xref: /openbmc/linux/fs/xfs/xfs_log.c (revision 9d56dd3b083a3bec56e9da35ce07baca81030b03)
1 /*
2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_log_priv.h"
32 #include "xfs_buf_item.h"
33 #include "xfs_bmap_btree.h"
34 #include "xfs_alloc_btree.h"
35 #include "xfs_ialloc_btree.h"
36 #include "xfs_log_recover.h"
37 #include "xfs_trans_priv.h"
38 #include "xfs_dir2_sf.h"
39 #include "xfs_attr_sf.h"
40 #include "xfs_dinode.h"
41 #include "xfs_inode.h"
42 #include "xfs_rw.h"
43 #include "xfs_trace.h"
44 
45 kmem_zone_t	*xfs_log_ticket_zone;
46 
47 #define xlog_write_adv_cnt(ptr, len, off, bytes) \
48 	{ (ptr) += (bytes); \
49 	  (len) -= (bytes); \
50 	  (off) += (bytes);}
51 
52 /* Local miscellaneous function prototypes */
53 STATIC int	 xlog_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_dealloc_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 int  xlog_state_release_iclog(xlog_t		*log,
79 				     xlog_in_core_t	*iclog);
80 STATIC void xlog_state_switch_iclogs(xlog_t		*log,
81 				     xlog_in_core_t *iclog,
82 				     int		eventual_size);
83 STATIC int  xlog_state_sync(xlog_t			*log,
84 			    xfs_lsn_t 			lsn,
85 			    uint			flags,
86 			    int				*log_flushed);
87 STATIC int  xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed);
88 STATIC void xlog_state_want_sync(xlog_t	*log, xlog_in_core_t *iclog);
89 
90 /* local functions to manipulate grant head */
91 STATIC int  xlog_grant_log_space(xlog_t		*log,
92 				 xlog_ticket_t	*xtic);
93 STATIC void xlog_grant_push_ail(xfs_mount_t	*mp,
94 				int		need_bytes);
95 STATIC void xlog_regrant_reserve_log_space(xlog_t	 *log,
96 					   xlog_ticket_t *ticket);
97 STATIC int xlog_regrant_write_log_space(xlog_t		*log,
98 					 xlog_ticket_t  *ticket);
99 STATIC void xlog_ungrant_log_space(xlog_t	 *log,
100 				   xlog_ticket_t *ticket);
101 
102 
103 /* local ticket functions */
104 STATIC xlog_ticket_t	*xlog_ticket_alloc(xlog_t *log,
105 					 int	unit_bytes,
106 					 int	count,
107 					 char	clientid,
108 					 uint	flags);
109 
110 #if defined(DEBUG)
111 STATIC void	xlog_verify_dest_ptr(xlog_t *log, __psint_t ptr);
112 STATIC void	xlog_verify_grant_head(xlog_t *log, int equals);
113 STATIC void	xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
114 				  int count, boolean_t syncing);
115 STATIC void	xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
116 				     xfs_lsn_t tail_lsn);
117 #else
118 #define xlog_verify_dest_ptr(a,b)
119 #define xlog_verify_grant_head(a,b)
120 #define xlog_verify_iclog(a,b,c,d)
121 #define xlog_verify_tail_lsn(a,b,c)
122 #endif
123 
124 STATIC int	xlog_iclogs_empty(xlog_t *log);
125 
126 
127 static void
128 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
129 {
130 	if (*qp) {
131 		tic->t_next	    = (*qp);
132 		tic->t_prev	    = (*qp)->t_prev;
133 		(*qp)->t_prev->t_next = tic;
134 		(*qp)->t_prev	    = tic;
135 	} else {
136 		tic->t_prev = tic->t_next = tic;
137 		*qp = tic;
138 	}
139 
140 	tic->t_flags |= XLOG_TIC_IN_Q;
141 }
142 
143 static void
144 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
145 {
146 	if (tic == tic->t_next) {
147 		*qp = NULL;
148 	} else {
149 		*qp = tic->t_next;
150 		tic->t_next->t_prev = tic->t_prev;
151 		tic->t_prev->t_next = tic->t_next;
152 	}
153 
154 	tic->t_next = tic->t_prev = NULL;
155 	tic->t_flags &= ~XLOG_TIC_IN_Q;
156 }
157 
158 static void
159 xlog_grant_sub_space(struct log *log, int bytes)
160 {
161 	log->l_grant_write_bytes -= bytes;
162 	if (log->l_grant_write_bytes < 0) {
163 		log->l_grant_write_bytes += log->l_logsize;
164 		log->l_grant_write_cycle--;
165 	}
166 
167 	log->l_grant_reserve_bytes -= bytes;
168 	if ((log)->l_grant_reserve_bytes < 0) {
169 		log->l_grant_reserve_bytes += log->l_logsize;
170 		log->l_grant_reserve_cycle--;
171 	}
172 
173 }
174 
175 static void
176 xlog_grant_add_space_write(struct log *log, int bytes)
177 {
178 	int tmp = log->l_logsize - log->l_grant_write_bytes;
179 	if (tmp > bytes)
180 		log->l_grant_write_bytes += bytes;
181 	else {
182 		log->l_grant_write_cycle++;
183 		log->l_grant_write_bytes = bytes - tmp;
184 	}
185 }
186 
187 static void
188 xlog_grant_add_space_reserve(struct log *log, int bytes)
189 {
190 	int tmp = log->l_logsize - log->l_grant_reserve_bytes;
191 	if (tmp > bytes)
192 		log->l_grant_reserve_bytes += bytes;
193 	else {
194 		log->l_grant_reserve_cycle++;
195 		log->l_grant_reserve_bytes = bytes - tmp;
196 	}
197 }
198 
199 static inline void
200 xlog_grant_add_space(struct log *log, int bytes)
201 {
202 	xlog_grant_add_space_write(log, bytes);
203 	xlog_grant_add_space_reserve(log, bytes);
204 }
205 
206 static void
207 xlog_tic_reset_res(xlog_ticket_t *tic)
208 {
209 	tic->t_res_num = 0;
210 	tic->t_res_arr_sum = 0;
211 	tic->t_res_num_ophdrs = 0;
212 }
213 
214 static void
215 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
216 {
217 	if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
218 		/* add to overflow and start again */
219 		tic->t_res_o_flow += tic->t_res_arr_sum;
220 		tic->t_res_num = 0;
221 		tic->t_res_arr_sum = 0;
222 	}
223 
224 	tic->t_res_arr[tic->t_res_num].r_len = len;
225 	tic->t_res_arr[tic->t_res_num].r_type = type;
226 	tic->t_res_arr_sum += len;
227 	tic->t_res_num++;
228 }
229 
230 /*
231  * NOTES:
232  *
233  *	1. currblock field gets updated at startup and after in-core logs
234  *		marked as with WANT_SYNC.
235  */
236 
237 /*
238  * This routine is called when a user of a log manager ticket is done with
239  * the reservation.  If the ticket was ever used, then a commit record for
240  * the associated transaction is written out as a log operation header with
241  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
242  * a given ticket.  If the ticket was one with a permanent reservation, then
243  * a few operations are done differently.  Permanent reservation tickets by
244  * default don't release the reservation.  They just commit the current
245  * transaction with the belief that the reservation is still needed.  A flag
246  * must be passed in before permanent reservations are actually released.
247  * When these type of tickets are not released, they need to be set into
248  * the inited state again.  By doing this, a start record will be written
249  * out when the next write occurs.
250  */
251 xfs_lsn_t
252 xfs_log_done(xfs_mount_t	*mp,
253 	     xfs_log_ticket_t	xtic,
254 	     void		**iclog,
255 	     uint		flags)
256 {
257 	xlog_t		*log    = mp->m_log;
258 	xlog_ticket_t	*ticket = (xfs_log_ticket_t) xtic;
259 	xfs_lsn_t	lsn	= 0;
260 
261 	if (XLOG_FORCED_SHUTDOWN(log) ||
262 	    /*
263 	     * If nothing was ever written, don't write out commit record.
264 	     * If we get an error, just continue and give back the log ticket.
265 	     */
266 	    (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
267 	     (xlog_commit_record(mp, ticket,
268 				 (xlog_in_core_t **)iclog, &lsn)))) {
269 		lsn = (xfs_lsn_t) -1;
270 		if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
271 			flags |= XFS_LOG_REL_PERM_RESERV;
272 		}
273 	}
274 
275 
276 	if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
277 	    (flags & XFS_LOG_REL_PERM_RESERV)) {
278 		trace_xfs_log_done_nonperm(log, ticket);
279 
280 		/*
281 		 * Release ticket if not permanent reservation or a specific
282 		 * request has been made to release a permanent reservation.
283 		 */
284 		xlog_ungrant_log_space(log, ticket);
285 		xfs_log_ticket_put(ticket);
286 	} else {
287 		trace_xfs_log_done_perm(log, ticket);
288 
289 		xlog_regrant_reserve_log_space(log, ticket);
290 		/* If this ticket was a permanent reservation and we aren't
291 		 * trying to release it, reset the inited flags; so next time
292 		 * we write, a start record will be written out.
293 		 */
294 		ticket->t_flags |= XLOG_TIC_INITED;
295 	}
296 
297 	return lsn;
298 }	/* xfs_log_done */
299 
300 
301 /*
302  * Force the in-core log to disk.  If flags == XFS_LOG_SYNC,
303  *	the force is done synchronously.
304  *
305  * Asynchronous forces are implemented by setting the WANT_SYNC
306  * bit in the appropriate in-core log and then returning.
307  *
308  * Synchronous forces are implemented with a signal variable. All callers
309  * to force a given lsn to disk will wait on a the sv attached to the
310  * specific in-core log.  When given in-core log finally completes its
311  * write to disk, that thread will wake up all threads waiting on the
312  * sv.
313  */
314 int
315 _xfs_log_force(
316 	xfs_mount_t	*mp,
317 	xfs_lsn_t	lsn,
318 	uint		flags,
319 	int		*log_flushed)
320 {
321 	xlog_t		*log = mp->m_log;
322 	int		dummy;
323 
324 	if (!log_flushed)
325 		log_flushed = &dummy;
326 
327 	ASSERT(flags & XFS_LOG_FORCE);
328 
329 	XFS_STATS_INC(xs_log_force);
330 
331 	if (log->l_flags & XLOG_IO_ERROR)
332 		return XFS_ERROR(EIO);
333 	if (lsn == 0)
334 		return xlog_state_sync_all(log, flags, log_flushed);
335 	else
336 		return xlog_state_sync(log, lsn, flags, log_flushed);
337 }	/* _xfs_log_force */
338 
339 /*
340  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
341  * about errors or whether the log was flushed or not. This is the normal
342  * interface to use when trying to unpin items or move the log forward.
343  */
344 void
345 xfs_log_force(
346 	xfs_mount_t	*mp,
347 	xfs_lsn_t	lsn,
348 	uint		flags)
349 {
350 	int	error;
351 	error = _xfs_log_force(mp, lsn, flags, NULL);
352 	if (error) {
353 		xfs_fs_cmn_err(CE_WARN, mp, "xfs_log_force: "
354 			"error %d returned.", error);
355 	}
356 }
357 
358 
359 /*
360  * Attaches a new iclog I/O completion callback routine during
361  * transaction commit.  If the log is in error state, a non-zero
362  * return code is handed back and the caller is responsible for
363  * executing the callback at an appropriate time.
364  */
365 int
366 xfs_log_notify(xfs_mount_t	  *mp,		/* mount of partition */
367 	       void		  *iclog_hndl,	/* iclog to hang callback off */
368 	       xfs_log_callback_t *cb)
369 {
370 	xlog_in_core_t	  *iclog = (xlog_in_core_t *)iclog_hndl;
371 	int	abortflg;
372 
373 	spin_lock(&iclog->ic_callback_lock);
374 	abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
375 	if (!abortflg) {
376 		ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
377 			      (iclog->ic_state == XLOG_STATE_WANT_SYNC));
378 		cb->cb_next = NULL;
379 		*(iclog->ic_callback_tail) = cb;
380 		iclog->ic_callback_tail = &(cb->cb_next);
381 	}
382 	spin_unlock(&iclog->ic_callback_lock);
383 	return abortflg;
384 }	/* xfs_log_notify */
385 
386 int
387 xfs_log_release_iclog(xfs_mount_t *mp,
388 		      void	  *iclog_hndl)
389 {
390 	xlog_t *log = mp->m_log;
391 	xlog_in_core_t	  *iclog = (xlog_in_core_t *)iclog_hndl;
392 
393 	if (xlog_state_release_iclog(log, iclog)) {
394 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
395 		return EIO;
396 	}
397 
398 	return 0;
399 }
400 
401 /*
402  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
403  *	to the reservation.
404  *  2. Potentially, push buffers at tail of log to disk.
405  *
406  * Each reservation is going to reserve extra space for a log record header.
407  * When writes happen to the on-disk log, we don't subtract the length of the
408  * log record header from any reservation.  By wasting space in each
409  * reservation, we prevent over allocation problems.
410  */
411 int
412 xfs_log_reserve(xfs_mount_t	 *mp,
413 		int		 unit_bytes,
414 		int		 cnt,
415 		xfs_log_ticket_t *ticket,
416 		__uint8_t	 client,
417 		uint		 flags,
418 		uint		 t_type)
419 {
420 	xlog_t		*log = mp->m_log;
421 	xlog_ticket_t	*internal_ticket;
422 	int		retval = 0;
423 
424 	ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
425 	ASSERT((flags & XFS_LOG_NOSLEEP) == 0);
426 
427 	if (XLOG_FORCED_SHUTDOWN(log))
428 		return XFS_ERROR(EIO);
429 
430 	XFS_STATS_INC(xs_try_logspace);
431 
432 
433 	if (*ticket != NULL) {
434 		ASSERT(flags & XFS_LOG_PERM_RESERV);
435 		internal_ticket = (xlog_ticket_t *)*ticket;
436 
437 		trace_xfs_log_reserve(log, internal_ticket);
438 
439 		xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
440 		retval = xlog_regrant_write_log_space(log, internal_ticket);
441 	} else {
442 		/* may sleep if need to allocate more tickets */
443 		internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
444 						  client, flags);
445 		if (!internal_ticket)
446 			return XFS_ERROR(ENOMEM);
447 		internal_ticket->t_trans_type = t_type;
448 		*ticket = internal_ticket;
449 
450 		trace_xfs_log_reserve(log, internal_ticket);
451 
452 		xlog_grant_push_ail(mp,
453 				    (internal_ticket->t_unit_res *
454 				     internal_ticket->t_cnt));
455 		retval = xlog_grant_log_space(log, internal_ticket);
456 	}
457 
458 	return retval;
459 }	/* xfs_log_reserve */
460 
461 
462 /*
463  * Mount a log filesystem
464  *
465  * mp		- ubiquitous xfs mount point structure
466  * log_target	- buftarg of on-disk log device
467  * blk_offset	- Start block # where block size is 512 bytes (BBSIZE)
468  * num_bblocks	- Number of BBSIZE blocks in on-disk log
469  *
470  * Return error or zero.
471  */
472 int
473 xfs_log_mount(
474 	xfs_mount_t	*mp,
475 	xfs_buftarg_t	*log_target,
476 	xfs_daddr_t	blk_offset,
477 	int		num_bblks)
478 {
479 	int		error;
480 
481 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
482 		cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
483 	else {
484 		cmn_err(CE_NOTE,
485 			"!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
486 			mp->m_fsname);
487 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
488 	}
489 
490 	mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
491 	if (IS_ERR(mp->m_log)) {
492 		error = -PTR_ERR(mp->m_log);
493 		goto out;
494 	}
495 
496 	/*
497 	 * Initialize the AIL now we have a log.
498 	 */
499 	error = xfs_trans_ail_init(mp);
500 	if (error) {
501 		cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
502 		goto out_free_log;
503 	}
504 	mp->m_log->l_ailp = mp->m_ail;
505 
506 	/*
507 	 * skip log recovery on a norecovery mount.  pretend it all
508 	 * just worked.
509 	 */
510 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
511 		int	readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
512 
513 		if (readonly)
514 			mp->m_flags &= ~XFS_MOUNT_RDONLY;
515 
516 		error = xlog_recover(mp->m_log);
517 
518 		if (readonly)
519 			mp->m_flags |= XFS_MOUNT_RDONLY;
520 		if (error) {
521 			cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
522 			goto out_destroy_ail;
523 		}
524 	}
525 
526 	/* Normal transactions can now occur */
527 	mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
528 
529 	return 0;
530 
531 out_destroy_ail:
532 	xfs_trans_ail_destroy(mp);
533 out_free_log:
534 	xlog_dealloc_log(mp->m_log);
535 out:
536 	return error;
537 }
538 
539 /*
540  * Finish the recovery of the file system.  This is separate from
541  * the xfs_log_mount() call, because it depends on the code in
542  * xfs_mountfs() to read in the root and real-time bitmap inodes
543  * between calling xfs_log_mount() and here.
544  *
545  * mp		- ubiquitous xfs mount point structure
546  */
547 int
548 xfs_log_mount_finish(xfs_mount_t *mp)
549 {
550 	int	error;
551 
552 	if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
553 		error = xlog_recover_finish(mp->m_log);
554 	else {
555 		error = 0;
556 		ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
557 	}
558 
559 	return error;
560 }
561 
562 /*
563  * Final log writes as part of unmount.
564  *
565  * Mark the filesystem clean as unmount happens.  Note that during relocation
566  * this routine needs to be executed as part of source-bag while the
567  * deallocation must not be done until source-end.
568  */
569 
570 /*
571  * Unmount record used to have a string "Unmount filesystem--" in the
572  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
573  * We just write the magic number now since that particular field isn't
574  * currently architecture converted and "nUmount" is a bit foo.
575  * As far as I know, there weren't any dependencies on the old behaviour.
576  */
577 
578 int
579 xfs_log_unmount_write(xfs_mount_t *mp)
580 {
581 	xlog_t		 *log = mp->m_log;
582 	xlog_in_core_t	 *iclog;
583 #ifdef DEBUG
584 	xlog_in_core_t	 *first_iclog;
585 #endif
586 	xfs_log_iovec_t  reg[1];
587 	xfs_log_ticket_t tic = NULL;
588 	xfs_lsn_t	 lsn;
589 	int		 error;
590 
591 	/* the data section must be 32 bit size aligned */
592 	struct {
593 	    __uint16_t magic;
594 	    __uint16_t pad1;
595 	    __uint32_t pad2; /* may as well make it 64 bits */
596 	} magic = { XLOG_UNMOUNT_TYPE, 0, 0 };
597 
598 	/*
599 	 * Don't write out unmount record on read-only mounts.
600 	 * Or, if we are doing a forced umount (typically because of IO errors).
601 	 */
602 	if (mp->m_flags & XFS_MOUNT_RDONLY)
603 		return 0;
604 
605 	error = _xfs_log_force(mp, 0, XFS_LOG_FORCE|XFS_LOG_SYNC, NULL);
606 	ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
607 
608 #ifdef DEBUG
609 	first_iclog = iclog = log->l_iclog;
610 	do {
611 		if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
612 			ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
613 			ASSERT(iclog->ic_offset == 0);
614 		}
615 		iclog = iclog->ic_next;
616 	} while (iclog != first_iclog);
617 #endif
618 	if (! (XLOG_FORCED_SHUTDOWN(log))) {
619 		reg[0].i_addr = (void*)&magic;
620 		reg[0].i_len  = sizeof(magic);
621 		XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_UNMOUNT);
622 
623 		error = xfs_log_reserve(mp, 600, 1, &tic,
624 					XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
625 		if (!error) {
626 			/* remove inited flag */
627 			((xlog_ticket_t *)tic)->t_flags = 0;
628 			error = xlog_write(mp, reg, 1, tic, &lsn,
629 					   NULL, XLOG_UNMOUNT_TRANS);
630 			/*
631 			 * At this point, we're umounting anyway,
632 			 * so there's no point in transitioning log state
633 			 * to IOERROR. Just continue...
634 			 */
635 		}
636 
637 		if (error) {
638 			xfs_fs_cmn_err(CE_ALERT, mp,
639 				"xfs_log_unmount: unmount record failed");
640 		}
641 
642 
643 		spin_lock(&log->l_icloglock);
644 		iclog = log->l_iclog;
645 		atomic_inc(&iclog->ic_refcnt);
646 		xlog_state_want_sync(log, iclog);
647 		spin_unlock(&log->l_icloglock);
648 		error = xlog_state_release_iclog(log, iclog);
649 
650 		spin_lock(&log->l_icloglock);
651 		if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
652 		      iclog->ic_state == XLOG_STATE_DIRTY)) {
653 			if (!XLOG_FORCED_SHUTDOWN(log)) {
654 				sv_wait(&iclog->ic_force_wait, PMEM,
655 					&log->l_icloglock, s);
656 			} else {
657 				spin_unlock(&log->l_icloglock);
658 			}
659 		} else {
660 			spin_unlock(&log->l_icloglock);
661 		}
662 		if (tic) {
663 			trace_xfs_log_umount_write(log, tic);
664 			xlog_ungrant_log_space(log, tic);
665 			xfs_log_ticket_put(tic);
666 		}
667 	} else {
668 		/*
669 		 * We're already in forced_shutdown mode, couldn't
670 		 * even attempt to write out the unmount transaction.
671 		 *
672 		 * Go through the motions of sync'ing and releasing
673 		 * the iclog, even though no I/O will actually happen,
674 		 * we need to wait for other log I/Os that may already
675 		 * be in progress.  Do this as a separate section of
676 		 * code so we'll know if we ever get stuck here that
677 		 * we're in this odd situation of trying to unmount
678 		 * a file system that went into forced_shutdown as
679 		 * the result of an unmount..
680 		 */
681 		spin_lock(&log->l_icloglock);
682 		iclog = log->l_iclog;
683 		atomic_inc(&iclog->ic_refcnt);
684 
685 		xlog_state_want_sync(log, iclog);
686 		spin_unlock(&log->l_icloglock);
687 		error =  xlog_state_release_iclog(log, iclog);
688 
689 		spin_lock(&log->l_icloglock);
690 
691 		if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
692 			|| iclog->ic_state == XLOG_STATE_DIRTY
693 			|| iclog->ic_state == XLOG_STATE_IOERROR) ) {
694 
695 				sv_wait(&iclog->ic_force_wait, PMEM,
696 					&log->l_icloglock, s);
697 		} else {
698 			spin_unlock(&log->l_icloglock);
699 		}
700 	}
701 
702 	return error;
703 }	/* xfs_log_unmount_write */
704 
705 /*
706  * Deallocate log structures for unmount/relocation.
707  *
708  * We need to stop the aild from running before we destroy
709  * and deallocate the log as the aild references the log.
710  */
711 void
712 xfs_log_unmount(xfs_mount_t *mp)
713 {
714 	xfs_trans_ail_destroy(mp);
715 	xlog_dealloc_log(mp->m_log);
716 }
717 
718 /*
719  * Write region vectors to log.  The write happens using the space reservation
720  * of the ticket (tic).  It is not a requirement that all writes for a given
721  * transaction occur with one call to xfs_log_write().
722  */
723 int
724 xfs_log_write(xfs_mount_t *	mp,
725 	      xfs_log_iovec_t	reg[],
726 	      int		nentries,
727 	      xfs_log_ticket_t	tic,
728 	      xfs_lsn_t		*start_lsn)
729 {
730 	int	error;
731 	xlog_t *log = mp->m_log;
732 
733 	if (XLOG_FORCED_SHUTDOWN(log))
734 		return XFS_ERROR(EIO);
735 
736 	if ((error = xlog_write(mp, reg, nentries, tic, start_lsn, NULL, 0))) {
737 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
738 	}
739 	return error;
740 }	/* xfs_log_write */
741 
742 
743 void
744 xfs_log_move_tail(xfs_mount_t	*mp,
745 		  xfs_lsn_t	tail_lsn)
746 {
747 	xlog_ticket_t	*tic;
748 	xlog_t		*log = mp->m_log;
749 	int		need_bytes, free_bytes, cycle, bytes;
750 
751 	if (XLOG_FORCED_SHUTDOWN(log))
752 		return;
753 
754 	if (tail_lsn == 0) {
755 		/* needed since sync_lsn is 64 bits */
756 		spin_lock(&log->l_icloglock);
757 		tail_lsn = log->l_last_sync_lsn;
758 		spin_unlock(&log->l_icloglock);
759 	}
760 
761 	spin_lock(&log->l_grant_lock);
762 
763 	/* Also an invalid lsn.  1 implies that we aren't passing in a valid
764 	 * tail_lsn.
765 	 */
766 	if (tail_lsn != 1) {
767 		log->l_tail_lsn = tail_lsn;
768 	}
769 
770 	if ((tic = log->l_write_headq)) {
771 #ifdef DEBUG
772 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
773 			panic("Recovery problem");
774 #endif
775 		cycle = log->l_grant_write_cycle;
776 		bytes = log->l_grant_write_bytes;
777 		free_bytes = xlog_space_left(log, cycle, bytes);
778 		do {
779 			ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
780 
781 			if (free_bytes < tic->t_unit_res && tail_lsn != 1)
782 				break;
783 			tail_lsn = 0;
784 			free_bytes -= tic->t_unit_res;
785 			sv_signal(&tic->t_wait);
786 			tic = tic->t_next;
787 		} while (tic != log->l_write_headq);
788 	}
789 	if ((tic = log->l_reserve_headq)) {
790 #ifdef DEBUG
791 		if (log->l_flags & XLOG_ACTIVE_RECOVERY)
792 			panic("Recovery problem");
793 #endif
794 		cycle = log->l_grant_reserve_cycle;
795 		bytes = log->l_grant_reserve_bytes;
796 		free_bytes = xlog_space_left(log, cycle, bytes);
797 		do {
798 			if (tic->t_flags & XLOG_TIC_PERM_RESERV)
799 				need_bytes = tic->t_unit_res*tic->t_cnt;
800 			else
801 				need_bytes = tic->t_unit_res;
802 			if (free_bytes < need_bytes && tail_lsn != 1)
803 				break;
804 			tail_lsn = 0;
805 			free_bytes -= need_bytes;
806 			sv_signal(&tic->t_wait);
807 			tic = tic->t_next;
808 		} while (tic != log->l_reserve_headq);
809 	}
810 	spin_unlock(&log->l_grant_lock);
811 }	/* xfs_log_move_tail */
812 
813 /*
814  * Determine if we have a transaction that has gone to disk
815  * that needs to be covered. Log activity needs to be idle (no AIL and
816  * nothing in the iclogs). And, we need to be in the right state indicating
817  * something has gone out.
818  */
819 int
820 xfs_log_need_covered(xfs_mount_t *mp)
821 {
822 	int		needed = 0;
823 	xlog_t		*log = mp->m_log;
824 
825 	if (!xfs_fs_writable(mp))
826 		return 0;
827 
828 	spin_lock(&log->l_icloglock);
829 	if (((log->l_covered_state == XLOG_STATE_COVER_NEED) ||
830 		(log->l_covered_state == XLOG_STATE_COVER_NEED2))
831 			&& !xfs_trans_ail_tail(log->l_ailp)
832 			&& xlog_iclogs_empty(log)) {
833 		if (log->l_covered_state == XLOG_STATE_COVER_NEED)
834 			log->l_covered_state = XLOG_STATE_COVER_DONE;
835 		else {
836 			ASSERT(log->l_covered_state == XLOG_STATE_COVER_NEED2);
837 			log->l_covered_state = XLOG_STATE_COVER_DONE2;
838 		}
839 		needed = 1;
840 	}
841 	spin_unlock(&log->l_icloglock);
842 	return needed;
843 }
844 
845 /******************************************************************************
846  *
847  *	local routines
848  *
849  ******************************************************************************
850  */
851 
852 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
853  * The log manager must keep track of the last LR which was committed
854  * to disk.  The lsn of this LR will become the new tail_lsn whenever
855  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
856  * the situation where stuff could be written into the log but nothing
857  * was ever in the AIL when asked.  Eventually, we panic since the
858  * tail hits the head.
859  *
860  * We may be holding the log iclog lock upon entering this routine.
861  */
862 xfs_lsn_t
863 xlog_assign_tail_lsn(xfs_mount_t *mp)
864 {
865 	xfs_lsn_t tail_lsn;
866 	xlog_t	  *log = mp->m_log;
867 
868 	tail_lsn = xfs_trans_ail_tail(mp->m_ail);
869 	spin_lock(&log->l_grant_lock);
870 	if (tail_lsn != 0) {
871 		log->l_tail_lsn = tail_lsn;
872 	} else {
873 		tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
874 	}
875 	spin_unlock(&log->l_grant_lock);
876 
877 	return tail_lsn;
878 }	/* xlog_assign_tail_lsn */
879 
880 
881 /*
882  * Return the space in the log between the tail and the head.  The head
883  * is passed in the cycle/bytes formal parms.  In the special case where
884  * the reserve head has wrapped passed the tail, this calculation is no
885  * longer valid.  In this case, just return 0 which means there is no space
886  * in the log.  This works for all places where this function is called
887  * with the reserve head.  Of course, if the write head were to ever
888  * wrap the tail, we should blow up.  Rather than catch this case here,
889  * we depend on other ASSERTions in other parts of the code.   XXXmiken
890  *
891  * This code also handles the case where the reservation head is behind
892  * the tail.  The details of this case are described below, but the end
893  * result is that we return the size of the log as the amount of space left.
894  */
895 STATIC int
896 xlog_space_left(xlog_t *log, int cycle, int bytes)
897 {
898 	int free_bytes;
899 	int tail_bytes;
900 	int tail_cycle;
901 
902 	tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
903 	tail_cycle = CYCLE_LSN(log->l_tail_lsn);
904 	if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
905 		free_bytes = log->l_logsize - (bytes - tail_bytes);
906 	} else if ((tail_cycle + 1) < cycle) {
907 		return 0;
908 	} else if (tail_cycle < cycle) {
909 		ASSERT(tail_cycle == (cycle - 1));
910 		free_bytes = tail_bytes - bytes;
911 	} else {
912 		/*
913 		 * The reservation head is behind the tail.
914 		 * In this case we just want to return the size of the
915 		 * log as the amount of space left.
916 		 */
917 		xfs_fs_cmn_err(CE_ALERT, log->l_mp,
918 			"xlog_space_left: head behind tail\n"
919 			"  tail_cycle = %d, tail_bytes = %d\n"
920 			"  GH   cycle = %d, GH   bytes = %d",
921 			tail_cycle, tail_bytes, cycle, bytes);
922 		ASSERT(0);
923 		free_bytes = log->l_logsize;
924 	}
925 	return free_bytes;
926 }	/* xlog_space_left */
927 
928 
929 /*
930  * Log function which is called when an io completes.
931  *
932  * The log manager needs its own routine, in order to control what
933  * happens with the buffer after the write completes.
934  */
935 void
936 xlog_iodone(xfs_buf_t *bp)
937 {
938 	xlog_in_core_t	*iclog;
939 	xlog_t		*l;
940 	int		aborted;
941 
942 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
943 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
944 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
945 	aborted = 0;
946 	l = iclog->ic_log;
947 
948 	/*
949 	 * If the _XFS_BARRIER_FAILED flag was set by a lower
950 	 * layer, it means the underlying device no longer supports
951 	 * barrier I/O. Warn loudly and turn off barriers.
952 	 */
953 	if (bp->b_flags & _XFS_BARRIER_FAILED) {
954 		bp->b_flags &= ~_XFS_BARRIER_FAILED;
955 		l->l_mp->m_flags &= ~XFS_MOUNT_BARRIER;
956 		xfs_fs_cmn_err(CE_WARN, l->l_mp,
957 				"xlog_iodone: Barriers are no longer supported"
958 				" by device. Disabling barriers\n");
959 	}
960 
961 	/*
962 	 * Race to shutdown the filesystem if we see an error.
963 	 */
964 	if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
965 			XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
966 		xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
967 		XFS_BUF_STALE(bp);
968 		xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
969 		/*
970 		 * This flag will be propagated to the trans-committed
971 		 * callback routines to let them know that the log-commit
972 		 * didn't succeed.
973 		 */
974 		aborted = XFS_LI_ABORTED;
975 	} else if (iclog->ic_state & XLOG_STATE_IOERROR) {
976 		aborted = XFS_LI_ABORTED;
977 	}
978 
979 	/* log I/O is always issued ASYNC */
980 	ASSERT(XFS_BUF_ISASYNC(bp));
981 	xlog_state_done_syncing(iclog, aborted);
982 	/*
983 	 * do not reference the buffer (bp) here as we could race
984 	 * with it being freed after writing the unmount record to the
985 	 * log.
986 	 */
987 
988 }	/* xlog_iodone */
989 
990 /*
991  * The bdstrat callback function for log bufs. This gives us a central
992  * place to trap bufs in case we get hit by a log I/O error and need to
993  * shutdown. Actually, in practice, even when we didn't get a log error,
994  * we transition the iclogs to IOERROR state *after* flushing all existing
995  * iclogs to disk. This is because we don't want anymore new transactions to be
996  * started or completed afterwards.
997  */
998 STATIC int
999 xlog_bdstrat_cb(struct xfs_buf *bp)
1000 {
1001 	xlog_in_core_t *iclog;
1002 
1003 	iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1004 
1005 	if ((iclog->ic_state & XLOG_STATE_IOERROR) == 0) {
1006 	  /* note for irix bstrat will need  struct bdevsw passed
1007 	   * Fix the following macro if the code ever is merged
1008 	   */
1009 	    XFS_bdstrat(bp);
1010 		return 0;
1011 	}
1012 
1013 	XFS_BUF_ERROR(bp, EIO);
1014 	XFS_BUF_STALE(bp);
1015 	xfs_biodone(bp);
1016 	return XFS_ERROR(EIO);
1017 }
1018 
1019 /*
1020  * Return size of each in-core log record buffer.
1021  *
1022  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1023  *
1024  * If the filesystem blocksize is too large, we may need to choose a
1025  * larger size since the directory code currently logs entire blocks.
1026  */
1027 
1028 STATIC void
1029 xlog_get_iclog_buffer_size(xfs_mount_t	*mp,
1030 			   xlog_t	*log)
1031 {
1032 	int size;
1033 	int xhdrs;
1034 
1035 	if (mp->m_logbufs <= 0)
1036 		log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1037 	else
1038 		log->l_iclog_bufs = mp->m_logbufs;
1039 
1040 	/*
1041 	 * Buffer size passed in from mount system call.
1042 	 */
1043 	if (mp->m_logbsize > 0) {
1044 		size = log->l_iclog_size = mp->m_logbsize;
1045 		log->l_iclog_size_log = 0;
1046 		while (size != 1) {
1047 			log->l_iclog_size_log++;
1048 			size >>= 1;
1049 		}
1050 
1051 		if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1052 			/* # headers = size / 32k
1053 			 * one header holds cycles from 32k of data
1054 			 */
1055 
1056 			xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1057 			if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1058 				xhdrs++;
1059 			log->l_iclog_hsize = xhdrs << BBSHIFT;
1060 			log->l_iclog_heads = xhdrs;
1061 		} else {
1062 			ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1063 			log->l_iclog_hsize = BBSIZE;
1064 			log->l_iclog_heads = 1;
1065 		}
1066 		goto done;
1067 	}
1068 
1069 	/* All machines use 32kB buffers by default. */
1070 	log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1071 	log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1072 
1073 	/* the default log size is 16k or 32k which is one header sector */
1074 	log->l_iclog_hsize = BBSIZE;
1075 	log->l_iclog_heads = 1;
1076 
1077 done:
1078 	/* are we being asked to make the sizes selected above visible? */
1079 	if (mp->m_logbufs == 0)
1080 		mp->m_logbufs = log->l_iclog_bufs;
1081 	if (mp->m_logbsize == 0)
1082 		mp->m_logbsize = log->l_iclog_size;
1083 }	/* xlog_get_iclog_buffer_size */
1084 
1085 
1086 /*
1087  * This routine initializes some of the log structure for a given mount point.
1088  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1089  * some other stuff may be filled in too.
1090  */
1091 STATIC xlog_t *
1092 xlog_alloc_log(xfs_mount_t	*mp,
1093 	       xfs_buftarg_t	*log_target,
1094 	       xfs_daddr_t	blk_offset,
1095 	       int		num_bblks)
1096 {
1097 	xlog_t			*log;
1098 	xlog_rec_header_t	*head;
1099 	xlog_in_core_t		**iclogp;
1100 	xlog_in_core_t		*iclog, *prev_iclog=NULL;
1101 	xfs_buf_t		*bp;
1102 	int			i;
1103 	int			iclogsize;
1104 	int			error = ENOMEM;
1105 
1106 	log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1107 	if (!log) {
1108 		xlog_warn("XFS: Log allocation failed: No memory!");
1109 		goto out;
1110 	}
1111 
1112 	log->l_mp	   = mp;
1113 	log->l_targ	   = log_target;
1114 	log->l_logsize     = BBTOB(num_bblks);
1115 	log->l_logBBstart  = blk_offset;
1116 	log->l_logBBsize   = num_bblks;
1117 	log->l_covered_state = XLOG_STATE_COVER_IDLE;
1118 	log->l_flags	   |= XLOG_ACTIVE_RECOVERY;
1119 
1120 	log->l_prev_block  = -1;
1121 	log->l_tail_lsn	   = xlog_assign_lsn(1, 0);
1122 	/* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1123 	log->l_last_sync_lsn = log->l_tail_lsn;
1124 	log->l_curr_cycle  = 1;	    /* 0 is bad since this is initial value */
1125 	log->l_grant_reserve_cycle = 1;
1126 	log->l_grant_write_cycle = 1;
1127 
1128 	error = EFSCORRUPTED;
1129 	if (xfs_sb_version_hassector(&mp->m_sb)) {
1130 		log->l_sectbb_log = mp->m_sb.sb_logsectlog - BBSHIFT;
1131 		if (log->l_sectbb_log < 0 ||
1132 		    log->l_sectbb_log > mp->m_sectbb_log) {
1133 			xlog_warn("XFS: Log sector size (0x%x) out of range.",
1134 						log->l_sectbb_log);
1135 			goto out_free_log;
1136 		}
1137 
1138 		/* for larger sector sizes, must have v2 or external log */
1139 		if (log->l_sectbb_log != 0 &&
1140 		    (log->l_logBBstart != 0 &&
1141 		     !xfs_sb_version_haslogv2(&mp->m_sb))) {
1142 			xlog_warn("XFS: log sector size (0x%x) invalid "
1143 				  "for configuration.", log->l_sectbb_log);
1144 			goto out_free_log;
1145 		}
1146 		if (mp->m_sb.sb_logsectlog < BBSHIFT) {
1147 			xlog_warn("XFS: Log sector log (0x%x) too small.",
1148 						mp->m_sb.sb_logsectlog);
1149 			goto out_free_log;
1150 		}
1151 	}
1152 	log->l_sectbb_mask = (1 << log->l_sectbb_log) - 1;
1153 
1154 	xlog_get_iclog_buffer_size(mp, log);
1155 
1156 	error = ENOMEM;
1157 	bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1158 	if (!bp)
1159 		goto out_free_log;
1160 	XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1161 	XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1162 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1163 	ASSERT(XFS_BUF_ISBUSY(bp));
1164 	ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1165 	log->l_xbuf = bp;
1166 
1167 	spin_lock_init(&log->l_icloglock);
1168 	spin_lock_init(&log->l_grant_lock);
1169 	sv_init(&log->l_flush_wait, 0, "flush_wait");
1170 
1171 	/* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1172 	ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1173 
1174 	iclogp = &log->l_iclog;
1175 	/*
1176 	 * The amount of memory to allocate for the iclog structure is
1177 	 * rather funky due to the way the structure is defined.  It is
1178 	 * done this way so that we can use different sizes for machines
1179 	 * with different amounts of memory.  See the definition of
1180 	 * xlog_in_core_t in xfs_log_priv.h for details.
1181 	 */
1182 	iclogsize = log->l_iclog_size;
1183 	ASSERT(log->l_iclog_size >= 4096);
1184 	for (i=0; i < log->l_iclog_bufs; i++) {
1185 		*iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1186 		if (!*iclogp)
1187 			goto out_free_iclog;
1188 
1189 		iclog = *iclogp;
1190 		iclog->ic_prev = prev_iclog;
1191 		prev_iclog = iclog;
1192 
1193 		bp = xfs_buf_get_noaddr(log->l_iclog_size, mp->m_logdev_targp);
1194 		if (!bp)
1195 			goto out_free_iclog;
1196 		if (!XFS_BUF_CPSEMA(bp))
1197 			ASSERT(0);
1198 		XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1199 		XFS_BUF_SET_BDSTRAT_FUNC(bp, xlog_bdstrat_cb);
1200 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1201 		iclog->ic_bp = bp;
1202 		iclog->ic_data = bp->b_addr;
1203 #ifdef DEBUG
1204 		log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1205 #endif
1206 		head = &iclog->ic_header;
1207 		memset(head, 0, sizeof(xlog_rec_header_t));
1208 		head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1209 		head->h_version = cpu_to_be32(
1210 			xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1211 		head->h_size = cpu_to_be32(log->l_iclog_size);
1212 		/* new fields */
1213 		head->h_fmt = cpu_to_be32(XLOG_FMT);
1214 		memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1215 
1216 		iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1217 		iclog->ic_state = XLOG_STATE_ACTIVE;
1218 		iclog->ic_log = log;
1219 		atomic_set(&iclog->ic_refcnt, 0);
1220 		spin_lock_init(&iclog->ic_callback_lock);
1221 		iclog->ic_callback_tail = &(iclog->ic_callback);
1222 		iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1223 
1224 		ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1225 		ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1226 		sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1227 		sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1228 
1229 		iclogp = &iclog->ic_next;
1230 	}
1231 	*iclogp = log->l_iclog;			/* complete ring */
1232 	log->l_iclog->ic_prev = prev_iclog;	/* re-write 1st prev ptr */
1233 
1234 	return log;
1235 
1236 out_free_iclog:
1237 	for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1238 		prev_iclog = iclog->ic_next;
1239 		if (iclog->ic_bp) {
1240 			sv_destroy(&iclog->ic_force_wait);
1241 			sv_destroy(&iclog->ic_write_wait);
1242 			xfs_buf_free(iclog->ic_bp);
1243 		}
1244 		kmem_free(iclog);
1245 	}
1246 	spinlock_destroy(&log->l_icloglock);
1247 	spinlock_destroy(&log->l_grant_lock);
1248 	xfs_buf_free(log->l_xbuf);
1249 out_free_log:
1250 	kmem_free(log);
1251 out:
1252 	return ERR_PTR(-error);
1253 }	/* xlog_alloc_log */
1254 
1255 
1256 /*
1257  * Write out the commit record of a transaction associated with the given
1258  * ticket.  Return the lsn of the commit record.
1259  */
1260 STATIC int
1261 xlog_commit_record(xfs_mount_t  *mp,
1262 		   xlog_ticket_t *ticket,
1263 		   xlog_in_core_t **iclog,
1264 		   xfs_lsn_t	*commitlsnp)
1265 {
1266 	int		error;
1267 	xfs_log_iovec_t	reg[1];
1268 
1269 	reg[0].i_addr = NULL;
1270 	reg[0].i_len = 0;
1271 	XLOG_VEC_SET_TYPE(&reg[0], XLOG_REG_TYPE_COMMIT);
1272 
1273 	ASSERT_ALWAYS(iclog);
1274 	if ((error = xlog_write(mp, reg, 1, ticket, commitlsnp,
1275 			       iclog, XLOG_COMMIT_TRANS))) {
1276 		xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1277 	}
1278 	return error;
1279 }	/* xlog_commit_record */
1280 
1281 
1282 /*
1283  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1284  * log space.  This code pushes on the lsn which would supposedly free up
1285  * the 25% which we want to leave free.  We may need to adopt a policy which
1286  * pushes on an lsn which is further along in the log once we reach the high
1287  * water mark.  In this manner, we would be creating a low water mark.
1288  */
1289 STATIC void
1290 xlog_grant_push_ail(xfs_mount_t	*mp,
1291 		    int		need_bytes)
1292 {
1293     xlog_t	*log = mp->m_log;	/* pointer to the log */
1294     xfs_lsn_t	tail_lsn;		/* lsn of the log tail */
1295     xfs_lsn_t	threshold_lsn = 0;	/* lsn we'd like to be at */
1296     int		free_blocks;		/* free blocks left to write to */
1297     int		free_bytes;		/* free bytes left to write to */
1298     int		threshold_block;	/* block in lsn we'd like to be at */
1299     int		threshold_cycle;	/* lsn cycle we'd like to be at */
1300     int		free_threshold;
1301 
1302     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1303 
1304     spin_lock(&log->l_grant_lock);
1305     free_bytes = xlog_space_left(log,
1306 				 log->l_grant_reserve_cycle,
1307 				 log->l_grant_reserve_bytes);
1308     tail_lsn = log->l_tail_lsn;
1309     free_blocks = BTOBBT(free_bytes);
1310 
1311     /*
1312      * Set the threshold for the minimum number of free blocks in the
1313      * log to the maximum of what the caller needs, one quarter of the
1314      * log, and 256 blocks.
1315      */
1316     free_threshold = BTOBB(need_bytes);
1317     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1318     free_threshold = MAX(free_threshold, 256);
1319     if (free_blocks < free_threshold) {
1320 	threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1321 	threshold_cycle = CYCLE_LSN(tail_lsn);
1322 	if (threshold_block >= log->l_logBBsize) {
1323 	    threshold_block -= log->l_logBBsize;
1324 	    threshold_cycle += 1;
1325 	}
1326 	threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1327 
1328 	/* Don't pass in an lsn greater than the lsn of the last
1329 	 * log record known to be on disk.
1330 	 */
1331 	if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1332 	    threshold_lsn = log->l_last_sync_lsn;
1333     }
1334     spin_unlock(&log->l_grant_lock);
1335 
1336     /*
1337      * Get the transaction layer to kick the dirty buffers out to
1338      * disk asynchronously. No point in trying to do this if
1339      * the filesystem is shutting down.
1340      */
1341     if (threshold_lsn &&
1342 	!XLOG_FORCED_SHUTDOWN(log))
1343 	    xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1344 }	/* xlog_grant_push_ail */
1345 
1346 
1347 /*
1348  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1349  * fashion.  Previously, we should have moved the current iclog
1350  * ptr in the log to point to the next available iclog.  This allows further
1351  * write to continue while this code syncs out an iclog ready to go.
1352  * Before an in-core log can be written out, the data section must be scanned
1353  * to save away the 1st word of each BBSIZE block into the header.  We replace
1354  * it with the current cycle count.  Each BBSIZE block is tagged with the
1355  * cycle count because there in an implicit assumption that drives will
1356  * guarantee that entire 512 byte blocks get written at once.  In other words,
1357  * we can't have part of a 512 byte block written and part not written.  By
1358  * tagging each block, we will know which blocks are valid when recovering
1359  * after an unclean shutdown.
1360  *
1361  * This routine is single threaded on the iclog.  No other thread can be in
1362  * this routine with the same iclog.  Changing contents of iclog can there-
1363  * fore be done without grabbing the state machine lock.  Updating the global
1364  * log will require grabbing the lock though.
1365  *
1366  * The entire log manager uses a logical block numbering scheme.  Only
1367  * log_sync (and then only bwrite()) know about the fact that the log may
1368  * not start with block zero on a given device.  The log block start offset
1369  * is added immediately before calling bwrite().
1370  */
1371 
1372 STATIC int
1373 xlog_sync(xlog_t		*log,
1374 	  xlog_in_core_t	*iclog)
1375 {
1376 	xfs_caddr_t	dptr;		/* pointer to byte sized element */
1377 	xfs_buf_t	*bp;
1378 	int		i;
1379 	uint		count;		/* byte count of bwrite */
1380 	uint		count_init;	/* initial count before roundup */
1381 	int		roundoff;       /* roundoff to BB or stripe */
1382 	int		split = 0;	/* split write into two regions */
1383 	int		error;
1384 	int		v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1385 
1386 	XFS_STATS_INC(xs_log_writes);
1387 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1388 
1389 	/* Add for LR header */
1390 	count_init = log->l_iclog_hsize + iclog->ic_offset;
1391 
1392 	/* Round out the log write size */
1393 	if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1394 		/* we have a v2 stripe unit to use */
1395 		count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1396 	} else {
1397 		count = BBTOB(BTOBB(count_init));
1398 	}
1399 	roundoff = count - count_init;
1400 	ASSERT(roundoff >= 0);
1401 	ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1402                 roundoff < log->l_mp->m_sb.sb_logsunit)
1403 		||
1404 		(log->l_mp->m_sb.sb_logsunit <= 1 &&
1405 		 roundoff < BBTOB(1)));
1406 
1407 	/* move grant heads by roundoff in sync */
1408 	spin_lock(&log->l_grant_lock);
1409 	xlog_grant_add_space(log, roundoff);
1410 	spin_unlock(&log->l_grant_lock);
1411 
1412 	/* put cycle number in every block */
1413 	xlog_pack_data(log, iclog, roundoff);
1414 
1415 	/* real byte length */
1416 	if (v2) {
1417 		iclog->ic_header.h_len =
1418 			cpu_to_be32(iclog->ic_offset + roundoff);
1419 	} else {
1420 		iclog->ic_header.h_len =
1421 			cpu_to_be32(iclog->ic_offset);
1422 	}
1423 
1424 	bp = iclog->ic_bp;
1425 	ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1426 	XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1427 	XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1428 
1429 	XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1430 
1431 	/* Do we need to split this write into 2 parts? */
1432 	if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1433 		split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1434 		count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1435 		iclog->ic_bwritecnt = 2;	/* split into 2 writes */
1436 	} else {
1437 		iclog->ic_bwritecnt = 1;
1438 	}
1439 	XFS_BUF_SET_COUNT(bp, count);
1440 	XFS_BUF_SET_FSPRIVATE(bp, iclog);	/* save for later */
1441 	XFS_BUF_ZEROFLAGS(bp);
1442 	XFS_BUF_BUSY(bp);
1443 	XFS_BUF_ASYNC(bp);
1444 	bp->b_flags |= XBF_LOG_BUFFER;
1445 	/*
1446 	 * Do an ordered write for the log block.
1447 	 * Its unnecessary to flush the first split block in the log wrap case.
1448 	 */
1449 	if (!split && (log->l_mp->m_flags & XFS_MOUNT_BARRIER))
1450 		XFS_BUF_ORDERED(bp);
1451 
1452 	ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1453 	ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1454 
1455 	xlog_verify_iclog(log, iclog, count, B_TRUE);
1456 
1457 	/* account for log which doesn't start at block #0 */
1458 	XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1459 	/*
1460 	 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1461 	 * is shutting down.
1462 	 */
1463 	XFS_BUF_WRITE(bp);
1464 
1465 	if ((error = XFS_bwrite(bp))) {
1466 		xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1467 				  XFS_BUF_ADDR(bp));
1468 		return error;
1469 	}
1470 	if (split) {
1471 		bp = iclog->ic_log->l_xbuf;
1472 		ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1473 							(unsigned long)1);
1474 		XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1475 		XFS_BUF_SET_ADDR(bp, 0);	     /* logical 0 */
1476 		XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1477 					    (__psint_t)count), split);
1478 		XFS_BUF_SET_FSPRIVATE(bp, iclog);
1479 		XFS_BUF_ZEROFLAGS(bp);
1480 		XFS_BUF_BUSY(bp);
1481 		XFS_BUF_ASYNC(bp);
1482 		bp->b_flags |= XBF_LOG_BUFFER;
1483 		if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1484 			XFS_BUF_ORDERED(bp);
1485 		dptr = XFS_BUF_PTR(bp);
1486 		/*
1487 		 * Bump the cycle numbers at the start of each block
1488 		 * since this part of the buffer is at the start of
1489 		 * a new cycle.  Watch out for the header magic number
1490 		 * case, though.
1491 		 */
1492 		for (i = 0; i < split; i += BBSIZE) {
1493 			be32_add_cpu((__be32 *)dptr, 1);
1494 			if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1495 				be32_add_cpu((__be32 *)dptr, 1);
1496 			dptr += BBSIZE;
1497 		}
1498 
1499 		ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1500 		ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1501 
1502 		/* account for internal log which doesn't start at block #0 */
1503 		XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1504 		XFS_BUF_WRITE(bp);
1505 		if ((error = XFS_bwrite(bp))) {
1506 			xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1507 					  bp, XFS_BUF_ADDR(bp));
1508 			return error;
1509 		}
1510 	}
1511 	return 0;
1512 }	/* xlog_sync */
1513 
1514 
1515 /*
1516  * Deallocate a log structure
1517  */
1518 STATIC void
1519 xlog_dealloc_log(xlog_t *log)
1520 {
1521 	xlog_in_core_t	*iclog, *next_iclog;
1522 	int		i;
1523 
1524 	iclog = log->l_iclog;
1525 	for (i=0; i<log->l_iclog_bufs; i++) {
1526 		sv_destroy(&iclog->ic_force_wait);
1527 		sv_destroy(&iclog->ic_write_wait);
1528 		xfs_buf_free(iclog->ic_bp);
1529 		next_iclog = iclog->ic_next;
1530 		kmem_free(iclog);
1531 		iclog = next_iclog;
1532 	}
1533 	spinlock_destroy(&log->l_icloglock);
1534 	spinlock_destroy(&log->l_grant_lock);
1535 
1536 	xfs_buf_free(log->l_xbuf);
1537 	log->l_mp->m_log = NULL;
1538 	kmem_free(log);
1539 }	/* xlog_dealloc_log */
1540 
1541 /*
1542  * Update counters atomically now that memcpy is done.
1543  */
1544 /* ARGSUSED */
1545 static inline void
1546 xlog_state_finish_copy(xlog_t		*log,
1547 		       xlog_in_core_t	*iclog,
1548 		       int		record_cnt,
1549 		       int		copy_bytes)
1550 {
1551 	spin_lock(&log->l_icloglock);
1552 
1553 	be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1554 	iclog->ic_offset += copy_bytes;
1555 
1556 	spin_unlock(&log->l_icloglock);
1557 }	/* xlog_state_finish_copy */
1558 
1559 
1560 
1561 
1562 /*
1563  * print out info relating to regions written which consume
1564  * the reservation
1565  */
1566 STATIC void
1567 xlog_print_tic_res(xfs_mount_t *mp, xlog_ticket_t *ticket)
1568 {
1569 	uint i;
1570 	uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1571 
1572 	/* match with XLOG_REG_TYPE_* in xfs_log.h */
1573 	static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1574 	    "bformat",
1575 	    "bchunk",
1576 	    "efi_format",
1577 	    "efd_format",
1578 	    "iformat",
1579 	    "icore",
1580 	    "iext",
1581 	    "ibroot",
1582 	    "ilocal",
1583 	    "iattr_ext",
1584 	    "iattr_broot",
1585 	    "iattr_local",
1586 	    "qformat",
1587 	    "dquot",
1588 	    "quotaoff",
1589 	    "LR header",
1590 	    "unmount",
1591 	    "commit",
1592 	    "trans header"
1593 	};
1594 	static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1595 	    "SETATTR_NOT_SIZE",
1596 	    "SETATTR_SIZE",
1597 	    "INACTIVE",
1598 	    "CREATE",
1599 	    "CREATE_TRUNC",
1600 	    "TRUNCATE_FILE",
1601 	    "REMOVE",
1602 	    "LINK",
1603 	    "RENAME",
1604 	    "MKDIR",
1605 	    "RMDIR",
1606 	    "SYMLINK",
1607 	    "SET_DMATTRS",
1608 	    "GROWFS",
1609 	    "STRAT_WRITE",
1610 	    "DIOSTRAT",
1611 	    "WRITE_SYNC",
1612 	    "WRITEID",
1613 	    "ADDAFORK",
1614 	    "ATTRINVAL",
1615 	    "ATRUNCATE",
1616 	    "ATTR_SET",
1617 	    "ATTR_RM",
1618 	    "ATTR_FLAG",
1619 	    "CLEAR_AGI_BUCKET",
1620 	    "QM_SBCHANGE",
1621 	    "DUMMY1",
1622 	    "DUMMY2",
1623 	    "QM_QUOTAOFF",
1624 	    "QM_DQALLOC",
1625 	    "QM_SETQLIM",
1626 	    "QM_DQCLUSTER",
1627 	    "QM_QINOCREATE",
1628 	    "QM_QUOTAOFF_END",
1629 	    "SB_UNIT",
1630 	    "FSYNC_TS",
1631 	    "GROWFSRT_ALLOC",
1632 	    "GROWFSRT_ZERO",
1633 	    "GROWFSRT_FREE",
1634 	    "SWAPEXT"
1635 	};
1636 
1637 	xfs_fs_cmn_err(CE_WARN, mp,
1638 			"xfs_log_write: reservation summary:\n"
1639 			"  trans type  = %s (%u)\n"
1640 			"  unit res    = %d bytes\n"
1641 			"  current res = %d bytes\n"
1642 			"  total reg   = %u bytes (o/flow = %u bytes)\n"
1643 			"  ophdrs      = %u (ophdr space = %u bytes)\n"
1644 			"  ophdr + reg = %u bytes\n"
1645 			"  num regions = %u\n",
1646 			((ticket->t_trans_type <= 0 ||
1647 			  ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1648 			  "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1649 			ticket->t_trans_type,
1650 			ticket->t_unit_res,
1651 			ticket->t_curr_res,
1652 			ticket->t_res_arr_sum, ticket->t_res_o_flow,
1653 			ticket->t_res_num_ophdrs, ophdr_spc,
1654 			ticket->t_res_arr_sum +
1655 			ticket->t_res_o_flow + ophdr_spc,
1656 			ticket->t_res_num);
1657 
1658 	for (i = 0; i < ticket->t_res_num; i++) {
1659 		uint r_type = ticket->t_res_arr[i].r_type;
1660 		cmn_err(CE_WARN,
1661 			    "region[%u]: %s - %u bytes\n",
1662 			    i,
1663 			    ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1664 			    "bad-rtype" : res_type_str[r_type-1]),
1665 			    ticket->t_res_arr[i].r_len);
1666 	}
1667 }
1668 
1669 /*
1670  * Write some region out to in-core log
1671  *
1672  * This will be called when writing externally provided regions or when
1673  * writing out a commit record for a given transaction.
1674  *
1675  * General algorithm:
1676  *	1. Find total length of this write.  This may include adding to the
1677  *		lengths passed in.
1678  *	2. Check whether we violate the tickets reservation.
1679  *	3. While writing to this iclog
1680  *	    A. Reserve as much space in this iclog as can get
1681  *	    B. If this is first write, save away start lsn
1682  *	    C. While writing this region:
1683  *		1. If first write of transaction, write start record
1684  *		2. Write log operation header (header per region)
1685  *		3. Find out if we can fit entire region into this iclog
1686  *		4. Potentially, verify destination memcpy ptr
1687  *		5. Memcpy (partial) region
1688  *		6. If partial copy, release iclog; otherwise, continue
1689  *			copying more regions into current iclog
1690  *	4. Mark want sync bit (in simulation mode)
1691  *	5. Release iclog for potential flush to on-disk log.
1692  *
1693  * ERRORS:
1694  * 1.	Panic if reservation is overrun.  This should never happen since
1695  *	reservation amounts are generated internal to the filesystem.
1696  * NOTES:
1697  * 1. Tickets are single threaded data structures.
1698  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1699  *	syncing routine.  When a single log_write region needs to span
1700  *	multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1701  *	on all log operation writes which don't contain the end of the
1702  *	region.  The XLOG_END_TRANS bit is used for the in-core log
1703  *	operation which contains the end of the continued log_write region.
1704  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1705  *	we don't really know exactly how much space will be used.  As a result,
1706  *	we don't update ic_offset until the end when we know exactly how many
1707  *	bytes have been written out.
1708  */
1709 STATIC int
1710 xlog_write(xfs_mount_t *	mp,
1711 	   xfs_log_iovec_t	reg[],
1712 	   int			nentries,
1713 	   xfs_log_ticket_t	tic,
1714 	   xfs_lsn_t		*start_lsn,
1715 	   xlog_in_core_t	**commit_iclog,
1716 	   uint			flags)
1717 {
1718     xlog_t	     *log = mp->m_log;
1719     xlog_ticket_t    *ticket = (xlog_ticket_t *)tic;
1720     xlog_in_core_t   *iclog = NULL;  /* ptr to current in-core log */
1721     xlog_op_header_t *logop_head;    /* ptr to log operation header */
1722     __psint_t	     ptr;	     /* copy address into data region */
1723     int		     len;	     /* # xlog_write() bytes 2 still copy */
1724     int		     index;	     /* region index currently copying */
1725     int		     log_offset;     /* offset (from 0) into data region */
1726     int		     start_rec_copy; /* # bytes to copy for start record */
1727     int		     partial_copy;   /* did we split a region? */
1728     int		     partial_copy_len;/* # bytes copied if split region */
1729     int		     need_copy;	     /* # bytes need to memcpy this region */
1730     int		     copy_len;	     /* # bytes actually memcpy'ing */
1731     int		     copy_off;	     /* # bytes from entry start */
1732     int		     contwr;	     /* continued write of in-core log? */
1733     int		     error;
1734     int		     record_cnt = 0, data_cnt = 0;
1735 
1736     partial_copy_len = partial_copy = 0;
1737 
1738     /* Calculate potential maximum space.  Each region gets its own
1739      * xlog_op_header_t and may need to be double word aligned.
1740      */
1741     len = 0;
1742     if (ticket->t_flags & XLOG_TIC_INITED) {    /* acct for start rec of xact */
1743 	len += sizeof(xlog_op_header_t);
1744 	ticket->t_res_num_ophdrs++;
1745     }
1746 
1747     for (index = 0; index < nentries; index++) {
1748 	len += sizeof(xlog_op_header_t);	    /* each region gets >= 1 */
1749 	ticket->t_res_num_ophdrs++;
1750 	len += reg[index].i_len;
1751 	xlog_tic_add_region(ticket, reg[index].i_len, reg[index].i_type);
1752     }
1753     contwr = *start_lsn = 0;
1754 
1755     if (ticket->t_curr_res < len) {
1756 	xlog_print_tic_res(mp, ticket);
1757 #ifdef DEBUG
1758 	xlog_panic(
1759 		"xfs_log_write: reservation ran out. Need to up reservation");
1760 #else
1761 	/* Customer configurable panic */
1762 	xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1763 		"xfs_log_write: reservation ran out. Need to up reservation");
1764 	/* If we did not panic, shutdown the filesystem */
1765 	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1766 #endif
1767     } else
1768 	ticket->t_curr_res -= len;
1769 
1770     for (index = 0; index < nentries; ) {
1771 	if ((error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1772 					       &contwr, &log_offset)))
1773 		return error;
1774 
1775 	ASSERT(log_offset <= iclog->ic_size - 1);
1776 	ptr = (__psint_t) ((char *)iclog->ic_datap+log_offset);
1777 
1778 	/* start_lsn is the first lsn written to. That's all we need. */
1779 	if (! *start_lsn)
1780 	    *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1781 
1782 	/* This loop writes out as many regions as can fit in the amount
1783 	 * of space which was allocated by xlog_state_get_iclog_space().
1784 	 */
1785 	while (index < nentries) {
1786 	    ASSERT(reg[index].i_len % sizeof(__int32_t) == 0);
1787 	    ASSERT((__psint_t)ptr % sizeof(__int32_t) == 0);
1788 	    start_rec_copy = 0;
1789 
1790 	    /* If first write for transaction, insert start record.
1791 	     * We can't be trying to commit if we are inited.  We can't
1792 	     * have any "partial_copy" if we are inited.
1793 	     */
1794 	    if (ticket->t_flags & XLOG_TIC_INITED) {
1795 		logop_head		= (xlog_op_header_t *)ptr;
1796 		logop_head->oh_tid	= cpu_to_be32(ticket->t_tid);
1797 		logop_head->oh_clientid = ticket->t_clientid;
1798 		logop_head->oh_len	= 0;
1799 		logop_head->oh_flags    = XLOG_START_TRANS;
1800 		logop_head->oh_res2	= 0;
1801 		ticket->t_flags		&= ~XLOG_TIC_INITED;	/* clear bit */
1802 		record_cnt++;
1803 
1804 		start_rec_copy = sizeof(xlog_op_header_t);
1805 		xlog_write_adv_cnt(ptr, len, log_offset, start_rec_copy);
1806 	    }
1807 
1808 	    /* Copy log operation header directly into data section */
1809 	    logop_head			= (xlog_op_header_t *)ptr;
1810 	    logop_head->oh_tid		= cpu_to_be32(ticket->t_tid);
1811 	    logop_head->oh_clientid	= ticket->t_clientid;
1812 	    logop_head->oh_res2		= 0;
1813 
1814 	    /* header copied directly */
1815 	    xlog_write_adv_cnt(ptr, len, log_offset, sizeof(xlog_op_header_t));
1816 
1817 	    /* are we copying a commit or unmount record? */
1818 	    logop_head->oh_flags = flags;
1819 
1820 	    /*
1821 	     * We've seen logs corrupted with bad transaction client
1822 	     * ids.  This makes sure that XFS doesn't generate them on.
1823 	     * Turn this into an EIO and shut down the filesystem.
1824 	     */
1825 	    switch (logop_head->oh_clientid)  {
1826 	    case XFS_TRANSACTION:
1827 	    case XFS_VOLUME:
1828 	    case XFS_LOG:
1829 		break;
1830 	    default:
1831 		xfs_fs_cmn_err(CE_WARN, mp,
1832 		    "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1833 		    logop_head->oh_clientid, tic);
1834 		return XFS_ERROR(EIO);
1835 	    }
1836 
1837 	    /* Partial write last time? => (partial_copy != 0)
1838 	     * need_copy is the amount we'd like to copy if everything could
1839 	     * fit in the current memcpy.
1840 	     */
1841 	    need_copy =	reg[index].i_len - partial_copy_len;
1842 
1843 	    copy_off = partial_copy_len;
1844 	    if (need_copy <= iclog->ic_size - log_offset) { /*complete write */
1845 	        copy_len = need_copy;
1846 		logop_head->oh_len = cpu_to_be32(copy_len);
1847 		if (partial_copy)
1848 		    logop_head->oh_flags|= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1849 		partial_copy_len = partial_copy = 0;
1850 	    } else {					    /* partial write */
1851 		copy_len = iclog->ic_size - log_offset;
1852 		logop_head->oh_len = cpu_to_be32(copy_len);
1853 		logop_head->oh_flags |= XLOG_CONTINUE_TRANS;
1854 		if (partial_copy)
1855 			logop_head->oh_flags |= XLOG_WAS_CONT_TRANS;
1856 		partial_copy_len += copy_len;
1857 		partial_copy++;
1858 		len += sizeof(xlog_op_header_t); /* from splitting of region */
1859 		/* account for new log op header */
1860 		ticket->t_curr_res -= sizeof(xlog_op_header_t);
1861 		ticket->t_res_num_ophdrs++;
1862 	    }
1863 	    xlog_verify_dest_ptr(log, ptr);
1864 
1865 	    /* copy region */
1866 	    ASSERT(copy_len >= 0);
1867 	    memcpy((xfs_caddr_t)ptr, reg[index].i_addr + copy_off, copy_len);
1868 	    xlog_write_adv_cnt(ptr, len, log_offset, copy_len);
1869 
1870 	    /* make copy_len total bytes copied, including headers */
1871 	    copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1872 	    record_cnt++;
1873 	    data_cnt += contwr ? copy_len : 0;
1874 	    if (partial_copy) {			/* copied partial region */
1875 		    /* already marked WANT_SYNC by xlog_state_get_iclog_space */
1876 		    xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1877 		    record_cnt = data_cnt = 0;
1878 		    if ((error = xlog_state_release_iclog(log, iclog)))
1879 			    return error;
1880 		    break;			/* don't increment index */
1881 	    } else {				/* copied entire region */
1882 		index++;
1883 		partial_copy_len = partial_copy = 0;
1884 
1885 		if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1886 		    xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1887 		    record_cnt = data_cnt = 0;
1888 		    spin_lock(&log->l_icloglock);
1889 		    xlog_state_want_sync(log, iclog);
1890 		    spin_unlock(&log->l_icloglock);
1891 		    if (commit_iclog) {
1892 			ASSERT(flags & XLOG_COMMIT_TRANS);
1893 			*commit_iclog = iclog;
1894 		    } else if ((error = xlog_state_release_iclog(log, iclog)))
1895 			   return error;
1896 		    if (index == nentries)
1897 			    return 0;		/* we are done */
1898 		    else
1899 			    break;
1900 		}
1901 	    } /* if (partial_copy) */
1902 	} /* while (index < nentries) */
1903     } /* for (index = 0; index < nentries; ) */
1904     ASSERT(len == 0);
1905 
1906     xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1907     if (commit_iclog) {
1908 	ASSERT(flags & XLOG_COMMIT_TRANS);
1909 	*commit_iclog = iclog;
1910 	return 0;
1911     }
1912     return xlog_state_release_iclog(log, iclog);
1913 }	/* xlog_write */
1914 
1915 
1916 /*****************************************************************************
1917  *
1918  *		State Machine functions
1919  *
1920  *****************************************************************************
1921  */
1922 
1923 /* Clean iclogs starting from the head.  This ordering must be
1924  * maintained, so an iclog doesn't become ACTIVE beyond one that
1925  * is SYNCING.  This is also required to maintain the notion that we use
1926  * a ordered wait queue to hold off would be writers to the log when every
1927  * iclog is trying to sync to disk.
1928  *
1929  * State Change: DIRTY -> ACTIVE
1930  */
1931 STATIC void
1932 xlog_state_clean_log(xlog_t *log)
1933 {
1934 	xlog_in_core_t	*iclog;
1935 	int changed = 0;
1936 
1937 	iclog = log->l_iclog;
1938 	do {
1939 		if (iclog->ic_state == XLOG_STATE_DIRTY) {
1940 			iclog->ic_state	= XLOG_STATE_ACTIVE;
1941 			iclog->ic_offset       = 0;
1942 			ASSERT(iclog->ic_callback == NULL);
1943 			/*
1944 			 * If the number of ops in this iclog indicate it just
1945 			 * contains the dummy transaction, we can
1946 			 * change state into IDLE (the second time around).
1947 			 * Otherwise we should change the state into
1948 			 * NEED a dummy.
1949 			 * We don't need to cover the dummy.
1950 			 */
1951 			if (!changed &&
1952 			   (be32_to_cpu(iclog->ic_header.h_num_logops) ==
1953 			   		XLOG_COVER_OPS)) {
1954 				changed = 1;
1955 			} else {
1956 				/*
1957 				 * We have two dirty iclogs so start over
1958 				 * This could also be num of ops indicates
1959 				 * this is not the dummy going out.
1960 				 */
1961 				changed = 2;
1962 			}
1963 			iclog->ic_header.h_num_logops = 0;
1964 			memset(iclog->ic_header.h_cycle_data, 0,
1965 			      sizeof(iclog->ic_header.h_cycle_data));
1966 			iclog->ic_header.h_lsn = 0;
1967 		} else if (iclog->ic_state == XLOG_STATE_ACTIVE)
1968 			/* do nothing */;
1969 		else
1970 			break;	/* stop cleaning */
1971 		iclog = iclog->ic_next;
1972 	} while (iclog != log->l_iclog);
1973 
1974 	/* log is locked when we are called */
1975 	/*
1976 	 * Change state for the dummy log recording.
1977 	 * We usually go to NEED. But we go to NEED2 if the changed indicates
1978 	 * we are done writing the dummy record.
1979 	 * If we are done with the second dummy recored (DONE2), then
1980 	 * we go to IDLE.
1981 	 */
1982 	if (changed) {
1983 		switch (log->l_covered_state) {
1984 		case XLOG_STATE_COVER_IDLE:
1985 		case XLOG_STATE_COVER_NEED:
1986 		case XLOG_STATE_COVER_NEED2:
1987 			log->l_covered_state = XLOG_STATE_COVER_NEED;
1988 			break;
1989 
1990 		case XLOG_STATE_COVER_DONE:
1991 			if (changed == 1)
1992 				log->l_covered_state = XLOG_STATE_COVER_NEED2;
1993 			else
1994 				log->l_covered_state = XLOG_STATE_COVER_NEED;
1995 			break;
1996 
1997 		case XLOG_STATE_COVER_DONE2:
1998 			if (changed == 1)
1999 				log->l_covered_state = XLOG_STATE_COVER_IDLE;
2000 			else
2001 				log->l_covered_state = XLOG_STATE_COVER_NEED;
2002 			break;
2003 
2004 		default:
2005 			ASSERT(0);
2006 		}
2007 	}
2008 }	/* xlog_state_clean_log */
2009 
2010 STATIC xfs_lsn_t
2011 xlog_get_lowest_lsn(
2012 	xlog_t		*log)
2013 {
2014 	xlog_in_core_t  *lsn_log;
2015 	xfs_lsn_t	lowest_lsn, lsn;
2016 
2017 	lsn_log = log->l_iclog;
2018 	lowest_lsn = 0;
2019 	do {
2020 	    if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2021 		lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2022 		if ((lsn && !lowest_lsn) ||
2023 		    (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2024 			lowest_lsn = lsn;
2025 		}
2026 	    }
2027 	    lsn_log = lsn_log->ic_next;
2028 	} while (lsn_log != log->l_iclog);
2029 	return lowest_lsn;
2030 }
2031 
2032 
2033 STATIC void
2034 xlog_state_do_callback(
2035 	xlog_t		*log,
2036 	int		aborted,
2037 	xlog_in_core_t	*ciclog)
2038 {
2039 	xlog_in_core_t	   *iclog;
2040 	xlog_in_core_t	   *first_iclog;	/* used to know when we've
2041 						 * processed all iclogs once */
2042 	xfs_log_callback_t *cb, *cb_next;
2043 	int		   flushcnt = 0;
2044 	xfs_lsn_t	   lowest_lsn;
2045 	int		   ioerrors;	/* counter: iclogs with errors */
2046 	int		   loopdidcallbacks; /* flag: inner loop did callbacks*/
2047 	int		   funcdidcallbacks; /* flag: function did callbacks */
2048 	int		   repeats;	/* for issuing console warnings if
2049 					 * looping too many times */
2050 	int		   wake = 0;
2051 
2052 	spin_lock(&log->l_icloglock);
2053 	first_iclog = iclog = log->l_iclog;
2054 	ioerrors = 0;
2055 	funcdidcallbacks = 0;
2056 	repeats = 0;
2057 
2058 	do {
2059 		/*
2060 		 * Scan all iclogs starting with the one pointed to by the
2061 		 * log.  Reset this starting point each time the log is
2062 		 * unlocked (during callbacks).
2063 		 *
2064 		 * Keep looping through iclogs until one full pass is made
2065 		 * without running any callbacks.
2066 		 */
2067 		first_iclog = log->l_iclog;
2068 		iclog = log->l_iclog;
2069 		loopdidcallbacks = 0;
2070 		repeats++;
2071 
2072 		do {
2073 
2074 			/* skip all iclogs in the ACTIVE & DIRTY states */
2075 			if (iclog->ic_state &
2076 			    (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2077 				iclog = iclog->ic_next;
2078 				continue;
2079 			}
2080 
2081 			/*
2082 			 * Between marking a filesystem SHUTDOWN and stopping
2083 			 * the log, we do flush all iclogs to disk (if there
2084 			 * wasn't a log I/O error). So, we do want things to
2085 			 * go smoothly in case of just a SHUTDOWN  w/o a
2086 			 * LOG_IO_ERROR.
2087 			 */
2088 			if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2089 				/*
2090 				 * Can only perform callbacks in order.  Since
2091 				 * this iclog is not in the DONE_SYNC/
2092 				 * DO_CALLBACK state, we skip the rest and
2093 				 * just try to clean up.  If we set our iclog
2094 				 * to DO_CALLBACK, we will not process it when
2095 				 * we retry since a previous iclog is in the
2096 				 * CALLBACK and the state cannot change since
2097 				 * we are holding the l_icloglock.
2098 				 */
2099 				if (!(iclog->ic_state &
2100 					(XLOG_STATE_DONE_SYNC |
2101 						 XLOG_STATE_DO_CALLBACK))) {
2102 					if (ciclog && (ciclog->ic_state ==
2103 							XLOG_STATE_DONE_SYNC)) {
2104 						ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2105 					}
2106 					break;
2107 				}
2108 				/*
2109 				 * We now have an iclog that is in either the
2110 				 * DO_CALLBACK or DONE_SYNC states. The other
2111 				 * states (WANT_SYNC, SYNCING, or CALLBACK were
2112 				 * caught by the above if and are going to
2113 				 * clean (i.e. we aren't doing their callbacks)
2114 				 * see the above if.
2115 				 */
2116 
2117 				/*
2118 				 * We will do one more check here to see if we
2119 				 * have chased our tail around.
2120 				 */
2121 
2122 				lowest_lsn = xlog_get_lowest_lsn(log);
2123 				if (lowest_lsn &&
2124 				    XFS_LSN_CMP(lowest_lsn,
2125 				    		be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2126 					iclog = iclog->ic_next;
2127 					continue; /* Leave this iclog for
2128 						   * another thread */
2129 				}
2130 
2131 				iclog->ic_state = XLOG_STATE_CALLBACK;
2132 
2133 				spin_unlock(&log->l_icloglock);
2134 
2135 				/* l_last_sync_lsn field protected by
2136 				 * l_grant_lock. Don't worry about iclog's lsn.
2137 				 * No one else can be here except us.
2138 				 */
2139 				spin_lock(&log->l_grant_lock);
2140 				ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2141 				       be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2142 				log->l_last_sync_lsn =
2143 					be64_to_cpu(iclog->ic_header.h_lsn);
2144 				spin_unlock(&log->l_grant_lock);
2145 
2146 			} else {
2147 				spin_unlock(&log->l_icloglock);
2148 				ioerrors++;
2149 			}
2150 
2151 			/*
2152 			 * Keep processing entries in the callback list until
2153 			 * we come around and it is empty.  We need to
2154 			 * atomically see that the list is empty and change the
2155 			 * state to DIRTY so that we don't miss any more
2156 			 * callbacks being added.
2157 			 */
2158 			spin_lock(&iclog->ic_callback_lock);
2159 			cb = iclog->ic_callback;
2160 			while (cb) {
2161 				iclog->ic_callback_tail = &(iclog->ic_callback);
2162 				iclog->ic_callback = NULL;
2163 				spin_unlock(&iclog->ic_callback_lock);
2164 
2165 				/* perform callbacks in the order given */
2166 				for (; cb; cb = cb_next) {
2167 					cb_next = cb->cb_next;
2168 					cb->cb_func(cb->cb_arg, aborted);
2169 				}
2170 				spin_lock(&iclog->ic_callback_lock);
2171 				cb = iclog->ic_callback;
2172 			}
2173 
2174 			loopdidcallbacks++;
2175 			funcdidcallbacks++;
2176 
2177 			spin_lock(&log->l_icloglock);
2178 			ASSERT(iclog->ic_callback == NULL);
2179 			spin_unlock(&iclog->ic_callback_lock);
2180 			if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2181 				iclog->ic_state = XLOG_STATE_DIRTY;
2182 
2183 			/*
2184 			 * Transition from DIRTY to ACTIVE if applicable.
2185 			 * NOP if STATE_IOERROR.
2186 			 */
2187 			xlog_state_clean_log(log);
2188 
2189 			/* wake up threads waiting in xfs_log_force() */
2190 			sv_broadcast(&iclog->ic_force_wait);
2191 
2192 			iclog = iclog->ic_next;
2193 		} while (first_iclog != iclog);
2194 
2195 		if (repeats > 5000) {
2196 			flushcnt += repeats;
2197 			repeats = 0;
2198 			xfs_fs_cmn_err(CE_WARN, log->l_mp,
2199 				"%s: possible infinite loop (%d iterations)",
2200 				__func__, flushcnt);
2201 		}
2202 	} while (!ioerrors && loopdidcallbacks);
2203 
2204 	/*
2205 	 * make one last gasp attempt to see if iclogs are being left in
2206 	 * limbo..
2207 	 */
2208 #ifdef DEBUG
2209 	if (funcdidcallbacks) {
2210 		first_iclog = iclog = log->l_iclog;
2211 		do {
2212 			ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2213 			/*
2214 			 * Terminate the loop if iclogs are found in states
2215 			 * which will cause other threads to clean up iclogs.
2216 			 *
2217 			 * SYNCING - i/o completion will go through logs
2218 			 * DONE_SYNC - interrupt thread should be waiting for
2219 			 *              l_icloglock
2220 			 * IOERROR - give up hope all ye who enter here
2221 			 */
2222 			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2223 			    iclog->ic_state == XLOG_STATE_SYNCING ||
2224 			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2225 			    iclog->ic_state == XLOG_STATE_IOERROR )
2226 				break;
2227 			iclog = iclog->ic_next;
2228 		} while (first_iclog != iclog);
2229 	}
2230 #endif
2231 
2232 	if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2233 		wake = 1;
2234 	spin_unlock(&log->l_icloglock);
2235 
2236 	if (wake)
2237 		sv_broadcast(&log->l_flush_wait);
2238 }
2239 
2240 
2241 /*
2242  * Finish transitioning this iclog to the dirty state.
2243  *
2244  * Make sure that we completely execute this routine only when this is
2245  * the last call to the iclog.  There is a good chance that iclog flushes,
2246  * when we reach the end of the physical log, get turned into 2 separate
2247  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2248  * routine.  By using the reference count bwritecnt, we guarantee that only
2249  * the second completion goes through.
2250  *
2251  * Callbacks could take time, so they are done outside the scope of the
2252  * global state machine log lock.
2253  */
2254 STATIC void
2255 xlog_state_done_syncing(
2256 	xlog_in_core_t	*iclog,
2257 	int		aborted)
2258 {
2259 	xlog_t		   *log = iclog->ic_log;
2260 
2261 	spin_lock(&log->l_icloglock);
2262 
2263 	ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2264 	       iclog->ic_state == XLOG_STATE_IOERROR);
2265 	ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2266 	ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2267 
2268 
2269 	/*
2270 	 * If we got an error, either on the first buffer, or in the case of
2271 	 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2272 	 * and none should ever be attempted to be written to disk
2273 	 * again.
2274 	 */
2275 	if (iclog->ic_state != XLOG_STATE_IOERROR) {
2276 		if (--iclog->ic_bwritecnt == 1) {
2277 			spin_unlock(&log->l_icloglock);
2278 			return;
2279 		}
2280 		iclog->ic_state = XLOG_STATE_DONE_SYNC;
2281 	}
2282 
2283 	/*
2284 	 * Someone could be sleeping prior to writing out the next
2285 	 * iclog buffer, we wake them all, one will get to do the
2286 	 * I/O, the others get to wait for the result.
2287 	 */
2288 	sv_broadcast(&iclog->ic_write_wait);
2289 	spin_unlock(&log->l_icloglock);
2290 	xlog_state_do_callback(log, aborted, iclog);	/* also cleans log */
2291 }	/* xlog_state_done_syncing */
2292 
2293 
2294 /*
2295  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2296  * sleep.  We wait on the flush queue on the head iclog as that should be
2297  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2298  * we will wait here and all new writes will sleep until a sync completes.
2299  *
2300  * The in-core logs are used in a circular fashion. They are not used
2301  * out-of-order even when an iclog past the head is free.
2302  *
2303  * return:
2304  *	* log_offset where xlog_write() can start writing into the in-core
2305  *		log's data space.
2306  *	* in-core log pointer to which xlog_write() should write.
2307  *	* boolean indicating this is a continued write to an in-core log.
2308  *		If this is the last write, then the in-core log's offset field
2309  *		needs to be incremented, depending on the amount of data which
2310  *		is copied.
2311  */
2312 STATIC int
2313 xlog_state_get_iclog_space(xlog_t	  *log,
2314 			   int		  len,
2315 			   xlog_in_core_t **iclogp,
2316 			   xlog_ticket_t  *ticket,
2317 			   int		  *continued_write,
2318 			   int		  *logoffsetp)
2319 {
2320 	int		  log_offset;
2321 	xlog_rec_header_t *head;
2322 	xlog_in_core_t	  *iclog;
2323 	int		  error;
2324 
2325 restart:
2326 	spin_lock(&log->l_icloglock);
2327 	if (XLOG_FORCED_SHUTDOWN(log)) {
2328 		spin_unlock(&log->l_icloglock);
2329 		return XFS_ERROR(EIO);
2330 	}
2331 
2332 	iclog = log->l_iclog;
2333 	if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2334 		XFS_STATS_INC(xs_log_noiclogs);
2335 
2336 		/* Wait for log writes to have flushed */
2337 		sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2338 		goto restart;
2339 	}
2340 
2341 	head = &iclog->ic_header;
2342 
2343 	atomic_inc(&iclog->ic_refcnt);	/* prevents sync */
2344 	log_offset = iclog->ic_offset;
2345 
2346 	/* On the 1st write to an iclog, figure out lsn.  This works
2347 	 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2348 	 * committing to.  If the offset is set, that's how many blocks
2349 	 * must be written.
2350 	 */
2351 	if (log_offset == 0) {
2352 		ticket->t_curr_res -= log->l_iclog_hsize;
2353 		xlog_tic_add_region(ticket,
2354 				    log->l_iclog_hsize,
2355 				    XLOG_REG_TYPE_LRHEADER);
2356 		head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2357 		head->h_lsn = cpu_to_be64(
2358 			xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2359 		ASSERT(log->l_curr_block >= 0);
2360 	}
2361 
2362 	/* If there is enough room to write everything, then do it.  Otherwise,
2363 	 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2364 	 * bit is on, so this will get flushed out.  Don't update ic_offset
2365 	 * until you know exactly how many bytes get copied.  Therefore, wait
2366 	 * until later to update ic_offset.
2367 	 *
2368 	 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2369 	 * can fit into remaining data section.
2370 	 */
2371 	if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2372 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2373 
2374 		/*
2375 		 * If I'm the only one writing to this iclog, sync it to disk.
2376 		 * We need to do an atomic compare and decrement here to avoid
2377 		 * racing with concurrent atomic_dec_and_lock() calls in
2378 		 * xlog_state_release_iclog() when there is more than one
2379 		 * reference to the iclog.
2380 		 */
2381 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2382 			/* we are the only one */
2383 			spin_unlock(&log->l_icloglock);
2384 			error = xlog_state_release_iclog(log, iclog);
2385 			if (error)
2386 				return error;
2387 		} else {
2388 			spin_unlock(&log->l_icloglock);
2389 		}
2390 		goto restart;
2391 	}
2392 
2393 	/* Do we have enough room to write the full amount in the remainder
2394 	 * of this iclog?  Or must we continue a write on the next iclog and
2395 	 * mark this iclog as completely taken?  In the case where we switch
2396 	 * iclogs (to mark it taken), this particular iclog will release/sync
2397 	 * to disk in xlog_write().
2398 	 */
2399 	if (len <= iclog->ic_size - iclog->ic_offset) {
2400 		*continued_write = 0;
2401 		iclog->ic_offset += len;
2402 	} else {
2403 		*continued_write = 1;
2404 		xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2405 	}
2406 	*iclogp = iclog;
2407 
2408 	ASSERT(iclog->ic_offset <= iclog->ic_size);
2409 	spin_unlock(&log->l_icloglock);
2410 
2411 	*logoffsetp = log_offset;
2412 	return 0;
2413 }	/* xlog_state_get_iclog_space */
2414 
2415 /*
2416  * Atomically get the log space required for a log ticket.
2417  *
2418  * Once a ticket gets put onto the reserveq, it will only return after
2419  * the needed reservation is satisfied.
2420  */
2421 STATIC int
2422 xlog_grant_log_space(xlog_t	   *log,
2423 		     xlog_ticket_t *tic)
2424 {
2425 	int		 free_bytes;
2426 	int		 need_bytes;
2427 #ifdef DEBUG
2428 	xfs_lsn_t	 tail_lsn;
2429 #endif
2430 
2431 
2432 #ifdef DEBUG
2433 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2434 		panic("grant Recovery problem");
2435 #endif
2436 
2437 	/* Is there space or do we need to sleep? */
2438 	spin_lock(&log->l_grant_lock);
2439 
2440 	trace_xfs_log_grant_enter(log, tic);
2441 
2442 	/* something is already sleeping; insert new transaction at end */
2443 	if (log->l_reserve_headq) {
2444 		xlog_ins_ticketq(&log->l_reserve_headq, tic);
2445 
2446 		trace_xfs_log_grant_sleep1(log, tic);
2447 
2448 		/*
2449 		 * Gotta check this before going to sleep, while we're
2450 		 * holding the grant lock.
2451 		 */
2452 		if (XLOG_FORCED_SHUTDOWN(log))
2453 			goto error_return;
2454 
2455 		XFS_STATS_INC(xs_sleep_logspace);
2456 		sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2457 		/*
2458 		 * If we got an error, and the filesystem is shutting down,
2459 		 * we'll catch it down below. So just continue...
2460 		 */
2461 		trace_xfs_log_grant_wake1(log, tic);
2462 		spin_lock(&log->l_grant_lock);
2463 	}
2464 	if (tic->t_flags & XFS_LOG_PERM_RESERV)
2465 		need_bytes = tic->t_unit_res*tic->t_ocnt;
2466 	else
2467 		need_bytes = tic->t_unit_res;
2468 
2469 redo:
2470 	if (XLOG_FORCED_SHUTDOWN(log))
2471 		goto error_return;
2472 
2473 	free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2474 				     log->l_grant_reserve_bytes);
2475 	if (free_bytes < need_bytes) {
2476 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2477 			xlog_ins_ticketq(&log->l_reserve_headq, tic);
2478 
2479 		trace_xfs_log_grant_sleep2(log, tic);
2480 
2481 		spin_unlock(&log->l_grant_lock);
2482 		xlog_grant_push_ail(log->l_mp, need_bytes);
2483 		spin_lock(&log->l_grant_lock);
2484 
2485 		XFS_STATS_INC(xs_sleep_logspace);
2486 		sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2487 
2488 		spin_lock(&log->l_grant_lock);
2489 		if (XLOG_FORCED_SHUTDOWN(log))
2490 			goto error_return;
2491 
2492 		trace_xfs_log_grant_wake2(log, tic);
2493 
2494 		goto redo;
2495 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2496 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2497 
2498 	/* we've got enough space */
2499 	xlog_grant_add_space(log, need_bytes);
2500 #ifdef DEBUG
2501 	tail_lsn = log->l_tail_lsn;
2502 	/*
2503 	 * Check to make sure the grant write head didn't just over lap the
2504 	 * tail.  If the cycles are the same, we can't be overlapping.
2505 	 * Otherwise, make sure that the cycles differ by exactly one and
2506 	 * check the byte count.
2507 	 */
2508 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2509 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2510 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2511 	}
2512 #endif
2513 	trace_xfs_log_grant_exit(log, tic);
2514 	xlog_verify_grant_head(log, 1);
2515 	spin_unlock(&log->l_grant_lock);
2516 	return 0;
2517 
2518  error_return:
2519 	if (tic->t_flags & XLOG_TIC_IN_Q)
2520 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2521 
2522 	trace_xfs_log_grant_error(log, tic);
2523 
2524 	/*
2525 	 * If we are failing, make sure the ticket doesn't have any
2526 	 * current reservations. We don't want to add this back when
2527 	 * the ticket/transaction gets cancelled.
2528 	 */
2529 	tic->t_curr_res = 0;
2530 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2531 	spin_unlock(&log->l_grant_lock);
2532 	return XFS_ERROR(EIO);
2533 }	/* xlog_grant_log_space */
2534 
2535 
2536 /*
2537  * Replenish the byte reservation required by moving the grant write head.
2538  *
2539  *
2540  */
2541 STATIC int
2542 xlog_regrant_write_log_space(xlog_t	   *log,
2543 			     xlog_ticket_t *tic)
2544 {
2545 	int		free_bytes, need_bytes;
2546 	xlog_ticket_t	*ntic;
2547 #ifdef DEBUG
2548 	xfs_lsn_t	tail_lsn;
2549 #endif
2550 
2551 	tic->t_curr_res = tic->t_unit_res;
2552 	xlog_tic_reset_res(tic);
2553 
2554 	if (tic->t_cnt > 0)
2555 		return 0;
2556 
2557 #ifdef DEBUG
2558 	if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2559 		panic("regrant Recovery problem");
2560 #endif
2561 
2562 	spin_lock(&log->l_grant_lock);
2563 
2564 	trace_xfs_log_regrant_write_enter(log, tic);
2565 
2566 	if (XLOG_FORCED_SHUTDOWN(log))
2567 		goto error_return;
2568 
2569 	/* If there are other waiters on the queue then give them a
2570 	 * chance at logspace before us. Wake up the first waiters,
2571 	 * if we do not wake up all the waiters then go to sleep waiting
2572 	 * for more free space, otherwise try to get some space for
2573 	 * this transaction.
2574 	 */
2575 	need_bytes = tic->t_unit_res;
2576 	if ((ntic = log->l_write_headq)) {
2577 		free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2578 					     log->l_grant_write_bytes);
2579 		do {
2580 			ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2581 
2582 			if (free_bytes < ntic->t_unit_res)
2583 				break;
2584 			free_bytes -= ntic->t_unit_res;
2585 			sv_signal(&ntic->t_wait);
2586 			ntic = ntic->t_next;
2587 		} while (ntic != log->l_write_headq);
2588 
2589 		if (ntic != log->l_write_headq) {
2590 			if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2591 				xlog_ins_ticketq(&log->l_write_headq, tic);
2592 
2593 			trace_xfs_log_regrant_write_sleep1(log, tic);
2594 
2595 			spin_unlock(&log->l_grant_lock);
2596 			xlog_grant_push_ail(log->l_mp, need_bytes);
2597 			spin_lock(&log->l_grant_lock);
2598 
2599 			XFS_STATS_INC(xs_sleep_logspace);
2600 			sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2601 				&log->l_grant_lock, s);
2602 
2603 			/* If we're shutting down, this tic is already
2604 			 * off the queue */
2605 			spin_lock(&log->l_grant_lock);
2606 			if (XLOG_FORCED_SHUTDOWN(log))
2607 				goto error_return;
2608 
2609 			trace_xfs_log_regrant_write_wake1(log, tic);
2610 		}
2611 	}
2612 
2613 redo:
2614 	if (XLOG_FORCED_SHUTDOWN(log))
2615 		goto error_return;
2616 
2617 	free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2618 				     log->l_grant_write_bytes);
2619 	if (free_bytes < need_bytes) {
2620 		if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2621 			xlog_ins_ticketq(&log->l_write_headq, tic);
2622 		spin_unlock(&log->l_grant_lock);
2623 		xlog_grant_push_ail(log->l_mp, need_bytes);
2624 		spin_lock(&log->l_grant_lock);
2625 
2626 		XFS_STATS_INC(xs_sleep_logspace);
2627 		trace_xfs_log_regrant_write_sleep2(log, tic);
2628 
2629 		sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2630 
2631 		/* If we're shutting down, this tic is already off the queue */
2632 		spin_lock(&log->l_grant_lock);
2633 		if (XLOG_FORCED_SHUTDOWN(log))
2634 			goto error_return;
2635 
2636 		trace_xfs_log_regrant_write_wake2(log, tic);
2637 		goto redo;
2638 	} else if (tic->t_flags & XLOG_TIC_IN_Q)
2639 		xlog_del_ticketq(&log->l_write_headq, tic);
2640 
2641 	/* we've got enough space */
2642 	xlog_grant_add_space_write(log, need_bytes);
2643 #ifdef DEBUG
2644 	tail_lsn = log->l_tail_lsn;
2645 	if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2646 		ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2647 		ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2648 	}
2649 #endif
2650 
2651 	trace_xfs_log_regrant_write_exit(log, tic);
2652 
2653 	xlog_verify_grant_head(log, 1);
2654 	spin_unlock(&log->l_grant_lock);
2655 	return 0;
2656 
2657 
2658  error_return:
2659 	if (tic->t_flags & XLOG_TIC_IN_Q)
2660 		xlog_del_ticketq(&log->l_reserve_headq, tic);
2661 
2662 	trace_xfs_log_regrant_write_error(log, tic);
2663 
2664 	/*
2665 	 * If we are failing, make sure the ticket doesn't have any
2666 	 * current reservations. We don't want to add this back when
2667 	 * the ticket/transaction gets cancelled.
2668 	 */
2669 	tic->t_curr_res = 0;
2670 	tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2671 	spin_unlock(&log->l_grant_lock);
2672 	return XFS_ERROR(EIO);
2673 }	/* xlog_regrant_write_log_space */
2674 
2675 
2676 /* The first cnt-1 times through here we don't need to
2677  * move the grant write head because the permanent
2678  * reservation has reserved cnt times the unit amount.
2679  * Release part of current permanent unit reservation and
2680  * reset current reservation to be one units worth.  Also
2681  * move grant reservation head forward.
2682  */
2683 STATIC void
2684 xlog_regrant_reserve_log_space(xlog_t	     *log,
2685 			       xlog_ticket_t *ticket)
2686 {
2687 	trace_xfs_log_regrant_reserve_enter(log, ticket);
2688 
2689 	if (ticket->t_cnt > 0)
2690 		ticket->t_cnt--;
2691 
2692 	spin_lock(&log->l_grant_lock);
2693 	xlog_grant_sub_space(log, ticket->t_curr_res);
2694 	ticket->t_curr_res = ticket->t_unit_res;
2695 	xlog_tic_reset_res(ticket);
2696 
2697 	trace_xfs_log_regrant_reserve_sub(log, ticket);
2698 
2699 	xlog_verify_grant_head(log, 1);
2700 
2701 	/* just return if we still have some of the pre-reserved space */
2702 	if (ticket->t_cnt > 0) {
2703 		spin_unlock(&log->l_grant_lock);
2704 		return;
2705 	}
2706 
2707 	xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2708 
2709 	trace_xfs_log_regrant_reserve_exit(log, ticket);
2710 
2711 	xlog_verify_grant_head(log, 0);
2712 	spin_unlock(&log->l_grant_lock);
2713 	ticket->t_curr_res = ticket->t_unit_res;
2714 	xlog_tic_reset_res(ticket);
2715 }	/* xlog_regrant_reserve_log_space */
2716 
2717 
2718 /*
2719  * Give back the space left from a reservation.
2720  *
2721  * All the information we need to make a correct determination of space left
2722  * is present.  For non-permanent reservations, things are quite easy.  The
2723  * count should have been decremented to zero.  We only need to deal with the
2724  * space remaining in the current reservation part of the ticket.  If the
2725  * ticket contains a permanent reservation, there may be left over space which
2726  * needs to be released.  A count of N means that N-1 refills of the current
2727  * reservation can be done before we need to ask for more space.  The first
2728  * one goes to fill up the first current reservation.  Once we run out of
2729  * space, the count will stay at zero and the only space remaining will be
2730  * in the current reservation field.
2731  */
2732 STATIC void
2733 xlog_ungrant_log_space(xlog_t	     *log,
2734 		       xlog_ticket_t *ticket)
2735 {
2736 	if (ticket->t_cnt > 0)
2737 		ticket->t_cnt--;
2738 
2739 	spin_lock(&log->l_grant_lock);
2740 	trace_xfs_log_ungrant_enter(log, ticket);
2741 
2742 	xlog_grant_sub_space(log, ticket->t_curr_res);
2743 
2744 	trace_xfs_log_ungrant_sub(log, ticket);
2745 
2746 	/* If this is a permanent reservation ticket, we may be able to free
2747 	 * up more space based on the remaining count.
2748 	 */
2749 	if (ticket->t_cnt > 0) {
2750 		ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2751 		xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2752 	}
2753 
2754 	trace_xfs_log_ungrant_exit(log, ticket);
2755 
2756 	xlog_verify_grant_head(log, 1);
2757 	spin_unlock(&log->l_grant_lock);
2758 	xfs_log_move_tail(log->l_mp, 1);
2759 }	/* xlog_ungrant_log_space */
2760 
2761 
2762 /*
2763  * Flush iclog to disk if this is the last reference to the given iclog and
2764  * the WANT_SYNC bit is set.
2765  *
2766  * When this function is entered, the iclog is not necessarily in the
2767  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2768  *
2769  *
2770  */
2771 STATIC int
2772 xlog_state_release_iclog(
2773 	xlog_t		*log,
2774 	xlog_in_core_t	*iclog)
2775 {
2776 	int		sync = 0;	/* do we sync? */
2777 
2778 	if (iclog->ic_state & XLOG_STATE_IOERROR)
2779 		return XFS_ERROR(EIO);
2780 
2781 	ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2782 	if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2783 		return 0;
2784 
2785 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2786 		spin_unlock(&log->l_icloglock);
2787 		return XFS_ERROR(EIO);
2788 	}
2789 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2790 	       iclog->ic_state == XLOG_STATE_WANT_SYNC);
2791 
2792 	if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2793 		/* update tail before writing to iclog */
2794 		xlog_assign_tail_lsn(log->l_mp);
2795 		sync++;
2796 		iclog->ic_state = XLOG_STATE_SYNCING;
2797 		iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2798 		xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2799 		/* cycle incremented when incrementing curr_block */
2800 	}
2801 	spin_unlock(&log->l_icloglock);
2802 
2803 	/*
2804 	 * We let the log lock go, so it's possible that we hit a log I/O
2805 	 * error or some other SHUTDOWN condition that marks the iclog
2806 	 * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2807 	 * this iclog has consistent data, so we ignore IOERROR
2808 	 * flags after this point.
2809 	 */
2810 	if (sync)
2811 		return xlog_sync(log, iclog);
2812 	return 0;
2813 }	/* xlog_state_release_iclog */
2814 
2815 
2816 /*
2817  * This routine will mark the current iclog in the ring as WANT_SYNC
2818  * and move the current iclog pointer to the next iclog in the ring.
2819  * When this routine is called from xlog_state_get_iclog_space(), the
2820  * exact size of the iclog has not yet been determined.  All we know is
2821  * that every data block.  We have run out of space in this log record.
2822  */
2823 STATIC void
2824 xlog_state_switch_iclogs(xlog_t		*log,
2825 			 xlog_in_core_t *iclog,
2826 			 int		eventual_size)
2827 {
2828 	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2829 	if (!eventual_size)
2830 		eventual_size = iclog->ic_offset;
2831 	iclog->ic_state = XLOG_STATE_WANT_SYNC;
2832 	iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2833 	log->l_prev_block = log->l_curr_block;
2834 	log->l_prev_cycle = log->l_curr_cycle;
2835 
2836 	/* roll log?: ic_offset changed later */
2837 	log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2838 
2839 	/* Round up to next log-sunit */
2840 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2841 	    log->l_mp->m_sb.sb_logsunit > 1) {
2842 		__uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2843 		log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2844 	}
2845 
2846 	if (log->l_curr_block >= log->l_logBBsize) {
2847 		log->l_curr_cycle++;
2848 		if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2849 			log->l_curr_cycle++;
2850 		log->l_curr_block -= log->l_logBBsize;
2851 		ASSERT(log->l_curr_block >= 0);
2852 	}
2853 	ASSERT(iclog == log->l_iclog);
2854 	log->l_iclog = iclog->ic_next;
2855 }	/* xlog_state_switch_iclogs */
2856 
2857 
2858 /*
2859  * Write out all data in the in-core log as of this exact moment in time.
2860  *
2861  * Data may be written to the in-core log during this call.  However,
2862  * we don't guarantee this data will be written out.  A change from past
2863  * implementation means this routine will *not* write out zero length LRs.
2864  *
2865  * Basically, we try and perform an intelligent scan of the in-core logs.
2866  * If we determine there is no flushable data, we just return.  There is no
2867  * flushable data if:
2868  *
2869  *	1. the current iclog is active and has no data; the previous iclog
2870  *		is in the active or dirty state.
2871  *	2. the current iclog is drity, and the previous iclog is in the
2872  *		active or dirty state.
2873  *
2874  * We may sleep if:
2875  *
2876  *	1. the current iclog is not in the active nor dirty state.
2877  *	2. the current iclog dirty, and the previous iclog is not in the
2878  *		active nor dirty state.
2879  *	3. the current iclog is active, and there is another thread writing
2880  *		to this particular iclog.
2881  *	4. a) the current iclog is active and has no other writers
2882  *	   b) when we return from flushing out this iclog, it is still
2883  *		not in the active nor dirty state.
2884  */
2885 STATIC int
2886 xlog_state_sync_all(xlog_t *log, uint flags, int *log_flushed)
2887 {
2888 	xlog_in_core_t	*iclog;
2889 	xfs_lsn_t	lsn;
2890 
2891 	spin_lock(&log->l_icloglock);
2892 
2893 	iclog = log->l_iclog;
2894 	if (iclog->ic_state & XLOG_STATE_IOERROR) {
2895 		spin_unlock(&log->l_icloglock);
2896 		return XFS_ERROR(EIO);
2897 	}
2898 
2899 	/* If the head iclog is not active nor dirty, we just attach
2900 	 * ourselves to the head and go to sleep.
2901 	 */
2902 	if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2903 	    iclog->ic_state == XLOG_STATE_DIRTY) {
2904 		/*
2905 		 * If the head is dirty or (active and empty), then
2906 		 * we need to look at the previous iclog.  If the previous
2907 		 * iclog is active or dirty we are done.  There is nothing
2908 		 * to sync out.  Otherwise, we attach ourselves to the
2909 		 * previous iclog and go to sleep.
2910 		 */
2911 		if (iclog->ic_state == XLOG_STATE_DIRTY ||
2912 		    (atomic_read(&iclog->ic_refcnt) == 0
2913 		     && iclog->ic_offset == 0)) {
2914 			iclog = iclog->ic_prev;
2915 			if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2916 			    iclog->ic_state == XLOG_STATE_DIRTY)
2917 				goto no_sleep;
2918 			else
2919 				goto maybe_sleep;
2920 		} else {
2921 			if (atomic_read(&iclog->ic_refcnt) == 0) {
2922 				/* We are the only one with access to this
2923 				 * iclog.  Flush it out now.  There should
2924 				 * be a roundoff of zero to show that someone
2925 				 * has already taken care of the roundoff from
2926 				 * the previous sync.
2927 				 */
2928 				atomic_inc(&iclog->ic_refcnt);
2929 				lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2930 				xlog_state_switch_iclogs(log, iclog, 0);
2931 				spin_unlock(&log->l_icloglock);
2932 
2933 				if (xlog_state_release_iclog(log, iclog))
2934 					return XFS_ERROR(EIO);
2935 				*log_flushed = 1;
2936 				spin_lock(&log->l_icloglock);
2937 				if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2938 				    iclog->ic_state != XLOG_STATE_DIRTY)
2939 					goto maybe_sleep;
2940 				else
2941 					goto no_sleep;
2942 			} else {
2943 				/* Someone else is writing to this iclog.
2944 				 * Use its call to flush out the data.  However,
2945 				 * the other thread may not force out this LR,
2946 				 * so we mark it WANT_SYNC.
2947 				 */
2948 				xlog_state_switch_iclogs(log, iclog, 0);
2949 				goto maybe_sleep;
2950 			}
2951 		}
2952 	}
2953 
2954 	/* By the time we come around again, the iclog could've been filled
2955 	 * which would give it another lsn.  If we have a new lsn, just
2956 	 * return because the relevant data has been flushed.
2957 	 */
2958 maybe_sleep:
2959 	if (flags & XFS_LOG_SYNC) {
2960 		/*
2961 		 * We must check if we're shutting down here, before
2962 		 * we wait, while we're holding the l_icloglock.
2963 		 * Then we check again after waking up, in case our
2964 		 * sleep was disturbed by a bad news.
2965 		 */
2966 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
2967 			spin_unlock(&log->l_icloglock);
2968 			return XFS_ERROR(EIO);
2969 		}
2970 		XFS_STATS_INC(xs_log_force_sleep);
2971 		sv_wait(&iclog->ic_force_wait, PINOD, &log->l_icloglock, s);
2972 		/*
2973 		 * No need to grab the log lock here since we're
2974 		 * only deciding whether or not to return EIO
2975 		 * and the memory read should be atomic.
2976 		 */
2977 		if (iclog->ic_state & XLOG_STATE_IOERROR)
2978 			return XFS_ERROR(EIO);
2979 		*log_flushed = 1;
2980 
2981 	} else {
2982 
2983 no_sleep:
2984 		spin_unlock(&log->l_icloglock);
2985 	}
2986 	return 0;
2987 }	/* xlog_state_sync_all */
2988 
2989 
2990 /*
2991  * Used by code which implements synchronous log forces.
2992  *
2993  * Find in-core log with lsn.
2994  *	If it is in the DIRTY state, just return.
2995  *	If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2996  *		state and go to sleep or return.
2997  *	If it is in any other state, go to sleep or return.
2998  *
2999  * If filesystem activity goes to zero, the iclog will get flushed only by
3000  * bdflush().
3001  */
3002 STATIC int
3003 xlog_state_sync(xlog_t	  *log,
3004 		xfs_lsn_t lsn,
3005 		uint	  flags,
3006 		int	  *log_flushed)
3007 {
3008     xlog_in_core_t	*iclog;
3009     int			already_slept = 0;
3010 
3011 try_again:
3012     spin_lock(&log->l_icloglock);
3013     iclog = log->l_iclog;
3014 
3015     if (iclog->ic_state & XLOG_STATE_IOERROR) {
3016 	    spin_unlock(&log->l_icloglock);
3017 	    return XFS_ERROR(EIO);
3018     }
3019 
3020     do {
3021 	if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3022 		iclog = iclog->ic_next;
3023 		continue;
3024 	}
3025 
3026 	if (iclog->ic_state == XLOG_STATE_DIRTY) {
3027 		spin_unlock(&log->l_icloglock);
3028 		return 0;
3029 	}
3030 
3031 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3032 		/*
3033 		 * We sleep here if we haven't already slept (e.g.
3034 		 * this is the first time we've looked at the correct
3035 		 * iclog buf) and the buffer before us is going to
3036 		 * be sync'ed. The reason for this is that if we
3037 		 * are doing sync transactions here, by waiting for
3038 		 * the previous I/O to complete, we can allow a few
3039 		 * more transactions into this iclog before we close
3040 		 * it down.
3041 		 *
3042 		 * Otherwise, we mark the buffer WANT_SYNC, and bump
3043 		 * up the refcnt so we can release the log (which drops
3044 		 * the ref count).  The state switch keeps new transaction
3045 		 * commits from using this buffer.  When the current commits
3046 		 * finish writing into the buffer, the refcount will drop to
3047 		 * zero and the buffer will go out then.
3048 		 */
3049 		if (!already_slept &&
3050 		    (iclog->ic_prev->ic_state & (XLOG_STATE_WANT_SYNC |
3051 						 XLOG_STATE_SYNCING))) {
3052 			ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3053 			XFS_STATS_INC(xs_log_force_sleep);
3054 			sv_wait(&iclog->ic_prev->ic_write_wait, PSWP,
3055 				&log->l_icloglock, s);
3056 			*log_flushed = 1;
3057 			already_slept = 1;
3058 			goto try_again;
3059 		} else {
3060 			atomic_inc(&iclog->ic_refcnt);
3061 			xlog_state_switch_iclogs(log, iclog, 0);
3062 			spin_unlock(&log->l_icloglock);
3063 			if (xlog_state_release_iclog(log, iclog))
3064 				return XFS_ERROR(EIO);
3065 			*log_flushed = 1;
3066 			spin_lock(&log->l_icloglock);
3067 		}
3068 	}
3069 
3070 	if ((flags & XFS_LOG_SYNC) && /* sleep */
3071 	    !(iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3072 
3073 		/*
3074 		 * Don't wait on completion if we know that we've
3075 		 * gotten a log write error.
3076 		 */
3077 		if (iclog->ic_state & XLOG_STATE_IOERROR) {
3078 			spin_unlock(&log->l_icloglock);
3079 			return XFS_ERROR(EIO);
3080 		}
3081 		XFS_STATS_INC(xs_log_force_sleep);
3082 		sv_wait(&iclog->ic_force_wait, PSWP, &log->l_icloglock, s);
3083 		/*
3084 		 * No need to grab the log lock here since we're
3085 		 * only deciding whether or not to return EIO
3086 		 * and the memory read should be atomic.
3087 		 */
3088 		if (iclog->ic_state & XLOG_STATE_IOERROR)
3089 			return XFS_ERROR(EIO);
3090 		*log_flushed = 1;
3091 	} else {		/* just return */
3092 		spin_unlock(&log->l_icloglock);
3093 	}
3094 	return 0;
3095 
3096     } while (iclog != log->l_iclog);
3097 
3098     spin_unlock(&log->l_icloglock);
3099     return 0;
3100 }	/* xlog_state_sync */
3101 
3102 
3103 /*
3104  * Called when we want to mark the current iclog as being ready to sync to
3105  * disk.
3106  */
3107 STATIC void
3108 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3109 {
3110 	assert_spin_locked(&log->l_icloglock);
3111 
3112 	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3113 		xlog_state_switch_iclogs(log, iclog, 0);
3114 	} else {
3115 		ASSERT(iclog->ic_state &
3116 			(XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3117 	}
3118 }
3119 
3120 
3121 /*****************************************************************************
3122  *
3123  *		TICKET functions
3124  *
3125  *****************************************************************************
3126  */
3127 
3128 /*
3129  * Free a used ticket when its refcount falls to zero.
3130  */
3131 void
3132 xfs_log_ticket_put(
3133 	xlog_ticket_t	*ticket)
3134 {
3135 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3136 	if (atomic_dec_and_test(&ticket->t_ref)) {
3137 		sv_destroy(&ticket->t_wait);
3138 		kmem_zone_free(xfs_log_ticket_zone, ticket);
3139 	}
3140 }
3141 
3142 xlog_ticket_t *
3143 xfs_log_ticket_get(
3144 	xlog_ticket_t	*ticket)
3145 {
3146 	ASSERT(atomic_read(&ticket->t_ref) > 0);
3147 	atomic_inc(&ticket->t_ref);
3148 	return ticket;
3149 }
3150 
3151 /*
3152  * Allocate and initialise a new log ticket.
3153  */
3154 STATIC xlog_ticket_t *
3155 xlog_ticket_alloc(xlog_t		*log,
3156 		int		unit_bytes,
3157 		int		cnt,
3158 		char		client,
3159 		uint		xflags)
3160 {
3161 	xlog_ticket_t	*tic;
3162 	uint		num_headers;
3163 
3164 	tic = kmem_zone_zalloc(xfs_log_ticket_zone, KM_SLEEP|KM_MAYFAIL);
3165 	if (!tic)
3166 		return NULL;
3167 
3168 	/*
3169 	 * Permanent reservations have up to 'cnt'-1 active log operations
3170 	 * in the log.  A unit in this case is the amount of space for one
3171 	 * of these log operations.  Normal reservations have a cnt of 1
3172 	 * and their unit amount is the total amount of space required.
3173 	 *
3174 	 * The following lines of code account for non-transaction data
3175 	 * which occupy space in the on-disk log.
3176 	 *
3177 	 * Normal form of a transaction is:
3178 	 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3179 	 * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3180 	 *
3181 	 * We need to account for all the leadup data and trailer data
3182 	 * around the transaction data.
3183 	 * And then we need to account for the worst case in terms of using
3184 	 * more space.
3185 	 * The worst case will happen if:
3186 	 * - the placement of the transaction happens to be such that the
3187 	 *   roundoff is at its maximum
3188 	 * - the transaction data is synced before the commit record is synced
3189 	 *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3190 	 *   Therefore the commit record is in its own Log Record.
3191 	 *   This can happen as the commit record is called with its
3192 	 *   own region to xlog_write().
3193 	 *   This then means that in the worst case, roundoff can happen for
3194 	 *   the commit-rec as well.
3195 	 *   The commit-rec is smaller than padding in this scenario and so it is
3196 	 *   not added separately.
3197 	 */
3198 
3199 	/* for trans header */
3200 	unit_bytes += sizeof(xlog_op_header_t);
3201 	unit_bytes += sizeof(xfs_trans_header_t);
3202 
3203 	/* for start-rec */
3204 	unit_bytes += sizeof(xlog_op_header_t);
3205 
3206 	/* for LR headers */
3207 	num_headers = ((unit_bytes + log->l_iclog_size-1) >> log->l_iclog_size_log);
3208 	unit_bytes += log->l_iclog_hsize * num_headers;
3209 
3210 	/* for commit-rec LR header - note: padding will subsume the ophdr */
3211 	unit_bytes += log->l_iclog_hsize;
3212 
3213 	/* for split-recs - ophdrs added when data split over LRs */
3214 	unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3215 
3216 	/* for roundoff padding for transaction data and one for commit record */
3217 	if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3218 	    log->l_mp->m_sb.sb_logsunit > 1) {
3219 		/* log su roundoff */
3220 		unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3221 	} else {
3222 		/* BB roundoff */
3223 		unit_bytes += 2*BBSIZE;
3224         }
3225 
3226 	atomic_set(&tic->t_ref, 1);
3227 	tic->t_unit_res		= unit_bytes;
3228 	tic->t_curr_res		= unit_bytes;
3229 	tic->t_cnt		= cnt;
3230 	tic->t_ocnt		= cnt;
3231 	tic->t_tid		= (xlog_tid_t)((__psint_t)tic & 0xffffffff);
3232 	tic->t_clientid		= client;
3233 	tic->t_flags		= XLOG_TIC_INITED;
3234 	tic->t_trans_type	= 0;
3235 	if (xflags & XFS_LOG_PERM_RESERV)
3236 		tic->t_flags |= XLOG_TIC_PERM_RESERV;
3237 	sv_init(&(tic->t_wait), SV_DEFAULT, "logtick");
3238 
3239 	xlog_tic_reset_res(tic);
3240 
3241 	return tic;
3242 }
3243 
3244 
3245 /******************************************************************************
3246  *
3247  *		Log debug routines
3248  *
3249  ******************************************************************************
3250  */
3251 #if defined(DEBUG)
3252 /*
3253  * Make sure that the destination ptr is within the valid data region of
3254  * one of the iclogs.  This uses backup pointers stored in a different
3255  * part of the log in case we trash the log structure.
3256  */
3257 void
3258 xlog_verify_dest_ptr(xlog_t     *log,
3259 		     __psint_t  ptr)
3260 {
3261 	int i;
3262 	int good_ptr = 0;
3263 
3264 	for (i=0; i < log->l_iclog_bufs; i++) {
3265 		if (ptr >= (__psint_t)log->l_iclog_bak[i] &&
3266 		    ptr <= (__psint_t)log->l_iclog_bak[i]+log->l_iclog_size)
3267 			good_ptr++;
3268 	}
3269 	if (! good_ptr)
3270 		xlog_panic("xlog_verify_dest_ptr: invalid ptr");
3271 }	/* xlog_verify_dest_ptr */
3272 
3273 STATIC void
3274 xlog_verify_grant_head(xlog_t *log, int equals)
3275 {
3276     if (log->l_grant_reserve_cycle == log->l_grant_write_cycle) {
3277 	if (equals)
3278 	    ASSERT(log->l_grant_reserve_bytes >= log->l_grant_write_bytes);
3279 	else
3280 	    ASSERT(log->l_grant_reserve_bytes > log->l_grant_write_bytes);
3281     } else {
3282 	ASSERT(log->l_grant_reserve_cycle-1 == log->l_grant_write_cycle);
3283 	ASSERT(log->l_grant_write_bytes >= log->l_grant_reserve_bytes);
3284     }
3285 }	/* xlog_verify_grant_head */
3286 
3287 /* check if it will fit */
3288 STATIC void
3289 xlog_verify_tail_lsn(xlog_t	    *log,
3290 		     xlog_in_core_t *iclog,
3291 		     xfs_lsn_t	    tail_lsn)
3292 {
3293     int blocks;
3294 
3295     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3296 	blocks =
3297 	    log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3298 	if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3299 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3300     } else {
3301 	ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3302 
3303 	if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3304 	    xlog_panic("xlog_verify_tail_lsn: tail wrapped");
3305 
3306 	blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3307 	if (blocks < BTOBB(iclog->ic_offset) + 1)
3308 	    xlog_panic("xlog_verify_tail_lsn: ran out of log space");
3309     }
3310 }	/* xlog_verify_tail_lsn */
3311 
3312 /*
3313  * Perform a number of checks on the iclog before writing to disk.
3314  *
3315  * 1. Make sure the iclogs are still circular
3316  * 2. Make sure we have a good magic number
3317  * 3. Make sure we don't have magic numbers in the data
3318  * 4. Check fields of each log operation header for:
3319  *	A. Valid client identifier
3320  *	B. tid ptr value falls in valid ptr space (user space code)
3321  *	C. Length in log record header is correct according to the
3322  *		individual operation headers within record.
3323  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3324  *	log, check the preceding blocks of the physical log to make sure all
3325  *	the cycle numbers agree with the current cycle number.
3326  */
3327 STATIC void
3328 xlog_verify_iclog(xlog_t	 *log,
3329 		  xlog_in_core_t *iclog,
3330 		  int		 count,
3331 		  boolean_t	 syncing)
3332 {
3333 	xlog_op_header_t	*ophead;
3334 	xlog_in_core_t		*icptr;
3335 	xlog_in_core_2_t	*xhdr;
3336 	xfs_caddr_t		ptr;
3337 	xfs_caddr_t		base_ptr;
3338 	__psint_t		field_offset;
3339 	__uint8_t		clientid;
3340 	int			len, i, j, k, op_len;
3341 	int			idx;
3342 
3343 	/* check validity of iclog pointers */
3344 	spin_lock(&log->l_icloglock);
3345 	icptr = log->l_iclog;
3346 	for (i=0; i < log->l_iclog_bufs; i++) {
3347 		if (icptr == NULL)
3348 			xlog_panic("xlog_verify_iclog: invalid ptr");
3349 		icptr = icptr->ic_next;
3350 	}
3351 	if (icptr != log->l_iclog)
3352 		xlog_panic("xlog_verify_iclog: corrupt iclog ring");
3353 	spin_unlock(&log->l_icloglock);
3354 
3355 	/* check log magic numbers */
3356 	if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM)
3357 		xlog_panic("xlog_verify_iclog: invalid magic num");
3358 
3359 	ptr = (xfs_caddr_t) &iclog->ic_header;
3360 	for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3361 	     ptr += BBSIZE) {
3362 		if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM)
3363 			xlog_panic("xlog_verify_iclog: unexpected magic num");
3364 	}
3365 
3366 	/* check fields */
3367 	len = be32_to_cpu(iclog->ic_header.h_num_logops);
3368 	ptr = iclog->ic_datap;
3369 	base_ptr = ptr;
3370 	ophead = (xlog_op_header_t *)ptr;
3371 	xhdr = iclog->ic_data;
3372 	for (i = 0; i < len; i++) {
3373 		ophead = (xlog_op_header_t *)ptr;
3374 
3375 		/* clientid is only 1 byte */
3376 		field_offset = (__psint_t)
3377 			       ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3378 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3379 			clientid = ophead->oh_clientid;
3380 		} else {
3381 			idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3382 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3383 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3384 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3385 				clientid = xlog_get_client_id(
3386 					xhdr[j].hic_xheader.xh_cycle_data[k]);
3387 			} else {
3388 				clientid = xlog_get_client_id(
3389 					iclog->ic_header.h_cycle_data[idx]);
3390 			}
3391 		}
3392 		if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3393 			cmn_err(CE_WARN, "xlog_verify_iclog: "
3394 				"invalid clientid %d op 0x%p offset 0x%lx",
3395 				clientid, ophead, (unsigned long)field_offset);
3396 
3397 		/* check length */
3398 		field_offset = (__psint_t)
3399 			       ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3400 		if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3401 			op_len = be32_to_cpu(ophead->oh_len);
3402 		} else {
3403 			idx = BTOBBT((__psint_t)&ophead->oh_len -
3404 				    (__psint_t)iclog->ic_datap);
3405 			if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3406 				j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3407 				k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3408 				op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3409 			} else {
3410 				op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3411 			}
3412 		}
3413 		ptr += sizeof(xlog_op_header_t) + op_len;
3414 	}
3415 }	/* xlog_verify_iclog */
3416 #endif
3417 
3418 /*
3419  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3420  */
3421 STATIC int
3422 xlog_state_ioerror(
3423 	xlog_t	*log)
3424 {
3425 	xlog_in_core_t	*iclog, *ic;
3426 
3427 	iclog = log->l_iclog;
3428 	if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3429 		/*
3430 		 * Mark all the incore logs IOERROR.
3431 		 * From now on, no log flushes will result.
3432 		 */
3433 		ic = iclog;
3434 		do {
3435 			ic->ic_state = XLOG_STATE_IOERROR;
3436 			ic = ic->ic_next;
3437 		} while (ic != iclog);
3438 		return 0;
3439 	}
3440 	/*
3441 	 * Return non-zero, if state transition has already happened.
3442 	 */
3443 	return 1;
3444 }
3445 
3446 /*
3447  * This is called from xfs_force_shutdown, when we're forcibly
3448  * shutting down the filesystem, typically because of an IO error.
3449  * Our main objectives here are to make sure that:
3450  *	a. the filesystem gets marked 'SHUTDOWN' for all interested
3451  *	   parties to find out, 'atomically'.
3452  *	b. those who're sleeping on log reservations, pinned objects and
3453  *	    other resources get woken up, and be told the bad news.
3454  *	c. nothing new gets queued up after (a) and (b) are done.
3455  *	d. if !logerror, flush the iclogs to disk, then seal them off
3456  *	   for business.
3457  */
3458 int
3459 xfs_log_force_umount(
3460 	struct xfs_mount	*mp,
3461 	int			logerror)
3462 {
3463 	xlog_ticket_t	*tic;
3464 	xlog_t		*log;
3465 	int		retval;
3466 	int		dummy;
3467 
3468 	log = mp->m_log;
3469 
3470 	/*
3471 	 * If this happens during log recovery, don't worry about
3472 	 * locking; the log isn't open for business yet.
3473 	 */
3474 	if (!log ||
3475 	    log->l_flags & XLOG_ACTIVE_RECOVERY) {
3476 		mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3477 		if (mp->m_sb_bp)
3478 			XFS_BUF_DONE(mp->m_sb_bp);
3479 		return 0;
3480 	}
3481 
3482 	/*
3483 	 * Somebody could've already done the hard work for us.
3484 	 * No need to get locks for this.
3485 	 */
3486 	if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3487 		ASSERT(XLOG_FORCED_SHUTDOWN(log));
3488 		return 1;
3489 	}
3490 	retval = 0;
3491 	/*
3492 	 * We must hold both the GRANT lock and the LOG lock,
3493 	 * before we mark the filesystem SHUTDOWN and wake
3494 	 * everybody up to tell the bad news.
3495 	 */
3496 	spin_lock(&log->l_icloglock);
3497 	spin_lock(&log->l_grant_lock);
3498 	mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3499 	if (mp->m_sb_bp)
3500 		XFS_BUF_DONE(mp->m_sb_bp);
3501 
3502 	/*
3503 	 * This flag is sort of redundant because of the mount flag, but
3504 	 * it's good to maintain the separation between the log and the rest
3505 	 * of XFS.
3506 	 */
3507 	log->l_flags |= XLOG_IO_ERROR;
3508 
3509 	/*
3510 	 * If we hit a log error, we want to mark all the iclogs IOERROR
3511 	 * while we're still holding the loglock.
3512 	 */
3513 	if (logerror)
3514 		retval = xlog_state_ioerror(log);
3515 	spin_unlock(&log->l_icloglock);
3516 
3517 	/*
3518 	 * We don't want anybody waiting for log reservations
3519 	 * after this. That means we have to wake up everybody
3520 	 * queued up on reserve_headq as well as write_headq.
3521 	 * In addition, we make sure in xlog_{re}grant_log_space
3522 	 * that we don't enqueue anything once the SHUTDOWN flag
3523 	 * is set, and this action is protected by the GRANTLOCK.
3524 	 */
3525 	if ((tic = log->l_reserve_headq)) {
3526 		do {
3527 			sv_signal(&tic->t_wait);
3528 			tic = tic->t_next;
3529 		} while (tic != log->l_reserve_headq);
3530 	}
3531 
3532 	if ((tic = log->l_write_headq)) {
3533 		do {
3534 			sv_signal(&tic->t_wait);
3535 			tic = tic->t_next;
3536 		} while (tic != log->l_write_headq);
3537 	}
3538 	spin_unlock(&log->l_grant_lock);
3539 
3540 	if (! (log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3541 		ASSERT(!logerror);
3542 		/*
3543 		 * Force the incore logs to disk before shutting the
3544 		 * log down completely.
3545 		 */
3546 		xlog_state_sync_all(log, XFS_LOG_FORCE|XFS_LOG_SYNC, &dummy);
3547 		spin_lock(&log->l_icloglock);
3548 		retval = xlog_state_ioerror(log);
3549 		spin_unlock(&log->l_icloglock);
3550 	}
3551 	/*
3552 	 * Wake up everybody waiting on xfs_log_force.
3553 	 * Callback all log item committed functions as if the
3554 	 * log writes were completed.
3555 	 */
3556 	xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3557 
3558 #ifdef XFSERRORDEBUG
3559 	{
3560 		xlog_in_core_t	*iclog;
3561 
3562 		spin_lock(&log->l_icloglock);
3563 		iclog = log->l_iclog;
3564 		do {
3565 			ASSERT(iclog->ic_callback == 0);
3566 			iclog = iclog->ic_next;
3567 		} while (iclog != log->l_iclog);
3568 		spin_unlock(&log->l_icloglock);
3569 	}
3570 #endif
3571 	/* return non-zero if log IOERROR transition had already happened */
3572 	return retval;
3573 }
3574 
3575 STATIC int
3576 xlog_iclogs_empty(xlog_t *log)
3577 {
3578 	xlog_in_core_t	*iclog;
3579 
3580 	iclog = log->l_iclog;
3581 	do {
3582 		/* endianness does not matter here, zero is zero in
3583 		 * any language.
3584 		 */
3585 		if (iclog->ic_header.h_num_logops)
3586 			return 0;
3587 		iclog = iclog->ic_next;
3588 	} while (iclog != log->l_iclog);
3589 	return 1;
3590 }
3591