xref: /linux/fs/xfs/xfs_log.c (revision 2624f124b3b5d550ab2fbef7ee3bc0e1fed09722)
1 /*
2  * Copyright (c) 2000-2004 Silicon Graphics, Inc.  All Rights Reserved.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms of version 2 of the GNU General Public License as
6  * published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it would be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
11  *
12  * Further, this software is distributed without any warranty that it is
13  * free of the rightful claim of any third person regarding infringement
14  * or the like.  Any license provided herein, whether implied or
15  * otherwise, applies only to this software file.  Patent licenses, if
16  * any, provided herein do not apply to combinations of this program with
17  * other software, or any other product whatsoever.
18  *
19  * You should have received a copy of the GNU General Public License along
20  * with this program; if not, write the Free Software Foundation, Inc., 59
21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
22  *
23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
24  * Mountain View, CA  94043, or:
25  *
26  * http://www.sgi.com
27  *
28  * For further information regarding this notice, see:
29  *
30  * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
31  */
32 
33 /*
34  * High level interface routines for log manager
35  */
36 
37 #include "xfs.h"
38 #include "xfs_macros.h"
39 #include "xfs_types.h"
40 #include "xfs_inum.h"
41 #include "xfs_ag.h"
42 #include "xfs_sb.h"
43 #include "xfs_log.h"
44 #include "xfs_trans.h"
45 #include "xfs_dir.h"
46 #include "xfs_dmapi.h"
47 #include "xfs_mount.h"
48 #include "xfs_error.h"
49 #include "xfs_log_priv.h"
50 #include "xfs_buf_item.h"
51 #include "xfs_alloc_btree.h"
52 #include "xfs_log_recover.h"
53 #include "xfs_bit.h"
54 #include "xfs_rw.h"
55 #include "xfs_trans_priv.h"
56 
57 
58 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
59 	{ (ptr) += (bytes); \
60 	  (len) -= (bytes); \
61 	  (off) += (bytes);}
62 
63 /* Local miscellaneous function prototypes */
64 STATIC int	 xlog_bdstrat_cb(struct xfs_buf *);
65 STATIC int	 xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
66 				    xlog_in_core_t **, xfs_lsn_t *);
67 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t	*mp,
68 				xfs_buftarg_t	*log_target,
69 				xfs_daddr_t	blk_offset,
70 				int		num_bblks);
71 STATIC int	 xlog_space_left(xlog_t *log, int cycle, int bytes);
72 STATIC int	 xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
73 STATIC void	 xlog_unalloc_log(xlog_t *log);
74 STATIC int	 xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
75 			    int nentries, xfs_log_ticket_t tic,
76 			    xfs_lsn_t *start_lsn,
77 			    xlog_in_core_t **commit_iclog,
78 			    uint flags);
79 
80 /* local state machine functions */
81 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
82 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
83 STATIC int  xlog_state_get_iclog_space(xlog_t		*log,
84 				       int		len,
85 				       xlog_in_core_t	**iclog,
86 				       xlog_ticket_t	*ticket,
87 				       int		*continued_write,
88 				       int		*logoffsetp);
89 STATIC void xlog_state_put_ticket(xlog_t	*log,
90 				  xlog_ticket_t *tic);
91 STATIC int  xlog_state_release_iclog(xlog_t		*log,
92 				     xlog_in_core_t	*iclog);
93 STATIC void xlog_state_switch_iclogs(xlog_t		*log,
94 				     xlog_in_core_t *iclog,
95 				     int		eventual_size);
96 STATIC int  xlog_state_sync(xlog_t *log, xfs_lsn_t lsn, uint flags);
97 STATIC int  xlog_state_sync_all(xlog_t *log, uint flags);
98 STATIC void xlog_state_want_sync(xlog_t	*log, xlog_in_core_t *iclog);
99 
100 /* local functions to manipulate grant head */
101 STATIC int  xlog_grant_log_space(xlog_t		*log,
102 				 xlog_ticket_t	*xtic);
103 STATIC void xlog_grant_push_ail(xfs_mount_t	*mp,
104 				int		need_bytes);
105 STATIC void xlog_regrant_reserve_log_space(xlog_t	 *log,
106 					   xlog_ticket_t *ticket);
107 STATIC int xlog_regrant_write_log_space(xlog_t		*log,
108 					 xlog_ticket_t  *ticket);
109 STATIC void xlog_ungrant_log_space(xlog_t	 *log,
110 				   xlog_ticket_t *ticket);
111 
112 
113 /* local ticket functions */
114 STATIC void		xlog_state_ticket_alloc(xlog_t *log);
115 STATIC xlog_ticket_t	*xlog_ticket_get(xlog_t *log,
116 					 int	unit_bytes,
117 					 int	count,
118 					 char	clientid,
119 					 uint	flags);
120 STATIC void		xlog_ticket_put(xlog_t *log, xlog_ticket_t *ticket);
121 
122 /* local debug functions */
123 #if defined(DEBUG) && !defined(XLOG_NOLOG)
124 STATIC void	xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
125 STATIC void	xlog_verify_grant_head(xlog_t *log, int equals);
126 STATIC void	xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
127 				  int count, boolean_t syncing);
128 STATIC void	xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
129 				     xfs_lsn_t tail_lsn);
130 #else
131 #define xlog_verify_dest_ptr(a,b)
132 #define xlog_verify_grant_head(a,b)
133 #define xlog_verify_iclog(a,b,c,d)
134 #define xlog_verify_tail_lsn(a,b,c)
135 #endif
136 
137 STATIC int	xlog_iclogs_empty(xlog_t *log);
138 
139 #ifdef DEBUG
140 int xlog_do_error = 0;
141 int xlog_req_num  = 0;
142 int xlog_error_mod = 33;
143 #endif
144 
145 #define XLOG_FORCED_SHUTDOWN(log)	(log->l_flags & XLOG_IO_ERROR)
146 
147 /*
148  * 0 => disable log manager
149  * 1 => enable log manager
150  * 2 => enable log manager and log debugging
151  */
152 #if defined(XLOG_NOLOG) || defined(DEBUG)
153 int   xlog_debug = 1;
154 xfs_buftarg_t *xlog_target;
155 #endif
156 
157 #if defined(XFS_LOG_TRACE)
158 
159 void
160 xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
161 {
162 	unsigned long cnts;
163 
164 	if (!log->l_grant_trace) {
165 		log->l_grant_trace = ktrace_alloc(2048, KM_NOSLEEP);
166 		if (!log->l_grant_trace)
167 			return;
168 	}
169 	/* ticket counts are 1 byte each */
170 	cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
171 
172 	ktrace_enter(log->l_grant_trace,
173 		     (void *)tic,
174 		     (void *)log->l_reserve_headq,
175 		     (void *)log->l_write_headq,
176 		     (void *)((unsigned long)log->l_grant_reserve_cycle),
177 		     (void *)((unsigned long)log->l_grant_reserve_bytes),
178 		     (void *)((unsigned long)log->l_grant_write_cycle),
179 		     (void *)((unsigned long)log->l_grant_write_bytes),
180 		     (void *)((unsigned long)log->l_curr_cycle),
181 		     (void *)((unsigned long)log->l_curr_block),
182 		     (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
183 		     (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
184 		     (void *)string,
185 		     (void *)((unsigned long)tic->t_trans_type),
186 		     (void *)cnts,
187 		     (void *)((unsigned long)tic->t_curr_res),
188 		     (void *)((unsigned long)tic->t_unit_res));
189 }
190 
191 void
192 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
193 {
194 	pid_t pid;
195 
196 	pid = current_pid();
197 
198 	if (!iclog->ic_trace)
199 		iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
200 	ktrace_enter(iclog->ic_trace,
201 		     (void *)((unsigned long)state),
202 		     (void *)((unsigned long)pid),
203 		     (void *)0,
204 		     (void *)0,
205 		     (void *)0,
206 		     (void *)0,
207 		     (void *)0,
208 		     (void *)0,
209 		     (void *)0,
210 		     (void *)0,
211 		     (void *)0,
212 		     (void *)0,
213 		     (void *)0,
214 		     (void *)0,
215 		     (void *)0,
216 		     (void *)0);
217 }
218 
219 #else
220 #define	xlog_trace_loggrant(log,tic,string)
221 #define	xlog_trace_iclog(iclog,state)
222 #endif /* XFS_LOG_TRACE */
223 
224 /*
225  * NOTES:
226  *
227  *	1. currblock field gets updated at startup and after in-core logs
228  *		marked as with WANT_SYNC.
229  */
230 
231 /*
232  * This routine is called when a user of a log manager ticket is done with
233  * the reservation.  If the ticket was ever used, then a commit record for
234  * the associated transaction is written out as a log operation header with
235  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
236  * a given ticket.  If the ticket was one with a permanent reservation, then
237  * a few operations are done differently.  Permanent reservation tickets by
238  * default don't release the reservation.  They just commit the current
239  * transaction with the belief that the reservation is still needed.  A flag
240  * must be passed in before permanent reservations are actually released.
241  * When these type of tickets are not released, they need to be set into
242  * the inited state again.  By doing this, a start record will be written
243  * out when the next write occurs.
244  */
245 xfs_lsn_t
246 xfs_log_done(xfs_mount_t	*mp,
247 	     xfs_log_ticket_t	xtic,
248 	     void		**iclog,
249 	     uint		flags)
250 {
251 	xlog_t		*log    = mp->m_log;
252 	xlog_ticket_t	*ticket = (xfs_log_ticket_t) xtic;
253 	xfs_lsn_t	lsn	= 0;
254 
255 #if defined(DEBUG) || defined(XLOG_NOLOG)
256 	if (!xlog_debug && xlog_target == log->l_targ)
257 		return 0;
258 #endif
259 
260 	if (XLOG_FORCED_SHUTDOWN(log) ||
261 	    /*
262 	     * If nothing was ever written, don't write out commit record.
263 	     * If we get an error, just continue and give back the log ticket.
264 	     */
265 	    (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
266 	     (xlog_commit_record(mp, ticket,
267 				 (xlog_in_core_t **)iclog, &lsn)))) {
268 		lsn = (xfs_lsn_t) -1;
269 		if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
270 			flags |= XFS_LOG_REL_PERM_RESERV;
271 		}
272 	}
273 
274 
275 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
276 	    (flags & XFS_LOG_REL_PERM_RESERV)) {
277 		/*
278 		 * Release ticket if not permanent reservation or a specifc
279 		 * request has been made to release a permanent reservation.
280 		 */
281 		xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
282 		xlog_ungrant_log_space(log, ticket);
283 		xlog_state_put_ticket(log, ticket);
284 	} else {
285 		xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
286 		xlog_regrant_reserve_log_space(log, ticket);
287 	}
288 
289 	/* If this ticket was a permanent reservation and we aren't
290 	 * trying to release it, reset the inited flags; so next time
291 	 * we write, a start record will be written out.
292 	 */
293 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) &&
294 	    (flags & XFS_LOG_REL_PERM_RESERV) == 0)
295 		ticket->t_flags |= XLOG_TIC_INITED;
296 
297 	return lsn;
298 }	/* xfs_log_done */
299 
300 
301 /*
302  * Force the in-core log to disk.  If flags == XFS_LOG_SYNC,
303  *	the force is done synchronously.
304  *
305  * Asynchronous forces are implemented by setting the WANT_SYNC
306  * bit in the appropriate in-core log and then returning.
307  *
308  * Synchronous forces are implemented with a semaphore.  All callers
309  * to force a given lsn to disk will wait on a semaphore attached to the
310  * specific in-core log.  When given in-core log finally completes its
311  * write to disk, that thread will wake up all threads waiting on the
312  * semaphore.
313  */
314 int
315 xfs_log_force(xfs_mount_t *mp,
316 	      xfs_lsn_t	  lsn,
317 	      uint	  flags)
318 {
319 	int	rval;
320 	xlog_t *log = mp->m_log;
321 
322 #if defined(DEBUG) || defined(XLOG_NOLOG)
323 	if (!xlog_debug && xlog_target == log->l_targ)
324 		return 0;
325 #endif
326 
327 	ASSERT(flags & XFS_LOG_FORCE);
328 
329 	XFS_STATS_INC(xs_log_force);
330 
331 	if ((log->l_flags & XLOG_IO_ERROR) == 0) {
332 		if (lsn == 0)
333 			rval = xlog_state_sync_all(log, flags);
334 		else
335 			rval = xlog_state_sync(log, lsn, flags);
336 	} else {
337 		rval = XFS_ERROR(EIO);
338 	}
339 
340 	return rval;
341 
342 }	/* xfs_log_force */
343 
344 /*
345  * Attaches a new iclog I/O completion callback routine during
346  * transaction commit.  If the log is in error state, a non-zero
347  * return code is handed back and the caller is responsible for
348  * executing the callback at an appropriate time.
349  */
350 int
351 xfs_log_notify(xfs_mount_t	  *mp,		/* mount of partition */
352 	       void		  *iclog_hndl,	/* iclog to hang callback off */
353 	       xfs_log_callback_t *cb)
354 {
355 	xlog_t *log = mp->m_log;
356 	xlog_in_core_t	  *iclog = (xlog_in_core_t *)iclog_hndl;
357 	int	abortflg, spl;
358 
359 #if defined(DEBUG) || defined(XLOG_NOLOG)
360 	if (!xlog_debug && xlog_target == log->l_targ)
361 		return 0;
362 #endif
363 	cb->cb_next = NULL;
364 	spl = LOG_LOCK(log);
365 	abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
366 	if (!abortflg) {
367 		ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
368 			      (iclog->ic_state == XLOG_STATE_WANT_SYNC));
369 		cb->cb_next = NULL;
370 		*(iclog->ic_callback_tail) = cb;
371 		iclog->ic_callback_tail = &(cb->cb_next);
372 	}
373 	LOG_UNLOCK(log, spl);
374 	return abortflg;
375 }	/* xfs_log_notify */
376 
377 int
378 xfs_log_release_iclog(xfs_mount_t *mp,
379 		      void	  *iclog_hndl)
380 {
381 	xlog_t *log = mp->m_log;
382 	xlog_in_core_t	  *iclog = (xlog_in_core_t *)iclog_hndl;
383 
384 	if (xlog_state_release_iclog(log, iclog)) {
385 		xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
386 		return(EIO);
387 	}
388 
389 	return 0;
390 }
391 
392 /*
393  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
394  *	to the reservation.
395  *  2. Potentially, push buffers at tail of log to disk.
396  *
397  * Each reservation is going to reserve extra space for a log record header.
398  * When writes happen to the on-disk log, we don't subtract the length of the
399  * log record header from any reservation.  By wasting space in each
400  * reservation, we prevent over allocation problems.
401  */
402 int
403 xfs_log_reserve(xfs_mount_t	 *mp,
404 		int		 unit_bytes,
405 		int		 cnt,
406 		xfs_log_ticket_t *ticket,
407 		__uint8_t	 client,
408 		uint		 flags,
409 		uint		 t_type)
410 {
411 	xlog_t		*log = mp->m_log;
412 	xlog_ticket_t	*internal_ticket;
413 	int		retval;
414 
415 #if defined(DEBUG) || defined(XLOG_NOLOG)
416 	if (!xlog_debug && xlog_target == log->l_targ)
417 		return 0;
418 #endif
419 	retval = 0;
420 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
421 	ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
422 
423 	if (XLOG_FORCED_SHUTDOWN(log))
424 		return XFS_ERROR(EIO);
425 
426 	XFS_STATS_INC(xs_try_logspace);
427 
428 	if (*ticket != NULL) {
429 		ASSERT(flags & XFS_LOG_PERM_RESERV);
430 		internal_ticket = (xlog_ticket_t *)*ticket;
431 		xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
432 		xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
433 		retval = xlog_regrant_write_log_space(log, internal_ticket);
434 	} else {
435 		/* may sleep if need to allocate more tickets */
436 		internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
437 						  client, flags);
438 		internal_ticket->t_trans_type = t_type;
439 		*ticket = internal_ticket;
440 		xlog_trace_loggrant(log, internal_ticket,
441 			(internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
442 			"xfs_log_reserve: create new ticket (permanent trans)" :
443 			"xfs_log_reserve: create new ticket");
444 		xlog_grant_push_ail(mp,
445 				    (internal_ticket->t_unit_res *
446 				     internal_ticket->t_cnt));
447 		retval = xlog_grant_log_space(log, internal_ticket);
448 	}
449 
450 	return retval;
451 }	/* xfs_log_reserve */
452 
453 
454 /*
455  * Mount a log filesystem
456  *
457  * mp		- ubiquitous xfs mount point structure
458  * log_target	- buftarg of on-disk log device
459  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
460  * num_bblocks	- Number of BBSIZE blocks in on-disk log
461  *
462  * Return error or zero.
463  */
464 int
465 xfs_log_mount(xfs_mount_t	*mp,
466 	      xfs_buftarg_t	*log_target,
467 	      xfs_daddr_t	blk_offset,
468 	      int		num_bblks)
469 {
470 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
471 		cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
472 	else {
473 		cmn_err(CE_NOTE,
474 			"!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
475 			mp->m_fsname);
476 		ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
477 	}
478 
479 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
480 
481 #if defined(DEBUG) || defined(XLOG_NOLOG)
482 	if (!xlog_debug) {
483 		cmn_err(CE_NOTE, "log dev: %s", XFS_BUFTARG_NAME(log_target));
484 		return 0;
485 	}
486 #endif
487 	/*
488 	 * skip log recovery on a norecovery mount.  pretend it all
489 	 * just worked.
490 	 */
491 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
492 		int	error;
493 		vfs_t	*vfsp = XFS_MTOVFS(mp);
494 		int	readonly = (vfsp->vfs_flag & VFS_RDONLY);
495 
496 		if (readonly)
497 			vfsp->vfs_flag &= ~VFS_RDONLY;
498 
499 		error = xlog_recover(mp->m_log, readonly);
500 
501 		if (readonly)
502 			vfsp->vfs_flag |= VFS_RDONLY;
503 		if (error) {
504 			cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
505 			xlog_unalloc_log(mp->m_log);
506 			return error;
507 		}
508 	}
509 
510 	/* Normal transactions can now occur */
511 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
512 
513 	/* End mounting message in xfs_log_mount_finish */
514 	return 0;
515 }	/* xfs_log_mount */
516 
517 /*
518  * Finish the recovery of the file system.  This is separate from
519  * the xfs_log_mount() call, because it depends on the code in
520  * xfs_mountfs() to read in the root and real-time bitmap inodes
521  * between calling xfs_log_mount() and here.
522  *
523  * mp		- ubiquitous xfs mount point structure
524  */
525 int
526 xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
527 {
528 	int	error;
529 
530 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
531 		error = xlog_recover_finish(mp->m_log, mfsi_flags);
532 	else {
533 		error = 0;
534 		ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
535 	}
536 
537 	return error;
538 }
539 
540 /*
541  * Unmount processing for the log.
542  */
543 int
544 xfs_log_unmount(xfs_mount_t *mp)
545 {
546 	int		error;
547 
548 	error = xfs_log_unmount_write(mp);
549 	xfs_log_unmount_dealloc(mp);
550 	return (error);
551 }
552 
553 /*
554  * Final log writes as part of unmount.
555  *
556  * Mark the filesystem clean as unmount happens.  Note that during relocation
557  * this routine needs to be executed as part of source-bag while the
558  * deallocation must not be done until source-end.
559  */
560 
561 /*
562  * Unmount record used to have a string "Unmount filesystem--" in the
563  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
564  * We just write the magic number now since that particular field isn't
565  * currently architecture converted and "nUmount" is a bit foo.
566  * As far as I know, there weren't any dependencies on the old behaviour.
567  */
568 
569 int
570 xfs_log_unmount_write(xfs_mount_t *mp)
571 {
572 	xlog_t		 *log = mp->m_log;
573 	xlog_in_core_t	 *iclog;
574 #ifdef DEBUG
575 	xlog_in_core_t	 *first_iclog;
576 #endif
577 	xfs_log_iovec_t  reg[1];
578 	xfs_log_ticket_t tic = NULL;
579 	xfs_lsn_t	 lsn;
580 	int		 error;
581 	SPLDECL(s);
582 
583 	/* the data section must be 32 bit size aligned */
584 	struct {
585 	    __uint16_t magic;
586 	    __uint16_t pad1;
587 	    __uint32_t pad2; /* may as well make it 64 bits */
588 	} magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
589 
590 #if defined(DEBUG) || defined(XLOG_NOLOG)
591 	if (!xlog_debug && xlog_target == log->l_targ)
592 		return 0;
593 #endif
594 
595 	/*
596 	 * Don't write out unmount record on read-only mounts.
597 	 * Or, if we are doing a forced umount (typically because of IO errors).
598 	 */
599 	if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
600 		return 0;
601 
602 	xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
603 
604 #ifdef DEBUG
605 	first_iclog = iclog = log->l_iclog;
606 	do {
607 		if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
608 			ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
609 			ASSERT(iclog->ic_offset == 0);
610 		}
611 		iclog = iclog->ic_next;
612 	} while (iclog != first_iclog);
613 #endif
614 	if (! (XLOG_FORCED_SHUTDOWN(log))) {
615 		reg[0].i_addr = (void*)&magic;
616 		reg[0].i_len  = sizeof(magic);
617 		XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
618 
619 		error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0, 0);
620 		if (!error) {
621 			/* remove inited flag */
622 			((xlog_ticket_t *)tic)->t_flags = 0;
623 			error = xlog_write(mp, reg, 1, tic, &lsn,
624 					   NULL, XLOG_UNMOUNT_TRANS);
625 			/*
626 			 * At this point, we're umounting anyway,
627 			 * so there's no point in transitioning log state
628 			 * to IOERROR. Just continue...
629 			 */
630 		}
631 
632 		if (error) {
633 			xfs_fs_cmn_err(CE_ALERT, mp,
634 				"xfs_log_unmount: unmount record failed");
635 		}
636 
637 
638 		s = LOG_LOCK(log);
639 		iclog = log->l_iclog;
640 		iclog->ic_refcnt++;
641 		LOG_UNLOCK(log, s);
642 		xlog_state_want_sync(log, iclog);
643 		(void) xlog_state_release_iclog(log, iclog);
644 
645 		s = LOG_LOCK(log);
646 		if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
647 		      iclog->ic_state == XLOG_STATE_DIRTY)) {
648 			if (!XLOG_FORCED_SHUTDOWN(log)) {
649 				sv_wait(&iclog->ic_forcesema, PMEM,
650 					&log->l_icloglock, s);
651 			} else {
652 				LOG_UNLOCK(log, s);
653 			}
654 		} else {
655 			LOG_UNLOCK(log, s);
656 		}
657 		if (tic)
658 			xlog_state_put_ticket(log, tic);
659 	} else {
660 		/*
661 		 * We're already in forced_shutdown mode, couldn't
662 		 * even attempt to write out the unmount transaction.
663 		 *
664 		 * Go through the motions of sync'ing and releasing
665 		 * the iclog, even though no I/O will actually happen,
666 		 * we need to wait for other log I/O's that may already
667 		 * be in progress.  Do this as a separate section of
668 		 * code so we'll know if we ever get stuck here that
669 		 * we're in this odd situation of trying to unmount
670 		 * a file system that went into forced_shutdown as
671 		 * the result of an unmount..
672 		 */
673 		s = LOG_LOCK(log);
674 		iclog = log->l_iclog;
675 		iclog->ic_refcnt++;
676 		LOG_UNLOCK(log, s);
677 
678 		xlog_state_want_sync(log, iclog);
679 		(void) xlog_state_release_iclog(log, iclog);
680 
681 		s = LOG_LOCK(log);
682 
683 		if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
684 			|| iclog->ic_state == XLOG_STATE_DIRTY
685 			|| iclog->ic_state == XLOG_STATE_IOERROR) ) {
686 
687 				sv_wait(&iclog->ic_forcesema, PMEM,
688 					&log->l_icloglock, s);
689 		} else {
690 			LOG_UNLOCK(log, s);
691 		}
692 	}
693 
694 	return 0;
695 }	/* xfs_log_unmount_write */
696 
697 /*
698  * Deallocate log structures for unmount/relocation.
699  */
700 void
701 xfs_log_unmount_dealloc(xfs_mount_t *mp)
702 {
703 	xlog_unalloc_log(mp->m_log);
704 }
705 
706 /*
707  * Write region vectors to log.  The write happens using the space reservation
708  * of the ticket (tic).  It is not a requirement that all writes for a given
709  * transaction occur with one call to xfs_log_write().
710  */
711 int
712 xfs_log_write(xfs_mount_t *	mp,
713 	      xfs_log_iovec_t	reg[],
714 	      int		nentries,
715 	      xfs_log_ticket_t	tic,
716 	      xfs_lsn_t		*start_lsn)
717 {
718 	int	error;
719 	xlog_t *log = mp->m_log;
720 
721 #if defined(DEBUG) || defined(XLOG_NOLOG)
722 	if (!xlog_debug && xlog_target == log->l_targ) {
723 		*start_lsn = 0;
724 		return 0;
725 	}
726 #endif
727 	if (XLOG_FORCED_SHUTDOWN(log))
728 		return XFS_ERROR(EIO);
729 
730 	if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
731 		xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
732 	}
733 	return (error);
734 }	/* xfs_log_write */
735 
736 
737 void
738 xfs_log_move_tail(xfs_mount_t	*mp,
739 		  xfs_lsn_t	tail_lsn)
740 {
741 	xlog_ticket_t	*tic;
742 	xlog_t		*log = mp->m_log;
743 	int		need_bytes, free_bytes, cycle, bytes;
744 	SPLDECL(s);
745 
746 #if defined(DEBUG) || defined(XLOG_NOLOG)
747 	if (!xlog_debug && xlog_target == log->l_targ)
748 		return;
749 #endif
750 	/* XXXsup tmp */
751 	if (XLOG_FORCED_SHUTDOWN(log))
752 		return;
753 	ASSERT(!XFS_FORCED_SHUTDOWN(mp));
754 
755 	if (tail_lsn == 0) {
756 		/* needed since sync_lsn is 64 bits */
757 		s = LOG_LOCK(log);
758 		tail_lsn = log->l_last_sync_lsn;
759 		LOG_UNLOCK(log, s);
760 	}
761 
762 	s = GRANT_LOCK(log);
763 
764 	/* Also an invalid lsn.  1 implies that we aren't passing in a valid
765 	 * tail_lsn.
766 	 */
767 	if (tail_lsn != 1) {
768 		log->l_tail_lsn = tail_lsn;
769 	}
770 
771 	if ((tic = log->l_write_headq)) {
772 #ifdef DEBUG
773 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
774 			panic("Recovery problem");
775 #endif
776 		cycle = log->l_grant_write_cycle;
777 		bytes = log->l_grant_write_bytes;
778 		free_bytes = xlog_space_left(log, cycle, bytes);
779 		do {
780 			ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
781 
782 			if (free_bytes < tic->t_unit_res && tail_lsn != 1)
783 				break;
784 			tail_lsn = 0;
785 			free_bytes -= tic->t_unit_res;
786 			sv_signal(&tic->t_sema);
787 			tic = tic->t_next;
788 		} while (tic != log->l_write_headq);
789 	}
790 	if ((tic = log->l_reserve_headq)) {
791 #ifdef DEBUG
792 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
793 			panic("Recovery problem");
794 #endif
795 		cycle = log->l_grant_reserve_cycle;
796 		bytes = log->l_grant_reserve_bytes;
797 		free_bytes = xlog_space_left(log, cycle, bytes);
798 		do {
799 			if (tic->t_flags & XLOG_TIC_PERM_RESERV)
800 				need_bytes = tic->t_unit_res*tic->t_cnt;
801 			else
802 				need_bytes = tic->t_unit_res;
803 			if (free_bytes < need_bytes && tail_lsn != 1)
804 				break;
805 			tail_lsn = 0;
806 			free_bytes -= need_bytes;
807 			sv_signal(&tic->t_sema);
808 			tic = tic->t_next;
809 		} while (tic != log->l_reserve_headq);
810 	}
811 	GRANT_UNLOCK(log, s);
812 }	/* xfs_log_move_tail */
813 
814 /*
815  * Determine if we have a transaction that has gone to disk
816  * that needs to be covered. Log activity needs to be idle (no AIL and
817  * nothing in the iclogs). And, we need to be in the right state indicating
818  * something has gone out.
819  */
820 int
821 xfs_log_need_covered(xfs_mount_t *mp)
822 {
823 	SPLDECL(s);
824 	int		needed = 0, gen;
825 	xlog_t		*log = mp->m_log;
826 	vfs_t		*vfsp = XFS_MTOVFS(mp);
827 
828 	if (fs_frozen(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
829 	    (vfsp->vfs_flag & VFS_RDONLY))
830 		return 0;
831 
832 	s = LOG_LOCK(log);
833 	if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
834 		(log->l_covered_state == XLOG_STATE_COVER_NEED2))
835 			&& !xfs_trans_first_ail(mp, &gen)
836 			&& xlog_iclogs_empty(log)) {
837 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
838 			log->l_covered_state = XLOG_STATE_COVER_DONE;
839 		else {
840 			ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
841 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
842 		}
843 		needed = 1;
844 	}
845 	LOG_UNLOCK(log, s);
846 	return(needed);
847 }
848 
849 /******************************************************************************
850  *
851  *	local routines
852  *
853  ******************************************************************************
854  */
855 
856 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
857  * The log manager must keep track of the last LR which was committed
858  * to disk.  The lsn of this LR will become the new tail_lsn whenever
859  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
860  * the situation where stuff could be written into the log but nothing
861  * was ever in the AIL when asked.  Eventually, we panic since the
862  * tail hits the head.
863  *
864  * We may be holding the log iclog lock upon entering this routine.
865  */
866 xfs_lsn_t
867 xlog_assign_tail_lsn(xfs_mount_t *mp)
868 {
869 	xfs_lsn_t tail_lsn;
870 	SPLDECL(s);
871 	xlog_t	  *log = mp->m_log;
872 
873 	tail_lsn = xfs_trans_tail_ail(mp);
874 	s = GRANT_LOCK(log);
875 	if (tail_lsn != 0) {
876 		log->l_tail_lsn = tail_lsn;
877 	} else {
878 		tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
879 	}
880 	GRANT_UNLOCK(log, s);
881 
882 	return tail_lsn;
883 }	/* xlog_assign_tail_lsn */
884 
885 
886 /*
887  * Return the space in the log between the tail and the head.  The head
888  * is passed in the cycle/bytes formal parms.  In the special case where
889  * the reserve head has wrapped passed the tail, this calculation is no
890  * longer valid.  In this case, just return 0 which means there is no space
891  * in the log.  This works for all places where this function is called
892  * with the reserve head.  Of course, if the write head were to ever
893  * wrap the tail, we should blow up.  Rather than catch this case here,
894  * we depend on other ASSERTions in other parts of the code.   XXXmiken
895  *
896  * This code also handles the case where the reservation head is behind
897  * the tail.  The details of this case are described below, but the end
898  * result is that we return the size of the log as the amount of space left.
899  */
900 int
901 xlog_space_left(xlog_t *log, int cycle, int bytes)
902 {
903 	int free_bytes;
904 	int tail_bytes;
905 	int tail_cycle;
906 
907 	tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
908 	tail_cycle = CYCLE_LSN(log->l_tail_lsn);
909 	if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
910 		free_bytes = log->l_logsize - (bytes - tail_bytes);
911 	} else if ((tail_cycle + 1) < cycle) {
912 		return 0;
913 	} else if (tail_cycle < cycle) {
914 		ASSERT(tail_cycle == (cycle - 1));
915 		free_bytes = tail_bytes - bytes;
916 	} else {
917 		/*
918 		 * The reservation head is behind the tail.
919 		 * In this case we just want to return the size of the
920 		 * log as the amount of space left.
921 		 */
922 		xfs_fs_cmn_err(CE_ALERT, log->l_mp,
923 			"xlog_space_left: head behind tail\n"
924 			"  tail_cycle = %d, tail_bytes = %d\n"
925 			"  GH   cycle = %d, GH   bytes = %d",
926 			tail_cycle, tail_bytes, cycle, bytes);
927 		ASSERT(0);
928 		free_bytes = log->l_logsize;
929 	}
930 	return free_bytes;
931 }	/* xlog_space_left */
932 
933 
934 /*
935  * Log function which is called when an io completes.
936  *
937  * The log manager needs its own routine, in order to control what
938  * happens with the buffer after the write completes.
939  */
940 void
941 xlog_iodone(xfs_buf_t *bp)
942 {
943 	xlog_in_core_t	*iclog;
944 	xlog_t		*l;
945 	int		aborted;
946 
947 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
948 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
949 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
950 	aborted = 0;
951 
952 	/*
953 	 * Some versions of cpp barf on the recursive definition of
954 	 * ic_log -> hic_fields.ic_log and expand ic_log twice when
955 	 * it is passed through two macros.  Workaround broken cpp.
956 	 */
957 	l = iclog->ic_log;
958 
959 	/*
960 	 * Race to shutdown the filesystem if we see an error.
961 	 */
962 	if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
963 			XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
964 		xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
965 		XFS_BUF_STALE(bp);
966 		xfs_force_shutdown(l->l_mp, XFS_LOG_IO_ERROR);
967 		/*
968 		 * This flag will be propagated to the trans-committed
969 		 * callback routines to let them know that the log-commit
970 		 * didn't succeed.
971 		 */
972 		aborted = XFS_LI_ABORTED;
973 	} else if (iclog->ic_state & XLOG_STATE_IOERROR) {
974 		aborted = XFS_LI_ABORTED;
975 	}
976 	xlog_state_done_syncing(iclog, aborted);
977 	if (!(XFS_BUF_ISASYNC(bp))) {
978 		/*
979 		 * Corresponding psema() will be done in bwrite().  If we don't
980 		 * vsema() here, panic.
981 		 */
982 		XFS_BUF_V_IODONESEMA(bp);
983 	}
984 }	/* xlog_iodone */
985 
986 /*
987  * The bdstrat callback function for log bufs. This gives us a central
988  * place to trap bufs in case we get hit by a log I/O error and need to
989  * shutdown. Actually, in practice, even when we didn't get a log error,
990  * we transition the iclogs to IOERROR state *after* flushing all existing
991  * iclogs to disk. This is because we don't want anymore new transactions to be
992  * started or completed afterwards.
993  */
994 STATIC int
995 xlog_bdstrat_cb(struct xfs_buf *bp)
996 {
997 	xlog_in_core_t *iclog;
998 
999 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1000 
1001 	if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
1002 	  /* note for irix bstrat will need  struct bdevsw passed
1003 	   * Fix the following macro if the code ever is merged
1004 	   */
1005 	    XFS_bdstrat(bp);
1006 		return 0;
1007 	}
1008 
1009 	xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
1010 	XFS_BUF_ERROR(bp, EIO);
1011 	XFS_BUF_STALE(bp);
1012 	xfs_biodone(bp);
1013 	return (XFS_ERROR(EIO));
1014 
1015 
1016 }
1017 
1018 /*
1019  * Return size of each in-core log record buffer.
1020  *
1021  * Low memory machines only get 2 16KB buffers.  We don't want to waste
1022  * memory here.  However, all other machines get at least 2 32KB buffers.
1023  * The number is hard coded because we don't care about the minimum
1024  * memory size, just 32MB systems.
1025  *
1026  * If the filesystem blocksize is too large, we may need to choose a
1027  * larger size since the directory code currently logs entire blocks.
1028  */
1029 
1030 STATIC void
1031 xlog_get_iclog_buffer_size(xfs_mount_t	*mp,
1032 			   xlog_t	*log)
1033 {
1034 	int size;
1035 	int xhdrs;
1036 
1037 #if defined(DEBUG) || defined(XLOG_NOLOG)
1038 	/*
1039 	 * When logbufs == 0, someone has disabled the log from the FSTAB
1040 	 * file.  This is not a documented feature.  We need to set xlog_debug
1041 	 * to zero (this deactivates the log) and set xlog_target to the
1042 	 * appropriate device.  Only one filesystem may be affected as such
1043 	 * since this is just a performance hack to test what we might be able
1044 	 * to get if the log were not present.
1045 	 */
1046 	if (mp->m_logbufs == 0) {
1047 		xlog_debug = 0;
1048 		xlog_target = log->l_targ;
1049 		log->l_iclog_bufs = XLOG_MIN_ICLOGS;
1050 	} else
1051 #endif
1052 	{
1053 		/*
1054 		 * This is the normal path.  If m_logbufs == -1, then the
1055 		 * admin has chosen to use the system defaults for logbuffers.
1056 		 */
1057 		if (mp->m_logbufs == -1) {
1058 			if (xfs_physmem <= btoc(128*1024*1024)) {
1059 				log->l_iclog_bufs = XLOG_MIN_ICLOGS;
1060 			} else if (xfs_physmem <= btoc(400*1024*1024)) {
1061 				log->l_iclog_bufs = XLOG_MED_ICLOGS;
1062 			} else {
1063 				/* 256K with 32K bufs */
1064 				log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1065 			}
1066 		} else
1067 			log->l_iclog_bufs = mp->m_logbufs;
1068 
1069 #if defined(DEBUG) || defined(XLOG_NOLOG)
1070 		/* We are reactivating a filesystem after it was inactive */
1071 		if (log->l_targ == xlog_target) {
1072 			xlog_target = NULL;
1073 			xlog_debug = 1;
1074 		}
1075 #endif
1076 	}
1077 
1078 	/*
1079 	 * Buffer size passed in from mount system call.
1080 	 */
1081 	if (mp->m_logbsize != -1) {
1082 		size = log->l_iclog_size = mp->m_logbsize;
1083 		log->l_iclog_size_log = 0;
1084 		while (size != 1) {
1085 			log->l_iclog_size_log++;
1086 			size >>= 1;
1087 		}
1088 
1089 		if (XFS_SB_VERSION_HASLOGV2(&mp->m_sb)) {
1090 			/* # headers = size / 32K
1091 			 * one header holds cycles from 32K of data
1092 			 */
1093 
1094 			xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1095 			if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1096 				xhdrs++;
1097 			log->l_iclog_hsize = xhdrs << BBSHIFT;
1098 			log->l_iclog_heads = xhdrs;
1099 		} else {
1100 			ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1101 			log->l_iclog_hsize = BBSIZE;
1102 			log->l_iclog_heads = 1;
1103 		}
1104 		return;
1105 	}
1106 
1107 	/*
1108 	 * Special case machines that have less than 32MB of memory.
1109 	 * All machines with more memory use 32KB buffers.
1110 	 */
1111 	if (xfs_physmem <= btoc(32*1024*1024)) {
1112 		/* Don't change; min configuration */
1113 		log->l_iclog_size = XLOG_RECORD_BSIZE;		/* 16k */
1114 		log->l_iclog_size_log = XLOG_RECORD_BSHIFT;
1115 	} else {
1116 		log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;	/* 32k */
1117 		log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1118 	}
1119 
1120 	/* the default log size is 16k or 32k which is one header sector */
1121 	log->l_iclog_hsize = BBSIZE;
1122 	log->l_iclog_heads = 1;
1123 
1124 	/*
1125 	 * For 16KB, we use 3 32KB buffers.  For 32KB block sizes, we use
1126 	 * 4 32KB buffers.  For 64KB block sizes, we use 8 32KB buffers.
1127 	 */
1128 	if (mp->m_sb.sb_blocksize >= 16*1024) {
1129 		log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1130 		log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1131 		if (mp->m_logbufs == -1) {
1132 			switch (mp->m_sb.sb_blocksize) {
1133 			    case 16*1024:			/* 16 KB */
1134 				log->l_iclog_bufs = 3;
1135 				break;
1136 			    case 32*1024:			/* 32 KB */
1137 				log->l_iclog_bufs = 4;
1138 				break;
1139 			    case 64*1024:			/* 64 KB */
1140 				log->l_iclog_bufs = 8;
1141 				break;
1142 			    default:
1143 				xlog_panic("XFS: Invalid blocksize");
1144 				break;
1145 			}
1146 		}
1147 	}
1148 }	/* xlog_get_iclog_buffer_size */
1149 
1150 
1151 /*
1152  * This routine initializes some of the log structure for a given mount point.
1153  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1154  * some other stuff may be filled in too.
1155  */
1156 STATIC xlog_t *
1157 xlog_alloc_log(xfs_mount_t	*mp,
1158 	       xfs_buftarg_t	*log_target,
1159 	       xfs_daddr_t	blk_offset,
1160 	       int		num_bblks)
1161 {
1162 	xlog_t			*log;
1163 	xlog_rec_header_t	*head;
1164 	xlog_in_core_t		**iclogp;
1165 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1166 	xfs_buf_t		*bp;
1167 	int			i;
1168 	int			iclogsize;
1169 
1170 	log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1171 
1172 	log->l_mp	   = mp;
1173 	log->l_targ	   = log_target;
1174 	log->l_logsize     = BBTOB(num_bblks);
1175 	log->l_logBBstart  = blk_offset;
1176 	log->l_logBBsize   = num_bblks;
1177 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1178 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1179 
1180 	log->l_prev_block  = -1;
1181 	ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, 1, 0);
1182 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1183 	log->l_last_sync_lsn = log->l_tail_lsn;
1184 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1185 	log->l_grant_reserve_cycle = 1;
1186 	log->l_grant_write_cycle = 1;
1187 
1188 	if (XFS_SB_VERSION_HASSECTOR(&mp->m_sb)) {
1189 		log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1190 		ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1191 		/* for larger sector sizes, must have v2 or external log */
1192 		ASSERT(log->l_sectbb_log == 0 ||
1193 			log->l_logBBstart == 0 ||
1194 			XFS_SB_VERSION_HASLOGV2(&mp->m_sb));
1195 		ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1196 	}
1197 	log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1198 
1199 	xlog_get_iclog_buffer_size(mp, log);
1200 
1201 	bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1202 	XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1203 	XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1204 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1205 	ASSERT(XFS_BUF_ISBUSY(bp));
1206 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1207 	log->l_xbuf = bp;
1208 
1209 	spinlock_init(&log->l_icloglock, "iclog");
1210 	spinlock_init(&log->l_grant_lock, "grhead_iclog");
1211 	initnsema(&log->l_flushsema, 0, "ic-flush");
1212 	xlog_state_ticket_alloc(log);  /* wait until after icloglock inited */
1213 
1214 	/* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1215 	ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1216 
1217 	iclogp = &log->l_iclog;
1218 	/*
1219 	 * The amount of memory to allocate for the iclog structure is
1220 	 * rather funky due to the way the structure is defined.  It is
1221 	 * done this way so that we can use different sizes for machines
1222 	 * with different amounts of memory.  See the definition of
1223 	 * xlog_in_core_t in xfs_log_priv.h for details.
1224 	 */
1225 	iclogsize = log->l_iclog_size;
1226 	ASSERT(log->l_iclog_size >= 4096);
1227 	for (i=0; i < log->l_iclog_bufs; i++) {
1228 		*iclogp = (xlog_in_core_t *)
1229 			  kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1230 		iclog = *iclogp;
1231 		iclog->hic_data = (xlog_in_core_2_t *)
1232 			  kmem_zalloc(iclogsize, KM_SLEEP);
1233 
1234 		iclog->ic_prev = prev_iclog;
1235 		prev_iclog = iclog;
1236 		log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1237 
1238 		head = &iclog->ic_header;
1239 		memset(head, 0, sizeof(xlog_rec_header_t));
1240 		INT_SET(head->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1241 		INT_SET(head->h_version, ARCH_CONVERT,
1242 			XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1243 		INT_SET(head->h_size, ARCH_CONVERT, log->l_iclog_size);
1244 		/* new fields */
1245 		INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
1246 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1247 
1248 		bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1249 		XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1250 		XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1251 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1252 		iclog->ic_bp = bp;
1253 
1254 		iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1255 		iclog->ic_state = XLOG_STATE_ACTIVE;
1256 		iclog->ic_log = log;
1257 		iclog->ic_callback_tail = &(iclog->ic_callback);
1258 		iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1259 
1260 		ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1261 		ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1262 		sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1263 		sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1264 
1265 		iclogp = &iclog->ic_next;
1266 	}
1267 	*iclogp = log->l_iclog;			/* complete ring */
1268 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1269 
1270 	return log;
1271 }	/* xlog_alloc_log */
1272 
1273 
1274 /*
1275  * Write out the commit record of a transaction associated with the given
1276  * ticket.  Return the lsn of the commit record.
1277  */
1278 STATIC int
1279 xlog_commit_record(xfs_mount_t  *mp,
1280 		   xlog_ticket_t *ticket,
1281 		   xlog_in_core_t **iclog,
1282 		   xfs_lsn_t	*commitlsnp)
1283 {
1284 	int		error;
1285 	xfs_log_iovec_t	reg[1];
1286 
1287 	reg[0].i_addr = NULL;
1288 	reg[0].i_len = 0;
1289 	XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
1290 
1291 	ASSERT_ALWAYS(iclog);
1292 	if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1293 			       iclog, XLOG_COMMIT_TRANS))) {
1294 		xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
1295 	}
1296 	return (error);
1297 }	/* xlog_commit_record */
1298 
1299 
1300 /*
1301  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1302  * log space.  This code pushes on the lsn which would supposedly free up
1303  * the 25% which we want to leave free.  We may need to adopt a policy which
1304  * pushes on an lsn which is further along in the log once we reach the high
1305  * water mark.  In this manner, we would be creating a low water mark.
1306  */
1307 void
1308 xlog_grant_push_ail(xfs_mount_t	*mp,
1309 		    int		need_bytes)
1310 {
1311     xlog_t	*log = mp->m_log;	/* pointer to the log */
1312     xfs_lsn_t	tail_lsn;		/* lsn of the log tail */
1313     xfs_lsn_t	threshold_lsn = 0;	/* lsn we'd like to be at */
1314     int		free_blocks;		/* free blocks left to write to */
1315     int		free_bytes;		/* free bytes left to write to */
1316     int		threshold_block;	/* block in lsn we'd like to be at */
1317     int		threshold_cycle;	/* lsn cycle we'd like to be at */
1318     int		free_threshold;
1319     SPLDECL(s);
1320 
1321     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1322 
1323     s = GRANT_LOCK(log);
1324     free_bytes = xlog_space_left(log,
1325 				 log->l_grant_reserve_cycle,
1326 				 log->l_grant_reserve_bytes);
1327     tail_lsn = log->l_tail_lsn;
1328     free_blocks = BTOBBT(free_bytes);
1329 
1330     /*
1331      * Set the threshold for the minimum number of free blocks in the
1332      * log to the maximum of what the caller needs, one quarter of the
1333      * log, and 256 blocks.
1334      */
1335     free_threshold = BTOBB(need_bytes);
1336     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1337     free_threshold = MAX(free_threshold, 256);
1338     if (free_blocks < free_threshold) {
1339 	threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1340 	threshold_cycle = CYCLE_LSN(tail_lsn);
1341 	if (threshold_block >= log->l_logBBsize) {
1342 	    threshold_block -= log->l_logBBsize;
1343 	    threshold_cycle += 1;
1344 	}
1345 	ASSIGN_ANY_LSN_HOST(threshold_lsn, threshold_cycle,
1346 		       threshold_block);
1347 
1348 	/* Don't pass in an lsn greater than the lsn of the last
1349 	 * log record known to be on disk.
1350 	 */
1351 	if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1352 	    threshold_lsn = log->l_last_sync_lsn;
1353     }
1354     GRANT_UNLOCK(log, s);
1355 
1356     /*
1357      * Get the transaction layer to kick the dirty buffers out to
1358      * disk asynchronously. No point in trying to do this if
1359      * the filesystem is shutting down.
1360      */
1361     if (threshold_lsn &&
1362 	!XLOG_FORCED_SHUTDOWN(log))
1363 	    xfs_trans_push_ail(mp, threshold_lsn);
1364 }	/* xlog_grant_push_ail */
1365 
1366 
1367 /*
1368  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1369  * fashion.  Previously, we should have moved the current iclog
1370  * ptr in the log to point to the next available iclog.  This allows further
1371  * write to continue while this code syncs out an iclog ready to go.
1372  * Before an in-core log can be written out, the data section must be scanned
1373  * to save away the 1st word of each BBSIZE block into the header.  We replace
1374  * it with the current cycle count.  Each BBSIZE block is tagged with the
1375  * cycle count because there in an implicit assumption that drives will
1376  * guarantee that entire 512 byte blocks get written at once.  In other words,
1377  * we can't have part of a 512 byte block written and part not written.  By
1378  * tagging each block, we will know which blocks are valid when recovering
1379  * after an unclean shutdown.
1380  *
1381  * This routine is single threaded on the iclog.  No other thread can be in
1382  * this routine with the same iclog.  Changing contents of iclog can there-
1383  * fore be done without grabbing the state machine lock.  Updating the global
1384  * log will require grabbing the lock though.
1385  *
1386  * The entire log manager uses a logical block numbering scheme.  Only
1387  * log_sync (and then only bwrite()) know about the fact that the log may
1388  * not start with block zero on a given device.  The log block start offset
1389  * is added immediately before calling bwrite().
1390  */
1391 
1392 int
1393 xlog_sync(xlog_t		*log,
1394 	  xlog_in_core_t	*iclog)
1395 {
1396 	xfs_caddr_t	dptr;		/* pointer to byte sized element */
1397 	xfs_buf_t	*bp;
1398 	int		i, ops;
1399 	uint		count;		/* byte count of bwrite */
1400 	uint		count_init;	/* initial count before roundup */
1401 	int		roundoff;       /* roundoff to BB or stripe */
1402 	int		split = 0;	/* split write into two regions */
1403 	int		error;
1404 	SPLDECL(s);
1405 	int		v2 = XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb);
1406 
1407 	XFS_STATS_INC(xs_log_writes);
1408 	ASSERT(iclog->ic_refcnt == 0);
1409 
1410 	/* Add for LR header */
1411 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1412 
1413 	/* Round out the log write size */
1414 	if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1415 		/* we have a v2 stripe unit to use */
1416 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1417 	} else {
1418 		count = BBTOB(BTOBB(count_init));
1419 	}
1420 	roundoff = count - count_init;
1421 	ASSERT(roundoff >= 0);
1422 	ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1423                 roundoff < log->l_mp->m_sb.sb_logsunit)
1424 		||
1425 		(log->l_mp->m_sb.sb_logsunit <= 1 &&
1426 		 roundoff < BBTOB(1)));
1427 
1428 	/* move grant heads by roundoff in sync */
1429 	s = GRANT_LOCK(log);
1430 	XLOG_GRANT_ADD_SPACE(log, roundoff, 'w');
1431 	XLOG_GRANT_ADD_SPACE(log, roundoff, 'r');
1432 	GRANT_UNLOCK(log, s);
1433 
1434 	/* put cycle number in every block */
1435 	xlog_pack_data(log, iclog, roundoff);
1436 
1437 	/* real byte length */
1438 	if (v2) {
1439 		INT_SET(iclog->ic_header.h_len,
1440 			ARCH_CONVERT,
1441 			iclog->ic_offset + roundoff);
1442 	} else {
1443 		INT_SET(iclog->ic_header.h_len, ARCH_CONVERT, iclog->ic_offset);
1444 	}
1445 
1446 	/* put ops count in correct order */
1447 	ops = iclog->ic_header.h_num_logops;
1448 	INT_SET(iclog->ic_header.h_num_logops, ARCH_CONVERT, ops);
1449 
1450 	bp	    = iclog->ic_bp;
1451 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1452 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1453 	XFS_BUF_SET_ADDR(bp, BLOCK_LSN(INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)));
1454 
1455 	XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1456 
1457 	/* Do we need to split this write into 2 parts? */
1458 	if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1459 		split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1460 		count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1461 		iclog->ic_bwritecnt = 2;	/* split into 2 writes */
1462 	} else {
1463 		iclog->ic_bwritecnt = 1;
1464 	}
1465 	XFS_BUF_SET_PTR(bp, (xfs_caddr_t) &(iclog->ic_header), count);
1466 	XFS_BUF_SET_FSPRIVATE(bp, iclog);	/* save for later */
1467 	XFS_BUF_BUSY(bp);
1468 	XFS_BUF_ASYNC(bp);
1469 	/*
1470 	 * Do a disk write cache flush for the log block.
1471 	 * This is a bit of a sledgehammer, it would be better
1472 	 * to use a tag barrier here that just prevents reordering.
1473 	 * It may not be needed to flush the first split block in the log wrap
1474 	 * case, but do it anyways to be safe -AK
1475 	 */
1476 	if (!(log->l_mp->m_flags & XFS_MOUNT_NOLOGFLUSH))
1477 		XFS_BUF_FLUSH(bp);
1478 
1479 	ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1480 	ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1481 
1482 	xlog_verify_iclog(log, iclog, count, B_TRUE);
1483 
1484 	/* account for log which doesn't start at block #0 */
1485 	XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1486 	/*
1487 	 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1488 	 * is shutting down.
1489 	 */
1490 	XFS_BUF_WRITE(bp);
1491 
1492 	if ((error = XFS_bwrite(bp))) {
1493 		xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1494 				  XFS_BUF_ADDR(bp));
1495 		return (error);
1496 	}
1497 	if (split) {
1498 		bp		= iclog->ic_log->l_xbuf;
1499 		ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1500 							(unsigned long)1);
1501 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1502 		XFS_BUF_SET_ADDR(bp, 0);	     /* logical 0 */
1503 		XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1504 					    (__psint_t)count), split);
1505 		XFS_BUF_SET_FSPRIVATE(bp, iclog);
1506 		XFS_BUF_BUSY(bp);
1507 		XFS_BUF_ASYNC(bp);
1508 		if (!(log->l_mp->m_flags & XFS_MOUNT_NOLOGFLUSH))
1509 			XFS_BUF_FLUSH(bp);
1510 		dptr = XFS_BUF_PTR(bp);
1511 		/*
1512 		 * Bump the cycle numbers at the start of each block
1513 		 * since this part of the buffer is at the start of
1514 		 * a new cycle.  Watch out for the header magic number
1515 		 * case, though.
1516 		 */
1517 		for (i=0; i<split; i += BBSIZE) {
1518 			INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1519 			if (INT_GET(*(uint *)dptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
1520 				INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1521 			dptr += BBSIZE;
1522 		}
1523 
1524 		ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1525 		ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1526 
1527 		/* account for internal log which does't start at block #0 */
1528 		XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1529 		XFS_BUF_WRITE(bp);
1530 		if ((error = XFS_bwrite(bp))) {
1531 			xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1532 					  bp, XFS_BUF_ADDR(bp));
1533 			return (error);
1534 		}
1535 	}
1536 	return (0);
1537 }	/* xlog_sync */
1538 
1539 
1540 /*
1541  * Unallocate a log structure
1542  */
1543 void
1544 xlog_unalloc_log(xlog_t *log)
1545 {
1546 	xlog_in_core_t	*iclog, *next_iclog;
1547 	xlog_ticket_t	*tic, *next_tic;
1548 	int		i;
1549 
1550 
1551 	iclog = log->l_iclog;
1552 	for (i=0; i<log->l_iclog_bufs; i++) {
1553 		sv_destroy(&iclog->ic_forcesema);
1554 		sv_destroy(&iclog->ic_writesema);
1555 		xfs_buf_free(iclog->ic_bp);
1556 #ifdef XFS_LOG_TRACE
1557 		if (iclog->ic_trace != NULL) {
1558 			ktrace_free(iclog->ic_trace);
1559 		}
1560 #endif
1561 		next_iclog = iclog->ic_next;
1562 		kmem_free(iclog->hic_data, log->l_iclog_size);
1563 		kmem_free(iclog, sizeof(xlog_in_core_t));
1564 		iclog = next_iclog;
1565 	}
1566 	freesema(&log->l_flushsema);
1567 	spinlock_destroy(&log->l_icloglock);
1568 	spinlock_destroy(&log->l_grant_lock);
1569 
1570 	/* XXXsup take a look at this again. */
1571 	if ((log->l_ticket_cnt != log->l_ticket_tcnt)  &&
1572 	    !XLOG_FORCED_SHUTDOWN(log)) {
1573 		xfs_fs_cmn_err(CE_WARN, log->l_mp,
1574 			"xlog_unalloc_log: (cnt: %d, total: %d)",
1575 			log->l_ticket_cnt, log->l_ticket_tcnt);
1576 		/* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
1577 
1578 	} else {
1579 		tic = log->l_unmount_free;
1580 		while (tic) {
1581 			next_tic = tic->t_next;
1582 			kmem_free(tic, NBPP);
1583 			tic = next_tic;
1584 		}
1585 	}
1586 	xfs_buf_free(log->l_xbuf);
1587 #ifdef XFS_LOG_TRACE
1588 	if (log->l_trace != NULL) {
1589 		ktrace_free(log->l_trace);
1590 	}
1591 	if (log->l_grant_trace != NULL) {
1592 		ktrace_free(log->l_grant_trace);
1593 	}
1594 #endif
1595 	log->l_mp->m_log = NULL;
1596 	kmem_free(log, sizeof(xlog_t));
1597 }	/* xlog_unalloc_log */
1598 
1599 /*
1600  * Update counters atomically now that memcpy is done.
1601  */
1602 /* ARGSUSED */
1603 static inline void
1604 xlog_state_finish_copy(xlog_t		*log,
1605 		       xlog_in_core_t	*iclog,
1606 		       int		record_cnt,
1607 		       int		copy_bytes)
1608 {
1609 	SPLDECL(s);
1610 
1611 	s = LOG_LOCK(log);
1612 
1613 	iclog->ic_header.h_num_logops += record_cnt;
1614 	iclog->ic_offset += copy_bytes;
1615 
1616 	LOG_UNLOCK(log, s);
1617 }	/* xlog_state_finish_copy */
1618 
1619 
1620 
1621 
1622 /*
1623  * print out info relating to regions written which consume
1624  * the reservation
1625  */
1626 #if defined(XFS_LOG_RES_DEBUG)
1627 STATIC void
1628 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1629 {
1630 	uint i;
1631 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1632 
1633 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
1634 	static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1635 	    "bformat",
1636 	    "bchunk",
1637 	    "efi_format",
1638 	    "efd_format",
1639 	    "iformat",
1640 	    "icore",
1641 	    "iext",
1642 	    "ibroot",
1643 	    "ilocal",
1644 	    "iattr_ext",
1645 	    "iattr_broot",
1646 	    "iattr_local",
1647 	    "qformat",
1648 	    "dquot",
1649 	    "quotaoff",
1650 	    "LR header",
1651 	    "unmount",
1652 	    "commit",
1653 	    "trans header"
1654 	};
1655 	static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1656 	    "SETATTR_NOT_SIZE",
1657 	    "SETATTR_SIZE",
1658 	    "INACTIVE",
1659 	    "CREATE",
1660 	    "CREATE_TRUNC",
1661 	    "TRUNCATE_FILE",
1662 	    "REMOVE",
1663 	    "LINK",
1664 	    "RENAME",
1665 	    "MKDIR",
1666 	    "RMDIR",
1667 	    "SYMLINK",
1668 	    "SET_DMATTRS",
1669 	    "GROWFS",
1670 	    "STRAT_WRITE",
1671 	    "DIOSTRAT",
1672 	    "WRITE_SYNC",
1673 	    "WRITEID",
1674 	    "ADDAFORK",
1675 	    "ATTRINVAL",
1676 	    "ATRUNCATE",
1677 	    "ATTR_SET",
1678 	    "ATTR_RM",
1679 	    "ATTR_FLAG",
1680 	    "CLEAR_AGI_BUCKET",
1681 	    "QM_SBCHANGE",
1682 	    "DUMMY1",
1683 	    "DUMMY2",
1684 	    "QM_QUOTAOFF",
1685 	    "QM_DQALLOC",
1686 	    "QM_SETQLIM",
1687 	    "QM_DQCLUSTER",
1688 	    "QM_QINOCREATE",
1689 	    "QM_QUOTAOFF_END",
1690 	    "SB_UNIT",
1691 	    "FSYNC_TS",
1692 	    "GROWFSRT_ALLOC",
1693 	    "GROWFSRT_ZERO",
1694 	    "GROWFSRT_FREE",
1695 	    "SWAPEXT"
1696 	};
1697 
1698 	xfs_fs_cmn_err(CE_WARN, mp,
1699 			"xfs_log_write: reservation summary:\n"
1700 			"  trans type  = %s (%u)\n"
1701 			"  unit res    = %d bytes\n"
1702 			"  current res = %d bytes\n"
1703 			"  total reg   = %u bytes (o/flow = %u bytes)\n"
1704 			"  ophdrs      = %u (ophdr space = %u bytes)\n"
1705 			"  ophdr + reg = %u bytes\n"
1706 			"  num regions = %u\n",
1707 			((ticket->t_trans_type <= 0 ||
1708 			  ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1709 			  "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1710 			ticket->t_trans_type,
1711 			ticket->t_unit_res,
1712 			ticket->t_curr_res,
1713 			ticket->t_res_arr_sum, ticket->t_res_o_flow,
1714 			ticket->t_res_num_ophdrs, ophdr_spc,
1715 			ticket->t_res_arr_sum +
1716 			  ticket->t_res_o_flow + ophdr_spc,
1717 			ticket->t_res_num);
1718 
1719 	for (i = 0; i < ticket->t_res_num; i++) {
1720 	   	uint r_type = ticket->t_res_arr[i].r_type;
1721 		cmn_err(CE_WARN,
1722 			    "region[%u]: %s - %u bytes\n",
1723 			    i,
1724 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1725 			    "bad-rtype" : res_type_str[r_type-1]),
1726 			    ticket->t_res_arr[i].r_len);
1727 	}
1728 }
1729 #else
1730 #define xlog_print_tic_res(mp, ticket)
1731 #endif
1732 
1733 /*
1734  * Write some region out to in-core log
1735  *
1736  * This will be called when writing externally provided regions or when
1737  * writing out a commit record for a given transaction.
1738  *
1739  * General algorithm:
1740  *	1. Find total length of this write.  This may include adding to the
1741  *		lengths passed in.
1742  *	2. Check whether we violate the tickets reservation.
1743  *	3. While writing to this iclog
1744  *	    A. Reserve as much space in this iclog as can get
1745  *	    B. If this is first write, save away start lsn
1746  *	    C. While writing this region:
1747  *		1. If first write of transaction, write start record
1748  *		2. Write log operation header (header per region)
1749  *		3. Find out if we can fit entire region into this iclog
1750  *		4. Potentially, verify destination memcpy ptr
1751  *		5. Memcpy (partial) region
1752  *		6. If partial copy, release iclog; otherwise, continue
1753  *			copying more regions into current iclog
1754  *	4. Mark want sync bit (in simulation mode)
1755  *	5. Release iclog for potential flush to on-disk log.
1756  *
1757  * ERRORS:
1758  * 1.	Panic if reservation is overrun.  This should never happen since
1759  *	reservation amounts are generated internal to the filesystem.
1760  * NOTES:
1761  * 1. Tickets are single threaded data structures.
1762  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1763  *	syncing routine.  When a single log_write region needs to span
1764  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1765  *	on all log operation writes which don't contain the end of the
1766  *	region.  The XLOG_END_TRANS bit is used for the in-core log
1767  *	operation which contains the end of the continued log_write region.
1768  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1769  *	we don't really know exactly how much space will be used.  As a result,
1770  *	we don't update ic_offset until the end when we know exactly how many
1771  *	bytes have been written out.
1772  */
1773 int
1774 xlog_write(xfs_mount_t *	mp,
1775 	   xfs_log_iovec_t	reg[],
1776 	   int			nentries,
1777 	   xfs_log_ticket_t	tic,
1778 	   xfs_lsn_t		*start_lsn,
1779 	   xlog_in_core_t	**commit_iclog,
1780 	   uint			flags)
1781 {
1782     xlog_t	     *log    = mp->m_log;
1783     xlog_ticket_t    *ticket = (xlog_ticket_t *)tic;
1784     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1785     xlog_in_core_t   *iclog;	     /* ptr to current in-core log */
1786     __psint_t	     ptr;	     /* copy address into data region */
1787     int		     len;	     /* # xlog_write() bytes 2 still copy */
1788     int		     index;	     /* region index currently copying */
1789     int		     log_offset;     /* offset (from 0) into data region */
1790     int		     start_rec_copy; /* # bytes to copy for start record */
1791     int		     partial_copy;   /* did we split a region? */
1792     int		     partial_copy_len;/* # bytes copied if split region */
1793     int		     need_copy;	     /* # bytes need to memcpy this region */
1794     int		     copy_len;	     /* # bytes actually memcpy'ing */
1795     int		     copy_off;	     /* # bytes from entry start */
1796     int		     contwr;	     /* continued write of in-core log? */
1797     int		     error;
1798     int		     record_cnt = 0, data_cnt = 0;
1799 
1800     partial_copy_len = partial_copy = 0;
1801 
1802     /* Calculate potential maximum space.  Each region gets its own
1803      * xlog_op_header_t and may need to be double word aligned.
1804      */
1805     len = 0;
1806     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1807 	len += sizeof(xlog_op_header_t);
1808 	XLOG_TIC_ADD_OPHDR(ticket);
1809     }
1810 
1811     for (index = 0; index < nentries; index++) {
1812 	len += sizeof(xlog_op_header_t);	    /* each region gets >= 1 */
1813 	XLOG_TIC_ADD_OPHDR(ticket);
1814 	len += reg[index].i_len;
1815 	XLOG_TIC_ADD_REGION(ticket, reg[index].i_len, reg[index].i_type);
1816     }
1817     contwr = *start_lsn = 0;
1818 
1819     if (ticket->t_curr_res < len) {
1820 	xlog_print_tic_res(mp, ticket);
1821 #ifdef DEBUG
1822 	xlog_panic(
1823 		"xfs_log_write: reservation ran out. Need to up reservation");
1824 #else
1825 	/* Customer configurable panic */
1826 	xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1827 		"xfs_log_write: reservation ran out. Need to up reservation");
1828 	/* If we did not panic, shutdown the filesystem */
1829 	xfs_force_shutdown(mp, XFS_CORRUPT_INCORE);
1830 #endif
1831     } else
1832 	ticket->t_curr_res -= len;
1833 
1834     for (index = 0; index < nentries; ) {
1835 	if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1836 					       &contwr, &log_offset)))
1837 		return (error);
1838 
1839 	ASSERT(log_offset <= iclog->ic_size - 1);
1840 	ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1841 
1842 	/* start_lsn is the first lsn written to. That's all we need. */
1843 	if (! *start_lsn)
1844 	    *start_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
1845 
1846 	/* This loop writes out as many regions as can fit in the amount
1847 	 * of space which was allocated by xlog_state_get_iclog_space().
1848 	 */
1849 	while (index < nentries) {
1850 	    ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1851 	    ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1852 	    start_rec_copy = 0;
1853 
1854 	    /* If first write for transaction, insert start record.
1855 	     * We can't be trying to commit if we are inited.  We can't
1856 	     * have any "partial_copy" if we are inited.
1857 	     */
1858 	    if (ticket->t_flags & XLOG_TIC_INITED) {
1859 		logop_head		= (xlog_op_header_t *)ptr;
1860 		INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1861 		logop_head->oh_clientid = ticket->t_clientid;
1862 		logop_head->oh_len	= 0;
1863 		logop_head->oh_flags    = XLOG_START_TRANS;
1864 		logop_head->oh_res2	= 0;
1865 		ticket->t_flags		&= ~XLOG_TIC_INITED;	/* clear bit */
1866 		record_cnt++;
1867 
1868 		start_rec_copy = sizeof(xlog_op_header_t);
1869 		xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1870 	    }
1871 
1872 	    /* Copy log operation header directly into data section */
1873 	    logop_head			= (xlog_op_header_t *)ptr;
1874 	    INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1875 	    logop_head->oh_clientid	= ticket->t_clientid;
1876 	    logop_head->oh_res2		= 0;
1877 
1878 	    /* header copied directly */
1879 	    xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1880 
1881 	    /* are we copying a commit or unmount record? */
1882 	    logop_head->oh_flags = flags;
1883 
1884 	    /*
1885 	     * We've seen logs corrupted with bad transaction client
1886 	     * ids.  This makes sure that XFS doesn't generate them on.
1887 	     * Turn this into an EIO and shut down the filesystem.
1888 	     */
1889 	    switch (logop_head->oh_clientid)  {
1890 	    case XFS_TRANSACTION:
1891 	    case XFS_VOLUME:
1892 	    case XFS_LOG:
1893 		break;
1894 	    default:
1895 		xfs_fs_cmn_err(CE_WARN, mp,
1896 		    "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1897 		    logop_head->oh_clientid, tic);
1898 		return XFS_ERROR(EIO);
1899 	    }
1900 
1901 	    /* Partial write last time? => (partial_copy != 0)
1902 	     * need_copy is the amount we'd like to copy if everything could
1903 	     * fit in the current memcpy.
1904 	     */
1905 	    need_copy =	reg[index].i_len - partial_copy_len;
1906 
1907 	    copy_off = partial_copy_len;
1908 	    if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1909 		INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len = need_copy);
1910 		if (partial_copy)
1911 		    logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1912 		partial_copy_len = partial_copy = 0;
1913 	    } else {					    /* partial write */
1914 		copy_len = iclog->ic_size - log_offset;
1915 		INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len);
1916 		logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1917 		if (partial_copy)
1918 			logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1919 		partial_copy_len += copy_len;
1920 		partial_copy++;
1921 		len += sizeof(xlog_op_header_t); /* from splitting of region */
1922 		/* account for new log op header */
1923 		ticket->t_curr_res -= sizeof(xlog_op_header_t);
1924 		XLOG_TIC_ADD_OPHDR(ticket);
1925 	    }
1926 	    xlog_verify_dest_ptr(log, ptr);
1927 
1928 	    /* copy region */
1929 	    ASSERT(copy_len >= 0);
1930 	    memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1931 	    xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1932 
1933 	    /* make copy_len total bytes copied, including headers */
1934 	    copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1935 	    record_cnt++;
1936 	    data_cnt += contwr ? copy_len : 0;
1937 	    if (partial_copy) {			/* copied partial region */
1938 		    /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1939 		    xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1940 		    record_cnt = data_cnt = 0;
1941 		    if ((error = xlog_state_release_iclog(log, iclog)))
1942 			    return (error);
1943 		    break;			/* don't increment index */
1944 	    } else {				/* copied entire region */
1945 		index++;
1946 		partial_copy_len = partial_copy = 0;
1947 
1948 		if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1949 		    xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1950 		    record_cnt = data_cnt = 0;
1951 		    xlog_state_want_sync(log, iclog);
1952 		    if (commit_iclog) {
1953 			ASSERT(flags & XLOG_COMMIT_TRANS);
1954 			*commit_iclog = iclog;
1955 		    } else if ((error = xlog_state_release_iclog(log, iclog)))
1956 			   return (error);
1957 		    if (index == nentries)
1958 			    return 0;		/* we are done */
1959 		    else
1960 			    break;
1961 		}
1962 	    } /* if (partial_copy) */
1963 	} /* while (index < nentries) */
1964     } /* for (index = 0; index < nentries; ) */
1965     ASSERT(len == 0);
1966 
1967     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1968     if (commit_iclog) {
1969 	ASSERT(flags & XLOG_COMMIT_TRANS);
1970 	*commit_iclog = iclog;
1971 	return 0;
1972     }
1973     return (xlog_state_release_iclog(log, iclog));
1974 }	/* xlog_write */
1975 
1976 
1977 /*****************************************************************************
1978  *
1979  *		State Machine functions
1980  *
1981  *****************************************************************************
1982  */
1983 
1984 /* Clean iclogs starting from the head.  This ordering must be
1985  * maintained, so an iclog doesn't become ACTIVE beyond one that
1986  * is SYNCING.  This is also required to maintain the notion that we use
1987  * a counting semaphore to hold off would be writers to the log when every
1988  * iclog is trying to sync to disk.
1989  *
1990  * State Change: DIRTY -> ACTIVE
1991  */
1992 STATIC void
1993 xlog_state_clean_log(xlog_t *log)
1994 {
1995 	xlog_in_core_t	*iclog;
1996 	int changed = 0;
1997 
1998 	iclog = log->l_iclog;
1999 	do {
2000 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
2001 			iclog->ic_state	= XLOG_STATE_ACTIVE;
2002 			iclog->ic_offset       = 0;
2003 			iclog->ic_callback	= NULL;   /* don't need to free */
2004 			/*
2005 			 * If the number of ops in this iclog indicate it just
2006 			 * contains the dummy transaction, we can
2007 			 * change state into IDLE (the second time around).
2008 			 * Otherwise we should change the state into
2009 			 * NEED a dummy.
2010 			 * We don't need to cover the dummy.
2011 			 */
2012 			if (!changed &&
2013 			   (INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT) == XLOG_COVER_OPS)) {
2014 				changed = 1;
2015 			} else {
2016 				/*
2017 				 * We have two dirty iclogs so start over
2018 				 * This could also be num of ops indicates
2019 				 * this is not the dummy going out.
2020 				 */
2021 				changed = 2;
2022 			}
2023 			iclog->ic_header.h_num_logops = 0;
2024 			memset(iclog->ic_header.h_cycle_data, 0,
2025 			      sizeof(iclog->ic_header.h_cycle_data));
2026 			iclog->ic_header.h_lsn = 0;
2027 		} else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2028 			/* do nothing */;
2029 		else
2030 			break;	/* stop cleaning */
2031 		iclog = iclog->ic_next;
2032 	} while (iclog != log->l_iclog);
2033 
2034 	/* log is locked when we are called */
2035 	/*
2036 	 * Change state for the dummy log recording.
2037 	 * We usually go to NEED. But we go to NEED2 if the changed indicates
2038 	 * we are done writing the dummy record.
2039 	 * If we are done with the second dummy recored (DONE2), then
2040 	 * we go to IDLE.
2041 	 */
2042 	if (changed) {
2043 		switch (log->l_covered_state) {
2044 		case XLOG_STATE_COVER_IDLE:
2045 		case XLOG_STATE_COVER_NEED:
2046 		case XLOG_STATE_COVER_NEED2:
2047 			log->l_covered_state = XLOG_STATE_COVER_NEED;
2048 			break;
2049 
2050 		case XLOG_STATE_COVER_DONE:
2051 			if (changed == 1)
2052 				log->l_covered_state = XLOG_STATE_COVER_NEED2;
2053 			else
2054 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2055 			break;
2056 
2057 		case XLOG_STATE_COVER_DONE2:
2058 			if (changed == 1)
2059 				log->l_covered_state = XLOG_STATE_COVER_IDLE;
2060 			else
2061 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2062 			break;
2063 
2064 		default:
2065 			ASSERT(0);
2066 		}
2067 	}
2068 }	/* xlog_state_clean_log */
2069 
2070 STATIC xfs_lsn_t
2071 xlog_get_lowest_lsn(
2072 	xlog_t		*log)
2073 {
2074 	xlog_in_core_t  *lsn_log;
2075 	xfs_lsn_t	lowest_lsn, lsn;
2076 
2077 	lsn_log = log->l_iclog;
2078 	lowest_lsn = 0;
2079 	do {
2080 	    if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2081 		lsn = INT_GET(lsn_log->ic_header.h_lsn, ARCH_CONVERT);
2082 		if ((lsn && !lowest_lsn) ||
2083 		    (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2084 			lowest_lsn = lsn;
2085 		}
2086 	    }
2087 	    lsn_log = lsn_log->ic_next;
2088 	} while (lsn_log != log->l_iclog);
2089 	return(lowest_lsn);
2090 }
2091 
2092 
2093 STATIC void
2094 xlog_state_do_callback(
2095 	xlog_t		*log,
2096 	int		aborted,
2097 	xlog_in_core_t	*ciclog)
2098 {
2099 	xlog_in_core_t	   *iclog;
2100 	xlog_in_core_t	   *first_iclog;	/* used to know when we've
2101 						 * processed all iclogs once */
2102 	xfs_log_callback_t *cb, *cb_next;
2103 	int		   flushcnt = 0;
2104 	xfs_lsn_t	   lowest_lsn;
2105 	int		   ioerrors;	/* counter: iclogs with errors */
2106 	int		   loopdidcallbacks; /* flag: inner loop did callbacks*/
2107 	int		   funcdidcallbacks; /* flag: function did callbacks */
2108 	int		   repeats;	/* for issuing console warnings if
2109 					 * looping too many times */
2110 	SPLDECL(s);
2111 
2112 	s = LOG_LOCK(log);
2113 	first_iclog = iclog = log->l_iclog;
2114 	ioerrors = 0;
2115 	funcdidcallbacks = 0;
2116 	repeats = 0;
2117 
2118 	do {
2119 		/*
2120 		 * Scan all iclogs starting with the one pointed to by the
2121 		 * log.  Reset this starting point each time the log is
2122 		 * unlocked (during callbacks).
2123 		 *
2124 		 * Keep looping through iclogs until one full pass is made
2125 		 * without running any callbacks.
2126 		 */
2127 		first_iclog = log->l_iclog;
2128 		iclog = log->l_iclog;
2129 		loopdidcallbacks = 0;
2130 		repeats++;
2131 
2132 		do {
2133 
2134 			/* skip all iclogs in the ACTIVE & DIRTY states */
2135 			if (iclog->ic_state &
2136 			    (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2137 				iclog = iclog->ic_next;
2138 				continue;
2139 			}
2140 
2141 			/*
2142 			 * Between marking a filesystem SHUTDOWN and stopping
2143 			 * the log, we do flush all iclogs to disk (if there
2144 			 * wasn't a log I/O error). So, we do want things to
2145 			 * go smoothly in case of just a SHUTDOWN  w/o a
2146 			 * LOG_IO_ERROR.
2147 			 */
2148 			if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2149 				/*
2150 				 * Can only perform callbacks in order.  Since
2151 				 * this iclog is not in the DONE_SYNC/
2152 				 * DO_CALLBACK state, we skip the rest and
2153 				 * just try to clean up.  If we set our iclog
2154 				 * to DO_CALLBACK, we will not process it when
2155 				 * we retry since a previous iclog is in the
2156 				 * CALLBACK and the state cannot change since
2157 				 * we are holding the LOG_LOCK.
2158 				 */
2159 				if (!(iclog->ic_state &
2160 					(XLOG_STATE_DONE_SYNC |
2161 						 XLOG_STATE_DO_CALLBACK))) {
2162 					if (ciclog && (ciclog->ic_state ==
2163 							XLOG_STATE_DONE_SYNC)) {
2164 						ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2165 					}
2166 					break;
2167 				}
2168 				/*
2169 				 * We now have an iclog that is in either the
2170 				 * DO_CALLBACK or DONE_SYNC states. The other
2171 				 * states (WANT_SYNC, SYNCING, or CALLBACK were
2172 				 * caught by the above if and are going to
2173 				 * clean (i.e. we aren't doing their callbacks)
2174 				 * see the above if.
2175 				 */
2176 
2177 				/*
2178 				 * We will do one more check here to see if we
2179 				 * have chased our tail around.
2180 				 */
2181 
2182 				lowest_lsn = xlog_get_lowest_lsn(log);
2183 				if (lowest_lsn && (
2184 					XFS_LSN_CMP(
2185 						lowest_lsn,
2186 						INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2187 					)<0)) {
2188 					iclog = iclog->ic_next;
2189 					continue; /* Leave this iclog for
2190 						   * another thread */
2191 				}
2192 
2193 				iclog->ic_state = XLOG_STATE_CALLBACK;
2194 
2195 				LOG_UNLOCK(log, s);
2196 
2197 				/* l_last_sync_lsn field protected by
2198 				 * GRANT_LOCK. Don't worry about iclog's lsn.
2199 				 * No one else can be here except us.
2200 				 */
2201 				s = GRANT_LOCK(log);
2202 				ASSERT(XFS_LSN_CMP(
2203 						log->l_last_sync_lsn,
2204 						INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2205 					)<=0);
2206 				log->l_last_sync_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2207 				GRANT_UNLOCK(log, s);
2208 
2209 				/*
2210 				 * Keep processing entries in the callback list
2211 				 * until we come around and it is empty.  We
2212 				 * need to atomically see that the list is
2213 				 * empty and change the state to DIRTY so that
2214 				 * we don't miss any more callbacks being added.
2215 				 */
2216 				s = LOG_LOCK(log);
2217 			} else {
2218 				ioerrors++;
2219 			}
2220 			cb = iclog->ic_callback;
2221 
2222 			while (cb != 0) {
2223 				iclog->ic_callback_tail = &(iclog->ic_callback);
2224 				iclog->ic_callback = NULL;
2225 				LOG_UNLOCK(log, s);
2226 
2227 				/* perform callbacks in the order given */
2228 				for (; cb != 0; cb = cb_next) {
2229 					cb_next = cb->cb_next;
2230 					cb->cb_func(cb->cb_arg, aborted);
2231 				}
2232 				s = LOG_LOCK(log);
2233 				cb = iclog->ic_callback;
2234 			}
2235 
2236 			loopdidcallbacks++;
2237 			funcdidcallbacks++;
2238 
2239 			ASSERT(iclog->ic_callback == 0);
2240 			if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2241 				iclog->ic_state = XLOG_STATE_DIRTY;
2242 
2243 			/*
2244 			 * Transition from DIRTY to ACTIVE if applicable.
2245 			 * NOP if STATE_IOERROR.
2246 			 */
2247 			xlog_state_clean_log(log);
2248 
2249 			/* wake up threads waiting in xfs_log_force() */
2250 			sv_broadcast(&iclog->ic_forcesema);
2251 
2252 			iclog = iclog->ic_next;
2253 		} while (first_iclog != iclog);
2254 		if (repeats && (repeats % 10) == 0) {
2255 			xfs_fs_cmn_err(CE_WARN, log->l_mp,
2256 				"xlog_state_do_callback: looping %d", repeats);
2257 		}
2258 	} while (!ioerrors && loopdidcallbacks);
2259 
2260 	/*
2261 	 * make one last gasp attempt to see if iclogs are being left in
2262 	 * limbo..
2263 	 */
2264 #ifdef DEBUG
2265 	if (funcdidcallbacks) {
2266 		first_iclog = iclog = log->l_iclog;
2267 		do {
2268 			ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2269 			/*
2270 			 * Terminate the loop if iclogs are found in states
2271 			 * which will cause other threads to clean up iclogs.
2272 			 *
2273 			 * SYNCING - i/o completion will go through logs
2274 			 * DONE_SYNC - interrupt thread should be waiting for
2275 			 *              LOG_LOCK
2276 			 * IOERROR - give up hope all ye who enter here
2277 			 */
2278 			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2279 			    iclog->ic_state == XLOG_STATE_SYNCING ||
2280 			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2281 			    iclog->ic_state == XLOG_STATE_IOERROR )
2282 				break;
2283 			iclog = iclog->ic_next;
2284 		} while (first_iclog != iclog);
2285 	}
2286 #endif
2287 
2288 	if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2289 		flushcnt = log->l_flushcnt;
2290 		log->l_flushcnt = 0;
2291 	}
2292 	LOG_UNLOCK(log, s);
2293 	while (flushcnt--)
2294 		vsema(&log->l_flushsema);
2295 }	/* xlog_state_do_callback */
2296 
2297 
2298 /*
2299  * Finish transitioning this iclog to the dirty state.
2300  *
2301  * Make sure that we completely execute this routine only when this is
2302  * the last call to the iclog.  There is a good chance that iclog flushes,
2303  * when we reach the end of the physical log, get turned into 2 separate
2304  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2305  * routine.  By using the reference count bwritecnt, we guarantee that only
2306  * the second completion goes through.
2307  *
2308  * Callbacks could take time, so they are done outside the scope of the
2309  * global state machine log lock.  Assume that the calls to cvsema won't
2310  * take a long time.  At least we know it won't sleep.
2311  */
2312 void
2313 xlog_state_done_syncing(
2314 	xlog_in_core_t	*iclog,
2315 	int		aborted)
2316 {
2317 	xlog_t		   *log = iclog->ic_log;
2318 	SPLDECL(s);
2319 
2320 	s = LOG_LOCK(log);
2321 
2322 	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2323 	       iclog->ic_state == XLOG_STATE_IOERROR);
2324 	ASSERT(iclog->ic_refcnt == 0);
2325 	ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2326 
2327 
2328 	/*
2329 	 * If we got an error, either on the first buffer, or in the case of
2330 	 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2331 	 * and none should ever be attempted to be written to disk
2332 	 * again.
2333 	 */
2334 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
2335 		if (--iclog->ic_bwritecnt == 1) {
2336 			LOG_UNLOCK(log, s);
2337 			return;
2338 		}
2339 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2340 	}
2341 
2342 	/*
2343 	 * Someone could be sleeping prior to writing out the next
2344 	 * iclog buffer, we wake them all, one will get to do the
2345 	 * I/O, the others get to wait for the result.
2346 	 */
2347 	sv_broadcast(&iclog->ic_writesema);
2348 	LOG_UNLOCK(log, s);
2349 	xlog_state_do_callback(log, aborted, iclog);	/* also cleans log */
2350 }	/* xlog_state_done_syncing */
2351 
2352 
2353 /*
2354  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2355  * sleep.  The flush semaphore is set to the number of in-core buffers and
2356  * decremented around disk syncing.  Therefore, if all buffers are syncing,
2357  * this semaphore will cause new writes to sleep until a sync completes.
2358  * Otherwise, this code just does p() followed by v().  This approximates
2359  * a sleep/wakeup except we can't race.
2360  *
2361  * The in-core logs are used in a circular fashion. They are not used
2362  * out-of-order even when an iclog past the head is free.
2363  *
2364  * return:
2365  *	* log_offset where xlog_write() can start writing into the in-core
2366  *		log's data space.
2367  *	* in-core log pointer to which xlog_write() should write.
2368  *	* boolean indicating this is a continued write to an in-core log.
2369  *		If this is the last write, then the in-core log's offset field
2370  *		needs to be incremented, depending on the amount of data which
2371  *		is copied.
2372  */
2373 int
2374 xlog_state_get_iclog_space(xlog_t	  *log,
2375 			   int		  len,
2376 			   xlog_in_core_t **iclogp,
2377 			   xlog_ticket_t  *ticket,
2378 			   int		  *continued_write,
2379 			   int		  *logoffsetp)
2380 {
2381 	SPLDECL(s);
2382 	int		  log_offset;
2383 	xlog_rec_header_t *head;
2384 	xlog_in_core_t	  *iclog;
2385 	int		  error;
2386 
2387 restart:
2388 	s = LOG_LOCK(log);
2389 	if (XLOG_FORCED_SHUTDOWN(log)) {
2390 		LOG_UNLOCK(log, s);
2391 		return XFS_ERROR(EIO);
2392 	}
2393 
2394 	iclog = log->l_iclog;
2395 	if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2396 		log->l_flushcnt++;
2397 		LOG_UNLOCK(log, s);
2398 		xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2399 		XFS_STATS_INC(xs_log_noiclogs);
2400 		/* Ensure that log writes happen */
2401 		psema(&log->l_flushsema, PINOD);
2402 		goto restart;
2403 	}
2404 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2405 	head = &iclog->ic_header;
2406 
2407 	iclog->ic_refcnt++;			/* prevents sync */
2408 	log_offset = iclog->ic_offset;
2409 
2410 	/* On the 1st write to an iclog, figure out lsn.  This works
2411 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2412 	 * committing to.  If the offset is set, that's how many blocks
2413 	 * must be written.
2414 	 */
2415 	if (log_offset == 0) {
2416 		ticket->t_curr_res -= log->l_iclog_hsize;
2417 		XLOG_TIC_ADD_REGION(ticket,
2418 				    log->l_iclog_hsize,
2419 				    XLOG_REG_TYPE_LRHEADER);
2420 		INT_SET(head->h_cycle, ARCH_CONVERT, log->l_curr_cycle);
2421 		ASSIGN_LSN(head->h_lsn, log);
2422 		ASSERT(log->l_curr_block >= 0);
2423 	}
2424 
2425 	/* If there is enough room to write everything, then do it.  Otherwise,
2426 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2427 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2428 	 * until you know exactly how many bytes get copied.  Therefore, wait
2429 	 * until later to update ic_offset.
2430 	 *
2431 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2432 	 * can fit into remaining data section.
2433 	 */
2434 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2435 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2436 
2437 		/* If I'm the only one writing to this iclog, sync it to disk */
2438 		if (iclog->ic_refcnt == 1) {
2439 			LOG_UNLOCK(log, s);
2440 			if ((error = xlog_state_release_iclog(log, iclog)))
2441 				return (error);
2442 		} else {
2443 			iclog->ic_refcnt--;
2444 			LOG_UNLOCK(log, s);
2445 		}
2446 		goto restart;
2447 	}
2448 
2449 	/* Do we have enough room to write the full amount in the remainder
2450 	 * of this iclog?  Or must we continue a write on the next iclog and
2451 	 * mark this iclog as completely taken?  In the case where we switch
2452 	 * iclogs (to mark it taken), this particular iclog will release/sync
2453 	 * to disk in xlog_write().
2454 	 */
2455 	if (len <= iclog->ic_size - iclog->ic_offset) {
2456 		*continued_write = 0;
2457 		iclog->ic_offset += len;
2458 	} else {
2459 		*continued_write = 1;
2460 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2461 	}
2462 	*iclogp = iclog;
2463 
2464 	ASSERT(iclog->ic_offset <= iclog->ic_size);
2465 	LOG_UNLOCK(log, s);
2466 
2467 	*logoffsetp = log_offset;
2468 	return 0;
2469 }	/* xlog_state_get_iclog_space */
2470 
2471 /*
2472  * Atomically get the log space required for a log ticket.
2473  *
2474  * Once a ticket gets put onto the reserveq, it will only return after
2475  * the needed reservation is satisfied.
2476  */
2477 STATIC int
2478 xlog_grant_log_space(xlog_t	   *log,
2479 		     xlog_ticket_t *tic)
2480 {
2481 	int		 free_bytes;
2482 	int		 need_bytes;
2483 	SPLDECL(s);
2484 #ifdef DEBUG
2485 	xfs_lsn_t	 tail_lsn;
2486 #endif
2487 
2488 
2489 #ifdef DEBUG
2490 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2491 		panic("grant Recovery problem");
2492 #endif
2493 
2494 	/* Is there space or do we need to sleep? */
2495 	s = GRANT_LOCK(log);
2496 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2497 
2498 	/* something is already sleeping; insert new transaction at end */
2499 	if (log->l_reserve_headq) {
2500 		XLOG_INS_TICKETQ(log->l_reserve_headq, tic);
2501 		xlog_trace_loggrant(log, tic,
2502 				    "xlog_grant_log_space: sleep 1");
2503 		/*
2504 		 * Gotta check this before going to sleep, while we're
2505 		 * holding the grant lock.
2506 		 */
2507 		if (XLOG_FORCED_SHUTDOWN(log))
2508 			goto error_return;
2509 
2510 		XFS_STATS_INC(xs_sleep_logspace);
2511 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2512 		/*
2513 		 * If we got an error, and the filesystem is shutting down,
2514 		 * we'll catch it down below. So just continue...
2515 		 */
2516 		xlog_trace_loggrant(log, tic,
2517 				    "xlog_grant_log_space: wake 1");
2518 		s = GRANT_LOCK(log);
2519 	}
2520 	if (tic->t_flags & XFS_LOG_PERM_RESERV)
2521 		need_bytes = tic->t_unit_res*tic->t_ocnt;
2522 	else
2523 		need_bytes = tic->t_unit_res;
2524 
2525 redo:
2526 	if (XLOG_FORCED_SHUTDOWN(log))
2527 		goto error_return;
2528 
2529 	free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2530 				     log->l_grant_reserve_bytes);
2531 	if (free_bytes < need_bytes) {
2532 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2533 			XLOG_INS_TICKETQ(log->l_reserve_headq, tic);
2534 		xlog_trace_loggrant(log, tic,
2535 				    "xlog_grant_log_space: sleep 2");
2536 		XFS_STATS_INC(xs_sleep_logspace);
2537 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2538 
2539 		if (XLOG_FORCED_SHUTDOWN(log)) {
2540 			s = GRANT_LOCK(log);
2541 			goto error_return;
2542 		}
2543 
2544 		xlog_trace_loggrant(log, tic,
2545 				    "xlog_grant_log_space: wake 2");
2546 		xlog_grant_push_ail(log->l_mp, need_bytes);
2547 		s = GRANT_LOCK(log);
2548 		goto redo;
2549 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2550 		XLOG_DEL_TICKETQ(log->l_reserve_headq, tic);
2551 
2552 	/* we've got enough space */
2553 	XLOG_GRANT_ADD_SPACE(log, need_bytes, 'w');
2554 	XLOG_GRANT_ADD_SPACE(log, need_bytes, 'r');
2555 #ifdef DEBUG
2556 	tail_lsn = log->l_tail_lsn;
2557 	/*
2558 	 * Check to make sure the grant write head didn't just over lap the
2559 	 * tail.  If the cycles are the same, we can't be overlapping.
2560 	 * Otherwise, make sure that the cycles differ by exactly one and
2561 	 * check the byte count.
2562 	 */
2563 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2564 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2565 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2566 	}
2567 #endif
2568 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2569 	xlog_verify_grant_head(log, 1);
2570 	GRANT_UNLOCK(log, s);
2571 	return 0;
2572 
2573  error_return:
2574 	if (tic->t_flags & XLOG_TIC_IN_Q)
2575 		XLOG_DEL_TICKETQ(log->l_reserve_headq, tic);
2576 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2577 	/*
2578 	 * If we are failing, make sure the ticket doesn't have any
2579 	 * current reservations. We don't want to add this back when
2580 	 * the ticket/transaction gets cancelled.
2581 	 */
2582 	tic->t_curr_res = 0;
2583 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2584 	GRANT_UNLOCK(log, s);
2585 	return XFS_ERROR(EIO);
2586 }	/* xlog_grant_log_space */
2587 
2588 
2589 /*
2590  * Replenish the byte reservation required by moving the grant write head.
2591  *
2592  *
2593  */
2594 STATIC int
2595 xlog_regrant_write_log_space(xlog_t	   *log,
2596 			     xlog_ticket_t *tic)
2597 {
2598 	SPLDECL(s);
2599 	int		free_bytes, need_bytes;
2600 	xlog_ticket_t	*ntic;
2601 #ifdef DEBUG
2602 	xfs_lsn_t	tail_lsn;
2603 #endif
2604 
2605 	tic->t_curr_res = tic->t_unit_res;
2606 	XLOG_TIC_RESET_RES(tic);
2607 
2608 	if (tic->t_cnt > 0)
2609 		return (0);
2610 
2611 #ifdef DEBUG
2612 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2613 		panic("regrant Recovery problem");
2614 #endif
2615 
2616 	s = GRANT_LOCK(log);
2617 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2618 
2619 	if (XLOG_FORCED_SHUTDOWN(log))
2620 		goto error_return;
2621 
2622 	/* If there are other waiters on the queue then give them a
2623 	 * chance at logspace before us. Wake up the first waiters,
2624 	 * if we do not wake up all the waiters then go to sleep waiting
2625 	 * for more free space, otherwise try to get some space for
2626 	 * this transaction.
2627 	 */
2628 
2629 	if ((ntic = log->l_write_headq)) {
2630 		free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2631 					     log->l_grant_write_bytes);
2632 		do {
2633 			ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2634 
2635 			if (free_bytes < ntic->t_unit_res)
2636 				break;
2637 			free_bytes -= ntic->t_unit_res;
2638 			sv_signal(&ntic->t_sema);
2639 			ntic = ntic->t_next;
2640 		} while (ntic != log->l_write_headq);
2641 
2642 		if (ntic != log->l_write_headq) {
2643 			if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2644 				XLOG_INS_TICKETQ(log->l_write_headq, tic);
2645 
2646 			xlog_trace_loggrant(log, tic,
2647 				    "xlog_regrant_write_log_space: sleep 1");
2648 			XFS_STATS_INC(xs_sleep_logspace);
2649 			sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2650 				&log->l_grant_lock, s);
2651 
2652 			/* If we're shutting down, this tic is already
2653 			 * off the queue */
2654 			if (XLOG_FORCED_SHUTDOWN(log)) {
2655 				s = GRANT_LOCK(log);
2656 				goto error_return;
2657 			}
2658 
2659 			xlog_trace_loggrant(log, tic,
2660 				    "xlog_regrant_write_log_space: wake 1");
2661 			xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2662 			s = GRANT_LOCK(log);
2663 		}
2664 	}
2665 
2666 	need_bytes = tic->t_unit_res;
2667 
2668 redo:
2669 	if (XLOG_FORCED_SHUTDOWN(log))
2670 		goto error_return;
2671 
2672 	free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2673 				     log->l_grant_write_bytes);
2674 	if (free_bytes < need_bytes) {
2675 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2676 			XLOG_INS_TICKETQ(log->l_write_headq, tic);
2677 		XFS_STATS_INC(xs_sleep_logspace);
2678 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2679 
2680 		/* If we're shutting down, this tic is already off the queue */
2681 		if (XLOG_FORCED_SHUTDOWN(log)) {
2682 			s = GRANT_LOCK(log);
2683 			goto error_return;
2684 		}
2685 
2686 		xlog_trace_loggrant(log, tic,
2687 				    "xlog_regrant_write_log_space: wake 2");
2688 		xlog_grant_push_ail(log->l_mp, need_bytes);
2689 		s = GRANT_LOCK(log);
2690 		goto redo;
2691 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2692 		XLOG_DEL_TICKETQ(log->l_write_headq, tic);
2693 
2694 	XLOG_GRANT_ADD_SPACE(log, need_bytes, 'w'); /* we've got enough space */
2695 #ifdef DEBUG
2696 	tail_lsn = log->l_tail_lsn;
2697 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2698 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2699 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2700 	}
2701 #endif
2702 
2703 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2704 	xlog_verify_grant_head(log, 1);
2705 	GRANT_UNLOCK(log, s);
2706 	return (0);
2707 
2708 
2709  error_return:
2710 	if (tic->t_flags & XLOG_TIC_IN_Q)
2711 		XLOG_DEL_TICKETQ(log->l_reserve_headq, tic);
2712 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2713 	/*
2714 	 * If we are failing, make sure the ticket doesn't have any
2715 	 * current reservations. We don't want to add this back when
2716 	 * the ticket/transaction gets cancelled.
2717 	 */
2718 	tic->t_curr_res = 0;
2719 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2720 	GRANT_UNLOCK(log, s);
2721 	return XFS_ERROR(EIO);
2722 }	/* xlog_regrant_write_log_space */
2723 
2724 
2725 /* The first cnt-1 times through here we don't need to
2726  * move the grant write head because the permanent
2727  * reservation has reserved cnt times the unit amount.
2728  * Release part of current permanent unit reservation and
2729  * reset current reservation to be one units worth.  Also
2730  * move grant reservation head forward.
2731  */
2732 STATIC void
2733 xlog_regrant_reserve_log_space(xlog_t	     *log,
2734 			       xlog_ticket_t *ticket)
2735 {
2736 	SPLDECL(s);
2737 
2738 	xlog_trace_loggrant(log, ticket,
2739 			    "xlog_regrant_reserve_log_space: enter");
2740 	if (ticket->t_cnt > 0)
2741 		ticket->t_cnt--;
2742 
2743 	s = GRANT_LOCK(log);
2744 	XLOG_GRANT_SUB_SPACE(log, ticket->t_curr_res, 'w');
2745 	XLOG_GRANT_SUB_SPACE(log, ticket->t_curr_res, 'r');
2746 	ticket->t_curr_res = ticket->t_unit_res;
2747 	XLOG_TIC_RESET_RES(ticket);
2748 	xlog_trace_loggrant(log, ticket,
2749 			    "xlog_regrant_reserve_log_space: sub current res");
2750 	xlog_verify_grant_head(log, 1);
2751 
2752 	/* just return if we still have some of the pre-reserved space */
2753 	if (ticket->t_cnt > 0) {
2754 		GRANT_UNLOCK(log, s);
2755 		return;
2756 	}
2757 
2758 	XLOG_GRANT_ADD_SPACE(log, ticket->t_unit_res, 'r');
2759 	xlog_trace_loggrant(log, ticket,
2760 			    "xlog_regrant_reserve_log_space: exit");
2761 	xlog_verify_grant_head(log, 0);
2762 	GRANT_UNLOCK(log, s);
2763 	ticket->t_curr_res = ticket->t_unit_res;
2764 	XLOG_TIC_RESET_RES(ticket);
2765 }	/* xlog_regrant_reserve_log_space */
2766 
2767 
2768 /*
2769  * Give back the space left from a reservation.
2770  *
2771  * All the information we need to make a correct determination of space left
2772  * is present.  For non-permanent reservations, things are quite easy.  The
2773  * count should have been decremented to zero.  We only need to deal with the
2774  * space remaining in the current reservation part of the ticket.  If the
2775  * ticket contains a permanent reservation, there may be left over space which
2776  * needs to be released.  A count of N means that N-1 refills of the current
2777  * reservation can be done before we need to ask for more space.  The first
2778  * one goes to fill up the first current reservation.  Once we run out of
2779  * space, the count will stay at zero and the only space remaining will be
2780  * in the current reservation field.
2781  */
2782 STATIC void
2783 xlog_ungrant_log_space(xlog_t	     *log,
2784 		       xlog_ticket_t *ticket)
2785 {
2786 	SPLDECL(s);
2787 
2788 	if (ticket->t_cnt > 0)
2789 		ticket->t_cnt--;
2790 
2791 	s = GRANT_LOCK(log);
2792 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2793 
2794 	XLOG_GRANT_SUB_SPACE(log, ticket->t_curr_res, 'w');
2795 	XLOG_GRANT_SUB_SPACE(log, ticket->t_curr_res, 'r');
2796 
2797 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2798 
2799 	/* If this is a permanent reservation ticket, we may be able to free
2800 	 * up more space based on the remaining count.
2801 	 */
2802 	if (ticket->t_cnt > 0) {
2803 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2804 		XLOG_GRANT_SUB_SPACE(log, ticket->t_unit_res*ticket->t_cnt,'w');
2805 		XLOG_GRANT_SUB_SPACE(log, ticket->t_unit_res*ticket->t_cnt,'r');
2806 	}
2807 
2808 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2809 	xlog_verify_grant_head(log, 1);
2810 	GRANT_UNLOCK(log, s);
2811 	xfs_log_move_tail(log->l_mp, 1);
2812 }	/* xlog_ungrant_log_space */
2813 
2814 
2815 /*
2816  * Atomically put back used ticket.
2817  */
2818 void
2819 xlog_state_put_ticket(xlog_t	    *log,
2820 		      xlog_ticket_t *tic)
2821 {
2822 	unsigned long s;
2823 
2824 	s = LOG_LOCK(log);
2825 	xlog_ticket_put(log, tic);
2826 	LOG_UNLOCK(log, s);
2827 }	/* xlog_state_put_ticket */
2828 
2829 /*
2830  * Flush iclog to disk if this is the last reference to the given iclog and
2831  * the WANT_SYNC bit is set.
2832  *
2833  * When this function is entered, the iclog is not necessarily in the
2834  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2835  *
2836  *
2837  */
2838 int
2839 xlog_state_release_iclog(xlog_t		*log,
2840 			 xlog_in_core_t	*iclog)
2841 {
2842 	SPLDECL(s);
2843 	int		sync = 0;	/* do we sync? */
2844 
2845 	xlog_assign_tail_lsn(log->l_mp);
2846 
2847 	s = LOG_LOCK(log);
2848 
2849 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2850 		LOG_UNLOCK(log, s);
2851 		return XFS_ERROR(EIO);
2852 	}
2853 
2854 	ASSERT(iclog->ic_refcnt > 0);
2855 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2856 	       iclog->ic_state == XLOG_STATE_WANT_SYNC);
2857 
2858 	if (--iclog->ic_refcnt == 0 &&
2859 	    iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2860 		sync++;
2861 		iclog->ic_state = XLOG_STATE_SYNCING;
2862 		INT_SET(iclog->ic_header.h_tail_lsn, ARCH_CONVERT, log->l_tail_lsn);
2863 		xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2864 		/* cycle incremented when incrementing curr_block */
2865 	}
2866 
2867 	LOG_UNLOCK(log, s);
2868 
2869 	/*
2870 	 * We let the log lock go, so it's possible that we hit a log I/O
2871 	 * error or someother SHUTDOWN condition that marks the iclog
2872 	 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2873 	 * this iclog has consistent data, so we ignore IOERROR
2874 	 * flags after this point.
2875 	 */
2876 	if (sync) {
2877 		return xlog_sync(log, iclog);
2878 	}
2879 	return (0);
2880 
2881 }	/* xlog_state_release_iclog */
2882 
2883 
2884 /*
2885  * This routine will mark the current iclog in the ring as WANT_SYNC
2886  * and move the current iclog pointer to the next iclog in the ring.
2887  * When this routine is called from xlog_state_get_iclog_space(), the
2888  * exact size of the iclog has not yet been determined.  All we know is
2889  * that every data block.  We have run out of space in this log record.
2890  */
2891 STATIC void
2892 xlog_state_switch_iclogs(xlog_t		*log,
2893 			 xlog_in_core_t *iclog,
2894 			 int		eventual_size)
2895 {
2896 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2897 	if (!eventual_size)
2898 		eventual_size = iclog->ic_offset;
2899 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2900 	INT_SET(iclog->ic_header.h_prev_block, ARCH_CONVERT, log->l_prev_block);
2901 	log->l_prev_block = log->l_curr_block;
2902 	log->l_prev_cycle = log->l_curr_cycle;
2903 
2904 	/* roll log?: ic_offset changed later */
2905 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2906 
2907 	/* Round up to next log-sunit */
2908 	if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
2909 	    log->l_mp->m_sb.sb_logsunit > 1) {
2910 		__uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2911 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2912 	}
2913 
2914 	if (log->l_curr_block >= log->l_logBBsize) {
2915 		log->l_curr_cycle++;
2916 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2917 			log->l_curr_cycle++;
2918 		log->l_curr_block -= log->l_logBBsize;
2919 		ASSERT(log->l_curr_block >= 0);
2920 	}
2921 	ASSERT(iclog == log->l_iclog);
2922 	log->l_iclog = iclog->ic_next;
2923 }	/* xlog_state_switch_iclogs */
2924 
2925 
2926 /*
2927  * Write out all data in the in-core log as of this exact moment in time.
2928  *
2929  * Data may be written to the in-core log during this call.  However,
2930  * we don't guarantee this data will be written out.  A change from past
2931  * implementation means this routine will *not* write out zero length LRs.
2932  *
2933  * Basically, we try and perform an intelligent scan of the in-core logs.
2934  * If we determine there is no flushable data, we just return.  There is no
2935  * flushable data if:
2936  *
2937  *	1. the current iclog is active and has no data; the previous iclog
2938  *		is in the active or dirty state.
2939  *	2. the current iclog is drity, and the previous iclog is in the
2940  *		active or dirty state.
2941  *
2942  * We may sleep (call psema) if:
2943  *
2944  *	1. the current iclog is not in the active nor dirty state.
2945  *	2. the current iclog dirty, and the previous iclog is not in the
2946  *		active nor dirty state.
2947  *	3. the current iclog is active, and there is another thread writing
2948  *		to this particular iclog.
2949  *	4. a) the current iclog is active and has no other writers
2950  *	   b) when we return from flushing out this iclog, it is still
2951  *		not in the active nor dirty state.
2952  */
2953 STATIC int
2954 xlog_state_sync_all(xlog_t *log, uint flags)
2955 {
2956 	xlog_in_core_t	*iclog;
2957 	xfs_lsn_t	lsn;
2958 	SPLDECL(s);
2959 
2960 	s = LOG_LOCK(log);
2961 
2962 	iclog = log->l_iclog;
2963 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2964 		LOG_UNLOCK(log, s);
2965 		return XFS_ERROR(EIO);
2966 	}
2967 
2968 	/* If the head iclog is not active nor dirty, we just attach
2969 	 * ourselves to the head and go to sleep.
2970 	 */
2971 	if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2972 	    iclog->ic_state == XLOG_STATE_DIRTY) {
2973 		/*
2974 		 * If the head is dirty or (active and empty), then
2975 		 * we need to look at the previous iclog.  If the previous
2976 		 * iclog is active or dirty we are done.  There is nothing
2977 		 * to sync out.  Otherwise, we attach ourselves to the
2978 		 * previous iclog and go to sleep.
2979 		 */
2980 		if (iclog->ic_state == XLOG_STATE_DIRTY ||
2981 		    (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
2982 			iclog = iclog->ic_prev;
2983 			if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2984 			    iclog->ic_state == XLOG_STATE_DIRTY)
2985 				goto no_sleep;
2986 			else
2987 				goto maybe_sleep;
2988 		} else {
2989 			if (iclog->ic_refcnt == 0) {
2990 				/* We are the only one with access to this
2991 				 * iclog.  Flush it out now.  There should
2992 				 * be a roundoff of zero to show that someone
2993 				 * has already taken care of the roundoff from
2994 				 * the previous sync.
2995 				 */
2996 				iclog->ic_refcnt++;
2997 				lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2998 				xlog_state_switch_iclogs(log, iclog, 0);
2999 				LOG_UNLOCK(log, s);
3000 
3001 				if (xlog_state_release_iclog(log, iclog))
3002 					return XFS_ERROR(EIO);
3003 				s = LOG_LOCK(log);
3004 				if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) == lsn &&
3005 				    iclog->ic_state != XLOG_STATE_DIRTY)
3006 					goto maybe_sleep;
3007 				else
3008 					goto no_sleep;
3009 			} else {
3010 				/* Someone else is writing to this iclog.
3011 				 * Use its call to flush out the data.  However,
3012 				 * the other thread may not force out this LR,
3013 				 * so we mark it WANT_SYNC.
3014 				 */
3015 				xlog_state_switch_iclogs(log, iclog, 0);
3016 				goto maybe_sleep;
3017 			}
3018 		}
3019 	}
3020 
3021 	/* By the time we come around again, the iclog could've been filled
3022 	 * which would give it another lsn.  If we have a new lsn, just
3023 	 * return because the relevant data has been flushed.
3024 	 */
3025 maybe_sleep:
3026 	if (flags & XFS_LOG_SYNC) {
3027 		/*
3028 		 * We must check if we're shutting down here, before
3029 		 * we wait, while we're holding the LOG_LOCK.
3030 		 * Then we check again after waking up, in case our
3031 		 * sleep was disturbed by a bad news.
3032 		 */
3033 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
3034 			LOG_UNLOCK(log, s);
3035 			return XFS_ERROR(EIO);
3036 		}
3037 		XFS_STATS_INC(xs_log_force_sleep);
3038 		sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
3039 		/*
3040 		 * No need to grab the log lock here since we're
3041 		 * only deciding whether or not to return EIO
3042 		 * and the memory read should be atomic.
3043 		 */
3044 		if (iclog->ic_state & XLOG_STATE_IOERROR)
3045 			return XFS_ERROR(EIO);
3046 
3047 	} else {
3048 
3049 no_sleep:
3050 		LOG_UNLOCK(log, s);
3051 	}
3052 	return 0;
3053 }	/* xlog_state_sync_all */
3054 
3055 
3056 /*
3057  * Used by code which implements synchronous log forces.
3058  *
3059  * Find in-core log with lsn.
3060  *	If it is in the DIRTY state, just return.
3061  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3062  *		state and go to sleep or return.
3063  *	If it is in any other state, go to sleep or return.
3064  *
3065  * If filesystem activity goes to zero, the iclog will get flushed only by
3066  * bdflush().
3067  */
3068 int
3069 xlog_state_sync(xlog_t	  *log,
3070 		xfs_lsn_t lsn,
3071 		uint	  flags)
3072 {
3073     xlog_in_core_t	*iclog;
3074     int			already_slept = 0;
3075     SPLDECL(s);
3076 
3077 
3078 try_again:
3079     s = LOG_LOCK(log);
3080     iclog = log->l_iclog;
3081 
3082     if (iclog->ic_state & XLOG_STATE_IOERROR) {
3083 	    LOG_UNLOCK(log, s);
3084 	    return XFS_ERROR(EIO);
3085     }
3086 
3087     do {
3088 	if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) != lsn) {
3089 	    iclog = iclog->ic_next;
3090 	    continue;
3091 	}
3092 
3093 	if (iclog->ic_state == XLOG_STATE_DIRTY) {
3094 		LOG_UNLOCK(log, s);
3095 		return 0;
3096 	}
3097 
3098 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3099 		/*
3100 		 * We sleep here if we haven't already slept (e.g.
3101 		 * this is the first time we've looked at the correct
3102 		 * iclog buf) and the buffer before us is going to
3103 		 * be sync'ed. The reason for this is that if we
3104 		 * are doing sync transactions here, by waiting for
3105 		 * the previous I/O to complete, we can allow a few
3106 		 * more transactions into this iclog before we close
3107 		 * it down.
3108 		 *
3109 		 * Otherwise, we mark the buffer WANT_SYNC, and bump
3110 		 * up the refcnt so we can release the log (which drops
3111 		 * the ref count).  The state switch keeps new transaction
3112 		 * commits from using this buffer.  When the current commits
3113 		 * finish writing into the buffer, the refcount will drop to
3114 		 * zero and the buffer will go out then.
3115 		 */
3116 		if (!already_slept &&
3117 		    (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3118 						 XLOG_STATE_SYNCING))) {
3119 			ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3120 			XFS_STATS_INC(xs_log_force_sleep);
3121 			sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3122 				&log->l_icloglock, s);
3123 			already_slept = 1;
3124 			goto try_again;
3125 		} else {
3126 			iclog->ic_refcnt++;
3127 			xlog_state_switch_iclogs(log, iclog, 0);
3128 			LOG_UNLOCK(log, s);
3129 			if (xlog_state_release_iclog(log, iclog))
3130 				return XFS_ERROR(EIO);
3131 			s = LOG_LOCK(log);
3132 		}
3133 	}
3134 
3135 	if ((flags & XFS_LOG_SYNC) && /* sleep */
3136 	    !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3137 
3138 		/*
3139 		 * Don't wait on the forcesema if we know that we've
3140 		 * gotten a log write error.
3141 		 */
3142 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
3143 			LOG_UNLOCK(log, s);
3144 			return XFS_ERROR(EIO);
3145 		}
3146 		XFS_STATS_INC(xs_log_force_sleep);
3147 		sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3148 		/*
3149 		 * No need to grab the log lock here since we're
3150 		 * only deciding whether or not to return EIO
3151 		 * and the memory read should be atomic.
3152 		 */
3153 		if (iclog->ic_state & XLOG_STATE_IOERROR)
3154 			return XFS_ERROR(EIO);
3155 	} else {		/* just return */
3156 		LOG_UNLOCK(log, s);
3157 	}
3158 	return 0;
3159 
3160     } while (iclog != log->l_iclog);
3161 
3162     LOG_UNLOCK(log, s);
3163     return (0);
3164 }	/* xlog_state_sync */
3165 
3166 
3167 /*
3168  * Called when we want to mark the current iclog as being ready to sync to
3169  * disk.
3170  */
3171 void
3172 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3173 {
3174 	SPLDECL(s);
3175 
3176 	s = LOG_LOCK(log);
3177 
3178 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3179 		xlog_state_switch_iclogs(log, iclog, 0);
3180 	} else {
3181 		ASSERT(iclog->ic_state &
3182 			(XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3183 	}
3184 
3185 	LOG_UNLOCK(log, s);
3186 }	/* xlog_state_want_sync */
3187 
3188 
3189 
3190 /*****************************************************************************
3191  *
3192  *		TICKET functions
3193  *
3194  *****************************************************************************
3195  */
3196 
3197 /*
3198  *	Algorithm doesn't take into account page size. ;-(
3199  */
3200 STATIC void
3201 xlog_state_ticket_alloc(xlog_t *log)
3202 {
3203 	xlog_ticket_t	*t_list;
3204 	xlog_ticket_t	*next;
3205 	xfs_caddr_t	buf;
3206 	uint		i = (NBPP / sizeof(xlog_ticket_t)) - 2;
3207 	SPLDECL(s);
3208 
3209 	/*
3210 	 * The kmem_zalloc may sleep, so we shouldn't be holding the
3211 	 * global lock.  XXXmiken: may want to use zone allocator.
3212 	 */
3213 	buf = (xfs_caddr_t) kmem_zalloc(NBPP, KM_SLEEP);
3214 
3215 	s = LOG_LOCK(log);
3216 
3217 	/* Attach 1st ticket to Q, so we can keep track of allocated memory */
3218 	t_list = (xlog_ticket_t *)buf;
3219 	t_list->t_next = log->l_unmount_free;
3220 	log->l_unmount_free = t_list++;
3221 	log->l_ticket_cnt++;
3222 	log->l_ticket_tcnt++;
3223 
3224 	/* Next ticket becomes first ticket attached to ticket free list */
3225 	if (log->l_freelist != NULL) {
3226 		ASSERT(log->l_tail != NULL);
3227 		log->l_tail->t_next = t_list;
3228 	} else {
3229 		log->l_freelist = t_list;
3230 	}
3231 	log->l_ticket_cnt++;
3232 	log->l_ticket_tcnt++;
3233 
3234 	/* Cycle through rest of alloc'ed memory, building up free Q */
3235 	for ( ; i > 0; i--) {
3236 		next = t_list + 1;
3237 		t_list->t_next = next;
3238 		t_list = next;
3239 		log->l_ticket_cnt++;
3240 		log->l_ticket_tcnt++;
3241 	}
3242 	t_list->t_next = NULL;
3243 	log->l_tail = t_list;
3244 	LOG_UNLOCK(log, s);
3245 }	/* xlog_state_ticket_alloc */
3246 
3247 
3248 /*
3249  * Put ticket into free list
3250  *
3251  * Assumption: log lock is held around this call.
3252  */
3253 STATIC void
3254 xlog_ticket_put(xlog_t		*log,
3255 		xlog_ticket_t	*ticket)
3256 {
3257 	sv_destroy(&ticket->t_sema);
3258 
3259 	/*
3260 	 * Don't think caching will make that much difference.  It's
3261 	 * more important to make debug easier.
3262 	 */
3263 #if 0
3264 	/* real code will want to use LIFO for caching */
3265 	ticket->t_next = log->l_freelist;
3266 	log->l_freelist = ticket;
3267 	/* no need to clear fields */
3268 #else
3269 	/* When we debug, it is easier if tickets are cycled */
3270 	ticket->t_next     = NULL;
3271 	if (log->l_tail != 0) {
3272 		log->l_tail->t_next = ticket;
3273 	} else {
3274 		ASSERT(log->l_freelist == 0);
3275 		log->l_freelist = ticket;
3276 	}
3277 	log->l_tail	    = ticket;
3278 #endif /* DEBUG */
3279 	log->l_ticket_cnt++;
3280 }	/* xlog_ticket_put */
3281 
3282 
3283 /*
3284  * Grab ticket off freelist or allocation some more
3285  */
3286 xlog_ticket_t *
3287 xlog_ticket_get(xlog_t		*log,
3288 		int		unit_bytes,
3289 		int		cnt,
3290 		char		client,
3291 		uint		xflags)
3292 {
3293 	xlog_ticket_t	*tic;
3294 	uint		num_headers;
3295 	SPLDECL(s);
3296 
3297  alloc:
3298 	if (log->l_freelist == NULL)
3299 		xlog_state_ticket_alloc(log);		/* potentially sleep */
3300 
3301 	s = LOG_LOCK(log);
3302 	if (log->l_freelist == NULL) {
3303 		LOG_UNLOCK(log, s);
3304 		goto alloc;
3305 	}
3306 	tic		= log->l_freelist;
3307 	log->l_freelist	= tic->t_next;
3308 	if (log->l_freelist == NULL)
3309 		log->l_tail = NULL;
3310 	log->l_ticket_cnt--;
3311 	LOG_UNLOCK(log, s);
3312 
3313 	/*
3314 	 * Permanent reservations have up to 'cnt'-1 active log operations
3315 	 * in the log.  A unit in this case is the amount of space for one
3316 	 * of these log operations.  Normal reservations have a cnt of 1
3317 	 * and their unit amount is the total amount of space required.
3318 	 *
3319 	 * The following lines of code account for non-transaction data
3320 	 * which occupy space in the on-disk log.
3321 	 *
3322 	 * Normal form of a transaction is:
3323 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3324 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3325 	 *
3326 	 * We need to account for all the leadup data and trailer data
3327 	 * around the transaction data.
3328 	 * And then we need to account for the worst case in terms of using
3329 	 * more space.
3330 	 * The worst case will happen if:
3331 	 * - the placement of the transaction happens to be such that the
3332 	 *   roundoff is at its maximum
3333 	 * - the transaction data is synced before the commit record is synced
3334 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3335 	 *   Therefore the commit record is in its own Log Record.
3336 	 *   This can happen as the commit record is called with its
3337 	 *   own region to xlog_write().
3338 	 *   This then means that in the worst case, roundoff can happen for
3339 	 *   the commit-rec as well.
3340 	 *   The commit-rec is smaller than padding in this scenario and so it is
3341 	 *   not added separately.
3342 	 */
3343 
3344 	/* for trans header */
3345 	unit_bytes += sizeof(xlog_op_header_t);
3346 	unit_bytes += sizeof(xfs_trans_header_t);
3347 
3348 	/* for start-rec */
3349 	unit_bytes += sizeof(xlog_op_header_t);
3350 
3351 	/* for LR headers */
3352 	num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3353 	unit_bytes += log->l_iclog_hsize * num_headers;
3354 
3355 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3356 	unit_bytes += log->l_iclog_hsize;
3357 
3358 	/* for split-recs - ophdrs added when data split over LRs */
3359 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3360 
3361 	/* for roundoff padding for transaction data and one for commit record */
3362 	if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
3363 	    log->l_mp->m_sb.sb_logsunit > 1) {
3364 		/* log su roundoff */
3365 		unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3366 	} else {
3367 		/* BB roundoff */
3368 		unit_bytes += 2*BBSIZE;
3369         }
3370 
3371 	tic->t_unit_res		= unit_bytes;
3372 	tic->t_curr_res		= unit_bytes;
3373 	tic->t_cnt		= cnt;
3374 	tic->t_ocnt		= cnt;
3375 	tic->t_tid		= (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3376 	tic->t_clientid		= client;
3377 	tic->t_flags		= XLOG_TIC_INITED;
3378 	tic->t_trans_type	= 0;
3379 	if (xflags & XFS_LOG_PERM_RESERV)
3380 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3381 	sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3382 
3383 	XLOG_TIC_RESET_RES(tic);
3384 
3385 	return tic;
3386 }	/* xlog_ticket_get */
3387 
3388 
3389 /******************************************************************************
3390  *
3391  *		Log debug routines
3392  *
3393  ******************************************************************************
3394  */
3395 #if defined(DEBUG) && !defined(XLOG_NOLOG)
3396 /*
3397  * Make sure that the destination ptr is within the valid data region of
3398  * one of the iclogs.  This uses backup pointers stored in a different
3399  * part of the log in case we trash the log structure.
3400  */
3401 void
3402 xlog_verify_dest_ptr(xlog_t     *log,
3403 		     __psint_t  ptr)
3404 {
3405 	int i;
3406 	int good_ptr = 0;
3407 
3408 	for (i=0; i < log->l_iclog_bufs; i++) {
3409 		if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3410 		    ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3411 			good_ptr++;
3412 	}
3413 	if (! good_ptr)
3414 		xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3415 }	/* xlog_verify_dest_ptr */
3416 
3417 STATIC void
3418 xlog_verify_grant_head(xlog_t *log, int equals)
3419 {
3420     if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3421 	if (equals)
3422 	    ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3423 	else
3424 	    ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3425     } else {
3426 	ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3427 	ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3428     }
3429 }	/* xlog_verify_grant_head */
3430 
3431 /* check if it will fit */
3432 STATIC void
3433 xlog_verify_tail_lsn(xlog_t	    *log,
3434 		     xlog_in_core_t *iclog,
3435 		     xfs_lsn_t	    tail_lsn)
3436 {
3437     int blocks;
3438 
3439     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3440 	blocks =
3441 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3442 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3443 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3444     } else {
3445 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3446 
3447 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3448 	    xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3449 
3450 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3451 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3452 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3453     }
3454 }	/* xlog_verify_tail_lsn */
3455 
3456 /*
3457  * Perform a number of checks on the iclog before writing to disk.
3458  *
3459  * 1. Make sure the iclogs are still circular
3460  * 2. Make sure we have a good magic number
3461  * 3. Make sure we don't have magic numbers in the data
3462  * 4. Check fields of each log operation header for:
3463  *	A. Valid client identifier
3464  *	B. tid ptr value falls in valid ptr space (user space code)
3465  *	C. Length in log record header is correct according to the
3466  *		individual operation headers within record.
3467  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3468  *	log, check the preceding blocks of the physical log to make sure all
3469  *	the cycle numbers agree with the current cycle number.
3470  */
3471 STATIC void
3472 xlog_verify_iclog(xlog_t	 *log,
3473 		  xlog_in_core_t *iclog,
3474 		  int		 count,
3475 		  boolean_t	 syncing)
3476 {
3477 	xlog_op_header_t	*ophead;
3478 	xlog_in_core_t		*icptr;
3479 	xlog_in_core_2_t	*xhdr;
3480 	xfs_caddr_t		ptr;
3481 	xfs_caddr_t		base_ptr;
3482 	__psint_t		field_offset;
3483 	__uint8_t		clientid;
3484 	int			len, i, j, k, op_len;
3485 	int			idx;
3486 	SPLDECL(s);
3487 
3488 	/* check validity of iclog pointers */
3489 	s = LOG_LOCK(log);
3490 	icptr = log->l_iclog;
3491 	for (i=0; i < log->l_iclog_bufs; i++) {
3492 		if (icptr == 0)
3493 			xlog_panic("xlog_verify_iclog: invalid ptr");
3494 		icptr = icptr->ic_next;
3495 	}
3496 	if (icptr != log->l_iclog)
3497 		xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3498 	LOG_UNLOCK(log, s);
3499 
3500 	/* check log magic numbers */
3501 	ptr = (xfs_caddr_t) &(iclog->ic_header);
3502 	if (INT_GET(*(uint *)ptr, ARCH_CONVERT) != XLOG_HEADER_MAGIC_NUM)
3503 		xlog_panic("xlog_verify_iclog: invalid magic num");
3504 
3505 	for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&(iclog->ic_header))+count;
3506 	     ptr += BBSIZE) {
3507 		if (INT_GET(*(uint *)ptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
3508 			xlog_panic("xlog_verify_iclog: unexpected magic num");
3509 	}
3510 
3511 	/* check fields */
3512 	len = INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT);
3513 	ptr = iclog->ic_datap;
3514 	base_ptr = ptr;
3515 	ophead = (xlog_op_header_t *)ptr;
3516 	xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3517 	for (i = 0; i < len; i++) {
3518 		ophead = (xlog_op_header_t *)ptr;
3519 
3520 		/* clientid is only 1 byte */
3521 		field_offset = (__psint_t)
3522 			       ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3523 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3524 			clientid = ophead->oh_clientid;
3525 		} else {
3526 			idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3527 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3528 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3529 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3530 				clientid = GET_CLIENT_ID(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3531 			} else {
3532 				clientid = GET_CLIENT_ID(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3533 			}
3534 		}
3535 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3536 			cmn_err(CE_WARN, "xlog_verify_iclog: invalid clientid %d op 0x%p offset 0x%x", clientid, ophead, field_offset);
3537 
3538 		/* check length */
3539 		field_offset = (__psint_t)
3540 			       ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3541 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3542 			op_len = INT_GET(ophead->oh_len, ARCH_CONVERT);
3543 		} else {
3544 			idx = BTOBBT((__psint_t)&ophead->oh_len -
3545 				    (__psint_t)iclog->ic_datap);
3546 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3547 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3548 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3549 				op_len = INT_GET(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3550 			} else {
3551 				op_len = INT_GET(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3552 			}
3553 		}
3554 		ptr += sizeof(xlog_op_header_t) + op_len;
3555 	}
3556 }	/* xlog_verify_iclog */
3557 #endif /* DEBUG && !XLOG_NOLOG */
3558 
3559 /*
3560  * Mark all iclogs IOERROR. LOG_LOCK is held by the caller.
3561  */
3562 STATIC int
3563 xlog_state_ioerror(
3564 	xlog_t	*log)
3565 {
3566 	xlog_in_core_t	*iclog, *ic;
3567 
3568 	iclog = log->l_iclog;
3569 	if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3570 		/*
3571 		 * Mark all the incore logs IOERROR.
3572 		 * From now on, no log flushes will result.
3573 		 */
3574 		ic = iclog;
3575 		do {
3576 			ic->ic_state = XLOG_STATE_IOERROR;
3577 			ic = ic->ic_next;
3578 		} while (ic != iclog);
3579 		return (0);
3580 	}
3581 	/*
3582 	 * Return non-zero, if state transition has already happened.
3583 	 */
3584 	return (1);
3585 }
3586 
3587 /*
3588  * This is called from xfs_force_shutdown, when we're forcibly
3589  * shutting down the filesystem, typically because of an IO error.
3590  * Our main objectives here are to make sure that:
3591  *	a. the filesystem gets marked 'SHUTDOWN' for all interested
3592  *	   parties to find out, 'atomically'.
3593  *	b. those who're sleeping on log reservations, pinned objects and
3594  *	    other resources get woken up, and be told the bad news.
3595  *	c. nothing new gets queued up after (a) and (b) are done.
3596  *	d. if !logerror, flush the iclogs to disk, then seal them off
3597  *	   for business.
3598  */
3599 int
3600 xfs_log_force_umount(
3601 	struct xfs_mount	*mp,
3602 	int			logerror)
3603 {
3604 	xlog_ticket_t	*tic;
3605 	xlog_t		*log;
3606 	int		retval;
3607 	SPLDECL(s);
3608 	SPLDECL(s2);
3609 
3610 	log = mp->m_log;
3611 
3612 	/*
3613 	 * If this happens during log recovery, don't worry about
3614 	 * locking; the log isn't open for business yet.
3615 	 */
3616 	if (!log ||
3617 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3618 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3619 		XFS_BUF_DONE(mp->m_sb_bp);
3620 		return (0);
3621 	}
3622 
3623 	/*
3624 	 * Somebody could've already done the hard work for us.
3625 	 * No need to get locks for this.
3626 	 */
3627 	if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3628 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3629 		return (1);
3630 	}
3631 	retval = 0;
3632 	/*
3633 	 * We must hold both the GRANT lock and the LOG lock,
3634 	 * before we mark the filesystem SHUTDOWN and wake
3635 	 * everybody up to tell the bad news.
3636 	 */
3637 	s = GRANT_LOCK(log);
3638 	s2 = LOG_LOCK(log);
3639 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3640 	XFS_BUF_DONE(mp->m_sb_bp);
3641 	/*
3642 	 * This flag is sort of redundant because of the mount flag, but
3643 	 * it's good to maintain the separation between the log and the rest
3644 	 * of XFS.
3645 	 */
3646 	log->l_flags |= XLOG_IO_ERROR;
3647 
3648 	/*
3649 	 * If we hit a log error, we want to mark all the iclogs IOERROR
3650 	 * while we're still holding the loglock.
3651 	 */
3652 	if (logerror)
3653 		retval = xlog_state_ioerror(log);
3654 	LOG_UNLOCK(log, s2);
3655 
3656 	/*
3657 	 * We don't want anybody waiting for log reservations
3658 	 * after this. That means we have to wake up everybody
3659 	 * queued up on reserve_headq as well as write_headq.
3660 	 * In addition, we make sure in xlog_{re}grant_log_space
3661 	 * that we don't enqueue anything once the SHUTDOWN flag
3662 	 * is set, and this action is protected by the GRANTLOCK.
3663 	 */
3664 	if ((tic = log->l_reserve_headq)) {
3665 		do {
3666 			sv_signal(&tic->t_sema);
3667 			tic = tic->t_next;
3668 		} while (tic != log->l_reserve_headq);
3669 	}
3670 
3671 	if ((tic = log->l_write_headq)) {
3672 		do {
3673 			sv_signal(&tic->t_sema);
3674 			tic = tic->t_next;
3675 		} while (tic != log->l_write_headq);
3676 	}
3677 	GRANT_UNLOCK(log, s);
3678 
3679 	if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3680 		ASSERT(!logerror);
3681 		/*
3682 		 * Force the incore logs to disk before shutting the
3683 		 * log down completely.
3684 		 */
3685 		xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC);
3686 		s2 = LOG_LOCK(log);
3687 		retval = xlog_state_ioerror(log);
3688 		LOG_UNLOCK(log, s2);
3689 	}
3690 	/*
3691 	 * Wake up everybody waiting on xfs_log_force.
3692 	 * Callback all log item committed functions as if the
3693 	 * log writes were completed.
3694 	 */
3695 	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3696 
3697 #ifdef XFSERRORDEBUG
3698 	{
3699 		xlog_in_core_t	*iclog;
3700 
3701 		s = LOG_LOCK(log);
3702 		iclog = log->l_iclog;
3703 		do {
3704 			ASSERT(iclog->ic_callback == 0);
3705 			iclog = iclog->ic_next;
3706 		} while (iclog != log->l_iclog);
3707 		LOG_UNLOCK(log, s);
3708 	}
3709 #endif
3710 	/* return non-zero if log IOERROR transition had already happened */
3711 	return (retval);
3712 }
3713 
3714 STATIC int
3715 xlog_iclogs_empty(xlog_t *log)
3716 {
3717 	xlog_in_core_t	*iclog;
3718 
3719 	iclog = log->l_iclog;
3720 	do {
3721 		/* endianness does not matter here, zero is zero in
3722 		 * any language.
3723 		 */
3724 		if (iclog->ic_header.h_num_logops)
3725 			return(0);
3726 		iclog = iclog->ic_next;
3727 	} while (iclog != log->l_iclog);
3728 	return(1);
3729 }
3730