zero and the buffer will
3196 			 * go out then.
3197 			 */
3198 			if (!already_slept &&
3199 			    (iclog->ic_prev->ic_state &
3200 			     (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3201 				ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3202 
3203 				XFS_STATS_INC(xs_log_force_sleep);
3204 
3205 				sv_wait(&iclog->ic_prev->ic_write_wait,
3206 					PSWP, &log->l_icloglock, s);
3207 				if (log_flushed)
3208 					*log_flushed = 1;
3209 				already_slept = 1;
3210 				goto try_again;
3211 			}
3212 			atomic_inc(&iclog->ic_refcnt);
3213 			xlog_state_switch_iclogs(log, iclog, 0);
3214 			spin_unlock(&log->l_icloglock);
3215 			if (xlog_state_release_iclog(log, iclog))
3216 				return XFS_ERROR(EIO);
3217 			if (log_flushed)
3218 				*log_flushed = 1;
3219 			spin_lock(&log->l_icloglock);
3220 		}
3221 
3222 		if ((flags & XFS_LOG_SYNC) && /* sleep */
3223 		    !(iclog->ic_state &
3224 		      (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3225 			/*
3226 			 * Don't wait on completion if we know that we've
3227 			 * gotten a log write error.
3228 			 */
3229 			if (iclog->ic_state & XLOG_STATE_IOERROR) {
3230 				spin_unlock(&log->l_icloglock);
3231 				return XFS_ERROR(EIO);
3232 			}
3233 			XFS_STATS_INC(xs_log_force_sleep);
3234 			sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3235 			/*
3236 			 * No need to grab the log lock here since we're
3237 			 * only deciding whether or not to return EIO
3238 			 * and the memory read should be atomic.
3239 			 */
3240 			if (iclog->ic_state & XLOG_STATE_IOERROR)
3241 				return XFS_ERROR(EIO);
3242 
3243 			if (log_flushed)
3244 				*log_flushed = 1;
3245 		} else {		/* just return */
3246 			spin_unlock(&log->l_icloglock);
3247 		}
3248 
3249 		return 0;
3250 	} while (iclog != log->l_iclog);
3251 
3252 	spin_unlock(&log->l_icloglock);
3253 	return 0;
3254 }
3255 
3256 /*
3257  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3258  * about errors or whether the log was flushed or not. This is the normal
3259  * interface to use when trying to unpin items or move the log forward.
3260  */
3261 void
3262 xfs_log_force_lsn(
3263 	xfs_mount_t	*mp,
3264 	xfs_lsn_t	lsn,
3265 	uint		flags)
3266 {
3267 	int	error;
3268 
3269 	error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3270 	if (error) {
3271 		xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
3272 			"error %d returned.", error);
3273 	}
3274 }
3275 
3276 /*
3277  * Called when we want to mark the current iclog as being ready to sync to
3278  * disk.
3279  */
3280 STATIC void
3281 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3282 {
3283 	assert_spin_locked(&log->l_icloglock);
3284 
3285 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3286 		xlog_state_switch_iclogs(log, iclog, 0);
3287 	} else {
3288 		ASSERT(iclog->ic_state &
3289 			(XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3290 	}
3291 }
3292 
3293 
3294 /*****************************************************************************
3295  *
3296  *		TICKET functions
3297  *
3298  *****************************************************************************
3299  */
3300 
3301 /*
3302  * Free a used ticket when its refcount falls to zero.
3303  */
3304 void
3305 xfs_log_ticket_put(
3306 	xlog_ticket_t	*ticket)
3307 {
3308 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3309 	if (atomic_dec_and_test(&ticket->t_ref)) {
3310 		sv_destroy(&ticket->t_wait);
3311 		kmem_zone_free(xfs_log_ticket_zone, ticket);
3312 	}
3313 }
3314 
3315 xlog_ticket_t *
3316 xfs_log_ticket_get(
3317 	xlog_ticket_t	*ticket)
3318 {
3319 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3320 	atomic_inc(&ticket->t_ref);
3321 	return ticket;
3322 }
3323 
3324 /*
3325  * Allocate and initialise a new log ticket.
3326  */
3327 STATIC xlog_ticket_t *
3328 xlog_ticket_alloc(
3329 	struct log	*log,
3330 	int		unit_bytes,
3331 	int		cnt,
3332 	char		client,
3333 	uint		xflags)
3334 {
3335 	struct xlog_ticket *tic;
3336 	uint		num_headers;
3337 	int		iclog_space;
3338 
3339 	tic = kmem_zone_zalloc(xfs_log_ticket_zone, KM_SLEEP|KM_MAYFAIL);
3340 	if (!tic)
3341 		return NULL;
3342 
3343 	/*
3344 	 * Permanent reservations have up to 'cnt'-1 active log operations
3345 	 * in the log.  A unit in this case is the amount of space for one
3346 	 * of these log operations.  Normal reservations have a cnt of 1
3347 	 * and their unit amount is the total amount of space required.
3348 	 *
3349 	 * The following lines of code account for non-transaction data
3350 	 * which occupy space in the on-disk log.
3351 	 *
3352 	 * Normal form of a transaction is:
3353 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3354 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3355 	 *
3356 	 * We need to account for all the leadup data and trailer data
3357 	 * around the transaction data.
3358 	 * And then we need to account for the worst case in terms of using
3359 	 * more space.
3360 	 * The worst case will happen if:
3361 	 * - the placement of the transaction happens to be such that the
3362 	 *   roundoff is at its maximum
3363 	 * - the transaction data is synced before the commit record is synced
3364 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3365 	 *   Therefore the commit record is in its own Log Record.
3366 	 *   This can happen as the commit record is called with its
3367 	 *   own region to xlog_write().
3368 	 *   This then means that in the worst case, roundoff can happen for
3369 	 *   the commit-rec as well.
3370 	 *   The commit-rec is smaller than padding in this scenario and so it is
3371 	 *   not added separately.
3372 	 */
3373 
3374 	/* for trans header */
3375 	unit_bytes += sizeof(xlog_op_header_t);
3376 	unit_bytes += sizeof(xfs_trans_header_t);
3377 
3378 	/* for start-rec */
3379 	unit_bytes += sizeof(xlog_op_header_t);
3380 
3381 	/*
3382 	 * for LR headers - the space for data in an iclog is the size minus
3383 	 * the space used for the headers. If we use the iclog size, then we
3384 	 * undercalculate the number of headers required.
3385 	 *
3386 	 * Furthermore - the addition of op headers for split-recs might
3387 	 * increase the space required enough to require more log and op
3388 	 * headers, so take that into account too.
3389 	 *
3390 	 * IMPORTANT: This reservation makes the assumption that if this
3391 	 * transaction is the first in an iclog and hence has the LR headers
3392 	 * accounted to it, then the remaining space in the iclog is
3393 	 * exclusively for this transaction.  i.e. if the transaction is larger
3394 	 * than the iclog, it will be the only thing in that iclog.
3395 	 * Fundamentally, this means we must pass the entire log vector to
3396 	 * xlog_write to guarantee this.
3397 	 */
3398 	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3399 	num_headers = howmany(unit_bytes, iclog_space);
3400 
3401 	/* for split-recs - ophdrs added when data split over LRs */
3402 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3403 
3404 	/* add extra header reservations if we overrun */
3405 	while (!num_headers ||
3406 	       howmany(unit_bytes, iclog_space) > num_headers) {
3407 		unit_bytes += sizeof(xlog_op_header_t);
3408 		num_headers++;
3409 	}
3410 	unit_bytes += log->l_iclog_hsize * num_headers;
3411 
3412 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3413 	unit_bytes += log->l_iclog_hsize;
3414 
3415 	/* for roundoff padding for transaction data and one for commit record */
3416 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3417 	    log->l_mp->m_sb.sb_logsunit > 1) {
3418 		/* log su roundoff */
3419 		unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3420 	} else {
3421 		/* BB roundoff */
3422 		unit_bytes += 2*BBSIZE;
3423         }
3424 
3425 	atomic_set(&tic->t_ref, 1);
3426 	tic->t_unit_res		= unit_bytes;
3427 	tic->t_curr_res		= unit_bytes;
3428 	tic->t_cnt		= cnt;
3429 	tic->t_ocnt		= cnt;
3430 	tic->t_tid		= (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3431 	tic->t_clientid		= client;
3432 	tic->t_flags		= XLOG_TIC_INITED;
3433 	tic->t_trans_type	= 0;
3434 	if (xflags & XFS_LOG_PERM_RESERV)
3435 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3436 	sv_init(&tic->t_wait, SV_DEFAULT, "logtick");
3437 
3438 	xlog_tic_reset_res(tic);
3439 
3440 	return tic;
3441 }
3442 
3443 
3444 /******************************************************************************
3445  *
3446  *		Log debug routines
3447  *
3448  ******************************************************************************
3449  */
3450 #if defined(DEBUG)
3451 /*
3452  * Make sure that the destination ptr is within the valid data region of
3453  * one of the iclogs.  This uses backup pointers stored in a different
3454  * part of the log in case we trash the log structure.
3455  */
3456 void
3457 xlog_verify_dest_ptr(xlog_t     *log,
3458 		     __psint_t  ptr)
3459 {
3460 	int i;
3461 	int good_ptr = 0;
3462 
3463 	for (i=0; i < log->l_iclog_bufs; i++) {
3464 		if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3465 		    ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3466 			good_ptr++;
3467 	}
3468 	if (! good_ptr)
3469 		xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3470 }	/* xlog_verify_dest_ptr */
3471 
3472 STATIC void
3473 xlog_verify_grant_head(xlog_t *log, int equals)
3474 {
3475     if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3476 	if (equals)
3477 	    ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3478 	else
3479 	    ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3480     } else {
3481 	ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3482 	ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3483     }
3484 }	/* xlog_verify_grant_head */
3485 
3486 /* check if it will fit */
3487 STATIC void
3488 xlog_verify_tail_lsn(xlog_t	    *log,
3489 		     xlog_in_core_t *iclog,
3490 		     xfs_lsn_t	    tail_lsn)
3491 {
3492     int blocks;
3493 
3494     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3495 	blocks =
3496 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3497 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3498 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3499     } else {
3500 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3501 
3502 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3503 	    xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3504 
3505 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3506 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3507 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3508     }
3509 }	/* xlog_verify_tail_lsn */
3510 
3511 /*
3512  * Perform a number of checks on the iclog before writing to disk.
3513  *
3514  * 1. Make sure the iclogs are still circular
3515  * 2. Make sure we have a good magic number
3516  * 3. Make sure we don't have magic numbers in the data
3517  * 4. Check fields of each log operation header for:
3518  *	A. Valid client identifier
3519  *	B. tid ptr value falls in valid ptr space (user space code)
3520  *	C. Length in log record header is correct according to the
3521  *		individual operation headers within record.
3522  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3523  *	log, check the preceding blocks of the physical log to make sure all
3524  *	the cycle numbers agree with the current cycle number.
3525  */
3526 STATIC void
3527 xlog_verify_iclog(xlog_t	 *log,
3528 		  xlog_in_core_t *iclog,
3529 		  int		 count,
3530 		  boolean_t	 syncing)
3531 {
3532 	xlog_op_header_t	*ophead;
3533 	xlog_in_core_t		*icptr;
3534 	xlog_in_core_2_t	*xhdr;
3535 	xfs_caddr_t		ptr;
3536 	xfs_caddr_t		base_ptr;
3537 	__psint_t		field_offset;
3538 	__uint8_t		clientid;
3539 	int			len, i, j, k, op_len;
3540 	int			idx;
3541 
3542 	/* check validity of iclog pointers */
3543 	spin_lock(&log->l_icloglock);
3544 	icptr = log->l_iclog;
3545 	for (i=0; i < log->l_iclog_bufs; i++) {
3546 		if (icptr == NULL)
3547 			xlog_panic("xlog_verify_iclog: invalid ptr");
3548 		icptr = icptr->ic_next;
3549 	}
3550 	if (icptr != log->l_iclog)
3551 		xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3552 	spin_unlock(&log->l_icloglock);
3553 
3554 	/* check log magic numbers */
3555 	if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3556 		xlog_panic("xlog_verify_iclog: invalid magic num");
3557 
3558 	ptr = (xfs_caddr_t) &iclog->ic_header;
3559 	for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3560 	     ptr += BBSIZE) {
3561 		if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3562 			xlog_panic("xlog_verify_iclog: unexpected magic num");
3563 	}
3564 
3565 	/* check fields */
3566 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3567 	ptr = iclog->ic_datap;
3568 	base_ptr = ptr;
3569 	ophead = (xlog_op_header_t *)ptr;
3570 	xhdr = iclog->ic_data;
3571 	for (i = 0; i < len; i++) {
3572 		ophead = (xlog_op_header_t *)ptr;
3573 
3574 		/* clientid is only 1 byte */
3575 		field_offset = (__psint_t)
3576 			       ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3577 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3578 			clientid = ophead->oh_clientid;
3579 		} else {
3580 			idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3581 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3582 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3583 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3584 				clientid = xlog_get_client_id(
3585 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3586 			} else {
3587 				clientid = xlog_get_client_id(
3588 					iclog->ic_header.h_cycle_data[idx]);
3589 			}
3590 		}
3591 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3592 			cmn_err(CE_WARN, "xlog_verify_iclog: "
3593 				"invalid clientid %d op 0x%p offset 0x%lx",
3594 				clientid, ophead, (unsigned long)field_offset);
3595 
3596 		/* check length */
3597 		field_offset = (__psint_t)
3598 			       ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3599 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3600 			op_len = be32_to_cpu(ophead->oh_len);
3601 		} else {
3602 			idx = BTOBBT((__psint_t)&ophead->oh_len -
3603 				    (__psint_t)iclog->ic_datap);
3604 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3605 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3606 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3607 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3608 			} else {
3609 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3610 			}
3611 		}
3612 		ptr += sizeof(xlog_op_header_t) + op_len;
3613 	}
3614 }	/* xlog_verify_iclog */
3615 #endif
3616 
3617 /*
3618  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3619  */
3620 STATIC int
3621 xlog_state_ioerror(
3622 	xlog_t	*log)
3623 {
3624 	xlog_in_core_t	*iclog, *ic;
3625 
3626 	iclog = log->l_iclog;
3627 	if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3628 		/*
3629 		 * Mark all the incore logs IOERROR.
3630 		 * From now on, no log flushes will result.
3631 		 */
3632 		ic = iclog;
3633 		do {
3634 			ic->ic_state = XLOG_STATE_IOERROR;
3635 			ic = ic->ic_next;
3636 		} while (ic != iclog);
3637 		return 0;
3638 	}
3639 	/*
3640 	 * Return non-zero, if state transition has already happened.
3641 	 */
3642 	return 1;
3643 }
3644 
3645 /*
3646  * This is called from xfs_force_shutdown, when we're forcibly
3647  * shutting down the filesystem, typically because of an IO error.
3648  * Our main objectives here are to make sure that:
3649  *	a. the filesystem gets marked 'SHUTDOWN' for all interested
3650  *	   parties to find out, 'atomically'.
3651  *	b. those who're sleeping on log reservations, pinned objects and
3652  *	    other resources get woken up, and be told the bad news.
3653  *	c. nothing new gets queued up after (a) and (b) are done.
3654  *	d. if !logerror, flush the iclogs to disk, then seal them off
3655  *	   for business.
3656  */
3657 int
3658 xfs_log_force_umount(
3659 	struct xfs_mount	*mp,
3660 	int			logerror)
3661 {
3662 	xlog_ticket_t	*tic;
3663 	xlog_t		*log;
3664 	int		retval;
3665 
3666 	log = mp->m_log;
3667 
3668 	/*
3669 	 * If this happens during log recovery, don't worry about
3670 	 * locking; the log isn't open for business yet.
3671 	 */
3672 	if (!log ||
3673 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3674 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3675 		if (mp->m_sb_bp)
3676 			XFS_BUF_DONE(mp->m_sb_bp);
3677 		return 0;
3678 	}
3679 
3680 	/*
3681 	 * Somebody could've already done the hard work for us.
3682 	 * No need to get locks for this.
3683 	 */
3684 	if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3685 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3686 		return 1;
3687 	}
3688 	retval = 0;
3689 	/*
3690 	 * We must hold both the GRANT lock and the LOG lock,
3691 	 * before we mark the filesystem SHUTDOWN and wake
3692 	 * everybody up to tell the bad news.
3693 	 */
3694 	spin_lock(&log->l_icloglock);
3695 	spin_lock(&log->l_grant_lock);
3696 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3697 	if (mp->m_sb_bp)
3698 		XFS_BUF_DONE(mp->m_sb_bp);
3699 
3700 	/*
3701 	 * This flag is sort of redundant because of the mount flag, but
3702 	 * it's good to maintain the separation between the log and the rest
3703 	 * of XFS.
3704 	 */
3705 	log->l_flags |= XLOG_IO_ERROR;
3706 
3707 	/*
3708 	 * If we hit a log error, we want to mark all the iclogs IOERROR
3709 	 * while we're still holding the loglock.
3710 	 */
3711 	if (logerror)
3712 		retval = xlog_state_ioerror(log);
3713 	spin_unlock(&log->l_icloglock);
3714 
3715 	/*
3716 	 * We don't want anybody waiting for log reservations
3717 	 * after this. That means we have to wake up everybody
3718 	 * queued up on reserve_headq as well as write_headq.
3719 	 * In addition, we make sure in xlog_{re}grant_log_space
3720 	 * that we don't enqueue anything once the SHUTDOWN flag
3721 	 * is set, and this action is protected by the GRANTLOCK.
3722 	 */
3723 	if ((tic = log->l_reserve_headq)) {
3724 		do {
3725 			sv_signal(&tic->t_wait);
3726 			tic = tic->t_next;
3727 		} while (tic != log->l_reserve_headq);
3728 	}
3729 
3730 	if ((tic = log->l_write_headq)) {
3731 		do {
3732 			sv_signal(&tic->t_wait);
3733 			tic = tic->t_next;
3734 		} while (tic != log->l_write_headq);
3735 	}
3736 	spin_unlock(&log->l_grant_lock);
3737 
3738 	if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3739 		ASSERT(!logerror);
3740 		/*
3741 		 * Force the incore logs to disk before shutting the
3742 		 * log down completely.
3743 		 */
3744 		_xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3745 
3746 		spin_lock(&log->l_icloglock);
3747 		retval = xlog_state_ioerror(log);
3748 		spin_unlock(&log->l_icloglock);
3749 	}
3750 	/*
3751 	 * Wake up everybody waiting on xfs_log_force.
3752 	 * Callback all log item committed functions as if the
3753 	 * log writes were completed.
3754 	 */
3755 	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3756 
3757 #ifdef XFSERRORDEBUG
3758 	{
3759 		xlog_in_core_t	*iclog;
3760 
3761 		spin_lock(&log->l_icloglock);
3762 		iclog = log->l_iclog;
3763 		do {
3764 			ASSERT(iclog->ic_callback == 0);
3765 			iclog = iclog->ic_next;
3766 		} while (iclog != log->l_iclog);
3767 		spin_unlock(&log->l_icloglock);
3768 	}
3769 #endif
3770 	/* return non-zero if log IOERROR transition had already happened */
3771 	return retval;
3772 }
3773 
3774 STATIC int
3775 xlog_iclogs_empty(xlog_t *log)
3776 {
3777 	xlog_in_core_t	*iclog;
3778 
3779 	iclog = log->l_iclog;
3780 	do {
3781 		/* endianness does not matter here, zero is zero in
3782 		 * any language.
3783 		 */
3784 		if (iclog->ic_header.h_num_logops)
3785 			return 0;
3786 		iclog = iclog->ic_next;
3787 	} while (iclog != log->l_iclog);
3788 	return 1;
3789 }
3790