Merge branch 'for-linus' of git://oss.sgi.com/xfs/xfs
[pandora-kernel.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_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_log_priv.h"
30 #include "xfs_buf_item.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_log_recover.h"
35 #include "xfs_trans_priv.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
38 #include "xfs_rw.h"
39 #include "xfs_trace.h"
40
41 kmem_zone_t     *xfs_log_ticket_zone;
42
43 /* Local miscellaneous function prototypes */
44 STATIC int       xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
45                                     xlog_in_core_t **, xfs_lsn_t *);
46 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t     *mp,
47                                 xfs_buftarg_t   *log_target,
48                                 xfs_daddr_t     blk_offset,
49                                 int             num_bblks);
50 STATIC int       xlog_space_left(xlog_t *log, int cycle, int bytes);
51 STATIC int       xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
52 STATIC void      xlog_dealloc_log(xlog_t *log);
53
54 /* local state machine functions */
55 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
56 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
57 STATIC int  xlog_state_get_iclog_space(xlog_t           *log,
58                                        int              len,
59                                        xlog_in_core_t   **iclog,
60                                        xlog_ticket_t    *ticket,
61                                        int              *continued_write,
62                                        int              *logoffsetp);
63 STATIC int  xlog_state_release_iclog(xlog_t             *log,
64                                      xlog_in_core_t     *iclog);
65 STATIC void xlog_state_switch_iclogs(xlog_t             *log,
66                                      xlog_in_core_t *iclog,
67                                      int                eventual_size);
68 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
69
70 /* local functions to manipulate grant head */
71 STATIC int  xlog_grant_log_space(xlog_t         *log,
72                                  xlog_ticket_t  *xtic);
73 STATIC void xlog_grant_push_ail(xfs_mount_t     *mp,
74                                 int             need_bytes);
75 STATIC void xlog_regrant_reserve_log_space(xlog_t        *log,
76                                            xlog_ticket_t *ticket);
77 STATIC int xlog_regrant_write_log_space(xlog_t          *log,
78                                          xlog_ticket_t  *ticket);
79 STATIC void xlog_ungrant_log_space(xlog_t        *log,
80                                    xlog_ticket_t *ticket);
81
82 #if defined(DEBUG)
83 STATIC void     xlog_verify_dest_ptr(xlog_t *log, char *ptr);
84 STATIC void     xlog_verify_grant_head(xlog_t *log, int equals);
85 STATIC void     xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
86                                   int count, boolean_t syncing);
87 STATIC void     xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
88                                      xfs_lsn_t tail_lsn);
89 #else
90 #define xlog_verify_dest_ptr(a,b)
91 #define xlog_verify_grant_head(a,b)
92 #define xlog_verify_iclog(a,b,c,d)
93 #define xlog_verify_tail_lsn(a,b,c)
94 #endif
95
96 STATIC int      xlog_iclogs_empty(xlog_t *log);
97
98
99 static void
100 xlog_ins_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
101 {
102         if (*qp) {
103                 tic->t_next         = (*qp);
104                 tic->t_prev         = (*qp)->t_prev;
105                 (*qp)->t_prev->t_next = tic;
106                 (*qp)->t_prev       = tic;
107         } else {
108                 tic->t_prev = tic->t_next = tic;
109                 *qp = tic;
110         }
111
112         tic->t_flags |= XLOG_TIC_IN_Q;
113 }
114
115 static void
116 xlog_del_ticketq(struct xlog_ticket **qp, struct xlog_ticket *tic)
117 {
118         if (tic == tic->t_next) {
119                 *qp = NULL;
120         } else {
121                 *qp = tic->t_next;
122                 tic->t_next->t_prev = tic->t_prev;
123                 tic->t_prev->t_next = tic->t_next;
124         }
125
126         tic->t_next = tic->t_prev = NULL;
127         tic->t_flags &= ~XLOG_TIC_IN_Q;
128 }
129
130 static void
131 xlog_grant_sub_space(struct log *log, int bytes)
132 {
133         log->l_grant_write_bytes -= bytes;
134         if (log->l_grant_write_bytes < 0) {
135                 log->l_grant_write_bytes += log->l_logsize;
136                 log->l_grant_write_cycle--;
137         }
138
139         log->l_grant_reserve_bytes -= bytes;
140         if ((log)->l_grant_reserve_bytes < 0) {
141                 log->l_grant_reserve_bytes += log->l_logsize;
142                 log->l_grant_reserve_cycle--;
143         }
144
145 }
146
147 static void
148 xlog_grant_add_space_write(struct log *log, int bytes)
149 {
150         int tmp = log->l_logsize - log->l_grant_write_bytes;
151         if (tmp > bytes)
152                 log->l_grant_write_bytes += bytes;
153         else {
154                 log->l_grant_write_cycle++;
155                 log->l_grant_write_bytes = bytes - tmp;
156         }
157 }
158
159 static void
160 xlog_grant_add_space_reserve(struct log *log, int bytes)
161 {
162         int tmp = log->l_logsize - log->l_grant_reserve_bytes;
163         if (tmp > bytes)
164                 log->l_grant_reserve_bytes += bytes;
165         else {
166                 log->l_grant_reserve_cycle++;
167                 log->l_grant_reserve_bytes = bytes - tmp;
168         }
169 }
170
171 static inline void
172 xlog_grant_add_space(struct log *log, int bytes)
173 {
174         xlog_grant_add_space_write(log, bytes);
175         xlog_grant_add_space_reserve(log, bytes);
176 }
177
178 static void
179 xlog_tic_reset_res(xlog_ticket_t *tic)
180 {
181         tic->t_res_num = 0;
182         tic->t_res_arr_sum = 0;
183         tic->t_res_num_ophdrs = 0;
184 }
185
186 static void
187 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
188 {
189         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
190                 /* add to overflow and start again */
191                 tic->t_res_o_flow += tic->t_res_arr_sum;
192                 tic->t_res_num = 0;
193                 tic->t_res_arr_sum = 0;
194         }
195
196         tic->t_res_arr[tic->t_res_num].r_len = len;
197         tic->t_res_arr[tic->t_res_num].r_type = type;
198         tic->t_res_arr_sum += len;
199         tic->t_res_num++;
200 }
201
202 /*
203  * NOTES:
204  *
205  *      1. currblock field gets updated at startup and after in-core logs
206  *              marked as with WANT_SYNC.
207  */
208
209 /*
210  * This routine is called when a user of a log manager ticket is done with
211  * the reservation.  If the ticket was ever used, then a commit record for
212  * the associated transaction is written out as a log operation header with
213  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
214  * a given ticket.  If the ticket was one with a permanent reservation, then
215  * a few operations are done differently.  Permanent reservation tickets by
216  * default don't release the reservation.  They just commit the current
217  * transaction with the belief that the reservation is still needed.  A flag
218  * must be passed in before permanent reservations are actually released.
219  * When these type of tickets are not released, they need to be set into
220  * the inited state again.  By doing this, a start record will be written
221  * out when the next write occurs.
222  */
223 xfs_lsn_t
224 xfs_log_done(
225         struct xfs_mount        *mp,
226         struct xlog_ticket      *ticket,
227         struct xlog_in_core     **iclog,
228         uint                    flags)
229 {
230         struct log              *log = mp->m_log;
231         xfs_lsn_t               lsn = 0;
232
233         if (XLOG_FORCED_SHUTDOWN(log) ||
234             /*
235              * If nothing was ever written, don't write out commit record.
236              * If we get an error, just continue and give back the log ticket.
237              */
238             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
239              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
240                 lsn = (xfs_lsn_t) -1;
241                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
242                         flags |= XFS_LOG_REL_PERM_RESERV;
243                 }
244         }
245
246
247         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
248             (flags & XFS_LOG_REL_PERM_RESERV)) {
249                 trace_xfs_log_done_nonperm(log, ticket);
250
251                 /*
252                  * Release ticket if not permanent reservation or a specific
253                  * request has been made to release a permanent reservation.
254                  */
255                 xlog_ungrant_log_space(log, ticket);
256                 xfs_log_ticket_put(ticket);
257         } else {
258                 trace_xfs_log_done_perm(log, ticket);
259
260                 xlog_regrant_reserve_log_space(log, ticket);
261                 /* If this ticket was a permanent reservation and we aren't
262                  * trying to release it, reset the inited flags; so next time
263                  * we write, a start record will be written out.
264                  */
265                 ticket->t_flags |= XLOG_TIC_INITED;
266         }
267
268         return lsn;
269 }
270
271 /*
272  * Attaches a new iclog I/O completion callback routine during
273  * transaction commit.  If the log is in error state, a non-zero
274  * return code is handed back and the caller is responsible for
275  * executing the callback at an appropriate time.
276  */
277 int
278 xfs_log_notify(
279         struct xfs_mount        *mp,
280         struct xlog_in_core     *iclog,
281         xfs_log_callback_t      *cb)
282 {
283         int     abortflg;
284
285         spin_lock(&iclog->ic_callback_lock);
286         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
287         if (!abortflg) {
288                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
289                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
290                 cb->cb_next = NULL;
291                 *(iclog->ic_callback_tail) = cb;
292                 iclog->ic_callback_tail = &(cb->cb_next);
293         }
294         spin_unlock(&iclog->ic_callback_lock);
295         return abortflg;
296 }
297
298 int
299 xfs_log_release_iclog(
300         struct xfs_mount        *mp,
301         struct xlog_in_core     *iclog)
302 {
303         if (xlog_state_release_iclog(mp->m_log, iclog)) {
304                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
305                 return EIO;
306         }
307
308         return 0;
309 }
310
311 /*
312  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
313  *      to the reservation.
314  *  2. Potentially, push buffers at tail of log to disk.
315  *
316  * Each reservation is going to reserve extra space for a log record header.
317  * When writes happen to the on-disk log, we don't subtract the length of the
318  * log record header from any reservation.  By wasting space in each
319  * reservation, we prevent over allocation problems.
320  */
321 int
322 xfs_log_reserve(
323         struct xfs_mount        *mp,
324         int                     unit_bytes,
325         int                     cnt,
326         struct xlog_ticket      **ticket,
327         __uint8_t               client,
328         uint                    flags,
329         uint                    t_type)
330 {
331         struct log              *log = mp->m_log;
332         struct xlog_ticket      *internal_ticket;
333         int                     retval = 0;
334
335         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
336
337         if (XLOG_FORCED_SHUTDOWN(log))
338                 return XFS_ERROR(EIO);
339
340         XFS_STATS_INC(xs_try_logspace);
341
342
343         if (*ticket != NULL) {
344                 ASSERT(flags & XFS_LOG_PERM_RESERV);
345                 internal_ticket = *ticket;
346
347                 /*
348                  * this is a new transaction on the ticket, so we need to
349                  * change the transaction ID so that the next transaction has a
350                  * different TID in the log. Just add one to the existing tid
351                  * so that we can see chains of rolling transactions in the log
352                  * easily.
353                  */
354                 internal_ticket->t_tid++;
355
356                 trace_xfs_log_reserve(log, internal_ticket);
357
358                 xlog_grant_push_ail(mp, internal_ticket->t_unit_res);
359                 retval = xlog_regrant_write_log_space(log, internal_ticket);
360         } else {
361                 /* may sleep if need to allocate more tickets */
362                 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
363                                                   client, flags,
364                                                   KM_SLEEP|KM_MAYFAIL);
365                 if (!internal_ticket)
366                         return XFS_ERROR(ENOMEM);
367                 internal_ticket->t_trans_type = t_type;
368                 *ticket = internal_ticket;
369
370                 trace_xfs_log_reserve(log, internal_ticket);
371
372                 xlog_grant_push_ail(mp,
373                                     (internal_ticket->t_unit_res *
374                                      internal_ticket->t_cnt));
375                 retval = xlog_grant_log_space(log, internal_ticket);
376         }
377
378         return retval;
379 }       /* xfs_log_reserve */
380
381
382 /*
383  * Mount a log filesystem
384  *
385  * mp           - ubiquitous xfs mount point structure
386  * log_target   - buftarg of on-disk log device
387  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
388  * num_bblocks  - Number of BBSIZE blocks in on-disk log
389  *
390  * Return error or zero.
391  */
392 int
393 xfs_log_mount(
394         xfs_mount_t     *mp,
395         xfs_buftarg_t   *log_target,
396         xfs_daddr_t     blk_offset,
397         int             num_bblks)
398 {
399         int             error;
400
401         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
402                 cmn_err(CE_NOTE, "XFS mounting filesystem %s", mp->m_fsname);
403         else {
404                 cmn_err(CE_NOTE,
405                         "!Mounting filesystem \"%s\" in no-recovery mode.  Filesystem will be inconsistent.",
406                         mp->m_fsname);
407                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
408         }
409
410         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
411         if (IS_ERR(mp->m_log)) {
412                 error = -PTR_ERR(mp->m_log);
413                 goto out;
414         }
415
416         /*
417          * Initialize the AIL now we have a log.
418          */
419         error = xfs_trans_ail_init(mp);
420         if (error) {
421                 cmn_err(CE_WARN, "XFS: AIL initialisation failed: error %d", error);
422                 goto out_free_log;
423         }
424         mp->m_log->l_ailp = mp->m_ail;
425
426         /*
427          * skip log recovery on a norecovery mount.  pretend it all
428          * just worked.
429          */
430         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
431                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
432
433                 if (readonly)
434                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
435
436                 error = xlog_recover(mp->m_log);
437
438                 if (readonly)
439                         mp->m_flags |= XFS_MOUNT_RDONLY;
440                 if (error) {
441                         cmn_err(CE_WARN, "XFS: log mount/recovery failed: error %d", error);
442                         goto out_destroy_ail;
443                 }
444         }
445
446         /* Normal transactions can now occur */
447         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
448
449         /*
450          * Now the log has been fully initialised and we know were our
451          * space grant counters are, we can initialise the permanent ticket
452          * needed for delayed logging to work.
453          */
454         xlog_cil_init_post_recovery(mp->m_log);
455
456         return 0;
457
458 out_destroy_ail:
459         xfs_trans_ail_destroy(mp);
460 out_free_log:
461         xlog_dealloc_log(mp->m_log);
462 out:
463         return error;
464 }
465
466 /*
467  * Finish the recovery of the file system.  This is separate from
468  * the xfs_log_mount() call, because it depends on the code in
469  * xfs_mountfs() to read in the root and real-time bitmap inodes
470  * between calling xfs_log_mount() and here.
471  *
472  * mp           - ubiquitous xfs mount point structure
473  */
474 int
475 xfs_log_mount_finish(xfs_mount_t *mp)
476 {
477         int     error;
478
479         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
480                 error = xlog_recover_finish(mp->m_log);
481         else {
482                 error = 0;
483                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
484         }
485
486         return error;
487 }
488
489 /*
490  * Final log writes as part of unmount.
491  *
492  * Mark the filesystem clean as unmount happens.  Note that during relocation
493  * this routine needs to be executed as part of source-bag while the
494  * deallocation must not be done until source-end.
495  */
496
497 /*
498  * Unmount record used to have a string "Unmount filesystem--" in the
499  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
500  * We just write the magic number now since that particular field isn't
501  * currently architecture converted and "nUmount" is a bit foo.
502  * As far as I know, there weren't any dependencies on the old behaviour.
503  */
504
505 int
506 xfs_log_unmount_write(xfs_mount_t *mp)
507 {
508         xlog_t           *log = mp->m_log;
509         xlog_in_core_t   *iclog;
510 #ifdef DEBUG
511         xlog_in_core_t   *first_iclog;
512 #endif
513         xlog_ticket_t   *tic = NULL;
514         xfs_lsn_t        lsn;
515         int              error;
516
517         /*
518          * Don't write out unmount record on read-only mounts.
519          * Or, if we are doing a forced umount (typically because of IO errors).
520          */
521         if (mp->m_flags & XFS_MOUNT_RDONLY)
522                 return 0;
523
524         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
525         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
526
527 #ifdef DEBUG
528         first_iclog = iclog = log->l_iclog;
529         do {
530                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
531                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
532                         ASSERT(iclog->ic_offset == 0);
533                 }
534                 iclog = iclog->ic_next;
535         } while (iclog != first_iclog);
536 #endif
537         if (! (XLOG_FORCED_SHUTDOWN(log))) {
538                 error = xfs_log_reserve(mp, 600, 1, &tic,
539                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
540                 if (!error) {
541                         /* the data section must be 32 bit size aligned */
542                         struct {
543                             __uint16_t magic;
544                             __uint16_t pad1;
545                             __uint32_t pad2; /* may as well make it 64 bits */
546                         } magic = {
547                                 .magic = XLOG_UNMOUNT_TYPE,
548                         };
549                         struct xfs_log_iovec reg = {
550                                 .i_addr = &magic,
551                                 .i_len = sizeof(magic),
552                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
553                         };
554                         struct xfs_log_vec vec = {
555                                 .lv_niovecs = 1,
556                                 .lv_iovecp = &reg,
557                         };
558
559                         /* remove inited flag */
560                         tic->t_flags = 0;
561                         error = xlog_write(log, &vec, tic, &lsn,
562                                            NULL, XLOG_UNMOUNT_TRANS);
563                         /*
564                          * At this point, we're umounting anyway,
565                          * so there's no point in transitioning log state
566                          * to IOERROR. Just continue...
567                          */
568                 }
569
570                 if (error) {
571                         xfs_fs_cmn_err(CE_ALERT, mp,
572                                 "xfs_log_unmount: unmount record failed");
573                 }
574
575
576                 spin_lock(&log->l_icloglock);
577                 iclog = log->l_iclog;
578                 atomic_inc(&iclog->ic_refcnt);
579                 xlog_state_want_sync(log, iclog);
580                 spin_unlock(&log->l_icloglock);
581                 error = xlog_state_release_iclog(log, iclog);
582
583                 spin_lock(&log->l_icloglock);
584                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
585                       iclog->ic_state == XLOG_STATE_DIRTY)) {
586                         if (!XLOG_FORCED_SHUTDOWN(log)) {
587                                 sv_wait(&iclog->ic_force_wait, PMEM,
588                                         &log->l_icloglock, s);
589                         } else {
590                                 spin_unlock(&log->l_icloglock);
591                         }
592                 } else {
593                         spin_unlock(&log->l_icloglock);
594                 }
595                 if (tic) {
596                         trace_xfs_log_umount_write(log, tic);
597                         xlog_ungrant_log_space(log, tic);
598                         xfs_log_ticket_put(tic);
599                 }
600         } else {
601                 /*
602                  * We're already in forced_shutdown mode, couldn't
603                  * even attempt to write out the unmount transaction.
604                  *
605                  * Go through the motions of sync'ing and releasing
606                  * the iclog, even though no I/O will actually happen,
607                  * we need to wait for other log I/Os that may already
608                  * be in progress.  Do this as a separate section of
609                  * code so we'll know if we ever get stuck here that
610                  * we're in this odd situation of trying to unmount
611                  * a file system that went into forced_shutdown as
612                  * the result of an unmount..
613                  */
614                 spin_lock(&log->l_icloglock);
615                 iclog = log->l_iclog;
616                 atomic_inc(&iclog->ic_refcnt);
617
618                 xlog_state_want_sync(log, iclog);
619                 spin_unlock(&log->l_icloglock);
620                 error =  xlog_state_release_iclog(log, iclog);
621
622                 spin_lock(&log->l_icloglock);
623
624                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
625                         || iclog->ic_state == XLOG_STATE_DIRTY
626                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
627
628                                 sv_wait(&iclog->ic_force_wait, PMEM,
629                                         &log->l_icloglock, s);
630                 } else {
631                         spin_unlock(&log->l_icloglock);
632                 }
633         }
634
635         return error;
636 }       /* xfs_log_unmount_write */
637
638 /*
639  * Deallocate log structures for unmount/relocation.
640  *
641  * We need to stop the aild from running before we destroy
642  * and deallocate the log as the aild references the log.
643  */
644 void
645 xfs_log_unmount(xfs_mount_t *mp)
646 {
647         xfs_trans_ail_destroy(mp);
648         xlog_dealloc_log(mp->m_log);
649 }
650
651 void
652 xfs_log_item_init(
653         struct xfs_mount        *mp,
654         struct xfs_log_item     *item,
655         int                     type,
656         struct xfs_item_ops     *ops)
657 {
658         item->li_mountp = mp;
659         item->li_ailp = mp->m_ail;
660         item->li_type = type;
661         item->li_ops = ops;
662         item->li_lv = NULL;
663
664         INIT_LIST_HEAD(&item->li_ail);
665         INIT_LIST_HEAD(&item->li_cil);
666 }
667
668 /*
669  * Write region vectors to log.  The write happens using the space reservation
670  * of the ticket (tic).  It is not a requirement that all writes for a given
671  * transaction occur with one call to xfs_log_write(). However, it is important
672  * to note that the transaction reservation code makes an assumption about the
673  * number of log headers a transaction requires that may be violated if you
674  * don't pass all the transaction vectors in one call....
675  */
676 int
677 xfs_log_write(
678         struct xfs_mount        *mp,
679         struct xfs_log_iovec    reg[],
680         int                     nentries,
681         struct xlog_ticket      *tic,
682         xfs_lsn_t               *start_lsn)
683 {
684         struct log              *log = mp->m_log;
685         int                     error;
686         struct xfs_log_vec      vec = {
687                 .lv_niovecs = nentries,
688                 .lv_iovecp = reg,
689         };
690
691         if (XLOG_FORCED_SHUTDOWN(log))
692                 return XFS_ERROR(EIO);
693
694         error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
695         if (error)
696                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
697         return error;
698 }
699
700 void
701 xfs_log_move_tail(xfs_mount_t   *mp,
702                   xfs_lsn_t     tail_lsn)
703 {
704         xlog_ticket_t   *tic;
705         xlog_t          *log = mp->m_log;
706         int             need_bytes, free_bytes, cycle, bytes;
707
708         if (XLOG_FORCED_SHUTDOWN(log))
709                 return;
710
711         if (tail_lsn == 0) {
712                 /* needed since sync_lsn is 64 bits */
713                 spin_lock(&log->l_icloglock);
714                 tail_lsn = log->l_last_sync_lsn;
715                 spin_unlock(&log->l_icloglock);
716         }
717
718         spin_lock(&log->l_grant_lock);
719
720         /* Also an invalid lsn.  1 implies that we aren't passing in a valid
721          * tail_lsn.
722          */
723         if (tail_lsn != 1) {
724                 log->l_tail_lsn = tail_lsn;
725         }
726
727         if ((tic = log->l_write_headq)) {
728 #ifdef DEBUG
729                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
730                         panic("Recovery problem");
731 #endif
732                 cycle = log->l_grant_write_cycle;
733                 bytes = log->l_grant_write_bytes;
734                 free_bytes = xlog_space_left(log, cycle, bytes);
735                 do {
736                         ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
737
738                         if (free_bytes < tic->t_unit_res && tail_lsn != 1)
739                                 break;
740                         tail_lsn = 0;
741                         free_bytes -= tic->t_unit_res;
742                         sv_signal(&tic->t_wait);
743                         tic = tic->t_next;
744                 } while (tic != log->l_write_headq);
745         }
746         if ((tic = log->l_reserve_headq)) {
747 #ifdef DEBUG
748                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
749                         panic("Recovery problem");
750 #endif
751                 cycle = log->l_grant_reserve_cycle;
752                 bytes = log->l_grant_reserve_bytes;
753                 free_bytes = xlog_space_left(log, cycle, bytes);
754                 do {
755                         if (tic->t_flags & XLOG_TIC_PERM_RESERV)
756                                 need_bytes = tic->t_unit_res*tic->t_cnt;
757                         else
758                                 need_bytes = tic->t_unit_res;
759                         if (free_bytes < need_bytes && tail_lsn != 1)
760                                 break;
761                         tail_lsn = 0;
762                         free_bytes -= need_bytes;
763                         sv_signal(&tic->t_wait);
764                         tic = tic->t_next;
765                 } while (tic != log->l_reserve_headq);
766         }
767         spin_unlock(&log->l_grant_lock);
768 }       /* xfs_log_move_tail */
769
770 /*
771  * Determine if we have a transaction that has gone to disk
772  * that needs to be covered. To begin the transition to the idle state
773  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
774  * If we are then in a state where covering is needed, the caller is informed
775  * that dummy transactions are required to move the log into the idle state.
776  *
777  * Because this is called as part of the sync process, we should also indicate
778  * that dummy transactions should be issued in anything but the covered or
779  * idle states. This ensures that the log tail is accurately reflected in
780  * the log at the end of the sync, hence if a crash occurrs avoids replay
781  * of transactions where the metadata is already on disk.
782  */
783 int
784 xfs_log_need_covered(xfs_mount_t *mp)
785 {
786         int             needed = 0;
787         xlog_t          *log = mp->m_log;
788
789         if (!xfs_fs_writable(mp))
790                 return 0;
791
792         spin_lock(&log->l_icloglock);
793         switch (log->l_covered_state) {
794         case XLOG_STATE_COVER_DONE:
795         case XLOG_STATE_COVER_DONE2:
796         case XLOG_STATE_COVER_IDLE:
797                 break;
798         case XLOG_STATE_COVER_NEED:
799         case XLOG_STATE_COVER_NEED2:
800                 if (!xfs_trans_ail_tail(log->l_ailp) &&
801                     xlog_iclogs_empty(log)) {
802                         if (log->l_covered_state == XLOG_STATE_COVER_NEED)
803                                 log->l_covered_state = XLOG_STATE_COVER_DONE;
804                         else
805                                 log->l_covered_state = XLOG_STATE_COVER_DONE2;
806                 }
807                 /* FALLTHRU */
808         default:
809                 needed = 1;
810                 break;
811         }
812         spin_unlock(&log->l_icloglock);
813         return needed;
814 }
815
816 /******************************************************************************
817  *
818  *      local routines
819  *
820  ******************************************************************************
821  */
822
823 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
824  * The log manager must keep track of the last LR which was committed
825  * to disk.  The lsn of this LR will become the new tail_lsn whenever
826  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
827  * the situation where stuff could be written into the log but nothing
828  * was ever in the AIL when asked.  Eventually, we panic since the
829  * tail hits the head.
830  *
831  * We may be holding the log iclog lock upon entering this routine.
832  */
833 xfs_lsn_t
834 xlog_assign_tail_lsn(xfs_mount_t *mp)
835 {
836         xfs_lsn_t tail_lsn;
837         xlog_t    *log = mp->m_log;
838
839         tail_lsn = xfs_trans_ail_tail(mp->m_ail);
840         spin_lock(&log->l_grant_lock);
841         if (tail_lsn != 0) {
842                 log->l_tail_lsn = tail_lsn;
843         } else {
844                 tail_lsn = log->l_tail_lsn = log->l_last_sync_lsn;
845         }
846         spin_unlock(&log->l_grant_lock);
847
848         return tail_lsn;
849 }       /* xlog_assign_tail_lsn */
850
851
852 /*
853  * Return the space in the log between the tail and the head.  The head
854  * is passed in the cycle/bytes formal parms.  In the special case where
855  * the reserve head has wrapped passed the tail, this calculation is no
856  * longer valid.  In this case, just return 0 which means there is no space
857  * in the log.  This works for all places where this function is called
858  * with the reserve head.  Of course, if the write head were to ever
859  * wrap the tail, we should blow up.  Rather than catch this case here,
860  * we depend on other ASSERTions in other parts of the code.   XXXmiken
861  *
862  * This code also handles the case where the reservation head is behind
863  * the tail.  The details of this case are described below, but the end
864  * result is that we return the size of the log as the amount of space left.
865  */
866 STATIC int
867 xlog_space_left(xlog_t *log, int cycle, int bytes)
868 {
869         int free_bytes;
870         int tail_bytes;
871         int tail_cycle;
872
873         tail_bytes = BBTOB(BLOCK_LSN(log->l_tail_lsn));
874         tail_cycle = CYCLE_LSN(log->l_tail_lsn);
875         if ((tail_cycle == cycle) && (bytes >= tail_bytes)) {
876                 free_bytes = log->l_logsize - (bytes - tail_bytes);
877         } else if ((tail_cycle + 1) < cycle) {
878                 return 0;
879         } else if (tail_cycle < cycle) {
880                 ASSERT(tail_cycle == (cycle - 1));
881                 free_bytes = tail_bytes - bytes;
882         } else {
883                 /*
884                  * The reservation head is behind the tail.
885                  * In this case we just want to return the size of the
886                  * log as the amount of space left.
887                  */
888                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
889                         "xlog_space_left: head behind tail\n"
890                         "  tail_cycle = %d, tail_bytes = %d\n"
891                         "  GH   cycle = %d, GH   bytes = %d",
892                         tail_cycle, tail_bytes, cycle, bytes);
893                 ASSERT(0);
894                 free_bytes = log->l_logsize;
895         }
896         return free_bytes;
897 }       /* xlog_space_left */
898
899
900 /*
901  * Log function which is called when an io completes.
902  *
903  * The log manager needs its own routine, in order to control what
904  * happens with the buffer after the write completes.
905  */
906 void
907 xlog_iodone(xfs_buf_t *bp)
908 {
909         xlog_in_core_t  *iclog;
910         xlog_t          *l;
911         int             aborted;
912
913         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
914         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
915         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
916         aborted = 0;
917         l = iclog->ic_log;
918
919         /*
920          * Race to shutdown the filesystem if we see an error.
921          */
922         if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
923                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
924                 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
925                 XFS_BUF_STALE(bp);
926                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
927                 /*
928                  * This flag will be propagated to the trans-committed
929                  * callback routines to let them know that the log-commit
930                  * didn't succeed.
931                  */
932                 aborted = XFS_LI_ABORTED;
933         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
934                 aborted = XFS_LI_ABORTED;
935         }
936
937         /* log I/O is always issued ASYNC */
938         ASSERT(XFS_BUF_ISASYNC(bp));
939         xlog_state_done_syncing(iclog, aborted);
940         /*
941          * do not reference the buffer (bp) here as we could race
942          * with it being freed after writing the unmount record to the
943          * log.
944          */
945
946 }       /* xlog_iodone */
947
948 /*
949  * Return size of each in-core log record buffer.
950  *
951  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
952  *
953  * If the filesystem blocksize is too large, we may need to choose a
954  * larger size since the directory code currently logs entire blocks.
955  */
956
957 STATIC void
958 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
959                            xlog_t       *log)
960 {
961         int size;
962         int xhdrs;
963
964         if (mp->m_logbufs <= 0)
965                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
966         else
967                 log->l_iclog_bufs = mp->m_logbufs;
968
969         /*
970          * Buffer size passed in from mount system call.
971          */
972         if (mp->m_logbsize > 0) {
973                 size = log->l_iclog_size = mp->m_logbsize;
974                 log->l_iclog_size_log = 0;
975                 while (size != 1) {
976                         log->l_iclog_size_log++;
977                         size >>= 1;
978                 }
979
980                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
981                         /* # headers = size / 32k
982                          * one header holds cycles from 32k of data
983                          */
984
985                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
986                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
987                                 xhdrs++;
988                         log->l_iclog_hsize = xhdrs << BBSHIFT;
989                         log->l_iclog_heads = xhdrs;
990                 } else {
991                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
992                         log->l_iclog_hsize = BBSIZE;
993                         log->l_iclog_heads = 1;
994                 }
995                 goto done;
996         }
997
998         /* All machines use 32kB buffers by default. */
999         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1000         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1001
1002         /* the default log size is 16k or 32k which is one header sector */
1003         log->l_iclog_hsize = BBSIZE;
1004         log->l_iclog_heads = 1;
1005
1006 done:
1007         /* are we being asked to make the sizes selected above visible? */
1008         if (mp->m_logbufs == 0)
1009                 mp->m_logbufs = log->l_iclog_bufs;
1010         if (mp->m_logbsize == 0)
1011                 mp->m_logbsize = log->l_iclog_size;
1012 }       /* xlog_get_iclog_buffer_size */
1013
1014
1015 /*
1016  * This routine initializes some of the log structure for a given mount point.
1017  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1018  * some other stuff may be filled in too.
1019  */
1020 STATIC xlog_t *
1021 xlog_alloc_log(xfs_mount_t      *mp,
1022                xfs_buftarg_t    *log_target,
1023                xfs_daddr_t      blk_offset,
1024                int              num_bblks)
1025 {
1026         xlog_t                  *log;
1027         xlog_rec_header_t       *head;
1028         xlog_in_core_t          **iclogp;
1029         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1030         xfs_buf_t               *bp;
1031         int                     i;
1032         int                     error = ENOMEM;
1033         uint                    log2_size = 0;
1034
1035         log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1036         if (!log) {
1037                 xlog_warn("XFS: Log allocation failed: No memory!");
1038                 goto out;
1039         }
1040
1041         log->l_mp          = mp;
1042         log->l_targ        = log_target;
1043         log->l_logsize     = BBTOB(num_bblks);
1044         log->l_logBBstart  = blk_offset;
1045         log->l_logBBsize   = num_bblks;
1046         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1047         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1048
1049         log->l_prev_block  = -1;
1050         log->l_tail_lsn    = xlog_assign_lsn(1, 0);
1051         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1052         log->l_last_sync_lsn = log->l_tail_lsn;
1053         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1054         log->l_grant_reserve_cycle = 1;
1055         log->l_grant_write_cycle = 1;
1056
1057         error = EFSCORRUPTED;
1058         if (xfs_sb_version_hassector(&mp->m_sb)) {
1059                 log2_size = mp->m_sb.sb_logsectlog;
1060                 if (log2_size < BBSHIFT) {
1061                         xlog_warn("XFS: Log sector size too small "
1062                                 "(0x%x < 0x%x)", log2_size, BBSHIFT);
1063                         goto out_free_log;
1064                 }
1065
1066                 log2_size -= BBSHIFT;
1067                 if (log2_size > mp->m_sectbb_log) {
1068                         xlog_warn("XFS: Log sector size too large "
1069                                 "(0x%x > 0x%x)", log2_size, mp->m_sectbb_log);
1070                         goto out_free_log;
1071                 }
1072
1073                 /* for larger sector sizes, must have v2 or external log */
1074                 if (log2_size && log->l_logBBstart > 0 &&
1075                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1076
1077                         xlog_warn("XFS: log sector size (0x%x) invalid "
1078                                   "for configuration.", log2_size);
1079                         goto out_free_log;
1080                 }
1081         }
1082         log->l_sectBBsize = 1 << log2_size;
1083
1084         xlog_get_iclog_buffer_size(mp, log);
1085
1086         error = ENOMEM;
1087         bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1088         if (!bp)
1089                 goto out_free_log;
1090         XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1091         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1092         ASSERT(XFS_BUF_ISBUSY(bp));
1093         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1094         log->l_xbuf = bp;
1095
1096         spin_lock_init(&log->l_icloglock);
1097         spin_lock_init(&log->l_grant_lock);
1098         sv_init(&log->l_flush_wait, 0, "flush_wait");
1099
1100         /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1101         ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1102
1103         iclogp = &log->l_iclog;
1104         /*
1105          * The amount of memory to allocate for the iclog structure is
1106          * rather funky due to the way the structure is defined.  It is
1107          * done this way so that we can use different sizes for machines
1108          * with different amounts of memory.  See the definition of
1109          * xlog_in_core_t in xfs_log_priv.h for details.
1110          */
1111         ASSERT(log->l_iclog_size >= 4096);
1112         for (i=0; i < log->l_iclog_bufs; i++) {
1113                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1114                 if (!*iclogp)
1115                         goto out_free_iclog;
1116
1117                 iclog = *iclogp;
1118                 iclog->ic_prev = prev_iclog;
1119                 prev_iclog = iclog;
1120
1121                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1122                                                 log->l_iclog_size, 0);
1123                 if (!bp)
1124                         goto out_free_iclog;
1125                 if (!XFS_BUF_CPSEMA(bp))
1126                         ASSERT(0);
1127                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1128                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1129                 iclog->ic_bp = bp;
1130                 iclog->ic_data = bp->b_addr;
1131 #ifdef DEBUG
1132                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1133 #endif
1134                 head = &iclog->ic_header;
1135                 memset(head, 0, sizeof(xlog_rec_header_t));
1136                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1137                 head->h_version = cpu_to_be32(
1138                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1139                 head->h_size = cpu_to_be32(log->l_iclog_size);
1140                 /* new fields */
1141                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1142                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1143
1144                 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1145                 iclog->ic_state = XLOG_STATE_ACTIVE;
1146                 iclog->ic_log = log;
1147                 atomic_set(&iclog->ic_refcnt, 0);
1148                 spin_lock_init(&iclog->ic_callback_lock);
1149                 iclog->ic_callback_tail = &(iclog->ic_callback);
1150                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1151
1152                 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1153                 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1154                 sv_init(&iclog->ic_force_wait, SV_DEFAULT, "iclog-force");
1155                 sv_init(&iclog->ic_write_wait, SV_DEFAULT, "iclog-write");
1156
1157                 iclogp = &iclog->ic_next;
1158         }
1159         *iclogp = log->l_iclog;                 /* complete ring */
1160         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1161
1162         error = xlog_cil_init(log);
1163         if (error)
1164                 goto out_free_iclog;
1165         return log;
1166
1167 out_free_iclog:
1168         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1169                 prev_iclog = iclog->ic_next;
1170                 if (iclog->ic_bp) {
1171                         sv_destroy(&iclog->ic_force_wait);
1172                         sv_destroy(&iclog->ic_write_wait);
1173                         xfs_buf_free(iclog->ic_bp);
1174                 }
1175                 kmem_free(iclog);
1176         }
1177         spinlock_destroy(&log->l_icloglock);
1178         spinlock_destroy(&log->l_grant_lock);
1179         xfs_buf_free(log->l_xbuf);
1180 out_free_log:
1181         kmem_free(log);
1182 out:
1183         return ERR_PTR(-error);
1184 }       /* xlog_alloc_log */
1185
1186
1187 /*
1188  * Write out the commit record of a transaction associated with the given
1189  * ticket.  Return the lsn of the commit record.
1190  */
1191 STATIC int
1192 xlog_commit_record(
1193         struct log              *log,
1194         struct xlog_ticket      *ticket,
1195         struct xlog_in_core     **iclog,
1196         xfs_lsn_t               *commitlsnp)
1197 {
1198         struct xfs_mount *mp = log->l_mp;
1199         int     error;
1200         struct xfs_log_iovec reg = {
1201                 .i_addr = NULL,
1202                 .i_len = 0,
1203                 .i_type = XLOG_REG_TYPE_COMMIT,
1204         };
1205         struct xfs_log_vec vec = {
1206                 .lv_niovecs = 1,
1207                 .lv_iovecp = &reg,
1208         };
1209
1210         ASSERT_ALWAYS(iclog);
1211         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1212                                         XLOG_COMMIT_TRANS);
1213         if (error)
1214                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1215         return error;
1216 }
1217
1218 /*
1219  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1220  * log space.  This code pushes on the lsn which would supposedly free up
1221  * the 25% which we want to leave free.  We may need to adopt a policy which
1222  * pushes on an lsn which is further along in the log once we reach the high
1223  * water mark.  In this manner, we would be creating a low water mark.
1224  */
1225 STATIC void
1226 xlog_grant_push_ail(xfs_mount_t *mp,
1227                     int         need_bytes)
1228 {
1229     xlog_t      *log = mp->m_log;       /* pointer to the log */
1230     xfs_lsn_t   tail_lsn;               /* lsn of the log tail */
1231     xfs_lsn_t   threshold_lsn = 0;      /* lsn we'd like to be at */
1232     int         free_blocks;            /* free blocks left to write to */
1233     int         free_bytes;             /* free bytes left to write to */
1234     int         threshold_block;        /* block in lsn we'd like to be at */
1235     int         threshold_cycle;        /* lsn cycle we'd like to be at */
1236     int         free_threshold;
1237
1238     ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1239
1240     spin_lock(&log->l_grant_lock);
1241     free_bytes = xlog_space_left(log,
1242                                  log->l_grant_reserve_cycle,
1243                                  log->l_grant_reserve_bytes);
1244     tail_lsn = log->l_tail_lsn;
1245     free_blocks = BTOBBT(free_bytes);
1246
1247     /*
1248      * Set the threshold for the minimum number of free blocks in the
1249      * log to the maximum of what the caller needs, one quarter of the
1250      * log, and 256 blocks.
1251      */
1252     free_threshold = BTOBB(need_bytes);
1253     free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1254     free_threshold = MAX(free_threshold, 256);
1255     if (free_blocks < free_threshold) {
1256         threshold_block = BLOCK_LSN(tail_lsn) + free_threshold;
1257         threshold_cycle = CYCLE_LSN(tail_lsn);
1258         if (threshold_block >= log->l_logBBsize) {
1259             threshold_block -= log->l_logBBsize;
1260             threshold_cycle += 1;
1261         }
1262         threshold_lsn = xlog_assign_lsn(threshold_cycle, threshold_block);
1263
1264         /* Don't pass in an lsn greater than the lsn of the last
1265          * log record known to be on disk.
1266          */
1267         if (XFS_LSN_CMP(threshold_lsn, log->l_last_sync_lsn) > 0)
1268             threshold_lsn = log->l_last_sync_lsn;
1269     }
1270     spin_unlock(&log->l_grant_lock);
1271
1272     /*
1273      * Get the transaction layer to kick the dirty buffers out to
1274      * disk asynchronously. No point in trying to do this if
1275      * the filesystem is shutting down.
1276      */
1277     if (threshold_lsn &&
1278         !XLOG_FORCED_SHUTDOWN(log))
1279             xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1280 }       /* xlog_grant_push_ail */
1281
1282 /*
1283  * The bdstrat callback function for log bufs. This gives us a central
1284  * place to trap bufs in case we get hit by a log I/O error and need to
1285  * shutdown. Actually, in practice, even when we didn't get a log error,
1286  * we transition the iclogs to IOERROR state *after* flushing all existing
1287  * iclogs to disk. This is because we don't want anymore new transactions to be
1288  * started or completed afterwards.
1289  */
1290 STATIC int
1291 xlog_bdstrat(
1292         struct xfs_buf          *bp)
1293 {
1294         struct xlog_in_core     *iclog;
1295
1296         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1297         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1298                 XFS_BUF_ERROR(bp, EIO);
1299                 XFS_BUF_STALE(bp);
1300                 xfs_buf_ioend(bp, 0);
1301                 /*
1302                  * It would seem logical to return EIO here, but we rely on
1303                  * the log state machine to propagate I/O errors instead of
1304                  * doing it here.
1305                  */
1306                 return 0;
1307         }
1308
1309         bp->b_flags |= _XBF_RUN_QUEUES;
1310         xfs_buf_iorequest(bp);
1311         return 0;
1312 }
1313
1314 /*
1315  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1316  * fashion.  Previously, we should have moved the current iclog
1317  * ptr in the log to point to the next available iclog.  This allows further
1318  * write to continue while this code syncs out an iclog ready to go.
1319  * Before an in-core log can be written out, the data section must be scanned
1320  * to save away the 1st word of each BBSIZE block into the header.  We replace
1321  * it with the current cycle count.  Each BBSIZE block is tagged with the
1322  * cycle count because there in an implicit assumption that drives will
1323  * guarantee that entire 512 byte blocks get written at once.  In other words,
1324  * we can't have part of a 512 byte block written and part not written.  By
1325  * tagging each block, we will know which blocks are valid when recovering
1326  * after an unclean shutdown.
1327  *
1328  * This routine is single threaded on the iclog.  No other thread can be in
1329  * this routine with the same iclog.  Changing contents of iclog can there-
1330  * fore be done without grabbing the state machine lock.  Updating the global
1331  * log will require grabbing the lock though.
1332  *
1333  * The entire log manager uses a logical block numbering scheme.  Only
1334  * log_sync (and then only bwrite()) know about the fact that the log may
1335  * not start with block zero on a given device.  The log block start offset
1336  * is added immediately before calling bwrite().
1337  */
1338
1339 STATIC int
1340 xlog_sync(xlog_t                *log,
1341           xlog_in_core_t        *iclog)
1342 {
1343         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1344         xfs_buf_t       *bp;
1345         int             i;
1346         uint            count;          /* byte count of bwrite */
1347         uint            count_init;     /* initial count before roundup */
1348         int             roundoff;       /* roundoff to BB or stripe */
1349         int             split = 0;      /* split write into two regions */
1350         int             error;
1351         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1352
1353         XFS_STATS_INC(xs_log_writes);
1354         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1355
1356         /* Add for LR header */
1357         count_init = log->l_iclog_hsize + iclog->ic_offset;
1358
1359         /* Round out the log write size */
1360         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1361                 /* we have a v2 stripe unit to use */
1362                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1363         } else {
1364                 count = BBTOB(BTOBB(count_init));
1365         }
1366         roundoff = count - count_init;
1367         ASSERT(roundoff >= 0);
1368         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1369                 roundoff < log->l_mp->m_sb.sb_logsunit)
1370                 || 
1371                 (log->l_mp->m_sb.sb_logsunit <= 1 && 
1372                  roundoff < BBTOB(1)));
1373
1374         /* move grant heads by roundoff in sync */
1375         spin_lock(&log->l_grant_lock);
1376         xlog_grant_add_space(log, roundoff);
1377         spin_unlock(&log->l_grant_lock);
1378
1379         /* put cycle number in every block */
1380         xlog_pack_data(log, iclog, roundoff); 
1381
1382         /* real byte length */
1383         if (v2) {
1384                 iclog->ic_header.h_len =
1385                         cpu_to_be32(iclog->ic_offset + roundoff);
1386         } else {
1387                 iclog->ic_header.h_len =
1388                         cpu_to_be32(iclog->ic_offset);
1389         }
1390
1391         bp = iclog->ic_bp;
1392         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1393         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1394         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1395
1396         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1397
1398         /* Do we need to split this write into 2 parts? */
1399         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1400                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1401                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1402                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1403         } else {
1404                 iclog->ic_bwritecnt = 1;
1405         }
1406         XFS_BUF_SET_COUNT(bp, count);
1407         XFS_BUF_SET_FSPRIVATE(bp, iclog);       /* save for later */
1408         XFS_BUF_ZEROFLAGS(bp);
1409         XFS_BUF_BUSY(bp);
1410         XFS_BUF_ASYNC(bp);
1411         bp->b_flags |= XBF_LOG_BUFFER;
1412
1413         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1414                 XFS_BUF_ORDERED(bp);
1415
1416         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1417         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1418
1419         xlog_verify_iclog(log, iclog, count, B_TRUE);
1420
1421         /* account for log which doesn't start at block #0 */
1422         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1423         /*
1424          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1425          * is shutting down.
1426          */
1427         XFS_BUF_WRITE(bp);
1428
1429         if ((error = xlog_bdstrat(bp))) {
1430                 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1431                                   XFS_BUF_ADDR(bp));
1432                 return error;
1433         }
1434         if (split) {
1435                 bp = iclog->ic_log->l_xbuf;
1436                 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1437                                                         (unsigned long)1);
1438                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1439                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1440                 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1441                                             (__psint_t)count), split);
1442                 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1443                 XFS_BUF_ZEROFLAGS(bp);
1444                 XFS_BUF_BUSY(bp);
1445                 XFS_BUF_ASYNC(bp);
1446                 bp->b_flags |= XBF_LOG_BUFFER;
1447                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1448                         XFS_BUF_ORDERED(bp);
1449                 dptr = XFS_BUF_PTR(bp);
1450                 /*
1451                  * Bump the cycle numbers at the start of each block
1452                  * since this part of the buffer is at the start of
1453                  * a new cycle.  Watch out for the header magic number
1454                  * case, though.
1455                  */
1456                 for (i = 0; i < split; i += BBSIZE) {
1457                         be32_add_cpu((__be32 *)dptr, 1);
1458                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1459                                 be32_add_cpu((__be32 *)dptr, 1);
1460                         dptr += BBSIZE;
1461                 }
1462
1463                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1464                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1465
1466                 /* account for internal log which doesn't start at block #0 */
1467                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1468                 XFS_BUF_WRITE(bp);
1469                 if ((error = xlog_bdstrat(bp))) {
1470                         xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1471                                           bp, XFS_BUF_ADDR(bp));
1472                         return error;
1473                 }
1474         }
1475         return 0;
1476 }       /* xlog_sync */
1477
1478
1479 /*
1480  * Deallocate a log structure
1481  */
1482 STATIC void
1483 xlog_dealloc_log(xlog_t *log)
1484 {
1485         xlog_in_core_t  *iclog, *next_iclog;
1486         int             i;
1487
1488         xlog_cil_destroy(log);
1489
1490         iclog = log->l_iclog;
1491         for (i=0; i<log->l_iclog_bufs; i++) {
1492                 sv_destroy(&iclog->ic_force_wait);
1493                 sv_destroy(&iclog->ic_write_wait);
1494                 xfs_buf_free(iclog->ic_bp);
1495                 next_iclog = iclog->ic_next;
1496                 kmem_free(iclog);
1497                 iclog = next_iclog;
1498         }
1499         spinlock_destroy(&log->l_icloglock);
1500         spinlock_destroy(&log->l_grant_lock);
1501
1502         xfs_buf_free(log->l_xbuf);
1503         log->l_mp->m_log = NULL;
1504         kmem_free(log);
1505 }       /* xlog_dealloc_log */
1506
1507 /*
1508  * Update counters atomically now that memcpy is done.
1509  */
1510 /* ARGSUSED */
1511 static inline void
1512 xlog_state_finish_copy(xlog_t           *log,
1513                        xlog_in_core_t   *iclog,
1514                        int              record_cnt,
1515                        int              copy_bytes)
1516 {
1517         spin_lock(&log->l_icloglock);
1518
1519         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1520         iclog->ic_offset += copy_bytes;
1521
1522         spin_unlock(&log->l_icloglock);
1523 }       /* xlog_state_finish_copy */
1524
1525
1526
1527
1528 /*
1529  * print out info relating to regions written which consume
1530  * the reservation
1531  */
1532 void
1533 xlog_print_tic_res(
1534         struct xfs_mount        *mp,
1535         struct xlog_ticket      *ticket)
1536 {
1537         uint i;
1538         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1539
1540         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1541         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1542             "bformat",
1543             "bchunk",
1544             "efi_format",
1545             "efd_format",
1546             "iformat",
1547             "icore",
1548             "iext",
1549             "ibroot",
1550             "ilocal",
1551             "iattr_ext",
1552             "iattr_broot",
1553             "iattr_local",
1554             "qformat",
1555             "dquot",
1556             "quotaoff",
1557             "LR header",
1558             "unmount",
1559             "commit",
1560             "trans header"
1561         };
1562         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1563             "SETATTR_NOT_SIZE",
1564             "SETATTR_SIZE",
1565             "INACTIVE",
1566             "CREATE",
1567             "CREATE_TRUNC",
1568             "TRUNCATE_FILE",
1569             "REMOVE",
1570             "LINK",
1571             "RENAME",
1572             "MKDIR",
1573             "RMDIR",
1574             "SYMLINK",
1575             "SET_DMATTRS",
1576             "GROWFS",
1577             "STRAT_WRITE",
1578             "DIOSTRAT",
1579             "WRITE_SYNC",
1580             "WRITEID",
1581             "ADDAFORK",
1582             "ATTRINVAL",
1583             "ATRUNCATE",
1584             "ATTR_SET",
1585             "ATTR_RM",
1586             "ATTR_FLAG",
1587             "CLEAR_AGI_BUCKET",
1588             "QM_SBCHANGE",
1589             "DUMMY1",
1590             "DUMMY2",
1591             "QM_QUOTAOFF",
1592             "QM_DQALLOC",
1593             "QM_SETQLIM",
1594             "QM_DQCLUSTER",
1595             "QM_QINOCREATE",
1596             "QM_QUOTAOFF_END",
1597             "SB_UNIT",
1598             "FSYNC_TS",
1599             "GROWFSRT_ALLOC",
1600             "GROWFSRT_ZERO",
1601             "GROWFSRT_FREE",
1602             "SWAPEXT"
1603         };
1604
1605         xfs_fs_cmn_err(CE_WARN, mp,
1606                         "xfs_log_write: reservation summary:\n"
1607                         "  trans type  = %s (%u)\n"
1608                         "  unit res    = %d bytes\n"
1609                         "  current res = %d bytes\n"
1610                         "  total reg   = %u bytes (o/flow = %u bytes)\n"
1611                         "  ophdrs      = %u (ophdr space = %u bytes)\n"
1612                         "  ophdr + reg = %u bytes\n"
1613                         "  num regions = %u\n",
1614                         ((ticket->t_trans_type <= 0 ||
1615                           ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1616                           "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1617                         ticket->t_trans_type,
1618                         ticket->t_unit_res,
1619                         ticket->t_curr_res,
1620                         ticket->t_res_arr_sum, ticket->t_res_o_flow,
1621                         ticket->t_res_num_ophdrs, ophdr_spc,
1622                         ticket->t_res_arr_sum + 
1623                         ticket->t_res_o_flow + ophdr_spc,
1624                         ticket->t_res_num);
1625
1626         for (i = 0; i < ticket->t_res_num; i++) {
1627                 uint r_type = ticket->t_res_arr[i].r_type; 
1628                 cmn_err(CE_WARN,
1629                             "region[%u]: %s - %u bytes\n",
1630                             i, 
1631                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1632                             "bad-rtype" : res_type_str[r_type-1]),
1633                             ticket->t_res_arr[i].r_len);
1634         }
1635
1636         xfs_cmn_err(XFS_PTAG_LOGRES, CE_ALERT, mp,
1637                 "xfs_log_write: reservation ran out. Need to up reservation");
1638         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1639 }
1640
1641 /*
1642  * Calculate the potential space needed by the log vector.  Each region gets
1643  * its own xlog_op_header_t and may need to be double word aligned.
1644  */
1645 static int
1646 xlog_write_calc_vec_length(
1647         struct xlog_ticket      *ticket,
1648         struct xfs_log_vec      *log_vector)
1649 {
1650         struct xfs_log_vec      *lv;
1651         int                     headers = 0;
1652         int                     len = 0;
1653         int                     i;
1654
1655         /* acct for start rec of xact */
1656         if (ticket->t_flags & XLOG_TIC_INITED)
1657                 headers++;
1658
1659         for (lv = log_vector; lv; lv = lv->lv_next) {
1660                 headers += lv->lv_niovecs;
1661
1662                 for (i = 0; i < lv->lv_niovecs; i++) {
1663                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
1664
1665                         len += vecp->i_len;
1666                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1667                 }
1668         }
1669
1670         ticket->t_res_num_ophdrs += headers;
1671         len += headers * sizeof(struct xlog_op_header);
1672
1673         return len;
1674 }
1675
1676 /*
1677  * If first write for transaction, insert start record  We can't be trying to
1678  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1679  */
1680 static int
1681 xlog_write_start_rec(
1682         struct xlog_op_header   *ophdr,
1683         struct xlog_ticket      *ticket)
1684 {
1685         if (!(ticket->t_flags & XLOG_TIC_INITED))
1686                 return 0;
1687
1688         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
1689         ophdr->oh_clientid = ticket->t_clientid;
1690         ophdr->oh_len = 0;
1691         ophdr->oh_flags = XLOG_START_TRANS;
1692         ophdr->oh_res2 = 0;
1693
1694         ticket->t_flags &= ~XLOG_TIC_INITED;
1695
1696         return sizeof(struct xlog_op_header);
1697 }
1698
1699 static xlog_op_header_t *
1700 xlog_write_setup_ophdr(
1701         struct log              *log,
1702         struct xlog_op_header   *ophdr,
1703         struct xlog_ticket      *ticket,
1704         uint                    flags)
1705 {
1706         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1707         ophdr->oh_clientid = ticket->t_clientid;
1708         ophdr->oh_res2 = 0;
1709
1710         /* are we copying a commit or unmount record? */
1711         ophdr->oh_flags = flags;
1712
1713         /*
1714          * We've seen logs corrupted with bad transaction client ids.  This
1715          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
1716          * and shut down the filesystem.
1717          */
1718         switch (ophdr->oh_clientid)  {
1719         case XFS_TRANSACTION:
1720         case XFS_VOLUME:
1721         case XFS_LOG:
1722                 break;
1723         default:
1724                 xfs_fs_cmn_err(CE_WARN, log->l_mp,
1725                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1726                         ophdr->oh_clientid, ticket);
1727                 return NULL;
1728         }
1729
1730         return ophdr;
1731 }
1732
1733 /*
1734  * Set up the parameters of the region copy into the log. This has
1735  * to handle region write split across multiple log buffers - this
1736  * state is kept external to this function so that this code can
1737  * can be written in an obvious, self documenting manner.
1738  */
1739 static int
1740 xlog_write_setup_copy(
1741         struct xlog_ticket      *ticket,
1742         struct xlog_op_header   *ophdr,
1743         int                     space_available,
1744         int                     space_required,
1745         int                     *copy_off,
1746         int                     *copy_len,
1747         int                     *last_was_partial_copy,
1748         int                     *bytes_consumed)
1749 {
1750         int                     still_to_copy;
1751
1752         still_to_copy = space_required - *bytes_consumed;
1753         *copy_off = *bytes_consumed;
1754
1755         if (still_to_copy <= space_available) {
1756                 /* write of region completes here */
1757                 *copy_len = still_to_copy;
1758                 ophdr->oh_len = cpu_to_be32(*copy_len);
1759                 if (*last_was_partial_copy)
1760                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1761                 *last_was_partial_copy = 0;
1762                 *bytes_consumed = 0;
1763                 return 0;
1764         }
1765
1766         /* partial write of region, needs extra log op header reservation */
1767         *copy_len = space_available;
1768         ophdr->oh_len = cpu_to_be32(*copy_len);
1769         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1770         if (*last_was_partial_copy)
1771                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1772         *bytes_consumed += *copy_len;
1773         (*last_was_partial_copy)++;
1774
1775         /* account for new log op header */
1776         ticket->t_curr_res -= sizeof(struct xlog_op_header);
1777         ticket->t_res_num_ophdrs++;
1778
1779         return sizeof(struct xlog_op_header);
1780 }
1781
1782 static int
1783 xlog_write_copy_finish(
1784         struct log              *log,
1785         struct xlog_in_core     *iclog,
1786         uint                    flags,
1787         int                     *record_cnt,
1788         int                     *data_cnt,
1789         int                     *partial_copy,
1790         int                     *partial_copy_len,
1791         int                     log_offset,
1792         struct xlog_in_core     **commit_iclog)
1793 {
1794         if (*partial_copy) {
1795                 /*
1796                  * This iclog has already been marked WANT_SYNC by
1797                  * xlog_state_get_iclog_space.
1798                  */
1799                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1800                 *record_cnt = 0;
1801                 *data_cnt = 0;
1802                 return xlog_state_release_iclog(log, iclog);
1803         }
1804
1805         *partial_copy = 0;
1806         *partial_copy_len = 0;
1807
1808         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1809                 /* no more space in this iclog - push it. */
1810                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1811                 *record_cnt = 0;
1812                 *data_cnt = 0;
1813
1814                 spin_lock(&log->l_icloglock);
1815                 xlog_state_want_sync(log, iclog);
1816                 spin_unlock(&log->l_icloglock);
1817
1818                 if (!commit_iclog)
1819                         return xlog_state_release_iclog(log, iclog);
1820                 ASSERT(flags & XLOG_COMMIT_TRANS);
1821                 *commit_iclog = iclog;
1822         }
1823
1824         return 0;
1825 }
1826
1827 /*
1828  * Write some region out to in-core log
1829  *
1830  * This will be called when writing externally provided regions or when
1831  * writing out a commit record for a given transaction.
1832  *
1833  * General algorithm:
1834  *      1. Find total length of this write.  This may include adding to the
1835  *              lengths passed in.
1836  *      2. Check whether we violate the tickets reservation.
1837  *      3. While writing to this iclog
1838  *          A. Reserve as much space in this iclog as can get
1839  *          B. If this is first write, save away start lsn
1840  *          C. While writing this region:
1841  *              1. If first write of transaction, write start record
1842  *              2. Write log operation header (header per region)
1843  *              3. Find out if we can fit entire region into this iclog
1844  *              4. Potentially, verify destination memcpy ptr
1845  *              5. Memcpy (partial) region
1846  *              6. If partial copy, release iclog; otherwise, continue
1847  *                      copying more regions into current iclog
1848  *      4. Mark want sync bit (in simulation mode)
1849  *      5. Release iclog for potential flush to on-disk log.
1850  *
1851  * ERRORS:
1852  * 1.   Panic if reservation is overrun.  This should never happen since
1853  *      reservation amounts are generated internal to the filesystem.
1854  * NOTES:
1855  * 1. Tickets are single threaded data structures.
1856  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1857  *      syncing routine.  When a single log_write region needs to span
1858  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1859  *      on all log operation writes which don't contain the end of the
1860  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1861  *      operation which contains the end of the continued log_write region.
1862  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1863  *      we don't really know exactly how much space will be used.  As a result,
1864  *      we don't update ic_offset until the end when we know exactly how many
1865  *      bytes have been written out.
1866  */
1867 int
1868 xlog_write(
1869         struct log              *log,
1870         struct xfs_log_vec      *log_vector,
1871         struct xlog_ticket      *ticket,
1872         xfs_lsn_t               *start_lsn,
1873         struct xlog_in_core     **commit_iclog,
1874         uint                    flags)
1875 {
1876         struct xlog_in_core     *iclog = NULL;
1877         struct xfs_log_iovec    *vecp;
1878         struct xfs_log_vec      *lv;
1879         int                     len;
1880         int                     index;
1881         int                     partial_copy = 0;
1882         int                     partial_copy_len = 0;
1883         int                     contwr = 0;
1884         int                     record_cnt = 0;
1885         int                     data_cnt = 0;
1886         int                     error;
1887
1888         *start_lsn = 0;
1889
1890         len = xlog_write_calc_vec_length(ticket, log_vector);
1891         if (log->l_cilp) {
1892                 /*
1893                  * Region headers and bytes are already accounted for.
1894                  * We only need to take into account start records and
1895                  * split regions in this function.
1896                  */
1897                 if (ticket->t_flags & XLOG_TIC_INITED)
1898                         ticket->t_curr_res -= sizeof(xlog_op_header_t);
1899
1900                 /*
1901                  * Commit record headers need to be accounted for. These
1902                  * come in as separate writes so are easy to detect.
1903                  */
1904                 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1905                         ticket->t_curr_res -= sizeof(xlog_op_header_t);
1906         } else
1907                 ticket->t_curr_res -= len;
1908
1909         if (ticket->t_curr_res < 0)
1910                 xlog_print_tic_res(log->l_mp, ticket);
1911
1912         index = 0;
1913         lv = log_vector;
1914         vecp = lv->lv_iovecp;
1915         while (lv && index < lv->lv_niovecs) {
1916                 void            *ptr;
1917                 int             log_offset;
1918
1919                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1920                                                    &contwr, &log_offset);
1921                 if (error)
1922                         return error;
1923
1924                 ASSERT(log_offset <= iclog->ic_size - 1);
1925                 ptr = iclog->ic_datap + log_offset;
1926
1927                 /* start_lsn is the first lsn written to. That's all we need. */
1928                 if (!*start_lsn)
1929                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1930
1931                 /*
1932                  * This loop writes out as many regions as can fit in the amount
1933                  * of space which was allocated by xlog_state_get_iclog_space().
1934                  */
1935                 while (lv && index < lv->lv_niovecs) {
1936                         struct xfs_log_iovec    *reg = &vecp[index];
1937                         struct xlog_op_header   *ophdr;
1938                         int                     start_rec_copy;
1939                         int                     copy_len;
1940                         int                     copy_off;
1941
1942                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1943                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1944
1945                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
1946                         if (start_rec_copy) {
1947                                 record_cnt++;
1948                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1949                                                    start_rec_copy);
1950                         }
1951
1952                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1953                         if (!ophdr)
1954                                 return XFS_ERROR(EIO);
1955
1956                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
1957                                            sizeof(struct xlog_op_header));
1958
1959                         len += xlog_write_setup_copy(ticket, ophdr,
1960                                                      iclog->ic_size-log_offset,
1961                                                      reg->i_len,
1962                                                      &copy_off, &copy_len,
1963                                                      &partial_copy,
1964                                                      &partial_copy_len);
1965                         xlog_verify_dest_ptr(log, ptr);
1966
1967                         /* copy region */
1968                         ASSERT(copy_len >= 0);
1969                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
1970                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1971
1972                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1973                         record_cnt++;
1974                         data_cnt += contwr ? copy_len : 0;
1975
1976                         error = xlog_write_copy_finish(log, iclog, flags,
1977                                                        &record_cnt, &data_cnt,
1978                                                        &partial_copy,
1979                                                        &partial_copy_len,
1980                                                        log_offset,
1981                                                        commit_iclog);
1982                         if (error)
1983                                 return error;
1984
1985                         /*
1986                          * if we had a partial copy, we need to get more iclog
1987                          * space but we don't want to increment the region
1988                          * index because there is still more is this region to
1989                          * write.
1990                          *
1991                          * If we completed writing this region, and we flushed
1992                          * the iclog (indicated by resetting of the record
1993                          * count), then we also need to get more log space. If
1994                          * this was the last record, though, we are done and
1995                          * can just return.
1996                          */
1997                         if (partial_copy)
1998                                 break;
1999
2000                         if (++index == lv->lv_niovecs) {
2001                                 lv = lv->lv_next;
2002                                 index = 0;
2003                                 if (lv)
2004                                         vecp = lv->lv_iovecp;
2005                         }
2006                         if (record_cnt == 0) {
2007                                 if (!lv)
2008                                         return 0;
2009                                 break;
2010                         }
2011                 }
2012         }
2013
2014         ASSERT(len == 0);
2015
2016         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2017         if (!commit_iclog)
2018                 return xlog_state_release_iclog(log, iclog);
2019
2020         ASSERT(flags & XLOG_COMMIT_TRANS);
2021         *commit_iclog = iclog;
2022         return 0;
2023 }
2024
2025
2026 /*****************************************************************************
2027  *
2028  *              State Machine functions
2029  *
2030  *****************************************************************************
2031  */
2032
2033 /* Clean iclogs starting from the head.  This ordering must be
2034  * maintained, so an iclog doesn't become ACTIVE beyond one that
2035  * is SYNCING.  This is also required to maintain the notion that we use
2036  * a ordered wait queue to hold off would be writers to the log when every
2037  * iclog is trying to sync to disk.
2038  *
2039  * State Change: DIRTY -> ACTIVE
2040  */
2041 STATIC void
2042 xlog_state_clean_log(xlog_t *log)
2043 {
2044         xlog_in_core_t  *iclog;
2045         int changed = 0;
2046
2047         iclog = log->l_iclog;
2048         do {
2049                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2050                         iclog->ic_state = XLOG_STATE_ACTIVE;
2051                         iclog->ic_offset       = 0;
2052                         ASSERT(iclog->ic_callback == NULL);
2053                         /*
2054                          * If the number of ops in this iclog indicate it just
2055                          * contains the dummy transaction, we can
2056                          * change state into IDLE (the second time around).
2057                          * Otherwise we should change the state into
2058                          * NEED a dummy.
2059                          * We don't need to cover the dummy.
2060                          */
2061                         if (!changed &&
2062                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2063                                         XLOG_COVER_OPS)) {
2064                                 changed = 1;
2065                         } else {
2066                                 /*
2067                                  * We have two dirty iclogs so start over
2068                                  * This could also be num of ops indicates
2069                                  * this is not the dummy going out.
2070                                  */
2071                                 changed = 2;
2072                         }
2073                         iclog->ic_header.h_num_logops = 0;
2074                         memset(iclog->ic_header.h_cycle_data, 0,
2075                               sizeof(iclog->ic_header.h_cycle_data));
2076                         iclog->ic_header.h_lsn = 0;
2077                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2078                         /* do nothing */;
2079                 else
2080                         break;  /* stop cleaning */
2081                 iclog = iclog->ic_next;
2082         } while (iclog != log->l_iclog);
2083
2084         /* log is locked when we are called */
2085         /*
2086          * Change state for the dummy log recording.
2087          * We usually go to NEED. But we go to NEED2 if the changed indicates
2088          * we are done writing the dummy record.
2089          * If we are done with the second dummy recored (DONE2), then
2090          * we go to IDLE.
2091          */
2092         if (changed) {
2093                 switch (log->l_covered_state) {
2094                 case XLOG_STATE_COVER_IDLE:
2095                 case XLOG_STATE_COVER_NEED:
2096                 case XLOG_STATE_COVER_NEED2:
2097                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2098                         break;
2099
2100                 case XLOG_STATE_COVER_DONE:
2101                         if (changed == 1)
2102                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2103                         else
2104                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2105                         break;
2106
2107                 case XLOG_STATE_COVER_DONE2:
2108                         if (changed == 1)
2109                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2110                         else
2111                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2112                         break;
2113
2114                 default:
2115                         ASSERT(0);
2116                 }
2117         }
2118 }       /* xlog_state_clean_log */
2119
2120 STATIC xfs_lsn_t
2121 xlog_get_lowest_lsn(
2122         xlog_t          *log)
2123 {
2124         xlog_in_core_t  *lsn_log;
2125         xfs_lsn_t       lowest_lsn, lsn;
2126
2127         lsn_log = log->l_iclog;
2128         lowest_lsn = 0;
2129         do {
2130             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2131                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2132                 if ((lsn && !lowest_lsn) ||
2133                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2134                         lowest_lsn = lsn;
2135                 }
2136             }
2137             lsn_log = lsn_log->ic_next;
2138         } while (lsn_log != log->l_iclog);
2139         return lowest_lsn;
2140 }
2141
2142
2143 STATIC void
2144 xlog_state_do_callback(
2145         xlog_t          *log,
2146         int             aborted,
2147         xlog_in_core_t  *ciclog)
2148 {
2149         xlog_in_core_t     *iclog;
2150         xlog_in_core_t     *first_iclog;        /* used to know when we've
2151                                                  * processed all iclogs once */
2152         xfs_log_callback_t *cb, *cb_next;
2153         int                flushcnt = 0;
2154         xfs_lsn_t          lowest_lsn;
2155         int                ioerrors;    /* counter: iclogs with errors */
2156         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2157         int                funcdidcallbacks; /* flag: function did callbacks */
2158         int                repeats;     /* for issuing console warnings if
2159                                          * looping too many times */
2160         int                wake = 0;
2161
2162         spin_lock(&log->l_icloglock);
2163         first_iclog = iclog = log->l_iclog;
2164         ioerrors = 0;
2165         funcdidcallbacks = 0;
2166         repeats = 0;
2167
2168         do {
2169                 /*
2170                  * Scan all iclogs starting with the one pointed to by the
2171                  * log.  Reset this starting point each time the log is
2172                  * unlocked (during callbacks).
2173                  *
2174                  * Keep looping through iclogs until one full pass is made
2175                  * without running any callbacks.
2176                  */
2177                 first_iclog = log->l_iclog;
2178                 iclog = log->l_iclog;
2179                 loopdidcallbacks = 0;
2180                 repeats++;
2181
2182                 do {
2183
2184                         /* skip all iclogs in the ACTIVE & DIRTY states */
2185                         if (iclog->ic_state &
2186                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2187                                 iclog = iclog->ic_next;
2188                                 continue;
2189                         }
2190
2191                         /*
2192                          * Between marking a filesystem SHUTDOWN and stopping
2193                          * the log, we do flush all iclogs to disk (if there
2194                          * wasn't a log I/O error). So, we do want things to
2195                          * go smoothly in case of just a SHUTDOWN  w/o a
2196                          * LOG_IO_ERROR.
2197                          */
2198                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2199                                 /*
2200                                  * Can only perform callbacks in order.  Since
2201                                  * this iclog is not in the DONE_SYNC/
2202                                  * DO_CALLBACK state, we skip the rest and
2203                                  * just try to clean up.  If we set our iclog
2204                                  * to DO_CALLBACK, we will not process it when
2205                                  * we retry since a previous iclog is in the
2206                                  * CALLBACK and the state cannot change since
2207                                  * we are holding the l_icloglock.
2208                                  */
2209                                 if (!(iclog->ic_state &
2210                                         (XLOG_STATE_DONE_SYNC |
2211                                                  XLOG_STATE_DO_CALLBACK))) {
2212                                         if (ciclog && (ciclog->ic_state ==
2213                                                         XLOG_STATE_DONE_SYNC)) {
2214                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2215                                         }
2216                                         break;
2217                                 }
2218                                 /*
2219                                  * We now have an iclog that is in either the
2220                                  * DO_CALLBACK or DONE_SYNC states. The other
2221                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2222                                  * caught by the above if and are going to
2223                                  * clean (i.e. we aren't doing their callbacks)
2224                                  * see the above if.
2225                                  */
2226
2227                                 /*
2228                                  * We will do one more check here to see if we
2229                                  * have chased our tail around.
2230                                  */
2231
2232                                 lowest_lsn = xlog_get_lowest_lsn(log);
2233                                 if (lowest_lsn &&
2234                                     XFS_LSN_CMP(lowest_lsn,
2235                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2236                                         iclog = iclog->ic_next;
2237                                         continue; /* Leave this iclog for
2238                                                    * another thread */
2239                                 }
2240
2241                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2242
2243                                 spin_unlock(&log->l_icloglock);
2244
2245                                 /* l_last_sync_lsn field protected by
2246                                  * l_grant_lock. Don't worry about iclog's lsn.
2247                                  * No one else can be here except us.
2248                                  */
2249                                 spin_lock(&log->l_grant_lock);
2250                                 ASSERT(XFS_LSN_CMP(log->l_last_sync_lsn,
2251                                        be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2252                                 log->l_last_sync_lsn =
2253                                         be64_to_cpu(iclog->ic_header.h_lsn);
2254                                 spin_unlock(&log->l_grant_lock);
2255
2256                         } else {
2257                                 spin_unlock(&log->l_icloglock);
2258                                 ioerrors++;
2259                         }
2260
2261                         /*
2262                          * Keep processing entries in the callback list until
2263                          * we come around and it is empty.  We need to
2264                          * atomically see that the list is empty and change the
2265                          * state to DIRTY so that we don't miss any more
2266                          * callbacks being added.
2267                          */
2268                         spin_lock(&iclog->ic_callback_lock);
2269                         cb = iclog->ic_callback;
2270                         while (cb) {
2271                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2272                                 iclog->ic_callback = NULL;
2273                                 spin_unlock(&iclog->ic_callback_lock);
2274
2275                                 /* perform callbacks in the order given */
2276                                 for (; cb; cb = cb_next) {
2277                                         cb_next = cb->cb_next;
2278                                         cb->cb_func(cb->cb_arg, aborted);
2279                                 }
2280                                 spin_lock(&iclog->ic_callback_lock);
2281                                 cb = iclog->ic_callback;
2282                         }
2283
2284                         loopdidcallbacks++;
2285                         funcdidcallbacks++;
2286
2287                         spin_lock(&log->l_icloglock);
2288                         ASSERT(iclog->ic_callback == NULL);
2289                         spin_unlock(&iclog->ic_callback_lock);
2290                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2291                                 iclog->ic_state = XLOG_STATE_DIRTY;
2292
2293                         /*
2294                          * Transition from DIRTY to ACTIVE if applicable.
2295                          * NOP if STATE_IOERROR.
2296                          */
2297                         xlog_state_clean_log(log);
2298
2299                         /* wake up threads waiting in xfs_log_force() */
2300                         sv_broadcast(&iclog->ic_force_wait);
2301
2302                         iclog = iclog->ic_next;
2303                 } while (first_iclog != iclog);
2304
2305                 if (repeats > 5000) {
2306                         flushcnt += repeats;
2307                         repeats = 0;
2308                         xfs_fs_cmn_err(CE_WARN, log->l_mp,
2309                                 "%s: possible infinite loop (%d iterations)",
2310                                 __func__, flushcnt);
2311                 }
2312         } while (!ioerrors && loopdidcallbacks);
2313
2314         /*
2315          * make one last gasp attempt to see if iclogs are being left in
2316          * limbo..
2317          */
2318 #ifdef DEBUG
2319         if (funcdidcallbacks) {
2320                 first_iclog = iclog = log->l_iclog;
2321                 do {
2322                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2323                         /*
2324                          * Terminate the loop if iclogs are found in states
2325                          * which will cause other threads to clean up iclogs.
2326                          *
2327                          * SYNCING - i/o completion will go through logs
2328                          * DONE_SYNC - interrupt thread should be waiting for
2329                          *              l_icloglock
2330                          * IOERROR - give up hope all ye who enter here
2331                          */
2332                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2333                             iclog->ic_state == XLOG_STATE_SYNCING ||
2334                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2335                             iclog->ic_state == XLOG_STATE_IOERROR )
2336                                 break;
2337                         iclog = iclog->ic_next;
2338                 } while (first_iclog != iclog);
2339         }
2340 #endif
2341
2342         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2343                 wake = 1;
2344         spin_unlock(&log->l_icloglock);
2345
2346         if (wake)
2347                 sv_broadcast(&log->l_flush_wait);
2348 }
2349
2350
2351 /*
2352  * Finish transitioning this iclog to the dirty state.
2353  *
2354  * Make sure that we completely execute this routine only when this is
2355  * the last call to the iclog.  There is a good chance that iclog flushes,
2356  * when we reach the end of the physical log, get turned into 2 separate
2357  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2358  * routine.  By using the reference count bwritecnt, we guarantee that only
2359  * the second completion goes through.
2360  *
2361  * Callbacks could take time, so they are done outside the scope of the
2362  * global state machine log lock.
2363  */
2364 STATIC void
2365 xlog_state_done_syncing(
2366         xlog_in_core_t  *iclog,
2367         int             aborted)
2368 {
2369         xlog_t             *log = iclog->ic_log;
2370
2371         spin_lock(&log->l_icloglock);
2372
2373         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2374                iclog->ic_state == XLOG_STATE_IOERROR);
2375         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2376         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2377
2378
2379         /*
2380          * If we got an error, either on the first buffer, or in the case of
2381          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2382          * and none should ever be attempted to be written to disk
2383          * again.
2384          */
2385         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2386                 if (--iclog->ic_bwritecnt == 1) {
2387                         spin_unlock(&log->l_icloglock);
2388                         return;
2389                 }
2390                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2391         }
2392
2393         /*
2394          * Someone could be sleeping prior to writing out the next
2395          * iclog buffer, we wake them all, one will get to do the
2396          * I/O, the others get to wait for the result.
2397          */
2398         sv_broadcast(&iclog->ic_write_wait);
2399         spin_unlock(&log->l_icloglock);
2400         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2401 }       /* xlog_state_done_syncing */
2402
2403
2404 /*
2405  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2406  * sleep.  We wait on the flush queue on the head iclog as that should be
2407  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2408  * we will wait here and all new writes will sleep until a sync completes.
2409  *
2410  * The in-core logs are used in a circular fashion. They are not used
2411  * out-of-order even when an iclog past the head is free.
2412  *
2413  * return:
2414  *      * log_offset where xlog_write() can start writing into the in-core
2415  *              log's data space.
2416  *      * in-core log pointer to which xlog_write() should write.
2417  *      * boolean indicating this is a continued write to an in-core log.
2418  *              If this is the last write, then the in-core log's offset field
2419  *              needs to be incremented, depending on the amount of data which
2420  *              is copied.
2421  */
2422 STATIC int
2423 xlog_state_get_iclog_space(xlog_t         *log,
2424                            int            len,
2425                            xlog_in_core_t **iclogp,
2426                            xlog_ticket_t  *ticket,
2427                            int            *continued_write,
2428                            int            *logoffsetp)
2429 {
2430         int               log_offset;
2431         xlog_rec_header_t *head;
2432         xlog_in_core_t    *iclog;
2433         int               error;
2434
2435 restart:
2436         spin_lock(&log->l_icloglock);
2437         if (XLOG_FORCED_SHUTDOWN(log)) {
2438                 spin_unlock(&log->l_icloglock);
2439                 return XFS_ERROR(EIO);
2440         }
2441
2442         iclog = log->l_iclog;
2443         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2444                 XFS_STATS_INC(xs_log_noiclogs);
2445
2446                 /* Wait for log writes to have flushed */
2447                 sv_wait(&log->l_flush_wait, 0, &log->l_icloglock, 0);
2448                 goto restart;
2449         }
2450
2451         head = &iclog->ic_header;
2452
2453         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2454         log_offset = iclog->ic_offset;
2455
2456         /* On the 1st write to an iclog, figure out lsn.  This works
2457          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2458          * committing to.  If the offset is set, that's how many blocks
2459          * must be written.
2460          */
2461         if (log_offset == 0) {
2462                 ticket->t_curr_res -= log->l_iclog_hsize;
2463                 xlog_tic_add_region(ticket,
2464                                     log->l_iclog_hsize,
2465                                     XLOG_REG_TYPE_LRHEADER);
2466                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2467                 head->h_lsn = cpu_to_be64(
2468                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2469                 ASSERT(log->l_curr_block >= 0);
2470         }
2471
2472         /* If there is enough room to write everything, then do it.  Otherwise,
2473          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2474          * bit is on, so this will get flushed out.  Don't update ic_offset
2475          * until you know exactly how many bytes get copied.  Therefore, wait
2476          * until later to update ic_offset.
2477          *
2478          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2479          * can fit into remaining data section.
2480          */
2481         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2482                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2483
2484                 /*
2485                  * If I'm the only one writing to this iclog, sync it to disk.
2486                  * We need to do an atomic compare and decrement here to avoid
2487                  * racing with concurrent atomic_dec_and_lock() calls in
2488                  * xlog_state_release_iclog() when there is more than one
2489                  * reference to the iclog.
2490                  */
2491                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2492                         /* we are the only one */
2493                         spin_unlock(&log->l_icloglock);
2494                         error = xlog_state_release_iclog(log, iclog);
2495                         if (error)
2496                                 return error;
2497                 } else {
2498                         spin_unlock(&log->l_icloglock);
2499                 }
2500                 goto restart;
2501         }
2502
2503         /* Do we have enough room to write the full amount in the remainder
2504          * of this iclog?  Or must we continue a write on the next iclog and
2505          * mark this iclog as completely taken?  In the case where we switch
2506          * iclogs (to mark it taken), this particular iclog will release/sync
2507          * to disk in xlog_write().
2508          */
2509         if (len <= iclog->ic_size - iclog->ic_offset) {
2510                 *continued_write = 0;
2511                 iclog->ic_offset += len;
2512         } else {
2513                 *continued_write = 1;
2514                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2515         }
2516         *iclogp = iclog;
2517
2518         ASSERT(iclog->ic_offset <= iclog->ic_size);
2519         spin_unlock(&log->l_icloglock);
2520
2521         *logoffsetp = log_offset;
2522         return 0;
2523 }       /* xlog_state_get_iclog_space */
2524
2525 /*
2526  * Atomically get the log space required for a log ticket.
2527  *
2528  * Once a ticket gets put onto the reserveq, it will only return after
2529  * the needed reservation is satisfied.
2530  */
2531 STATIC int
2532 xlog_grant_log_space(xlog_t        *log,
2533                      xlog_ticket_t *tic)
2534 {
2535         int              free_bytes;
2536         int              need_bytes;
2537 #ifdef DEBUG
2538         xfs_lsn_t        tail_lsn;
2539 #endif
2540
2541
2542 #ifdef DEBUG
2543         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2544                 panic("grant Recovery problem");
2545 #endif
2546
2547         /* Is there space or do we need to sleep? */
2548         spin_lock(&log->l_grant_lock);
2549
2550         trace_xfs_log_grant_enter(log, tic);
2551
2552         /* something is already sleeping; insert new transaction at end */
2553         if (log->l_reserve_headq) {
2554                 xlog_ins_ticketq(&log->l_reserve_headq, tic);
2555
2556                 trace_xfs_log_grant_sleep1(log, tic);
2557
2558                 /*
2559                  * Gotta check this before going to sleep, while we're
2560                  * holding the grant lock.
2561                  */
2562                 if (XLOG_FORCED_SHUTDOWN(log))
2563                         goto error_return;
2564
2565                 XFS_STATS_INC(xs_sleep_logspace);
2566                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2567                 /*
2568                  * If we got an error, and the filesystem is shutting down,
2569                  * we'll catch it down below. So just continue...
2570                  */
2571                 trace_xfs_log_grant_wake1(log, tic);
2572                 spin_lock(&log->l_grant_lock);
2573         }
2574         if (tic->t_flags & XFS_LOG_PERM_RESERV)
2575                 need_bytes = tic->t_unit_res*tic->t_ocnt;
2576         else
2577                 need_bytes = tic->t_unit_res;
2578
2579 redo:
2580         if (XLOG_FORCED_SHUTDOWN(log))
2581                 goto error_return;
2582
2583         free_bytes = xlog_space_left(log, log->l_grant_reserve_cycle,
2584                                      log->l_grant_reserve_bytes);
2585         if (free_bytes < need_bytes) {
2586                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2587                         xlog_ins_ticketq(&log->l_reserve_headq, tic);
2588
2589                 trace_xfs_log_grant_sleep2(log, tic);
2590
2591                 spin_unlock(&log->l_grant_lock);
2592                 xlog_grant_push_ail(log->l_mp, need_bytes);
2593                 spin_lock(&log->l_grant_lock);
2594
2595                 XFS_STATS_INC(xs_sleep_logspace);
2596                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2597
2598                 spin_lock(&log->l_grant_lock);
2599                 if (XLOG_FORCED_SHUTDOWN(log))
2600                         goto error_return;
2601
2602                 trace_xfs_log_grant_wake2(log, tic);
2603
2604                 goto redo;
2605         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2606                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2607
2608         /* we've got enough space */
2609         xlog_grant_add_space(log, need_bytes);
2610 #ifdef DEBUG
2611         tail_lsn = log->l_tail_lsn;
2612         /*
2613          * Check to make sure the grant write head didn't just over lap the
2614          * tail.  If the cycles are the same, we can't be overlapping.
2615          * Otherwise, make sure that the cycles differ by exactly one and
2616          * check the byte count.
2617          */
2618         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2619                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2620                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2621         }
2622 #endif
2623         trace_xfs_log_grant_exit(log, tic);
2624         xlog_verify_grant_head(log, 1);
2625         spin_unlock(&log->l_grant_lock);
2626         return 0;
2627
2628  error_return:
2629         if (tic->t_flags & XLOG_TIC_IN_Q)
2630                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2631
2632         trace_xfs_log_grant_error(log, tic);
2633
2634         /*
2635          * If we are failing, make sure the ticket doesn't have any
2636          * current reservations. We don't want to add this back when
2637          * the ticket/transaction gets cancelled.
2638          */
2639         tic->t_curr_res = 0;
2640         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2641         spin_unlock(&log->l_grant_lock);
2642         return XFS_ERROR(EIO);
2643 }       /* xlog_grant_log_space */
2644
2645
2646 /*
2647  * Replenish the byte reservation required by moving the grant write head.
2648  *
2649  *
2650  */
2651 STATIC int
2652 xlog_regrant_write_log_space(xlog_t        *log,
2653                              xlog_ticket_t *tic)
2654 {
2655         int             free_bytes, need_bytes;
2656         xlog_ticket_t   *ntic;
2657 #ifdef DEBUG
2658         xfs_lsn_t       tail_lsn;
2659 #endif
2660
2661         tic->t_curr_res = tic->t_unit_res;
2662         xlog_tic_reset_res(tic);
2663
2664         if (tic->t_cnt > 0)
2665                 return 0;
2666
2667 #ifdef DEBUG
2668         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2669                 panic("regrant Recovery problem");
2670 #endif
2671
2672         spin_lock(&log->l_grant_lock);
2673
2674         trace_xfs_log_regrant_write_enter(log, tic);
2675
2676         if (XLOG_FORCED_SHUTDOWN(log))
2677                 goto error_return;
2678
2679         /* If there are other waiters on the queue then give them a
2680          * chance at logspace before us. Wake up the first waiters,
2681          * if we do not wake up all the waiters then go to sleep waiting
2682          * for more free space, otherwise try to get some space for
2683          * this transaction.
2684          */
2685         need_bytes = tic->t_unit_res;
2686         if ((ntic = log->l_write_headq)) {
2687                 free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2688                                              log->l_grant_write_bytes);
2689                 do {
2690                         ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2691
2692                         if (free_bytes < ntic->t_unit_res)
2693                                 break;
2694                         free_bytes -= ntic->t_unit_res;
2695                         sv_signal(&ntic->t_wait);
2696                         ntic = ntic->t_next;
2697                 } while (ntic != log->l_write_headq);
2698
2699                 if (ntic != log->l_write_headq) {
2700                         if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2701                                 xlog_ins_ticketq(&log->l_write_headq, tic);
2702
2703                         trace_xfs_log_regrant_write_sleep1(log, tic);
2704
2705                         spin_unlock(&log->l_grant_lock);
2706                         xlog_grant_push_ail(log->l_mp, need_bytes);
2707                         spin_lock(&log->l_grant_lock);
2708
2709                         XFS_STATS_INC(xs_sleep_logspace);
2710                         sv_wait(&tic->t_wait, PINOD|PLTWAIT,
2711                                 &log->l_grant_lock, s);
2712
2713                         /* If we're shutting down, this tic is already
2714                          * off the queue */
2715                         spin_lock(&log->l_grant_lock);
2716                         if (XLOG_FORCED_SHUTDOWN(log))
2717                                 goto error_return;
2718
2719                         trace_xfs_log_regrant_write_wake1(log, tic);
2720                 }
2721         }
2722
2723 redo:
2724         if (XLOG_FORCED_SHUTDOWN(log))
2725                 goto error_return;
2726
2727         free_bytes = xlog_space_left(log, log->l_grant_write_cycle,
2728                                      log->l_grant_write_bytes);
2729         if (free_bytes < need_bytes) {
2730                 if ((tic->t_flags & XLOG_TIC_IN_Q) == 0)
2731                         xlog_ins_ticketq(&log->l_write_headq, tic);
2732                 spin_unlock(&log->l_grant_lock);
2733                 xlog_grant_push_ail(log->l_mp, need_bytes);
2734                 spin_lock(&log->l_grant_lock);
2735
2736                 XFS_STATS_INC(xs_sleep_logspace);
2737                 trace_xfs_log_regrant_write_sleep2(log, tic);
2738
2739                 sv_wait(&tic->t_wait, PINOD|PLTWAIT, &log->l_grant_lock, s);
2740
2741                 /* If we're shutting down, this tic is already off the queue */
2742                 spin_lock(&log->l_grant_lock);
2743                 if (XLOG_FORCED_SHUTDOWN(log))
2744                         goto error_return;
2745
2746                 trace_xfs_log_regrant_write_wake2(log, tic);
2747                 goto redo;
2748         } else if (tic->t_flags & XLOG_TIC_IN_Q)
2749                 xlog_del_ticketq(&log->l_write_headq, tic);
2750
2751         /* we've got enough space */
2752         xlog_grant_add_space_write(log, need_bytes);
2753 #ifdef DEBUG
2754         tail_lsn = log->l_tail_lsn;
2755         if (CYCLE_LSN(tail_lsn) != log->l_grant_write_cycle) {
2756                 ASSERT(log->l_grant_write_cycle-1 == CYCLE_LSN(tail_lsn));
2757                 ASSERT(log->l_grant_write_bytes <= BBTOB(BLOCK_LSN(tail_lsn)));
2758         }
2759 #endif
2760
2761         trace_xfs_log_regrant_write_exit(log, tic);
2762
2763         xlog_verify_grant_head(log, 1);
2764         spin_unlock(&log->l_grant_lock);
2765         return 0;
2766
2767
2768  error_return:
2769         if (tic->t_flags & XLOG_TIC_IN_Q)
2770                 xlog_del_ticketq(&log->l_reserve_headq, tic);
2771
2772         trace_xfs_log_regrant_write_error(log, tic);
2773
2774         /*
2775          * If we are failing, make sure the ticket doesn't have any
2776          * current reservations. We don't want to add this back when
2777          * the ticket/transaction gets cancelled.
2778          */
2779         tic->t_curr_res = 0;
2780         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2781         spin_unlock(&log->l_grant_lock);
2782         return XFS_ERROR(EIO);
2783 }       /* xlog_regrant_write_log_space */
2784
2785
2786 /* The first cnt-1 times through here we don't need to
2787  * move the grant write head because the permanent
2788  * reservation has reserved cnt times the unit amount.
2789  * Release part of current permanent unit reservation and
2790  * reset current reservation to be one units worth.  Also
2791  * move grant reservation head forward.
2792  */
2793 STATIC void
2794 xlog_regrant_reserve_log_space(xlog_t        *log,
2795                                xlog_ticket_t *ticket)
2796 {
2797         trace_xfs_log_regrant_reserve_enter(log, ticket);
2798
2799         if (ticket->t_cnt > 0)
2800                 ticket->t_cnt--;
2801
2802         spin_lock(&log->l_grant_lock);
2803         xlog_grant_sub_space(log, ticket->t_curr_res);
2804         ticket->t_curr_res = ticket->t_unit_res;
2805         xlog_tic_reset_res(ticket);
2806
2807         trace_xfs_log_regrant_reserve_sub(log, ticket);
2808
2809         xlog_verify_grant_head(log, 1);
2810
2811         /* just return if we still have some of the pre-reserved space */
2812         if (ticket->t_cnt > 0) {
2813                 spin_unlock(&log->l_grant_lock);
2814                 return;
2815         }
2816
2817         xlog_grant_add_space_reserve(log, ticket->t_unit_res);
2818
2819         trace_xfs_log_regrant_reserve_exit(log, ticket);
2820
2821         xlog_verify_grant_head(log, 0);
2822         spin_unlock(&log->l_grant_lock);
2823         ticket->t_curr_res = ticket->t_unit_res;
2824         xlog_tic_reset_res(ticket);
2825 }       /* xlog_regrant_reserve_log_space */
2826
2827
2828 /*
2829  * Give back the space left from a reservation.
2830  *
2831  * All the information we need to make a correct determination of space left
2832  * is present.  For non-permanent reservations, things are quite easy.  The
2833  * count should have been decremented to zero.  We only need to deal with the
2834  * space remaining in the current reservation part of the ticket.  If the
2835  * ticket contains a permanent reservation, there may be left over space which
2836  * needs to be released.  A count of N means that N-1 refills of the current
2837  * reservation can be done before we need to ask for more space.  The first
2838  * one goes to fill up the first current reservation.  Once we run out of
2839  * space, the count will stay at zero and the only space remaining will be
2840  * in the current reservation field.
2841  */
2842 STATIC void
2843 xlog_ungrant_log_space(xlog_t        *log,
2844                        xlog_ticket_t *ticket)
2845 {
2846         if (ticket->t_cnt > 0)
2847                 ticket->t_cnt--;
2848
2849         spin_lock(&log->l_grant_lock);
2850         trace_xfs_log_ungrant_enter(log, ticket);
2851
2852         xlog_grant_sub_space(log, ticket->t_curr_res);
2853
2854         trace_xfs_log_ungrant_sub(log, ticket);
2855
2856         /* If this is a permanent reservation ticket, we may be able to free
2857          * up more space based on the remaining count.
2858          */
2859         if (ticket->t_cnt > 0) {
2860                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2861                 xlog_grant_sub_space(log, ticket->t_unit_res*ticket->t_cnt);
2862         }
2863
2864         trace_xfs_log_ungrant_exit(log, ticket);
2865
2866         xlog_verify_grant_head(log, 1);
2867         spin_unlock(&log->l_grant_lock);
2868         xfs_log_move_tail(log->l_mp, 1);
2869 }       /* xlog_ungrant_log_space */
2870
2871
2872 /*
2873  * Flush iclog to disk if this is the last reference to the given iclog and
2874  * the WANT_SYNC bit is set.
2875  *
2876  * When this function is entered, the iclog is not necessarily in the
2877  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2878  *
2879  *
2880  */
2881 STATIC int
2882 xlog_state_release_iclog(
2883         xlog_t          *log,
2884         xlog_in_core_t  *iclog)
2885 {
2886         int             sync = 0;       /* do we sync? */
2887
2888         if (iclog->ic_state & XLOG_STATE_IOERROR)
2889                 return XFS_ERROR(EIO);
2890
2891         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2892         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2893                 return 0;
2894
2895         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2896                 spin_unlock(&log->l_icloglock);
2897                 return XFS_ERROR(EIO);
2898         }
2899         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2900                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2901
2902         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2903                 /* update tail before writing to iclog */
2904                 xlog_assign_tail_lsn(log->l_mp);
2905                 sync++;
2906                 iclog->ic_state = XLOG_STATE_SYNCING;
2907                 iclog->ic_header.h_tail_lsn = cpu_to_be64(log->l_tail_lsn);
2908                 xlog_verify_tail_lsn(log, iclog, log->l_tail_lsn);
2909                 /* cycle incremented when incrementing curr_block */
2910         }
2911         spin_unlock(&log->l_icloglock);
2912
2913         /*
2914          * We let the log lock go, so it's possible that we hit a log I/O
2915          * error or some other SHUTDOWN condition that marks the iclog
2916          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2917          * this iclog has consistent data, so we ignore IOERROR
2918          * flags after this point.
2919          */
2920         if (sync)
2921                 return xlog_sync(log, iclog);
2922         return 0;
2923 }       /* xlog_state_release_iclog */
2924
2925
2926 /*
2927  * This routine will mark the current iclog in the ring as WANT_SYNC
2928  * and move the current iclog pointer to the next iclog in the ring.
2929  * When this routine is called from xlog_state_get_iclog_space(), the
2930  * exact size of the iclog has not yet been determined.  All we know is
2931  * that every data block.  We have run out of space in this log record.
2932  */
2933 STATIC void
2934 xlog_state_switch_iclogs(xlog_t         *log,
2935                          xlog_in_core_t *iclog,
2936                          int            eventual_size)
2937 {
2938         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2939         if (!eventual_size)
2940                 eventual_size = iclog->ic_offset;
2941         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2942         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2943         log->l_prev_block = log->l_curr_block;
2944         log->l_prev_cycle = log->l_curr_cycle;
2945
2946         /* roll log?: ic_offset changed later */
2947         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2948
2949         /* Round up to next log-sunit */
2950         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2951             log->l_mp->m_sb.sb_logsunit > 1) {
2952                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2953                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2954         }
2955
2956         if (log->l_curr_block >= log->l_logBBsize) {
2957                 log->l_curr_cycle++;
2958                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2959                         log->l_curr_cycle++;
2960                 log->l_curr_block -= log->l_logBBsize;
2961                 ASSERT(log->l_curr_block >= 0);
2962         }
2963         ASSERT(iclog == log->l_iclog);
2964         log->l_iclog = iclog->ic_next;
2965 }       /* xlog_state_switch_iclogs */
2966
2967 /*
2968  * Write out all data in the in-core log as of this exact moment in time.
2969  *
2970  * Data may be written to the in-core log during this call.  However,
2971  * we don't guarantee this data will be written out.  A change from past
2972  * implementation means this routine will *not* write out zero length LRs.
2973  *
2974  * Basically, we try and perform an intelligent scan of the in-core logs.
2975  * If we determine there is no flushable data, we just return.  There is no
2976  * flushable data if:
2977  *
2978  *      1. the current iclog is active and has no data; the previous iclog
2979  *              is in the active or dirty state.
2980  *      2. the current iclog is drity, and the previous iclog is in the
2981  *              active or dirty state.
2982  *
2983  * We may sleep if:
2984  *
2985  *      1. the current iclog is not in the active nor dirty state.
2986  *      2. the current iclog dirty, and the previous iclog is not in the
2987  *              active nor dirty state.
2988  *      3. the current iclog is active, and there is another thread writing
2989  *              to this particular iclog.
2990  *      4. a) the current iclog is active and has no other writers
2991  *         b) when we return from flushing out this iclog, it is still
2992  *              not in the active nor dirty state.
2993  */
2994 int
2995 _xfs_log_force(
2996         struct xfs_mount        *mp,
2997         uint                    flags,
2998         int                     *log_flushed)
2999 {
3000         struct log              *log = mp->m_log;
3001         struct xlog_in_core     *iclog;
3002         xfs_lsn_t               lsn;
3003
3004         XFS_STATS_INC(xs_log_force);
3005
3006         if (log->l_cilp)
3007                 xlog_cil_force(log);
3008
3009         spin_lock(&log->l_icloglock);
3010
3011         iclog = log->l_iclog;
3012         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3013                 spin_unlock(&log->l_icloglock);
3014                 return XFS_ERROR(EIO);
3015         }
3016
3017         /* If the head iclog is not active nor dirty, we just attach
3018          * ourselves to the head and go to sleep.
3019          */
3020         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3021             iclog->ic_state == XLOG_STATE_DIRTY) {
3022                 /*
3023                  * If the head is dirty or (active and empty), then
3024                  * we need to look at the previous iclog.  If the previous
3025                  * iclog is active or dirty we are done.  There is nothing
3026                  * to sync out.  Otherwise, we attach ourselves to the
3027                  * previous iclog and go to sleep.
3028                  */
3029                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3030                     (atomic_read(&iclog->ic_refcnt) == 0
3031                      && iclog->ic_offset == 0)) {
3032                         iclog = iclog->ic_prev;
3033                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3034                             iclog->ic_state == XLOG_STATE_DIRTY)
3035                                 goto no_sleep;
3036                         else
3037                                 goto maybe_sleep;
3038                 } else {
3039                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3040                                 /* We are the only one with access to this
3041                                  * iclog.  Flush it out now.  There should
3042                                  * be a roundoff of zero to show that someone
3043                                  * has already taken care of the roundoff from
3044                                  * the previous sync.
3045                                  */
3046                                 atomic_inc(&iclog->ic_refcnt);
3047                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3048                                 xlog_state_switch_iclogs(log, iclog, 0);
3049                                 spin_unlock(&log->l_icloglock);
3050
3051                                 if (xlog_state_release_iclog(log, iclog))
3052                                         return XFS_ERROR(EIO);
3053
3054                                 if (log_flushed)
3055                                         *log_flushed = 1;
3056                                 spin_lock(&log->l_icloglock);
3057                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3058                                     iclog->ic_state != XLOG_STATE_DIRTY)
3059                                         goto maybe_sleep;
3060                                 else
3061                                         goto no_sleep;
3062                         } else {
3063                                 /* Someone else is writing to this iclog.
3064                                  * Use its call to flush out the data.  However,
3065                                  * the other thread may not force out this LR,
3066                                  * so we mark it WANT_SYNC.
3067                                  */
3068                                 xlog_state_switch_iclogs(log, iclog, 0);
3069                                 goto maybe_sleep;
3070                         }
3071                 }
3072         }
3073
3074         /* By the time we come around again, the iclog could've been filled
3075          * which would give it another lsn.  If we have a new lsn, just
3076          * return because the relevant data has been flushed.
3077          */
3078 maybe_sleep:
3079         if (flags & XFS_LOG_SYNC) {
3080                 /*
3081                  * We must check if we're shutting down here, before
3082                  * we wait, while we're holding the l_icloglock.
3083                  * Then we check again after waking up, in case our
3084                  * sleep was disturbed by a bad news.
3085                  */
3086                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3087                         spin_unlock(&log->l_icloglock);