xref: /openbmc/linux/fs/xfs/xfs_log.c (revision 87c2ce3b)
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir.h"
28 #include "xfs_dir2.h"
29 #include "xfs_dmapi.h"
30 #include "xfs_mount.h"
31 #include "xfs_error.h"
32 #include "xfs_log_priv.h"
33 #include "xfs_buf_item.h"
34 #include "xfs_bmap_btree.h"
35 #include "xfs_alloc_btree.h"
36 #include "xfs_ialloc_btree.h"
37 #include "xfs_log_recover.h"
38 #include "xfs_trans_priv.h"
39 #include "xfs_dir_sf.h"
40 #include "xfs_dir2_sf.h"
41 #include "xfs_attr_sf.h"
42 #include "xfs_dinode.h"
43 #include "xfs_inode.h"
44 #include "xfs_rw.h"
45 
46 
47 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
48 	{ (ptr) += (bytes); \
49 	  (len) -= (bytes); \
50 	  (off) += (bytes);}
51 
52 /* Local miscellaneous function prototypes */
53 STATIC int	 xlog_bdstrat_cb(struct xfs_buf *);
54 STATIC int	 xlog_commit_record(xfs_mount_t *mp, xlog_ticket_t *ticket,
55 				    xlog_in_core_t **, xfs_lsn_t *);
56 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t	*mp,
57 				xfs_buftarg_t	*log_target,
58 				xfs_daddr_t	blk_offset,
59 				int		num_bblks);
60 STATIC int	 xlog_space_left(xlog_t *log, int cycle, int bytes);
61 STATIC int	 xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
62 STATIC void	 xlog_unalloc_log(xlog_t *log);
63 STATIC int	 xlog_write(xfs_mount_t *mp, xfs_log_iovec_t region[],
64 			    int nentries, xfs_log_ticket_t tic,
65 			    xfs_lsn_t *start_lsn,
66 			    xlog_in_core_t **commit_iclog,
67 			    uint flags);
68 
69 /* local state machine functions */
70 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
71 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
72 STATIC int  xlog_state_get_iclog_space(xlog_t		*log,
73 				       int		len,
74 				       xlog_in_core_t	**iclog,
75 				       xlog_ticket_t	*ticket,
76 				       int		*continued_write,
77 				       int		*logoffsetp);
78 STATIC void xlog_state_put_ticket(xlog_t	*log,
79 				  xlog_ticket_t *tic);
80 STATIC int  xlog_state_release_iclog(xlog_t		*log,
81 				     xlog_in_core_t	*iclog);
82 STATIC void xlog_state_switch_iclogs(xlog_t		*log,
83 				     xlog_in_core_t *iclog,
84 				     int		eventual_size);
85 STATIC int  xlog_state_sync(xlog_t			*log,
86 			    xfs_lsn_t 			lsn,
87 			    uint			flags,
88 			    int				*log_flushed);
89 STATIC int  xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
90 STATIC void xlog_state_want_sync(xlog_t	*log, xlog_in_core_t *iclog);
91 
92 /* local functions to manipulate grant head */
93 STATIC int  xlog_grant_log_space(xlog_t		*log,
94 				 xlog_ticket_t	*xtic);
95 STATIC void xlog_grant_push_ail(xfs_mount_t	*mp,
96 				int		need_bytes);
97 STATIC void xlog_regrant_reserve_log_space(xlog_t	 *log,
98 					   xlog_ticket_t *ticket);
99 STATIC int xlog_regrant_write_log_space(xlog_t		*log,
100 					 xlog_ticket_t  *ticket);
101 STATIC void xlog_ungrant_log_space(xlog_t	 *log,
102 				   xlog_ticket_t *ticket);
103 
104 
105 /* local ticket functions */
106 STATIC void		xlog_state_ticket_alloc(xlog_t *log);
107 STATIC xlog_ticket_t	*xlog_ticket_get(xlog_t *log,
108 					 int	unit_bytes,
109 					 int	count,
110 					 char	clientid,
111 					 uint	flags);
112 STATIC void		xlog_ticket_put(xlog_t *log, xlog_ticket_t *ticket);
113 
114 #if defined(DEBUG)
115 STATIC void	xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
116 STATIC void	xlog_verify_grant_head(xlog_t *log, int equals);
117 STATIC void	xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
118 				  int count, boolean_t syncing);
119 STATIC void	xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
120 				     xfs_lsn_t tail_lsn);
121 #else
122 #define xlog_verify_dest_ptr(a,b)
123 #define xlog_verify_grant_head(a,b)
124 #define xlog_verify_iclog(a,b,c,d)
125 #define xlog_verify_tail_lsn(a,b,c)
126 #endif
127 
128 STATIC int	xlog_iclogs_empty(xlog_t *log);
129 
130 #if defined(XFS_LOG_TRACE)
131 void
132 xlog_trace_loggrant(xlog_t *log, xlog_ticket_t *tic, xfs_caddr_t string)
133 {
134 	unsigned long cnts;
135 
136 	if (!log->l_grant_trace) {
137 		log->l_grant_trace = ktrace_alloc(2048, KM_NOSLEEP);
138 		if (!log->l_grant_trace)
139 			return;
140 	}
141 	/* ticket counts are 1 byte each */
142 	cnts = ((unsigned long)tic->t_ocnt) | ((unsigned long)tic->t_cnt) << 8;
143 
144 	ktrace_enter(log->l_grant_trace,
145 		     (void *)tic,
146 		     (void *)log->l_reserve_headq,
147 		     (void *)log->l_write_headq,
148 		     (void *)((unsigned long)log->l_grant_reserve_cycle),
149 		     (void *)((unsigned long)log->l_grant_reserve_bytes),
150 		     (void *)((unsigned long)log->l_grant_write_cycle),
151 		     (void *)((unsigned long)log->l_grant_write_bytes),
152 		     (void *)((unsigned long)log->l_curr_cycle),
153 		     (void *)((unsigned long)log->l_curr_block),
154 		     (void *)((unsigned long)CYCLE_LSN(log->l_tail_lsn)),
155 		     (void *)((unsigned long)BLOCK_LSN(log->l_tail_lsn)),
156 		     (void *)string,
157 		     (void *)((unsigned long)tic->t_trans_type),
158 		     (void *)cnts,
159 		     (void *)((unsigned long)tic->t_curr_res),
160 		     (void *)((unsigned long)tic->t_unit_res));
161 }
162 
163 void
164 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
165 {
166 	if (!iclog->ic_trace)
167 		iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
168 	ktrace_enter(iclog->ic_trace,
169 		     (void *)((unsigned long)state),
170 		     (void *)((unsigned long)current_pid()),
171 		     (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
172 		     (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
173 		     (void *)NULL, (void *)NULL, (void *)NULL, (void *)NULL,
174 		     (void *)NULL, (void *)NULL);
175 }
176 #else
177 #define	xlog_trace_loggrant(log,tic,string)
178 #define	xlog_trace_iclog(iclog,state)
179 #endif /* XFS_LOG_TRACE */
180 
181 /*
182  * NOTES:
183  *
184  *	1. currblock field gets updated at startup and after in-core logs
185  *		marked as with WANT_SYNC.
186  */
187 
188 /*
189  * This routine is called when a user of a log manager ticket is done with
190  * the reservation.  If the ticket was ever used, then a commit record for
191  * the associated transaction is written out as a log operation header with
192  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
193  * a given ticket.  If the ticket was one with a permanent reservation, then
194  * a few operations are done differently.  Permanent reservation tickets by
195  * default don't release the reservation.  They just commit the current
196  * transaction with the belief that the reservation is still needed.  A flag
197  * must be passed in before permanent reservations are actually released.
198  * When these type of tickets are not released, they need to be set into
199  * the inited state again.  By doing this, a start record will be written
200  * out when the next write occurs.
201  */
202 xfs_lsn_t
203 xfs_log_done(xfs_mount_t	*mp,
204 	     xfs_log_ticket_t	xtic,
205 	     void		**iclog,
206 	     uint		flags)
207 {
208 	xlog_t		*log    = mp->m_log;
209 	xlog_ticket_t	*ticket = (xfs_log_ticket_t) xtic;
210 	xfs_lsn_t	lsn	= 0;
211 
212 	if (XLOG_FORCED_SHUTDOWN(log) ||
213 	    /*
214 	     * If nothing was ever written, don't write out commit record.
215 	     * If we get an error, just continue and give back the log ticket.
216 	     */
217 	    (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
218 	     (xlog_commit_record(mp, ticket,
219 				 (xlog_in_core_t **)iclog, &lsn)))) {
220 		lsn = (xfs_lsn_t) -1;
221 		if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
222 			flags |= XFS_LOG_REL_PERM_RESERV;
223 		}
224 	}
225 
226 
227 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
228 	    (flags & XFS_LOG_REL_PERM_RESERV)) {
229 		/*
230 		 * Release ticket if not permanent reservation or a specifc
231 		 * request has been made to release a permanent reservation.
232 		 */
233 		xlog_trace_loggrant(log, ticket, "xfs_log_done: (non-permanent)");
234 		xlog_ungrant_log_space(log, ticket);
235 		xlog_state_put_ticket(log, ticket);
236 	} else {
237 		xlog_trace_loggrant(log, ticket, "xfs_log_done: (permanent)");
238 		xlog_regrant_reserve_log_space(log, ticket);
239 	}
240 
241 	/* If this ticket was a permanent reservation and we aren't
242 	 * trying to release it, reset the inited flags; so next time
243 	 * we write, a start record will be written out.
244 	 */
245 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) &&
246 	    (flags & XFS_LOG_REL_PERM_RESERV) == 0)
247 		ticket->t_flags |= XLOG_TIC_INITED;
248 
249 	return lsn;
250 }	/* xfs_log_done */
251 
252 
253 /*
254  * Force the in-core log to disk.  If flags == XFS_LOG_SYNC,
255  *	the force is done synchronously.
256  *
257  * Asynchronous forces are implemented by setting the WANT_SYNC
258  * bit in the appropriate in-core log and then returning.
259  *
260  * Synchronous forces are implemented with a semaphore.  All callers
261  * to force a given lsn to disk will wait on a semaphore attached to the
262  * specific in-core log.  When given in-core log finally completes its
263  * write to disk, that thread will wake up all threads waiting on the
264  * semaphore.
265  */
266 int
267 _xfs_log_force(
268 	xfs_mount_t	*mp,
269 	xfs_lsn_t	lsn,
270 	uint		flags,
271 	int		*log_flushed)
272 {
273 	xlog_t		*log = mp->m_log;
274 	int		dummy;
275 
276 	if (!log_flushed)
277 		log_flushed = &dummy;
278 
279 	ASSERT(flags & XFS_LOG_FORCE);
280 
281 	XFS_STATS_INC(xs_log_force);
282 
283 	if (log->l_flags & XLOG_IO_ERROR)
284 		return XFS_ERROR(EIO);
285 	if (lsn == 0)
286 		return xlog_state_sync_all(log, flags, log_flushed);
287 	else
288 		return xlog_state_sync(log, lsn, flags, log_flushed);
289 }	/* xfs_log_force */
290 
291 /*
292  * Attaches a new iclog I/O completion callback routine during
293  * transaction commit.  If the log is in error state, a non-zero
294  * return code is handed back and the caller is responsible for
295  * executing the callback at an appropriate time.
296  */
297 int
298 xfs_log_notify(xfs_mount_t	  *mp,		/* mount of partition */
299 	       void		  *iclog_hndl,	/* iclog to hang callback off */
300 	       xfs_log_callback_t *cb)
301 {
302 	xlog_t *log = mp->m_log;
303 	xlog_in_core_t	  *iclog = (xlog_in_core_t *)iclog_hndl;
304 	int	abortflg, spl;
305 
306 	cb->cb_next = NULL;
307 	spl = LOG_LOCK(log);
308 	abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
309 	if (!abortflg) {
310 		ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
311 			      (iclog->ic_state == XLOG_STATE_WANT_SYNC));
312 		cb->cb_next = NULL;
313 		*(iclog->ic_callback_tail) = cb;
314 		iclog->ic_callback_tail = &(cb->cb_next);
315 	}
316 	LOG_UNLOCK(log, spl);
317 	return abortflg;
318 }	/* xfs_log_notify */
319 
320 int
321 xfs_log_release_iclog(xfs_mount_t *mp,
322 		      void	  *iclog_hndl)
323 {
324 	xlog_t *log = mp->m_log;
325 	xlog_in_core_t	  *iclog = (xlog_in_core_t *)iclog_hndl;
326 
327 	if (xlog_state_release_iclog(log, iclog)) {
328 		xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
329 		return(EIO);
330 	}
331 
332 	return 0;
333 }
334 
335 /*
336  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
337  *	to the reservation.
338  *  2. Potentially, push buffers at tail of log to disk.
339  *
340  * Each reservation is going to reserve extra space for a log record header.
341  * When writes happen to the on-disk log, we don't subtract the length of the
342  * log record header from any reservation.  By wasting space in each
343  * reservation, we prevent over allocation problems.
344  */
345 int
346 xfs_log_reserve(xfs_mount_t	 *mp,
347 		int		 unit_bytes,
348 		int		 cnt,
349 		xfs_log_ticket_t *ticket,
350 		__uint8_t	 client,
351 		uint		 flags,
352 		uint		 t_type)
353 {
354 	xlog_t		*log = mp->m_log;
355 	xlog_ticket_t	*internal_ticket;
356 	int		retval = 0;
357 
358 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
359 	ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
360 
361 	if (XLOG_FORCED_SHUTDOWN(log))
362 		return XFS_ERROR(EIO);
363 
364 	XFS_STATS_INC(xs_try_logspace);
365 
366 	if (*ticket != NULL) {
367 		ASSERT(flags & XFS_LOG_PERM_RESERV);
368 		internal_ticket = (xlog_ticket_t *)*ticket;
369 		xlog_trace_loggrant(log, internal_ticket, "xfs_log_reserve: existing ticket (permanent trans)");
370 		xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
371 		retval = xlog_regrant_write_log_space(log, internal_ticket);
372 	} else {
373 		/* may sleep if need to allocate more tickets */
374 		internal_ticket = xlog_ticket_get(log, unit_bytes, cnt,
375 						  client, flags);
376 		internal_ticket->t_trans_type = t_type;
377 		*ticket = internal_ticket;
378 		xlog_trace_loggrant(log, internal_ticket,
379 			(internal_ticket->t_flags & XLOG_TIC_PERM_RESERV) ?
380 			"xfs_log_reserve: create new ticket (permanent trans)" :
381 			"xfs_log_reserve: create new ticket");
382 		xlog_grant_push_ail(mp,
383 				    (internal_ticket->t_unit_res *
384 				     internal_ticket->t_cnt));
385 		retval = xlog_grant_log_space(log, internal_ticket);
386 	}
387 
388 	return retval;
389 }	/* xfs_log_reserve */
390 
391 
392 /*
393  * Mount a log filesystem
394  *
395  * mp		- ubiquitous xfs mount point structure
396  * log_target	- buftarg of on-disk log device
397  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
398  * num_bblocks	- Number of BBSIZE blocks in on-disk log
399  *
400  * Return error or zero.
401  */
402 int
403 xfs_log_mount(xfs_mount_t	*mp,
404 	      xfs_buftarg_t	*log_target,
405 	      xfs_daddr_t	blk_offset,
406 	      int		num_bblks)
407 {
408 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
409 		cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
410 	else {
411 		cmn_err(CE_NOTE,
412 			"!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
413 			mp->m_fsname);
414 		ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
415 	}
416 
417 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
418 
419 	/*
420 	 * skip log recovery on a norecovery mount.  pretend it all
421 	 * just worked.
422 	 */
423 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
424 		int	error;
425 		vfs_t	*vfsp = XFS_MTOVFS(mp);
426 		int	readonly = (vfsp->vfs_flag & VFS_RDONLY);
427 
428 		if (readonly)
429 			vfsp->vfs_flag &= ~VFS_RDONLY;
430 
431 		error = xlog_recover(mp->m_log, readonly);
432 
433 		if (readonly)
434 			vfsp->vfs_flag |= VFS_RDONLY;
435 		if (error) {
436 			cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
437 			xlog_unalloc_log(mp->m_log);
438 			return error;
439 		}
440 	}
441 
442 	/* Normal transactions can now occur */
443 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
444 
445 	/* End mounting message in xfs_log_mount_finish */
446 	return 0;
447 }	/* xfs_log_mount */
448 
449 /*
450  * Finish the recovery of the file system.  This is separate from
451  * the xfs_log_mount() call, because it depends on the code in
452  * xfs_mountfs() to read in the root and real-time bitmap inodes
453  * between calling xfs_log_mount() and here.
454  *
455  * mp		- ubiquitous xfs mount point structure
456  */
457 int
458 xfs_log_mount_finish(xfs_mount_t *mp, int mfsi_flags)
459 {
460 	int	error;
461 
462 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
463 		error = xlog_recover_finish(mp->m_log, mfsi_flags);
464 	else {
465 		error = 0;
466 		ASSERT(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY);
467 	}
468 
469 	return error;
470 }
471 
472 /*
473  * Unmount processing for the log.
474  */
475 int
476 xfs_log_unmount(xfs_mount_t *mp)
477 {
478 	int		error;
479 
480 	error = xfs_log_unmount_write(mp);
481 	xfs_log_unmount_dealloc(mp);
482 	return (error);
483 }
484 
485 /*
486  * Final log writes as part of unmount.
487  *
488  * Mark the filesystem clean as unmount happens.  Note that during relocation
489  * this routine needs to be executed as part of source-bag while the
490  * deallocation must not be done until source-end.
491  */
492 
493 /*
494  * Unmount record used to have a string "Unmount filesystem--" in the
495  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
496  * We just write the magic number now since that particular field isn't
497  * currently architecture converted and "nUmount" is a bit foo.
498  * As far as I know, there weren't any dependencies on the old behaviour.
499  */
500 
501 int
502 xfs_log_unmount_write(xfs_mount_t *mp)
503 {
504 	xlog_t		 *log = mp->m_log;
505 	xlog_in_core_t	 *iclog;
506 #ifdef DEBUG
507 	xlog_in_core_t	 *first_iclog;
508 #endif
509 	xfs_log_iovec_t  reg[1];
510 	xfs_log_ticket_t tic = NULL;
511 	xfs_lsn_t	 lsn;
512 	int		 error;
513 	SPLDECL(s);
514 
515 	/* the data section must be 32 bit size aligned */
516 	struct {
517 	    __uint16_t magic;
518 	    __uint16_t pad1;
519 	    __uint32_t pad2; /* may as well make it 64 bits */
520 	} magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
521 
522 	/*
523 	 * Don't write out unmount record on read-only mounts.
524 	 * Or, if we are doing a forced umount (typically because of IO errors).
525 	 */
526 	if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
527 		return 0;
528 
529 	xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC);
530 
531 #ifdef DEBUG
532 	first_iclog = iclog = log->l_iclog;
533 	do {
534 		if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
535 			ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
536 			ASSERT(iclog->ic_offset == 0);
537 		}
538 		iclog = iclog->ic_next;
539 	} while (iclog != first_iclog);
540 #endif
541 	if (! (XLOG_FORCED_SHUTDOWN(log))) {
542 		reg[0].i_addr = (void*)&magic;
543 		reg[0].i_len  = sizeof(magic);
544 		XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
545 
546 		error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0, 0);
547 		if (!error) {
548 			/* remove inited flag */
549 			((xlog_ticket_t *)tic)->t_flags = 0;
550 			error = xlog_write(mp, reg, 1, tic, &lsn,
551 					   NULL, XLOG_UNMOUNT_TRANS);
552 			/*
553 			 * At this point, we're umounting anyway,
554 			 * so there's no point in transitioning log state
555 			 * to IOERROR. Just continue...
556 			 */
557 		}
558 
559 		if (error) {
560 			xfs_fs_cmn_err(CE_ALERT, mp,
561 				"xfs_log_unmount: unmount record failed");
562 		}
563 
564 
565 		s = LOG_LOCK(log);
566 		iclog = log->l_iclog;
567 		iclog->ic_refcnt++;
568 		LOG_UNLOCK(log, s);
569 		xlog_state_want_sync(log, iclog);
570 		(void) xlog_state_release_iclog(log, iclog);
571 
572 		s = LOG_LOCK(log);
573 		if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
574 		      iclog->ic_state == XLOG_STATE_DIRTY)) {
575 			if (!XLOG_FORCED_SHUTDOWN(log)) {
576 				sv_wait(&iclog->ic_forcesema, PMEM,
577 					&log->l_icloglock, s);
578 			} else {
579 				LOG_UNLOCK(log, s);
580 			}
581 		} else {
582 			LOG_UNLOCK(log, s);
583 		}
584 		if (tic)
585 			xlog_state_put_ticket(log, tic);
586 	} else {
587 		/*
588 		 * We're already in forced_shutdown mode, couldn't
589 		 * even attempt to write out the unmount transaction.
590 		 *
591 		 * Go through the motions of sync'ing and releasing
592 		 * the iclog, even though no I/O will actually happen,
593 		 * we need to wait for other log I/O's that may already
594 		 * be in progress.  Do this as a separate section of
595 		 * code so we'll know if we ever get stuck here that
596 		 * we're in this odd situation of trying to unmount
597 		 * a file system that went into forced_shutdown as
598 		 * the result of an unmount..
599 		 */
600 		s = LOG_LOCK(log);
601 		iclog = log->l_iclog;
602 		iclog->ic_refcnt++;
603 		LOG_UNLOCK(log, s);
604 
605 		xlog_state_want_sync(log, iclog);
606 		(void) xlog_state_release_iclog(log, iclog);
607 
608 		s = LOG_LOCK(log);
609 
610 		if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
611 			|| iclog->ic_state == XLOG_STATE_DIRTY
612 			|| iclog->ic_state == XLOG_STATE_IOERROR) ) {
613 
614 				sv_wait(&iclog->ic_forcesema, PMEM,
615 					&log->l_icloglock, s);
616 		} else {
617 			LOG_UNLOCK(log, s);
618 		}
619 	}
620 
621 	return 0;
622 }	/* xfs_log_unmount_write */
623 
624 /*
625  * Deallocate log structures for unmount/relocation.
626  */
627 void
628 xfs_log_unmount_dealloc(xfs_mount_t *mp)
629 {
630 	xlog_unalloc_log(mp->m_log);
631 }
632 
633 /*
634  * Write region vectors to log.  The write happens using the space reservation
635  * of the ticket (tic).  It is not a requirement that all writes for a given
636  * transaction occur with one call to xfs_log_write().
637  */
638 int
639 xfs_log_write(xfs_mount_t *	mp,
640 	      xfs_log_iovec_t	reg[],
641 	      int		nentries,
642 	      xfs_log_ticket_t	tic,
643 	      xfs_lsn_t		*start_lsn)
644 {
645 	int	error;
646 	xlog_t *log = mp->m_log;
647 
648 	if (XLOG_FORCED_SHUTDOWN(log))
649 		return XFS_ERROR(EIO);
650 
651 	if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
652 		xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
653 	}
654 	return (error);
655 }	/* xfs_log_write */
656 
657 
658 void
659 xfs_log_move_tail(xfs_mount_t	*mp,
660 		  xfs_lsn_t	tail_lsn)
661 {
662 	xlog_ticket_t	*tic;
663 	xlog_t		*log = mp->m_log;
664 	int		need_bytes, free_bytes, cycle, bytes;
665 	SPLDECL(s);
666 
667 	if (XLOG_FORCED_SHUTDOWN(log))
668 		return;
669 	ASSERT(!XFS_FORCED_SHUTDOWN(mp));
670 
671 	if (tail_lsn == 0) {
672 		/* needed since sync_lsn is 64 bits */
673 		s = LOG_LOCK(log);
674 		tail_lsn = log->l_last_sync_lsn;
675 		LOG_UNLOCK(log, s);
676 	}
677 
678 	s = GRANT_LOCK(log);
679 
680 	/* Also an invalid lsn.  1 implies that we aren't passing in a valid
681 	 * tail_lsn.
682 	 */
683 	if (tail_lsn != 1) {
684 		log->l_tail_lsn = tail_lsn;
685 	}
686 
687 	if ((tic = log->l_write_headq)) {
688 #ifdef DEBUG
689 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
690 			panic("Recovery problem");
691 #endif
692 		cycle = log->l_grant_write_cycle;
693 		bytes = log->l_grant_write_bytes;
694 		free_bytes = xlog_space_left(log, cycle, bytes);
695 		do {
696 			ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
697 
698 			if (free_bytes < tic->t_unit_res && tail_lsn != 1)
699 				break;
700 			tail_lsn = 0;
701 			free_bytes -= tic->t_unit_res;
702 			sv_signal(&tic->t_sema);
703 			tic = tic->t_next;
704 		} while (tic != log->l_write_headq);
705 	}
706 	if ((tic = log->l_reserve_headq)) {
707 #ifdef DEBUG
708 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
709 			panic("Recovery problem");
710 #endif
711 		cycle = log->l_grant_reserve_cycle;
712 		bytes = log->l_grant_reserve_bytes;
713 		free_bytes = xlog_space_left(log, cycle, bytes);
714 		do {
715 			if (tic->t_flags & XLOG_TIC_PERM_RESERV)
716 				need_bytes = tic->t_unit_res*tic->t_cnt;
717 			else
718 				need_bytes = tic->t_unit_res;
719 			if (free_bytes < need_bytes && tail_lsn != 1)
720 				break;
721 			tail_lsn = 0;
722 			free_bytes -= need_bytes;
723 			sv_signal(&tic->t_sema);
724 			tic = tic->t_next;
725 		} while (tic != log->l_reserve_headq);
726 	}
727 	GRANT_UNLOCK(log, s);
728 }	/* xfs_log_move_tail */
729 
730 /*
731  * Determine if we have a transaction that has gone to disk
732  * that needs to be covered. Log activity needs to be idle (no AIL and
733  * nothing in the iclogs). And, we need to be in the right state indicating
734  * something has gone out.
735  */
736 int
737 xfs_log_need_covered(xfs_mount_t *mp)
738 {
739 	SPLDECL(s);
740 	int		needed = 0, gen;
741 	xlog_t		*log = mp->m_log;
742 	vfs_t		*vfsp = XFS_MTOVFS(mp);
743 
744 	if (fs_frozen(vfsp) || XFS_FORCED_SHUTDOWN(mp) ||
745 	    (vfsp->vfs_flag & VFS_RDONLY))
746 		return 0;
747 
748 	s = LOG_LOCK(log);
749 	if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
750 		(log->l_covered_state == XLOG_STATE_COVER_NEED2))
751 			&& !xfs_trans_first_ail(mp, &gen)
752 			&& xlog_iclogs_empty(log)) {
753 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
754 			log->l_covered_state = XLOG_STATE_COVER_DONE;
755 		else {
756 			ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
757 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
758 		}
759 		needed = 1;
760 	}
761 	LOG_UNLOCK(log, s);
762 	return(needed);
763 }
764 
765 /******************************************************************************
766  *
767  *	local routines
768  *
769  ******************************************************************************
770  */
771 
772 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
773  * The log manager must keep track of the last LR which was committed
774  * to disk.  The lsn of this LR will become the new tail_lsn whenever
775  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
776  * the situation where stuff could be written into the log but nothing
777  * was ever in the AIL when asked.  Eventually, we panic since the
778  * tail hits the head.
779  *
780  * We may be holding the log iclog lock upon entering this routine.
781  */
782 xfs_lsn_t
783 xlog_assign_tail_lsn(xfs_mount_t *mp)
784 {
785 	xfs_lsn_t tail_lsn;
786 	SPLDECL(s);
787 	xlog_t	  *log = mp->m_log;
788 
789 	tail_lsn = xfs_trans_tail_ail(mp);
790 	s = GRANT_LOCK(log);
791 	if (tail_lsn != 0) {
792 		log->l_tail_lsn = tail_lsn;
793 	} else {
794 		tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
795 	}
796 	GRANT_UNLOCK(log, s);
797 
798 	return tail_lsn;
799 }	/* xlog_assign_tail_lsn */
800 
801 
802 /*
803  * Return the space in the log between the tail and the head.  The head
804  * is passed in the cycle/bytes formal parms.  In the special case where
805  * the reserve head has wrapped passed the tail, this calculation is no
806  * longer valid.  In this case, just return 0 which means there is no space
807  * in the log.  This works for all places where this function is called
808  * with the reserve head.  Of course, if the write head were to ever
809  * wrap the tail, we should blow up.  Rather than catch this case here,
810  * we depend on other ASSERTions in other parts of the code.   XXXmiken
811  *
812  * This code also handles the case where the reservation head is behind
813  * the tail.  The details of this case are described below, but the end
814  * result is that we return the size of the log as the amount of space left.
815  */
816 int
817 xlog_space_left(xlog_t *log, int cycle, int bytes)
818 {
819 	int free_bytes;
820 	int tail_bytes;
821 	int tail_cycle;
822 
823 	tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
824 	tail_cycle = CYCLE_LSN(log->l_tail_lsn);
825 	if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
826 		free_bytes = log->l_logsize - (bytes - tail_bytes);
827 	} else if ((tail_cycle + 1) < cycle) {
828 		return 0;
829 	} else if (tail_cycle < cycle) {
830 		ASSERT(tail_cycle == (cycle - 1));
831 		free_bytes = tail_bytes - bytes;
832 	} else {
833 		/*
834 		 * The reservation head is behind the tail.
835 		 * In this case we just want to return the size of the
836 		 * log as the amount of space left.
837 		 */
838 		xfs_fs_cmn_err(CE_ALERT, log->l_mp,
839 			"xlog_space_left: head behind tail\n"
840 			"  tail_cycle = %d, tail_bytes = %d\n"
841 			"  GH   cycle = %d, GH   bytes = %d",
842 			tail_cycle, tail_bytes, cycle, bytes);
843 		ASSERT(0);
844 		free_bytes = log->l_logsize;
845 	}
846 	return free_bytes;
847 }	/* xlog_space_left */
848 
849 
850 /*
851  * Log function which is called when an io completes.
852  *
853  * The log manager needs its own routine, in order to control what
854  * happens with the buffer after the write completes.
855  */
856 void
857 xlog_iodone(xfs_buf_t *bp)
858 {
859 	xlog_in_core_t	*iclog;
860 	xlog_t		*l;
861 	int		aborted;
862 
863 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
864 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
865 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
866 	aborted = 0;
867 
868 	/*
869 	 * Some versions of cpp barf on the recursive definition of
870 	 * ic_log -> hic_fields.ic_log and expand ic_log twice when
871 	 * it is passed through two macros.  Workaround broken cpp.
872 	 */
873 	l = iclog->ic_log;
874 
875 	/*
876 	 * Race to shutdown the filesystem if we see an error.
877 	 */
878 	if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
879 			XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
880 		xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
881 		XFS_BUF_STALE(bp);
882 		xfs_force_shutdown(l->l_mp, XFS_LOG_IO_ERROR);
883 		/*
884 		 * This flag will be propagated to the trans-committed
885 		 * callback routines to let them know that the log-commit
886 		 * didn't succeed.
887 		 */
888 		aborted = XFS_LI_ABORTED;
889 	} else if (iclog->ic_state & XLOG_STATE_IOERROR) {
890 		aborted = XFS_LI_ABORTED;
891 	}
892 	xlog_state_done_syncing(iclog, aborted);
893 	if (!(XFS_BUF_ISASYNC(bp))) {
894 		/*
895 		 * Corresponding psema() will be done in bwrite().  If we don't
896 		 * vsema() here, panic.
897 		 */
898 		XFS_BUF_V_IODONESEMA(bp);
899 	}
900 }	/* xlog_iodone */
901 
902 /*
903  * The bdstrat callback function for log bufs. This gives us a central
904  * place to trap bufs in case we get hit by a log I/O error and need to
905  * shutdown. Actually, in practice, even when we didn't get a log error,
906  * we transition the iclogs to IOERROR state *after* flushing all existing
907  * iclogs to disk. This is because we don't want anymore new transactions to be
908  * started or completed afterwards.
909  */
910 STATIC int
911 xlog_bdstrat_cb(struct xfs_buf *bp)
912 {
913 	xlog_in_core_t *iclog;
914 
915 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
916 
917 	if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
918 	  /* note for irix bstrat will need  struct bdevsw passed
919 	   * Fix the following macro if the code ever is merged
920 	   */
921 	    XFS_bdstrat(bp);
922 		return 0;
923 	}
924 
925 	xfs_buftrace("XLOG__BDSTRAT IOERROR", bp);
926 	XFS_BUF_ERROR(bp, EIO);
927 	XFS_BUF_STALE(bp);
928 	xfs_biodone(bp);
929 	return (XFS_ERROR(EIO));
930 
931 
932 }
933 
934 /*
935  * Return size of each in-core log record buffer.
936  *
937  * Low memory machines only get 2 16KB buffers.  We don't want to waste
938  * memory here.  However, all other machines get at least 2 32KB buffers.
939  * The number is hard coded because we don't care about the minimum
940  * memory size, just 32MB systems.
941  *
942  * If the filesystem blocksize is too large, we may need to choose a
943  * larger size since the directory code currently logs entire blocks.
944  */
945 
946 STATIC void
947 xlog_get_iclog_buffer_size(xfs_mount_t	*mp,
948 			   xlog_t	*log)
949 {
950 	int size;
951 	int xhdrs;
952 
953 	if (mp->m_logbufs <= 0) {
954 		if (xfs_physmem <= btoc(128*1024*1024)) {
955 			log->l_iclog_bufs = XLOG_MIN_ICLOGS;
956 		} else if (xfs_physmem <= btoc(400*1024*1024)) {
957 			log->l_iclog_bufs = XLOG_MED_ICLOGS;
958 		} else {	/* 256K with 32K bufs */
959 			log->l_iclog_bufs = XLOG_MAX_ICLOGS;
960 		}
961 	} else {
962 		log->l_iclog_bufs = mp->m_logbufs;
963 	}
964 
965 	/*
966 	 * Buffer size passed in from mount system call.
967 	 */
968 	if (mp->m_logbsize > 0) {
969 		size = log->l_iclog_size = mp->m_logbsize;
970 		log->l_iclog_size_log = 0;
971 		while (size != 1) {
972 			log->l_iclog_size_log++;
973 			size >>= 1;
974 		}
975 
976 		if (XFS_SB_VERSION_HASLOGV2(&mp->m_sb)) {
977 			/* # headers = size / 32K
978 			 * one header holds cycles from 32K of data
979 			 */
980 
981 			xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
982 			if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
983 				xhdrs++;
984 			log->l_iclog_hsize = xhdrs << BBSHIFT;
985 			log->l_iclog_heads = xhdrs;
986 		} else {
987 			ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
988 			log->l_iclog_hsize = BBSIZE;
989 			log->l_iclog_heads = 1;
990 		}
991 		goto done;
992 	}
993 
994 	/*
995 	 * Special case machines that have less than 32MB of memory.
996 	 * All machines with more memory use 32KB buffers.
997 	 */
998 	if (xfs_physmem <= btoc(32*1024*1024)) {
999 		/* Don't change; min configuration */
1000 		log->l_iclog_size = XLOG_RECORD_BSIZE;		/* 16k */
1001 		log->l_iclog_size_log = XLOG_RECORD_BSHIFT;
1002 	} else {
1003 		log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;	/* 32k */
1004 		log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1005 	}
1006 
1007 	/* the default log size is 16k or 32k which is one header sector */
1008 	log->l_iclog_hsize = BBSIZE;
1009 	log->l_iclog_heads = 1;
1010 
1011 	/*
1012 	 * For 16KB, we use 3 32KB buffers.  For 32KB block sizes, we use
1013 	 * 4 32KB buffers.  For 64KB block sizes, we use 8 32KB buffers.
1014 	 */
1015 	if (mp->m_sb.sb_blocksize >= 16*1024) {
1016 		log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1017 		log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1018 		if (mp->m_logbufs <= 0) {
1019 			switch (mp->m_sb.sb_blocksize) {
1020 			    case 16*1024:			/* 16 KB */
1021 				log->l_iclog_bufs = 3;
1022 				break;
1023 			    case 32*1024:			/* 32 KB */
1024 				log->l_iclog_bufs = 4;
1025 				break;
1026 			    case 64*1024:			/* 64 KB */
1027 				log->l_iclog_bufs = 8;
1028 				break;
1029 			    default:
1030 				xlog_panic("XFS: Invalid blocksize");
1031 				break;
1032 			}
1033 		}
1034 	}
1035 
1036 done:	/* are we being asked to make the sizes selected above visible? */
1037 	if (mp->m_logbufs == 0)
1038 		mp->m_logbufs = log->l_iclog_bufs;
1039 	if (mp->m_logbsize == 0)
1040 		mp->m_logbsize = log->l_iclog_size;
1041 }	/* xlog_get_iclog_buffer_size */
1042 
1043 
1044 /*
1045  * This routine initializes some of the log structure for a given mount point.
1046  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1047  * some other stuff may be filled in too.
1048  */
1049 STATIC xlog_t *
1050 xlog_alloc_log(xfs_mount_t	*mp,
1051 	       xfs_buftarg_t	*log_target,
1052 	       xfs_daddr_t	blk_offset,
1053 	       int		num_bblks)
1054 {
1055 	xlog_t			*log;
1056 	xlog_rec_header_t	*head;
1057 	xlog_in_core_t		**iclogp;
1058 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1059 	xfs_buf_t		*bp;
1060 	int			i;
1061 	int			iclogsize;
1062 
1063 	log = (xlog_t *)kmem_zalloc(sizeof(xlog_t), KM_SLEEP);
1064 
1065 	log->l_mp	   = mp;
1066 	log->l_targ	   = log_target;
1067 	log->l_logsize     = BBTOB(num_bblks);
1068 	log->l_logBBstart  = blk_offset;
1069 	log->l_logBBsize   = num_bblks;
1070 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1071 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1072 
1073 	log->l_prev_block  = -1;
1074 	ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, 1, 0);
1075 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1076 	log->l_last_sync_lsn = log->l_tail_lsn;
1077 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1078 	log->l_grant_reserve_cycle = 1;
1079 	log->l_grant_write_cycle = 1;
1080 
1081 	if (XFS_SB_VERSION_HASSECTOR(&mp->m_sb)) {
1082 		log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1083 		ASSERT(log->l_sectbb_log <= mp->m_sectbb_log);
1084 		/* for larger sector sizes, must have v2 or external log */
1085 		ASSERT(log->l_sectbb_log == 0 ||
1086 			log->l_logBBstart == 0 ||
1087 			XFS_SB_VERSION_HASLOGV2(&mp->m_sb));
1088 		ASSERT(mp->m_sb.sb_logsectlog >= BBSHIFT);
1089 	}
1090 	log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1091 
1092 	xlog_get_iclog_buffer_size(mp, log);
1093 
1094 	bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1095 	XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1096 	XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1097 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1098 	ASSERT(XFS_BUF_ISBUSY(bp));
1099 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1100 	log->l_xbuf = bp;
1101 
1102 	spinlock_init(&log->l_icloglock, "iclog");
1103 	spinlock_init(&log->l_grant_lock, "grhead_iclog");
1104 	initnsema(&log->l_flushsema, 0, "ic-flush");
1105 	xlog_state_ticket_alloc(log);  /* wait until after icloglock inited */
1106 
1107 	/* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1108 	ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1109 
1110 	iclogp = &log->l_iclog;
1111 	/*
1112 	 * The amount of memory to allocate for the iclog structure is
1113 	 * rather funky due to the way the structure is defined.  It is
1114 	 * done this way so that we can use different sizes for machines
1115 	 * with different amounts of memory.  See the definition of
1116 	 * xlog_in_core_t in xfs_log_priv.h for details.
1117 	 */
1118 	iclogsize = log->l_iclog_size;
1119 	ASSERT(log->l_iclog_size >= 4096);
1120 	for (i=0; i < log->l_iclog_bufs; i++) {
1121 		*iclogp = (xlog_in_core_t *)
1122 			  kmem_zalloc(sizeof(xlog_in_core_t), KM_SLEEP);
1123 		iclog = *iclogp;
1124 		iclog->hic_data = (xlog_in_core_2_t *)
1125 			  kmem_zalloc(iclogsize, KM_SLEEP);
1126 
1127 		iclog->ic_prev = prev_iclog;
1128 		prev_iclog = iclog;
1129 		log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1130 
1131 		head = &iclog->ic_header;
1132 		memset(head, 0, sizeof(xlog_rec_header_t));
1133 		INT_SET(head->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1134 		INT_SET(head->h_version, ARCH_CONVERT,
1135 			XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1136 		INT_SET(head->h_size, ARCH_CONVERT, log->l_iclog_size);
1137 		/* new fields */
1138 		INT_SET(head->h_fmt, ARCH_CONVERT, XLOG_FMT);
1139 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1140 
1141 		bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1142 		XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1143 		XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1144 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1145 		iclog->ic_bp = bp;
1146 
1147 		iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1148 		iclog->ic_state = XLOG_STATE_ACTIVE;
1149 		iclog->ic_log = log;
1150 		iclog->ic_callback_tail = &(iclog->ic_callback);
1151 		iclog->ic_datap = (char *)iclog->hic_data + log->l_iclog_hsize;
1152 
1153 		ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1154 		ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1155 		sv_init(&iclog->ic_forcesema, SV_DEFAULT, "iclog-force");
1156 		sv_init(&iclog->ic_writesema, SV_DEFAULT, "iclog-write");
1157 
1158 		iclogp = &iclog->ic_next;
1159 	}
1160 	*iclogp = log->l_iclog;			/* complete ring */
1161 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1162 
1163 	return log;
1164 }	/* xlog_alloc_log */
1165 
1166 
1167 /*
1168  * Write out the commit record of a transaction associated with the given
1169  * ticket.  Return the lsn of the commit record.
1170  */
1171 STATIC int
1172 xlog_commit_record(xfs_mount_t  *mp,
1173 		   xlog_ticket_t *ticket,
1174 		   xlog_in_core_t **iclog,
1175 		   xfs_lsn_t	*commitlsnp)
1176 {
1177 	int		error;
1178 	xfs_log_iovec_t	reg[1];
1179 
1180 	reg[0].i_addr = NULL;
1181 	reg[0].i_len = 0;
1182 	XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
1183 
1184 	ASSERT_ALWAYS(iclog);
1185 	if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1186 			       iclog, XLOG_COMMIT_TRANS))) {
1187 		xfs_force_shutdown(mp, XFS_LOG_IO_ERROR);
1188 	}
1189 	return (error);
1190 }	/* xlog_commit_record */
1191 
1192 
1193 /*
1194  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1195  * log space.  This code pushes on the lsn which would supposedly free up
1196  * the 25% which we want to leave free.  We may need to adopt a policy which
1197  * pushes on an lsn which is further along in the log once we reach the high
1198  * water mark.  In this manner, we would be creating a low water mark.
1199  */
1200 void
1201 xlog_grant_push_ail(xfs_mount_t	*mp,
1202 		    int		need_bytes)
1203 {
1204     xlog_t	*log = mp->m_log;	/* pointer to the log */
1205     xfs_lsn_t	tail_lsn;		/* lsn of the log tail */
1206     xfs_lsn_t	threshold_lsn = 0;	/* lsn we'd like to be at */
1207     int		free_blocks;		/* free blocks left to write to */
1208     int		free_bytes;		/* free bytes left to write to */
1209     int		threshold_block;	/* block in lsn we'd like to be at */
1210     int		threshold_cycle;	/* lsn cycle we'd like to be at */
1211     int		free_threshold;
1212     SPLDECL(s);
1213 
1214     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1215 
1216     s = GRANT_LOCK(log);
1217     free_bytes = xlog_space_left(log,
1218 				 log->l_grant_reserve_cycle,
1219 				 log->l_grant_reserve_bytes);
1220     tail_lsn = log->l_tail_lsn;
1221     free_blocks = BTOBBT(free_bytes);
1222 
1223     /*
1224      * Set the threshold for the minimum number of free blocks in the
1225      * log to the maximum of what the caller needs, one quarter of the
1226      * log, and 256 blocks.
1227      */
1228     free_threshold = BTOBB(need_bytes);
1229     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1230     free_threshold = MAX(free_threshold, 256);
1231     if (free_blocks < free_threshold) {
1232 	threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1233 	threshold_cycle = CYCLE_LSN(tail_lsn);
1234 	if (threshold_block >= log->l_logBBsize) {
1235 	    threshold_block -= log->l_logBBsize;
1236 	    threshold_cycle += 1;
1237 	}
1238 	ASSIGN_ANY_LSN_HOST(threshold_lsn, threshold_cycle,
1239 		       threshold_block);
1240 
1241 	/* Don't pass in an lsn greater than the lsn of the last
1242 	 * log record known to be on disk.
1243 	 */
1244 	if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1245 	    threshold_lsn = log->l_last_sync_lsn;
1246     }
1247     GRANT_UNLOCK(log, s);
1248 
1249     /*
1250      * Get the transaction layer to kick the dirty buffers out to
1251      * disk asynchronously. No point in trying to do this if
1252      * the filesystem is shutting down.
1253      */
1254     if (threshold_lsn &&
1255 	!XLOG_FORCED_SHUTDOWN(log))
1256 	    xfs_trans_push_ail(mp, threshold_lsn);
1257 }	/* xlog_grant_push_ail */
1258 
1259 
1260 /*
1261  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1262  * fashion.  Previously, we should have moved the current iclog
1263  * ptr in the log to point to the next available iclog.  This allows further
1264  * write to continue while this code syncs out an iclog ready to go.
1265  * Before an in-core log can be written out, the data section must be scanned
1266  * to save away the 1st word of each BBSIZE block into the header.  We replace
1267  * it with the current cycle count.  Each BBSIZE block is tagged with the
1268  * cycle count because there in an implicit assumption that drives will
1269  * guarantee that entire 512 byte blocks get written at once.  In other words,
1270  * we can't have part of a 512 byte block written and part not written.  By
1271  * tagging each block, we will know which blocks are valid when recovering
1272  * after an unclean shutdown.
1273  *
1274  * This routine is single threaded on the iclog.  No other thread can be in
1275  * this routine with the same iclog.  Changing contents of iclog can there-
1276  * fore be done without grabbing the state machine lock.  Updating the global
1277  * log will require grabbing the lock though.
1278  *
1279  * The entire log manager uses a logical block numbering scheme.  Only
1280  * log_sync (and then only bwrite()) know about the fact that the log may
1281  * not start with block zero on a given device.  The log block start offset
1282  * is added immediately before calling bwrite().
1283  */
1284 
1285 int
1286 xlog_sync(xlog_t		*log,
1287 	  xlog_in_core_t	*iclog)
1288 {
1289 	xfs_caddr_t	dptr;		/* pointer to byte sized element */
1290 	xfs_buf_t	*bp;
1291 	int		i, ops;
1292 	uint		count;		/* byte count of bwrite */
1293 	uint		count_init;	/* initial count before roundup */
1294 	int		roundoff;       /* roundoff to BB or stripe */
1295 	int		split = 0;	/* split write into two regions */
1296 	int		error;
1297 	SPLDECL(s);
1298 	int		v2 = XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb);
1299 
1300 	XFS_STATS_INC(xs_log_writes);
1301 	ASSERT(iclog->ic_refcnt == 0);
1302 
1303 	/* Add for LR header */
1304 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1305 
1306 	/* Round out the log write size */
1307 	if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1308 		/* we have a v2 stripe unit to use */
1309 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1310 	} else {
1311 		count = BBTOB(BTOBB(count_init));
1312 	}
1313 	roundoff = count - count_init;
1314 	ASSERT(roundoff >= 0);
1315 	ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1316                 roundoff < log->l_mp->m_sb.sb_logsunit)
1317 		||
1318 		(log->l_mp->m_sb.sb_logsunit <= 1 &&
1319 		 roundoff < BBTOB(1)));
1320 
1321 	/* move grant heads by roundoff in sync */
1322 	s = GRANT_LOCK(log);
1323 	XLOG_GRANT_ADD_SPACE(log, roundoff, 'w');
1324 	XLOG_GRANT_ADD_SPACE(log, roundoff, 'r');
1325 	GRANT_UNLOCK(log, s);
1326 
1327 	/* put cycle number in every block */
1328 	xlog_pack_data(log, iclog, roundoff);
1329 
1330 	/* real byte length */
1331 	if (v2) {
1332 		INT_SET(iclog->ic_header.h_len,
1333 			ARCH_CONVERT,
1334 			iclog->ic_offset + roundoff);
1335 	} else {
1336 		INT_SET(iclog->ic_header.h_len, ARCH_CONVERT, iclog->ic_offset);
1337 	}
1338 
1339 	/* put ops count in correct order */
1340 	ops = iclog->ic_header.h_num_logops;
1341 	INT_SET(iclog->ic_header.h_num_logops, ARCH_CONVERT, ops);
1342 
1343 	bp	    = iclog->ic_bp;
1344 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1345 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1346 	XFS_BUF_SET_ADDR(bp, BLOCK_LSN(INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)));
1347 
1348 	XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1349 
1350 	/* Do we need to split this write into 2 parts? */
1351 	if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1352 		split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1353 		count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1354 		iclog->ic_bwritecnt = 2;	/* split into 2 writes */
1355 	} else {
1356 		iclog->ic_bwritecnt = 1;
1357 	}
1358 	XFS_BUF_SET_PTR(bp, (xfs_caddr_t) &(iclog->ic_header), count);
1359 	XFS_BUF_SET_FSPRIVATE(bp, iclog);	/* save for later */
1360 	XFS_BUF_BUSY(bp);
1361 	XFS_BUF_ASYNC(bp);
1362 	/*
1363 	 * Do an ordered write for the log block.
1364 	 *
1365 	 * It may not be needed to flush the first split block in the log wrap
1366 	 * case, but do it anyways to be safe -AK
1367 	 */
1368 	if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1369 		XFS_BUF_ORDERED(bp);
1370 
1371 	ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1372 	ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1373 
1374 	xlog_verify_iclog(log, iclog, count, B_TRUE);
1375 
1376 	/* account for log which doesn't start at block #0 */
1377 	XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1378 	/*
1379 	 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1380 	 * is shutting down.
1381 	 */
1382 	XFS_BUF_WRITE(bp);
1383 
1384 	if ((error = XFS_bwrite(bp))) {
1385 		xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1386 				  XFS_BUF_ADDR(bp));
1387 		return (error);
1388 	}
1389 	if (split) {
1390 		bp		= iclog->ic_log->l_xbuf;
1391 		ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1392 							(unsigned long)1);
1393 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1394 		XFS_BUF_SET_ADDR(bp, 0);	     /* logical 0 */
1395 		XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1396 					    (__psint_t)count), split);
1397 		XFS_BUF_SET_FSPRIVATE(bp, iclog);
1398 		XFS_BUF_BUSY(bp);
1399 		XFS_BUF_ASYNC(bp);
1400 		if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1401 			XFS_BUF_ORDERED(bp);
1402 		dptr = XFS_BUF_PTR(bp);
1403 		/*
1404 		 * Bump the cycle numbers at the start of each block
1405 		 * since this part of the buffer is at the start of
1406 		 * a new cycle.  Watch out for the header magic number
1407 		 * case, though.
1408 		 */
1409 		for (i=0; i<split; i += BBSIZE) {
1410 			INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1411 			if (INT_GET(*(uint *)dptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
1412 				INT_MOD(*(uint *)dptr, ARCH_CONVERT, +1);
1413 			dptr += BBSIZE;
1414 		}
1415 
1416 		ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1417 		ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1418 
1419 		/* account for internal log which does't start at block #0 */
1420 		XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1421 		XFS_BUF_WRITE(bp);
1422 		if ((error = XFS_bwrite(bp))) {
1423 			xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1424 					  bp, XFS_BUF_ADDR(bp));
1425 			return (error);
1426 		}
1427 	}
1428 	return (0);
1429 }	/* xlog_sync */
1430 
1431 
1432 /*
1433  * Unallocate a log structure
1434  */
1435 void
1436 xlog_unalloc_log(xlog_t *log)
1437 {
1438 	xlog_in_core_t	*iclog, *next_iclog;
1439 	xlog_ticket_t	*tic, *next_tic;
1440 	int		i;
1441 
1442 
1443 	iclog = log->l_iclog;
1444 	for (i=0; i<log->l_iclog_bufs; i++) {
1445 		sv_destroy(&iclog->ic_forcesema);
1446 		sv_destroy(&iclog->ic_writesema);
1447 		xfs_buf_free(iclog->ic_bp);
1448 #ifdef XFS_LOG_TRACE
1449 		if (iclog->ic_trace != NULL) {
1450 			ktrace_free(iclog->ic_trace);
1451 		}
1452 #endif
1453 		next_iclog = iclog->ic_next;
1454 		kmem_free(iclog->hic_data, log->l_iclog_size);
1455 		kmem_free(iclog, sizeof(xlog_in_core_t));
1456 		iclog = next_iclog;
1457 	}
1458 	freesema(&log->l_flushsema);
1459 	spinlock_destroy(&log->l_icloglock);
1460 	spinlock_destroy(&log->l_grant_lock);
1461 
1462 	/* XXXsup take a look at this again. */
1463 	if ((log->l_ticket_cnt != log->l_ticket_tcnt)  &&
1464 	    !XLOG_FORCED_SHUTDOWN(log)) {
1465 		xfs_fs_cmn_err(CE_WARN, log->l_mp,
1466 			"xlog_unalloc_log: (cnt: %d, total: %d)",
1467 			log->l_ticket_cnt, log->l_ticket_tcnt);
1468 		/* ASSERT(log->l_ticket_cnt == log->l_ticket_tcnt); */
1469 
1470 	} else {
1471 		tic = log->l_unmount_free;
1472 		while (tic) {
1473 			next_tic = tic->t_next;
1474 			kmem_free(tic, NBPP);
1475 			tic = next_tic;
1476 		}
1477 	}
1478 	xfs_buf_free(log->l_xbuf);
1479 #ifdef XFS_LOG_TRACE
1480 	if (log->l_trace != NULL) {
1481 		ktrace_free(log->l_trace);
1482 	}
1483 	if (log->l_grant_trace != NULL) {
1484 		ktrace_free(log->l_grant_trace);
1485 	}
1486 #endif
1487 	log->l_mp->m_log = NULL;
1488 	kmem_free(log, sizeof(xlog_t));
1489 }	/* xlog_unalloc_log */
1490 
1491 /*
1492  * Update counters atomically now that memcpy is done.
1493  */
1494 /* ARGSUSED */
1495 static inline void
1496 xlog_state_finish_copy(xlog_t		*log,
1497 		       xlog_in_core_t	*iclog,
1498 		       int		record_cnt,
1499 		       int		copy_bytes)
1500 {
1501 	SPLDECL(s);
1502 
1503 	s = LOG_LOCK(log);
1504 
1505 	iclog->ic_header.h_num_logops += record_cnt;
1506 	iclog->ic_offset += copy_bytes;
1507 
1508 	LOG_UNLOCK(log, s);
1509 }	/* xlog_state_finish_copy */
1510 
1511 
1512 
1513 
1514 /*
1515  * print out info relating to regions written which consume
1516  * the reservation
1517  */
1518 #if defined(XFS_LOG_RES_DEBUG)
1519 STATIC void
1520 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1521 {
1522 	uint i;
1523 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1524 
1525 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
1526 	static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1527 	    "bformat",
1528 	    "bchunk",
1529 	    "efi_format",
1530 	    "efd_format",
1531 	    "iformat",
1532 	    "icore",
1533 	    "iext",
1534 	    "ibroot",
1535 	    "ilocal",
1536 	    "iattr_ext",
1537 	    "iattr_broot",
1538 	    "iattr_local",
1539 	    "qformat",
1540 	    "dquot",
1541 	    "quotaoff",
1542 	    "LR header",
1543 	    "unmount",
1544 	    "commit",
1545 	    "trans header"
1546 	};
1547 	static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1548 	    "SETATTR_NOT_SIZE",
1549 	    "SETATTR_SIZE",
1550 	    "INACTIVE",
1551 	    "CREATE",
1552 	    "CREATE_TRUNC",
1553 	    "TRUNCATE_FILE",
1554 	    "REMOVE",
1555 	    "LINK",
1556 	    "RENAME",
1557 	    "MKDIR",
1558 	    "RMDIR",
1559 	    "SYMLINK",
1560 	    "SET_DMATTRS",
1561 	    "GROWFS",
1562 	    "STRAT_WRITE",
1563 	    "DIOSTRAT",
1564 	    "WRITE_SYNC",
1565 	    "WRITEID",
1566 	    "ADDAFORK",
1567 	    "ATTRINVAL",
1568 	    "ATRUNCATE",
1569 	    "ATTR_SET",
1570 	    "ATTR_RM",
1571 	    "ATTR_FLAG",
1572 	    "CLEAR_AGI_BUCKET",
1573 	    "QM_SBCHANGE",
1574 	    "DUMMY1",
1575 	    "DUMMY2",
1576 	    "QM_QUOTAOFF",
1577 	    "QM_DQALLOC",
1578 	    "QM_SETQLIM",
1579 	    "QM_DQCLUSTER",
1580 	    "QM_QINOCREATE",
1581 	    "QM_QUOTAOFF_END",
1582 	    "SB_UNIT",
1583 	    "FSYNC_TS",
1584 	    "GROWFSRT_ALLOC",
1585 	    "GROWFSRT_ZERO",
1586 	    "GROWFSRT_FREE",
1587 	    "SWAPEXT"
1588 	};
1589 
1590 	xfs_fs_cmn_err(CE_WARN, mp,
1591 			"xfs_log_write: reservation summary:\n"
1592 			"  trans type  = %s (%u)\n"
1593 			"  unit res    = %d bytes\n"
1594 			"  current res = %d bytes\n"
1595 			"  total reg   = %u bytes (o/flow = %u bytes)\n"
1596 			"  ophdrs      = %u (ophdr space = %u bytes)\n"
1597 			"  ophdr + reg = %u bytes\n"
1598 			"  num regions = %u\n",
1599 			((ticket->t_trans_type <= 0 ||
1600 			  ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1601 			  "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1602 			ticket->t_trans_type,
1603 			ticket->t_unit_res,
1604 			ticket->t_curr_res,
1605 			ticket->t_res_arr_sum, ticket->t_res_o_flow,
1606 			ticket->t_res_num_ophdrs, ophdr_spc,
1607 			ticket->t_res_arr_sum +
1608 			  ticket->t_res_o_flow + ophdr_spc,
1609 			ticket->t_res_num);
1610 
1611 	for (i = 0; i < ticket->t_res_num; i++) {
1612 	   	uint r_type = ticket->t_res_arr[i].r_type;
1613 		cmn_err(CE_WARN,
1614 			    "region[%u]: %s - %u bytes\n",
1615 			    i,
1616 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1617 			    "bad-rtype" : res_type_str[r_type-1]),
1618 			    ticket->t_res_arr[i].r_len);
1619 	}
1620 }
1621 #else
1622 #define xlog_print_tic_res(mp, ticket)
1623 #endif
1624 
1625 /*
1626  * Write some region out to in-core log
1627  *
1628  * This will be called when writing externally provided regions or when
1629  * writing out a commit record for a given transaction.
1630  *
1631  * General algorithm:
1632  *	1. Find total length of this write.  This may include adding to the
1633  *		lengths passed in.
1634  *	2. Check whether we violate the tickets reservation.
1635  *	3. While writing to this iclog
1636  *	    A. Reserve as much space in this iclog as can get
1637  *	    B. If this is first write, save away start lsn
1638  *	    C. While writing this region:
1639  *		1. If first write of transaction, write start record
1640  *		2. Write log operation header (header per region)
1641  *		3. Find out if we can fit entire region into this iclog
1642  *		4. Potentially, verify destination memcpy ptr
1643  *		5. Memcpy (partial) region
1644  *		6. If partial copy, release iclog; otherwise, continue
1645  *			copying more regions into current iclog
1646  *	4. Mark want sync bit (in simulation mode)
1647  *	5. Release iclog for potential flush to on-disk log.
1648  *
1649  * ERRORS:
1650  * 1.	Panic if reservation is overrun.  This should never happen since
1651  *	reservation amounts are generated internal to the filesystem.
1652  * NOTES:
1653  * 1. Tickets are single threaded data structures.
1654  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1655  *	syncing routine.  When a single log_write region needs to span
1656  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1657  *	on all log operation writes which don't contain the end of the
1658  *	region.  The XLOG_END_TRANS bit is used for the in-core log
1659  *	operation which contains the end of the continued log_write region.
1660  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1661  *	we don't really know exactly how much space will be used.  As a result,
1662  *	we don't update ic_offset until the end when we know exactly how many
1663  *	bytes have been written out.
1664  */
1665 int
1666 xlog_write(xfs_mount_t *	mp,
1667 	   xfs_log_iovec_t	reg[],
1668 	   int			nentries,
1669 	   xfs_log_ticket_t	tic,
1670 	   xfs_lsn_t		*start_lsn,
1671 	   xlog_in_core_t	**commit_iclog,
1672 	   uint			flags)
1673 {
1674     xlog_t	     *log    = mp->m_log;
1675     xlog_ticket_t    *ticket = (xlog_ticket_t *)tic;
1676     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1677     xlog_in_core_t   *iclog;	     /* ptr to current in-core log */
1678     __psint_t	     ptr;	     /* copy address into data region */
1679     int		     len;	     /* # xlog_write() bytes 2 still copy */
1680     int		     index;	     /* region index currently copying */
1681     int		     log_offset;     /* offset (from 0) into data region */
1682     int		     start_rec_copy; /* # bytes to copy for start record */
1683     int		     partial_copy;   /* did we split a region? */
1684     int		     partial_copy_len;/* # bytes copied if split region */
1685     int		     need_copy;	     /* # bytes need to memcpy this region */
1686     int		     copy_len;	     /* # bytes actually memcpy'ing */
1687     int		     copy_off;	     /* # bytes from entry start */
1688     int		     contwr;	     /* continued write of in-core log? */
1689     int		     error;
1690     int		     record_cnt = 0, data_cnt = 0;
1691 
1692     partial_copy_len = partial_copy = 0;
1693 
1694     /* Calculate potential maximum space.  Each region gets its own
1695      * xlog_op_header_t and may need to be double word aligned.
1696      */
1697     len = 0;
1698     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1699 	len += sizeof(xlog_op_header_t);
1700 	XLOG_TIC_ADD_OPHDR(ticket);
1701     }
1702 
1703     for (index = 0; index < nentries; index++) {
1704 	len += sizeof(xlog_op_header_t);	    /* each region gets >= 1 */
1705 	XLOG_TIC_ADD_OPHDR(ticket);
1706 	len += reg[index].i_len;
1707 	XLOG_TIC_ADD_REGION(ticket, reg[index].i_len, reg[index].i_type);
1708     }
1709     contwr = *start_lsn = 0;
1710 
1711     if (ticket->t_curr_res < len) {
1712 	xlog_print_tic_res(mp, ticket);
1713 #ifdef DEBUG
1714 	xlog_panic(
1715 		"xfs_log_write: reservation ran out. Need to up reservation");
1716 #else
1717 	/* Customer configurable panic */
1718 	xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1719 		"xfs_log_write: reservation ran out. Need to up reservation");
1720 	/* If we did not panic, shutdown the filesystem */
1721 	xfs_force_shutdown(mp, XFS_CORRUPT_INCORE);
1722 #endif
1723     } else
1724 	ticket->t_curr_res -= len;
1725 
1726     for (index = 0; index < nentries; ) {
1727 	if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1728 					       &contwr, &log_offset)))
1729 		return (error);
1730 
1731 	ASSERT(log_offset <= iclog->ic_size - 1);
1732 	ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1733 
1734 	/* start_lsn is the first lsn written to. That's all we need. */
1735 	if (! *start_lsn)
1736 	    *start_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
1737 
1738 	/* This loop writes out as many regions as can fit in the amount
1739 	 * of space which was allocated by xlog_state_get_iclog_space().
1740 	 */
1741 	while (index < nentries) {
1742 	    ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1743 	    ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1744 	    start_rec_copy = 0;
1745 
1746 	    /* If first write for transaction, insert start record.
1747 	     * We can't be trying to commit if we are inited.  We can't
1748 	     * have any "partial_copy" if we are inited.
1749 	     */
1750 	    if (ticket->t_flags & XLOG_TIC_INITED) {
1751 		logop_head		= (xlog_op_header_t *)ptr;
1752 		INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1753 		logop_head->oh_clientid = ticket->t_clientid;
1754 		logop_head->oh_len	= 0;
1755 		logop_head->oh_flags    = XLOG_START_TRANS;
1756 		logop_head->oh_res2	= 0;
1757 		ticket->t_flags		&= ~XLOG_TIC_INITED;	/* clear bit */
1758 		record_cnt++;
1759 
1760 		start_rec_copy = sizeof(xlog_op_header_t);
1761 		xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1762 	    }
1763 
1764 	    /* Copy log operation header directly into data section */
1765 	    logop_head			= (xlog_op_header_t *)ptr;
1766 	    INT_SET(logop_head->oh_tid, ARCH_CONVERT, ticket->t_tid);
1767 	    logop_head->oh_clientid	= ticket->t_clientid;
1768 	    logop_head->oh_res2		= 0;
1769 
1770 	    /* header copied directly */
1771 	    xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1772 
1773 	    /* are we copying a commit or unmount record? */
1774 	    logop_head->oh_flags = flags;
1775 
1776 	    /*
1777 	     * We've seen logs corrupted with bad transaction client
1778 	     * ids.  This makes sure that XFS doesn't generate them on.
1779 	     * Turn this into an EIO and shut down the filesystem.
1780 	     */
1781 	    switch (logop_head->oh_clientid)  {
1782 	    case XFS_TRANSACTION:
1783 	    case XFS_VOLUME:
1784 	    case XFS_LOG:
1785 		break;
1786 	    default:
1787 		xfs_fs_cmn_err(CE_WARN, mp,
1788 		    "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1789 		    logop_head->oh_clientid, tic);
1790 		return XFS_ERROR(EIO);
1791 	    }
1792 
1793 	    /* Partial write last time? => (partial_copy != 0)
1794 	     * need_copy is the amount we'd like to copy if everything could
1795 	     * fit in the current memcpy.
1796 	     */
1797 	    need_copy =	reg[index].i_len - partial_copy_len;
1798 
1799 	    copy_off = partial_copy_len;
1800 	    if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1801 		INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len = need_copy);
1802 		if (partial_copy)
1803 		    logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1804 		partial_copy_len = partial_copy = 0;
1805 	    } else {					    /* partial write */
1806 		copy_len = iclog->ic_size - log_offset;
1807 		INT_SET(logop_head->oh_len, ARCH_CONVERT, copy_len);
1808 		logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1809 		if (partial_copy)
1810 			logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1811 		partial_copy_len += copy_len;
1812 		partial_copy++;
1813 		len += sizeof(xlog_op_header_t); /* from splitting of region */
1814 		/* account for new log op header */
1815 		ticket->t_curr_res -= sizeof(xlog_op_header_t);
1816 		XLOG_TIC_ADD_OPHDR(ticket);
1817 	    }
1818 	    xlog_verify_dest_ptr(log, ptr);
1819 
1820 	    /* copy region */
1821 	    ASSERT(copy_len >= 0);
1822 	    memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1823 	    xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1824 
1825 	    /* make copy_len total bytes copied, including headers */
1826 	    copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1827 	    record_cnt++;
1828 	    data_cnt += contwr ? copy_len : 0;
1829 	    if (partial_copy) {			/* copied partial region */
1830 		    /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1831 		    xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1832 		    record_cnt = data_cnt = 0;
1833 		    if ((error = xlog_state_release_iclog(log, iclog)))
1834 			    return (error);
1835 		    break;			/* don't increment index */
1836 	    } else {				/* copied entire region */
1837 		index++;
1838 		partial_copy_len = partial_copy = 0;
1839 
1840 		if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1841 		    xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1842 		    record_cnt = data_cnt = 0;
1843 		    xlog_state_want_sync(log, iclog);
1844 		    if (commit_iclog) {
1845 			ASSERT(flags & XLOG_COMMIT_TRANS);
1846 			*commit_iclog = iclog;
1847 		    } else if ((error = xlog_state_release_iclog(log, iclog)))
1848 			   return (error);
1849 		    if (index == nentries)
1850 			    return 0;		/* we are done */
1851 		    else
1852 			    break;
1853 		}
1854 	    } /* if (partial_copy) */
1855 	} /* while (index < nentries) */
1856     } /* for (index = 0; index < nentries; ) */
1857     ASSERT(len == 0);
1858 
1859     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1860     if (commit_iclog) {
1861 	ASSERT(flags & XLOG_COMMIT_TRANS);
1862 	*commit_iclog = iclog;
1863 	return 0;
1864     }
1865     return (xlog_state_release_iclog(log, iclog));
1866 }	/* xlog_write */
1867 
1868 
1869 /*****************************************************************************
1870  *
1871  *		State Machine functions
1872  *
1873  *****************************************************************************
1874  */
1875 
1876 /* Clean iclogs starting from the head.  This ordering must be
1877  * maintained, so an iclog doesn't become ACTIVE beyond one that
1878  * is SYNCING.  This is also required to maintain the notion that we use
1879  * a counting semaphore to hold off would be writers to the log when every
1880  * iclog is trying to sync to disk.
1881  *
1882  * State Change: DIRTY -> ACTIVE
1883  */
1884 STATIC void
1885 xlog_state_clean_log(xlog_t *log)
1886 {
1887 	xlog_in_core_t	*iclog;
1888 	int changed = 0;
1889 
1890 	iclog = log->l_iclog;
1891 	do {
1892 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
1893 			iclog->ic_state	= XLOG_STATE_ACTIVE;
1894 			iclog->ic_offset       = 0;
1895 			iclog->ic_callback	= NULL;   /* don't need to free */
1896 			/*
1897 			 * If the number of ops in this iclog indicate it just
1898 			 * contains the dummy transaction, we can
1899 			 * change state into IDLE (the second time around).
1900 			 * Otherwise we should change the state into
1901 			 * NEED a dummy.
1902 			 * We don't need to cover the dummy.
1903 			 */
1904 			if (!changed &&
1905 			   (INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT) == XLOG_COVER_OPS)) {
1906 				changed = 1;
1907 			} else {
1908 				/*
1909 				 * We have two dirty iclogs so start over
1910 				 * This could also be num of ops indicates
1911 				 * this is not the dummy going out.
1912 				 */
1913 				changed = 2;
1914 			}
1915 			iclog->ic_header.h_num_logops = 0;
1916 			memset(iclog->ic_header.h_cycle_data, 0,
1917 			      sizeof(iclog->ic_header.h_cycle_data));
1918 			iclog->ic_header.h_lsn = 0;
1919 		} else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1920 			/* do nothing */;
1921 		else
1922 			break;	/* stop cleaning */
1923 		iclog = iclog->ic_next;
1924 	} while (iclog != log->l_iclog);
1925 
1926 	/* log is locked when we are called */
1927 	/*
1928 	 * Change state for the dummy log recording.
1929 	 * We usually go to NEED. But we go to NEED2 if the changed indicates
1930 	 * we are done writing the dummy record.
1931 	 * If we are done with the second dummy recored (DONE2), then
1932 	 * we go to IDLE.
1933 	 */
1934 	if (changed) {
1935 		switch (log->l_covered_state) {
1936 		case XLOG_STATE_COVER_IDLE:
1937 		case XLOG_STATE_COVER_NEED:
1938 		case XLOG_STATE_COVER_NEED2:
1939 			log->l_covered_state = XLOG_STATE_COVER_NEED;
1940 			break;
1941 
1942 		case XLOG_STATE_COVER_DONE:
1943 			if (changed == 1)
1944 				log->l_covered_state = XLOG_STATE_COVER_NEED2;
1945 			else
1946 				log->l_covered_state = XLOG_STATE_COVER_NEED;
1947 			break;
1948 
1949 		case XLOG_STATE_COVER_DONE2:
1950 			if (changed == 1)
1951 				log->l_covered_state = XLOG_STATE_COVER_IDLE;
1952 			else
1953 				log->l_covered_state = XLOG_STATE_COVER_NEED;
1954 			break;
1955 
1956 		default:
1957 			ASSERT(0);
1958 		}
1959 	}
1960 }	/* xlog_state_clean_log */
1961 
1962 STATIC xfs_lsn_t
1963 xlog_get_lowest_lsn(
1964 	xlog_t		*log)
1965 {
1966 	xlog_in_core_t  *lsn_log;
1967 	xfs_lsn_t	lowest_lsn, lsn;
1968 
1969 	lsn_log = log->l_iclog;
1970 	lowest_lsn = 0;
1971 	do {
1972 	    if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
1973 		lsn = INT_GET(lsn_log->ic_header.h_lsn, ARCH_CONVERT);
1974 		if ((lsn && !lowest_lsn) ||
1975 		    (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
1976 			lowest_lsn = lsn;
1977 		}
1978 	    }
1979 	    lsn_log = lsn_log->ic_next;
1980 	} while (lsn_log != log->l_iclog);
1981 	return(lowest_lsn);
1982 }
1983 
1984 
1985 STATIC void
1986 xlog_state_do_callback(
1987 	xlog_t		*log,
1988 	int		aborted,
1989 	xlog_in_core_t	*ciclog)
1990 {
1991 	xlog_in_core_t	   *iclog;
1992 	xlog_in_core_t	   *first_iclog;	/* used to know when we've
1993 						 * processed all iclogs once */
1994 	xfs_log_callback_t *cb, *cb_next;
1995 	int		   flushcnt = 0;
1996 	xfs_lsn_t	   lowest_lsn;
1997 	int		   ioerrors;	/* counter: iclogs with errors */
1998 	int		   loopdidcallbacks; /* flag: inner loop did callbacks*/
1999 	int		   funcdidcallbacks; /* flag: function did callbacks */
2000 	int		   repeats;	/* for issuing console warnings if
2001 					 * looping too many times */
2002 	SPLDECL(s);
2003 
2004 	s = LOG_LOCK(log);
2005 	first_iclog = iclog = log->l_iclog;
2006 	ioerrors = 0;
2007 	funcdidcallbacks = 0;
2008 	repeats = 0;
2009 
2010 	do {
2011 		/*
2012 		 * Scan all iclogs starting with the one pointed to by the
2013 		 * log.  Reset this starting point each time the log is
2014 		 * unlocked (during callbacks).
2015 		 *
2016 		 * Keep looping through iclogs until one full pass is made
2017 		 * without running any callbacks.
2018 		 */
2019 		first_iclog = log->l_iclog;
2020 		iclog = log->l_iclog;
2021 		loopdidcallbacks = 0;
2022 		repeats++;
2023 
2024 		do {
2025 
2026 			/* skip all iclogs in the ACTIVE & DIRTY states */
2027 			if (iclog->ic_state &
2028 			    (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2029 				iclog = iclog->ic_next;
2030 				continue;
2031 			}
2032 
2033 			/*
2034 			 * Between marking a filesystem SHUTDOWN and stopping
2035 			 * the log, we do flush all iclogs to disk (if there
2036 			 * wasn't a log I/O error). So, we do want things to
2037 			 * go smoothly in case of just a SHUTDOWN  w/o a
2038 			 * LOG_IO_ERROR.
2039 			 */
2040 			if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2041 				/*
2042 				 * Can only perform callbacks in order.  Since
2043 				 * this iclog is not in the DONE_SYNC/
2044 				 * DO_CALLBACK state, we skip the rest and
2045 				 * just try to clean up.  If we set our iclog
2046 				 * to DO_CALLBACK, we will not process it when
2047 				 * we retry since a previous iclog is in the
2048 				 * CALLBACK and the state cannot change since
2049 				 * we are holding the LOG_LOCK.
2050 				 */
2051 				if (!(iclog->ic_state &
2052 					(XLOG_STATE_DONE_SYNC |
2053 						 XLOG_STATE_DO_CALLBACK))) {
2054 					if (ciclog && (ciclog->ic_state ==
2055 							XLOG_STATE_DONE_SYNC)) {
2056 						ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2057 					}
2058 					break;
2059 				}
2060 				/*
2061 				 * We now have an iclog that is in either the
2062 				 * DO_CALLBACK or DONE_SYNC states. The other
2063 				 * states (WANT_SYNC, SYNCING, or CALLBACK were
2064 				 * caught by the above if and are going to
2065 				 * clean (i.e. we aren't doing their callbacks)
2066 				 * see the above if.
2067 				 */
2068 
2069 				/*
2070 				 * We will do one more check here to see if we
2071 				 * have chased our tail around.
2072 				 */
2073 
2074 				lowest_lsn = xlog_get_lowest_lsn(log);
2075 				if (lowest_lsn && (
2076 					XFS_LSN_CMP(
2077 						lowest_lsn,
2078 						INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2079 					)<0)) {
2080 					iclog = iclog->ic_next;
2081 					continue; /* Leave this iclog for
2082 						   * another thread */
2083 				}
2084 
2085 				iclog->ic_state = XLOG_STATE_CALLBACK;
2086 
2087 				LOG_UNLOCK(log, s);
2088 
2089 				/* l_last_sync_lsn field protected by
2090 				 * GRANT_LOCK. Don't worry about iclog's lsn.
2091 				 * No one else can be here except us.
2092 				 */
2093 				s = GRANT_LOCK(log);
2094 				ASSERT(XFS_LSN_CMP(
2095 						log->l_last_sync_lsn,
2096 						INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT)
2097 					)<=0);
2098 				log->l_last_sync_lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2099 				GRANT_UNLOCK(log, s);
2100 
2101 				/*
2102 				 * Keep processing entries in the callback list
2103 				 * until we come around and it is empty.  We
2104 				 * need to atomically see that the list is
2105 				 * empty and change the state to DIRTY so that
2106 				 * we don't miss any more callbacks being added.
2107 				 */
2108 				s = LOG_LOCK(log);
2109 			} else {
2110 				ioerrors++;
2111 			}
2112 			cb = iclog->ic_callback;
2113 
2114 			while (cb != 0) {
2115 				iclog->ic_callback_tail = &(iclog->ic_callback);
2116 				iclog->ic_callback = NULL;
2117 				LOG_UNLOCK(log, s);
2118 
2119 				/* perform callbacks in the order given */
2120 				for (; cb != 0; cb = cb_next) {
2121 					cb_next = cb->cb_next;
2122 					cb->cb_func(cb->cb_arg, aborted);
2123 				}
2124 				s = LOG_LOCK(log);
2125 				cb = iclog->ic_callback;
2126 			}
2127 
2128 			loopdidcallbacks++;
2129 			funcdidcallbacks++;
2130 
2131 			ASSERT(iclog->ic_callback == 0);
2132 			if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2133 				iclog->ic_state = XLOG_STATE_DIRTY;
2134 
2135 			/*
2136 			 * Transition from DIRTY to ACTIVE if applicable.
2137 			 * NOP if STATE_IOERROR.
2138 			 */
2139 			xlog_state_clean_log(log);
2140 
2141 			/* wake up threads waiting in xfs_log_force() */
2142 			sv_broadcast(&iclog->ic_forcesema);
2143 
2144 			iclog = iclog->ic_next;
2145 		} while (first_iclog != iclog);
2146 		if (repeats && (repeats % 10) == 0) {
2147 			xfs_fs_cmn_err(CE_WARN, log->l_mp,
2148 				"xlog_state_do_callback: looping %d", repeats);
2149 		}
2150 	} while (!ioerrors && loopdidcallbacks);
2151 
2152 	/*
2153 	 * make one last gasp attempt to see if iclogs are being left in
2154 	 * limbo..
2155 	 */
2156 #ifdef DEBUG
2157 	if (funcdidcallbacks) {
2158 		first_iclog = iclog = log->l_iclog;
2159 		do {
2160 			ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2161 			/*
2162 			 * Terminate the loop if iclogs are found in states
2163 			 * which will cause other threads to clean up iclogs.
2164 			 *
2165 			 * SYNCING - i/o completion will go through logs
2166 			 * DONE_SYNC - interrupt thread should be waiting for
2167 			 *              LOG_LOCK
2168 			 * IOERROR - give up hope all ye who enter here
2169 			 */
2170 			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2171 			    iclog->ic_state == XLOG_STATE_SYNCING ||
2172 			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2173 			    iclog->ic_state == XLOG_STATE_IOERROR )
2174 				break;
2175 			iclog = iclog->ic_next;
2176 		} while (first_iclog != iclog);
2177 	}
2178 #endif
2179 
2180 	if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) {
2181 		flushcnt = log->l_flushcnt;
2182 		log->l_flushcnt = 0;
2183 	}
2184 	LOG_UNLOCK(log, s);
2185 	while (flushcnt--)
2186 		vsema(&log->l_flushsema);
2187 }	/* xlog_state_do_callback */
2188 
2189 
2190 /*
2191  * Finish transitioning this iclog to the dirty state.
2192  *
2193  * Make sure that we completely execute this routine only when this is
2194  * the last call to the iclog.  There is a good chance that iclog flushes,
2195  * when we reach the end of the physical log, get turned into 2 separate
2196  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2197  * routine.  By using the reference count bwritecnt, we guarantee that only
2198  * the second completion goes through.
2199  *
2200  * Callbacks could take time, so they are done outside the scope of the
2201  * global state machine log lock.  Assume that the calls to cvsema won't
2202  * take a long time.  At least we know it won't sleep.
2203  */
2204 void
2205 xlog_state_done_syncing(
2206 	xlog_in_core_t	*iclog,
2207 	int		aborted)
2208 {
2209 	xlog_t		   *log = iclog->ic_log;
2210 	SPLDECL(s);
2211 
2212 	s = LOG_LOCK(log);
2213 
2214 	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2215 	       iclog->ic_state == XLOG_STATE_IOERROR);
2216 	ASSERT(iclog->ic_refcnt == 0);
2217 	ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2218 
2219 
2220 	/*
2221 	 * If we got an error, either on the first buffer, or in the case of
2222 	 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2223 	 * and none should ever be attempted to be written to disk
2224 	 * again.
2225 	 */
2226 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
2227 		if (--iclog->ic_bwritecnt == 1) {
2228 			LOG_UNLOCK(log, s);
2229 			return;
2230 		}
2231 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2232 	}
2233 
2234 	/*
2235 	 * Someone could be sleeping prior to writing out the next
2236 	 * iclog buffer, we wake them all, one will get to do the
2237 	 * I/O, the others get to wait for the result.
2238 	 */
2239 	sv_broadcast(&iclog->ic_writesema);
2240 	LOG_UNLOCK(log, s);
2241 	xlog_state_do_callback(log, aborted, iclog);	/* also cleans log */
2242 }	/* xlog_state_done_syncing */
2243 
2244 
2245 /*
2246  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2247  * sleep.  The flush semaphore is set to the number of in-core buffers and
2248  * decremented around disk syncing.  Therefore, if all buffers are syncing,
2249  * this semaphore will cause new writes to sleep until a sync completes.
2250  * Otherwise, this code just does p() followed by v().  This approximates
2251  * a sleep/wakeup except we can't race.
2252  *
2253  * The in-core logs are used in a circular fashion. They are not used
2254  * out-of-order even when an iclog past the head is free.
2255  *
2256  * return:
2257  *	* log_offset where xlog_write() can start writing into the in-core
2258  *		log's data space.
2259  *	* in-core log pointer to which xlog_write() should write.
2260  *	* boolean indicating this is a continued write to an in-core log.
2261  *		If this is the last write, then the in-core log's offset field
2262  *		needs to be incremented, depending on the amount of data which
2263  *		is copied.
2264  */
2265 int
2266 xlog_state_get_iclog_space(xlog_t	  *log,
2267 			   int		  len,
2268 			   xlog_in_core_t **iclogp,
2269 			   xlog_ticket_t  *ticket,
2270 			   int		  *continued_write,
2271 			   int		  *logoffsetp)
2272 {
2273 	SPLDECL(s);
2274 	int		  log_offset;
2275 	xlog_rec_header_t *head;
2276 	xlog_in_core_t	  *iclog;
2277 	int		  error;
2278 
2279 restart:
2280 	s = LOG_LOCK(log);
2281 	if (XLOG_FORCED_SHUTDOWN(log)) {
2282 		LOG_UNLOCK(log, s);
2283 		return XFS_ERROR(EIO);
2284 	}
2285 
2286 	iclog = log->l_iclog;
2287 	if (! (iclog->ic_state == XLOG_STATE_ACTIVE)) {
2288 		log->l_flushcnt++;
2289 		LOG_UNLOCK(log, s);
2290 		xlog_trace_iclog(iclog, XLOG_TRACE_SLEEP_FLUSH);
2291 		XFS_STATS_INC(xs_log_noiclogs);
2292 		/* Ensure that log writes happen */
2293 		psema(&log->l_flushsema, PINOD);
2294 		goto restart;
2295 	}
2296 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2297 	head = &iclog->ic_header;
2298 
2299 	iclog->ic_refcnt++;			/* prevents sync */
2300 	log_offset = iclog->ic_offset;
2301 
2302 	/* On the 1st write to an iclog, figure out lsn.  This works
2303 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2304 	 * committing to.  If the offset is set, that's how many blocks
2305 	 * must be written.
2306 	 */
2307 	if (log_offset == 0) {
2308 		ticket->t_curr_res -= log->l_iclog_hsize;
2309 		XLOG_TIC_ADD_REGION(ticket,
2310 				    log->l_iclog_hsize,
2311 				    XLOG_REG_TYPE_LRHEADER);
2312 		INT_SET(head->h_cycle, ARCH_CONVERT, log->l_curr_cycle);
2313 		ASSIGN_LSN(head->h_lsn, log);
2314 		ASSERT(log->l_curr_block >= 0);
2315 	}
2316 
2317 	/* If there is enough room to write everything, then do it.  Otherwise,
2318 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2319 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2320 	 * until you know exactly how many bytes get copied.  Therefore, wait
2321 	 * until later to update ic_offset.
2322 	 *
2323 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2324 	 * can fit into remaining data section.
2325 	 */
2326 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2327 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2328 
2329 		/* If I'm the only one writing to this iclog, sync it to disk */
2330 		if (iclog->ic_refcnt == 1) {
2331 			LOG_UNLOCK(log, s);
2332 			if ((error = xlog_state_release_iclog(log, iclog)))
2333 				return (error);
2334 		} else {
2335 			iclog->ic_refcnt--;
2336 			LOG_UNLOCK(log, s);
2337 		}
2338 		goto restart;
2339 	}
2340 
2341 	/* Do we have enough room to write the full amount in the remainder
2342 	 * of this iclog?  Or must we continue a write on the next iclog and
2343 	 * mark this iclog as completely taken?  In the case where we switch
2344 	 * iclogs (to mark it taken), this particular iclog will release/sync
2345 	 * to disk in xlog_write().
2346 	 */
2347 	if (len <= iclog->ic_size - iclog->ic_offset) {
2348 		*continued_write = 0;
2349 		iclog->ic_offset += len;
2350 	} else {
2351 		*continued_write = 1;
2352 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2353 	}
2354 	*iclogp = iclog;
2355 
2356 	ASSERT(iclog->ic_offset <= iclog->ic_size);
2357 	LOG_UNLOCK(log, s);
2358 
2359 	*logoffsetp = log_offset;
2360 	return 0;
2361 }	/* xlog_state_get_iclog_space */
2362 
2363 /*
2364  * Atomically get the log space required for a log ticket.
2365  *
2366  * Once a ticket gets put onto the reserveq, it will only return after
2367  * the needed reservation is satisfied.
2368  */
2369 STATIC int
2370 xlog_grant_log_space(xlog_t	   *log,
2371 		     xlog_ticket_t *tic)
2372 {
2373 	int		 free_bytes;
2374 	int		 need_bytes;
2375 	SPLDECL(s);
2376 #ifdef DEBUG
2377 	xfs_lsn_t	 tail_lsn;
2378 #endif
2379 
2380 
2381 #ifdef DEBUG
2382 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2383 		panic("grant Recovery problem");
2384 #endif
2385 
2386 	/* Is there space or do we need to sleep? */
2387 	s = GRANT_LOCK(log);
2388 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: enter");
2389 
2390 	/* something is already sleeping; insert new transaction at end */
2391 	if (log->l_reserve_headq) {
2392 		XLOG_INS_TICKETQ(log->l_reserve_headq, tic);
2393 		xlog_trace_loggrant(log, tic,
2394 				    "xlog_grant_log_space: sleep 1");
2395 		/*
2396 		 * Gotta check this before going to sleep, while we're
2397 		 * holding the grant lock.
2398 		 */
2399 		if (XLOG_FORCED_SHUTDOWN(log))
2400 			goto error_return;
2401 
2402 		XFS_STATS_INC(xs_sleep_logspace);
2403 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2404 		/*
2405 		 * If we got an error, and the filesystem is shutting down,
2406 		 * we'll catch it down below. So just continue...
2407 		 */
2408 		xlog_trace_loggrant(log, tic,
2409 				    "xlog_grant_log_space: wake 1");
2410 		s = GRANT_LOCK(log);
2411 	}
2412 	if (tic->t_flags & XFS_LOG_PERM_RESERV)
2413 		need_bytes = tic->t_unit_res*tic->t_ocnt;
2414 	else
2415 		need_bytes = tic->t_unit_res;
2416 
2417 redo:
2418 	if (XLOG_FORCED_SHUTDOWN(log))
2419 		goto error_return;
2420 
2421 	free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2422 				     log->l_grant_reserve_bytes);
2423 	if (free_bytes < need_bytes) {
2424 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2425 			XLOG_INS_TICKETQ(log->l_reserve_headq, tic);
2426 		xlog_trace_loggrant(log, tic,
2427 				    "xlog_grant_log_space: sleep 2");
2428 		XFS_STATS_INC(xs_sleep_logspace);
2429 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2430 
2431 		if (XLOG_FORCED_SHUTDOWN(log)) {
2432 			s = GRANT_LOCK(log);
2433 			goto error_return;
2434 		}
2435 
2436 		xlog_trace_loggrant(log, tic,
2437 				    "xlog_grant_log_space: wake 2");
2438 		xlog_grant_push_ail(log->l_mp, need_bytes);
2439 		s = GRANT_LOCK(log);
2440 		goto redo;
2441 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2442 		XLOG_DEL_TICKETQ(log->l_reserve_headq, tic);
2443 
2444 	/* we've got enough space */
2445 	XLOG_GRANT_ADD_SPACE(log, need_bytes, 'w');
2446 	XLOG_GRANT_ADD_SPACE(log, need_bytes, 'r');
2447 #ifdef DEBUG
2448 	tail_lsn = log->l_tail_lsn;
2449 	/*
2450 	 * Check to make sure the grant write head didn't just over lap the
2451 	 * tail.  If the cycles are the same, we can't be overlapping.
2452 	 * Otherwise, make sure that the cycles differ by exactly one and
2453 	 * check the byte count.
2454 	 */
2455 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2456 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2457 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2458 	}
2459 #endif
2460 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: exit");
2461 	xlog_verify_grant_head(log, 1);
2462 	GRANT_UNLOCK(log, s);
2463 	return 0;
2464 
2465  error_return:
2466 	if (tic->t_flags & XLOG_TIC_IN_Q)
2467 		XLOG_DEL_TICKETQ(log->l_reserve_headq, tic);
2468 	xlog_trace_loggrant(log, tic, "xlog_grant_log_space: err_ret");
2469 	/*
2470 	 * If we are failing, make sure the ticket doesn't have any
2471 	 * current reservations. We don't want to add this back when
2472 	 * the ticket/transaction gets cancelled.
2473 	 */
2474 	tic->t_curr_res = 0;
2475 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2476 	GRANT_UNLOCK(log, s);
2477 	return XFS_ERROR(EIO);
2478 }	/* xlog_grant_log_space */
2479 
2480 
2481 /*
2482  * Replenish the byte reservation required by moving the grant write head.
2483  *
2484  *
2485  */
2486 STATIC int
2487 xlog_regrant_write_log_space(xlog_t	   *log,
2488 			     xlog_ticket_t *tic)
2489 {
2490 	SPLDECL(s);
2491 	int		free_bytes, need_bytes;
2492 	xlog_ticket_t	*ntic;
2493 #ifdef DEBUG
2494 	xfs_lsn_t	tail_lsn;
2495 #endif
2496 
2497 	tic->t_curr_res = tic->t_unit_res;
2498 	XLOG_TIC_RESET_RES(tic);
2499 
2500 	if (tic->t_cnt > 0)
2501 		return (0);
2502 
2503 #ifdef DEBUG
2504 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2505 		panic("regrant Recovery problem");
2506 #endif
2507 
2508 	s = GRANT_LOCK(log);
2509 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: enter");
2510 
2511 	if (XLOG_FORCED_SHUTDOWN(log))
2512 		goto error_return;
2513 
2514 	/* If there are other waiters on the queue then give them a
2515 	 * chance at logspace before us. Wake up the first waiters,
2516 	 * if we do not wake up all the waiters then go to sleep waiting
2517 	 * for more free space, otherwise try to get some space for
2518 	 * this transaction.
2519 	 */
2520 
2521 	if ((ntic = log->l_write_headq)) {
2522 		free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2523 					     log->l_grant_write_bytes);
2524 		do {
2525 			ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2526 
2527 			if (free_bytes < ntic->t_unit_res)
2528 				break;
2529 			free_bytes -= ntic->t_unit_res;
2530 			sv_signal(&ntic->t_sema);
2531 			ntic = ntic->t_next;
2532 		} while (ntic != log->l_write_headq);
2533 
2534 		if (ntic != log->l_write_headq) {
2535 			if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2536 				XLOG_INS_TICKETQ(log->l_write_headq, tic);
2537 
2538 			xlog_trace_loggrant(log, tic,
2539 				    "xlog_regrant_write_log_space: sleep 1");
2540 			XFS_STATS_INC(xs_sleep_logspace);
2541 			sv_wait(&tic->t_sema, PINOD|PLTWAIT,
2542 				&log->l_grant_lock, s);
2543 
2544 			/* If we're shutting down, this tic is already
2545 			 * off the queue */
2546 			if (XLOG_FORCED_SHUTDOWN(log)) {
2547 				s = GRANT_LOCK(log);
2548 				goto error_return;
2549 			}
2550 
2551 			xlog_trace_loggrant(log, tic,
2552 				    "xlog_regrant_write_log_space: wake 1");
2553 			xlog_grant_push_ail(log->l_mp, tic->t_unit_res);
2554 			s = GRANT_LOCK(log);
2555 		}
2556 	}
2557 
2558 	need_bytes = tic->t_unit_res;
2559 
2560 redo:
2561 	if (XLOG_FORCED_SHUTDOWN(log))
2562 		goto error_return;
2563 
2564 	free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2565 				     log->l_grant_write_bytes);
2566 	if (free_bytes < need_bytes) {
2567 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2568 			XLOG_INS_TICKETQ(log->l_write_headq, tic);
2569 		XFS_STATS_INC(xs_sleep_logspace);
2570 		sv_wait(&tic->t_sema, PINOD|PLTWAIT, &log->l_grant_lock, s);
2571 
2572 		/* If we're shutting down, this tic is already off the queue */
2573 		if (XLOG_FORCED_SHUTDOWN(log)) {
2574 			s = GRANT_LOCK(log);
2575 			goto error_return;
2576 		}
2577 
2578 		xlog_trace_loggrant(log, tic,
2579 				    "xlog_regrant_write_log_space: wake 2");
2580 		xlog_grant_push_ail(log->l_mp, need_bytes);
2581 		s = GRANT_LOCK(log);
2582 		goto redo;
2583 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2584 		XLOG_DEL_TICKETQ(log->l_write_headq, tic);
2585 
2586 	XLOG_GRANT_ADD_SPACE(log, need_bytes, 'w'); /* we've got enough space */
2587 #ifdef DEBUG
2588 	tail_lsn = log->l_tail_lsn;
2589 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2590 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2591 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2592 	}
2593 #endif
2594 
2595 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: exit");
2596 	xlog_verify_grant_head(log, 1);
2597 	GRANT_UNLOCK(log, s);
2598 	return (0);
2599 
2600 
2601  error_return:
2602 	if (tic->t_flags & XLOG_TIC_IN_Q)
2603 		XLOG_DEL_TICKETQ(log->l_reserve_headq, tic);
2604 	xlog_trace_loggrant(log, tic, "xlog_regrant_write_log_space: err_ret");
2605 	/*
2606 	 * If we are failing, make sure the ticket doesn't have any
2607 	 * current reservations. We don't want to add this back when
2608 	 * the ticket/transaction gets cancelled.
2609 	 */
2610 	tic->t_curr_res = 0;
2611 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2612 	GRANT_UNLOCK(log, s);
2613 	return XFS_ERROR(EIO);
2614 }	/* xlog_regrant_write_log_space */
2615 
2616 
2617 /* The first cnt-1 times through here we don't need to
2618  * move the grant write head because the permanent
2619  * reservation has reserved cnt times the unit amount.
2620  * Release part of current permanent unit reservation and
2621  * reset current reservation to be one units worth.  Also
2622  * move grant reservation head forward.
2623  */
2624 STATIC void
2625 xlog_regrant_reserve_log_space(xlog_t	     *log,
2626 			       xlog_ticket_t *ticket)
2627 {
2628 	SPLDECL(s);
2629 
2630 	xlog_trace_loggrant(log, ticket,
2631 			    "xlog_regrant_reserve_log_space: enter");
2632 	if (ticket->t_cnt > 0)
2633 		ticket->t_cnt--;
2634 
2635 	s = GRANT_LOCK(log);
2636 	XLOG_GRANT_SUB_SPACE(log, ticket->t_curr_res, 'w');
2637 	XLOG_GRANT_SUB_SPACE(log, ticket->t_curr_res, 'r');
2638 	ticket->t_curr_res = ticket->t_unit_res;
2639 	XLOG_TIC_RESET_RES(ticket);
2640 	xlog_trace_loggrant(log, ticket,
2641 			    "xlog_regrant_reserve_log_space: sub current res");
2642 	xlog_verify_grant_head(log, 1);
2643 
2644 	/* just return if we still have some of the pre-reserved space */
2645 	if (ticket->t_cnt > 0) {
2646 		GRANT_UNLOCK(log, s);
2647 		return;
2648 	}
2649 
2650 	XLOG_GRANT_ADD_SPACE(log, ticket->t_unit_res, 'r');
2651 	xlog_trace_loggrant(log, ticket,
2652 			    "xlog_regrant_reserve_log_space: exit");
2653 	xlog_verify_grant_head(log, 0);
2654 	GRANT_UNLOCK(log, s);
2655 	ticket->t_curr_res = ticket->t_unit_res;
2656 	XLOG_TIC_RESET_RES(ticket);
2657 }	/* xlog_regrant_reserve_log_space */
2658 
2659 
2660 /*
2661  * Give back the space left from a reservation.
2662  *
2663  * All the information we need to make a correct determination of space left
2664  * is present.  For non-permanent reservations, things are quite easy.  The
2665  * count should have been decremented to zero.  We only need to deal with the
2666  * space remaining in the current reservation part of the ticket.  If the
2667  * ticket contains a permanent reservation, there may be left over space which
2668  * needs to be released.  A count of N means that N-1 refills of the current
2669  * reservation can be done before we need to ask for more space.  The first
2670  * one goes to fill up the first current reservation.  Once we run out of
2671  * space, the count will stay at zero and the only space remaining will be
2672  * in the current reservation field.
2673  */
2674 STATIC void
2675 xlog_ungrant_log_space(xlog_t	     *log,
2676 		       xlog_ticket_t *ticket)
2677 {
2678 	SPLDECL(s);
2679 
2680 	if (ticket->t_cnt > 0)
2681 		ticket->t_cnt--;
2682 
2683 	s = GRANT_LOCK(log);
2684 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: enter");
2685 
2686 	XLOG_GRANT_SUB_SPACE(log, ticket->t_curr_res, 'w');
2687 	XLOG_GRANT_SUB_SPACE(log, ticket->t_curr_res, 'r');
2688 
2689 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: sub current");
2690 
2691 	/* If this is a permanent reservation ticket, we may be able to free
2692 	 * up more space based on the remaining count.
2693 	 */
2694 	if (ticket->t_cnt > 0) {
2695 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2696 		XLOG_GRANT_SUB_SPACE(log, ticket->t_unit_res*ticket->t_cnt,'w');
2697 		XLOG_GRANT_SUB_SPACE(log, ticket->t_unit_res*ticket->t_cnt,'r');
2698 	}
2699 
2700 	xlog_trace_loggrant(log, ticket, "xlog_ungrant_log_space: exit");
2701 	xlog_verify_grant_head(log, 1);
2702 	GRANT_UNLOCK(log, s);
2703 	xfs_log_move_tail(log->l_mp, 1);
2704 }	/* xlog_ungrant_log_space */
2705 
2706 
2707 /*
2708  * Atomically put back used ticket.
2709  */
2710 void
2711 xlog_state_put_ticket(xlog_t	    *log,
2712 		      xlog_ticket_t *tic)
2713 {
2714 	unsigned long s;
2715 
2716 	s = LOG_LOCK(log);
2717 	xlog_ticket_put(log, tic);
2718 	LOG_UNLOCK(log, s);
2719 }	/* xlog_state_put_ticket */
2720 
2721 /*
2722  * Flush iclog to disk if this is the last reference to the given iclog and
2723  * the WANT_SYNC bit is set.
2724  *
2725  * When this function is entered, the iclog is not necessarily in the
2726  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2727  *
2728  *
2729  */
2730 int
2731 xlog_state_release_iclog(xlog_t		*log,
2732 			 xlog_in_core_t	*iclog)
2733 {
2734 	SPLDECL(s);
2735 	int		sync = 0;	/* do we sync? */
2736 
2737 	xlog_assign_tail_lsn(log->l_mp);
2738 
2739 	s = LOG_LOCK(log);
2740 
2741 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2742 		LOG_UNLOCK(log, s);
2743 		return XFS_ERROR(EIO);
2744 	}
2745 
2746 	ASSERT(iclog->ic_refcnt > 0);
2747 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2748 	       iclog->ic_state == XLOG_STATE_WANT_SYNC);
2749 
2750 	if (--iclog->ic_refcnt == 0 &&
2751 	    iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2752 		sync++;
2753 		iclog->ic_state = XLOG_STATE_SYNCING;
2754 		INT_SET(iclog->ic_header.h_tail_lsn, ARCH_CONVERT, log->l_tail_lsn);
2755 		xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2756 		/* cycle incremented when incrementing curr_block */
2757 	}
2758 
2759 	LOG_UNLOCK(log, s);
2760 
2761 	/*
2762 	 * We let the log lock go, so it's possible that we hit a log I/O
2763 	 * error or someother SHUTDOWN condition that marks the iclog
2764 	 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2765 	 * this iclog has consistent data, so we ignore IOERROR
2766 	 * flags after this point.
2767 	 */
2768 	if (sync) {
2769 		return xlog_sync(log, iclog);
2770 	}
2771 	return (0);
2772 
2773 }	/* xlog_state_release_iclog */
2774 
2775 
2776 /*
2777  * This routine will mark the current iclog in the ring as WANT_SYNC
2778  * and move the current iclog pointer to the next iclog in the ring.
2779  * When this routine is called from xlog_state_get_iclog_space(), the
2780  * exact size of the iclog has not yet been determined.  All we know is
2781  * that every data block.  We have run out of space in this log record.
2782  */
2783 STATIC void
2784 xlog_state_switch_iclogs(xlog_t		*log,
2785 			 xlog_in_core_t *iclog,
2786 			 int		eventual_size)
2787 {
2788 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2789 	if (!eventual_size)
2790 		eventual_size = iclog->ic_offset;
2791 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2792 	INT_SET(iclog->ic_header.h_prev_block, ARCH_CONVERT, log->l_prev_block);
2793 	log->l_prev_block = log->l_curr_block;
2794 	log->l_prev_cycle = log->l_curr_cycle;
2795 
2796 	/* roll log?: ic_offset changed later */
2797 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2798 
2799 	/* Round up to next log-sunit */
2800 	if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
2801 	    log->l_mp->m_sb.sb_logsunit > 1) {
2802 		__uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2803 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2804 	}
2805 
2806 	if (log->l_curr_block >= log->l_logBBsize) {
2807 		log->l_curr_cycle++;
2808 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2809 			log->l_curr_cycle++;
2810 		log->l_curr_block -= log->l_logBBsize;
2811 		ASSERT(log->l_curr_block >= 0);
2812 	}
2813 	ASSERT(iclog == log->l_iclog);
2814 	log->l_iclog = iclog->ic_next;
2815 }	/* xlog_state_switch_iclogs */
2816 
2817 
2818 /*
2819  * Write out all data in the in-core log as of this exact moment in time.
2820  *
2821  * Data may be written to the in-core log during this call.  However,
2822  * we don't guarantee this data will be written out.  A change from past
2823  * implementation means this routine will *not* write out zero length LRs.
2824  *
2825  * Basically, we try and perform an intelligent scan of the in-core logs.
2826  * If we determine there is no flushable data, we just return.  There is no
2827  * flushable data if:
2828  *
2829  *	1. the current iclog is active and has no data; the previous iclog
2830  *		is in the active or dirty state.
2831  *	2. the current iclog is drity, and the previous iclog is in the
2832  *		active or dirty state.
2833  *
2834  * We may sleep (call psema) if:
2835  *
2836  *	1. the current iclog is not in the active nor dirty state.
2837  *	2. the current iclog dirty, and the previous iclog is not in the
2838  *		active nor dirty state.
2839  *	3. the current iclog is active, and there is another thread writing
2840  *		to this particular iclog.
2841  *	4. a) the current iclog is active and has no other writers
2842  *	   b) when we return from flushing out this iclog, it is still
2843  *		not in the active nor dirty state.
2844  */
2845 STATIC int
2846 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2847 {
2848 	xlog_in_core_t	*iclog;
2849 	xfs_lsn_t	lsn;
2850 	SPLDECL(s);
2851 
2852 	s = LOG_LOCK(log);
2853 
2854 	iclog = log->l_iclog;
2855 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2856 		LOG_UNLOCK(log, s);
2857 		return XFS_ERROR(EIO);
2858 	}
2859 
2860 	/* If the head iclog is not active nor dirty, we just attach
2861 	 * ourselves to the head and go to sleep.
2862 	 */
2863 	if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2864 	    iclog->ic_state == XLOG_STATE_DIRTY) {
2865 		/*
2866 		 * If the head is dirty or (active and empty), then
2867 		 * we need to look at the previous iclog.  If the previous
2868 		 * iclog is active or dirty we are done.  There is nothing
2869 		 * to sync out.  Otherwise, we attach ourselves to the
2870 		 * previous iclog and go to sleep.
2871 		 */
2872 		if (iclog->ic_state == XLOG_STATE_DIRTY ||
2873 		    (iclog->ic_refcnt == 0 && iclog->ic_offset == 0)) {
2874 			iclog = iclog->ic_prev;
2875 			if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2876 			    iclog->ic_state == XLOG_STATE_DIRTY)
2877 				goto no_sleep;
2878 			else
2879 				goto maybe_sleep;
2880 		} else {
2881 			if (iclog->ic_refcnt == 0) {
2882 				/* We are the only one with access to this
2883 				 * iclog.  Flush it out now.  There should
2884 				 * be a roundoff of zero to show that someone
2885 				 * has already taken care of the roundoff from
2886 				 * the previous sync.
2887 				 */
2888 				iclog->ic_refcnt++;
2889 				lsn = INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT);
2890 				xlog_state_switch_iclogs(log, iclog, 0);
2891 				LOG_UNLOCK(log, s);
2892 
2893 				if (xlog_state_release_iclog(log, iclog))
2894 					return XFS_ERROR(EIO);
2895 				*log_flushed = 1;
2896 				s = LOG_LOCK(log);
2897 				if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) == lsn &&
2898 				    iclog->ic_state != XLOG_STATE_DIRTY)
2899 					goto maybe_sleep;
2900 				else
2901 					goto no_sleep;
2902 			} else {
2903 				/* Someone else is writing to this iclog.
2904 				 * Use its call to flush out the data.  However,
2905 				 * the other thread may not force out this LR,
2906 				 * so we mark it WANT_SYNC.
2907 				 */
2908 				xlog_state_switch_iclogs(log, iclog, 0);
2909 				goto maybe_sleep;
2910 			}
2911 		}
2912 	}
2913 
2914 	/* By the time we come around again, the iclog could've been filled
2915 	 * which would give it another lsn.  If we have a new lsn, just
2916 	 * return because the relevant data has been flushed.
2917 	 */
2918 maybe_sleep:
2919 	if (flags & XFS_LOG_SYNC) {
2920 		/*
2921 		 * We must check if we're shutting down here, before
2922 		 * we wait, while we're holding the LOG_LOCK.
2923 		 * Then we check again after waking up, in case our
2924 		 * sleep was disturbed by a bad news.
2925 		 */
2926 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
2927 			LOG_UNLOCK(log, s);
2928 			return XFS_ERROR(EIO);
2929 		}
2930 		XFS_STATS_INC(xs_log_force_sleep);
2931 		sv_wait(&iclog->ic_forcesema, PINOD, &log->l_icloglock, s);
2932 		/*
2933 		 * No need to grab the log lock here since we're
2934 		 * only deciding whether or not to return EIO
2935 		 * and the memory read should be atomic.
2936 		 */
2937 		if (iclog->ic_state & XLOG_STATE_IOERROR)
2938 			return XFS_ERROR(EIO);
2939 		*log_flushed = 1;
2940 
2941 	} else {
2942 
2943 no_sleep:
2944 		LOG_UNLOCK(log, s);
2945 	}
2946 	return 0;
2947 }	/* xlog_state_sync_all */
2948 
2949 
2950 /*
2951  * Used by code which implements synchronous log forces.
2952  *
2953  * Find in-core log with lsn.
2954  *	If it is in the DIRTY state, just return.
2955  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2956  *		state and go to sleep or return.
2957  *	If it is in any other state, go to sleep or return.
2958  *
2959  * If filesystem activity goes to zero, the iclog will get flushed only by
2960  * bdflush().
2961  */
2962 int
2963 xlog_state_sync(xlog_t	  *log,
2964 		xfs_lsn_t lsn,
2965 		uint	  flags,
2966 		int	  *log_flushed)
2967 {
2968     xlog_in_core_t	*iclog;
2969     int			already_slept = 0;
2970     SPLDECL(s);
2971 
2972 
2973 try_again:
2974     s = LOG_LOCK(log);
2975     iclog = log->l_iclog;
2976 
2977     if (iclog->ic_state & XLOG_STATE_IOERROR) {
2978 	    LOG_UNLOCK(log, s);
2979 	    return XFS_ERROR(EIO);
2980     }
2981 
2982     do {
2983 	if (INT_GET(iclog->ic_header.h_lsn, ARCH_CONVERT) != lsn) {
2984 	    iclog = iclog->ic_next;
2985 	    continue;
2986 	}
2987 
2988 	if (iclog->ic_state == XLOG_STATE_DIRTY) {
2989 		LOG_UNLOCK(log, s);
2990 		return 0;
2991 	}
2992 
2993 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
2994 		/*
2995 		 * We sleep here if we haven't already slept (e.g.
2996 		 * this is the first time we've looked at the correct
2997 		 * iclog buf) and the buffer before us is going to
2998 		 * be sync'ed. The reason for this is that if we
2999 		 * are doing sync transactions here, by waiting for
3000 		 * the previous I/O to complete, we can allow a few
3001 		 * more transactions into this iclog before we close
3002 		 * it down.
3003 		 *
3004 		 * Otherwise, we mark the buffer WANT_SYNC, and bump
3005 		 * up the refcnt so we can release the log (which drops
3006 		 * the ref count).  The state switch keeps new transaction
3007 		 * commits from using this buffer.  When the current commits
3008 		 * finish writing into the buffer, the refcount will drop to
3009 		 * zero and the buffer will go out then.
3010 		 */
3011 		if (!already_slept &&
3012 		    (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3013 						 XLOG_STATE_SYNCING))) {
3014 			ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3015 			XFS_STATS_INC(xs_log_force_sleep);
3016 			sv_wait(&iclog->ic_prev->ic_writesema, PSWP,
3017 				&log->l_icloglock, s);
3018 			*log_flushed = 1;
3019 			already_slept = 1;
3020 			goto try_again;
3021 		} else {
3022 			iclog->ic_refcnt++;
3023 			xlog_state_switch_iclogs(log, iclog, 0);
3024 			LOG_UNLOCK(log, s);
3025 			if (xlog_state_release_iclog(log, iclog))
3026 				return XFS_ERROR(EIO);
3027 			*log_flushed = 1;
3028 			s = LOG_LOCK(log);
3029 		}
3030 	}
3031 
3032 	if ((flags & XFS_LOG_SYNC) && /* sleep */
3033 	    !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3034 
3035 		/*
3036 		 * Don't wait on the forcesema if we know that we've
3037 		 * gotten a log write error.
3038 		 */
3039 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
3040 			LOG_UNLOCK(log, s);
3041 			return XFS_ERROR(EIO);
3042 		}
3043 		XFS_STATS_INC(xs_log_force_sleep);
3044 		sv_wait(&iclog->ic_forcesema, PSWP, &log->l_icloglock, s);
3045 		/*
3046 		 * No need to grab the log lock here since we're
3047 		 * only deciding whether or not to return EIO
3048 		 * and the memory read should be atomic.
3049 		 */
3050 		if (iclog->ic_state & XLOG_STATE_IOERROR)
3051 			return XFS_ERROR(EIO);
3052 		*log_flushed = 1;
3053 	} else {		/* just return */
3054 		LOG_UNLOCK(log, s);
3055 	}
3056 	return 0;
3057 
3058     } while (iclog != log->l_iclog);
3059 
3060     LOG_UNLOCK(log, s);
3061     return (0);
3062 }	/* xlog_state_sync */
3063 
3064 
3065 /*
3066  * Called when we want to mark the current iclog as being ready to sync to
3067  * disk.
3068  */
3069 void
3070 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3071 {
3072 	SPLDECL(s);
3073 
3074 	s = LOG_LOCK(log);
3075 
3076 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3077 		xlog_state_switch_iclogs(log, iclog, 0);
3078 	} else {
3079 		ASSERT(iclog->ic_state &
3080 			(XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3081 	}
3082 
3083 	LOG_UNLOCK(log, s);
3084 }	/* xlog_state_want_sync */
3085 
3086 
3087 
3088 /*****************************************************************************
3089  *
3090  *		TICKET functions
3091  *
3092  *****************************************************************************
3093  */
3094 
3095 /*
3096  *	Algorithm doesn't take into account page size. ;-(
3097  */
3098 STATIC void
3099 xlog_state_ticket_alloc(xlog_t *log)
3100 {
3101 	xlog_ticket_t	*t_list;
3102 	xlog_ticket_t	*next;
3103 	xfs_caddr_t	buf;
3104 	uint		i = (NBPP / sizeof(xlog_ticket_t)) - 2;
3105 	SPLDECL(s);
3106 
3107 	/*
3108 	 * The kmem_zalloc may sleep, so we shouldn't be holding the
3109 	 * global lock.  XXXmiken: may want to use zone allocator.
3110 	 */
3111 	buf = (xfs_caddr_t) kmem_zalloc(NBPP, KM_SLEEP);
3112 
3113 	s = LOG_LOCK(log);
3114 
3115 	/* Attach 1st ticket to Q, so we can keep track of allocated memory */
3116 	t_list = (xlog_ticket_t *)buf;
3117 	t_list->t_next = log->l_unmount_free;
3118 	log->l_unmount_free = t_list++;
3119 	log->l_ticket_cnt++;
3120 	log->l_ticket_tcnt++;
3121 
3122 	/* Next ticket becomes first ticket attached to ticket free list */
3123 	if (log->l_freelist != NULL) {
3124 		ASSERT(log->l_tail != NULL);
3125 		log->l_tail->t_next = t_list;
3126 	} else {
3127 		log->l_freelist = t_list;
3128 	}
3129 	log->l_ticket_cnt++;
3130 	log->l_ticket_tcnt++;
3131 
3132 	/* Cycle through rest of alloc'ed memory, building up free Q */
3133 	for ( ; i > 0; i--) {
3134 		next = t_list + 1;
3135 		t_list->t_next = next;
3136 		t_list = next;
3137 		log->l_ticket_cnt++;
3138 		log->l_ticket_tcnt++;
3139 	}
3140 	t_list->t_next = NULL;
3141 	log->l_tail = t_list;
3142 	LOG_UNLOCK(log, s);
3143 }	/* xlog_state_ticket_alloc */
3144 
3145 
3146 /*
3147  * Put ticket into free list
3148  *
3149  * Assumption: log lock is held around this call.
3150  */
3151 STATIC void
3152 xlog_ticket_put(xlog_t		*log,
3153 		xlog_ticket_t	*ticket)
3154 {
3155 	sv_destroy(&ticket->t_sema);
3156 
3157 	/*
3158 	 * Don't think caching will make that much difference.  It's
3159 	 * more important to make debug easier.
3160 	 */
3161 #if 0
3162 	/* real code will want to use LIFO for caching */
3163 	ticket->t_next = log->l_freelist;
3164 	log->l_freelist = ticket;
3165 	/* no need to clear fields */
3166 #else
3167 	/* When we debug, it is easier if tickets are cycled */
3168 	ticket->t_next     = NULL;
3169 	if (log->l_tail != 0) {
3170 		log->l_tail->t_next = ticket;
3171 	} else {
3172 		ASSERT(log->l_freelist == 0);
3173 		log->l_freelist = ticket;
3174 	}
3175 	log->l_tail	    = ticket;
3176 #endif /* DEBUG */
3177 	log->l_ticket_cnt++;
3178 }	/* xlog_ticket_put */
3179 
3180 
3181 /*
3182  * Grab ticket off freelist or allocation some more
3183  */
3184 xlog_ticket_t *
3185 xlog_ticket_get(xlog_t		*log,
3186 		int		unit_bytes,
3187 		int		cnt,
3188 		char		client,
3189 		uint		xflags)
3190 {
3191 	xlog_ticket_t	*tic;
3192 	uint		num_headers;
3193 	SPLDECL(s);
3194 
3195  alloc:
3196 	if (log->l_freelist == NULL)
3197 		xlog_state_ticket_alloc(log);		/* potentially sleep */
3198 
3199 	s = LOG_LOCK(log);
3200 	if (log->l_freelist == NULL) {
3201 		LOG_UNLOCK(log, s);
3202 		goto alloc;
3203 	}
3204 	tic		= log->l_freelist;
3205 	log->l_freelist	= tic->t_next;
3206 	if (log->l_freelist == NULL)
3207 		log->l_tail = NULL;
3208 	log->l_ticket_cnt--;
3209 	LOG_UNLOCK(log, s);
3210 
3211 	/*
3212 	 * Permanent reservations have up to 'cnt'-1 active log operations
3213 	 * in the log.  A unit in this case is the amount of space for one
3214 	 * of these log operations.  Normal reservations have a cnt of 1
3215 	 * and their unit amount is the total amount of space required.
3216 	 *
3217 	 * The following lines of code account for non-transaction data
3218 	 * which occupy space in the on-disk log.
3219 	 *
3220 	 * Normal form of a transaction is:
3221 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3222 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3223 	 *
3224 	 * We need to account for all the leadup data and trailer data
3225 	 * around the transaction data.
3226 	 * And then we need to account for the worst case in terms of using
3227 	 * more space.
3228 	 * The worst case will happen if:
3229 	 * - the placement of the transaction happens to be such that the
3230 	 *   roundoff is at its maximum
3231 	 * - the transaction data is synced before the commit record is synced
3232 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3233 	 *   Therefore the commit record is in its own Log Record.
3234 	 *   This can happen as the commit record is called with its
3235 	 *   own region to xlog_write().
3236 	 *   This then means that in the worst case, roundoff can happen for
3237 	 *   the commit-rec as well.
3238 	 *   The commit-rec is smaller than padding in this scenario and so it is
3239 	 *   not added separately.
3240 	 */
3241 
3242 	/* for trans header */
3243 	unit_bytes += sizeof(xlog_op_header_t);
3244 	unit_bytes += sizeof(xfs_trans_header_t);
3245 
3246 	/* for start-rec */
3247 	unit_bytes += sizeof(xlog_op_header_t);
3248 
3249 	/* for LR headers */
3250 	num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3251 	unit_bytes += log->l_iclog_hsize * num_headers;
3252 
3253 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3254 	unit_bytes += log->l_iclog_hsize;
3255 
3256 	/* for split-recs - ophdrs added when data split over LRs */
3257 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3258 
3259 	/* for roundoff padding for transaction data and one for commit record */
3260 	if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) &&
3261 	    log->l_mp->m_sb.sb_logsunit > 1) {
3262 		/* log su roundoff */
3263 		unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3264 	} else {
3265 		/* BB roundoff */
3266 		unit_bytes += 2*BBSIZE;
3267         }
3268 
3269 	tic->t_unit_res		= unit_bytes;
3270 	tic->t_curr_res		= unit_bytes;
3271 	tic->t_cnt		= cnt;
3272 	tic->t_ocnt		= cnt;
3273 	tic->t_tid		= (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3274 	tic->t_clientid		= client;
3275 	tic->t_flags		= XLOG_TIC_INITED;
3276 	tic->t_trans_type	= 0;
3277 	if (xflags & XFS_LOG_PERM_RESERV)
3278 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3279 	sv_init(&(tic->t_sema), SV_DEFAULT, "logtick");
3280 
3281 	XLOG_TIC_RESET_RES(tic);
3282 
3283 	return tic;
3284 }	/* xlog_ticket_get */
3285 
3286 
3287 /******************************************************************************
3288  *
3289  *		Log debug routines
3290  *
3291  ******************************************************************************
3292  */
3293 #if defined(DEBUG)
3294 /*
3295  * Make sure that the destination ptr is within the valid data region of
3296  * one of the iclogs.  This uses backup pointers stored in a different
3297  * part of the log in case we trash the log structure.
3298  */
3299 void
3300 xlog_verify_dest_ptr(xlog_t     *log,
3301 		     __psint_t  ptr)
3302 {
3303 	int i;
3304 	int good_ptr = 0;
3305 
3306 	for (i=0; i < log->l_iclog_bufs; i++) {
3307 		if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3308 		    ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3309 			good_ptr++;
3310 	}
3311 	if (! good_ptr)
3312 		xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3313 }	/* xlog_verify_dest_ptr */
3314 
3315 STATIC void
3316 xlog_verify_grant_head(xlog_t *log, int equals)
3317 {
3318     if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3319 	if (equals)
3320 	    ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3321 	else
3322 	    ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3323     } else {
3324 	ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3325 	ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3326     }
3327 }	/* xlog_verify_grant_head */
3328 
3329 /* check if it will fit */
3330 STATIC void
3331 xlog_verify_tail_lsn(xlog_t	    *log,
3332 		     xlog_in_core_t *iclog,
3333 		     xfs_lsn_t	    tail_lsn)
3334 {
3335     int blocks;
3336 
3337     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3338 	blocks =
3339 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3340 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3341 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3342     } else {
3343 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3344 
3345 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3346 	    xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3347 
3348 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3349 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3350 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3351     }
3352 }	/* xlog_verify_tail_lsn */
3353 
3354 /*
3355  * Perform a number of checks on the iclog before writing to disk.
3356  *
3357  * 1. Make sure the iclogs are still circular
3358  * 2. Make sure we have a good magic number
3359  * 3. Make sure we don't have magic numbers in the data
3360  * 4. Check fields of each log operation header for:
3361  *	A. Valid client identifier
3362  *	B. tid ptr value falls in valid ptr space (user space code)
3363  *	C. Length in log record header is correct according to the
3364  *		individual operation headers within record.
3365  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3366  *	log, check the preceding blocks of the physical log to make sure all
3367  *	the cycle numbers agree with the current cycle number.
3368  */
3369 STATIC void
3370 xlog_verify_iclog(xlog_t	 *log,
3371 		  xlog_in_core_t *iclog,
3372 		  int		 count,
3373 		  boolean_t	 syncing)
3374 {
3375 	xlog_op_header_t	*ophead;
3376 	xlog_in_core_t		*icptr;
3377 	xlog_in_core_2_t	*xhdr;
3378 	xfs_caddr_t		ptr;
3379 	xfs_caddr_t		base_ptr;
3380 	__psint_t		field_offset;
3381 	__uint8_t		clientid;
3382 	int			len, i, j, k, op_len;
3383 	int			idx;
3384 	SPLDECL(s);
3385 
3386 	/* check validity of iclog pointers */
3387 	s = LOG_LOCK(log);
3388 	icptr = log->l_iclog;
3389 	for (i=0; i < log->l_iclog_bufs; i++) {
3390 		if (icptr == 0)
3391 			xlog_panic("xlog_verify_iclog: invalid ptr");
3392 		icptr = icptr->ic_next;
3393 	}
3394 	if (icptr != log->l_iclog)
3395 		xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3396 	LOG_UNLOCK(log, s);
3397 
3398 	/* check log magic numbers */
3399 	ptr = (xfs_caddr_t) &(iclog->ic_header);
3400 	if (INT_GET(*(uint *)ptr, ARCH_CONVERT) != XLOG_HEADER_MAGIC_NUM)
3401 		xlog_panic("xlog_verify_iclog: invalid magic num");
3402 
3403 	for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&(iclog->ic_header))+count;
3404 	     ptr += BBSIZE) {
3405 		if (INT_GET(*(uint *)ptr, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM)
3406 			xlog_panic("xlog_verify_iclog: unexpected magic num");
3407 	}
3408 
3409 	/* check fields */
3410 	len = INT_GET(iclog->ic_header.h_num_logops, ARCH_CONVERT);
3411 	ptr = iclog->ic_datap;
3412 	base_ptr = ptr;
3413 	ophead = (xlog_op_header_t *)ptr;
3414 	xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3415 	for (i = 0; i < len; i++) {
3416 		ophead = (xlog_op_header_t *)ptr;
3417 
3418 		/* clientid is only 1 byte */
3419 		field_offset = (__psint_t)
3420 			       ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3421 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3422 			clientid = ophead->oh_clientid;
3423 		} else {
3424 			idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3425 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3426 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3427 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3428 				clientid = GET_CLIENT_ID(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3429 			} else {
3430 				clientid = GET_CLIENT_ID(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3431 			}
3432 		}
3433 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3434 			cmn_err(CE_WARN, "xlog_verify_iclog: "
3435 				"invalid clientid %d op 0x%p offset 0x%lx",
3436 				clientid, ophead, (unsigned long)field_offset);
3437 
3438 		/* check length */
3439 		field_offset = (__psint_t)
3440 			       ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3441 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3442 			op_len = INT_GET(ophead->oh_len, ARCH_CONVERT);
3443 		} else {
3444 			idx = BTOBBT((__psint_t)&ophead->oh_len -
3445 				    (__psint_t)iclog->ic_datap);
3446 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3447 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3448 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3449 				op_len = INT_GET(xhdr[j].hic_xheader.xh_cycle_data[k], ARCH_CONVERT);
3450 			} else {
3451 				op_len = INT_GET(iclog->ic_header.h_cycle_data[idx], ARCH_CONVERT);
3452 			}
3453 		}
3454 		ptr += sizeof(xlog_op_header_t) + op_len;
3455 	}
3456 }	/* xlog_verify_iclog */
3457 #endif
3458 
3459 /*
3460  * Mark all iclogs IOERROR. LOG_LOCK is held by the caller.
3461  */
3462 STATIC int
3463 xlog_state_ioerror(
3464 	xlog_t	*log)
3465 {
3466 	xlog_in_core_t	*iclog, *ic;
3467 
3468 	iclog = log->l_iclog;
3469 	if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3470 		/*
3471 		 * Mark all the incore logs IOERROR.
3472 		 * From now on, no log flushes will result.
3473 		 */
3474 		ic = iclog;
3475 		do {
3476 			ic->ic_state = XLOG_STATE_IOERROR;
3477 			ic = ic->ic_next;
3478 		} while (ic != iclog);
3479 		return (0);
3480 	}
3481 	/*
3482 	 * Return non-zero, if state transition has already happened.
3483 	 */
3484 	return (1);
3485 }
3486 
3487 /*
3488  * This is called from xfs_force_shutdown, when we're forcibly
3489  * shutting down the filesystem, typically because of an IO error.
3490  * Our main objectives here are to make sure that:
3491  *	a. the filesystem gets marked 'SHUTDOWN' for all interested
3492  *	   parties to find out, 'atomically'.
3493  *	b. those who're sleeping on log reservations, pinned objects and
3494  *	    other resources get woken up, and be told the bad news.
3495  *	c. nothing new gets queued up after (a) and (b) are done.
3496  *	d. if !logerror, flush the iclogs to disk, then seal them off
3497  *	   for business.
3498  */
3499 int
3500 xfs_log_force_umount(
3501 	struct xfs_mount	*mp,
3502 	int			logerror)
3503 {
3504 	xlog_ticket_t	*tic;
3505 	xlog_t		*log;
3506 	int		retval;
3507 	int		dummy;
3508 	SPLDECL(s);
3509 	SPLDECL(s2);
3510 
3511 	log = mp->m_log;
3512 
3513 	/*
3514 	 * If this happens during log recovery, don't worry about
3515 	 * locking; the log isn't open for business yet.
3516 	 */
3517 	if (!log ||
3518 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3519 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3520 		XFS_BUF_DONE(mp->m_sb_bp);
3521 		return (0);
3522 	}
3523 
3524 	/*
3525 	 * Somebody could've already done the hard work for us.
3526 	 * No need to get locks for this.
3527 	 */
3528 	if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3529 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3530 		return (1);
3531 	}
3532 	retval = 0;
3533 	/*
3534 	 * We must hold both the GRANT lock and the LOG lock,
3535 	 * before we mark the filesystem SHUTDOWN and wake
3536 	 * everybody up to tell the bad news.
3537 	 */
3538 	s = GRANT_LOCK(log);
3539 	s2 = LOG_LOCK(log);
3540 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3541 	XFS_BUF_DONE(mp->m_sb_bp);
3542 	/*
3543 	 * This flag is sort of redundant because of the mount flag, but
3544 	 * it's good to maintain the separation between the log and the rest
3545 	 * of XFS.
3546 	 */
3547 	log->l_flags |= XLOG_IO_ERROR;
3548 
3549 	/*
3550 	 * If we hit a log error, we want to mark all the iclogs IOERROR
3551 	 * while we're still holding the loglock.
3552 	 */
3553 	if (logerror)
3554 		retval = xlog_state_ioerror(log);
3555 	LOG_UNLOCK(log, s2);
3556 
3557 	/*
3558 	 * We don't want anybody waiting for log reservations
3559 	 * after this. That means we have to wake up everybody
3560 	 * queued up on reserve_headq as well as write_headq.
3561 	 * In addition, we make sure in xlog_{re}grant_log_space
3562 	 * that we don't enqueue anything once the SHUTDOWN flag
3563 	 * is set, and this action is protected by the GRANTLOCK.
3564 	 */
3565 	if ((tic = log->l_reserve_headq)) {
3566 		do {
3567 			sv_signal(&tic->t_sema);
3568 			tic = tic->t_next;
3569 		} while (tic != log->l_reserve_headq);
3570 	}
3571 
3572 	if ((tic = log->l_write_headq)) {
3573 		do {
3574 			sv_signal(&tic->t_sema);
3575 			tic = tic->t_next;
3576 		} while (tic != log->l_write_headq);
3577 	}
3578 	GRANT_UNLOCK(log, s);
3579 
3580 	if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3581 		ASSERT(!logerror);
3582 		/*
3583 		 * Force the incore logs to disk before shutting the
3584 		 * log down completely.
3585 		 */
3586 		xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
3587 		s2 = LOG_LOCK(log);
3588 		retval = xlog_state_ioerror(log);
3589 		LOG_UNLOCK(log, s2);
3590 	}
3591 	/*
3592 	 * Wake up everybody waiting on xfs_log_force.
3593 	 * Callback all log item committed functions as if the
3594 	 * log writes were completed.
3595 	 */
3596 	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3597 
3598 #ifdef XFSERRORDEBUG
3599 	{
3600 		xlog_in_core_t	*iclog;
3601 
3602 		s = LOG_LOCK(log);
3603 		iclog = log->l_iclog;
3604 		do {
3605 			ASSERT(iclog->ic_callback == 0);
3606 			iclog = iclog->ic_next;
3607 		} while (iclog != log->l_iclog);
3608 		LOG_UNLOCK(log, s);
3609 	}
3610 #endif
3611 	/* return non-zero if log IOERROR transition had already happened */
3612 	return (retval);
3613 }
3614 
3615 STATIC int
3616 xlog_iclogs_empty(xlog_t *log)
3617 {
3618 	xlog_in_core_t	*iclog;
3619 
3620 	iclog = log->l_iclog;
3621 	do {
3622 		/* endianness does not matter here, zero is zero in
3623 		 * any language.
3624 		 */
3625 		if (iclog->ic_header.h_num_logops)
3626 			return(0);
3627 		iclog = iclog->ic_next;
3628 	} while (iclog != log->l_iclog);
3629 	return(1);
3630 }
3631