c49e2c12dba4ec55600cc2f36a175526ae3aa503
[~shefty/rdma-dev.git] / fs / xfs / xfs_log.c
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_log.h"
22 #include "xfs_trans.h"
23 #include "xfs_sb.h"
24 #include "xfs_ag.h"
25 #include "xfs_mount.h"
26 #include "xfs_error.h"
27 #include "xfs_log_priv.h"
28 #include "xfs_buf_item.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_alloc_btree.h"
31 #include "xfs_ialloc_btree.h"
32 #include "xfs_log_recover.h"
33 #include "xfs_trans_priv.h"
34 #include "xfs_dinode.h"
35 #include "xfs_inode.h"
36 #include "xfs_trace.h"
37 #include "xfs_fsops.h"
38 #include "xfs_cksum.h"
39
40 kmem_zone_t     *xfs_log_ticket_zone;
41
42 /* Local miscellaneous function prototypes */
43 STATIC int
44 xlog_commit_record(
45         struct xlog             *log,
46         struct xlog_ticket      *ticket,
47         struct xlog_in_core     **iclog,
48         xfs_lsn_t               *commitlsnp);
49
50 STATIC struct xlog *
51 xlog_alloc_log(
52         struct xfs_mount        *mp,
53         struct xfs_buftarg      *log_target,
54         xfs_daddr_t             blk_offset,
55         int                     num_bblks);
56 STATIC int
57 xlog_space_left(
58         struct xlog             *log,
59         atomic64_t              *head);
60 STATIC int
61 xlog_sync(
62         struct xlog             *log,
63         struct xlog_in_core     *iclog);
64 STATIC void
65 xlog_dealloc_log(
66         struct xlog             *log);
67
68 /* local state machine functions */
69 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
70 STATIC void
71 xlog_state_do_callback(
72         struct xlog             *log,
73         int                     aborted,
74         struct xlog_in_core     *iclog);
75 STATIC int
76 xlog_state_get_iclog_space(
77         struct xlog             *log,
78         int                     len,
79         struct xlog_in_core     **iclog,
80         struct xlog_ticket      *ticket,
81         int                     *continued_write,
82         int                     *logoffsetp);
83 STATIC int
84 xlog_state_release_iclog(
85         struct xlog             *log,
86         struct xlog_in_core     *iclog);
87 STATIC void
88 xlog_state_switch_iclogs(
89         struct xlog             *log,
90         struct xlog_in_core     *iclog,
91         int                     eventual_size);
92 STATIC void
93 xlog_state_want_sync(
94         struct xlog             *log,
95         struct xlog_in_core     *iclog);
96
97 STATIC void
98 xlog_grant_push_ail(
99         struct xlog             *log,
100         int                     need_bytes);
101 STATIC void
102 xlog_regrant_reserve_log_space(
103         struct xlog             *log,
104         struct xlog_ticket      *ticket);
105 STATIC void
106 xlog_ungrant_log_space(
107         struct xlog             *log,
108         struct xlog_ticket      *ticket);
109
110 #if defined(DEBUG)
111 STATIC void
112 xlog_verify_dest_ptr(
113         struct xlog             *log,
114         char                    *ptr);
115 STATIC void
116 xlog_verify_grant_tail(
117         struct xlog *log);
118 STATIC void
119 xlog_verify_iclog(
120         struct xlog             *log,
121         struct xlog_in_core     *iclog,
122         int                     count,
123         boolean_t               syncing);
124 STATIC void
125 xlog_verify_tail_lsn(
126         struct xlog             *log,
127         struct xlog_in_core     *iclog,
128         xfs_lsn_t               tail_lsn);
129 #else
130 #define xlog_verify_dest_ptr(a,b)
131 #define xlog_verify_grant_tail(a)
132 #define xlog_verify_iclog(a,b,c,d)
133 #define xlog_verify_tail_lsn(a,b,c)
134 #endif
135
136 STATIC int
137 xlog_iclogs_empty(
138         struct xlog             *log);
139
140 static void
141 xlog_grant_sub_space(
142         struct xlog             *log,
143         atomic64_t              *head,
144         int                     bytes)
145 {
146         int64_t head_val = atomic64_read(head);
147         int64_t new, old;
148
149         do {
150                 int     cycle, space;
151
152                 xlog_crack_grant_head_val(head_val, &cycle, &space);
153
154                 space -= bytes;
155                 if (space < 0) {
156                         space += log->l_logsize;
157                         cycle--;
158                 }
159
160                 old = head_val;
161                 new = xlog_assign_grant_head_val(cycle, space);
162                 head_val = atomic64_cmpxchg(head, old, new);
163         } while (head_val != old);
164 }
165
166 static void
167 xlog_grant_add_space(
168         struct xlog             *log,
169         atomic64_t              *head,
170         int                     bytes)
171 {
172         int64_t head_val = atomic64_read(head);
173         int64_t new, old;
174
175         do {
176                 int             tmp;
177                 int             cycle, space;
178
179                 xlog_crack_grant_head_val(head_val, &cycle, &space);
180
181                 tmp = log->l_logsize - space;
182                 if (tmp > bytes)
183                         space += bytes;
184                 else {
185                         space = bytes - tmp;
186                         cycle++;
187                 }
188
189                 old = head_val;
190                 new = xlog_assign_grant_head_val(cycle, space);
191                 head_val = atomic64_cmpxchg(head, old, new);
192         } while (head_val != old);
193 }
194
195 STATIC void
196 xlog_grant_head_init(
197         struct xlog_grant_head  *head)
198 {
199         xlog_assign_grant_head(&head->grant, 1, 0);
200         INIT_LIST_HEAD(&head->waiters);
201         spin_lock_init(&head->lock);
202 }
203
204 STATIC void
205 xlog_grant_head_wake_all(
206         struct xlog_grant_head  *head)
207 {
208         struct xlog_ticket      *tic;
209
210         spin_lock(&head->lock);
211         list_for_each_entry(tic, &head->waiters, t_queue)
212                 wake_up_process(tic->t_task);
213         spin_unlock(&head->lock);
214 }
215
216 static inline int
217 xlog_ticket_reservation(
218         struct xlog             *log,
219         struct xlog_grant_head  *head,
220         struct xlog_ticket      *tic)
221 {
222         if (head == &log->l_write_head) {
223                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
224                 return tic->t_unit_res;
225         } else {
226                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
227                         return tic->t_unit_res * tic->t_cnt;
228                 else
229                         return tic->t_unit_res;
230         }
231 }
232
233 STATIC bool
234 xlog_grant_head_wake(
235         struct xlog             *log,
236         struct xlog_grant_head  *head,
237         int                     *free_bytes)
238 {
239         struct xlog_ticket      *tic;
240         int                     need_bytes;
241
242         list_for_each_entry(tic, &head->waiters, t_queue) {
243                 need_bytes = xlog_ticket_reservation(log, head, tic);
244                 if (*free_bytes < need_bytes)
245                         return false;
246
247                 *free_bytes -= need_bytes;
248                 trace_xfs_log_grant_wake_up(log, tic);
249                 wake_up_process(tic->t_task);
250         }
251
252         return true;
253 }
254
255 STATIC int
256 xlog_grant_head_wait(
257         struct xlog             *log,
258         struct xlog_grant_head  *head,
259         struct xlog_ticket      *tic,
260         int                     need_bytes)
261 {
262         list_add_tail(&tic->t_queue, &head->waiters);
263
264         do {
265                 if (XLOG_FORCED_SHUTDOWN(log))
266                         goto shutdown;
267                 xlog_grant_push_ail(log, need_bytes);
268
269                 __set_current_state(TASK_UNINTERRUPTIBLE);
270                 spin_unlock(&head->lock);
271
272                 XFS_STATS_INC(xs_sleep_logspace);
273
274                 trace_xfs_log_grant_sleep(log, tic);
275                 schedule();
276                 trace_xfs_log_grant_wake(log, tic);
277
278                 spin_lock(&head->lock);
279                 if (XLOG_FORCED_SHUTDOWN(log))
280                         goto shutdown;
281         } while (xlog_space_left(log, &head->grant) < need_bytes);
282
283         list_del_init(&tic->t_queue);
284         return 0;
285 shutdown:
286         list_del_init(&tic->t_queue);
287         return XFS_ERROR(EIO);
288 }
289
290 /*
291  * Atomically get the log space required for a log ticket.
292  *
293  * Once a ticket gets put onto head->waiters, it will only return after the
294  * needed reservation is satisfied.
295  *
296  * This function is structured so that it has a lock free fast path. This is
297  * necessary because every new transaction reservation will come through this
298  * path. Hence any lock will be globally hot if we take it unconditionally on
299  * every pass.
300  *
301  * As tickets are only ever moved on and off head->waiters under head->lock, we
302  * only need to take that lock if we are going to add the ticket to the queue
303  * and sleep. We can avoid taking the lock if the ticket was never added to
304  * head->waiters because the t_queue list head will be empty and we hold the
305  * only reference to it so it can safely be checked unlocked.
306  */
307 STATIC int
308 xlog_grant_head_check(
309         struct xlog             *log,
310         struct xlog_grant_head  *head,
311         struct xlog_ticket      *tic,
312         int                     *need_bytes)
313 {
314         int                     free_bytes;
315         int                     error = 0;
316
317         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
318
319         /*
320          * If there are other waiters on the queue then give them a chance at
321          * logspace before us.  Wake up the first waiters, if we do not wake
322          * up all the waiters then go to sleep waiting for more free space,
323          * otherwise try to get some space for this transaction.
324          */
325         *need_bytes = xlog_ticket_reservation(log, head, tic);
326         free_bytes = xlog_space_left(log, &head->grant);
327         if (!list_empty_careful(&head->waiters)) {
328                 spin_lock(&head->lock);
329                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
330                     free_bytes < *need_bytes) {
331                         error = xlog_grant_head_wait(log, head, tic,
332                                                      *need_bytes);
333                 }
334                 spin_unlock(&head->lock);
335         } else if (free_bytes < *need_bytes) {
336                 spin_lock(&head->lock);
337                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
338                 spin_unlock(&head->lock);
339         }
340
341         return error;
342 }
343
344 static void
345 xlog_tic_reset_res(xlog_ticket_t *tic)
346 {
347         tic->t_res_num = 0;
348         tic->t_res_arr_sum = 0;
349         tic->t_res_num_ophdrs = 0;
350 }
351
352 static void
353 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
354 {
355         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
356                 /* add to overflow and start again */
357                 tic->t_res_o_flow += tic->t_res_arr_sum;
358                 tic->t_res_num = 0;
359                 tic->t_res_arr_sum = 0;
360         }
361
362         tic->t_res_arr[tic->t_res_num].r_len = len;
363         tic->t_res_arr[tic->t_res_num].r_type = type;
364         tic->t_res_arr_sum += len;
365         tic->t_res_num++;
366 }
367
368 /*
369  * Replenish the byte reservation required by moving the grant write head.
370  */
371 int
372 xfs_log_regrant(
373         struct xfs_mount        *mp,
374         struct xlog_ticket      *tic)
375 {
376         struct xlog             *log = mp->m_log;
377         int                     need_bytes;
378         int                     error = 0;
379
380         if (XLOG_FORCED_SHUTDOWN(log))
381                 return XFS_ERROR(EIO);
382
383         XFS_STATS_INC(xs_try_logspace);
384
385         /*
386          * This is a new transaction on the ticket, so we need to change the
387          * transaction ID so that the next transaction has a different TID in
388          * the log. Just add one to the existing tid so that we can see chains
389          * of rolling transactions in the log easily.
390          */
391         tic->t_tid++;
392
393         xlog_grant_push_ail(log, tic->t_unit_res);
394
395         tic->t_curr_res = tic->t_unit_res;
396         xlog_tic_reset_res(tic);
397
398         if (tic->t_cnt > 0)
399                 return 0;
400
401         trace_xfs_log_regrant(log, tic);
402
403         error = xlog_grant_head_check(log, &log->l_write_head, tic,
404                                       &need_bytes);
405         if (error)
406                 goto out_error;
407
408         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
409         trace_xfs_log_regrant_exit(log, tic);
410         xlog_verify_grant_tail(log);
411         return 0;
412
413 out_error:
414         /*
415          * If we are failing, make sure the ticket doesn't have any current
416          * reservations.  We don't want to add this back when the ticket/
417          * transaction gets cancelled.
418          */
419         tic->t_curr_res = 0;
420         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
421         return error;
422 }
423
424 /*
425  * Reserve log space and return a ticket corresponding the reservation.
426  *
427  * Each reservation is going to reserve extra space for a log record header.
428  * When writes happen to the on-disk log, we don't subtract the length of the
429  * log record header from any reservation.  By wasting space in each
430  * reservation, we prevent over allocation problems.
431  */
432 int
433 xfs_log_reserve(
434         struct xfs_mount        *mp,
435         int                     unit_bytes,
436         int                     cnt,
437         struct xlog_ticket      **ticp,
438         __uint8_t               client,
439         bool                    permanent,
440         uint                    t_type)
441 {
442         struct xlog             *log = mp->m_log;
443         struct xlog_ticket      *tic;
444         int                     need_bytes;
445         int                     error = 0;
446
447         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
448
449         if (XLOG_FORCED_SHUTDOWN(log))
450                 return XFS_ERROR(EIO);
451
452         XFS_STATS_INC(xs_try_logspace);
453
454         ASSERT(*ticp == NULL);
455         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
456                                 KM_SLEEP | KM_MAYFAIL);
457         if (!tic)
458                 return XFS_ERROR(ENOMEM);
459
460         tic->t_trans_type = t_type;
461         *ticp = tic;
462
463         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
464                                             : tic->t_unit_res);
465
466         trace_xfs_log_reserve(log, tic);
467
468         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
469                                       &need_bytes);
470         if (error)
471                 goto out_error;
472
473         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
474         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
475         trace_xfs_log_reserve_exit(log, tic);
476         xlog_verify_grant_tail(log);
477         return 0;
478
479 out_error:
480         /*
481          * If we are failing, make sure the ticket doesn't have any current
482          * reservations.  We don't want to add this back when the ticket/
483          * transaction gets cancelled.
484          */
485         tic->t_curr_res = 0;
486         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
487         return error;
488 }
489
490
491 /*
492  * NOTES:
493  *
494  *      1. currblock field gets updated at startup and after in-core logs
495  *              marked as with WANT_SYNC.
496  */
497
498 /*
499  * This routine is called when a user of a log manager ticket is done with
500  * the reservation.  If the ticket was ever used, then a commit record for
501  * the associated transaction is written out as a log operation header with
502  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
503  * a given ticket.  If the ticket was one with a permanent reservation, then
504  * a few operations are done differently.  Permanent reservation tickets by
505  * default don't release the reservation.  They just commit the current
506  * transaction with the belief that the reservation is still needed.  A flag
507  * must be passed in before permanent reservations are actually released.
508  * When these type of tickets are not released, they need to be set into
509  * the inited state again.  By doing this, a start record will be written
510  * out when the next write occurs.
511  */
512 xfs_lsn_t
513 xfs_log_done(
514         struct xfs_mount        *mp,
515         struct xlog_ticket      *ticket,
516         struct xlog_in_core     **iclog,
517         uint                    flags)
518 {
519         struct xlog             *log = mp->m_log;
520         xfs_lsn_t               lsn = 0;
521
522         if (XLOG_FORCED_SHUTDOWN(log) ||
523             /*
524              * If nothing was ever written, don't write out commit record.
525              * If we get an error, just continue and give back the log ticket.
526              */
527             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
528              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
529                 lsn = (xfs_lsn_t) -1;
530                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
531                         flags |= XFS_LOG_REL_PERM_RESERV;
532                 }
533         }
534
535
536         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
537             (flags & XFS_LOG_REL_PERM_RESERV)) {
538                 trace_xfs_log_done_nonperm(log, ticket);
539
540                 /*
541                  * Release ticket if not permanent reservation or a specific
542                  * request has been made to release a permanent reservation.
543                  */
544                 xlog_ungrant_log_space(log, ticket);
545                 xfs_log_ticket_put(ticket);
546         } else {
547                 trace_xfs_log_done_perm(log, ticket);
548
549                 xlog_regrant_reserve_log_space(log, ticket);
550                 /* If this ticket was a permanent reservation and we aren't
551                  * trying to release it, reset the inited flags; so next time
552                  * we write, a start record will be written out.
553                  */
554                 ticket->t_flags |= XLOG_TIC_INITED;
555         }
556
557         return lsn;
558 }
559
560 /*
561  * Attaches a new iclog I/O completion callback routine during
562  * transaction commit.  If the log is in error state, a non-zero
563  * return code is handed back and the caller is responsible for
564  * executing the callback at an appropriate time.
565  */
566 int
567 xfs_log_notify(
568         struct xfs_mount        *mp,
569         struct xlog_in_core     *iclog,
570         xfs_log_callback_t      *cb)
571 {
572         int     abortflg;
573
574         spin_lock(&iclog->ic_callback_lock);
575         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
576         if (!abortflg) {
577                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
578                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
579                 cb->cb_next = NULL;
580                 *(iclog->ic_callback_tail) = cb;
581                 iclog->ic_callback_tail = &(cb->cb_next);
582         }
583         spin_unlock(&iclog->ic_callback_lock);
584         return abortflg;
585 }
586
587 int
588 xfs_log_release_iclog(
589         struct xfs_mount        *mp,
590         struct xlog_in_core     *iclog)
591 {
592         if (xlog_state_release_iclog(mp->m_log, iclog)) {
593                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
594                 return EIO;
595         }
596
597         return 0;
598 }
599
600 /*
601  * Mount a log filesystem
602  *
603  * mp           - ubiquitous xfs mount point structure
604  * log_target   - buftarg of on-disk log device
605  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
606  * num_bblocks  - Number of BBSIZE blocks in on-disk log
607  *
608  * Return error or zero.
609  */
610 int
611 xfs_log_mount(
612         xfs_mount_t     *mp,
613         xfs_buftarg_t   *log_target,
614         xfs_daddr_t     blk_offset,
615         int             num_bblks)
616 {
617         int             error;
618
619         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
620                 xfs_notice(mp, "Mounting Filesystem");
621         else {
622                 xfs_notice(mp,
623 "Mounting filesystem in no-recovery mode.  Filesystem will be inconsistent.");
624                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
625         }
626
627         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
628         if (IS_ERR(mp->m_log)) {
629                 error = -PTR_ERR(mp->m_log);
630                 goto out;
631         }
632
633         /*
634          * Initialize the AIL now we have a log.
635          */
636         error = xfs_trans_ail_init(mp);
637         if (error) {
638                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
639                 goto out_free_log;
640         }
641         mp->m_log->l_ailp = mp->m_ail;
642
643         /*
644          * skip log recovery on a norecovery mount.  pretend it all
645          * just worked.
646          */
647         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
648                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
649
650                 if (readonly)
651                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
652
653                 error = xlog_recover(mp->m_log);
654
655                 if (readonly)
656                         mp->m_flags |= XFS_MOUNT_RDONLY;
657                 if (error) {
658                         xfs_warn(mp, "log mount/recovery failed: error %d",
659                                 error);
660                         goto out_destroy_ail;
661                 }
662         }
663
664         /* Normal transactions can now occur */
665         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
666
667         /*
668          * Now the log has been fully initialised and we know were our
669          * space grant counters are, we can initialise the permanent ticket
670          * needed for delayed logging to work.
671          */
672         xlog_cil_init_post_recovery(mp->m_log);
673
674         return 0;
675
676 out_destroy_ail:
677         xfs_trans_ail_destroy(mp);
678 out_free_log:
679         xlog_dealloc_log(mp->m_log);
680 out:
681         return error;
682 }
683
684 /*
685  * Finish the recovery of the file system.  This is separate from the
686  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
687  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
688  * here.
689  *
690  * If we finish recovery successfully, start the background log work. If we are
691  * not doing recovery, then we have a RO filesystem and we don't need to start
692  * it.
693  */
694 int
695 xfs_log_mount_finish(xfs_mount_t *mp)
696 {
697         int     error = 0;
698
699         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
700                 error = xlog_recover_finish(mp->m_log);
701                 if (!error)
702                         xfs_log_work_queue(mp);
703         } else {
704                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
705         }
706
707
708         return error;
709 }
710
711 /*
712  * Final log writes as part of unmount.
713  *
714  * Mark the filesystem clean as unmount happens.  Note that during relocation
715  * this routine needs to be executed as part of source-bag while the
716  * deallocation must not be done until source-end.
717  */
718
719 /*
720  * Unmount record used to have a string "Unmount filesystem--" in the
721  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
722  * We just write the magic number now since that particular field isn't
723  * currently architecture converted and "nUmount" is a bit foo.
724  * As far as I know, there weren't any dependencies on the old behaviour.
725  */
726
727 int
728 xfs_log_unmount_write(xfs_mount_t *mp)
729 {
730         struct xlog      *log = mp->m_log;
731         xlog_in_core_t   *iclog;
732 #ifdef DEBUG
733         xlog_in_core_t   *first_iclog;
734 #endif
735         xlog_ticket_t   *tic = NULL;
736         xfs_lsn_t        lsn;
737         int              error;
738
739         /*
740          * Don't write out unmount record on read-only mounts.
741          * Or, if we are doing a forced umount (typically because of IO errors).
742          */
743         if (mp->m_flags & XFS_MOUNT_RDONLY)
744                 return 0;
745
746         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
747         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
748
749 #ifdef DEBUG
750         first_iclog = iclog = log->l_iclog;
751         do {
752                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
753                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
754                         ASSERT(iclog->ic_offset == 0);
755                 }
756                 iclog = iclog->ic_next;
757         } while (iclog != first_iclog);
758 #endif
759         if (! (XLOG_FORCED_SHUTDOWN(log))) {
760                 error = xfs_log_reserve(mp, 600, 1, &tic,
761                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
762                 if (!error) {
763                         /* the data section must be 32 bit size aligned */
764                         struct {
765                             __uint16_t magic;
766                             __uint16_t pad1;
767                             __uint32_t pad2; /* may as well make it 64 bits */
768                         } magic = {
769                                 .magic = XLOG_UNMOUNT_TYPE,
770                         };
771                         struct xfs_log_iovec reg = {
772                                 .i_addr = &magic,
773                                 .i_len = sizeof(magic),
774                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
775                         };
776                         struct xfs_log_vec vec = {
777                                 .lv_niovecs = 1,
778                                 .lv_iovecp = &reg,
779                         };
780
781                         /* remove inited flag, and account for space used */
782                         tic->t_flags = 0;
783                         tic->t_curr_res -= sizeof(magic);
784                         error = xlog_write(log, &vec, tic, &lsn,
785                                            NULL, XLOG_UNMOUNT_TRANS);
786                         /*
787                          * At this point, we're umounting anyway,
788                          * so there's no point in transitioning log state
789                          * to IOERROR. Just continue...
790                          */
791                 }
792
793                 if (error)
794                         xfs_alert(mp, "%s: unmount record failed", __func__);
795
796
797                 spin_lock(&log->l_icloglock);
798                 iclog = log->l_iclog;
799                 atomic_inc(&iclog->ic_refcnt);
800                 xlog_state_want_sync(log, iclog);
801                 spin_unlock(&log->l_icloglock);
802                 error = xlog_state_release_iclog(log, iclog);
803
804                 spin_lock(&log->l_icloglock);
805                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
806                       iclog->ic_state == XLOG_STATE_DIRTY)) {
807                         if (!XLOG_FORCED_SHUTDOWN(log)) {
808                                 xlog_wait(&iclog->ic_force_wait,
809                                                         &log->l_icloglock);
810                         } else {
811                                 spin_unlock(&log->l_icloglock);
812                         }
813                 } else {
814                         spin_unlock(&log->l_icloglock);
815                 }
816                 if (tic) {
817                         trace_xfs_log_umount_write(log, tic);
818                         xlog_ungrant_log_space(log, tic);
819                         xfs_log_ticket_put(tic);
820                 }
821         } else {
822                 /*
823                  * We're already in forced_shutdown mode, couldn't
824                  * even attempt to write out the unmount transaction.
825                  *
826                  * Go through the motions of sync'ing and releasing
827                  * the iclog, even though no I/O will actually happen,
828                  * we need to wait for other log I/Os that may already
829                  * be in progress.  Do this as a separate section of
830                  * code so we'll know if we ever get stuck here that
831                  * we're in this odd situation of trying to unmount
832                  * a file system that went into forced_shutdown as
833                  * the result of an unmount..
834                  */
835                 spin_lock(&log->l_icloglock);
836                 iclog = log->l_iclog;
837                 atomic_inc(&iclog->ic_refcnt);
838
839                 xlog_state_want_sync(log, iclog);
840                 spin_unlock(&log->l_icloglock);
841                 error =  xlog_state_release_iclog(log, iclog);
842
843                 spin_lock(&log->l_icloglock);
844
845                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
846                         || iclog->ic_state == XLOG_STATE_DIRTY
847                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
848
849                                 xlog_wait(&iclog->ic_force_wait,
850                                                         &log->l_icloglock);
851                 } else {
852                         spin_unlock(&log->l_icloglock);
853                 }
854         }
855
856         return error;
857 }       /* xfs_log_unmount_write */
858
859 /*
860  * Empty the log for unmount/freeze.
861  *
862  * To do this, we first need to shut down the background log work so it is not
863  * trying to cover the log as we clean up. We then need to unpin all objects in
864  * the log so we can then flush them out. Once they have completed their IO and
865  * run the callbacks removing themselves from the AIL, we can write the unmount
866  * record.
867  */
868 void
869 xfs_log_quiesce(
870         struct xfs_mount        *mp)
871 {
872         cancel_delayed_work_sync(&mp->m_log->l_work);
873         xfs_log_force(mp, XFS_LOG_SYNC);
874
875         /*
876          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
877          * will push it, xfs_wait_buftarg() will not wait for it. Further,
878          * xfs_buf_iowait() cannot be used because it was pushed with the
879          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
880          * the IO to complete.
881          */
882         xfs_ail_push_all_sync(mp->m_ail);
883         xfs_wait_buftarg(mp->m_ddev_targp);
884         xfs_buf_lock(mp->m_sb_bp);
885         xfs_buf_unlock(mp->m_sb_bp);
886
887         xfs_log_unmount_write(mp);
888 }
889
890 /*
891  * Shut down and release the AIL and Log.
892  *
893  * During unmount, we need to ensure we flush all the dirty metadata objects
894  * from the AIL so that the log is empty before we write the unmount record to
895  * the log. Once this is done, we can tear down the AIL and the log.
896  */
897 void
898 xfs_log_unmount(
899         struct xfs_mount        *mp)
900 {
901         xfs_log_quiesce(mp);
902
903         xfs_trans_ail_destroy(mp);
904         xlog_dealloc_log(mp->m_log);
905 }
906
907 void
908 xfs_log_item_init(
909         struct xfs_mount        *mp,
910         struct xfs_log_item     *item,
911         int                     type,
912         const struct xfs_item_ops *ops)
913 {
914         item->li_mountp = mp;
915         item->li_ailp = mp->m_ail;
916         item->li_type = type;
917         item->li_ops = ops;
918         item->li_lv = NULL;
919
920         INIT_LIST_HEAD(&item->li_ail);
921         INIT_LIST_HEAD(&item->li_cil);
922 }
923
924 /*
925  * Wake up processes waiting for log space after we have moved the log tail.
926  */
927 void
928 xfs_log_space_wake(
929         struct xfs_mount        *mp)
930 {
931         struct xlog             *log = mp->m_log;
932         int                     free_bytes;
933
934         if (XLOG_FORCED_SHUTDOWN(log))
935                 return;
936
937         if (!list_empty_careful(&log->l_write_head.waiters)) {
938                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
939
940                 spin_lock(&log->l_write_head.lock);
941                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
942                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
943                 spin_unlock(&log->l_write_head.lock);
944         }
945
946         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
947                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
948
949                 spin_lock(&log->l_reserve_head.lock);
950                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
951                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
952                 spin_unlock(&log->l_reserve_head.lock);
953         }
954 }
955
956 /*
957  * Determine if we have a transaction that has gone to disk
958  * that needs to be covered. To begin the transition to the idle state
959  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
960  * If we are then in a state where covering is needed, the caller is informed
961  * that dummy transactions are required to move the log into the idle state.
962  *
963  * Because this is called as part of the sync process, we should also indicate
964  * that dummy transactions should be issued in anything but the covered or
965  * idle states. This ensures that the log tail is accurately reflected in
966  * the log at the end of the sync, hence if a crash occurrs avoids replay
967  * of transactions where the metadata is already on disk.
968  */
969 int
970 xfs_log_need_covered(xfs_mount_t *mp)
971 {
972         int             needed = 0;
973         struct xlog     *log = mp->m_log;
974
975         if (!xfs_fs_writable(mp))
976                 return 0;
977
978         spin_lock(&log->l_icloglock);
979         switch (log->l_covered_state) {
980         case XLOG_STATE_COVER_DONE:
981         case XLOG_STATE_COVER_DONE2:
982         case XLOG_STATE_COVER_IDLE:
983                 break;
984         case XLOG_STATE_COVER_NEED:
985         case XLOG_STATE_COVER_NEED2:
986                 if (!xfs_ail_min_lsn(log->l_ailp) &&
987                     xlog_iclogs_empty(log)) {
988                         if (log->l_covered_state == XLOG_STATE_COVER_NEED)
989                                 log->l_covered_state = XLOG_STATE_COVER_DONE;
990                         else
991                                 log->l_covered_state = XLOG_STATE_COVER_DONE2;
992                 }
993                 /* FALLTHRU */
994         default:
995                 needed = 1;
996                 break;
997         }
998         spin_unlock(&log->l_icloglock);
999         return needed;
1000 }
1001
1002 /*
1003  * We may be holding the log iclog lock upon entering this routine.
1004  */
1005 xfs_lsn_t
1006 xlog_assign_tail_lsn_locked(
1007         struct xfs_mount        *mp)
1008 {
1009         struct xlog             *log = mp->m_log;
1010         struct xfs_log_item     *lip;
1011         xfs_lsn_t               tail_lsn;
1012
1013         assert_spin_locked(&mp->m_ail->xa_lock);
1014
1015         /*
1016          * To make sure we always have a valid LSN for the log tail we keep
1017          * track of the last LSN which was committed in log->l_last_sync_lsn,
1018          * and use that when the AIL was empty.
1019          */
1020         lip = xfs_ail_min(mp->m_ail);
1021         if (lip)
1022                 tail_lsn = lip->li_lsn;
1023         else
1024                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1025         atomic64_set(&log->l_tail_lsn, tail_lsn);
1026         return tail_lsn;
1027 }
1028
1029 xfs_lsn_t
1030 xlog_assign_tail_lsn(
1031         struct xfs_mount        *mp)
1032 {
1033         xfs_lsn_t               tail_lsn;
1034
1035         spin_lock(&mp->m_ail->xa_lock);
1036         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1037         spin_unlock(&mp->m_ail->xa_lock);
1038
1039         return tail_lsn;
1040 }
1041
1042 /*
1043  * Return the space in the log between the tail and the head.  The head
1044  * is passed in the cycle/bytes formal parms.  In the special case where
1045  * the reserve head has wrapped passed the tail, this calculation is no
1046  * longer valid.  In this case, just return 0 which means there is no space
1047  * in the log.  This works for all places where this function is called
1048  * with the reserve head.  Of course, if the write head were to ever
1049  * wrap the tail, we should blow up.  Rather than catch this case here,
1050  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1051  *
1052  * This code also handles the case where the reservation head is behind
1053  * the tail.  The details of this case are described below, but the end
1054  * result is that we return the size of the log as the amount of space left.
1055  */
1056 STATIC int
1057 xlog_space_left(
1058         struct xlog     *log,
1059         atomic64_t      *head)
1060 {
1061         int             free_bytes;
1062         int             tail_bytes;
1063         int             tail_cycle;
1064         int             head_cycle;
1065         int             head_bytes;
1066
1067         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1068         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1069         tail_bytes = BBTOB(tail_bytes);
1070         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1071                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1072         else if (tail_cycle + 1 < head_cycle)
1073                 return 0;
1074         else if (tail_cycle < head_cycle) {
1075                 ASSERT(tail_cycle == (head_cycle - 1));
1076                 free_bytes = tail_bytes - head_bytes;
1077         } else {
1078                 /*
1079                  * The reservation head is behind the tail.
1080                  * In this case we just want to return the size of the
1081                  * log as the amount of space left.
1082                  */
1083                 xfs_alert(log->l_mp,
1084                         "xlog_space_left: head behind tail\n"
1085                         "  tail_cycle = %d, tail_bytes = %d\n"
1086                         "  GH   cycle = %d, GH   bytes = %d",
1087                         tail_cycle, tail_bytes, head_cycle, head_bytes);
1088                 ASSERT(0);
1089                 free_bytes = log->l_logsize;
1090         }
1091         return free_bytes;
1092 }
1093
1094
1095 /*
1096  * Log function which is called when an io completes.
1097  *
1098  * The log manager needs its own routine, in order to control what
1099  * happens with the buffer after the write completes.
1100  */
1101 void
1102 xlog_iodone(xfs_buf_t *bp)
1103 {
1104         struct xlog_in_core     *iclog = bp->b_fspriv;
1105         struct xlog             *l = iclog->ic_log;
1106         int                     aborted = 0;
1107
1108         /*
1109          * Race to shutdown the filesystem if we see an error.
1110          */
1111         if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1112                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1113                 xfs_buf_ioerror_alert(bp, __func__);
1114                 xfs_buf_stale(bp);
1115                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1116                 /*
1117                  * This flag will be propagated to the trans-committed
1118                  * callback routines to let them know that the log-commit
1119                  * didn't succeed.
1120                  */
1121                 aborted = XFS_LI_ABORTED;
1122         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1123                 aborted = XFS_LI_ABORTED;
1124         }
1125
1126         /* log I/O is always issued ASYNC */
1127         ASSERT(XFS_BUF_ISASYNC(bp));
1128         xlog_state_done_syncing(iclog, aborted);
1129         /*
1130          * do not reference the buffer (bp) here as we could race
1131          * with it being freed after writing the unmount record to the
1132          * log.
1133          */
1134 }
1135
1136 /*
1137  * Return size of each in-core log record buffer.
1138  *
1139  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1140  *
1141  * If the filesystem blocksize is too large, we may need to choose a
1142  * larger size since the directory code currently logs entire blocks.
1143  */
1144
1145 STATIC void
1146 xlog_get_iclog_buffer_size(
1147         struct xfs_mount        *mp,
1148         struct xlog             *log)
1149 {
1150         int size;
1151         int xhdrs;
1152
1153         if (mp->m_logbufs <= 0)
1154                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1155         else
1156                 log->l_iclog_bufs = mp->m_logbufs;
1157
1158         /*
1159          * Buffer size passed in from mount system call.
1160          */
1161         if (mp->m_logbsize > 0) {
1162                 size = log->l_iclog_size = mp->m_logbsize;
1163                 log->l_iclog_size_log = 0;
1164                 while (size != 1) {
1165                         log->l_iclog_size_log++;
1166                         size >>= 1;
1167                 }
1168
1169                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1170                         /* # headers = size / 32k
1171                          * one header holds cycles from 32k of data
1172                          */
1173
1174                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1175                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1176                                 xhdrs++;
1177                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1178                         log->l_iclog_heads = xhdrs;
1179                 } else {
1180                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1181                         log->l_iclog_hsize = BBSIZE;
1182                         log->l_iclog_heads = 1;
1183                 }
1184                 goto done;
1185         }
1186
1187         /* All machines use 32kB buffers by default. */
1188         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1189         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1190
1191         /* the default log size is 16k or 32k which is one header sector */
1192         log->l_iclog_hsize = BBSIZE;
1193         log->l_iclog_heads = 1;
1194
1195 done:
1196         /* are we being asked to make the sizes selected above visible? */
1197         if (mp->m_logbufs == 0)
1198                 mp->m_logbufs = log->l_iclog_bufs;
1199         if (mp->m_logbsize == 0)
1200                 mp->m_logbsize = log->l_iclog_size;
1201 }       /* xlog_get_iclog_buffer_size */
1202
1203
1204 void
1205 xfs_log_work_queue(
1206         struct xfs_mount        *mp)
1207 {
1208         queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1209                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1210 }
1211
1212 /*
1213  * Every sync period we need to unpin all items in the AIL and push them to
1214  * disk. If there is nothing dirty, then we might need to cover the log to
1215  * indicate that the filesystem is idle.
1216  */
1217 void
1218 xfs_log_worker(
1219         struct work_struct      *work)
1220 {
1221         struct xlog             *log = container_of(to_delayed_work(work),
1222                                                 struct xlog, l_work);
1223         struct xfs_mount        *mp = log->l_mp;
1224
1225         /* dgc: errors ignored - not fatal and nowhere to report them */
1226         if (xfs_log_need_covered(mp))
1227                 xfs_fs_log_dummy(mp);
1228         else
1229                 xfs_log_force(mp, 0);
1230
1231         /* start pushing all the metadata that is currently dirty */
1232         xfs_ail_push_all(mp->m_ail);
1233
1234         /* queue us up again */
1235         xfs_log_work_queue(mp);
1236 }
1237
1238 /*
1239  * This routine initializes some of the log structure for a given mount point.
1240  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1241  * some other stuff may be filled in too.
1242  */
1243 STATIC struct xlog *
1244 xlog_alloc_log(
1245         struct xfs_mount        *mp,
1246         struct xfs_buftarg      *log_target,
1247         xfs_daddr_t             blk_offset,
1248         int                     num_bblks)
1249 {
1250         struct xlog             *log;
1251         xlog_rec_header_t       *head;
1252         xlog_in_core_t          **iclogp;
1253         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1254         xfs_buf_t               *bp;
1255         int                     i;
1256         int                     error = ENOMEM;
1257         uint                    log2_size = 0;
1258
1259         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1260         if (!log) {
1261                 xfs_warn(mp, "Log allocation failed: No memory!");
1262                 goto out;
1263         }
1264
1265         log->l_mp          = mp;
1266         log->l_targ        = log_target;
1267         log->l_logsize     = BBTOB(num_bblks);
1268         log->l_logBBstart  = blk_offset;
1269         log->l_logBBsize   = num_bblks;
1270         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1271         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1272         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1273
1274         log->l_prev_block  = -1;
1275         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1276         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1277         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1278         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1279
1280         xlog_grant_head_init(&log->l_reserve_head);
1281         xlog_grant_head_init(&log->l_write_head);
1282
1283         error = EFSCORRUPTED;
1284         if (xfs_sb_version_hassector(&mp->m_sb)) {
1285                 log2_size = mp->m_sb.sb_logsectlog;
1286                 if (log2_size < BBSHIFT) {
1287                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1288                                 log2_size, BBSHIFT);
1289                         goto out_free_log;
1290                 }
1291
1292                 log2_size -= BBSHIFT;
1293                 if (log2_size > mp->m_sectbb_log) {
1294                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1295                                 log2_size, mp->m_sectbb_log);
1296                         goto out_free_log;
1297                 }
1298
1299                 /* for larger sector sizes, must have v2 or external log */
1300                 if (log2_size && log->l_logBBstart > 0 &&
1301                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1302                         xfs_warn(mp,
1303                 "log sector size (0x%x) invalid for configuration.",
1304                                 log2_size);
1305                         goto out_free_log;
1306                 }
1307         }
1308         log->l_sectBBsize = 1 << log2_size;
1309
1310         xlog_get_iclog_buffer_size(mp, log);
1311
1312         error = ENOMEM;
1313         bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1314         if (!bp)
1315                 goto out_free_log;
1316         bp->b_iodone = xlog_iodone;
1317         ASSERT(xfs_buf_islocked(bp));
1318         log->l_xbuf = bp;
1319
1320         spin_lock_init(&log->l_icloglock);
1321         init_waitqueue_head(&log->l_flush_wait);
1322
1323         iclogp = &log->l_iclog;
1324         /*
1325          * The amount of memory to allocate for the iclog structure is
1326          * rather funky due to the way the structure is defined.  It is
1327          * done this way so that we can use different sizes for machines
1328          * with different amounts of memory.  See the definition of
1329          * xlog_in_core_t in xfs_log_priv.h for details.
1330          */
1331         ASSERT(log->l_iclog_size >= 4096);
1332         for (i=0; i < log->l_iclog_bufs; i++) {
1333                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1334                 if (!*iclogp)
1335                         goto out_free_iclog;
1336
1337                 iclog = *iclogp;
1338                 iclog->ic_prev = prev_iclog;
1339                 prev_iclog = iclog;
1340
1341                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1342                                                 BTOBB(log->l_iclog_size), 0);
1343                 if (!bp)
1344                         goto out_free_iclog;
1345
1346                 bp->b_iodone = xlog_iodone;
1347                 iclog->ic_bp = bp;
1348                 iclog->ic_data = bp->b_addr;
1349 #ifdef DEBUG
1350                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1351 #endif
1352                 head = &iclog->ic_header;
1353                 memset(head, 0, sizeof(xlog_rec_header_t));
1354                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1355                 head->h_version = cpu_to_be32(
1356                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1357                 head->h_size = cpu_to_be32(log->l_iclog_size);
1358                 /* new fields */
1359                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1360                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1361
1362                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1363                 iclog->ic_state = XLOG_STATE_ACTIVE;
1364                 iclog->ic_log = log;
1365                 atomic_set(&iclog->ic_refcnt, 0);
1366                 spin_lock_init(&iclog->ic_callback_lock);
1367                 iclog->ic_callback_tail = &(iclog->ic_callback);
1368                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1369
1370                 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1371                 init_waitqueue_head(&iclog->ic_force_wait);
1372                 init_waitqueue_head(&iclog->ic_write_wait);
1373
1374                 iclogp = &iclog->ic_next;
1375         }
1376         *iclogp = log->l_iclog;                 /* complete ring */
1377         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1378
1379         error = xlog_cil_init(log);
1380         if (error)
1381                 goto out_free_iclog;
1382         return log;
1383
1384 out_free_iclog:
1385         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1386                 prev_iclog = iclog->ic_next;
1387                 if (iclog->ic_bp)
1388                         xfs_buf_free(iclog->ic_bp);
1389                 kmem_free(iclog);
1390         }
1391         spinlock_destroy(&log->l_icloglock);
1392         xfs_buf_free(log->l_xbuf);
1393 out_free_log:
1394         kmem_free(log);
1395 out:
1396         return ERR_PTR(-error);
1397 }       /* xlog_alloc_log */
1398
1399
1400 /*
1401  * Write out the commit record of a transaction associated with the given
1402  * ticket.  Return the lsn of the commit record.
1403  */
1404 STATIC int
1405 xlog_commit_record(
1406         struct xlog             *log,
1407         struct xlog_ticket      *ticket,
1408         struct xlog_in_core     **iclog,
1409         xfs_lsn_t               *commitlsnp)
1410 {
1411         struct xfs_mount *mp = log->l_mp;
1412         int     error;
1413         struct xfs_log_iovec reg = {
1414                 .i_addr = NULL,
1415                 .i_len = 0,
1416                 .i_type = XLOG_REG_TYPE_COMMIT,
1417         };
1418         struct xfs_log_vec vec = {
1419                 .lv_niovecs = 1,
1420                 .lv_iovecp = &reg,
1421         };
1422
1423         ASSERT_ALWAYS(iclog);
1424         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1425                                         XLOG_COMMIT_TRANS);
1426         if (error)
1427                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1428         return error;
1429 }
1430
1431 /*
1432  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1433  * log space.  This code pushes on the lsn which would supposedly free up
1434  * the 25% which we want to leave free.  We may need to adopt a policy which
1435  * pushes on an lsn which is further along in the log once we reach the high
1436  * water mark.  In this manner, we would be creating a low water mark.
1437  */
1438 STATIC void
1439 xlog_grant_push_ail(
1440         struct xlog     *log,
1441         int             need_bytes)
1442 {
1443         xfs_lsn_t       threshold_lsn = 0;
1444         xfs_lsn_t       last_sync_lsn;
1445         int             free_blocks;
1446         int             free_bytes;
1447         int             threshold_block;
1448         int             threshold_cycle;
1449         int             free_threshold;
1450
1451         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1452
1453         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1454         free_blocks = BTOBBT(free_bytes);
1455
1456         /*
1457          * Set the threshold for the minimum number of free blocks in the
1458          * log to the maximum of what the caller needs, one quarter of the
1459          * log, and 256 blocks.
1460          */
1461         free_threshold = BTOBB(need_bytes);
1462         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1463         free_threshold = MAX(free_threshold, 256);
1464         if (free_blocks >= free_threshold)
1465                 return;
1466
1467         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1468                                                 &threshold_block);
1469         threshold_block += free_threshold;
1470         if (threshold_block >= log->l_logBBsize) {
1471                 threshold_block -= log->l_logBBsize;
1472                 threshold_cycle += 1;
1473         }
1474         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1475                                         threshold_block);
1476         /*
1477          * Don't pass in an lsn greater than the lsn of the last
1478          * log record known to be on disk. Use a snapshot of the last sync lsn
1479          * so that it doesn't change between the compare and the set.
1480          */
1481         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1482         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1483                 threshold_lsn = last_sync_lsn;
1484
1485         /*
1486          * Get the transaction layer to kick the dirty buffers out to
1487          * disk asynchronously. No point in trying to do this if
1488          * the filesystem is shutting down.
1489          */
1490         if (!XLOG_FORCED_SHUTDOWN(log))
1491                 xfs_ail_push(log->l_ailp, threshold_lsn);
1492 }
1493
1494 /*
1495  * Stamp cycle number in every block
1496  */
1497 STATIC void
1498 xlog_pack_data(
1499         struct xlog             *log,
1500         struct xlog_in_core     *iclog,
1501         int                     roundoff)
1502 {
1503         int                     i, j, k;
1504         int                     size = iclog->ic_offset + roundoff;
1505         __be32                  cycle_lsn;
1506         xfs_caddr_t             dp;
1507
1508         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1509
1510         dp = iclog->ic_datap;
1511         for (i = 0; i < BTOBB(size); i++) {
1512                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1513                         break;
1514                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1515                 *(__be32 *)dp = cycle_lsn;
1516                 dp += BBSIZE;
1517         }
1518
1519         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1520                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1521
1522                 for ( ; i < BTOBB(size); i++) {
1523                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1524                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1525                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1526                         *(__be32 *)dp = cycle_lsn;
1527                         dp += BBSIZE;
1528                 }
1529
1530                 for (i = 1; i < log->l_iclog_heads; i++)
1531                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1532         }
1533 }
1534
1535 /*
1536  * Calculate the checksum for a log buffer.
1537  *
1538  * This is a little more complicated than it should be because the various
1539  * headers and the actual data are non-contiguous.
1540  */
1541 __be32
1542 xlog_cksum(
1543         struct xlog             *log,
1544         struct xlog_rec_header  *rhead,
1545         char                    *dp,
1546         int                     size)
1547 {
1548         __uint32_t              crc;
1549
1550         /* first generate the crc for the record header ... */
1551         crc = xfs_start_cksum((char *)rhead,
1552                               sizeof(struct xlog_rec_header),
1553                               offsetof(struct xlog_rec_header, h_crc));
1554
1555         /* ... then for additional cycle data for v2 logs ... */
1556         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1557                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1558                 int             i;
1559
1560                 for (i = 1; i < log->l_iclog_heads; i++) {
1561                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1562                                      sizeof(struct xlog_rec_ext_header));
1563                 }
1564         }
1565
1566         /* ... and finally for the payload */
1567         crc = crc32c(crc, dp, size);
1568
1569         return xfs_end_cksum(crc);
1570 }
1571
1572 /*
1573  * The bdstrat callback function for log bufs. This gives us a central
1574  * place to trap bufs in case we get hit by a log I/O error and need to
1575  * shutdown. Actually, in practice, even when we didn't get a log error,
1576  * we transition the iclogs to IOERROR state *after* flushing all existing
1577  * iclogs to disk. This is because we don't want anymore new transactions to be
1578  * started or completed afterwards.
1579  */
1580 STATIC int
1581 xlog_bdstrat(
1582         struct xfs_buf          *bp)
1583 {
1584         struct xlog_in_core     *iclog = bp->b_fspriv;
1585
1586         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1587                 xfs_buf_ioerror(bp, EIO);
1588                 xfs_buf_stale(bp);
1589                 xfs_buf_ioend(bp, 0);
1590                 /*
1591                  * It would seem logical to return EIO here, but we rely on
1592                  * the log state machine to propagate I/O errors instead of
1593                  * doing it here.
1594                  */
1595                 return 0;
1596         }
1597
1598         xfs_buf_iorequest(bp);
1599         return 0;
1600 }
1601
1602 /*
1603  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1604  * fashion.  Previously, we should have moved the current iclog
1605  * ptr in the log to point to the next available iclog.  This allows further
1606  * write to continue while this code syncs out an iclog ready to go.
1607  * Before an in-core log can be written out, the data section must be scanned
1608  * to save away the 1st word of each BBSIZE block into the header.  We replace
1609  * it with the current cycle count.  Each BBSIZE block is tagged with the
1610  * cycle count because there in an implicit assumption that drives will
1611  * guarantee that entire 512 byte blocks get written at once.  In other words,
1612  * we can't have part of a 512 byte block written and part not written.  By
1613  * tagging each block, we will know which blocks are valid when recovering
1614  * after an unclean shutdown.
1615  *
1616  * This routine is single threaded on the iclog.  No other thread can be in
1617  * this routine with the same iclog.  Changing contents of iclog can there-
1618  * fore be done without grabbing the state machine lock.  Updating the global
1619  * log will require grabbing the lock though.
1620  *
1621  * The entire log manager uses a logical block numbering scheme.  Only
1622  * log_sync (and then only bwrite()) know about the fact that the log may
1623  * not start with block zero on a given device.  The log block start offset
1624  * is added immediately before calling bwrite().
1625  */
1626
1627 STATIC int
1628 xlog_sync(
1629         struct xlog             *log,
1630         struct xlog_in_core     *iclog)
1631 {
1632         xfs_buf_t       *bp;
1633         int             i;
1634         uint            count;          /* byte count of bwrite */
1635         uint            count_init;     /* initial count before roundup */
1636         int             roundoff;       /* roundoff to BB or stripe */
1637         int             split = 0;      /* split write into two regions */
1638         int             error;
1639         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1640         int             size;
1641
1642         XFS_STATS_INC(xs_log_writes);
1643         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1644
1645         /* Add for LR header */
1646         count_init = log->l_iclog_hsize + iclog->ic_offset;
1647
1648         /* Round out the log write size */
1649         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1650                 /* we have a v2 stripe unit to use */
1651                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1652         } else {
1653                 count = BBTOB(BTOBB(count_init));
1654         }
1655         roundoff = count - count_init;
1656         ASSERT(roundoff >= 0);
1657         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1658                 roundoff < log->l_mp->m_sb.sb_logsunit)
1659                 || 
1660                 (log->l_mp->m_sb.sb_logsunit <= 1 && 
1661                  roundoff < BBTOB(1)));
1662
1663         /* move grant heads by roundoff in sync */
1664         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1665         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1666
1667         /* put cycle number in every block */
1668         xlog_pack_data(log, iclog, roundoff); 
1669
1670         /* real byte length */
1671         size = iclog->ic_offset;
1672         if (v2)
1673                 size += roundoff;
1674         iclog->ic_header.h_len = cpu_to_be32(size);
1675
1676         bp = iclog->ic_bp;
1677         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1678
1679         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1680
1681         /* Do we need to split this write into 2 parts? */
1682         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1683                 char            *dptr;
1684
1685                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1686                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1687                 iclog->ic_bwritecnt = 2;
1688
1689                 /*
1690                  * Bump the cycle numbers at the start of each block in the
1691                  * part of the iclog that ends up in the buffer that gets
1692                  * written to the start of the log.
1693                  *
1694                  * Watch out for the header magic number case, though.
1695                  */
1696                 dptr = (char *)&iclog->ic_header + count;
1697                 for (i = 0; i < split; i += BBSIZE) {
1698                         __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1699                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1700                                 cycle++;
1701                         *(__be32 *)dptr = cpu_to_be32(cycle);
1702
1703                         dptr += BBSIZE;
1704                 }
1705         } else {
1706                 iclog->ic_bwritecnt = 1;
1707         }
1708
1709         /* calculcate the checksum */
1710         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1711                                             iclog->ic_datap, size);
1712
1713         bp->b_io_length = BTOBB(count);
1714         bp->b_fspriv = iclog;
1715         XFS_BUF_ZEROFLAGS(bp);
1716         XFS_BUF_ASYNC(bp);
1717         bp->b_flags |= XBF_SYNCIO;
1718
1719         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1720                 bp->b_flags |= XBF_FUA;
1721
1722                 /*
1723                  * Flush the data device before flushing the log to make
1724                  * sure all meta data written back from the AIL actually made
1725                  * it to disk before stamping the new log tail LSN into the
1726                  * log buffer.  For an external log we need to issue the
1727                  * flush explicitly, and unfortunately synchronously here;
1728                  * for an internal log we can simply use the block layer
1729                  * state machine for preflushes.
1730                  */
1731                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1732                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1733                 else
1734                         bp->b_flags |= XBF_FLUSH;
1735         }
1736
1737         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1738         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1739
1740         xlog_verify_iclog(log, iclog, count, B_TRUE);
1741
1742         /* account for log which doesn't start at block #0 */
1743         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1744         /*
1745          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1746          * is shutting down.
1747          */
1748         XFS_BUF_WRITE(bp);
1749
1750         error = xlog_bdstrat(bp);
1751         if (error) {
1752                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1753                 return error;
1754         }
1755         if (split) {
1756                 bp = iclog->ic_log->l_xbuf;
1757                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1758                 xfs_buf_associate_memory(bp,
1759                                 (char *)&iclog->ic_header + count, split);
1760                 bp->b_fspriv = iclog;
1761                 XFS_BUF_ZEROFLAGS(bp);
1762                 XFS_BUF_ASYNC(bp);
1763                 bp->b_flags |= XBF_SYNCIO;
1764                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1765                         bp->b_flags |= XBF_FUA;
1766
1767                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1768                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1769
1770                 /* account for internal log which doesn't start at block #0 */
1771                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1772                 XFS_BUF_WRITE(bp);
1773                 error = xlog_bdstrat(bp);
1774                 if (error) {
1775                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1776                         return error;
1777                 }
1778         }
1779         return 0;
1780 }       /* xlog_sync */
1781
1782 /*
1783  * Deallocate a log structure
1784  */
1785 STATIC void
1786 xlog_dealloc_log(
1787         struct xlog     *log)
1788 {
1789         xlog_in_core_t  *iclog, *next_iclog;
1790         int             i;
1791
1792         xlog_cil_destroy(log);
1793
1794         /*
1795          * always need to ensure that the extra buffer does not point to memory
1796          * owned by another log buffer before we free it.
1797          */
1798         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1799         xfs_buf_free(log->l_xbuf);
1800
1801         iclog = log->l_iclog;
1802         for (i=0; i<log->l_iclog_bufs; i++) {
1803                 xfs_buf_free(iclog->ic_bp);
1804                 next_iclog = iclog->ic_next;
1805                 kmem_free(iclog);
1806                 iclog = next_iclog;
1807         }
1808         spinlock_destroy(&log->l_icloglock);
1809
1810         log->l_mp->m_log = NULL;
1811         kmem_free(log);
1812 }       /* xlog_dealloc_log */
1813
1814 /*
1815  * Update counters atomically now that memcpy is done.
1816  */
1817 /* ARGSUSED */
1818 static inline void
1819 xlog_state_finish_copy(
1820         struct xlog             *log,
1821         struct xlog_in_core     *iclog,
1822         int                     record_cnt,
1823         int                     copy_bytes)
1824 {
1825         spin_lock(&log->l_icloglock);
1826
1827         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1828         iclog->ic_offset += copy_bytes;
1829
1830         spin_unlock(&log->l_icloglock);
1831 }       /* xlog_state_finish_copy */
1832
1833
1834
1835
1836 /*
1837  * print out info relating to regions written which consume
1838  * the reservation
1839  */
1840 void
1841 xlog_print_tic_res(
1842         struct xfs_mount        *mp,
1843         struct xlog_ticket      *ticket)
1844 {
1845         uint i;
1846         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1847
1848         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1849         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1850             "bformat",
1851             "bchunk",
1852             "efi_format",
1853             "efd_format",
1854             "iformat",
1855             "icore",
1856             "iext",
1857             "ibroot",
1858             "ilocal",
1859             "iattr_ext",
1860             "iattr_broot",
1861             "iattr_local",
1862             "qformat",
1863             "dquot",
1864             "quotaoff",
1865             "LR header",
1866             "unmount",
1867             "commit",
1868             "trans header"
1869         };
1870         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1871             "SETATTR_NOT_SIZE",
1872             "SETATTR_SIZE",
1873             "INACTIVE",
1874             "CREATE",
1875             "CREATE_TRUNC",
1876             "TRUNCATE_FILE",
1877             "REMOVE",
1878             "LINK",
1879             "RENAME",
1880             "MKDIR",
1881             "RMDIR",
1882             "SYMLINK",
1883             "SET_DMATTRS",
1884             "GROWFS",
1885             "STRAT_WRITE",
1886             "DIOSTRAT",
1887             "WRITE_SYNC",
1888             "WRITEID",
1889             "ADDAFORK",
1890             "ATTRINVAL",
1891             "ATRUNCATE",
1892             "ATTR_SET",
1893             "ATTR_RM",
1894             "ATTR_FLAG",
1895             "CLEAR_AGI_BUCKET",
1896             "QM_SBCHANGE",
1897             "DUMMY1",
1898             "DUMMY2",
1899             "QM_QUOTAOFF",
1900             "QM_DQALLOC",
1901             "QM_SETQLIM",
1902             "QM_DQCLUSTER",
1903             "QM_QINOCREATE",
1904             "QM_QUOTAOFF_END",
1905             "SB_UNIT",
1906             "FSYNC_TS",
1907             "GROWFSRT_ALLOC",
1908             "GROWFSRT_ZERO",
1909             "GROWFSRT_FREE",
1910             "SWAPEXT"
1911         };
1912
1913         xfs_warn(mp,
1914                 "xlog_write: reservation summary:\n"
1915                 "  trans type  = %s (%u)\n"
1916                 "  unit res    = %d bytes\n"
1917                 "  current res = %d bytes\n"
1918                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
1919                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
1920                 "  ophdr + reg = %u bytes\n"
1921                 "  num regions = %u\n",
1922                 ((ticket->t_trans_type <= 0 ||
1923                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1924                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1925                 ticket->t_trans_type,
1926                 ticket->t_unit_res,
1927                 ticket->t_curr_res,
1928                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1929                 ticket->t_res_num_ophdrs, ophdr_spc,
1930                 ticket->t_res_arr_sum +
1931                 ticket->t_res_o_flow + ophdr_spc,
1932                 ticket->t_res_num);
1933
1934         for (i = 0; i < ticket->t_res_num; i++) {
1935                 uint r_type = ticket->t_res_arr[i].r_type;
1936                 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1937                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1938                             "bad-rtype" : res_type_str[r_type-1]),
1939                             ticket->t_res_arr[i].r_len);
1940         }
1941
1942         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1943                 "xlog_write: reservation ran out. Need to up reservation");
1944         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1945 }
1946
1947 /*
1948  * Calculate the potential space needed by the log vector.  Each region gets
1949  * its own xlog_op_header_t and may need to be double word aligned.
1950  */
1951 static int
1952 xlog_write_calc_vec_length(
1953         struct xlog_ticket      *ticket,
1954         struct xfs_log_vec      *log_vector)
1955 {
1956         struct xfs_log_vec      *lv;
1957         int                     headers = 0;
1958         int                     len = 0;
1959         int                     i;
1960
1961         /* acct for start rec of xact */
1962         if (ticket->t_flags & XLOG_TIC_INITED)
1963                 headers++;
1964
1965         for (lv = log_vector; lv; lv = lv->lv_next) {
1966                 headers += lv->lv_niovecs;
1967
1968                 for (i = 0; i < lv->lv_niovecs; i++) {
1969                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
1970
1971                         len += vecp->i_len;
1972                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1973                 }
1974         }
1975
1976         ticket->t_res_num_ophdrs += headers;
1977         len += headers * sizeof(struct xlog_op_header);
1978
1979         return len;
1980 }
1981
1982 /*
1983  * If first write for transaction, insert start record  We can't be trying to
1984  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1985  */
1986 static int
1987 xlog_write_start_rec(
1988         struct xlog_op_header   *ophdr,
1989         struct xlog_ticket      *ticket)
1990 {
1991         if (!(ticket->t_flags & XLOG_TIC_INITED))
1992                 return 0;
1993
1994         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
1995         ophdr->oh_clientid = ticket->t_clientid;
1996         ophdr->oh_len = 0;
1997         ophdr->oh_flags = XLOG_START_TRANS;
1998         ophdr->oh_res2 = 0;
1999
2000         ticket->t_flags &= ~XLOG_TIC_INITED;
2001
2002         return sizeof(struct xlog_op_header);
2003 }
2004
2005 static xlog_op_header_t *
2006 xlog_write_setup_ophdr(
2007         struct xlog             *log,
2008         struct xlog_op_header   *ophdr,
2009         struct xlog_ticket      *ticket,
2010         uint                    flags)
2011 {
2012         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2013         ophdr->oh_clientid = ticket->t_clientid;
2014         ophdr->oh_res2 = 0;
2015
2016         /* are we copying a commit or unmount record? */
2017         ophdr->oh_flags = flags;
2018
2019         /*
2020          * We've seen logs corrupted with bad transaction client ids.  This
2021          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2022          * and shut down the filesystem.
2023          */
2024         switch (ophdr->oh_clientid)  {
2025         case XFS_TRANSACTION:
2026         case XFS_VOLUME:
2027         case XFS_LOG:
2028                 break;
2029         default:
2030                 xfs_warn(log->l_mp,
2031                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2032                         ophdr->oh_clientid, ticket);
2033                 return NULL;
2034         }
2035
2036         return ophdr;
2037 }
2038
2039 /*
2040  * Set up the parameters of the region copy into the log. This has
2041  * to handle region write split across multiple log buffers - this
2042  * state is kept external to this function so that this code can
2043  * can be written in an obvious, self documenting manner.
2044  */
2045 static int
2046 xlog_write_setup_copy(
2047         struct xlog_ticket      *ticket,
2048         struct xlog_op_header   *ophdr,
2049         int                     space_available,
2050         int                     space_required,
2051         int                     *copy_off,
2052         int                     *copy_len,
2053         int                     *last_was_partial_copy,
2054         int                     *bytes_consumed)
2055 {
2056         int                     still_to_copy;
2057
2058         still_to_copy = space_required - *bytes_consumed;
2059         *copy_off = *bytes_consumed;
2060
2061         if (still_to_copy <= space_available) {
2062                 /* write of region completes here */
2063                 *copy_len = still_to_copy;
2064                 ophdr->oh_len = cpu_to_be32(*copy_len);
2065                 if (*last_was_partial_copy)
2066                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2067                 *last_was_partial_copy = 0;
2068                 *bytes_consumed = 0;
2069                 return 0;
2070         }
2071
2072         /* partial write of region, needs extra log op header reservation */
2073         *copy_len = space_available;
2074         ophdr->oh_len = cpu_to_be32(*copy_len);
2075         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2076         if (*last_was_partial_copy)
2077                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2078         *bytes_consumed += *copy_len;
2079         (*last_was_partial_copy)++;
2080
2081         /* account for new log op header */
2082         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2083         ticket->t_res_num_ophdrs++;
2084
2085         return sizeof(struct xlog_op_header);
2086 }
2087
2088 static int
2089 xlog_write_copy_finish(
2090         struct xlog             *log,
2091         struct xlog_in_core     *iclog,
2092         uint                    flags,
2093         int                     *record_cnt,
2094         int                     *data_cnt,
2095         int                     *partial_copy,
2096         int                     *partial_copy_len,
2097         int                     log_offset,
2098         struct xlog_in_core     **commit_iclog)
2099 {
2100         if (*partial_copy) {
2101                 /*
2102                  * This iclog has already been marked WANT_SYNC by
2103                  * xlog_state_get_iclog_space.
2104                  */
2105                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2106                 *record_cnt = 0;
2107                 *data_cnt = 0;
2108                 return xlog_state_release_iclog(log, iclog);
2109         }
2110
2111         *partial_copy = 0;
2112         *partial_copy_len = 0;
2113
2114         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2115                 /* no more space in this iclog - push it. */
2116                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2117                 *record_cnt = 0;
2118                 *data_cnt = 0;
2119
2120                 spin_lock(&log->l_icloglock);
2121                 xlog_state_want_sync(log, iclog);
2122                 spin_unlock(&log->l_icloglock);
2123
2124                 if (!commit_iclog)
2125                         return xlog_state_release_iclog(log, iclog);
2126                 ASSERT(flags & XLOG_COMMIT_TRANS);
2127                 *commit_iclog = iclog;
2128         }
2129
2130         return 0;
2131 }
2132
2133 /*
2134  * Write some region out to in-core log
2135  *
2136  * This will be called when writing externally provided regions or when
2137  * writing out a commit record for a given transaction.
2138  *
2139  * General algorithm:
2140  *      1. Find total length of this write.  This may include adding to the
2141  *              lengths passed in.
2142  *      2. Check whether we violate the tickets reservation.
2143  *      3. While writing to this iclog
2144  *          A. Reserve as much space in this iclog as can get
2145  *          B. If this is first write, save away start lsn
2146  *          C. While writing this region:
2147  *              1. If first write of transaction, write start record
2148  *              2. Write log operation header (header per region)
2149  *              3. Find out if we can fit entire region into this iclog
2150  *              4. Potentially, verify destination memcpy ptr
2151  *              5. Memcpy (partial) region
2152  *              6. If partial copy, release iclog; otherwise, continue
2153  *                      copying more regions into current iclog
2154  *      4. Mark want sync bit (in simulation mode)
2155  *      5. Release iclog for potential flush to on-disk log.
2156  *
2157  * ERRORS:
2158  * 1.   Panic if reservation is overrun.  This should never happen since
2159  *      reservation amounts are generated internal to the filesystem.
2160  * NOTES:
2161  * 1. Tickets are single threaded data structures.
2162  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2163  *      syncing routine.  When a single log_write region needs to span
2164  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2165  *      on all log operation writes which don't contain the end of the
2166  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2167  *      operation which contains the end of the continued log_write region.
2168  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2169  *      we don't really know exactly how much space will be used.  As a result,
2170  *      we don't update ic_offset until the end when we know exactly how many
2171  *      bytes have been written out.
2172  */
2173 int
2174 xlog_write(
2175         struct xlog             *log,
2176         struct xfs_log_vec      *log_vector,
2177         struct xlog_ticket      *ticket,
2178         xfs_lsn_t               *start_lsn,
2179         struct xlog_in_core     **commit_iclog,
2180         uint                    flags)
2181 {
2182         struct xlog_in_core     *iclog = NULL;
2183         struct xfs_log_iovec    *vecp;
2184         struct xfs_log_vec      *lv;
2185         int                     len;
2186         int                     index;
2187         int                     partial_copy = 0;
2188         int                     partial_copy_len = 0;
2189         int                     contwr = 0;
2190         int                     record_cnt = 0;
2191         int                     data_cnt = 0;
2192         int                     error;
2193
2194         *start_lsn = 0;
2195
2196         len = xlog_write_calc_vec_length(ticket, log_vector);
2197
2198         /*
2199          * Region headers and bytes are already accounted for.
2200          * We only need to take into account start records and
2201          * split regions in this function.
2202          */
2203         if (ticket->t_flags & XLOG_TIC_INITED)
2204                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2205
2206         /*
2207          * Commit record headers need to be accounted for. These
2208          * come in as separate writes so are easy to detect.
2209          */
2210         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2211                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2212
2213         if (ticket->t_curr_res < 0)
2214                 xlog_print_tic_res(log->l_mp, ticket);
2215
2216         index = 0;
2217         lv = log_vector;
2218         vecp = lv->lv_iovecp;
2219         while (lv && index < lv->lv_niovecs) {
2220                 void            *ptr;
2221                 int             log_offset;
2222
2223                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2224                                                    &contwr, &log_offset);
2225                 if (error)
2226                         return error;
2227
2228                 ASSERT(log_offset <= iclog->ic_size - 1);
2229                 ptr = iclog->ic_datap + log_offset;
2230
2231                 /* start_lsn is the first lsn written to. That's all we need. */
2232                 if (!*start_lsn)
2233                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2234
2235                 /*
2236                  * This loop writes out as many regions as can fit in the amount
2237                  * of space which was allocated by xlog_state_get_iclog_space().
2238                  */
2239                 while (lv && index < lv->lv_niovecs) {
2240                         struct xfs_log_iovec    *reg = &vecp[index];
2241                         struct xlog_op_header   *ophdr;
2242                         int                     start_rec_copy;
2243                         int                     copy_len;
2244                         int                     copy_off;
2245
2246                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2247                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2248
2249                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2250                         if (start_rec_copy) {
2251                                 record_cnt++;
2252                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2253                                                    start_rec_copy);
2254                         }
2255
2256                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2257                         if (!ophdr)
2258                                 return XFS_ERROR(EIO);
2259
2260                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2261                                            sizeof(struct xlog_op_header));
2262
2263                         len += xlog_write_setup_copy(ticket, ophdr,
2264                                                      iclog->ic_size-log_offset,
2265                                                      reg->i_len,
2266                                                      &copy_off, &copy_len,
2267                                                      &partial_copy,
2268                                                      &partial_copy_len);
2269                         xlog_verify_dest_ptr(log, ptr);
2270
2271                         /* copy region */
2272                         ASSERT(copy_len >= 0);
2273                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2274                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2275
2276                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2277                         record_cnt++;
2278                         data_cnt += contwr ? copy_len : 0;
2279
2280                         error = xlog_write_copy_finish(log, iclog, flags,
2281                                                        &record_cnt, &data_cnt,
2282                                                        &partial_copy,
2283                                                        &partial_copy_len,
2284                                                        log_offset,
2285                                                        commit_iclog);
2286                         if (error)
2287                                 return error;
2288
2289                         /*
2290                          * if we had a partial copy, we need to get more iclog
2291                          * space but we don't want to increment the region
2292                          * index because there is still more is this region to
2293                          * write.
2294                          *
2295                          * If we completed writing this region, and we flushed
2296                          * the iclog (indicated by resetting of the record
2297                          * count), then we also need to get more log space. If
2298                          * this was the last record, though, we are done and
2299                          * can just return.
2300                          */
2301                         if (partial_copy)
2302                                 break;
2303
2304                         if (++index == lv->lv_niovecs) {
2305                                 lv = lv->lv_next;
2306                                 index = 0;
2307                                 if (lv)
2308                                         vecp = lv->lv_iovecp;
2309                         }
2310                         if (record_cnt == 0) {
2311                                 if (!lv)
2312                                         return 0;
2313                                 break;
2314                         }
2315                 }
2316         }
2317
2318         ASSERT(len == 0);
2319
2320         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2321         if (!commit_iclog)
2322                 return xlog_state_release_iclog(log, iclog);
2323
2324         ASSERT(flags & XLOG_COMMIT_TRANS);
2325         *commit_iclog = iclog;
2326         return 0;
2327 }
2328
2329
2330 /*****************************************************************************
2331  *
2332  *              State Machine functions
2333  *
2334  *****************************************************************************
2335  */
2336
2337 /* Clean iclogs starting from the head.  This ordering must be
2338  * maintained, so an iclog doesn't become ACTIVE beyond one that
2339  * is SYNCING.  This is also required to maintain the notion that we use
2340  * a ordered wait queue to hold off would be writers to the log when every
2341  * iclog is trying to sync to disk.
2342  *
2343  * State Change: DIRTY -> ACTIVE
2344  */
2345 STATIC void
2346 xlog_state_clean_log(
2347         struct xlog *log)
2348 {
2349         xlog_in_core_t  *iclog;
2350         int changed = 0;
2351
2352         iclog = log->l_iclog;
2353         do {
2354                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2355                         iclog->ic_state = XLOG_STATE_ACTIVE;
2356                         iclog->ic_offset       = 0;
2357                         ASSERT(iclog->ic_callback == NULL);
2358                         /*
2359                          * If the number of ops in this iclog indicate it just
2360                          * contains the dummy transaction, we can
2361                          * change state into IDLE (the second time around).
2362                          * Otherwise we should change the state into
2363                          * NEED a dummy.
2364                          * We don't need to cover the dummy.
2365                          */
2366                         if (!changed &&
2367                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2368                                         XLOG_COVER_OPS)) {
2369                                 changed = 1;
2370                         } else {
2371                                 /*
2372                                  * We have two dirty iclogs so start over
2373                                  * This could also be num of ops indicates
2374                                  * this is not the dummy going out.
2375                                  */
2376                                 changed = 2;
2377                         }
2378                         iclog->ic_header.h_num_logops = 0;
2379                         memset(iclog->ic_header.h_cycle_data, 0,
2380                               sizeof(iclog->ic_header.h_cycle_data));
2381                         iclog->ic_header.h_lsn = 0;
2382                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2383                         /* do nothing */;
2384                 else
2385                         break;  /* stop cleaning */
2386                 iclog = iclog->ic_next;
2387         } while (iclog != log->l_iclog);
2388
2389         /* log is locked when we are called */
2390         /*
2391          * Change state for the dummy log recording.
2392          * We usually go to NEED. But we go to NEED2 if the changed indicates
2393          * we are done writing the dummy record.
2394          * If we are done with the second dummy recored (DONE2), then
2395          * we go to IDLE.
2396          */
2397         if (changed) {
2398                 switch (log->l_covered_state) {
2399                 case XLOG_STATE_COVER_IDLE:
2400                 case XLOG_STATE_COVER_NEED:
2401                 case XLOG_STATE_COVER_NEED2:
2402                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2403                         break;
2404
2405                 case XLOG_STATE_COVER_DONE:
2406                         if (changed == 1)
2407                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2408                         else
2409                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2410                         break;
2411
2412                 case XLOG_STATE_COVER_DONE2:
2413                         if (changed == 1)
2414                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2415                         else
2416                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2417                         break;
2418
2419                 default:
2420                         ASSERT(0);
2421                 }
2422         }
2423 }       /* xlog_state_clean_log */
2424
2425 STATIC xfs_lsn_t
2426 xlog_get_lowest_lsn(
2427         struct xlog     *log)
2428 {
2429         xlog_in_core_t  *lsn_log;
2430         xfs_lsn_t       lowest_lsn, lsn;
2431
2432         lsn_log = log->l_iclog;
2433         lowest_lsn = 0;
2434         do {
2435             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2436                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2437                 if ((lsn && !lowest_lsn) ||
2438                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2439                         lowest_lsn = lsn;
2440                 }
2441             }
2442             lsn_log = lsn_log->ic_next;
2443         } while (lsn_log != log->l_iclog);
2444         return lowest_lsn;
2445 }
2446
2447
2448 STATIC void
2449 xlog_state_do_callback(
2450         struct xlog             *log,
2451         int                     aborted,
2452         struct xlog_in_core     *ciclog)
2453 {
2454         xlog_in_core_t     *iclog;
2455         xlog_in_core_t     *first_iclog;        /* used to know when we've
2456                                                  * processed all iclogs once */
2457         xfs_log_callback_t *cb, *cb_next;
2458         int                flushcnt = 0;
2459         xfs_lsn_t          lowest_lsn;
2460         int                ioerrors;    /* counter: iclogs with errors */
2461         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2462         int                funcdidcallbacks; /* flag: function did callbacks */
2463         int                repeats;     /* for issuing console warnings if
2464                                          * looping too many times */
2465         int                wake = 0;
2466
2467         spin_lock(&log->l_icloglock);
2468         first_iclog = iclog = log->l_iclog;
2469         ioerrors = 0;
2470         funcdidcallbacks = 0;
2471         repeats = 0;
2472
2473         do {
2474                 /*
2475                  * Scan all iclogs starting with the one pointed to by the
2476                  * log.  Reset this starting point each time the log is
2477                  * unlocked (during callbacks).
2478                  *
2479                  * Keep looping through iclogs until one full pass is made
2480                  * without running any callbacks.
2481                  */
2482                 first_iclog = log->l_iclog;
2483                 iclog = log->l_iclog;
2484                 loopdidcallbacks = 0;
2485                 repeats++;
2486
2487                 do {
2488
2489                         /* skip all iclogs in the ACTIVE & DIRTY states */
2490                         if (iclog->ic_state &
2491                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2492                                 iclog = iclog->ic_next;
2493                                 continue;
2494                         }
2495
2496                         /*
2497                          * Between marking a filesystem SHUTDOWN and stopping
2498                          * the log, we do flush all iclogs to disk (if there
2499                          * wasn't a log I/O error). So, we do want things to
2500                          * go smoothly in case of just a SHUTDOWN  w/o a
2501                          * LOG_IO_ERROR.
2502                          */
2503                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2504                                 /*
2505                                  * Can only perform callbacks in order.  Since
2506                                  * this iclog is not in the DONE_SYNC/
2507                                  * DO_CALLBACK state, we skip the rest and
2508                                  * just try to clean up.  If we set our iclog
2509                                  * to DO_CALLBACK, we will not process it when
2510                                  * we retry since a previous iclog is in the
2511                                  * CALLBACK and the state cannot change since
2512                                  * we are holding the l_icloglock.
2513                                  */
2514                                 if (!(iclog->ic_state &
2515                                         (XLOG_STATE_DONE_SYNC |
2516                                                  XLOG_STATE_DO_CALLBACK))) {
2517                                         if (ciclog && (ciclog->ic_state ==
2518                                                         XLOG_STATE_DONE_SYNC)) {
2519                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2520                                         }
2521                                         break;
2522                                 }
2523                                 /*
2524                                  * We now have an iclog that is in either the
2525                                  * DO_CALLBACK or DONE_SYNC states. The other
2526                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2527                                  * caught by the above if and are going to
2528                                  * clean (i.e. we aren't doing their callbacks)
2529                                  * see the above if.
2530                                  */
2531
2532                                 /*
2533                                  * We will do one more check here to see if we
2534                                  * have chased our tail around.
2535                                  */
2536
2537                                 lowest_lsn = xlog_get_lowest_lsn(log);
2538                                 if (lowest_lsn &&
2539                                     XFS_LSN_CMP(lowest_lsn,
2540                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2541                                         iclog = iclog->ic_next;
2542                                         continue; /* Leave this iclog for
2543                                                    * another thread */
2544                                 }
2545
2546                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2547
2548
2549                                 /*
2550                                  * Completion of a iclog IO does not imply that
2551                                  * a transaction has completed, as transactions
2552                                  * can be large enough to span many iclogs. We
2553                                  * cannot change the tail of the log half way
2554                                  * through a transaction as this may be the only
2555                                  * transaction in the log and moving th etail to
2556                                  * point to the middle of it will prevent
2557                                  * recovery from finding the start of the
2558                                  * transaction. Hence we should only update the
2559                                  * last_sync_lsn if this iclog contains
2560                                  * transaction completion callbacks on it.
2561                                  *
2562                                  * We have to do this before we drop the
2563                                  * icloglock to ensure we are the only one that
2564                                  * can update it.
2565                                  */
2566                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2567                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2568                                 if (iclog->ic_callback)
2569                                         atomic64_set(&log->l_last_sync_lsn,
2570                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2571
2572                         } else
2573                                 ioerrors++;
2574
2575                         spin_unlock(&log->l_icloglock);
2576
2577                         /*
2578                          * Keep processing entries in the callback list until
2579                          * we come around and it is empty.  We need to
2580                          * atomically see that the list is empty and change the
2581                          * state to DIRTY so that we don't miss any more
2582                          * callbacks being added.
2583                          */
2584                         spin_lock(&iclog->ic_callback_lock);
2585                         cb = iclog->ic_callback;
2586                         while (cb) {
2587                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2588                                 iclog->ic_callback = NULL;
2589                                 spin_unlock(&iclog->ic_callback_lock);
2590
2591                                 /* perform callbacks in the order given */
2592                                 for (; cb; cb = cb_next) {
2593                                         cb_next = cb->cb_next;
2594                                         cb->cb_func(cb->cb_arg, aborted);
2595                                 }
2596                                 spin_lock(&iclog->ic_callback_lock);
2597                                 cb = iclog->ic_callback;
2598                         }
2599
2600                         loopdidcallbacks++;
2601                         funcdidcallbacks++;
2602
2603                         spin_lock(&log->l_icloglock);
2604                         ASSERT(iclog->ic_callback == NULL);
2605                         spin_unlock(&iclog->ic_callback_lock);
2606                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2607                                 iclog->ic_state = XLOG_STATE_DIRTY;
2608
2609                         /*
2610                          * Transition from DIRTY to ACTIVE if applicable.
2611                          * NOP if STATE_IOERROR.
2612                          */
2613                         xlog_state_clean_log(log);
2614
2615                         /* wake up threads waiting in xfs_log_force() */
2616                         wake_up_all(&iclog->ic_force_wait);
2617
2618                         iclog = iclog->ic_next;
2619                 } while (first_iclog != iclog);
2620
2621                 if (repeats > 5000) {
2622                         flushcnt += repeats;
2623                         repeats = 0;
2624                         xfs_warn(log->l_mp,
2625                                 "%s: possible infinite loop (%d iterations)",
2626                                 __func__, flushcnt);
2627                 }
2628         } while (!ioerrors && loopdidcallbacks);
2629
2630         /*
2631          * make one last gasp attempt to see if iclogs are being left in
2632          * limbo..
2633          */
2634 #ifdef DEBUG
2635         if (funcdidcallbacks) {
2636                 first_iclog = iclog = log->l_iclog;
2637                 do {
2638                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2639                         /*
2640                          * Terminate the loop if iclogs are found in states
2641                          * which will cause other threads to clean up iclogs.
2642                          *
2643                          * SYNCING - i/o completion will go through logs
2644                          * DONE_SYNC - interrupt thread should be waiting for
2645                          *              l_icloglock
2646                          * IOERROR - give up hope all ye who enter here
2647                          */
2648                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2649                             iclog->ic_state == XLOG_STATE_SYNCING ||
2650                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2651                             iclog->ic_state == XLOG_STATE_IOERROR )
2652                                 break;
2653                         iclog = iclog->ic_next;
2654                 } while (first_iclog != iclog);
2655         }
2656 #endif
2657
2658         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2659                 wake = 1;
2660         spin_unlock(&log->l_icloglock);
2661
2662         if (wake)
2663                 wake_up_all(&log->l_flush_wait);
2664 }
2665
2666
2667 /*
2668  * Finish transitioning this iclog to the dirty state.
2669  *
2670  * Make sure that we completely execute this routine only when this is
2671  * the last call to the iclog.  There is a good chance that iclog flushes,
2672  * when we reach the end of the physical log, get turned into 2 separate
2673  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2674  * routine.  By using the reference count bwritecnt, we guarantee that only
2675  * the second completion goes through.
2676  *
2677  * Callbacks could take time, so they are done outside the scope of the
2678  * global state machine log lock.
2679  */
2680 STATIC void
2681 xlog_state_done_syncing(
2682         xlog_in_core_t  *iclog,
2683         int             aborted)
2684 {
2685         struct xlog        *log = iclog->ic_log;
2686
2687         spin_lock(&log->l_icloglock);
2688
2689         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2690                iclog->ic_state == XLOG_STATE_IOERROR);
2691         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2692         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2693
2694
2695         /*
2696          * If we got an error, either on the first buffer, or in the case of
2697          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2698          * and none should ever be attempted to be written to disk
2699          * again.
2700          */
2701         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2702                 if (--iclog->ic_bwritecnt == 1) {
2703                         spin_unlock(&log->l_icloglock);
2704                         return;
2705                 }
2706                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2707         }
2708
2709         /*
2710          * Someone could be sleeping prior to writing out the next
2711          * iclog buffer, we wake them all, one will get to do the
2712          * I/O, the others get to wait for the result.
2713          */
2714         wake_up_all(&iclog->ic_write_wait);
2715         spin_unlock(&log->l_icloglock);
2716         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2717 }       /* xlog_state_done_syncing */
2718
2719
2720 /*
2721  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2722  * sleep.  We wait on the flush queue on the head iclog as that should be
2723  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2724  * we will wait here and all new writes will sleep until a sync completes.
2725  *
2726  * The in-core logs are used in a circular fashion. They are not used
2727  * out-of-order even when an iclog past the head is free.
2728  *
2729  * return:
2730  *      * log_offset where xlog_write() can start writing into the in-core
2731  *              log's data space.
2732  *      * in-core log pointer to which xlog_write() should write.
2733  *      * boolean indicating this is a continued write to an in-core log.
2734  *              If this is the last write, then the in-core log's offset field
2735  *              needs to be incremented, depending on the amount of data which
2736  *              is copied.
2737  */
2738 STATIC int
2739 xlog_state_get_iclog_space(
2740         struct xlog             *log,
2741         int                     len,
2742         struct xlog_in_core     **iclogp,
2743         struct xlog_ticket      *ticket,
2744         int                     *continued_write,
2745         int                     *logoffsetp)
2746 {
2747         int               log_offset;
2748         xlog_rec_header_t *head;
2749         xlog_in_core_t    *iclog;
2750         int               error;
2751
2752 restart:
2753         spin_lock(&log->l_icloglock);
2754         if (XLOG_FORCED_SHUTDOWN(log)) {
2755                 spin_unlock(&log->l_icloglock);
2756                 return XFS_ERROR(EIO);
2757         }
2758
2759         iclog = log->l_iclog;
2760         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2761                 XFS_STATS_INC(xs_log_noiclogs);
2762
2763                 /* Wait for log writes to have flushed */
2764                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2765                 goto restart;
2766         }
2767
2768         head = &iclog->ic_header;
2769
2770         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2771         log_offset = iclog->ic_offset;
2772
2773         /* On the 1st write to an iclog, figure out lsn.  This works
2774          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2775          * committing to.  If the offset is set, that's how many blocks
2776          * must be written.
2777          */
2778         if (log_offset == 0) {
2779                 ticket->t_curr_res -= log->l_iclog_hsize;
2780                 xlog_tic_add_region(ticket,
2781                                     log->l_iclog_hsize,
2782                                     XLOG_REG_TYPE_LRHEADER);
2783                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2784                 head->h_lsn = cpu_to_be64(
2785                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2786                 ASSERT(log->l_curr_block >= 0);
2787         }
2788
2789         /* If there is enough room to write everything, then do it.  Otherwise,
2790          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2791          * bit is on, so this will get flushed out.  Don't update ic_offset
2792          * until you know exactly how many bytes get copied.  Therefore, wait
2793          * until later to update ic_offset.
2794          *
2795          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2796          * can fit into remaining data section.
2797          */
2798         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2799                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2800
2801                 /*
2802                  * If I'm the only one writing to this iclog, sync it to disk.
2803                  * We need to do an atomic compare and decrement here to avoid
2804                  * racing with concurrent atomic_dec_and_lock() calls in
2805                  * xlog_state_release_iclog() when there is more than one
2806                  * reference to the iclog.
2807                  */
2808                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2809                         /* we are the only one */
2810                         spin_unlock(&log->l_icloglock);
2811                         error = xlog_state_release_iclog(log, iclog);
2812                         if (error)
2813                                 return error;
2814                 } else {
2815                         spin_unlock(&log->l_icloglock);
2816                 }
2817                 goto restart;
2818         }
2819
2820         /* Do we have enough room to write the full amount in the remainder
2821          * of this iclog?  Or must we continue a write on the next iclog and
2822          * mark this iclog as completely taken?  In the case where we switch
2823          * iclogs (to mark it taken), this particular iclog will release/sync
2824          * to disk in xlog_write().
2825          */
2826         if (len <= iclog->ic_size - iclog->ic_offset) {
2827                 *continued_write = 0;
2828                 iclog->ic_offset += len;
2829         } else {
2830                 *continued_write = 1;
2831                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2832         }
2833         *iclogp = iclog;
2834
2835         ASSERT(iclog->ic_offset <= iclog->ic_size);
2836         spin_unlock(&log->l_icloglock);
2837
2838         *logoffsetp = log_offset;
2839         return 0;
2840 }       /* xlog_state_get_iclog_space */
2841
2842 /* The first cnt-1 times through here we don't need to
2843  * move the grant write head because the permanent
2844  * reservation has reserved cnt times the unit amount.
2845  * Release part of current permanent unit reservation and
2846  * reset current reservation to be one units worth.  Also
2847  * move grant reservation head forward.
2848  */
2849 STATIC void
2850 xlog_regrant_reserve_log_space(
2851         struct xlog             *log,
2852         struct xlog_ticket      *ticket)
2853 {
2854         trace_xfs_log_regrant_reserve_enter(log, ticket);
2855
2856         if (ticket->t_cnt > 0)
2857                 ticket->t_cnt--;
2858
2859         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2860                                         ticket->t_curr_res);
2861         xlog_grant_sub_space(log, &log->l_write_head.grant,
2862                                         ticket->t_curr_res);
2863         ticket->t_curr_res = ticket->t_unit_res;
2864         xlog_tic_reset_res(ticket);
2865
2866         trace_xfs_log_regrant_reserve_sub(log, ticket);
2867
2868         /* just return if we still have some of the pre-reserved space */
2869         if (ticket->t_cnt > 0)
2870                 return;
2871
2872         xlog_grant_add_space(log, &log->l_reserve_head.grant,
2873                                         ticket->t_unit_res);
2874
2875         trace_xfs_log_regrant_reserve_exit(log, ticket);
2876
2877         ticket->t_curr_res = ticket->t_unit_res;
2878         xlog_tic_reset_res(ticket);
2879 }       /* xlog_regrant_reserve_log_space */
2880
2881
2882 /*
2883  * Give back the space left from a reservation.
2884  *
2885  * All the information we need to make a correct determination of space left
2886  * is present.  For non-permanent reservations, things are quite easy.  The
2887  * count should have been decremented to zero.  We only need to deal with the
2888  * space remaining in the current reservation part of the ticket.  If the
2889  * ticket contains a permanent reservation, there may be left over space which
2890  * needs to be released.  A count of N means that N-1 refills of the current
2891  * reservation can be done before we need to ask for more space.  The first
2892  * one goes to fill up the first current reservation.  Once we run out of
2893  * space, the count will stay at zero and the only space remaining will be
2894  * in the current reservation field.
2895  */
2896 STATIC void
2897 xlog_ungrant_log_space(
2898         struct xlog             *log,
2899         struct xlog_ticket      *ticket)
2900 {
2901         int     bytes;
2902
2903         if (ticket->t_cnt > 0)
2904                 ticket->t_cnt--;
2905
2906         trace_xfs_log_ungrant_enter(log, ticket);
2907         trace_xfs_log_ungrant_sub(log, ticket);
2908
2909         /*
2910          * If this is a permanent reservation ticket, we may be able to free
2911          * up more space based on the remaining count.
2912          */
2913         bytes = ticket->t_curr_res;
2914         if (ticket->t_cnt > 0) {
2915                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2916                 bytes += ticket->t_unit_res*ticket->t_cnt;
2917         }
2918
2919         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2920         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2921
2922         trace_xfs_log_ungrant_exit(log, ticket);
2923
2924         xfs_log_space_wake(log->l_mp);
2925 }
2926
2927 /*
2928  * Flush iclog to disk if this is the last reference to the given iclog and
2929  * the WANT_SYNC bit is set.
2930  *
2931  * When this function is entered, the iclog is not necessarily in the
2932  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2933  *
2934  *
2935  */
2936 STATIC int
2937 xlog_state_release_iclog(
2938         struct xlog             *log,
2939         struct xlog_in_core     *iclog)
2940 {
2941         int             sync = 0;       /* do we sync? */
2942
2943         if (iclog->ic_state & XLOG_STATE_IOERROR)
2944                 return XFS_ERROR(EIO);
2945
2946         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2947         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2948                 return 0;
2949
2950         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2951                 spin_unlock(&log->l_icloglock);
2952                 return XFS_ERROR(EIO);
2953         }
2954         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2955                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2956
2957         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2958                 /* update tail before writing to iclog */
2959                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2960                 sync++;
2961                 iclog->ic_state = XLOG_STATE_SYNCING;
2962                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2963                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2964                 /* cycle incremented when incrementing curr_block */
2965         }
2966         spin_unlock(&log->l_icloglock);
2967
2968         /*
2969          * We let the log lock go, so it's possible that we hit a log I/O
2970          * error or some other SHUTDOWN condition that marks the iclog
2971          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2972          * this iclog has consistent data, so we ignore IOERROR
2973          * flags after this point.
2974          */
2975         if (sync)
2976                 return xlog_sync(log, iclog);
2977         return 0;
2978 }       /* xlog_state_release_iclog */
2979
2980
2981 /*
2982  * This routine will mark the current iclog in the ring as WANT_SYNC
2983  * and move the current iclog pointer to the next iclog in the ring.
2984  * When this routine is called from xlog_state_get_iclog_space(), the
2985  * exact size of the iclog has not yet been determined.  All we know is
2986  * that every data block.  We have run out of space in this log record.
2987  */
2988 STATIC void
2989 xlog_state_switch_iclogs(
2990         struct xlog             *log,
2991         struct xlog_in_core     *iclog,
2992         int                     eventual_size)
2993 {
2994         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2995         if (!eventual_size)
2996                 eventual_size = iclog->ic_offset;
2997         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2998         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2999         log->l_prev_block = log->l_curr_block;
3000         log->l_prev_cycle = log->l_curr_cycle;
3001
3002         /* roll log?: ic_offset changed later */
3003         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3004
3005         /* Round up to next log-sunit */
3006         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3007             log->l_mp->m_sb.sb_logsunit > 1) {
3008                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3009                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3010         }
3011
3012         if (log->l_curr_block >= log->l_logBBsize) {
3013                 log->l_curr_cycle++;
3014                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3015                         log->l_curr_cycle++;
3016                 log->l_curr_block -= log->l_logBBsize;
3017                 ASSERT(log->l_curr_block >= 0);
3018         }
3019         ASSERT(iclog == log->l_iclog);
3020         log->l_iclog = iclog->ic_next;
3021 }       /* xlog_state_switch_iclogs */
3022
3023 /*
3024  * Write out all data in the in-core log as of this exact moment in time.
3025  *
3026  * Data may be written to the in-core log during this call.  However,
3027  * we don't guarantee this data will be written out.  A change from past
3028  * implementation means this routine will *not* write out zero length LRs.
3029  *
3030  * Basically, we try and perform an intelligent scan of the in-core logs.
3031  * If we determine there is no flushable data, we just return.  There is no
3032  * flushable data if:
3033  *
3034  *      1. the current iclog is active and has no data; the previous iclog
3035  *              is in the active or dirty state.
3036  *      2. the current iclog is drity, and the previous iclog is in the
3037  *              active or dirty state.
3038  *
3039  * We may sleep if:
3040  *
3041  *      1. the current iclog is not in the active nor dirty state.
3042  *      2. the current iclog dirty, and the previous iclog is not in the
3043  *              active nor dirty state.
3044  *      3. the current iclog is active, and there is another thread writing
3045  *              to this particular iclog.
3046  *      4. a) the current iclog is active and has no other writers
3047  *         b) when we return from flushing out this iclog, it is still
3048  *              not in the active nor dirty state.
3049  */
3050 int
3051 _xfs_log_force(
3052         struct xfs_mount        *mp,
3053         uint                    flags,
3054         int                     *log_flushed)
3055 {
3056         struct xlog             *log = mp->m_log;
3057         struct xlog_in_core     *iclog;
3058         xfs_lsn_t               lsn;
3059
3060         XFS_STATS_INC(xs_log_force);
3061
3062         xlog_cil_force(log);
3063
3064         spin_lock(&log->l_icloglock);
3065
3066         iclog = log->l_iclog;
3067         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3068                 spin_unlock(&log->l_icloglock);
3069                 return XFS_ERROR(EIO);
3070         }
3071
3072         /* If the head iclog is not active nor dirty, we just attach
3073          * ourselves to the head and go to sleep.
3074          */
3075         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3076             iclog->ic_state == XLOG_STATE_DIRTY) {
3077                 /*
3078                  * If the head is dirty or (active and empty), then
3079                  * we need to look at the previous iclog.  If the previous
3080                  * iclog is active or dirty we are done.  There is nothing
3081                  * to sync out.  Otherwise, we attach ourselves to the
3082                  * previous iclog and go to sleep.
3083                  */
3084                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3085                     (atomic_read(&iclog->ic_refcnt) == 0
3086                      && iclog->ic_offset == 0)) {
3087                         iclog = iclog->ic_prev;
3088                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3089                             iclog->ic_state == XLOG_STATE_DIRTY)
3090                                 goto no_sleep;
3091                         else
3092                                 goto maybe_sleep;
3093                 } else {
3094                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3095                                 /* We are the only one with access to this
3096                                  * iclog.  Flush it out now.  There should
3097                                  * be a roundoff of zero to show that someone
3098                                  * has already taken care of the roundoff from
3099                                  * the previous sync.
3100                                  */
3101                                 atomic_inc(&iclog->ic_refcnt);
3102                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3103                                 xlog_state_switch_iclogs(log, iclog, 0);
3104                                 spin_unlock(&log->l_icloglock);
3105
3106                                 if (xlog_state_release_iclog(log, iclog))
3107                                         return XFS_ERROR(EIO);
3108
3109                                 if (log_flushed)
3110                                         *log_flushed = 1;
3111                                 spin_lock(&log->l_icloglock);
3112                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3113                                     iclog->ic_state != XLOG_STATE_DIRTY)
3114                                         goto maybe_sleep;
3115                                 else
3116                                         goto no_sleep;
3117                         } else {
3118                                 /* Someone else is writing to this iclog.
3119                                  * Use its call to flush out the data.  However,
3120                                  * the other thread may not force out this LR,
3121                                  * so we mark it WANT_SYNC.
3122                                  */
3123                                 xlog_state_switch_iclogs(log, iclog, 0);
3124                                 goto maybe_sleep;
3125                         }
3126                 }
3127         }
3128
3129         /* By the time we come around again, the iclog could've been filled
3130          * which would give it another lsn.  If we have a new lsn, just
3131          * return because the relevant data has been flushed.
3132          */
3133 maybe_sleep:
3134         if (flags & XFS_LOG_SYNC) {
3135                 /*
3136                  * We must check if we're shutting down here, before
3137                  * we wait, while we're holding the l_icloglock.
3138                  * Then we check again after waking up, in case our
3139                  * sleep was disturbed by a bad news.
3140                  */
3141                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3142                         spin_unlock(&log->l_icloglock);
3143                         return XFS_ERROR(EIO);
3144                 }
3145                 XFS_STATS_INC(xs_log_force_sleep);
3146                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3147                 /*
3148                  * No need to grab the log lock here since we're
3149                  * only deciding whether or not to return EIO
3150                  * and the memory read should be atomic.
3151                  */
3152                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3153                         return XFS_ERROR(EIO);
3154                 if (log_flushed)
3155                         *log_flushed = 1;
3156         } else {
3157
3158 no_sleep:
3159                 spin_unlock(&log->l_icloglock);
3160         }
3161         return 0;
3162 }
3163
3164 /*
3165  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3166  * about errors or whether the log was flushed or not. This is the normal
3167  * interface to use when trying to unpin items or move the log forward.
3168  */
3169 void
3170 xfs_log_force(
3171         xfs_mount_t     *mp,
3172         uint            flags)
3173 {
3174         int     error;
3175
3176         trace_xfs_log_force(mp, 0);
3177         error = _xfs_log_force(mp, flags, NULL);
3178         if (error)
3179                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3180 }
3181
3182 /*
3183  * Force the in-core log to disk for a specific LSN.
3184  *
3185  * Find in-core log with lsn.
3186  *      If it is in the DIRTY state, just return.
3187  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3188  *              state and go to sleep or return.
3189  *      If it is in any other state, go to sleep or return.
3190  *
3191  * Synchronous forces are implemented with a signal variable. All callers
3192  * to force a given lsn to disk will wait on a the sv attached to the
3193  * specific in-core log.  When given in-core log finally completes its
3194  * write to disk, that thread will wake up all threads waiting on the
3195  * sv.
3196  */
3197 int
3198 _xfs_log_force_lsn(
3199         struct xfs_mount        *mp,
3200         xfs_lsn_t               lsn,
3201         uint                    flags,
3202         int                     *log_flushed)
3203 {
3204         struct xlog             *log = mp->m_log;
3205         struct xlog_in_core     *iclog;
3206         int                     already_slept = 0;
3207
3208         ASSERT(lsn != 0);
3209
3210         XFS_STATS_INC(xs_log_force);
3211
3212         lsn = xlog_cil_force_lsn(log, lsn);
3213         if (lsn == NULLCOMMITLSN)
3214                 return 0;
3215
3216 try_again:
3217         spin_lock(&log->l_icloglock);
3218         iclog = log->l_iclog;
3219         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3220                 spin_unlock(&log->l_icloglock);
3221                 return XFS_ERROR(EIO);
3222         }
3223
3224         do {
3225                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3226                         iclog = iclog->ic_next;
3227                         continue;
3228                 }
3229
3230                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3231                         spin_unlock(&log->l_icloglock);
3232                         return 0;
3233                 }
3234
3235                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3236                         /*
3237                          * We sleep here if we haven't already slept (e.g.
3238                          * this is the first time we've looked at the correct
3239                          * iclog buf) and the buffer before us is going to
3240                          * be sync'ed. The reason for this is that if we
3241                          * are doing sync transactions here, by waiting for
3242                          * the previous I/O to complete, we can allow a few
3243                          * more transactions into this iclog before we close
3244                          * it down.
3245                          *
3246                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3247                          * up the refcnt so we can release the log (which
3248                          * drops the ref count).  The state switch keeps new
3249                          * transaction commits from using this buffer.  When
3250                          * the current commits finish writing into the buffer,
3251                          * the refcount will drop to zero and the buffer will
3252                          * go out then.
3253                          */
3254                         if (!already_slept &&
3255                             (iclog->ic_prev->ic_state &
3256                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3257                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3258
3259                                 XFS_STATS_INC(xs_log_force_sleep);
3260
3261                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3262                                                         &log->l_icloglock);
3263                                 if (log_flushed)
3264                                         *log_flushed = 1;
3265                                 already_slept = 1;
3266                                 goto try_again;
3267                         }
3268                         atomic_inc(&iclog->ic_refcnt);
3269                         xlog_state_switch_iclogs(log, iclog, 0);
3270                         spin_unlock(&log->l_icloglock);
3271                         if (xlog_state_release_iclog(log, iclog))
3272                                 return XFS_ERROR(EIO);
3273                         if (log_flushed)
3274                                 *log_flushed = 1;
3275                         spin_lock(&log->l_icloglock);
3276                 }
3277
3278                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3279                     !(iclog->ic_state &
3280                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3281                         /*
3282                          * Don't wait on completion if we know that we've
3283                          * gotten a log write error.
3284                          */
3285                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3286                                 spin_unlock(&log->l_icloglock);
3287                                 return XFS_ERROR(EIO);
3288                         }
3289                         XFS_STATS_INC(xs_log_force_sleep);
3290                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3291                         /*
3292                          * No need to grab the log lock here since we're
3293                          * only deciding whether or not to return EIO
3294                          * and the memory read should be atomic.
3295                          */
3296                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3297                                 return XFS_ERROR(EIO);
3298
3299                         if (log_flushed)
3300                                 *log_flushed = 1;
3301                 } else {                /* just return */
3302                         spin_unlock(&log->l_icloglock);
3303                 }
3304
3305                 return 0;
3306         } while (iclog != log->l_iclog);
3307
3308         spin_unlock(&log->l_icloglock);
3309         return 0;
3310 }
3311
3312 /*
3313  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3314  * about errors or whether the log was flushed or not. This is the normal
3315  * interface to use when trying to unpin items or move the log forward.
3316  */
3317 void
3318 xfs_log_force_lsn(
3319         xfs_mount_t     *mp,
3320         xfs_lsn_t       lsn,
3321         uint            flags)
3322 {
3323         int     error;
3324
3325         trace_xfs_log_force(mp, lsn);
3326         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3327         if (error)
3328                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3329 }
3330
3331 /*
3332  * Called when we want to mark the current iclog as being ready to sync to
3333  * disk.
3334  */
3335 STATIC void
3336 xlog_state_want_sync(
3337         struct xlog             *log,
3338         struct xlog_in_core     *iclog)
3339 {
3340         assert_spin_locked(&log->l_icloglock);
3341
3342         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3343                 xlog_state_switch_iclogs(log, iclog, 0);
3344         } else {
3345                 ASSERT(iclog->ic_state &
3346                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3347         }
3348 }
3349
3350
3351 /*****************************************************************************
3352  *
3353  *              TICKET functions
3354  *
3355  *****************************************************************************
3356  */
3357
3358 /*
3359  * Free a used ticket when its refcount falls to zero.
3360  */
3361 void
3362 xfs_log_ticket_put(
3363         xlog_ticket_t   *ticket)
3364 {
3365         ASSERT(atomic_read(&ticket->t_ref) > 0);
3366         if (atomic_dec_and_test(&ticket->t_ref))
3367                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3368 }
3369
3370 xlog_ticket_t *
3371 xfs_log_ticket_get(
3372         xlog_ticket_t   *ticket)
3373 {
3374         ASSERT(atomic_read(&ticket->t_ref) > 0);
3375         atomic_inc(&ticket->t_ref);
3376         return ticket;
3377 }
3378
3379 /*
3380  * Allocate and initialise a new log ticket.
3381  */
3382 struct xlog_ticket *
3383 xlog_ticket_alloc(
3384         struct xlog     *log,
3385         int             unit_bytes,
3386         int             cnt,
3387         char            client,
3388         bool            permanent,
3389         xfs_km_flags_t  alloc_flags)
3390 {
3391         struct xlog_ticket *tic;
3392         uint            num_headers;
3393         int             iclog_space;
3394
3395         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3396         if (!tic)
3397                 return NULL;
3398
3399         /*
3400          * Permanent reservations have up to 'cnt'-1 active log operations
3401          * in the log.  A unit in this case is the amount of space for one
3402          * of these log operations.  Normal reservations have a cnt of 1
3403          * and their unit amount is the total amount of space required.
3404          *
3405          * The following lines of code account for non-transaction data
3406          * which occupy space in the on-disk log.
3407          *
3408          * Normal form of a transaction is:
3409          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3410          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3411          *
3412          * We need to account for all the leadup data and trailer data
3413          * around the transaction data.
3414          * And then we need to account for the worst case in terms of using
3415          * more space.
3416          * The worst case will happen if:
3417          * - the placement of the transaction happens to be such that the
3418          *   roundoff is at its maximum
3419          * - the transaction data is synced before the commit record is synced
3420          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3421          *   Therefore the commit record is in its own Log Record.
3422          *   This can happen as the commit record is called with its
3423          *   own region to xlog_write().
3424          *   This then means that in the worst case, roundoff can happen for
3425          *   the commit-rec as well.
3426          *   The commit-rec is smaller than padding in this scenario and so it is
3427          *   not added separately.
3428          */
3429
3430         /* for trans header */
3431         unit_bytes += sizeof(xlog_op_header_t);
3432         unit_bytes += sizeof(xfs_trans_header_t);
3433
3434         /* for start-rec */
3435         unit_bytes += sizeof(xlog_op_header_t);
3436
3437         /*
3438          * for LR headers - the space for data in an iclog is the size minus
3439          * the space used for the headers. If we use the iclog size, then we
3440          * undercalculate the number of headers required.
3441          *
3442          * Furthermore - the addition of op headers for split-recs might
3443          * increase the space required enough to require more log and op
3444          * headers, so take that into account too.
3445          *
3446          * IMPORTANT: This reservation makes the assumption that if this
3447          * transaction is the first in an iclog and hence has the LR headers
3448          * accounted to it, then the remaining space in the iclog is
3449          * exclusively for this transaction.  i.e. if the transaction is larger
3450          * than the iclog, it will be the only thing in that iclog.
3451          * Fundamentally, this means we must pass the entire log vector to
3452          * xlog_write to guarantee this.
3453          */
3454         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3455         num_headers = howmany(unit_bytes, iclog_space);
3456
3457         /* for split-recs - ophdrs added when data split over LRs */
3458         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3459
3460         /* add extra header reservations if we overrun */
3461         while (!num_headers ||
3462                howmany(unit_bytes, iclog_space) > num_headers) {
3463                 unit_bytes += sizeof(xlog_op_header_t);
3464                 num_headers++;
3465         }
3466         unit_bytes += log->l_iclog_hsize * num_headers;
3467
3468         /* for commit-rec LR header - note: padding will subsume the ophdr */
3469         unit_bytes += log->l_iclog_hsize;
3470
3471         /* for roundoff padding for transaction data and one for commit record */
3472         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3473             log->l_mp->m_sb.sb_logsunit > 1) {
3474                 /* log su roundoff */
3475                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3476         } else {
3477                 /* BB roundoff */
3478                 unit_bytes += 2*BBSIZE;
3479         }
3480
3481         atomic_set(&tic->t_ref, 1);
3482         tic->t_task             = current;
3483         INIT_LIST_HEAD(&tic->t_queue);
3484         tic->t_unit_res         = unit_bytes;
3485         tic->t_curr_res         = unit_bytes;
3486         tic->t_cnt              = cnt;
3487         tic->t_ocnt             = cnt;
3488         tic->t_tid              = random32();
3489         tic->t_clientid         = client;
3490         tic->t_flags            = XLOG_TIC_INITED;
3491         tic->t_trans_type       = 0;
3492         if (permanent)
3493                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3494
3495         xlog_tic_reset_res(tic);
3496
3497         return tic;
3498 }
3499
3500
3501 /******************************************************************************
3502  *
3503  *              Log debug routines
3504  *
3505  ******************************************************************************
3506  */
3507 #if defined(DEBUG)
3508 /*
3509  * Make sure that the destination ptr is within the valid data region of
3510  * one of the iclogs.  This uses backup pointers stored in a different
3511  * part of the log in case we trash the log structure.
3512  */
3513 void
3514 xlog_verify_dest_ptr(
3515         struct xlog     *log,
3516         char            *ptr)
3517 {
3518         int i;
3519         int good_ptr = 0;
3520
3521         for (i = 0; i < log->l_iclog_bufs; i++) {
3522                 if (ptr >= log->l_iclog_bak[i] &&
3523                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3524                         good_ptr++;
3525         }
3526
3527         if (!good_ptr)
3528                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3529 }
3530
3531 /*
3532  * Check to make sure the grant write head didn't just over lap the tail.  If
3533  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3534  * the cycles differ by exactly one and check the byte count.
3535  *
3536  * This check is run unlocked, so can give false positives. Rather than assert
3537  * on failures, use a warn-once flag and a panic tag to allow the admin to
3538  * determine if they want to panic the machine when such an error occurs. For
3539  * debug kernels this will have the same effect as using an assert but, unlinke
3540  * an assert, it can be turned off at runtime.
3541  */
3542 STATIC void
3543 xlog_verify_grant_tail(
3544         struct xlog     *log)
3545 {
3546         int             tail_cycle, tail_blocks;
3547         int             cycle, space;
3548
3549         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3550         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3551         if (tail_cycle != cycle) {
3552                 if (cycle - 1 != tail_cycle &&
3553                     !(log->l_flags & XLOG_TAIL_WARN)) {
3554                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3555                                 "%s: cycle - 1 != tail_cycle", __func__);
3556                         log->l_flags |= XLOG_TAIL_WARN;
3557                 }
3558
3559                 if (space > BBTOB(tail_blocks) &&
3560                     !(log->l_flags & XLOG_TAIL_WARN)) {
3561                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3562                                 "%s: space > BBTOB(tail_blocks)", __func__);
3563                         log->l_flags |= XLOG_TAIL_WARN;
3564                 }
3565         }
3566 }
3567
3568 /* check if it will fit */
3569 STATIC void
3570 xlog_verify_tail_lsn(
3571         struct xlog             *log,
3572         struct xlog_in_core     *iclog,
3573         xfs_lsn_t               tail_lsn)
3574 {
3575     int blocks;
3576
3577     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3578         blocks =
3579             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3580         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3581                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3582     } else {
3583         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3584
3585         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3586                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3587
3588         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3589         if (blocks < BTOBB(iclog->ic_offset) + 1)
3590                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3591     }
3592 }       /* xlog_verify_tail_lsn */
3593
3594 /*
3595  * Perform a number of checks on the iclog before writing to disk.
3596  *
3597  * 1. Make sure the iclogs are still circular
3598  * 2. Make sure we have a good magic number
3599  * 3. Make sure we don't have magic numbers in the data
3600  * 4. Check fields of each log operation header for:
3601  *      A. Valid client identifier
3602  *      B. tid ptr value falls in valid ptr space (user space code)
3603  *      C. Length in log record header is correct according to the
3604  *              individual operation headers within record.
3605  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3606  *      log, check the preceding blocks of the physical log to make sure all
3607  *      the cycle numbers agree with the current cycle number.
3608  */
3609 STATIC void
3610 xlog_verify_iclog(
3611         struct xlog             *log,
3612         struct xlog_in_core     *iclog,
3613         int                     count,
3614         boolean_t               syncing)
3615 {
3616         xlog_op_header_t        *ophead;
3617         xlog_in_core_t          *icptr;
3618         xlog_in_core_2_t        *xhdr;
3619         xfs_caddr_t             ptr;
3620         xfs_caddr_t             base_ptr;
3621         __psint_t               field_offset;
3622         __uint8_t               clientid;
3623         int                     len, i, j, k, op_len;
3624         int                     idx;
3625
3626         /* check validity of iclog pointers */
3627         spin_lock(&log->l_icloglock);
3628         icptr = log->l_iclog;
3629         for (i=0; i < log->l_iclog_bufs; i++) {
3630                 if (icptr == NULL)
3631                         xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3632                 icptr = icptr->ic_next;
3633         }
3634         if (icptr != log->l_iclog)
3635                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3636         spin_unlock(&log->l_icloglock);
3637
3638         /* check log magic numbers */
3639         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3640                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3641
3642         ptr = (xfs_caddr_t) &iclog->ic_header;
3643         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3644              ptr += BBSIZE) {
3645                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3646                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3647                                 __func__);
3648         }
3649
3650         /* check fields */
3651         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3652         ptr = iclog->ic_datap;
3653         base_ptr = ptr;
3654         ophead = (xlog_op_header_t *)ptr;
3655         xhdr = iclog->ic_data;
3656         for (i = 0; i < len; i++) {
3657                 ophead = (xlog_op_header_t *)ptr;
3658
3659                 /* clientid is only 1 byte */
3660                 field_offset = (__psint_t)
3661                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3662                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3663                         clientid = ophead->oh_clientid;
3664                 } else {
3665                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3666                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3667                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3668                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3669                                 clientid = xlog_get_client_id(
3670                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3671                         } else {
3672                                 clientid = xlog_get_client_id(
3673                                         iclog->ic_header.h_cycle_data[idx]);
3674                         }
3675                 }
3676                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3677                         xfs_warn(log->l_mp,
3678                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3679                                 __func__, clientid, ophead,
3680                                 (unsigned long)field_offset);
3681
3682                 /* check length */
3683                 field_offset = (__psint_t)
3684                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3685                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3686                         op_len = be32_to_cpu(ophead->oh_len);
3687                 } else {
3688                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3689                                     (__psint_t)iclog->ic_datap);
3690                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3691                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3692                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3693                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3694                         } else {
3695                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3696                         }
3697                 }
3698                 ptr += sizeof(xlog_op_header_t) + op_len;
3699         }
3700 }       /* xlog_verify_iclog */
3701 #endif
3702
3703 /*
3704  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3705  */
3706 STATIC int
3707 xlog_state_ioerror(
3708         struct xlog     *log)
3709 {
3710         xlog_in_core_t  *iclog, *ic;
3711
3712         iclog = log->l_iclog;
3713         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3714                 /*
3715                  * Mark all the incore logs IOERROR.
3716                  * From now on, no log flushes will result.
3717                  */
3718                 ic = iclog;
3719                 do {
3720                         ic->ic_state = XLOG_STATE_IOERROR;
3721                         ic = ic->ic_next;
3722                 } while (ic != iclog);
3723                 return 0;
3724         }
3725         /*
3726          * Return non-zero, if state transition has already happened.
3727          */
3728         return 1;
3729 }
3730
3731 /*
3732  * This is called from xfs_force_shutdown, when we're forcibly
3733  * shutting down the filesystem, typically because of an IO error.
3734  * Our main objectives here are to make sure that:
3735  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3736  *         parties to find out, 'atomically'.
3737  *      b. those who're sleeping on log reservations, pinned objects and
3738  *          other resources get woken up, and be told the bad news.
3739  *      c. nothing new gets queued up after (a) and (b) are done.
3740  *      d. if !logerror, flush the iclogs to disk, then seal them off
3741  *         for business.
3742  *
3743  * Note: for delayed logging the !logerror case needs to flush the regions
3744  * held in memory out to the iclogs before flushing them to disk. This needs
3745  * to be done before the log is marked as shutdown, otherwise the flush to the
3746  * iclogs will fail.
3747  */
3748 int
3749 xfs_log_force_umount(
3750         struct xfs_mount        *mp,
3751         int                     logerror)
3752 {
3753         struct xlog     *log;
3754         int             retval;
3755
3756         log = mp->m_log;
3757
3758         /*
3759          * If this happens during log recovery, don't worry about
3760          * locking; the log isn't open for business yet.
3761          */
3762         if (!log ||
3763             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3764                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3765                 if (mp->m_sb_bp)
3766                         XFS_BUF_DONE(mp->m_sb_bp);
3767                 return 0;
3768         }
3769
3770         /*
3771          * Somebody could've already done the hard work for us.
3772          * No need to get locks for this.
3773          */
3774         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3775                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3776                 return 1;
3777         }
3778         retval = 0;
3779
3780         /*
3781          * Flush the in memory commit item list before marking the log as
3782          * being shut down. We need to do it in this order to ensure all the
3783          * completed transactions are flushed to disk with the xfs_log_force()
3784          * call below.
3785          */
3786         if (!logerror)
3787                 xlog_cil_force(log);
3788
3789         /*
3790          * mark the filesystem and the as in a shutdown state and wake
3791          * everybody up to tell them the bad news.
3792          */
3793         spin_lock(&log->l_icloglock);