xref: /linux/fs/xfs/xfs_log.c (revision 6e8331ac6973435b1e7604c30f2ad394035b46e1)
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 
44 
45 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
46 	{ (ptr) += (bytes); \
47 	  (len) -= (bytes); \
48 	  (off) += (bytes);}
49 
50 /* Local miscellaneous function prototypes */
51 STATIC int	 xlog_bdstrat_cb(struct xfs_buf *);
52 STATIC int	 xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
53 				    xlog_in_core_t **, xfs_lsn_t *);
54 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t	*mp,
55 				xfs_buftarg_t	*log_target,
56 				xfs_daddr_t	blk_offset,
57 				int		num_bblks);
58 STATIC int	 xlog_space_left(xlog_t *log, int cycle, int bytes);
59 STATIC int	 xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
60 STATIC void	 xlog_dealloc_log(xlog_t *log);
61 STATIC int	 xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
62 			    int nentries, xfs_log_ticket_t tic,
63 			    xfs_lsn_t *start_lsn,
64 			    xlog_in_core_t **commit_iclog,
65 			    uint flags);
66 
67 /* local state machine functions */
68 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
69 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
70 STATIC int  xlog_state_get_iclog_space(xlog_t		*log,
71 				       int		len,
72 				       xlog_in_core_t	**iclog,
73 				       xlog_ticket_t	*ticket,
74 				       int		*continued_write,
75 				       int		*logoffsetp);
76 STATIC void xlog_state_put_ticket(xlog_t	*log,
77 				  xlog_ticket_t *tic);
78 STATIC int  xlog_state_release_iclog(xlog_t		*log,
79 				     xlog_in_core_t	*iclog);
80 STATIC void xlog_state_switch_iclogs(xlog_t		*log,
81 				     xlog_in_core_t *iclog,
82 				     int		eventual_size);
83 STATIC int  xlog_state_sync(xlog_t			*log,
84 			    xfs_lsn_t 			lsn,
85 			    uint			flags,
86 			    int				*log_flushed);
87 STATIC int  xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
88 STATIC void xlog_state_want_sync(xlog_t	*log, xlog_in_core_t *iclog);
89 
90 /* local functions to manipulate grant head */
91 STATIC int  xlog_grant_log_space(xlog_t		*log,
92 				 xlog_ticket_t	*xtic);
93 STATIC void xlog_grant_push_ail(xfs_mount_t	*mp,
94 				int		need_bytes);
95 STATIC void xlog_regrant_reserve_log_space(xlog_t	 *log,
96 					   xlog_ticket_t *ticket);
97 STATIC int xlog_regrant_write_log_space(xlog_t		*log,
98 					 xlog_ticket_t  *ticket);
99 STATIC void xlog_ungrant_log_space(xlog_t	 *log,
100 				   xlog_ticket_t *ticket);
101 
102 
103 /* local ticket functions */
104 STATIC void		xlog_state_ticket_alloc(xlog_t *log);
105 STATIC xlog_ticket_t	*xlog_ticket_get(xlog_t *log,
106 					 int	unit_bytes,
107 					 int	count,
108 					 char	clientid,
109 					 uint	flags);
110 STATIC void		xlog_ticket_put(xlog_t *log, xlog_ticket_t *ticket);
111 
112 #if defined(DEBUG)
113 STATIC void	xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
114 STATIC void	xlog_verify_grant_head(xlog_t *log, int equals);
115 STATIC void	xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
116 				  int count, boolean_t syncing);
117 STATIC void	xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
118 				     xfs_lsn_t tail_lsn);
119 #else
120 #define xlog_verify_dest_ptr(a,b)
121 #define xlog_verify_grant_head(a,b)
122 #define xlog_verify_iclog(a,b,c,d)
123 #define xlog_verify_tail_lsn(a,b,c)
124 #endif
125 
126 STATIC int	xlog_iclogs_empty(xlog_t *log);
127 
128 #if defined(XFS_LOG_TRACE)
129 void
130 xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
131 {
132 	unsigned long cnts;
133 
134 	if (!log->l_grant_trace) {
135 		log->l_grant_trace = ktrace_alloc(2048, KM_NOSLEEP);
136 		if (!log->l_grant_trace)
137 			return;
138 	}
139 	/* ticket counts are 1 byte each */
140 	cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
141 
142 	ktrace_enter(log->l_grant_trace,
143 		     (void *)tic,
144 		     (void *)log->l_reserve_headq,
145 		     (void *)log->l_write_headq,
146 		     (void *)((unsigned long)log->l_grant_reserve_cycle),
147 		     (void *)((unsigned long)log->l_grant_reserve_bytes),
148 		     (void *)((unsigned long)log->l_grant_write_cycle),
149 		     (void *)((unsigned long)log->l_grant_write_bytes),
150 		     (void *)((unsigned long)log->l_curr_cycle),
151 		     (void *)((unsigned long)log->l_curr_block),
152 		     (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
153 		     (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
154 		     (void *)string,
155 		     (void *)((unsigned long)tic->t_trans_type),
156 		     (void *)cnts,
157 		     (void *)((unsigned long)tic->t_curr_res),
158 		     (void *)((unsigned long)tic->t_unit_res));
159 }
160 
161 void
162 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
163 {
164 	if (!iclog->ic_trace)
165 		iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
166 	ktrace_enter(iclog->ic_trace,
167 		     (void *)((unsigned long)state),
168 		     (void *)((unsigned long)current_pid()),
169 		     (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
170 		     (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
171 		     (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
172 		     (void *)NULL, (void *)NULL);
173 }
174 #else
175 #define	xlog_trace_loggrant(log,tic,string)
176 #define	xlog_trace_iclog(iclog,state)
177 #endif /* XFS_LOG_TRACE */
178 
179 
180 static void
181 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
182 {
183 	if (*qp) {
184 		tic->t_next	    = (*qp);
185 		tic->t_prev	    = (*qp)->t_prev;
186 		(*qp)->t_prev->t_next = tic;
187 		(*qp)->t_prev	    = tic;
188 	} else {
189 		tic->t_prev = tic->t_next = tic;
190 		*qp = tic;
191 	}
192 
193 	tic->t_flags |= XLOG_TIC_IN_Q;
194 }
195 
196 static void
197 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
198 {
199 	if (tic == tic->t_next) {
200 		*qp = NULL;
201 	} else {
202 		*qp = tic->t_next;
203 		tic->t_next->t_prev = tic->t_prev;
204 		tic->t_prev->t_next = tic->t_next;
205 	}
206 
207 	tic->t_next = tic->t_prev = NULL;
208 	tic->t_flags &= ~XLOG_TIC_IN_Q;
209 }
210 
211 static void
212 xlog_grant_sub_space(struct log *log, int bytes)
213 {
214 	log->l_grant_write_bytes -= bytes;
215 	if (log->l_grant_write_bytes < 0) {
216 		log->l_grant_write_bytes += log->l_logsize;
217 		log->l_grant_write_cycle--;
218 	}
219 
220 	log->l_grant_reserve_bytes -= bytes;
221 	if ((log)->l_grant_reserve_bytes < 0) {
222 		log->l_grant_reserve_bytes += log->l_logsize;
223 		log->l_grant_reserve_cycle--;
224 	}
225 
226 }
227 
228 static void
229 xlog_grant_add_space_write(struct log *log, int bytes)
230 {
231 	log->l_grant_write_bytes += bytes;
232 	if (log->l_grant_write_bytes > log->l_logsize) {
233 		log->l_grant_write_bytes -= log->l_logsize;
234 		log->l_grant_write_cycle++;
235 	}
236 }
237 
238 static void
239 xlog_grant_add_space_reserve(struct log *log, int bytes)
240 {
241 	log->l_grant_reserve_bytes += bytes;
242 	if (log->l_grant_reserve_bytes > log->l_logsize) {
243 		log->l_grant_reserve_bytes -= log->l_logsize;
244 		log->l_grant_reserve_cycle++;
245 	}
246 }
247 
248 static inline void
249 xlog_grant_add_space(struct log *log, int bytes)
250 {
251 	xlog_grant_add_space_write(log, bytes);
252 	xlog_grant_add_space_reserve(log, bytes);
253 }
254 
255 
256 /*
257  * NOTES:
258  *
259  *	1. currblock field gets updated at startup and after in-core logs
260  *		marked as with WANT_SYNC.
261  */
262 
263 /*
264  * This routine is called when a user of a log manager ticket is done with
265  * the reservation.  If the ticket was ever used, then a commit record for
266  * the associated transaction is written out as a log operation header with
267  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
268  * a given ticket.  If the ticket was one with a permanent reservation, then
269  * a few operations are done differently.  Permanent reservation tickets by
270  * default don't release the reservation.  They just commit the current
271  * transaction with the belief that the reservation is still needed.  A flag
272  * must be passed in before permanent reservations are actually released.
273  * When these type of tickets are not released, they need to be set into
274  * the inited state again.  By doing this, a start record will be written
275  * out when the next write occurs.
276  */
277 xfs_lsn_t
278 xfs_log_done(xfs_mount_t	*mp,
279 	     xfs_log_ticket_t	xtic,
280 	     void		**iclog,
281 	     uint		flags)
282 {
283 	xlog_t		*log    = mp->m_log;
284 	xlog_ticket_t	*ticket = (xfs_log_ticket_t) xtic;
285 	xfs_lsn_t	lsn	= 0;
286 
287 	if (XLOG_FORCED_SHUTDOWN(log) ||
288 	    /*
289 	     * If nothing was ever written, don't write out commit record.
290 	     * If we get an error, just continue and give back the log ticket.
291 	     */
292 	    (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
293 	     (xlog_commit_record(mp, ticket,
294 				 (xlog_in_core_t **)iclog, &lsn)))) {
295 		lsn = (xfs_lsn_t) -1;
296 		if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
297 			flags |= XFS_LOG_REL_PERM_RESERV;
298 		}
299 	}
300 
301 
302 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
303 	    (flags & XFS_LOG_REL_PERM_RESERV)) {
304 		/*
305 		 * Release ticket if not permanent reservation or a specific
306 		 * request has been made to release a permanent reservation.
307 		 */
308 		xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
309 		xlog_ungrant_log_space(log, ticket);
310 		xlog_state_put_ticket(log, ticket);
311 	} else {
312 		xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
313 		xlog_regrant_reserve_log_space(log, ticket);
314 	}
315 
316 	/* If this ticket was a permanent reservation and we aren't
317 	 * trying to release it, reset the inited flags; so next time
318 	 * we write, a start record will be written out.
319 	 */
320 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) &&
321 	    (flags & XFS_LOG_REL_PERM_RESERV) == 0)
322 		ticket->t_flags |= XLOG_TIC_INITED;
323 
324 	return lsn;
325 }	/* xfs_log_done */
326 
327 
328 /*
329  * Force the in-core log to disk.  If flags == XFS_LOG_SYNC,
330  *	the force is done synchronously.
331  *
332  * Asynchronous forces are implemented by setting the WANT_SYNC
333  * bit in the appropriate in-core log and then returning.
334  *
335  * Synchronous forces are implemented with a semaphore.  All callers
336  * to force a given lsn to disk will wait on a semaphore attached to the
337  * specific in-core log.  When given in-core log finally completes its
338  * write to disk, that thread will wake up all threads waiting on the
339  * semaphore.
340  */
341 int
342 _xfs_log_force(
343 	xfs_mount_t	*mp,
344 	xfs_lsn_t	lsn,
345 	uint		flags,
346 	int		*log_flushed)
347 {
348 	xlog_t		*log = mp->m_log;
349 	int		dummy;
350 
351 	if (!log_flushed)
352 		log_flushed = &dummy;
353 
354 	ASSERT(flags & XFS_LOG_FORCE);
355 
356 	XFS_STATS_INC(xs_log_force);
357 
358 	if (log->l_flags & XLOG_IO_ERROR)
359 		return XFS_ERROR(EIO);
360 	if (lsn == 0)
361 		return xlog_state_sync_all(log, flags, log_flushed);
362 	else
363 		return xlog_state_sync(log, lsn, flags, log_flushed);
364 }	/* xfs_log_force */
365 
366 /*
367  * Attaches a new iclog I/O completion callback routine during
368  * transaction commit.  If the log is in error state, a non-zero
369  * return code is handed back and the caller is responsible for
370  * executing the callback at an appropriate time.
371  */
372 int
373 xfs_log_notify(xfs_mount_t	  *mp,		/* mount of partition */
374 	       void		  *iclog_hndl,	/* iclog to hang callback off */
375 	       xfs_log_callback_t *cb)
376 {
377 	xlog_t *log = mp->m_log;
378 	xlog_in_core_t	  *iclog = (xlog_in_core_t *)iclog_hndl;
379 	int	abortflg, spl;
380 
381 	cb->cb_next = NULL;
382 	spl = LOG_LOCK(log);
383 	abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
384 	if (!abortflg) {
385 		ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
386 			      (iclog->ic_state == XLOG_STATE_WANT_SYNC));
387 		cb->cb_next = NULL;
388 		*(iclog->ic_callback_tail) = cb;
389 		iclog->ic_callback_tail = &(cb->cb_next);
390 	}
391 	LOG_UNLOCK(log, spl);
392 	return abortflg;
393 }	/* xfs_log_notify */
394 
395 int
396 xfs_log_release_iclog(xfs_mount_t *mp,
397 		      void	  *iclog_hndl)
398 {
399 	xlog_t *log = mp->m_log;
400 	xlog_in_core_t	  *iclog = (xlog_in_core_t *)iclog_hndl;
401 
402 	if (xlog_state_release_iclog(log, iclog)) {
403 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
404 		return EIO;
405 	}
406 
407 	return 0;
408 }
409 
410 /*
411  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
412  *	to the reservation.
413  *  2. Potentially, push buffers at tail of log to disk.
414  *
415  * Each reservation is going to reserve extra space for a log record header.
416  * When writes happen to the on-disk log, we don't subtract the length of the
417  * log record header from any reservation.  By wasting space in each
418  * reservation, we prevent over allocation problems.
419  */
420 int
421 xfs_log_reserve(xfs_mount_t	 *mp,
422 		int		 unit_bytes,
423 		int		 cnt,
424 		xfs_log_ticket_t *ticket,
425 		__uint8_t	 client,
426 		uint		 flags,
427 		uint		 t_type)
428 {
429 	xlog_t		*log = mp->m_log;
430 	xlog_ticket_t	*internal_ticket;
431 	int		retval = 0;
432 
433 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
434 	ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
435 
436 	if (XLOG_FORCED_SHUTDOWN(log))
437 		return XFS_ERROR(EIO);
438 
439 	XFS_STATS_INC(xs_try_logspace);
440 
441 	if (*ticket != NULL) {
442 		ASSERT(flags & XFS_LOG_PERM_RESERV);
443 		internal_ticket = (xlog_ticket_t *)*ticket;
444 		xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
445 		xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
446 		retval = xlog_regrant_write_log_space(log, internal_ticket);
447 	} else {
448 		/* may sleep if need to allocate more tickets */
449 		internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
450 						  client, flags);
451 		internal_ticket->t_trans_type = t_type;
452 		*ticket = internal_ticket;
453 		xlog_trace_loggrant(log, internal_ticket,
454 			(internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
455 			"xfs_log_reserve: create new ticket (permanent trans)" :
456 			"xfs_log_reserve: create new ticket");
457 		xlog_grant_push_ail(mp,
458 				    (internal_ticket->t_unit_res *
459 				     internal_ticket->t_cnt));
460 		retval = xlog_grant_log_space(log, internal_ticket);
461 	}
462 
463 	return retval;
464 }	/* xfs_log_reserve */
465 
466 
467 /*
468  * Mount a log filesystem
469  *
470  * mp		- ubiquitous xfs mount point structure
471  * log_target	- buftarg of on-disk log device
472  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
473  * num_bblocks	- Number of BBSIZE blocks in on-disk log
474  *
475  * Return error or zero.
476  */
477 int
478 xfs_log_mount(xfs_mount_t	*mp,
479 	      xfs_buftarg_t	*log_target,
480 	      xfs_daddr_t	blk_offset,
481 	      int		num_bblks)
482 {
483 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
484 		cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
485 	else {
486 		cmn_err(CE_NOTE,
487 			"!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
488 			mp->m_fsname);
489 		ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
490 	}
491 
492 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
493 
494 	/*
495 	 * skip log recovery on a norecovery mount.  pretend it all
496 	 * just worked.
497 	 */
498 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
499 		bhv_vfs_t	*vfsp = XFS_MTOVFS(mp);
500 		int		error, readonly = (vfsp->vfs_flag & VFS_RDONLY);
501 
502 		if (readonly)
503 			vfsp->vfs_flag &= ~VFS_RDONLY;
504 
505 		error = xlog_recover(mp->m_log);
506 
507 		if (readonly)
508 			vfsp->vfs_flag |= VFS_RDONLY;
509 		if (error) {
510 			cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
511 			xlog_dealloc_log(mp->m_log);
512 			return error;
513 		}
514 	}
515 
516 	/* Normal transactions can now occur */
517 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
518 
519 	/* End mounting message in xfs_log_mount_finish */
520 	return 0;
521 }	/* xfs_log_mount */
522 
523 /*
524  * Finish the recovery of the file system.  This is separate from
525  * the xfs_log_mount() call, because it depends on the code in
526  * xfs_mountfs() to read in the root and real-time bitmap inodes
527  * between calling xfs_log_mount() and here.
528  *
529  * mp		- ubiquitous xfs mount point structure
530  */
531 int
532 xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
533 {
534 	int	error;
535 
536 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
537 		error = xlog_recover_finish(mp->m_log, mfsi_flags);
538 	else {
539 		error = 0;
540 		ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
541 	}
542 
543 	return error;
544 }
545 
546 /*
547  * Unmount processing for the log.
548  */
549 int
550 xfs_log_unmount(xfs_mount_t *mp)
551 {
552 	int		error;
553 
554 	error = xfs_log_unmount_write(mp);
555 	xfs_log_unmount_dealloc(mp);
556 	return error;
557 }
558 
559 /*
560  * Final log writes as part of unmount.
561  *
562  * Mark the filesystem clean as unmount happens.  Note that during relocation
563  * this routine needs to be executed as part of source-bag while the
564  * deallocation must not be done until source-end.
565  */
566 
567 /*
568  * Unmount record used to have a string "Unmount filesystem--" in the
569  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
570  * We just write the magic number now since that particular field isn't
571  * currently architecture converted and "nUmount" is a bit foo.
572  * As far as I know, there weren't any dependencies on the old behaviour.
573  */
574 
575 int
576 xfs_log_unmount_write(xfs_mount_t *mp)
577 {
578 	xlog_t		 *log = mp->m_log;
579 	xlog_in_core_t	 *iclog;
580 #ifdef DEBUG
581 	xlog_in_core_t	 *first_iclog;
582 #endif
583 	xfs_log_iovec_t  reg[1];
584 	xfs_log_ticket_t tic = NULL;
585 	xfs_lsn_t	 lsn;
586 	int		 error;
587 	SPLDECL(s);
588 
589 	/* the data section must be 32 bit size aligned */
590 	struct {
591 	    __uint16_t magic;
592 	    __uint16_t pad1;
593 	    __uint32_t pad2; /* may as well make it 64 bits */
594 	} magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
595 
596 	/*
597 	 * Don't write out unmount record on read-only mounts.
598 	 * Or, if we are doing a forced umount (typically because of IO errors).
599 	 */
600 	if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
601 		return 0;
602 
603 	xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
604 
605 #ifdef DEBUG
606 	first_iclog = iclog = log->l_iclog;
607 	do {
608 		if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
609 			ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
610 			ASSERT(iclog->ic_offset == 0);
611 		}
612 		iclog = iclog->ic_next;
613 	} while (iclog != first_iclog);
614 #endif
615 	if (! (XLOG_FORCED_SHUTDOWN(log))) {
616 		reg[0].i_addr = (void*)&magic;
617 		reg[0].i_len  = sizeof(magic);
618 		XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
619 
620 		error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0, 0);
621 		if (!error) {
622 			/* remove inited flag */
623 			((xlog_ticket_t *)tic)->t_flags = 0;
624 			error = xlog_write(mp, reg, 1, tic, &lsn,
625 					   NULL, XLOG_UNMOUNT_TRANS);
626 			/*
627 			 * At this point, we're umounting anyway,
628 			 * so there's no point in transitioning log state
629 			 * to IOERROR. Just continue...
630 			 */
631 		}
632 
633 		if (error) {
634 			xfs_fs_cmn_err(CE_ALERT, mp,
635 				"xfs_log_unmount: unmount record failed");
636 		}
637 
638 
639 		s = LOG_LOCK(log);
640 		iclog = log->l_iclog;
641 		iclog->ic_refcnt++;
642 		LOG_UNLOCK(log, s);
643 		xlog_state_want_sync(log, iclog);
644 		(void) xlog_state_release_iclog(log, iclog);
645 
646 		s = LOG_LOCK(log);
647 		if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
648 		      iclog->ic_state == XLOG_STATE_DIRTY)) {
649 			if (!XLOG_FORCED_SHUTDOWN(log)) {
650 				sv_wait(&iclog->ic_forcesema, PMEM,
651 					&log->l_icloglock, s);
652 			} else {
653 				LOG_UNLOCK(log, s);
654 			}
655 		} else {
656 			LOG_UNLOCK(log, s);
657 		}
658 		if (tic)
659 			xlog_state_put_ticket(log, tic);
660 	} else {
661 		/*
662 		 * We're already in forced_shutdown mode, couldn't
663 		 * even attempt to write out the unmount transaction.
664 		 *
665 		 * Go through the motions of sync'ing and releasing
666 		 * the iclog, even though no I/O will actually happen,
667 		 * we need to wait for other log I/Os that may already
668 		 * be in progress.  Do this as a separate section of
669 		 * code so we'll know if we ever get stuck here that
670 		 * we're in this odd situation of trying to unmount
671 		 * a file system that went into forced_shutdown as
672 		 * the result of an unmount..
673 		 */
674 		s = LOG_LOCK(log);
675 		iclog = log->l_iclog;
676 		iclog->ic_refcnt++;
677 		LOG_UNLOCK(log, s);
678 
679 		xlog_state_want_sync(log, iclog);
680 		(void) xlog_state_release_iclog(log, iclog);
681 
682 		s = LOG_LOCK(log);
683 
684 		if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
685 			|| iclog->ic_state == XLOG_STATE_DIRTY
686 			|| iclog->ic_state == XLOG_STATE_IOERROR) ) {
687 
688 				sv_wait(&iclog->ic_forcesema, PMEM,
689 					&log->l_icloglock, s);
690 		} else {
691 			LOG_UNLOCK(log, s);
692 		}
693 	}
694 
695 	return 0;
696 }	/* xfs_log_unmount_write */
697 
698 /*
699  * Deallocate log structures for unmount/relocation.
700  */
701 void
702 xfs_log_unmount_dealloc(xfs_mount_t *mp)
703 {
704 	xlog_dealloc_log(mp->m_log);
705 }
706 
707 /*
708  * Write region vectors to log.  The write happens using the space reservation
709  * of the ticket (tic).  It is not a requirement that all writes for a given
710  * transaction occur with one call to xfs_log_write().
711  */
712 int
713 xfs_log_write(xfs_mount_t *	mp,
714 	      xfs_log_iovec_t	reg[],
715 	      int		nentries,
716 	      xfs_log_ticket_t	tic,
717 	      xfs_lsn_t		*start_lsn)
718 {
719 	int	error;
720 	xlog_t *log = mp->m_log;
721 
722 	if (XLOG_FORCED_SHUTDOWN(log))
723 		return XFS_ERROR(EIO);
724 
725 	if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
726 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
727 	}
728 	return error;
729 }	/* xfs_log_write */
730 
731 
732 void
733 xfs_log_move_tail(xfs_mount_t	*mp,
734 		  xfs_lsn_t	tail_lsn)
735 {
736 	xlog_ticket_t	*tic;
737 	xlog_t		*log = mp->m_log;
738 	int		need_bytes, free_bytes, cycle, bytes;
739 	SPLDECL(s);
740 
741 	if (XLOG_FORCED_SHUTDOWN(log))
742 		return;
743 	ASSERT(!XFS_FORCED_SHUTDOWN(mp));
744 
745 	if (tail_lsn == 0) {
746 		/* needed since sync_lsn is 64 bits */
747 		s = LOG_LOCK(log);
748 		tail_lsn = log->l_last_sync_lsn;
749 		LOG_UNLOCK(log, s);
750 	}
751 
752 	s = GRANT_LOCK(log);
753 
754 	/* Also an invalid lsn.  1 implies that we aren't passing in a valid
755 	 * tail_lsn.
756 	 */
757 	if (tail_lsn != 1) {
758 		log->l_tail_lsn = tail_lsn;
759 	}
760 
761 	if ((tic = log->l_write_headq)) {
762 #ifdef DEBUG
763 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
764 			panic("Recovery problem");
765 #endif
766 		cycle = log->l_grant_write_cycle;
767 		bytes = log->l_grant_write_bytes;
768 		free_bytes = xlog_space_left(log, cycle, bytes);
769 		do {
770 			ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
771 
772 			if (free_bytes < tic->t_unit_res && tail_lsn != 1)
773 				break;
774 			tail_lsn = 0;
775 			free_bytes -= tic->t_unit_res;
776 			sv_signal(&tic->t_sema);
777 			tic = tic->t_next;
778 		} while (tic != log->l_write_headq);
779 	}
780 	if ((tic = log->l_reserve_headq)) {
781 #ifdef DEBUG
782 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
783 			panic("Recovery problem");
784 #endif
785 		cycle = log->l_grant_reserve_cycle;
786 		bytes = log->l_grant_reserve_bytes;
787 		free_bytes = xlog_space_left(log, cycle, bytes);
788 		do {
789 			if (tic->t_flags & XLOG_TIC_PERM_RESERV)
790 				need_bytes = tic->t_unit_res*tic->t_cnt;
791 			else
792 				need_bytes = tic->t_unit_res;
793 			if (free_bytes < need_bytes && tail_lsn != 1)
794 				break;
795 			tail_lsn = 0;
796 			free_bytes -= need_bytes;
797 			sv_signal(&tic->t_sema);
798 			tic = tic->t_next;
799 		} while (tic != log->l_reserve_headq);
800 	}
801 	GRANT_UNLOCK(log, s);
802 }	/* xfs_log_move_tail */
803 
804 /*
805  * Determine if we have a transaction that has gone to disk
806  * that needs to be covered. Log activity needs to be idle (no AIL and
807  * nothing in the iclogs). And, we need to be in the right state indicating
808  * something has gone out.
809  */
810 int
811 xfs_log_need_covered(xfs_mount_t *mp)
812 {
813 	SPLDECL(s);
814 	int		needed = 0, gen;
815 	xlog_t		*log = mp->m_log;
816 	bhv_vfs_t	*vfsp = XFS_MTOVFS(mp);
817 
818 	if (vfs_test_for_freeze(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
819 	    (vfsp->vfs_flag & VFS_RDONLY))
820 		return 0;
821 
822 	s = LOG_LOCK(log);
823 	if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
824 		(log->l_covered_state == XLOG_STATE_COVER_NEED2))
825 			&& !xfs_trans_first_ail(mp, &gen)
826 			&& xlog_iclogs_empty(log)) {
827 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
828 			log->l_covered_state = XLOG_STATE_COVER_DONE;
829 		else {
830 			ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
831 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
832 		}
833 		needed = 1;
834 	}
835 	LOG_UNLOCK(log, s);
836 	return needed;
837 }
838 
839 /******************************************************************************
840  *
841  *	local routines
842  *
843  ******************************************************************************
844  */
845 
846 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
847  * The log manager must keep track of the last LR which was committed
848  * to disk.  The lsn of this LR will become the new tail_lsn whenever
849  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
850  * the situation where stuff could be written into the log but nothing
851  * was ever in the AIL when asked.  Eventually, we panic since the
852  * tail hits the head.
853  *
854  * We may be holding the log iclog lock upon entering this routine.
855  */
856 xfs_lsn_t
857 xlog_assign_tail_lsn(xfs_mount_t *mp)
858 {
859 	xfs_lsn_t tail_lsn;
860 	SPLDECL(s);
861 	xlog_t	  *log = mp->m_log;
862 
863 	tail_lsn = xfs_trans_tail_ail(mp);
864 	s = GRANT_LOCK(log);
865 	if (tail_lsn != 0) {
866 		log->l_tail_lsn = tail_lsn;
867 	} else {
868 		tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
869 	}
870 	GRANT_UNLOCK(log, s);
871 
872 	return tail_lsn;
873 }	/* xlog_assign_tail_lsn */
874 
875 
876 /*
877  * Return the space in the log between the tail and the head.  The head
878  * is passed in the cycle/bytes formal parms.  In the special case where
879  * the reserve head has wrapped passed the tail, this calculation is no
880  * longer valid.  In this case, just return 0 which means there is no space
881  * in the log.  This works for all places where this function is called
882  * with the reserve head.  Of course, if the write head were to ever
883  * wrap the tail, we should blow up.  Rather than catch this case here,
884  * we depend on other ASSERTions in other parts of the code.   XXXmiken
885  *
886  * This code also handles the case where the reservation head is behind
887  * the tail.  The details of this case are described below, but the end
888  * result is that we return the size of the log as the amount of space left.
889  */
890 int
891 xlog_space_left(xlog_t *log, int cycle, int bytes)
892 {
893 	int free_bytes;
894 	int tail_bytes;
895 	int tail_cycle;
896 
897 	tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
898 	tail_cycle = CYCLE_LSN(log->l_tail_lsn);
899 	if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
900 		free_bytes = log->l_logsize - (bytes - tail_bytes);
901 	} else if ((tail_cycle + 1) < cycle) {
902 		return 0;
903 	} else if (tail_cycle < cycle) {
904 		ASSERT(tail_cycle == (cycle - 1));
905 		free_bytes = tail_bytes - bytes;
906 	} else {
907 		/*
908 		 * The reservation head is behind the tail.
909 		 * In this case we just want to return the size of the
910 		 * log as the amount of space left.
911 		 */
912 		xfs_fs_cmn_err(CE_ALERT, log->l_mp,
913 			"xlog_space_left: head behind tail\n"
914 			"  tail_cycle = %d, tail_bytes = %d\n"
915 			"  GH   cycle = %d, GH   bytes = %d",
916 			tail_cycle, tail_bytes, cycle, bytes);
917 		ASSERT(0);
918 		free_bytes = log->l_logsize;
919 	}
920 	return free_bytes;
921 }	/* xlog_space_left */
922 
923 
924 /*
925  * Log function which is called when an io completes.
926  *
927  * The log manager needs its own routine, in order to control what
928  * happens with the buffer after the write completes.
929  */
930 void
931 xlog_iodone(xfs_buf_t *bp)
932 {
933 	xlog_in_core_t	*iclog;
934 	xlog_t		*l;
935 	int		aborted;
936 
937 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
938 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
939 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
940 	aborted = 0;
941 
942 	/*
943 	 * Some versions of cpp barf on the recursive definition of
944 	 * ic_log -> hic_fields.ic_log and expand ic_log twice when
945 	 * it is passed through two macros.  Workaround broken cpp.
946 	 */
947 	l = iclog->ic_log;
948 
949 	/*
950 	 * Race to shutdown the filesystem if we see an error.
951 	 */
952 	if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
953 			XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
954 		xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
955 		XFS_BUF_STALE(bp);
956 		xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
957 		/*
958 		 * This flag will be propagated to the trans-committed
959 		 * callback routines to let them know that the log-commit
960 		 * didn't succeed.
961 		 */
962 		aborted = XFS_LI_ABORTED;
963 	} else if (iclog->ic_state & XLOG_STATE_IOERROR) {
964 		aborted = XFS_LI_ABORTED;
965 	}
966 	xlog_state_done_syncing(iclog, aborted);
967 	if (!(XFS_BUF_ISASYNC(bp))) {
968 		/*
969 		 * Corresponding psema() will be done in bwrite().  If we don't
970 		 * vsema() here, panic.
971 		 */
972 		XFS_BUF_V_IODONESEMA(bp);
973 	}
974 }	/* xlog_iodone */
975 
976 /*
977  * The bdstrat callback function for log bufs. This gives us a central
978  * place to trap bufs in case we get hit by a log I/O error and need to
979  * shutdown. Actually, in practice, even when we didn't get a log error,
980  * we transition the iclogs to IOERROR state *after* flushing all existing
981  * iclogs to disk. This is because we don't want anymore new transactions to be
982  * started or completed afterwards.
983  */
984 STATIC int
985 xlog_bdstrat_cb(struct xfs_buf *bp)
986 {
987 	xlog_in_core_t *iclog;
988 
989 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
990 
991 	if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
992 	  /* note for irix bstrat will need  struct bdevsw passed
993 	   * Fix the following macro if the code ever is merged
994 	   */
995 	    XFS_bdstrat(bp);
996 		return 0;
997 	}
998 
999 	xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1000 	XFS_BUF_ERROR(bp, EIO);
1001 	XFS_BUF_STALE(bp);
1002 	xfs_biodone(bp);
1003 	return XFS_ERROR(EIO);
1004 
1005 
1006 }
1007 
1008 /*
1009  * Return size of each in-core log record buffer.
1010  *
1011  * Low memory machines only get 2 16KB buffers.  We don't want to waste
1012  * memory here.  However, all other machines get at least 2 32KB buffers.
1013  * The number is hard coded because we don't care about the minimum
1014  * memory size, just 32MB systems.
1015  *
1016  * If the filesystem blocksize is too large, we may need to choose a
1017  * larger size since the directory code currently logs entire blocks.
1018  */
1019 
1020 STATIC void
1021 xlog_get_iclog_buffer_size(xfs_mount_t	*mp,
1022 			   xlog_t	*log)
1023 {
1024 	int size;
1025 	int xhdrs;
1026 
1027 	if (mp->m_logbufs <= 0) {
1028 		if (xfs_physmem <= btoc(128*1024*1024)) {
1029 			log->l_iclog_bufs = XLOG_MIN_ICLOGS;
1030 		} else if (xfs_physmem <= btoc(400*1024*1024)) {
1031 			log->l_iclog_bufs = XLOG_MED_ICLOGS;
1032 		} else {	/* 256K with 32K bufs */
1033 			log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1034 		}
1035 	} else {
1036 		log->l_iclog_bufs = mp->m_logbufs;
1037 	}
1038 
1039 	/*
1040 	 * Buffer size passed in from mount system call.
1041 	 */
1042 	if (mp->m_logbsize > 0) {
1043 		size = log->l_iclog_size = mp->m_logbsize;
1044 		log->l_iclog_size_log = 0;
1045 		while (size != 1) {
1046 			log->l_iclog_size_log++;
1047 			size >>= 1;
1048 		}
1049 
1050 		if (XFS_SB_VERSION_HASLOGV2(&mp->m_sb)) {
1051 			/* # headers = size / 32K
1052 			 * one header holds cycles from 32K of data
1053 			 */
1054 
1055 			xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1056 			if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1057 				xhdrs++;
1058 			log->l_iclog_hsize = xhdrs << BBSHIFT;
1059 			log->l_iclog_heads = xhdrs;
1060 		} else {
1061 			ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1062 			log->l_iclog_hsize = BBSIZE;
1063 			log->l_iclog_heads = 1;
1064 		}
1065 		goto done;
1066 	}
1067 
1068 	/*
1069 	 * Special case machines that have less than 32MB of memory.
1070 	 * All machines with more memory use 32KB buffers.
1071 	 */
1072 	if (xfs_physmem <= btoc(32*1024*1024)) {
1073 		/* Don't change; min configuration */
1074 		log->l_iclog_size = XLOG_RECORD_BSIZE;		/* 16k */
1075 		log->l_iclog_size_log = XLOG_RECORD_BSHIFT;
1076 	} else {
1077 		log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;	/* 32k */
1078 		log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1079 	}
1080 
1081 	/* the default log size is 16k or 32k which is one header sector */
1082 	log->l_iclog_hsize = BBSIZE;
1083 	log->l_iclog_heads = 1;
1084 
1085 	/*
1086 	 * For 16KB, we use 3 32KB buffers.  For 32KB block sizes, we use
1087 	 * 4 32KB buffers.  For 64KB block sizes, we use 8 32KB buffers.
1088 	 */
1089 	if (mp->m_sb.sb_blocksize >= 16*1024) {
1090 		log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1091 		log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1092 		if (mp->m_logbufs <= 0) {
1093 			switch (mp->m_sb.sb_blocksize) {
1094 			    case 16*1024:			/* 16 KB */
1095 				log->l_iclog_bufs = 3;
1096 				break;
1097 			    case 32*1024:			/* 32 KB */
1098 				log->l_iclog_bufs = 4;
1099 				break;
1100 			    case 64*1024:			/* 64 KB */
1101 				log->l_iclog_bufs = 8;
1102 				break;
1103 			    default:
1104 				xlog_panic("XFS: Invalid blocksize");
1105 				break;
1106 			}
1107 		}
1108 	}
1109 
1110 done:	/* are we being asked to make the sizes selected above visible? */
1111 	if (mp->m_logbufs == 0)
1112 		mp->m_logbufs = log->l_iclog_bufs;
1113 	if (mp->m_logbsize == 0)
1114 		mp->m_logbsize = log->l_iclog_size;
1115 }	/* xlog_get_iclog_buffer_size */
1116 
1117 
1118 /*
1119  * This routine initializes some of the log structure for a given mount point.
1120  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1121  * some other stuff may be filled in too.
1122  */
1123 STATIC xlog_t *
1124 xlog_alloc_log(xfs_mount_t	*mp,
1125 	       xfs_buftarg_t	*log_target,
1126 	       xfs_daddr_t	blk_offset,
1127 	       int		num_bblks)
1128 {
1129 	xlog_t			*log;
1130 	xlog_rec_header_t	*head;
1131 	xlog_in_core_t		**iclogp;
1132 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1133 	xfs_buf_t		*bp;
1134 	int			i;
1135 	int			iclogsize;
1136 
1137 	log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1138 
1139 	log->l_mp	   = mp;
1140 	log->l_targ	   = log_target;
1141 	log->l_logsize     = BBTOB(num_bblks);
1142 	log->l_logBBstart  = blk_offset;
1143 	log->l_logBBsize   = num_bblks;
1144 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1145 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1146 
1147 	log->l_prev_block  = -1;
1148 	ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, 1, 0);
1149 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1150 	log->l_last_sync_lsn = log->l_tail_lsn;
1151 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1152 	log->l_grant_reserve_cycle = 1;
1153 	log->l_grant_write_cycle = 1;
1154 
1155 	if (XFS_SB_VERSION_HASSECTOR(&mp->m_sb)) {
1156 		log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1157 		ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1158 		/* for larger sector sizes, must have v2 or external log */
1159 		ASSERT(log->l_sectbb_log == 0 ||
1160 			log->l_logBBstart == 0 ||
1161 			XFS_SB_VERSION_HASLOGV2(&mp->m_sb));
1162 		ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1163 	}
1164 	log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1165 
1166 	xlog_get_iclog_buffer_size(mp, log);
1167 
1168 	bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1169 	XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1170 	XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1171 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1172 	ASSERT(XFS_BUF_ISBUSY(bp));
1173 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1174 	log->l_xbuf = bp;
1175 
1176 	spinlock_init(&log->l_icloglock, "iclog");
1177 	spinlock_init(&log->l_grant_lock, "grhead_iclog");
1178 	initnsema(&log->l_flushsema, 0, "ic-flush");
1179 	xlog_state_ticket_alloc(log);  /* wait until after icloglock inited */
1180 
1181 	/* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1182 	ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1183 
1184 	iclogp = &log->l_iclog;
1185 	/*
1186 	 * The amount of memory to allocate for the iclog structure is
1187 	 * rather funky due to the way the structure is defined.  It is
1188 	 * done this way so that we can use different sizes for machines
1189 	 * with different amounts of memory.  See the definition of
1190 	 * xlog_in_core_t in xfs_log_priv.h for details.
1191 	 */
1192 	iclogsize = log->l_iclog_size;
1193 	ASSERT(log->l_iclog_size >= 4096);
1194 	for (i=0; i < log->l_iclog_bufs; i++) {
1195 		*iclogp = (xlog_in_core_t *)
1196 			  kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1197 		iclog = *iclogp;
1198 		iclog->hic_data = (xlog_in_core_2_t *)
1199 			  kmem_zalloc(iclogsize, KM_SLEEP);
1200 
1201 		iclog->ic_prev = prev_iclog;
1202 		prev_iclog = iclog;
1203 		log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1204 
1205 		head = &iclog->ic_header;
1206 		memset(head, 0, sizeof(xlog_rec_header_t));
1207 		INT_SET(head->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1208 		INT_SET(head->h_version, ARCH_CONVERT,
1209 			XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1210 		INT_SET(head->h_size, ARCH_CONVERT, log->l_iclog_size);
1211 		/* new fields */
1212 		INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
1213 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1214 
1215 		bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1216 		XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1217 		XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1218 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1219 		iclog->ic_bp = bp;
1220 
1221 		iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1222 		iclog->ic_state = XLOG_STATE_ACTIVE;
1223 		iclog->ic_log = log;
1224 		iclog->ic_callback_tail = &(iclog->ic_callback);
1225 		iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1226 
1227 		ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1228 		ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1229 		sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1230 		sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1231 
1232 		iclogp = &iclog->ic_next;
1233 	}
1234 	*iclogp = log->l_iclog;			/* complete ring */
1235 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1236 
1237 	return log;
1238 }	/* xlog_alloc_log */
1239 
1240 
1241 /*
1242  * Write out the commit record of a transaction associated with the given
1243  * ticket.  Return the lsn of the commit record.
1244  */
1245 STATIC int
1246 xlog_commit_record(xfs_mount_t  *mp,
1247 		   xlog_ticket_t *ticket,
1248 		   xlog_in_core_t **iclog,
1249 		   xfs_lsn_t	*commitlsnp)
1250 {
1251 	int		error;
1252 	xfs_log_iovec_t	reg[1];
1253 
1254 	reg[0].i_addr = NULL;
1255 	reg[0].i_len = 0;
1256 	XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
1257 
1258 	ASSERT_ALWAYS(iclog);
1259 	if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1260 			       iclog, XLOG_COMMIT_TRANS))) {
1261 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1262 	}
1263 	return error;
1264 }	/* xlog_commit_record */
1265 
1266 
1267 /*
1268  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1269  * log space.  This code pushes on the lsn which would supposedly free up
1270  * the 25% which we want to leave free.  We may need to adopt a policy which
1271  * pushes on an lsn which is further along in the log once we reach the high
1272  * water mark.  In this manner, we would be creating a low water mark.
1273  */
1274 void
1275 xlog_grant_push_ail(xfs_mount_t	*mp,
1276 		    int		need_bytes)
1277 {
1278     xlog_t	*log = mp->m_log;	/* pointer to the log */
1279     xfs_lsn_t	tail_lsn;		/* lsn of the log tail */
1280     xfs_lsn_t	threshold_lsn = 0;	/* lsn we'd like to be at */
1281     int		free_blocks;		/* free blocks left to write to */
1282     int		free_bytes;		/* free bytes left to write to */
1283     int		threshold_block;	/* block in lsn we'd like to be at */
1284     int		threshold_cycle;	/* lsn cycle we'd like to be at */
1285     int		free_threshold;
1286     SPLDECL(s);
1287 
1288     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1289 
1290     s = GRANT_LOCK(log);
1291     free_bytes = xlog_space_left(log,
1292 				 log->l_grant_reserve_cycle,
1293 				 log->l_grant_reserve_bytes);
1294     tail_lsn = log->l_tail_lsn;
1295     free_blocks = BTOBBT(free_bytes);
1296 
1297     /*
1298      * Set the threshold for the minimum number of free blocks in the
1299      * log to the maximum of what the caller needs, one quarter of the
1300      * log, and 256 blocks.
1301      */
1302     free_threshold = BTOBB(need_bytes);
1303     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1304     free_threshold = MAX(free_threshold, 256);
1305     if (free_blocks < free_threshold) {
1306 	threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1307 	threshold_cycle = CYCLE_LSN(tail_lsn);
1308 	if (threshold_block >= log->l_logBBsize) {
1309 	    threshold_block -= log->l_logBBsize;
1310 	    threshold_cycle += 1;
1311 	}
1312 	ASSIGN_ANY_LSN_HOST(threshold_lsn, threshold_cycle,
1313 		       threshold_block);
1314 
1315 	/* Don't pass in an lsn greater than the lsn of the last
1316 	 * log record known to be on disk.
1317 	 */
1318 	if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1319 	    threshold_lsn = log->l_last_sync_lsn;
1320     }
1321     GRANT_UNLOCK(log, s);
1322 
1323     /*
1324      * Get the transaction layer to kick the dirty buffers out to
1325      * disk asynchronously. No point in trying to do this if
1326      * the filesystem is shutting down.
1327      */
1328     if (threshold_lsn &&
1329 	!XLOG_FORCED_SHUTDOWN(log))
1330 	    xfs_trans_push_ail(mp, threshold_lsn);
1331 }	/* xlog_grant_push_ail */
1332 
1333 
1334 /*
1335  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1336  * fashion.  Previously, we should have moved the current iclog
1337  * ptr in the log to point to the next available iclog.  This allows further
1338  * write to continue while this code syncs out an iclog ready to go.
1339  * Before an in-core log can be written out, the data section must be scanned
1340  * to save away the 1st word of each BBSIZE block into the header.  We replace
1341  * it with the current cycle count.  Each BBSIZE block is tagged with the
1342  * cycle count because there in an implicit assumption that drives will
1343  * guarantee that entire 512 byte blocks get written at once.  In other words,
1344  * we can't have part of a 512 byte block written and part not written.  By
1345  * tagging each block, we will know which blocks are valid when recovering
1346  * after an unclean shutdown.
1347  *
1348  * This routine is single threaded on the iclog.  No other thread can be in
1349  * this routine with the same iclog.  Changing contents of iclog can there-
1350  * fore be done without grabbing the state machine lock.  Updating the global
1351  * log will require grabbing the lock though.
1352  *
1353  * The entire log manager uses a logical block numbering scheme.  Only
1354  * log_sync (and then only bwrite()) know about the fact that the log may
1355  * not start with block zero on a given device.  The log block start offset
1356  * is added immediately before calling bwrite().
1357  */
1358 
1359 int
1360 xlog_sync(xlog_t		*log,
1361 	  xlog_in_core_t	*iclog)
1362 {
1363 	xfs_caddr_t	dptr;		/* pointer to byte sized element */
1364 	xfs_buf_t	*bp;
1365 	int		i, ops;
1366 	uint		count;		/* byte count of bwrite */
1367 	uint		count_init;	/* initial count before roundup */
1368 	int		roundoff;       /* roundoff to BB or stripe */
1369 	int		split = 0;	/* split write into two regions */
1370 	int		error;
1371 	SPLDECL(s);
1372 	int		v2 = XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb);
1373 
1374 	XFS_STATS_INC(xs_log_writes);
1375 	ASSERT(iclog->ic_refcnt == 0);
1376 
1377 	/* Add for LR header */
1378 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1379 
1380 	/* Round out the log write size */
1381 	if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1382 		/* we have a v2 stripe unit to use */
1383 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1384 	} else {
1385 		count = BBTOB(BTOBB(count_init));
1386 	}
1387 	roundoff = count - count_init;
1388 	ASSERT(roundoff >= 0);
1389 	ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1390                 roundoff < log->l_mp->m_sb.sb_logsunit)
1391 		||
1392 		(log->l_mp->m_sb.sb_logsunit <= 1 &&
1393 		 roundoff < BBTOB(1)));
1394 
1395 	/* move grant heads by roundoff in sync */
1396 	s = GRANT_LOCK(log);
1397 	xlog_grant_add_space(log, roundoff);
1398 	GRANT_UNLOCK(log, s);
1399 
1400 	/* put cycle number in every block */
1401 	xlog_pack_data(log, iclog, roundoff);
1402 
1403 	/* real byte length */
1404 	if (v2) {
1405 		INT_SET(iclog->ic_header.h_len,
1406 			ARCH_CONVERT,
1407 			iclog->ic_offset + roundoff);
1408 	} else {
1409 		INT_SET(iclog->ic_header.h_len, ARCH_CONVERT, iclog->ic_offset);
1410 	}
1411 
1412 	/* put ops count in correct order */
1413 	ops = iclog->ic_header.h_num_logops;
1414 	INT_SET(iclog->ic_header.h_num_logops, ARCH_CONVERT, ops);
1415 
1416 	bp = iclog->ic_bp;
1417 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1418 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1419 	XFS_BUF_SET_ADDR(bp, BLOCK_LSN(INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)));
1420 
1421 	XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1422 
1423 	/* Do we need to split this write into 2 parts? */
1424 	if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1425 		split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1426 		count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1427 		iclog->ic_bwritecnt = 2;	/* split into 2 writes */
1428 	} else {
1429 		iclog->ic_bwritecnt = 1;
1430 	}
1431 	XFS_BUF_SET_PTR(bp, (xfs_caddr_t) &(iclog->ic_header), count);
1432 	XFS_BUF_SET_FSPRIVATE(bp, iclog);	/* save for later */
1433 	XFS_BUF_ZEROFLAGS(bp);
1434 	XFS_BUF_BUSY(bp);
1435 	XFS_BUF_ASYNC(bp);
1436 	/*
1437 	 * Do an ordered write for the log block.
1438 	 * Its unnecessary to flush the first split block in the log wrap case.
1439 	 */
1440 	if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1441 		XFS_BUF_ORDERED(bp);
1442 
1443 	ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1444 	ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1445 
1446 	xlog_verify_iclog(log, iclog, count, B_TRUE);
1447 
1448 	/* account for log which doesn't start at block #0 */
1449 	XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1450 	/*
1451 	 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1452 	 * is shutting down.
1453 	 */
1454 	XFS_BUF_WRITE(bp);
1455 
1456 	if ((error = XFS_bwrite(bp))) {
1457 		xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1458 				  XFS_BUF_ADDR(bp));
1459 		return error;
1460 	}
1461 	if (split) {
1462 		bp = iclog->ic_log->l_xbuf;
1463 		ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1464 							(unsigned long)1);
1465 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1466 		XFS_BUF_SET_ADDR(bp, 0);	     /* logical 0 */
1467 		XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1468 					    (__psint_t)count), split);
1469 		XFS_BUF_SET_FSPRIVATE(bp, iclog);
1470 		XFS_BUF_ZEROFLAGS(bp);
1471 		XFS_BUF_BUSY(bp);
1472 		XFS_BUF_ASYNC(bp);
1473 		if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1474 			XFS_BUF_ORDERED(bp);
1475 		dptr = XFS_BUF_PTR(bp);
1476 		/*
1477 		 * Bump the cycle numbers at the start of each block
1478 		 * since this part of the buffer is at the start of
1479 		 * a new cycle.  Watch out for the header magic number
1480 		 * case, though.
1481 		 */
1482 		for (i=0; i<split; i += BBSIZE) {
1483 			INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1484 			if (INT_GET(*(uint *)dptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
1485 				INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1486 			dptr += BBSIZE;
1487 		}
1488 
1489 		ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1490 		ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1491 
1492 		/* account for internal log which doesn't start at block #0 */
1493 		XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1494 		XFS_BUF_WRITE(bp);
1495 		if ((error = XFS_bwrite(bp))) {
1496 			xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1497 					  bp, XFS_BUF_ADDR(bp));
1498 			return error;
1499 		}
1500 	}
1501 	return 0;
1502 }	/* xlog_sync */
1503 
1504 
1505 /*
1506  * Deallocate a log structure
1507  */
1508 void
1509 xlog_dealloc_log(xlog_t *log)
1510 {
1511 	xlog_in_core_t	*iclog, *next_iclog;
1512 	xlog_ticket_t	*tic, *next_tic;
1513 	int		i;
1514 
1515 
1516 	iclog = log->l_iclog;
1517 	for (i=0; i<log->l_iclog_bufs; i++) {
1518 		sv_destroy(&iclog->ic_forcesema);
1519 		sv_destroy(&iclog->ic_writesema);
1520 		xfs_buf_free(iclog->ic_bp);
1521 #ifdef XFS_LOG_TRACE
1522 		if (iclog->ic_trace != NULL) {
1523 			ktrace_free(iclog->ic_trace);
1524 		}
1525 #endif
1526 		next_iclog = iclog->ic_next;
1527 		kmem_free(iclog->hic_data, log->l_iclog_size);
1528 		kmem_free(iclog, sizeof(xlog_in_core_t));
1529 		iclog = next_iclog;
1530 	}
1531 	freesema(&log->l_flushsema);
1532 	spinlock_destroy(&log->l_icloglock);
1533 	spinlock_destroy(&log->l_grant_lock);
1534 
1535 	/* XXXsup take a look at this again. */
1536 	if ((log->l_ticket_cnt != log->l_ticket_tcnt)  &&
1537 	    !XLOG_FORCED_SHUTDOWN(log)) {
1538 		xfs_fs_cmn_err(CE_WARN, log->l_mp,
1539 			"xlog_dealloc_log: (cnt: %d, total: %d)",
1540 			log->l_ticket_cnt, log->l_ticket_tcnt);
1541 		/* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
1542 
1543 	} else {
1544 		tic = log->l_unmount_free;
1545 		while (tic) {
1546 			next_tic = tic->t_next;
1547 			kmem_free(tic, NBPP);
1548 			tic = next_tic;
1549 		}
1550 	}
1551 	xfs_buf_free(log->l_xbuf);
1552 #ifdef XFS_LOG_TRACE
1553 	if (log->l_trace != NULL) {
1554 		ktrace_free(log->l_trace);
1555 	}
1556 	if (log->l_grant_trace != NULL) {
1557 		ktrace_free(log->l_grant_trace);
1558 	}
1559 #endif
1560 	log->l_mp->m_log = NULL;
1561 	kmem_free(log, sizeof(xlog_t));
1562 }	/* xlog_dealloc_log */
1563 
1564 /*
1565  * Update counters atomically now that memcpy is done.
1566  */
1567 /* ARGSUSED */
1568 static inline void
1569 xlog_state_finish_copy(xlog_t		*log,
1570 		       xlog_in_core_t	*iclog,
1571 		       int		record_cnt,
1572 		       int		copy_bytes)
1573 {
1574 	SPLDECL(s);
1575 
1576 	s = LOG_LOCK(log);
1577 
1578 	iclog->ic_header.h_num_logops += record_cnt;
1579 	iclog->ic_offset += copy_bytes;
1580 
1581 	LOG_UNLOCK(log, s);
1582 }	/* xlog_state_finish_copy */
1583 
1584 
1585 
1586 
1587 /*
1588  * print out info relating to regions written which consume
1589  * the reservation
1590  */
1591 STATIC void
1592 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1593 {
1594 	uint i;
1595 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1596 
1597 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
1598 	static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1599 	    "bformat",
1600 	    "bchunk",
1601 	    "efi_format",
1602 	    "efd_format",
1603 	    "iformat",
1604 	    "icore",
1605 	    "iext",
1606 	    "ibroot",
1607 	    "ilocal",
1608 	    "iattr_ext",
1609 	    "iattr_broot",
1610 	    "iattr_local",
1611 	    "qformat",
1612 	    "dquot",
1613 	    "quotaoff",
1614 	    "LR header",
1615 	    "unmount",
1616 	    "commit",
1617 	    "trans header"
1618 	};
1619 	static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1620 	    "SETATTR_NOT_SIZE",
1621 	    "SETATTR_SIZE",
1622 	    "INACTIVE",
1623 	    "CREATE",
1624 	    "CREATE_TRUNC",
1625 	    "TRUNCATE_FILE",
1626 	    "REMOVE",
1627 	    "LINK",
1628 	    "RENAME",
1629 	    "MKDIR",
1630 	    "RMDIR",
1631 	    "SYMLINK",
1632 	    "SET_DMATTRS",
1633 	    "GROWFS",
1634 	    "STRAT_WRITE",
1635 	    "DIOSTRAT",
1636 	    "WRITE_SYNC",
1637 	    "WRITEID",
1638 	    "ADDAFORK",
1639 	    "ATTRINVAL",
1640 	    "ATRUNCATE",
1641 	    "ATTR_SET",
1642 	    "ATTR_RM",
1643 	    "ATTR_FLAG",
1644 	    "CLEAR_AGI_BUCKET",
1645 	    "QM_SBCHANGE",
1646 	    "DUMMY1",
1647 	    "DUMMY2",
1648 	    "QM_QUOTAOFF",
1649 	    "QM_DQALLOC",
1650 	    "QM_SETQLIM",
1651 	    "QM_DQCLUSTER",
1652 	    "QM_QINOCREATE",
1653 	    "QM_QUOTAOFF_END",
1654 	    "SB_UNIT",
1655 	    "FSYNC_TS",
1656 	    "GROWFSRT_ALLOC",
1657 	    "GROWFSRT_ZERO",
1658 	    "GROWFSRT_FREE",
1659 	    "SWAPEXT"
1660 	};
1661 
1662 	xfs_fs_cmn_err(CE_WARN, mp,
1663 			"xfs_log_write: reservation summary:\n"
1664 			"  trans type  = %s (%u)\n"
1665 			"  unit res    = %d bytes\n"
1666 			"  current res = %d bytes\n"
1667 			"  total reg   = %u bytes (o/flow = %u bytes)\n"
1668 			"  ophdrs      = %u (ophdr space = %u bytes)\n"
1669 			"  ophdr + reg = %u bytes\n"
1670 			"  num regions = %u\n",
1671 			((ticket->t_trans_type <= 0 ||
1672 			  ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1673 			  "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1674 			ticket->t_trans_type,
1675 			ticket->t_unit_res,
1676 			ticket->t_curr_res,
1677 			ticket->t_res_arr_sum, ticket->t_res_o_flow,
1678 			ticket->t_res_num_ophdrs, ophdr_spc,
1679 			ticket->t_res_arr_sum +
1680 			ticket->t_res_o_flow + ophdr_spc,
1681 			ticket->t_res_num);
1682 
1683 	for (i = 0; i < ticket->t_res_num; i++) {
1684 		uint r_type = ticket->t_res_arr[i].r_type;
1685 		cmn_err(CE_WARN,
1686 			    "region[%u]: %s - %u bytes\n",
1687 			    i,
1688 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1689 			    "bad-rtype" : res_type_str[r_type-1]),
1690 			    ticket->t_res_arr[i].r_len);
1691 	}
1692 }
1693 
1694 /*
1695  * Write some region out to in-core log
1696  *
1697  * This will be called when writing externally provided regions or when
1698  * writing out a commit record for a given transaction.
1699  *
1700  * General algorithm:
1701  *	1. Find total length of this write.  This may include adding to the
1702  *		lengths passed in.
1703  *	2. Check whether we violate the tickets reservation.
1704  *	3. While writing to this iclog
1705  *	    A. Reserve as much space in this iclog as can get
1706  *	    B. If this is first write, save away start lsn
1707  *	    C. While writing this region:
1708  *		1. If first write of transaction, write start record
1709  *		2. Write log operation header (header per region)
1710  *		3. Find out if we can fit entire region into this iclog
1711  *		4. Potentially, verify destination memcpy ptr
1712  *		5. Memcpy (partial) region
1713  *		6. If partial copy, release iclog; otherwise, continue
1714  *			copying more regions into current iclog
1715  *	4. Mark want sync bit (in simulation mode)
1716  *	5. Release iclog for potential flush to on-disk log.
1717  *
1718  * ERRORS:
1719  * 1.	Panic if reservation is overrun.  This should never happen since
1720  *	reservation amounts are generated internal to the filesystem.
1721  * NOTES:
1722  * 1. Tickets are single threaded data structures.
1723  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1724  *	syncing routine.  When a single log_write region needs to span
1725  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1726  *	on all log operation writes which don't contain the end of the
1727  *	region.  The XLOG_END_TRANS bit is used for the in-core log
1728  *	operation which contains the end of the continued log_write region.
1729  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1730  *	we don't really know exactly how much space will be used.  As a result,
1731  *	we don't update ic_offset until the end when we know exactly how many
1732  *	bytes have been written out.
1733  */
1734 int
1735 xlog_write(xfs_mount_t *	mp,
1736 	   xfs_log_iovec_t	reg[],
1737 	   int			nentries,
1738 	   xfs_log_ticket_t	tic,
1739 	   xfs_lsn_t		*start_lsn,
1740 	   xlog_in_core_t	**commit_iclog,
1741 	   uint			flags)
1742 {
1743     xlog_t	     *log = mp->m_log;
1744     xlog_ticket_t    *ticket = (xlog_ticket_t *)tic;
1745     xlog_in_core_t   *iclog = NULL;  /* ptr to current in-core log */
1746     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1747     __psint_t	     ptr;	     /* copy address into data region */
1748     int		     len;	     /* # xlog_write() bytes 2 still copy */
1749     int		     index;	     /* region index currently copying */
1750     int		     log_offset;     /* offset (from 0) into data region */
1751     int		     start_rec_copy; /* # bytes to copy for start record */
1752     int		     partial_copy;   /* did we split a region? */
1753     int		     partial_copy_len;/* # bytes copied if split region */
1754     int		     need_copy;	     /* # bytes need to memcpy this region */
1755     int		     copy_len;	     /* # bytes actually memcpy'ing */
1756     int		     copy_off;	     /* # bytes from entry start */
1757     int		     contwr;	     /* continued write of in-core log? */
1758     int		     error;
1759     int		     record_cnt = 0, data_cnt = 0;
1760 
1761     partial_copy_len = partial_copy = 0;
1762 
1763     /* Calculate potential maximum space.  Each region gets its own
1764      * xlog_op_header_t and may need to be double word aligned.
1765      */
1766     len = 0;
1767     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1768 	len += sizeof(xlog_op_header_t);
1769 	XLOG_TIC_ADD_OPHDR(ticket);
1770     }
1771 
1772     for (index = 0; index < nentries; index++) {
1773 	len += sizeof(xlog_op_header_t);	    /* each region gets >= 1 */
1774 	XLOG_TIC_ADD_OPHDR(ticket);
1775 	len += reg[index].i_len;
1776 	XLOG_TIC_ADD_REGION(ticket, reg[index].i_len, reg[index].i_type);
1777     }
1778     contwr = *start_lsn = 0;
1779 
1780     if (ticket->t_curr_res < len) {
1781 	xlog_print_tic_res(mp, ticket);
1782 #ifdef DEBUG
1783 	xlog_panic(
1784 		"xfs_log_write: reservation ran out. Need to up reservation");
1785 #else
1786 	/* Customer configurable panic */
1787 	xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1788 		"xfs_log_write: reservation ran out. Need to up reservation");
1789 	/* If we did not panic, shutdown the filesystem */
1790 	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1791 #endif
1792     } else
1793 	ticket->t_curr_res -= len;
1794 
1795     for (index = 0; index < nentries; ) {
1796 	if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1797 					       &contwr, &log_offset)))
1798 		return error;
1799 
1800 	ASSERT(log_offset <= iclog->ic_size - 1);
1801 	ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1802 
1803 	/* start_lsn is the first lsn written to. That's all we need. */
1804 	if (! *start_lsn)
1805 	    *start_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
1806 
1807 	/* This loop writes out as many regions as can fit in the amount
1808 	 * of space which was allocated by xlog_state_get_iclog_space().
1809 	 */
1810 	while (index < nentries) {
1811 	    ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1812 	    ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1813 	    start_rec_copy = 0;
1814 
1815 	    /* If first write for transaction, insert start record.
1816 	     * We can't be trying to commit if we are inited.  We can't
1817 	     * have any "partial_copy" if we are inited.
1818 	     */
1819 	    if (ticket->t_flags & XLOG_TIC_INITED) {
1820 		logop_head		= (xlog_op_header_t *)ptr;
1821 		INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1822 		logop_head->oh_clientid = ticket->t_clientid;
1823 		logop_head->oh_len	= 0;
1824 		logop_head->oh_flags    = XLOG_START_TRANS;
1825 		logop_head->oh_res2	= 0;
1826 		ticket->t_flags		&= ~XLOG_TIC_INITED;	/* clear bit */
1827 		record_cnt++;
1828 
1829 		start_rec_copy = sizeof(xlog_op_header_t);
1830 		xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1831 	    }
1832 
1833 	    /* Copy log operation header directly into data section */
1834 	    logop_head			= (xlog_op_header_t *)ptr;
1835 	    INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1836 	    logop_head->oh_clientid	= ticket->t_clientid;
1837 	    logop_head->oh_res2		= 0;
1838 
1839 	    /* header copied directly */
1840 	    xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1841 
1842 	    /* are we copying a commit or unmount record? */
1843 	    logop_head->oh_flags = flags;
1844 
1845 	    /*
1846 	     * We've seen logs corrupted with bad transaction client
1847 	     * ids.  This makes sure that XFS doesn't generate them on.
1848 	     * Turn this into an EIO and shut down the filesystem.
1849 	     */
1850 	    switch (logop_head->oh_clientid)  {
1851 	    case XFS_TRANSACTION:
1852 	    case XFS_VOLUME:
1853 	    case XFS_LOG:
1854 		break;
1855 	    default:
1856 		xfs_fs_cmn_err(CE_WARN, mp,
1857 		    "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1858 		    logop_head->oh_clientid, tic);
1859 		return XFS_ERROR(EIO);
1860 	    }
1861 
1862 	    /* Partial write last time? => (partial_copy != 0)
1863 	     * need_copy is the amount we'd like to copy if everything could
1864 	     * fit in the current memcpy.
1865 	     */
1866 	    need_copy =	reg[index].i_len - partial_copy_len;
1867 
1868 	    copy_off = partial_copy_len;
1869 	    if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1870 		INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len = need_copy);
1871 		if (partial_copy)
1872 		    logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1873 		partial_copy_len = partial_copy = 0;
1874 	    } else {					    /* partial write */
1875 		copy_len = iclog->ic_size - log_offset;
1876 		INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len);
1877 		logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1878 		if (partial_copy)
1879 			logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1880 		partial_copy_len += copy_len;
1881 		partial_copy++;
1882 		len += sizeof(xlog_op_header_t); /* from splitting of region */
1883 		/* account for new log op header */
1884 		ticket->t_curr_res -= sizeof(xlog_op_header_t);
1885 		XLOG_TIC_ADD_OPHDR(ticket);
1886 	    }
1887 	    xlog_verify_dest_ptr(log, ptr);
1888 
1889 	    /* copy region */
1890 	    ASSERT(copy_len >= 0);
1891 	    memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1892 	    xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1893 
1894 	    /* make copy_len total bytes copied, including headers */
1895 	    copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1896 	    record_cnt++;
1897 	    data_cnt += contwr ? copy_len : 0;
1898 	    if (partial_copy) {			/* copied partial region */
1899 		    /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1900 		    xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1901 		    record_cnt = data_cnt = 0;
1902 		    if ((error = xlog_state_release_iclog(log, iclog)))
1903 			    return error;
1904 		    break;			/* don't increment index */
1905 	    } else {				/* copied entire region */
1906 		index++;
1907 		partial_copy_len = partial_copy = 0;
1908 
1909 		if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1910 		    xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1911 		    record_cnt = data_cnt = 0;
1912 		    xlog_state_want_sync(log, iclog);
1913 		    if (commit_iclog) {
1914 			ASSERT(flags & XLOG_COMMIT_TRANS);
1915 			*commit_iclog = iclog;
1916 		    } else if ((error = xlog_state_release_iclog(log, iclog)))
1917 			   return error;
1918 		    if (index == nentries)
1919 			    return 0;		/* we are done */
1920 		    else
1921 			    break;
1922 		}
1923 	    } /* if (partial_copy) */
1924 	} /* while (index < nentries) */
1925     } /* for (index = 0; index < nentries; ) */
1926     ASSERT(len == 0);
1927 
1928     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1929     if (commit_iclog) {
1930 	ASSERT(flags & XLOG_COMMIT_TRANS);
1931 	*commit_iclog = iclog;
1932 	return 0;
1933     }
1934     return xlog_state_release_iclog(log, iclog);
1935 }	/* xlog_write */
1936 
1937 
1938 /*****************************************************************************
1939  *
1940  *		State Machine functions
1941  *
1942  *****************************************************************************
1943  */
1944 
1945 /* Clean iclogs starting from the head.  This ordering must be
1946  * maintained, so an iclog doesn't become ACTIVE beyond one that
1947  * is SYNCING.  This is also required to maintain the notion that we use
1948  * a counting semaphore to hold off would be writers to the log when every
1949  * iclog is trying to sync to disk.
1950  *
1951  * State Change: DIRTY -> ACTIVE
1952  */
1953 STATIC void
1954 xlog_state_clean_log(xlog_t *log)
1955 {
1956 	xlog_in_core_t	*iclog;
1957 	int changed = 0;
1958 
1959 	iclog = log->l_iclog;
1960 	do {
1961 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
1962 			iclog->ic_state	= XLOG_STATE_ACTIVE;
1963 			iclog->ic_offset       = 0;
1964 			iclog->ic_callback	= NULL;   /* don't need to free */
1965 			/*
1966 			 * If the number of ops in this iclog indicate it just
1967 			 * contains the dummy transaction, we can
1968 			 * change state into IDLE (the second time around).
1969 			 * Otherwise we should change the state into
1970 			 * NEED a dummy.
1971 			 * We don't need to cover the dummy.
1972 			 */
1973 			if (!changed &&
1974 			   (INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT) == XLOG_COVER_OPS)) {
1975 				changed = 1;
1976 			} else {
1977 				/*
1978 				 * We have two dirty iclogs so start over
1979 				 * This could also be num of ops indicates
1980 				 * this is not the dummy going out.
1981 				 */
1982 				changed = 2;
1983 			}
1984 			iclog->ic_header.h_num_logops = 0;
1985 			memset(iclog->ic_header.h_cycle_data, 0,
1986 			      sizeof(iclog->ic_header.h_cycle_data));
1987 			iclog->ic_header.h_lsn = 0;
1988 		} else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1989 			/* do nothing */;
1990 		else
1991 			break;	/* stop cleaning */
1992 		iclog = iclog->ic_next;
1993 	} while (iclog != log->l_iclog);
1994 
1995 	/* log is locked when we are called */
1996 	/*
1997 	 * Change state for the dummy log recording.
1998 	 * We usually go to NEED. But we go to NEED2 if the changed indicates
1999 	 * we are done writing the dummy record.
2000 	 * If we are done with the second dummy recored (DONE2), then
2001 	 * we go to IDLE.
2002 	 */
2003 	if (changed) {
2004 		switch (log->l_covered_state) {
2005 		case XLOG_STATE_COVER_IDLE:
2006 		case XLOG_STATE_COVER_NEED:
2007 		case XLOG_STATE_COVER_NEED2:
2008 			log->l_covered_state = XLOG_STATE_COVER_NEED;
2009 			break;
2010 
2011 		case XLOG_STATE_COVER_DONE:
2012 			if (changed == 1)
2013 				log->l_covered_state = XLOG_STATE_COVER_NEED2;
2014 			else
2015 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2016 			break;
2017 
2018 		case XLOG_STATE_COVER_DONE2:
2019 			if (changed == 1)
2020 				log->l_covered_state = XLOG_STATE_COVER_IDLE;
2021 			else
2022 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2023 			break;
2024 
2025 		default:
2026 			ASSERT(0);
2027 		}
2028 	}
2029 }	/* xlog_state_clean_log */
2030 
2031 STATIC xfs_lsn_t
2032 xlog_get_lowest_lsn(
2033 	xlog_t		*log)
2034 {
2035 	xlog_in_core_t  *lsn_log;
2036 	xfs_lsn_t	lowest_lsn, lsn;
2037 
2038 	lsn_log = log->l_iclog;
2039 	lowest_lsn = 0;
2040 	do {
2041 	    if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2042 		lsn = INT_GET(lsn_log->ic_header.h_lsn, ARCH_CONVERT);
2043 		if ((lsn && !lowest_lsn) ||
2044 		    (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2045 			lowest_lsn = lsn;
2046 		}
2047 	    }
2048 	    lsn_log = lsn_log->ic_next;
2049 	} while (lsn_log != log->l_iclog);
2050 	return lowest_lsn;
2051 }
2052 
2053 
2054 STATIC void
2055 xlog_state_do_callback(
2056 	xlog_t		*log,
2057 	int		aborted,
2058 	xlog_in_core_t	*ciclog)
2059 {
2060 	xlog_in_core_t	   *iclog;
2061 	xlog_in_core_t	   *first_iclog;	/* used to know when we've
2062 						 * processed all iclogs once */
2063 	xfs_log_callback_t *cb, *cb_next;
2064 	int		   flushcnt = 0;
2065 	xfs_lsn_t	   lowest_lsn;
2066 	int		   ioerrors;	/* counter: iclogs with errors */
2067 	int		   loopdidcallbacks; /* flag: inner loop did callbacks*/
2068 	int		   funcdidcallbacks; /* flag: function did callbacks */
2069 	int		   repeats;	/* for issuing console warnings if
2070 					 * looping too many times */
2071 	SPLDECL(s);
2072 
2073 	s = LOG_LOCK(log);
2074 	first_iclog = iclog = log->l_iclog;
2075 	ioerrors = 0;
2076 	funcdidcallbacks = 0;
2077 	repeats = 0;
2078 
2079 	do {
2080 		/*
2081 		 * Scan all iclogs starting with the one pointed to by the
2082 		 * log.  Reset this starting point each time the log is
2083 		 * unlocked (during callbacks).
2084 		 *
2085 		 * Keep looping through iclogs until one full pass is made
2086 		 * without running any callbacks.
2087 		 */
2088 		first_iclog = log->l_iclog;
2089 		iclog = log->l_iclog;
2090 		loopdidcallbacks = 0;
2091 		repeats++;
2092 
2093 		do {
2094 
2095 			/* skip all iclogs in the ACTIVE & DIRTY states */
2096 			if (iclog->ic_state &
2097 			    (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2098 				iclog = iclog->ic_next;
2099 				continue;
2100 			}
2101 
2102 			/*
2103 			 * Between marking a filesystem SHUTDOWN and stopping
2104 			 * the log, we do flush all iclogs to disk (if there
2105 			 * wasn't a log I/O error). So, we do want things to
2106 			 * go smoothly in case of just a SHUTDOWN  w/o a
2107 			 * LOG_IO_ERROR.
2108 			 */
2109 			if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2110 				/*
2111 				 * Can only perform callbacks in order.  Since
2112 				 * this iclog is not in the DONE_SYNC/
2113 				 * DO_CALLBACK state, we skip the rest and
2114 				 * just try to clean up.  If we set our iclog
2115 				 * to DO_CALLBACK, we will not process it when
2116 				 * we retry since a previous iclog is in the
2117 				 * CALLBACK and the state cannot change since
2118 				 * we are holding the LOG_LOCK.
2119 				 */
2120 				if (!(iclog->ic_state &
2121 					(XLOG_STATE_DONE_SYNC |
2122 						 XLOG_STATE_DO_CALLBACK))) {
2123 					if (ciclog && (ciclog->ic_state ==
2124 							XLOG_STATE_DONE_SYNC)) {
2125 						ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2126 					}
2127 					break;
2128 				}
2129 				/*
2130 				 * We now have an iclog that is in either the
2131 				 * DO_CALLBACK or DONE_SYNC states. The other
2132 				 * states (WANT_SYNC, SYNCING, or CALLBACK were
2133 				 * caught by the above if and are going to
2134 				 * clean (i.e. we aren't doing their callbacks)
2135 				 * see the above if.
2136 				 */
2137 
2138 				/*
2139 				 * We will do one more check here to see if we
2140 				 * have chased our tail around.
2141 				 */
2142 
2143 				lowest_lsn = xlog_get_lowest_lsn(log);
2144 				if (lowest_lsn && (
2145 					XFS_LSN_CMP(
2146 						lowest_lsn,
2147 						INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2148 					)<0)) {
2149 					iclog = iclog->ic_next;
2150 					continue; /* Leave this iclog for
2151 						   * another thread */
2152 				}
2153 
2154 				iclog->ic_state = XLOG_STATE_CALLBACK;
2155 
2156 				LOG_UNLOCK(log, s);
2157 
2158 				/* l_last_sync_lsn field protected by
2159 				 * GRANT_LOCK. Don't worry about iclog's lsn.
2160 				 * No one else can be here except us.
2161 				 */
2162 				s = GRANT_LOCK(log);
2163 				ASSERT(XFS_LSN_CMP(
2164 						log->l_last_sync_lsn,
2165 						INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2166 					)<=0);
2167 				log->l_last_sync_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2168 				GRANT_UNLOCK(log, s);
2169 
2170 				/*
2171 				 * Keep processing entries in the callback list
2172 				 * until we come around and it is empty.  We
2173 				 * need to atomically see that the list is
2174 				 * empty and change the state to DIRTY so that
2175 				 * we don't miss any more callbacks being added.
2176 				 */
2177 				s = LOG_LOCK(log);
2178 			} else {
2179 				ioerrors++;
2180 			}
2181 			cb = iclog->ic_callback;
2182 
2183 			while (cb != 0) {
2184 				iclog->ic_callback_tail = &(iclog->ic_callback);
2185 				iclog->ic_callback = NULL;
2186 				LOG_UNLOCK(log, s);
2187 
2188 				/* perform callbacks in the order given */
2189 				for (; cb != 0; cb = cb_next) {
2190 					cb_next = cb->cb_next;
2191 					cb->cb_func(cb->cb_arg, aborted);
2192 				}
2193 				s = LOG_LOCK(log);
2194 				cb = iclog->ic_callback;
2195 			}
2196 
2197 			loopdidcallbacks++;
2198 			funcdidcallbacks++;
2199 
2200 			ASSERT(iclog->ic_callback == 0);
2201 			if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2202 				iclog->ic_state = XLOG_STATE_DIRTY;
2203 
2204 			/*
2205 			 * Transition from DIRTY to ACTIVE if applicable.
2206 			 * NOP if STATE_IOERROR.
2207 			 */
2208 			xlog_state_clean_log(log);
2209 
2210 			/* wake up threads waiting in xfs_log_force() */
2211 			sv_broadcast(&iclog->ic_forcesema);
2212 
2213 			iclog = iclog->ic_next;
2214 		} while (first_iclog != iclog);
2215 		if (repeats && (repeats % 10) == 0) {
2216 			xfs_fs_cmn_err(CE_WARN, log->l_mp,
2217 				"xlog_state_do_callback: looping %d", repeats);
2218 		}
2219 	} while (!ioerrors && loopdidcallbacks);
2220 
2221 	/*
2222 	 * make one last gasp attempt to see if iclogs are being left in
2223 	 * limbo..
2224 	 */
2225 #ifdef DEBUG
2226 	if (funcdidcallbacks) {
2227 		first_iclog = iclog = log->l_iclog;
2228 		do {
2229 			ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2230 			/*
2231 			 * Terminate the loop if iclogs are found in states
2232 			 * which will cause other threads to clean up iclogs.
2233 			 *
2234 			 * SYNCING - i/o completion will go through logs
2235 			 * DONE_SYNC - interrupt thread should be waiting for
2236 			 *              LOG_LOCK
2237 			 * IOERROR - give up hope all ye who enter here
2238 			 */
2239 			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2240 			    iclog->ic_state == XLOG_STATE_SYNCING ||
2241 			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2242 			    iclog->ic_state == XLOG_STATE_IOERROR )
2243 				break;
2244 			iclog = iclog->ic_next;
2245 		} while (first_iclog != iclog);
2246 	}
2247 #endif
2248 
2249 	if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2250 		flushcnt = log->l_flushcnt;
2251 		log->l_flushcnt = 0;
2252 	}
2253 	LOG_UNLOCK(log, s);
2254 	while (flushcnt--)
2255 		vsema(&log->l_flushsema);
2256 }	/* xlog_state_do_callback */
2257 
2258 
2259 /*
2260  * Finish transitioning this iclog to the dirty state.
2261  *
2262  * Make sure that we completely execute this routine only when this is
2263  * the last call to the iclog.  There is a good chance that iclog flushes,
2264  * when we reach the end of the physical log, get turned into 2 separate
2265  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2266  * routine.  By using the reference count bwritecnt, we guarantee that only
2267  * the second completion goes through.
2268  *
2269  * Callbacks could take time, so they are done outside the scope of the
2270  * global state machine log lock.  Assume that the calls to cvsema won't
2271  * take a long time.  At least we know it won't sleep.
2272  */
2273 void
2274 xlog_state_done_syncing(
2275 	xlog_in_core_t	*iclog,
2276 	int		aborted)
2277 {
2278 	xlog_t		   *log = iclog->ic_log;
2279 	SPLDECL(s);
2280 
2281 	s = LOG_LOCK(log);
2282 
2283 	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2284 	       iclog->ic_state == XLOG_STATE_IOERROR);
2285 	ASSERT(iclog->ic_refcnt == 0);
2286 	ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2287 
2288 
2289 	/*
2290 	 * If we got an error, either on the first buffer, or in the case of
2291 	 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2292 	 * and none should ever be attempted to be written to disk
2293 	 * again.
2294 	 */
2295 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
2296 		if (--iclog->ic_bwritecnt == 1) {
2297 			LOG_UNLOCK(log, s);
2298 			return;
2299 		}
2300 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2301 	}
2302 
2303 	/*
2304 	 * Someone could be sleeping prior to writing out the next
2305 	 * iclog buffer, we wake them all, one will get to do the
2306 	 * I/O, the others get to wait for the result.
2307 	 */
2308 	sv_broadcast(&iclog->ic_writesema);
2309 	LOG_UNLOCK(log, s);
2310 	xlog_state_do_callback(log, aborted, iclog);	/* also cleans log */
2311 }	/* xlog_state_done_syncing */
2312 
2313 
2314 /*
2315  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2316  * sleep.  The flush semaphore is set to the number of in-core buffers and
2317  * decremented around disk syncing.  Therefore, if all buffers are syncing,
2318  * this semaphore will cause new writes to sleep until a sync completes.
2319  * Otherwise, this code just does p() followed by v().  This approximates
2320  * a sleep/wakeup except we can't race.
2321  *
2322  * The in-core logs are used in a circular fashion. They are not used
2323  * out-of-order even when an iclog past the head is free.
2324  *
2325  * return:
2326  *	* log_offset where xlog_write() can start writing into the in-core
2327  *		log's data space.
2328  *	* in-core log pointer to which xlog_write() should write.
2329  *	* boolean indicating this is a continued write to an in-core log.
2330  *		If this is the last write, then the in-core log's offset field
2331  *		needs to be incremented, depending on the amount of data which
2332  *		is copied.
2333  */
2334 int
2335 xlog_state_get_iclog_space(xlog_t	  *log,
2336 			   int		  len,
2337 			   xlog_in_core_t **iclogp,
2338 			   xlog_ticket_t  *ticket,
2339 			   int		  *continued_write,
2340 			   int		  *logoffsetp)
2341 {
2342 	SPLDECL(s);
2343 	int		  log_offset;
2344 	xlog_rec_header_t *head;
2345 	xlog_in_core_t	  *iclog;
2346 	int		  error;
2347 
2348 restart:
2349 	s = LOG_LOCK(log);
2350 	if (XLOG_FORCED_SHUTDOWN(log)) {
2351 		LOG_UNLOCK(log, s);
2352 		return XFS_ERROR(EIO);
2353 	}
2354 
2355 	iclog = log->l_iclog;
2356 	if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2357 		log->l_flushcnt++;
2358 		LOG_UNLOCK(log, s);
2359 		xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2360 		XFS_STATS_INC(xs_log_noiclogs);
2361 		/* Ensure that log writes happen */
2362 		psema(&log->l_flushsema, PINOD);
2363 		goto restart;
2364 	}
2365 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2366 	head = &iclog->ic_header;
2367 
2368 	iclog->ic_refcnt++;			/* prevents sync */
2369 	log_offset = iclog->ic_offset;
2370 
2371 	/* On the 1st write to an iclog, figure out lsn.  This works
2372 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2373 	 * committing to.  If the offset is set, that's how many blocks
2374 	 * must be written.
2375 	 */
2376 	if (log_offset == 0) {
2377 		ticket->t_curr_res -= log->l_iclog_hsize;
2378 		XLOG_TIC_ADD_REGION(ticket,
2379 				    log->l_iclog_hsize,
2380 				    XLOG_REG_TYPE_LRHEADER);
2381 		INT_SET(head->h_cycle, ARCH_CONVERT, log->l_curr_cycle);
2382 		ASSIGN_LSN(head->h_lsn, log);
2383 		ASSERT(log->l_curr_block >= 0);
2384 	}
2385 
2386 	/* If there is enough room to write everything, then do it.  Otherwise,
2387 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2388 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2389 	 * until you know exactly how many bytes get copied.  Therefore, wait
2390 	 * until later to update ic_offset.
2391 	 *
2392 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2393 	 * can fit into remaining data section.
2394 	 */
2395 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2396 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2397 
2398 		/* If I'm the only one writing to this iclog, sync it to disk */
2399 		if (iclog->ic_refcnt == 1) {
2400 			LOG_UNLOCK(log, s);
2401 			if ((error = xlog_state_release_iclog(log, iclog)))
2402 				return error;
2403 		} else {
2404 			iclog->ic_refcnt--;
2405 			LOG_UNLOCK(log, s);
2406 		}
2407 		goto restart;
2408 	}
2409 
2410 	/* Do we have enough room to write the full amount in the remainder
2411 	 * of this iclog?  Or must we continue a write on the next iclog and
2412 	 * mark this iclog as completely taken?  In the case where we switch
2413 	 * iclogs (to mark it taken), this particular iclog will release/sync
2414 	 * to disk in xlog_write().
2415 	 */
2416 	if (len <= iclog->ic_size - iclog->ic_offset) {
2417 		*continued_write = 0;
2418 		iclog->ic_offset += len;
2419 	} else {
2420 		*continued_write = 1;
2421 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2422 	}
2423 	*iclogp = iclog;
2424 
2425 	ASSERT(iclog->ic_offset <= iclog->ic_size);
2426 	LOG_UNLOCK(log, s);
2427 
2428 	*logoffsetp = log_offset;
2429 	return 0;
2430 }	/* xlog_state_get_iclog_space */
2431 
2432 /*
2433  * Atomically get the log space required for a log ticket.
2434  *
2435  * Once a ticket gets put onto the reserveq, it will only return after
2436  * the needed reservation is satisfied.
2437  */
2438 STATIC int
2439 xlog_grant_log_space(xlog_t	   *log,
2440 		     xlog_ticket_t *tic)
2441 {
2442 	int		 free_bytes;
2443 	int		 need_bytes;
2444 	SPLDECL(s);
2445 #ifdef DEBUG
2446 	xfs_lsn_t	 tail_lsn;
2447 #endif
2448 
2449 
2450 #ifdef DEBUG
2451 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2452 		panic("grant Recovery problem");
2453 #endif
2454 
2455 	/* Is there space or do we need to sleep? */
2456 	s = GRANT_LOCK(log);
2457 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2458 
2459 	/* something is already sleeping; insert new transaction at end */
2460 	if (log->l_reserve_headq) {
2461 		xlog_ins_ticketq(&log->l_reserve_headq, tic);
2462 		xlog_trace_loggrant(log, tic,
2463 				    "xlog_grant_log_space: sleep 1");
2464 		/*
2465 		 * Gotta check this before going to sleep, while we're
2466 		 * holding the grant lock.
2467 		 */
2468 		if (XLOG_FORCED_SHUTDOWN(log))
2469 			goto error_return;
2470 
2471 		XFS_STATS_INC(xs_sleep_logspace);
2472 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2473 		/*
2474 		 * If we got an error, and the filesystem is shutting down,
2475 		 * we'll catch it down below. So just continue...
2476 		 */
2477 		xlog_trace_loggrant(log, tic,
2478 				    "xlog_grant_log_space: wake 1");
2479 		s = GRANT_LOCK(log);
2480 	}
2481 	if (tic->t_flags & XFS_LOG_PERM_RESERV)
2482 		need_bytes = tic->t_unit_res*tic->t_ocnt;
2483 	else
2484 		need_bytes = tic->t_unit_res;
2485 
2486 redo:
2487 	if (XLOG_FORCED_SHUTDOWN(log))
2488 		goto error_return;
2489 
2490 	free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2491 				     log->l_grant_reserve_bytes);
2492 	if (free_bytes < need_bytes) {
2493 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2494 			xlog_ins_ticketq(&log->l_reserve_headq, tic);
2495 		xlog_trace_loggrant(log, tic,
2496 				    "xlog_grant_log_space: sleep 2");
2497 		XFS_STATS_INC(xs_sleep_logspace);
2498 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2499 
2500 		if (XLOG_FORCED_SHUTDOWN(log)) {
2501 			s = GRANT_LOCK(log);
2502 			goto error_return;
2503 		}
2504 
2505 		xlog_trace_loggrant(log, tic,
2506 				    "xlog_grant_log_space: wake 2");
2507 		xlog_grant_push_ail(log->l_mp, need_bytes);
2508 		s = GRANT_LOCK(log);
2509 		goto redo;
2510 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2511 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2512 
2513 	/* we've got enough space */
2514 	xlog_grant_add_space(log, need_bytes);
2515 #ifdef DEBUG
2516 	tail_lsn = log->l_tail_lsn;
2517 	/*
2518 	 * Check to make sure the grant write head didn't just over lap the
2519 	 * tail.  If the cycles are the same, we can't be overlapping.
2520 	 * Otherwise, make sure that the cycles differ by exactly one and
2521 	 * check the byte count.
2522 	 */
2523 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2524 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2525 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2526 	}
2527 #endif
2528 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2529 	xlog_verify_grant_head(log, 1);
2530 	GRANT_UNLOCK(log, s);
2531 	return 0;
2532 
2533  error_return:
2534 	if (tic->t_flags & XLOG_TIC_IN_Q)
2535 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2536 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2537 	/*
2538 	 * If we are failing, make sure the ticket doesn't have any
2539 	 * current reservations. We don't want to add this back when
2540 	 * the ticket/transaction gets cancelled.
2541 	 */
2542 	tic->t_curr_res = 0;
2543 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2544 	GRANT_UNLOCK(log, s);
2545 	return XFS_ERROR(EIO);
2546 }	/* xlog_grant_log_space */
2547 
2548 
2549 /*
2550  * Replenish the byte reservation required by moving the grant write head.
2551  *
2552  *
2553  */
2554 STATIC int
2555 xlog_regrant_write_log_space(xlog_t	   *log,
2556 			     xlog_ticket_t *tic)
2557 {
2558 	SPLDECL(s);
2559 	int		free_bytes, need_bytes;
2560 	xlog_ticket_t	*ntic;
2561 #ifdef DEBUG
2562 	xfs_lsn_t	tail_lsn;
2563 #endif
2564 
2565 	tic->t_curr_res = tic->t_unit_res;
2566 	XLOG_TIC_RESET_RES(tic);
2567 
2568 	if (tic->t_cnt > 0)
2569 		return 0;
2570 
2571 #ifdef DEBUG
2572 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2573 		panic("regrant Recovery problem");
2574 #endif
2575 
2576 	s = GRANT_LOCK(log);
2577 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2578 
2579 	if (XLOG_FORCED_SHUTDOWN(log))
2580 		goto error_return;
2581 
2582 	/* If there are other waiters on the queue then give them a
2583 	 * chance at logspace before us. Wake up the first waiters,
2584 	 * if we do not wake up all the waiters then go to sleep waiting
2585 	 * for more free space, otherwise try to get some space for
2586 	 * this transaction.
2587 	 */
2588 
2589 	if ((ntic = log->l_write_headq)) {
2590 		free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2591 					     log->l_grant_write_bytes);
2592 		do {
2593 			ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2594 
2595 			if (free_bytes < ntic->t_unit_res)
2596 				break;
2597 			free_bytes -= ntic->t_unit_res;
2598 			sv_signal(&ntic->t_sema);
2599 			ntic = ntic->t_next;
2600 		} while (ntic != log->l_write_headq);
2601 
2602 		if (ntic != log->l_write_headq) {
2603 			if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2604 				xlog_ins_ticketq(&log->l_write_headq, tic);
2605 
2606 			xlog_trace_loggrant(log, tic,
2607 				    "xlog_regrant_write_log_space: sleep 1");
2608 			XFS_STATS_INC(xs_sleep_logspace);
2609 			sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2610 				&log->l_grant_lock, s);
2611 
2612 			/* If we're shutting down, this tic is already
2613 			 * off the queue */
2614 			if (XLOG_FORCED_SHUTDOWN(log)) {
2615 				s = GRANT_LOCK(log);
2616 				goto error_return;
2617 			}
2618 
2619 			xlog_trace_loggrant(log, tic,
2620 				    "xlog_regrant_write_log_space: wake 1");
2621 			xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2622 			s = GRANT_LOCK(log);
2623 		}
2624 	}
2625 
2626 	need_bytes = tic->t_unit_res;
2627 
2628 redo:
2629 	if (XLOG_FORCED_SHUTDOWN(log))
2630 		goto error_return;
2631 
2632 	free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2633 				     log->l_grant_write_bytes);
2634 	if (free_bytes < need_bytes) {
2635 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2636 			xlog_ins_ticketq(&log->l_write_headq, tic);
2637 		XFS_STATS_INC(xs_sleep_logspace);
2638 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2639 
2640 		/* If we're shutting down, this tic is already off the queue */
2641 		if (XLOG_FORCED_SHUTDOWN(log)) {
2642 			s = GRANT_LOCK(log);
2643 			goto error_return;
2644 		}
2645 
2646 		xlog_trace_loggrant(log, tic,
2647 				    "xlog_regrant_write_log_space: wake 2");
2648 		xlog_grant_push_ail(log->l_mp, need_bytes);
2649 		s = GRANT_LOCK(log);
2650 		goto redo;
2651 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2652 		xlog_del_ticketq(&log->l_write_headq, tic);
2653 
2654 	/* we've got enough space */
2655 	xlog_grant_add_space_write(log, need_bytes);
2656 #ifdef DEBUG
2657 	tail_lsn = log->l_tail_lsn;
2658 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2659 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2660 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2661 	}
2662 #endif
2663 
2664 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2665 	xlog_verify_grant_head(log, 1);
2666 	GRANT_UNLOCK(log, s);
2667 	return 0;
2668 
2669 
2670  error_return:
2671 	if (tic->t_flags & XLOG_TIC_IN_Q)
2672 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2673 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2674 	/*
2675 	 * If we are failing, make sure the ticket doesn't have any
2676 	 * current reservations. We don't want to add this back when
2677 	 * the ticket/transaction gets cancelled.
2678 	 */
2679 	tic->t_curr_res = 0;
2680 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2681 	GRANT_UNLOCK(log, s);
2682 	return XFS_ERROR(EIO);
2683 }	/* xlog_regrant_write_log_space */
2684 
2685 
2686 /* The first cnt-1 times through here we don't need to
2687  * move the grant write head because the permanent
2688  * reservation has reserved cnt times the unit amount.
2689  * Release part of current permanent unit reservation and
2690  * reset current reservation to be one units worth.  Also
2691  * move grant reservation head forward.
2692  */
2693 STATIC void
2694 xlog_regrant_reserve_log_space(xlog_t	     *log,
2695 			       xlog_ticket_t *ticket)
2696 {
2697 	SPLDECL(s);
2698 
2699 	xlog_trace_loggrant(log, ticket,
2700 			    "xlog_regrant_reserve_log_space: enter");
2701 	if (ticket->t_cnt > 0)
2702 		ticket->t_cnt--;
2703 
2704 	s = GRANT_LOCK(log);
2705 	xlog_grant_sub_space(log, ticket->t_curr_res);
2706 	ticket->t_curr_res = ticket->t_unit_res;
2707 	XLOG_TIC_RESET_RES(ticket);
2708 	xlog_trace_loggrant(log, ticket,
2709 			    "xlog_regrant_reserve_log_space: sub current res");
2710 	xlog_verify_grant_head(log, 1);
2711 
2712 	/* just return if we still have some of the pre-reserved space */
2713 	if (ticket->t_cnt > 0) {
2714 		GRANT_UNLOCK(log, s);
2715 		return;
2716 	}
2717 
2718 	xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2719 	xlog_trace_loggrant(log, ticket,
2720 			    "xlog_regrant_reserve_log_space: exit");
2721 	xlog_verify_grant_head(log, 0);
2722 	GRANT_UNLOCK(log, s);
2723 	ticket->t_curr_res = ticket->t_unit_res;
2724 	XLOG_TIC_RESET_RES(ticket);
2725 }	/* xlog_regrant_reserve_log_space */
2726 
2727 
2728 /*
2729  * Give back the space left from a reservation.
2730  *
2731  * All the information we need to make a correct determination of space left
2732  * is present.  For non-permanent reservations, things are quite easy.  The
2733  * count should have been decremented to zero.  We only need to deal with the
2734  * space remaining in the current reservation part of the ticket.  If the
2735  * ticket contains a permanent reservation, there may be left over space which
2736  * needs to be released.  A count of N means that N-1 refills of the current
2737  * reservation can be done before we need to ask for more space.  The first
2738  * one goes to fill up the first current reservation.  Once we run out of
2739  * space, the count will stay at zero and the only space remaining will be
2740  * in the current reservation field.
2741  */
2742 STATIC void
2743 xlog_ungrant_log_space(xlog_t	     *log,
2744 		       xlog_ticket_t *ticket)
2745 {
2746 	SPLDECL(s);
2747 
2748 	if (ticket->t_cnt > 0)
2749 		ticket->t_cnt--;
2750 
2751 	s = GRANT_LOCK(log);
2752 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2753 
2754 	xlog_grant_sub_space(log, ticket->t_curr_res);
2755 
2756 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2757 
2758 	/* If this is a permanent reservation ticket, we may be able to free
2759 	 * up more space based on the remaining count.
2760 	 */
2761 	if (ticket->t_cnt > 0) {
2762 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2763 		xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2764 	}
2765 
2766 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2767 	xlog_verify_grant_head(log, 1);
2768 	GRANT_UNLOCK(log, s);
2769 	xfs_log_move_tail(log->l_mp, 1);
2770 }	/* xlog_ungrant_log_space */
2771 
2772 
2773 /*
2774  * Atomically put back used ticket.
2775  */
2776 void
2777 xlog_state_put_ticket(xlog_t	    *log,
2778 		      xlog_ticket_t *tic)
2779 {
2780 	unsigned long s;
2781 
2782 	s = LOG_LOCK(log);
2783 	xlog_ticket_put(log, tic);
2784 	LOG_UNLOCK(log, s);
2785 }	/* xlog_state_put_ticket */
2786 
2787 /*
2788  * Flush iclog to disk if this is the last reference to the given iclog and
2789  * the WANT_SYNC bit is set.
2790  *
2791  * When this function is entered, the iclog is not necessarily in the
2792  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2793  *
2794  *
2795  */
2796 int
2797 xlog_state_release_iclog(xlog_t		*log,
2798 			 xlog_in_core_t	*iclog)
2799 {
2800 	SPLDECL(s);
2801 	int		sync = 0;	/* do we sync? */
2802 
2803 	xlog_assign_tail_lsn(log->l_mp);
2804 
2805 	s = LOG_LOCK(log);
2806 
2807 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2808 		LOG_UNLOCK(log, s);
2809 		return XFS_ERROR(EIO);
2810 	}
2811 
2812 	ASSERT(iclog->ic_refcnt > 0);
2813 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2814 	       iclog->ic_state == XLOG_STATE_WANT_SYNC);
2815 
2816 	if (--iclog->ic_refcnt == 0 &&
2817 	    iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2818 		sync++;
2819 		iclog->ic_state = XLOG_STATE_SYNCING;
2820 		INT_SET(iclog->ic_header.h_tail_lsn, ARCH_CONVERT, log->l_tail_lsn);
2821 		xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2822 		/* cycle incremented when incrementing curr_block */
2823 	}
2824 
2825 	LOG_UNLOCK(log, s);
2826 
2827 	/*
2828 	 * We let the log lock go, so it's possible that we hit a log I/O
2829 	 * error or some other SHUTDOWN condition that marks the iclog
2830 	 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2831 	 * this iclog has consistent data, so we ignore IOERROR
2832 	 * flags after this point.
2833 	 */
2834 	if (sync) {
2835 		return xlog_sync(log, iclog);
2836 	}
2837 	return 0;
2838 
2839 }	/* xlog_state_release_iclog */
2840 
2841 
2842 /*
2843  * This routine will mark the current iclog in the ring as WANT_SYNC
2844  * and move the current iclog pointer to the next iclog in the ring.
2845  * When this routine is called from xlog_state_get_iclog_space(), the
2846  * exact size of the iclog has not yet been determined.  All we know is
2847  * that every data block.  We have run out of space in this log record.
2848  */
2849 STATIC void
2850 xlog_state_switch_iclogs(xlog_t		*log,
2851 			 xlog_in_core_t *iclog,
2852 			 int		eventual_size)
2853 {
2854 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2855 	if (!eventual_size)
2856 		eventual_size = iclog->ic_offset;
2857 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2858 	INT_SET(iclog->ic_header.h_prev_block, ARCH_CONVERT, log->l_prev_block);
2859 	log->l_prev_block = log->l_curr_block;
2860 	log->l_prev_cycle = log->l_curr_cycle;
2861 
2862 	/* roll log?: ic_offset changed later */
2863 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2864 
2865 	/* Round up to next log-sunit */
2866 	if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
2867 	    log->l_mp->m_sb.sb_logsunit > 1) {
2868 		__uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2869 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2870 	}
2871 
2872 	if (log->l_curr_block >= log->l_logBBsize) {
2873 		log->l_curr_cycle++;
2874 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2875 			log->l_curr_cycle++;
2876 		log->l_curr_block -= log->l_logBBsize;
2877 		ASSERT(log->l_curr_block >= 0);
2878 	}
2879 	ASSERT(iclog == log->l_iclog);
2880 	log->l_iclog = iclog->ic_next;
2881 }	/* xlog_state_switch_iclogs */
2882 
2883 
2884 /*
2885  * Write out all data in the in-core log as of this exact moment in time.
2886  *
2887  * Data may be written to the in-core log during this call.  However,
2888  * we don't guarantee this data will be written out.  A change from past
2889  * implementation means this routine will *not* write out zero length LRs.
2890  *
2891  * Basically, we try and perform an intelligent scan of the in-core logs.
2892  * If we determine there is no flushable data, we just return.  There is no
2893  * flushable data if:
2894  *
2895  *	1. the current iclog is active and has no data; the previous iclog
2896  *		is in the active or dirty state.
2897  *	2. the current iclog is drity, and the previous iclog is in the
2898  *		active or dirty state.
2899  *
2900  * We may sleep (call psema) if:
2901  *
2902  *	1. the current iclog is not in the active nor dirty state.
2903  *	2. the current iclog dirty, and the previous iclog is not in the
2904  *		active nor dirty state.
2905  *	3. the current iclog is active, and there is another thread writing
2906  *		to this particular iclog.
2907  *	4. a) the current iclog is active and has no other writers
2908  *	   b) when we return from flushing out this iclog, it is still
2909  *		not in the active nor dirty state.
2910  */
2911 STATIC int
2912 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2913 {
2914 	xlog_in_core_t	*iclog;
2915 	xfs_lsn_t	lsn;
2916 	SPLDECL(s);
2917 
2918 	s = LOG_LOCK(log);
2919 
2920 	iclog = log->l_iclog;
2921 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2922 		LOG_UNLOCK(log, s);
2923 		return XFS_ERROR(EIO);
2924 	}
2925 
2926 	/* If the head iclog is not active nor dirty, we just attach
2927 	 * ourselves to the head and go to sleep.
2928 	 */
2929 	if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2930 	    iclog->ic_state == XLOG_STATE_DIRTY) {
2931 		/*
2932 		 * If the head is dirty or (active and empty), then
2933 		 * we need to look at the previous iclog.  If the previous
2934 		 * iclog is active or dirty we are done.  There is nothing
2935 		 * to sync out.  Otherwise, we attach ourselves to the
2936 		 * previous iclog and go to sleep.
2937 		 */
2938 		if (iclog->ic_state == XLOG_STATE_DIRTY ||
2939 		    (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
2940 			iclog = iclog->ic_prev;
2941 			if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2942 			    iclog->ic_state == XLOG_STATE_DIRTY)
2943 				goto no_sleep;
2944 			else
2945 				goto maybe_sleep;
2946 		} else {
2947 			if (iclog->ic_refcnt == 0) {
2948 				/* We are the only one with access to this
2949 				 * iclog.  Flush it out now.  There should
2950 				 * be a roundoff of zero to show that someone
2951 				 * has already taken care of the roundoff from
2952 				 * the previous sync.
2953 				 */
2954 				iclog->ic_refcnt++;
2955 				lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2956 				xlog_state_switch_iclogs(log, iclog, 0);
2957 				LOG_UNLOCK(log, s);
2958 
2959 				if (xlog_state_release_iclog(log, iclog))
2960 					return XFS_ERROR(EIO);
2961 				*log_flushed = 1;
2962 				s = LOG_LOCK(log);
2963 				if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) == lsn &&
2964 				    iclog->ic_state != XLOG_STATE_DIRTY)
2965 					goto maybe_sleep;
2966 				else
2967 					goto no_sleep;
2968 			} else {
2969 				/* Someone else is writing to this iclog.
2970 				 * Use its call to flush out the data.  However,
2971 				 * the other thread may not force out this LR,
2972 				 * so we mark it WANT_SYNC.
2973 				 */
2974 				xlog_state_switch_iclogs(log, iclog, 0);
2975 				goto maybe_sleep;
2976 			}
2977 		}
2978 	}
2979 
2980 	/* By the time we come around again, the iclog could've been filled
2981 	 * which would give it another lsn.  If we have a new lsn, just
2982 	 * return because the relevant data has been flushed.
2983 	 */
2984 maybe_sleep:
2985 	if (flags & XFS_LOG_SYNC) {
2986 		/*
2987 		 * We must check if we're shutting down here, before
2988 		 * we wait, while we're holding the LOG_LOCK.
2989 		 * Then we check again after waking up, in case our
2990 		 * sleep was disturbed by a bad news.
2991 		 */
2992 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
2993 			LOG_UNLOCK(log, s);
2994 			return XFS_ERROR(EIO);
2995 		}
2996 		XFS_STATS_INC(xs_log_force_sleep);
2997 		sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
2998 		/*
2999 		 * No need to grab the log lock here since we're
3000 		 * only deciding whether or not to return EIO
3001 		 * and the memory read should be atomic.
3002 		 */
3003 		if (iclog->ic_state & XLOG_STATE_IOERROR)
3004 			return XFS_ERROR(EIO);
3005 		*log_flushed = 1;
3006 
3007 	} else {
3008 
3009 no_sleep:
3010 		LOG_UNLOCK(log, s);
3011 	}
3012 	return 0;
3013 }	/* xlog_state_sync_all */
3014 
3015 
3016 /*
3017  * Used by code which implements synchronous log forces.
3018  *
3019  * Find in-core log with lsn.
3020  *	If it is in the DIRTY state, just return.
3021  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3022  *		state and go to sleep or return.
3023  *	If it is in any other state, go to sleep or return.
3024  *
3025  * If filesystem activity goes to zero, the iclog will get flushed only by
3026  * bdflush().
3027  */
3028 int
3029 xlog_state_sync(xlog_t	  *log,
3030 		xfs_lsn_t lsn,
3031 		uint	  flags,
3032 		int	  *log_flushed)
3033 {
3034     xlog_in_core_t	*iclog;
3035     int			already_slept = 0;
3036     SPLDECL(s);
3037 
3038 
3039 try_again:
3040     s = LOG_LOCK(log);
3041     iclog = log->l_iclog;
3042 
3043     if (iclog->ic_state & XLOG_STATE_IOERROR) {
3044 	    LOG_UNLOCK(log, s);
3045 	    return XFS_ERROR(EIO);
3046     }
3047 
3048     do {
3049 	if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) != lsn) {
3050 	    iclog = iclog->ic_next;
3051 	    continue;
3052 	}
3053 
3054 	if (iclog->ic_state == XLOG_STATE_DIRTY) {
3055 		LOG_UNLOCK(log, s);
3056 		return 0;
3057 	}
3058 
3059 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3060 		/*
3061 		 * We sleep here if we haven't already slept (e.g.
3062 		 * this is the first time we've looked at the correct
3063 		 * iclog buf) and the buffer before us is going to
3064 		 * be sync'ed. The reason for this is that if we
3065 		 * are doing sync transactions here, by waiting for
3066 		 * the previous I/O to complete, we can allow a few
3067 		 * more transactions into this iclog before we close
3068 		 * it down.
3069 		 *
3070 		 * Otherwise, we mark the buffer WANT_SYNC, and bump
3071 		 * up the refcnt so we can release the log (which drops
3072 		 * the ref count).  The state switch keeps new transaction
3073 		 * commits from using this buffer.  When the current commits
3074 		 * finish writing into the buffer, the refcount will drop to
3075 		 * zero and the buffer will go out then.
3076 		 */
3077 		if (!already_slept &&
3078 		    (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3079 						 XLOG_STATE_SYNCING))) {
3080 			ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3081 			XFS_STATS_INC(xs_log_force_sleep);
3082 			sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3083 				&log->l_icloglock, s);
3084 			*log_flushed = 1;
3085 			already_slept = 1;
3086 			goto try_again;
3087 		} else {
3088 			iclog->ic_refcnt++;
3089 			xlog_state_switch_iclogs(log, iclog, 0);
3090 			LOG_UNLOCK(log, s);
3091 			if (xlog_state_release_iclog(log, iclog))
3092 				return XFS_ERROR(EIO);
3093 			*log_flushed = 1;
3094 			s = LOG_LOCK(log);
3095 		}
3096 	}
3097 
3098 	if ((flags & XFS_LOG_SYNC) && /* sleep */
3099 	    !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3100 
3101 		/*
3102 		 * Don't wait on the forcesema if we know that we've
3103 		 * gotten a log write error.
3104 		 */
3105 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
3106 			LOG_UNLOCK(log, s);
3107 			return XFS_ERROR(EIO);
3108 		}
3109 		XFS_STATS_INC(xs_log_force_sleep);
3110 		sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3111 		/*
3112 		 * No need to grab the log lock here since we're
3113 		 * only deciding whether or not to return EIO
3114 		 * and the memory read should be atomic.
3115 		 */
3116 		if (iclog->ic_state & XLOG_STATE_IOERROR)
3117 			return XFS_ERROR(EIO);
3118 		*log_flushed = 1;
3119 	} else {		/* just return */
3120 		LOG_UNLOCK(log, s);
3121 	}
3122 	return 0;
3123 
3124     } while (iclog != log->l_iclog);
3125 
3126     LOG_UNLOCK(log, s);
3127     return 0;
3128 }	/* xlog_state_sync */
3129 
3130 
3131 /*
3132  * Called when we want to mark the current iclog as being ready to sync to
3133  * disk.
3134  */
3135 void
3136 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3137 {
3138 	SPLDECL(s);
3139 
3140 	s = LOG_LOCK(log);
3141 
3142 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3143 		xlog_state_switch_iclogs(log, iclog, 0);
3144 	} else {
3145 		ASSERT(iclog->ic_state &
3146 			(XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3147 	}
3148 
3149 	LOG_UNLOCK(log, s);
3150 }	/* xlog_state_want_sync */
3151 
3152 
3153 
3154 /*****************************************************************************
3155  *
3156  *		TICKET functions
3157  *
3158  *****************************************************************************
3159  */
3160 
3161 /*
3162  *	Algorithm doesn't take into account page size. ;-(
3163  */
3164 STATIC void
3165 xlog_state_ticket_alloc(xlog_t *log)
3166 {
3167 	xlog_ticket_t	*t_list;
3168 	xlog_ticket_t	*next;
3169 	xfs_caddr_t	buf;
3170 	uint		i = (NBPP / sizeof(xlog_ticket_t)) - 2;
3171 	SPLDECL(s);
3172 
3173 	/*
3174 	 * The kmem_zalloc may sleep, so we shouldn't be holding the
3175 	 * global lock.  XXXmiken: may want to use zone allocator.
3176 	 */
3177 	buf = (xfs_caddr_t) kmem_zalloc(NBPP, KM_SLEEP);
3178 
3179 	s = LOG_LOCK(log);
3180 
3181 	/* Attach 1st ticket to Q, so we can keep track of allocated memory */
3182 	t_list = (xlog_ticket_t *)buf;
3183 	t_list->t_next = log->l_unmount_free;
3184 	log->l_unmount_free = t_list++;
3185 	log->l_ticket_cnt++;
3186 	log->l_ticket_tcnt++;
3187 
3188 	/* Next ticket becomes first ticket attached to ticket free list */
3189 	if (log->l_freelist != NULL) {
3190 		ASSERT(log->l_tail != NULL);
3191 		log->l_tail->t_next = t_list;
3192 	} else {
3193 		log->l_freelist = t_list;
3194 	}
3195 	log->l_ticket_cnt++;
3196 	log->l_ticket_tcnt++;
3197 
3198 	/* Cycle through rest of alloc'ed memory, building up free Q */
3199 	for ( ; i > 0; i--) {
3200 		next = t_list + 1;
3201 		t_list->t_next = next;
3202 		t_list = next;
3203 		log->l_ticket_cnt++;
3204 		log->l_ticket_tcnt++;
3205 	}
3206 	t_list->t_next = NULL;
3207 	log->l_tail = t_list;
3208 	LOG_UNLOCK(log, s);
3209 }	/* xlog_state_ticket_alloc */
3210 
3211 
3212 /*
3213  * Put ticket into free list
3214  *
3215  * Assumption: log lock is held around this call.
3216  */
3217 STATIC void
3218 xlog_ticket_put(xlog_t		*log,
3219 		xlog_ticket_t	*ticket)
3220 {
3221 	sv_destroy(&ticket->t_sema);
3222 
3223 	/*
3224 	 * Don't think caching will make that much difference.  It's
3225 	 * more important to make debug easier.
3226 	 */
3227 #if 0
3228 	/* real code will want to use LIFO for caching */
3229 	ticket->t_next = log->l_freelist;
3230 	log->l_freelist = ticket;
3231 	/* no need to clear fields */
3232 #else
3233 	/* When we debug, it is easier if tickets are cycled */
3234 	ticket->t_next     = NULL;
3235 	if (log->l_tail != 0) {
3236 		log->l_tail->t_next = ticket;
3237 	} else {
3238 		ASSERT(log->l_freelist == 0);
3239 		log->l_freelist = ticket;
3240 	}
3241 	log->l_tail	    = ticket;
3242 #endif /* DEBUG */
3243 	log->l_ticket_cnt++;
3244 }	/* xlog_ticket_put */
3245 
3246 
3247 /*
3248  * Grab ticket off freelist or allocation some more
3249  */
3250 xlog_ticket_t *
3251 xlog_ticket_get(xlog_t		*log,
3252 		int		unit_bytes,
3253 		int		cnt,
3254 		char		client,
3255 		uint		xflags)
3256 {
3257 	xlog_ticket_t	*tic;
3258 	uint		num_headers;
3259 	SPLDECL(s);
3260 
3261  alloc:
3262 	if (log->l_freelist == NULL)
3263 		xlog_state_ticket_alloc(log);		/* potentially sleep */
3264 
3265 	s = LOG_LOCK(log);
3266 	if (log->l_freelist == NULL) {
3267 		LOG_UNLOCK(log, s);
3268 		goto alloc;
3269 	}
3270 	tic		= log->l_freelist;
3271 	log->l_freelist	= tic->t_next;
3272 	if (log->l_freelist == NULL)
3273 		log->l_tail = NULL;
3274 	log->l_ticket_cnt--;
3275 	LOG_UNLOCK(log, s);
3276 
3277 	/*
3278 	 * Permanent reservations have up to 'cnt'-1 active log operations
3279 	 * in the log.  A unit in this case is the amount of space for one
3280 	 * of these log operations.  Normal reservations have a cnt of 1
3281 	 * and their unit amount is the total amount of space required.
3282 	 *
3283 	 * The following lines of code account for non-transaction data
3284 	 * which occupy space in the on-disk log.
3285 	 *
3286 	 * Normal form of a transaction is:
3287 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3288 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3289 	 *
3290 	 * We need to account for all the leadup data and trailer data
3291 	 * around the transaction data.
3292 	 * And then we need to account for the worst case in terms of using
3293 	 * more space.
3294 	 * The worst case will happen if:
3295 	 * - the placement of the transaction happens to be such that the
3296 	 *   roundoff is at its maximum
3297 	 * - the transaction data is synced before the commit record is synced
3298 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3299 	 *   Therefore the commit record is in its own Log Record.
3300 	 *   This can happen as the commit record is called with its
3301 	 *   own region to xlog_write().
3302 	 *   This then means that in the worst case, roundoff can happen for
3303 	 *   the commit-rec as well.
3304 	 *   The commit-rec is smaller than padding in this scenario and so it is
3305 	 *   not added separately.
3306 	 */
3307 
3308 	/* for trans header */
3309 	unit_bytes += sizeof(xlog_op_header_t);
3310 	unit_bytes += sizeof(xfs_trans_header_t);
3311 
3312 	/* for start-rec */
3313 	unit_bytes += sizeof(xlog_op_header_t);
3314 
3315 	/* for LR headers */
3316 	num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3317 	unit_bytes += log->l_iclog_hsize * num_headers;
3318 
3319 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3320 	unit_bytes += log->l_iclog_hsize;
3321 
3322 	/* for split-recs - ophdrs added when data split over LRs */
3323 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3324 
3325 	/* for roundoff padding for transaction data and one for commit record */
3326 	if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
3327 	    log->l_mp->m_sb.sb_logsunit > 1) {
3328 		/* log su roundoff */
3329 		unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3330 	} else {
3331 		/* BB roundoff */
3332 		unit_bytes += 2*BBSIZE;
3333         }
3334 
3335 	tic->t_unit_res		= unit_bytes;
3336 	tic->t_curr_res		= unit_bytes;
3337 	tic->t_cnt		= cnt;
3338 	tic->t_ocnt		= cnt;
3339 	tic->t_tid		= (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3340 	tic->t_clientid		= client;
3341 	tic->t_flags		= XLOG_TIC_INITED;
3342 	tic->t_trans_type	= 0;
3343 	if (xflags & XFS_LOG_PERM_RESERV)
3344 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3345 	sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3346 
3347 	XLOG_TIC_RESET_RES(tic);
3348 
3349 	return tic;
3350 }	/* xlog_ticket_get */
3351 
3352 
3353 /******************************************************************************
3354  *
3355  *		Log debug routines
3356  *
3357  ******************************************************************************
3358  */
3359 #if defined(DEBUG)
3360 /*
3361  * Make sure that the destination ptr is within the valid data region of
3362  * one of the iclogs.  This uses backup pointers stored in a different
3363  * part of the log in case we trash the log structure.
3364  */
3365 void
3366 xlog_verify_dest_ptr(xlog_t     *log,
3367 		     __psint_t  ptr)
3368 {
3369 	int i;
3370 	int good_ptr = 0;
3371 
3372 	for (i=0; i < log->l_iclog_bufs; i++) {
3373 		if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3374 		    ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3375 			good_ptr++;
3376 	}
3377 	if (! good_ptr)
3378 		xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3379 }	/* xlog_verify_dest_ptr */
3380 
3381 STATIC void
3382 xlog_verify_grant_head(xlog_t *log, int equals)
3383 {
3384     if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3385 	if (equals)
3386 	    ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3387 	else
3388 	    ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3389     } else {
3390 	ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3391 	ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3392     }
3393 }	/* xlog_verify_grant_head */
3394 
3395 /* check if it will fit */
3396 STATIC void
3397 xlog_verify_tail_lsn(xlog_t	    *log,
3398 		     xlog_in_core_t *iclog,
3399 		     xfs_lsn_t	    tail_lsn)
3400 {
3401     int blocks;
3402 
3403     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3404 	blocks =
3405 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3406 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3407 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3408     } else {
3409 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3410 
3411 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3412 	    xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3413 
3414 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3415 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3416 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3417     }
3418 }	/* xlog_verify_tail_lsn */
3419 
3420 /*
3421  * Perform a number of checks on the iclog before writing to disk.
3422  *
3423  * 1. Make sure the iclogs are still circular
3424  * 2. Make sure we have a good magic number
3425  * 3. Make sure we don't have magic numbers in the data
3426  * 4. Check fields of each log operation header for:
3427  *	A. Valid client identifier
3428  *	B. tid ptr value falls in valid ptr space (user space code)
3429  *	C. Length in log record header is correct according to the
3430  *		individual operation headers within record.
3431  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3432  *	log, check the preceding blocks of the physical log to make sure all
3433  *	the cycle numbers agree with the current cycle number.
3434  */
3435 STATIC void
3436 xlog_verify_iclog(xlog_t	 *log,
3437 		  xlog_in_core_t *iclog,
3438 		  int		 count,
3439 		  boolean_t	 syncing)
3440 {
3441 	xlog_op_header_t	*ophead;
3442 	xlog_in_core_t		*icptr;
3443 	xlog_in_core_2_t	*xhdr;
3444 	xfs_caddr_t		ptr;
3445 	xfs_caddr_t		base_ptr;
3446 	__psint_t		field_offset;
3447 	__uint8_t		clientid;
3448 	int			len, i, j, k, op_len;
3449 	int			idx;
3450 	SPLDECL(s);
3451 
3452 	/* check validity of iclog pointers */
3453 	s = LOG_LOCK(log);
3454 	icptr = log->l_iclog;
3455 	for (i=0; i < log->l_iclog_bufs; i++) {
3456 		if (icptr == 0)
3457 			xlog_panic("xlog_verify_iclog: invalid ptr");
3458 		icptr = icptr->ic_next;
3459 	}
3460 	if (icptr != log->l_iclog)
3461 		xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3462 	LOG_UNLOCK(log, s);
3463 
3464 	/* check log magic numbers */
3465 	ptr = (xfs_caddr_t) &(iclog->ic_header);
3466 	if (INT_GET(*(uint *)ptr, ARCH_CONVERT) != XLOG_HEADER_MAGIC_NUM)
3467 		xlog_panic("xlog_verify_iclog: invalid magic num");
3468 
3469 	for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&(iclog->ic_header))+count;
3470 	     ptr += BBSIZE) {
3471 		if (INT_GET(*(uint *)ptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
3472 			xlog_panic("xlog_verify_iclog: unexpected magic num");
3473 	}
3474 
3475 	/* check fields */
3476 	len = INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT);
3477 	ptr = iclog->ic_datap;
3478 	base_ptr = ptr;
3479 	ophead = (xlog_op_header_t *)ptr;
3480 	xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3481 	for (i = 0; i < len; i++) {
3482 		ophead = (xlog_op_header_t *)ptr;
3483 
3484 		/* clientid is only 1 byte */
3485 		field_offset = (__psint_t)
3486 			       ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3487 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3488 			clientid = ophead->oh_clientid;
3489 		} else {
3490 			idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3491 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3492 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3493 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3494 				clientid = GET_CLIENT_ID(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3495 			} else {
3496 				clientid = GET_CLIENT_ID(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3497 			}
3498 		}
3499 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3500 			cmn_err(CE_WARN, "xlog_verify_iclog: "
3501 				"invalid clientid %d op 0x%p offset 0x%lx",
3502 				clientid, ophead, (unsigned long)field_offset);
3503 
3504 		/* check length */
3505 		field_offset = (__psint_t)
3506 			       ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3507 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3508 			op_len = INT_GET(ophead->oh_len, ARCH_CONVERT);
3509 		} else {
3510 			idx = BTOBBT((__psint_t)&ophead->oh_len -
3511 				    (__psint_t)iclog->ic_datap);
3512 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3513 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3514 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3515 				op_len = INT_GET(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3516 			} else {
3517 				op_len = INT_GET(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3518 			}
3519 		}
3520 		ptr += sizeof(xlog_op_header_t) + op_len;
3521 	}
3522 }	/* xlog_verify_iclog */
3523 #endif
3524 
3525 /*
3526  * Mark all iclogs IOERROR. LOG_LOCK is held by the caller.
3527  */
3528 STATIC int
3529 xlog_state_ioerror(
3530 	xlog_t	*log)
3531 {
3532 	xlog_in_core_t	*iclog, *ic;
3533 
3534 	iclog = log->l_iclog;
3535 	if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3536 		/*
3537 		 * Mark all the incore logs IOERROR.
3538 		 * From now on, no log flushes will result.
3539 		 */
3540 		ic = iclog;
3541 		do {
3542 			ic->ic_state = XLOG_STATE_IOERROR;
3543 			ic = ic->ic_next;
3544 		} while (ic != iclog);
3545 		return 0;
3546 	}
3547 	/*
3548 	 * Return non-zero, if state transition has already happened.
3549 	 */
3550 	return 1;
3551 }
3552 
3553 /*
3554  * This is called from xfs_force_shutdown, when we're forcibly
3555  * shutting down the filesystem, typically because of an IO error.
3556  * Our main objectives here are to make sure that:
3557  *	a. the filesystem gets marked 'SHUTDOWN' for all interested
3558  *	   parties to find out, 'atomically'.
3559  *	b. those who're sleeping on log reservations, pinned objects and
3560  *	    other resources get woken up, and be told the bad news.
3561  *	c. nothing new gets queued up after (a) and (b) are done.
3562  *	d. if !logerror, flush the iclogs to disk, then seal them off
3563  *	   for business.
3564  */
3565 int
3566 xfs_log_force_umount(
3567 	struct xfs_mount	*mp,
3568 	int			logerror)
3569 {
3570 	xlog_ticket_t	*tic;
3571 	xlog_t		*log;
3572 	int		retval;
3573 	int		dummy;
3574 	SPLDECL(s);
3575 	SPLDECL(s2);
3576 
3577 	log = mp->m_log;
3578 
3579 	/*
3580 	 * If this happens during log recovery, don't worry about
3581 	 * locking; the log isn't open for business yet.
3582 	 */
3583 	if (!log ||
3584 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3585 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3586 		XFS_BUF_DONE(mp->m_sb_bp);
3587 		return 0;
3588 	}
3589 
3590 	/*
3591 	 * Somebody could've already done the hard work for us.
3592 	 * No need to get locks for this.
3593 	 */
3594 	if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3595 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3596 		return 1;
3597 	}
3598 	retval = 0;
3599 	/*
3600 	 * We must hold both the GRANT lock and the LOG lock,
3601 	 * before we mark the filesystem SHUTDOWN and wake
3602 	 * everybody up to tell the bad news.
3603 	 */
3604 	s = GRANT_LOCK(log);
3605 	s2 = LOG_LOCK(log);
3606 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3607 	XFS_BUF_DONE(mp->m_sb_bp);
3608 	/*
3609 	 * This flag is sort of redundant because of the mount flag, but
3610 	 * it's good to maintain the separation between the log and the rest
3611 	 * of XFS.
3612 	 */
3613 	log->l_flags |= XLOG_IO_ERROR;
3614 
3615 	/*
3616 	 * If we hit a log error, we want to mark all the iclogs IOERROR
3617 	 * while we're still holding the loglock.
3618 	 */
3619 	if (logerror)
3620 		retval = xlog_state_ioerror(log);
3621 	LOG_UNLOCK(log, s2);
3622 
3623 	/*
3624 	 * We don't want anybody waiting for log reservations
3625 	 * after this. That means we have to wake up everybody
3626 	 * queued up on reserve_headq as well as write_headq.
3627 	 * In addition, we make sure in xlog_{re}grant_log_space
3628 	 * that we don't enqueue anything once the SHUTDOWN flag
3629 	 * is set, and this action is protected by the GRANTLOCK.
3630 	 */
3631 	if ((tic = log->l_reserve_headq)) {
3632 		do {
3633 			sv_signal(&tic->t_sema);
3634 			tic = tic->t_next;
3635 		} while (tic != log->l_reserve_headq);
3636 	}
3637 
3638 	if ((tic = log->l_write_headq)) {
3639 		do {
3640 			sv_signal(&tic->t_sema);
3641 			tic = tic->t_next;
3642 		} while (tic != log->l_write_headq);
3643 	}
3644 	GRANT_UNLOCK(log, s);
3645 
3646 	if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3647 		ASSERT(!logerror);
3648 		/*
3649 		 * Force the incore logs to disk before shutting the
3650 		 * log down completely.
3651 		 */
3652 		xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
3653 		s2 = LOG_LOCK(log);
3654 		retval = xlog_state_ioerror(log);
3655 		LOG_UNLOCK(log, s2);
3656 	}
3657 	/*
3658 	 * Wake up everybody waiting on xfs_log_force.
3659 	 * Callback all log item committed functions as if the
3660 	 * log writes were completed.
3661 	 */
3662 	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3663 
3664 #ifdef XFSERRORDEBUG
3665 	{
3666 		xlog_in_core_t	*iclog;
3667 
3668 		s = LOG_LOCK(log);
3669 		iclog = log->l_iclog;
3670 		do {
3671 			ASSERT(iclog->ic_callback == 0);
3672 			iclog = iclog->ic_next;
3673 		} while (iclog != log->l_iclog);
3674 		LOG_UNLOCK(log, s);
3675 	}
3676 #endif
3677 	/* return non-zero if log IOERROR transition had already happened */
3678 	return retval;
3679 }
3680 
3681 STATIC int
3682 xlog_iclogs_empty(xlog_t *log)
3683 {
3684 	xlog_in_core_t	*iclog;
3685 
3686 	iclog = log->l_iclog;
3687 	do {
3688 		/* endianness does not matter here, zero is zero in
3689 		 * any language.
3690 		 */
3691 		if (iclog->ic_header.h_num_logops)
3692 			return 0;
3693 		iclog = iclog->ic_next;
3694 	} while (iclog != log->l_iclog);
3695 	return 1;
3696 }
3697