Merge branch 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4
[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(struct log *log, atomic64_t *head);
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(struct log      *log,
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_tail(struct log *log);
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_tail(a)
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 static void
99 xlog_grant_sub_space(
100         struct log      *log,
101         atomic64_t      *head,
102         int             bytes)
103 {
104         int64_t head_val = atomic64_read(head);
105         int64_t new, old;
106
107         do {
108                 int     cycle, space;
109
110                 xlog_crack_grant_head_val(head_val, &cycle, &space);
111
112                 space -= bytes;
113                 if (space < 0) {
114                         space += log->l_logsize;
115                         cycle--;
116                 }
117
118                 old = head_val;
119                 new = xlog_assign_grant_head_val(cycle, space);
120                 head_val = atomic64_cmpxchg(head, old, new);
121         } while (head_val != old);
122 }
123
124 static void
125 xlog_grant_add_space(
126         struct log      *log,
127         atomic64_t      *head,
128         int             bytes)
129 {
130         int64_t head_val = atomic64_read(head);
131         int64_t new, old;
132
133         do {
134                 int             tmp;
135                 int             cycle, space;
136
137                 xlog_crack_grant_head_val(head_val, &cycle, &space);
138
139                 tmp = log->l_logsize - space;
140                 if (tmp > bytes)
141                         space += bytes;
142                 else {
143                         space = bytes - tmp;
144                         cycle++;
145                 }
146
147                 old = head_val;
148                 new = xlog_assign_grant_head_val(cycle, space);
149                 head_val = atomic64_cmpxchg(head, old, new);
150         } while (head_val != old);
151 }
152
153 static void
154 xlog_tic_reset_res(xlog_ticket_t *tic)
155 {
156         tic->t_res_num = 0;
157         tic->t_res_arr_sum = 0;
158         tic->t_res_num_ophdrs = 0;
159 }
160
161 static void
162 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
163 {
164         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
165                 /* add to overflow and start again */
166                 tic->t_res_o_flow += tic->t_res_arr_sum;
167                 tic->t_res_num = 0;
168                 tic->t_res_arr_sum = 0;
169         }
170
171         tic->t_res_arr[tic->t_res_num].r_len = len;
172         tic->t_res_arr[tic->t_res_num].r_type = type;
173         tic->t_res_arr_sum += len;
174         tic->t_res_num++;
175 }
176
177 /*
178  * NOTES:
179  *
180  *      1. currblock field gets updated at startup and after in-core logs
181  *              marked as with WANT_SYNC.
182  */
183
184 /*
185  * This routine is called when a user of a log manager ticket is done with
186  * the reservation.  If the ticket was ever used, then a commit record for
187  * the associated transaction is written out as a log operation header with
188  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
189  * a given ticket.  If the ticket was one with a permanent reservation, then
190  * a few operations are done differently.  Permanent reservation tickets by
191  * default don't release the reservation.  They just commit the current
192  * transaction with the belief that the reservation is still needed.  A flag
193  * must be passed in before permanent reservations are actually released.
194  * When these type of tickets are not released, they need to be set into
195  * the inited state again.  By doing this, a start record will be written
196  * out when the next write occurs.
197  */
198 xfs_lsn_t
199 xfs_log_done(
200         struct xfs_mount        *mp,
201         struct xlog_ticket      *ticket,
202         struct xlog_in_core     **iclog,
203         uint                    flags)
204 {
205         struct log              *log = mp->m_log;
206         xfs_lsn_t               lsn = 0;
207
208         if (XLOG_FORCED_SHUTDOWN(log) ||
209             /*
210              * If nothing was ever written, don't write out commit record.
211              * If we get an error, just continue and give back the log ticket.
212              */
213             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
214              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
215                 lsn = (xfs_lsn_t) -1;
216                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
217                         flags |= XFS_LOG_REL_PERM_RESERV;
218                 }
219         }
220
221
222         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
223             (flags & XFS_LOG_REL_PERM_RESERV)) {
224                 trace_xfs_log_done_nonperm(log, ticket);
225
226                 /*
227                  * Release ticket if not permanent reservation or a specific
228                  * request has been made to release a permanent reservation.
229                  */
230                 xlog_ungrant_log_space(log, ticket);
231                 xfs_log_ticket_put(ticket);
232         } else {
233                 trace_xfs_log_done_perm(log, ticket);
234
235                 xlog_regrant_reserve_log_space(log, ticket);
236                 /* If this ticket was a permanent reservation and we aren't
237                  * trying to release it, reset the inited flags; so next time
238                  * we write, a start record will be written out.
239                  */
240                 ticket->t_flags |= XLOG_TIC_INITED;
241         }
242
243         return lsn;
244 }
245
246 /*
247  * Attaches a new iclog I/O completion callback routine during
248  * transaction commit.  If the log is in error state, a non-zero
249  * return code is handed back and the caller is responsible for
250  * executing the callback at an appropriate time.
251  */
252 int
253 xfs_log_notify(
254         struct xfs_mount        *mp,
255         struct xlog_in_core     *iclog,
256         xfs_log_callback_t      *cb)
257 {
258         int     abortflg;
259
260         spin_lock(&iclog->ic_callback_lock);
261         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
262         if (!abortflg) {
263                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
264                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
265                 cb->cb_next = NULL;
266                 *(iclog->ic_callback_tail) = cb;
267                 iclog->ic_callback_tail = &(cb->cb_next);
268         }
269         spin_unlock(&iclog->ic_callback_lock);
270         return abortflg;
271 }
272
273 int
274 xfs_log_release_iclog(
275         struct xfs_mount        *mp,
276         struct xlog_in_core     *iclog)
277 {
278         if (xlog_state_release_iclog(mp->m_log, iclog)) {
279                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
280                 return EIO;
281         }
282
283         return 0;
284 }
285
286 /*
287  *  1. Reserve an amount of on-disk log space and return a ticket corresponding
288  *      to the reservation.
289  *  2. Potentially, push buffers at tail of log to disk.
290  *
291  * Each reservation is going to reserve extra space for a log record header.
292  * When writes happen to the on-disk log, we don't subtract the length of the
293  * log record header from any reservation.  By wasting space in each
294  * reservation, we prevent over allocation problems.
295  */
296 int
297 xfs_log_reserve(
298         struct xfs_mount        *mp,
299         int                     unit_bytes,
300         int                     cnt,
301         struct xlog_ticket      **ticket,
302         __uint8_t               client,
303         uint                    flags,
304         uint                    t_type)
305 {
306         struct log              *log = mp->m_log;
307         struct xlog_ticket      *internal_ticket;
308         int                     retval = 0;
309
310         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
311
312         if (XLOG_FORCED_SHUTDOWN(log))
313                 return XFS_ERROR(EIO);
314
315         XFS_STATS_INC(xs_try_logspace);
316
317
318         if (*ticket != NULL) {
319                 ASSERT(flags & XFS_LOG_PERM_RESERV);
320                 internal_ticket = *ticket;
321
322                 /*
323                  * this is a new transaction on the ticket, so we need to
324                  * change the transaction ID so that the next transaction has a
325                  * different TID in the log. Just add one to the existing tid
326                  * so that we can see chains of rolling transactions in the log
327                  * easily.
328                  */
329                 internal_ticket->t_tid++;
330
331                 trace_xfs_log_reserve(log, internal_ticket);
332
333                 xlog_grant_push_ail(log, internal_ticket->t_unit_res);
334                 retval = xlog_regrant_write_log_space(log, internal_ticket);
335         } else {
336                 /* may sleep if need to allocate more tickets */
337                 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt,
338                                                   client, flags,
339                                                   KM_SLEEP|KM_MAYFAIL);
340                 if (!internal_ticket)
341                         return XFS_ERROR(ENOMEM);
342                 internal_ticket->t_trans_type = t_type;
343                 *ticket = internal_ticket;
344
345                 trace_xfs_log_reserve(log, internal_ticket);
346
347                 xlog_grant_push_ail(log,
348                                     (internal_ticket->t_unit_res *
349                                      internal_ticket->t_cnt));
350                 retval = xlog_grant_log_space(log, internal_ticket);
351         }
352
353         return retval;
354 }       /* xfs_log_reserve */
355
356
357 /*
358  * Mount a log filesystem
359  *
360  * mp           - ubiquitous xfs mount point structure
361  * log_target   - buftarg of on-disk log device
362  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
363  * num_bblocks  - Number of BBSIZE blocks in on-disk log
364  *
365  * Return error or zero.
366  */
367 int
368 xfs_log_mount(
369         xfs_mount_t     *mp,
370         xfs_buftarg_t   *log_target,
371         xfs_daddr_t     blk_offset,
372         int             num_bblks)
373 {
374         int             error;
375
376         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
377                 xfs_notice(mp, "Mounting Filesystem");
378         else {
379                 xfs_notice(mp,
380 "Mounting filesystem in no-recovery mode.  Filesystem will be inconsistent.");
381                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
382         }
383
384         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
385         if (IS_ERR(mp->m_log)) {
386                 error = -PTR_ERR(mp->m_log);
387                 goto out;
388         }
389
390         /*
391          * Initialize the AIL now we have a log.
392          */
393         error = xfs_trans_ail_init(mp);
394         if (error) {
395                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
396                 goto out_free_log;
397         }
398         mp->m_log->l_ailp = mp->m_ail;
399
400         /*
401          * skip log recovery on a norecovery mount.  pretend it all
402          * just worked.
403          */
404         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
405                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
406
407                 if (readonly)
408                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
409
410                 error = xlog_recover(mp->m_log);
411
412                 if (readonly)
413                         mp->m_flags |= XFS_MOUNT_RDONLY;
414                 if (error) {
415                         xfs_warn(mp, "log mount/recovery failed: error %d",
416                                 error);
417                         goto out_destroy_ail;
418                 }
419         }
420
421         /* Normal transactions can now occur */
422         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
423
424         /*
425          * Now the log has been fully initialised and we know were our
426          * space grant counters are, we can initialise the permanent ticket
427          * needed for delayed logging to work.
428          */
429         xlog_cil_init_post_recovery(mp->m_log);
430
431         return 0;
432
433 out_destroy_ail:
434         xfs_trans_ail_destroy(mp);
435 out_free_log:
436         xlog_dealloc_log(mp->m_log);
437 out:
438         return error;
439 }
440
441 /*
442  * Finish the recovery of the file system.  This is separate from
443  * the xfs_log_mount() call, because it depends on the code in
444  * xfs_mountfs() to read in the root and real-time bitmap inodes
445  * between calling xfs_log_mount() and here.
446  *
447  * mp           - ubiquitous xfs mount point structure
448  */
449 int
450 xfs_log_mount_finish(xfs_mount_t *mp)
451 {
452         int     error;
453
454         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
455                 error = xlog_recover_finish(mp->m_log);
456         else {
457                 error = 0;
458                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
459         }
460
461         return error;
462 }
463
464 /*
465  * Final log writes as part of unmount.
466  *
467  * Mark the filesystem clean as unmount happens.  Note that during relocation
468  * this routine needs to be executed as part of source-bag while the
469  * deallocation must not be done until source-end.
470  */
471
472 /*
473  * Unmount record used to have a string "Unmount filesystem--" in the
474  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
475  * We just write the magic number now since that particular field isn't
476  * currently architecture converted and "nUmount" is a bit foo.
477  * As far as I know, there weren't any dependencies on the old behaviour.
478  */
479
480 int
481 xfs_log_unmount_write(xfs_mount_t *mp)
482 {
483         xlog_t           *log = mp->m_log;
484         xlog_in_core_t   *iclog;
485 #ifdef DEBUG
486         xlog_in_core_t   *first_iclog;
487 #endif
488         xlog_ticket_t   *tic = NULL;
489         xfs_lsn_t        lsn;
490         int              error;
491
492         /*
493          * Don't write out unmount record on read-only mounts.
494          * Or, if we are doing a forced umount (typically because of IO errors).
495          */
496         if (mp->m_flags & XFS_MOUNT_RDONLY)
497                 return 0;
498
499         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
500         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
501
502 #ifdef DEBUG
503         first_iclog = iclog = log->l_iclog;
504         do {
505                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
506                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
507                         ASSERT(iclog->ic_offset == 0);
508                 }
509                 iclog = iclog->ic_next;
510         } while (iclog != first_iclog);
511 #endif
512         if (! (XLOG_FORCED_SHUTDOWN(log))) {
513                 error = xfs_log_reserve(mp, 600, 1, &tic,
514                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
515                 if (!error) {
516                         /* the data section must be 32 bit size aligned */
517                         struct {
518                             __uint16_t magic;
519                             __uint16_t pad1;
520                             __uint32_t pad2; /* may as well make it 64 bits */
521                         } magic = {
522                                 .magic = XLOG_UNMOUNT_TYPE,
523                         };
524                         struct xfs_log_iovec reg = {
525                                 .i_addr = &magic,
526                                 .i_len = sizeof(magic),
527                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
528                         };
529                         struct xfs_log_vec vec = {
530                                 .lv_niovecs = 1,
531                                 .lv_iovecp = &reg,
532                         };
533
534                         /* remove inited flag */
535                         tic->t_flags = 0;
536                         error = xlog_write(log, &vec, tic, &lsn,
537                                            NULL, XLOG_UNMOUNT_TRANS);
538                         /*
539                          * At this point, we're umounting anyway,
540                          * so there's no point in transitioning log state
541                          * to IOERROR. Just continue...
542                          */
543                 }
544
545                 if (error)
546                         xfs_alert(mp, "%s: unmount record failed", __func__);
547
548
549                 spin_lock(&log->l_icloglock);
550                 iclog = log->l_iclog;
551                 atomic_inc(&iclog->ic_refcnt);
552                 xlog_state_want_sync(log, iclog);
553                 spin_unlock(&log->l_icloglock);
554                 error = xlog_state_release_iclog(log, iclog);
555
556                 spin_lock(&log->l_icloglock);
557                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
558                       iclog->ic_state == XLOG_STATE_DIRTY)) {
559                         if (!XLOG_FORCED_SHUTDOWN(log)) {
560                                 xlog_wait(&iclog->ic_force_wait,
561                                                         &log->l_icloglock);
562                         } else {
563                                 spin_unlock(&log->l_icloglock);
564                         }
565                 } else {
566                         spin_unlock(&log->l_icloglock);
567                 }
568                 if (tic) {
569                         trace_xfs_log_umount_write(log, tic);
570                         xlog_ungrant_log_space(log, tic);
571                         xfs_log_ticket_put(tic);
572                 }
573         } else {
574                 /*
575                  * We're already in forced_shutdown mode, couldn't
576                  * even attempt to write out the unmount transaction.
577                  *
578                  * Go through the motions of sync'ing and releasing
579                  * the iclog, even though no I/O will actually happen,
580                  * we need to wait for other log I/Os that may already
581                  * be in progress.  Do this as a separate section of
582                  * code so we'll know if we ever get stuck here that
583                  * we're in this odd situation of trying to unmount
584                  * a file system that went into forced_shutdown as
585                  * the result of an unmount..
586                  */
587                 spin_lock(&log->l_icloglock);
588                 iclog = log->l_iclog;
589                 atomic_inc(&iclog->ic_refcnt);
590
591                 xlog_state_want_sync(log, iclog);
592                 spin_unlock(&log->l_icloglock);
593                 error =  xlog_state_release_iclog(log, iclog);
594
595                 spin_lock(&log->l_icloglock);
596
597                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
598                         || iclog->ic_state == XLOG_STATE_DIRTY
599                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
600
601                                 xlog_wait(&iclog->ic_force_wait,
602                                                         &log->l_icloglock);
603                 } else {
604                         spin_unlock(&log->l_icloglock);
605                 }
606         }
607
608         return error;
609 }       /* xfs_log_unmount_write */
610
611 /*
612  * Deallocate log structures for unmount/relocation.
613  *
614  * We need to stop the aild from running before we destroy
615  * and deallocate the log as the aild references the log.
616  */
617 void
618 xfs_log_unmount(xfs_mount_t *mp)
619 {
620         xfs_trans_ail_destroy(mp);
621         xlog_dealloc_log(mp->m_log);
622 }
623
624 void
625 xfs_log_item_init(
626         struct xfs_mount        *mp,
627         struct xfs_log_item     *item,
628         int                     type,
629         struct xfs_item_ops     *ops)
630 {
631         item->li_mountp = mp;
632         item->li_ailp = mp->m_ail;
633         item->li_type = type;
634         item->li_ops = ops;
635         item->li_lv = NULL;
636
637         INIT_LIST_HEAD(&item->li_ail);
638         INIT_LIST_HEAD(&item->li_cil);
639 }
640
641 /*
642  * Write region vectors to log.  The write happens using the space reservation
643  * of the ticket (tic).  It is not a requirement that all writes for a given
644  * transaction occur with one call to xfs_log_write(). However, it is important
645  * to note that the transaction reservation code makes an assumption about the
646  * number of log headers a transaction requires that may be violated if you
647  * don't pass all the transaction vectors in one call....
648  */
649 int
650 xfs_log_write(
651         struct xfs_mount        *mp,
652         struct xfs_log_iovec    reg[],
653         int                     nentries,
654         struct xlog_ticket      *tic,
655         xfs_lsn_t               *start_lsn)
656 {
657         struct log              *log = mp->m_log;
658         int                     error;
659         struct xfs_log_vec      vec = {
660                 .lv_niovecs = nentries,
661                 .lv_iovecp = reg,
662         };
663
664         if (XLOG_FORCED_SHUTDOWN(log))
665                 return XFS_ERROR(EIO);
666
667         error = xlog_write(log, &vec, tic, start_lsn, NULL, 0);
668         if (error)
669                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
670         return error;
671 }
672
673 void
674 xfs_log_move_tail(xfs_mount_t   *mp,
675                   xfs_lsn_t     tail_lsn)
676 {
677         xlog_ticket_t   *tic;
678         xlog_t          *log = mp->m_log;
679         int             need_bytes, free_bytes;
680
681         if (XLOG_FORCED_SHUTDOWN(log))
682                 return;
683
684         if (tail_lsn == 0)
685                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
686
687         /* tail_lsn == 1 implies that we weren't passed a valid value.  */
688         if (tail_lsn != 1)
689                 atomic64_set(&log->l_tail_lsn, tail_lsn);
690
691         if (!list_empty_careful(&log->l_writeq)) {
692 #ifdef DEBUG
693                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
694                         panic("Recovery problem");
695 #endif
696                 spin_lock(&log->l_grant_write_lock);
697                 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
698                 list_for_each_entry(tic, &log->l_writeq, t_queue) {
699                         ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
700
701                         if (free_bytes < tic->t_unit_res && tail_lsn != 1)
702                                 break;
703                         tail_lsn = 0;
704                         free_bytes -= tic->t_unit_res;
705                         trace_xfs_log_regrant_write_wake_up(log, tic);
706                         wake_up(&tic->t_wait);
707                 }
708                 spin_unlock(&log->l_grant_write_lock);
709         }
710
711         if (!list_empty_careful(&log->l_reserveq)) {
712 #ifdef DEBUG
713                 if (log->l_flags & XLOG_ACTIVE_RECOVERY)
714                         panic("Recovery problem");
715 #endif
716                 spin_lock(&log->l_grant_reserve_lock);
717                 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
718                 list_for_each_entry(tic, &log->l_reserveq, t_queue) {
719                         if (tic->t_flags & XLOG_TIC_PERM_RESERV)
720                                 need_bytes = tic->t_unit_res*tic->t_cnt;
721                         else
722                                 need_bytes = tic->t_unit_res;
723                         if (free_bytes < need_bytes && tail_lsn != 1)
724                                 break;
725                         tail_lsn = 0;
726                         free_bytes -= need_bytes;
727                         trace_xfs_log_grant_wake_up(log, tic);
728                         wake_up(&tic->t_wait);
729                 }
730                 spin_unlock(&log->l_grant_reserve_lock);
731         }
732 }
733
734 /*
735  * Determine if we have a transaction that has gone to disk
736  * that needs to be covered. To begin the transition to the idle state
737  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
738  * If we are then in a state where covering is needed, the caller is informed
739  * that dummy transactions are required to move the log into the idle state.
740  *
741  * Because this is called as part of the sync process, we should also indicate
742  * that dummy transactions should be issued in anything but the covered or
743  * idle states. This ensures that the log tail is accurately reflected in
744  * the log at the end of the sync, hence if a crash occurrs avoids replay
745  * of transactions where the metadata is already on disk.
746  */
747 int
748 xfs_log_need_covered(xfs_mount_t *mp)
749 {
750         int             needed = 0;
751         xlog_t          *log = mp->m_log;
752
753         if (!xfs_fs_writable(mp))
754                 return 0;
755
756         spin_lock(&log->l_icloglock);
757         switch (log->l_covered_state) {
758         case XLOG_STATE_COVER_DONE:
759         case XLOG_STATE_COVER_DONE2:
760         case XLOG_STATE_COVER_IDLE:
761                 break;
762         case XLOG_STATE_COVER_NEED:
763         case XLOG_STATE_COVER_NEED2:
764                 if (!xfs_trans_ail_tail(log->l_ailp) &&
765                     xlog_iclogs_empty(log)) {
766                         if (log->l_covered_state == XLOG_STATE_COVER_NEED)
767                                 log->l_covered_state = XLOG_STATE_COVER_DONE;
768                         else
769                                 log->l_covered_state = XLOG_STATE_COVER_DONE2;
770                 }
771                 /* FALLTHRU */
772         default:
773                 needed = 1;
774                 break;
775         }
776         spin_unlock(&log->l_icloglock);
777         return needed;
778 }
779
780 /******************************************************************************
781  *
782  *      local routines
783  *
784  ******************************************************************************
785  */
786
787 /* xfs_trans_tail_ail returns 0 when there is nothing in the list.
788  * The log manager must keep track of the last LR which was committed
789  * to disk.  The lsn of this LR will become the new tail_lsn whenever
790  * xfs_trans_tail_ail returns 0.  If we don't do this, we run into
791  * the situation where stuff could be written into the log but nothing
792  * was ever in the AIL when asked.  Eventually, we panic since the
793  * tail hits the head.
794  *
795  * We may be holding the log iclog lock upon entering this routine.
796  */
797 xfs_lsn_t
798 xlog_assign_tail_lsn(
799         struct xfs_mount        *mp)
800 {
801         xfs_lsn_t               tail_lsn;
802         struct log              *log = mp->m_log;
803
804         tail_lsn = xfs_trans_ail_tail(mp->m_ail);
805         if (!tail_lsn)
806                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
807
808         atomic64_set(&log->l_tail_lsn, tail_lsn);
809         return tail_lsn;
810 }
811
812 /*
813  * Return the space in the log between the tail and the head.  The head
814  * is passed in the cycle/bytes formal parms.  In the special case where
815  * the reserve head has wrapped passed the tail, this calculation is no
816  * longer valid.  In this case, just return 0 which means there is no space
817  * in the log.  This works for all places where this function is called
818  * with the reserve head.  Of course, if the write head were to ever
819  * wrap the tail, we should blow up.  Rather than catch this case here,
820  * we depend on other ASSERTions in other parts of the code.   XXXmiken
821  *
822  * This code also handles the case where the reservation head is behind
823  * the tail.  The details of this case are described below, but the end
824  * result is that we return the size of the log as the amount of space left.
825  */
826 STATIC int
827 xlog_space_left(
828         struct log      *log,
829         atomic64_t      *head)
830 {
831         int             free_bytes;
832         int             tail_bytes;
833         int             tail_cycle;
834         int             head_cycle;
835         int             head_bytes;
836
837         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
838         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
839         tail_bytes = BBTOB(tail_bytes);
840         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
841                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
842         else if (tail_cycle + 1 < head_cycle)
843                 return 0;
844         else if (tail_cycle < head_cycle) {
845                 ASSERT(tail_cycle == (head_cycle - 1));
846                 free_bytes = tail_bytes - head_bytes;
847         } else {
848                 /*
849                  * The reservation head is behind the tail.
850                  * In this case we just want to return the size of the
851                  * log as the amount of space left.
852                  */
853                 xfs_alert(log->l_mp,
854                         "xlog_space_left: head behind tail\n"
855                         "  tail_cycle = %d, tail_bytes = %d\n"
856                         "  GH   cycle = %d, GH   bytes = %d",
857                         tail_cycle, tail_bytes, head_cycle, head_bytes);
858                 ASSERT(0);
859                 free_bytes = log->l_logsize;
860         }
861         return free_bytes;
862 }
863
864
865 /*
866  * Log function which is called when an io completes.
867  *
868  * The log manager needs its own routine, in order to control what
869  * happens with the buffer after the write completes.
870  */
871 void
872 xlog_iodone(xfs_buf_t *bp)
873 {
874         xlog_in_core_t  *iclog;
875         xlog_t          *l;
876         int             aborted;
877
878         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
879         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2);
880         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
881         aborted = 0;
882         l = iclog->ic_log;
883
884         /*
885          * Race to shutdown the filesystem if we see an error.
886          */
887         if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp,
888                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
889                 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp));
890                 XFS_BUF_STALE(bp);
891                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
892                 /*
893                  * This flag will be propagated to the trans-committed
894                  * callback routines to let them know that the log-commit
895                  * didn't succeed.
896                  */
897                 aborted = XFS_LI_ABORTED;
898         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
899                 aborted = XFS_LI_ABORTED;
900         }
901
902         /* log I/O is always issued ASYNC */
903         ASSERT(XFS_BUF_ISASYNC(bp));
904         xlog_state_done_syncing(iclog, aborted);
905         /*
906          * do not reference the buffer (bp) here as we could race
907          * with it being freed after writing the unmount record to the
908          * log.
909          */
910
911 }       /* xlog_iodone */
912
913 /*
914  * Return size of each in-core log record buffer.
915  *
916  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
917  *
918  * If the filesystem blocksize is too large, we may need to choose a
919  * larger size since the directory code currently logs entire blocks.
920  */
921
922 STATIC void
923 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
924                            xlog_t       *log)
925 {
926         int size;
927         int xhdrs;
928
929         if (mp->m_logbufs <= 0)
930                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
931         else
932                 log->l_iclog_bufs = mp->m_logbufs;
933
934         /*
935          * Buffer size passed in from mount system call.
936          */
937         if (mp->m_logbsize > 0) {
938                 size = log->l_iclog_size = mp->m_logbsize;
939                 log->l_iclog_size_log = 0;
940                 while (size != 1) {
941                         log->l_iclog_size_log++;
942                         size >>= 1;
943                 }
944
945                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
946                         /* # headers = size / 32k
947                          * one header holds cycles from 32k of data
948                          */
949
950                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
951                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
952                                 xhdrs++;
953                         log->l_iclog_hsize = xhdrs << BBSHIFT;
954                         log->l_iclog_heads = xhdrs;
955                 } else {
956                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
957                         log->l_iclog_hsize = BBSIZE;
958                         log->l_iclog_heads = 1;
959                 }
960                 goto done;
961         }
962
963         /* All machines use 32kB buffers by default. */
964         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
965         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
966
967         /* the default log size is 16k or 32k which is one header sector */
968         log->l_iclog_hsize = BBSIZE;
969         log->l_iclog_heads = 1;
970
971 done:
972         /* are we being asked to make the sizes selected above visible? */
973         if (mp->m_logbufs == 0)
974                 mp->m_logbufs = log->l_iclog_bufs;
975         if (mp->m_logbsize == 0)
976                 mp->m_logbsize = log->l_iclog_size;
977 }       /* xlog_get_iclog_buffer_size */
978
979
980 /*
981  * This routine initializes some of the log structure for a given mount point.
982  * Its primary purpose is to fill in enough, so recovery can occur.  However,
983  * some other stuff may be filled in too.
984  */
985 STATIC xlog_t *
986 xlog_alloc_log(xfs_mount_t      *mp,
987                xfs_buftarg_t    *log_target,
988                xfs_daddr_t      blk_offset,
989                int              num_bblks)
990 {
991         xlog_t                  *log;
992         xlog_rec_header_t       *head;
993         xlog_in_core_t          **iclogp;
994         xlog_in_core_t          *iclog, *prev_iclog=NULL;
995         xfs_buf_t               *bp;
996         int                     i;
997         int                     error = ENOMEM;
998         uint                    log2_size = 0;
999
1000         log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1001         if (!log) {
1002                 xfs_warn(mp, "Log allocation failed: No memory!");
1003                 goto out;
1004         }
1005
1006         log->l_mp          = mp;
1007         log->l_targ        = log_target;
1008         log->l_logsize     = BBTOB(num_bblks);
1009         log->l_logBBstart  = blk_offset;
1010         log->l_logBBsize   = num_bblks;
1011         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1012         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1013
1014         log->l_prev_block  = -1;
1015         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1016         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1017         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1018         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1019         xlog_assign_grant_head(&log->l_grant_reserve_head, 1, 0);
1020         xlog_assign_grant_head(&log->l_grant_write_head, 1, 0);
1021         INIT_LIST_HEAD(&log->l_reserveq);
1022         INIT_LIST_HEAD(&log->l_writeq);
1023         spin_lock_init(&log->l_grant_reserve_lock);
1024         spin_lock_init(&log->l_grant_write_lock);
1025
1026         error = EFSCORRUPTED;
1027         if (xfs_sb_version_hassector(&mp->m_sb)) {
1028                 log2_size = mp->m_sb.sb_logsectlog;
1029                 if (log2_size < BBSHIFT) {
1030                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1031                                 log2_size, BBSHIFT);
1032                         goto out_free_log;
1033                 }
1034
1035                 log2_size -= BBSHIFT;
1036                 if (log2_size > mp->m_sectbb_log) {
1037                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1038                                 log2_size, mp->m_sectbb_log);
1039                         goto out_free_log;
1040                 }
1041
1042                 /* for larger sector sizes, must have v2 or external log */
1043                 if (log2_size && log->l_logBBstart > 0 &&
1044                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1045                         xfs_warn(mp,
1046                 "log sector size (0x%x) invalid for configuration.",
1047                                 log2_size);
1048                         goto out_free_log;
1049                 }
1050         }
1051         log->l_sectBBsize = 1 << log2_size;
1052
1053         xlog_get_iclog_buffer_size(mp, log);
1054
1055         error = ENOMEM;
1056         bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp);
1057         if (!bp)
1058                 goto out_free_log;
1059         XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1060         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1061         ASSERT(XFS_BUF_ISBUSY(bp));
1062         ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
1063         log->l_xbuf = bp;
1064
1065         spin_lock_init(&log->l_icloglock);
1066         init_waitqueue_head(&log->l_flush_wait);
1067
1068         /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1069         ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1070
1071         iclogp = &log->l_iclog;
1072         /*
1073          * The amount of memory to allocate for the iclog structure is
1074          * rather funky due to the way the structure is defined.  It is
1075          * done this way so that we can use different sizes for machines
1076          * with different amounts of memory.  See the definition of
1077          * xlog_in_core_t in xfs_log_priv.h for details.
1078          */
1079         ASSERT(log->l_iclog_size >= 4096);
1080         for (i=0; i < log->l_iclog_bufs; i++) {
1081                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1082                 if (!*iclogp)
1083                         goto out_free_iclog;
1084
1085                 iclog = *iclogp;
1086                 iclog->ic_prev = prev_iclog;
1087                 prev_iclog = iclog;
1088
1089                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1090                                                 log->l_iclog_size, 0);
1091                 if (!bp)
1092                         goto out_free_iclog;
1093                 if (!XFS_BUF_CPSEMA(bp))
1094                         ASSERT(0);
1095                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone);
1096                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1);
1097                 iclog->ic_bp = bp;
1098                 iclog->ic_data = bp->b_addr;
1099 #ifdef DEBUG
1100                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1101 #endif
1102                 head = &iclog->ic_header;
1103                 memset(head, 0, sizeof(xlog_rec_header_t));
1104                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1105                 head->h_version = cpu_to_be32(
1106                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1107                 head->h_size = cpu_to_be32(log->l_iclog_size);
1108                 /* new fields */
1109                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1110                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1111
1112                 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1113                 iclog->ic_state = XLOG_STATE_ACTIVE;
1114                 iclog->ic_log = log;
1115                 atomic_set(&iclog->ic_refcnt, 0);
1116                 spin_lock_init(&iclog->ic_callback_lock);
1117                 iclog->ic_callback_tail = &(iclog->ic_callback);
1118                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1119
1120                 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp));
1121                 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0);
1122                 init_waitqueue_head(&iclog->ic_force_wait);
1123                 init_waitqueue_head(&iclog->ic_write_wait);
1124
1125                 iclogp = &iclog->ic_next;
1126         }
1127         *iclogp = log->l_iclog;                 /* complete ring */
1128         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1129
1130         error = xlog_cil_init(log);
1131         if (error)
1132                 goto out_free_iclog;
1133         return log;
1134
1135 out_free_iclog:
1136         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1137                 prev_iclog = iclog->ic_next;
1138                 if (iclog->ic_bp)
1139                         xfs_buf_free(iclog->ic_bp);
1140                 kmem_free(iclog);
1141         }
1142         spinlock_destroy(&log->l_icloglock);
1143         xfs_buf_free(log->l_xbuf);
1144 out_free_log:
1145         kmem_free(log);
1146 out:
1147         return ERR_PTR(-error);
1148 }       /* xlog_alloc_log */
1149
1150
1151 /*
1152  * Write out the commit record of a transaction associated with the given
1153  * ticket.  Return the lsn of the commit record.
1154  */
1155 STATIC int
1156 xlog_commit_record(
1157         struct log              *log,
1158         struct xlog_ticket      *ticket,
1159         struct xlog_in_core     **iclog,
1160         xfs_lsn_t               *commitlsnp)
1161 {
1162         struct xfs_mount *mp = log->l_mp;
1163         int     error;
1164         struct xfs_log_iovec reg = {
1165                 .i_addr = NULL,
1166                 .i_len = 0,
1167                 .i_type = XLOG_REG_TYPE_COMMIT,
1168         };
1169         struct xfs_log_vec vec = {
1170                 .lv_niovecs = 1,
1171                 .lv_iovecp = &reg,
1172         };
1173
1174         ASSERT_ALWAYS(iclog);
1175         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1176                                         XLOG_COMMIT_TRANS);
1177         if (error)
1178                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1179         return error;
1180 }
1181
1182 /*
1183  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1184  * log space.  This code pushes on the lsn which would supposedly free up
1185  * the 25% which we want to leave free.  We may need to adopt a policy which
1186  * pushes on an lsn which is further along in the log once we reach the high
1187  * water mark.  In this manner, we would be creating a low water mark.
1188  */
1189 STATIC void
1190 xlog_grant_push_ail(
1191         struct log      *log,
1192         int             need_bytes)
1193 {
1194         xfs_lsn_t       threshold_lsn = 0;
1195         xfs_lsn_t       last_sync_lsn;
1196         int             free_blocks;
1197         int             free_bytes;
1198         int             threshold_block;
1199         int             threshold_cycle;
1200         int             free_threshold;
1201
1202         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1203
1204         free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
1205         free_blocks = BTOBBT(free_bytes);
1206
1207         /*
1208          * Set the threshold for the minimum number of free blocks in the
1209          * log to the maximum of what the caller needs, one quarter of the
1210          * log, and 256 blocks.
1211          */
1212         free_threshold = BTOBB(need_bytes);
1213         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1214         free_threshold = MAX(free_threshold, 256);
1215         if (free_blocks >= free_threshold)
1216                 return;
1217
1218         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1219                                                 &threshold_block);
1220         threshold_block += free_threshold;
1221         if (threshold_block >= log->l_logBBsize) {
1222                 threshold_block -= log->l_logBBsize;
1223                 threshold_cycle += 1;
1224         }
1225         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1226                                         threshold_block);
1227         /*
1228          * Don't pass in an lsn greater than the lsn of the last
1229          * log record known to be on disk. Use a snapshot of the last sync lsn
1230          * so that it doesn't change between the compare and the set.
1231          */
1232         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1233         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1234                 threshold_lsn = last_sync_lsn;
1235
1236         /*
1237          * Get the transaction layer to kick the dirty buffers out to
1238          * disk asynchronously. No point in trying to do this if
1239          * the filesystem is shutting down.
1240          */
1241         if (!XLOG_FORCED_SHUTDOWN(log))
1242                 xfs_trans_ail_push(log->l_ailp, threshold_lsn);
1243 }
1244
1245 /*
1246  * The bdstrat callback function for log bufs. This gives us a central
1247  * place to trap bufs in case we get hit by a log I/O error and need to
1248  * shutdown. Actually, in practice, even when we didn't get a log error,
1249  * we transition the iclogs to IOERROR state *after* flushing all existing
1250  * iclogs to disk. This is because we don't want anymore new transactions to be
1251  * started or completed afterwards.
1252  */
1253 STATIC int
1254 xlog_bdstrat(
1255         struct xfs_buf          *bp)
1256 {
1257         struct xlog_in_core     *iclog;
1258
1259         iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *);
1260         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1261                 XFS_BUF_ERROR(bp, EIO);
1262                 XFS_BUF_STALE(bp);
1263                 xfs_buf_ioend(bp, 0);
1264                 /*
1265                  * It would seem logical to return EIO here, but we rely on
1266                  * the log state machine to propagate I/O errors instead of
1267                  * doing it here.
1268                  */
1269                 return 0;
1270         }
1271
1272         bp->b_flags |= _XBF_RUN_QUEUES;
1273         xfs_buf_iorequest(bp);
1274         return 0;
1275 }
1276
1277 /*
1278  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 
1279  * fashion.  Previously, we should have moved the current iclog
1280  * ptr in the log to point to the next available iclog.  This allows further
1281  * write to continue while this code syncs out an iclog ready to go.
1282  * Before an in-core log can be written out, the data section must be scanned
1283  * to save away the 1st word of each BBSIZE block into the header.  We replace
1284  * it with the current cycle count.  Each BBSIZE block is tagged with the
1285  * cycle count because there in an implicit assumption that drives will
1286  * guarantee that entire 512 byte blocks get written at once.  In other words,
1287  * we can't have part of a 512 byte block written and part not written.  By
1288  * tagging each block, we will know which blocks are valid when recovering
1289  * after an unclean shutdown.
1290  *
1291  * This routine is single threaded on the iclog.  No other thread can be in
1292  * this routine with the same iclog.  Changing contents of iclog can there-
1293  * fore be done without grabbing the state machine lock.  Updating the global
1294  * log will require grabbing the lock though.
1295  *
1296  * The entire log manager uses a logical block numbering scheme.  Only
1297  * log_sync (and then only bwrite()) know about the fact that the log may
1298  * not start with block zero on a given device.  The log block start offset
1299  * is added immediately before calling bwrite().
1300  */
1301
1302 STATIC int
1303 xlog_sync(xlog_t                *log,
1304           xlog_in_core_t        *iclog)
1305 {
1306         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1307         xfs_buf_t       *bp;
1308         int             i;
1309         uint            count;          /* byte count of bwrite */
1310         uint            count_init;     /* initial count before roundup */
1311         int             roundoff;       /* roundoff to BB or stripe */
1312         int             split = 0;      /* split write into two regions */
1313         int             error;
1314         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1315
1316         XFS_STATS_INC(xs_log_writes);
1317         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1318
1319         /* Add for LR header */
1320         count_init = log->l_iclog_hsize + iclog->ic_offset;
1321
1322         /* Round out the log write size */
1323         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1324                 /* we have a v2 stripe unit to use */
1325                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1326         } else {
1327                 count = BBTOB(BTOBB(count_init));
1328         }
1329         roundoff = count - count_init;
1330         ASSERT(roundoff >= 0);
1331         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 
1332                 roundoff < log->l_mp->m_sb.sb_logsunit)
1333                 || 
1334                 (log->l_mp->m_sb.sb_logsunit <= 1 && 
1335                  roundoff < BBTOB(1)));
1336
1337         /* move grant heads by roundoff in sync */
1338         xlog_grant_add_space(log, &log->l_grant_reserve_head, roundoff);
1339         xlog_grant_add_space(log, &log->l_grant_write_head, roundoff);
1340
1341         /* put cycle number in every block */
1342         xlog_pack_data(log, iclog, roundoff); 
1343
1344         /* real byte length */
1345         if (v2) {
1346                 iclog->ic_header.h_len =
1347                         cpu_to_be32(iclog->ic_offset + roundoff);
1348         } else {
1349                 iclog->ic_header.h_len =
1350                         cpu_to_be32(iclog->ic_offset);
1351         }
1352
1353         bp = iclog->ic_bp;
1354         ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1);
1355         XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1356         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1357
1358         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1359
1360         /* Do we need to split this write into 2 parts? */
1361         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1362                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1363                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1364                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1365         } else {
1366                 iclog->ic_bwritecnt = 1;
1367         }
1368         XFS_BUF_SET_COUNT(bp, count);
1369         XFS_BUF_SET_FSPRIVATE(bp, iclog);       /* save for later */
1370         XFS_BUF_ZEROFLAGS(bp);
1371         XFS_BUF_BUSY(bp);
1372         XFS_BUF_ASYNC(bp);
1373         bp->b_flags |= XBF_LOG_BUFFER;
1374
1375         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1376                 XFS_BUF_ORDERED(bp);
1377
1378         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1379         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1380
1381         xlog_verify_iclog(log, iclog, count, B_TRUE);
1382
1383         /* account for log which doesn't start at block #0 */
1384         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1385         /*
1386          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1387          * is shutting down.
1388          */
1389         XFS_BUF_WRITE(bp);
1390
1391         if ((error = xlog_bdstrat(bp))) {
1392                 xfs_ioerror_alert("xlog_sync", log->l_mp, bp,
1393                                   XFS_BUF_ADDR(bp));
1394                 return error;
1395         }
1396         if (split) {
1397                 bp = iclog->ic_log->l_xbuf;
1398                 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) ==
1399                                                         (unsigned long)1);
1400                 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2);
1401                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1402                 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+
1403                                             (__psint_t)count), split);
1404                 XFS_BUF_SET_FSPRIVATE(bp, iclog);
1405                 XFS_BUF_ZEROFLAGS(bp);
1406                 XFS_BUF_BUSY(bp);
1407                 XFS_BUF_ASYNC(bp);
1408                 bp->b_flags |= XBF_LOG_BUFFER;
1409                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1410                         XFS_BUF_ORDERED(bp);
1411                 dptr = XFS_BUF_PTR(bp);
1412                 /*
1413                  * Bump the cycle numbers at the start of each block
1414                  * since this part of the buffer is at the start of
1415                  * a new cycle.  Watch out for the header magic number
1416                  * case, though.
1417                  */
1418                 for (i = 0; i < split; i += BBSIZE) {
1419                         be32_add_cpu((__be32 *)dptr, 1);
1420                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1421                                 be32_add_cpu((__be32 *)dptr, 1);
1422                         dptr += BBSIZE;
1423                 }
1424
1425                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1426                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1427
1428                 /* account for internal log which doesn't start at block #0 */
1429                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1430                 XFS_BUF_WRITE(bp);
1431                 if ((error = xlog_bdstrat(bp))) {
1432                         xfs_ioerror_alert("xlog_sync (split)", log->l_mp,
1433                                           bp, XFS_BUF_ADDR(bp));
1434                         return error;
1435                 }
1436         }
1437         return 0;
1438 }       /* xlog_sync */
1439
1440
1441 /*
1442  * Deallocate a log structure
1443  */
1444 STATIC void
1445 xlog_dealloc_log(xlog_t *log)
1446 {
1447         xlog_in_core_t  *iclog, *next_iclog;
1448         int             i;
1449
1450         xlog_cil_destroy(log);
1451
1452         iclog = log->l_iclog;
1453         for (i=0; i<log->l_iclog_bufs; i++) {
1454                 xfs_buf_free(iclog->ic_bp);
1455                 next_iclog = iclog->ic_next;
1456                 kmem_free(iclog);
1457                 iclog = next_iclog;
1458         }
1459         spinlock_destroy(&log->l_icloglock);
1460
1461         xfs_buf_free(log->l_xbuf);
1462         log->l_mp->m_log = NULL;
1463         kmem_free(log);
1464 }       /* xlog_dealloc_log */
1465
1466 /*
1467  * Update counters atomically now that memcpy is done.
1468  */
1469 /* ARGSUSED */
1470 static inline void
1471 xlog_state_finish_copy(xlog_t           *log,
1472                        xlog_in_core_t   *iclog,
1473                        int              record_cnt,
1474                        int              copy_bytes)
1475 {
1476         spin_lock(&log->l_icloglock);
1477
1478         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1479         iclog->ic_offset += copy_bytes;
1480
1481         spin_unlock(&log->l_icloglock);
1482 }       /* xlog_state_finish_copy */
1483
1484
1485
1486
1487 /*
1488  * print out info relating to regions written which consume
1489  * the reservation
1490  */
1491 void
1492 xlog_print_tic_res(
1493         struct xfs_mount        *mp,
1494         struct xlog_ticket      *ticket)
1495 {
1496         uint i;
1497         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1498
1499         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1500         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1501             "bformat",
1502             "bchunk",
1503             "efi_format",
1504             "efd_format",
1505             "iformat",
1506             "icore",
1507             "iext",
1508             "ibroot",
1509             "ilocal",
1510             "iattr_ext",
1511             "iattr_broot",
1512             "iattr_local",
1513             "qformat",
1514             "dquot",
1515             "quotaoff",
1516             "LR header",
1517             "unmount",
1518             "commit",
1519             "trans header"
1520         };
1521         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1522             "SETATTR_NOT_SIZE",
1523             "SETATTR_SIZE",
1524             "INACTIVE",
1525             "CREATE",
1526             "CREATE_TRUNC",
1527             "TRUNCATE_FILE",
1528             "REMOVE",
1529             "LINK",
1530             "RENAME",
1531             "MKDIR",
1532             "RMDIR",
1533             "SYMLINK",
1534             "SET_DMATTRS",
1535             "GROWFS",
1536             "STRAT_WRITE",
1537             "DIOSTRAT",
1538             "WRITE_SYNC",
1539             "WRITEID",
1540             "ADDAFORK",
1541             "ATTRINVAL",
1542             "ATRUNCATE",
1543             "ATTR_SET",
1544             "ATTR_RM",
1545             "ATTR_FLAG",
1546             "CLEAR_AGI_BUCKET",
1547             "QM_SBCHANGE",
1548             "DUMMY1",
1549             "DUMMY2",
1550             "QM_QUOTAOFF",
1551             "QM_DQALLOC",
1552             "QM_SETQLIM",
1553             "QM_DQCLUSTER",
1554             "QM_QINOCREATE",
1555             "QM_QUOTAOFF_END",
1556             "SB_UNIT",
1557             "FSYNC_TS",
1558             "GROWFSRT_ALLOC",
1559             "GROWFSRT_ZERO",
1560             "GROWFSRT_FREE",
1561             "SWAPEXT"
1562         };
1563
1564         xfs_warn(mp,
1565                 "xfs_log_write: reservation summary:\n"
1566                 "  trans type  = %s (%u)\n"
1567                 "  unit res    = %d bytes\n"
1568                 "  current res = %d bytes\n"
1569                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
1570                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
1571                 "  ophdr + reg = %u bytes\n"
1572                 "  num regions = %u\n",
1573                 ((ticket->t_trans_type <= 0 ||
1574                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1575                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1576                 ticket->t_trans_type,
1577                 ticket->t_unit_res,
1578                 ticket->t_curr_res,
1579                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1580                 ticket->t_res_num_ophdrs, ophdr_spc,
1581                 ticket->t_res_arr_sum +
1582                 ticket->t_res_o_flow + ophdr_spc,
1583                 ticket->t_res_num);
1584
1585         for (i = 0; i < ticket->t_res_num; i++) {
1586                 uint r_type = ticket->t_res_arr[i].r_type;
1587                 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1588                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1589                             "bad-rtype" : res_type_str[r_type-1]),
1590                             ticket->t_res_arr[i].r_len);
1591         }
1592
1593         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1594                 "xfs_log_write: reservation ran out. Need to up reservation");
1595         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1596 }
1597
1598 /*
1599  * Calculate the potential space needed by the log vector.  Each region gets
1600  * its own xlog_op_header_t and may need to be double word aligned.
1601  */
1602 static int
1603 xlog_write_calc_vec_length(
1604         struct xlog_ticket      *ticket,
1605         struct xfs_log_vec      *log_vector)
1606 {
1607         struct xfs_log_vec      *lv;
1608         int                     headers = 0;
1609         int                     len = 0;
1610         int                     i;
1611
1612         /* acct for start rec of xact */
1613         if (ticket->t_flags & XLOG_TIC_INITED)
1614                 headers++;
1615
1616         for (lv = log_vector; lv; lv = lv->lv_next) {
1617                 headers += lv->lv_niovecs;
1618
1619                 for (i = 0; i < lv->lv_niovecs; i++) {
1620                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
1621
1622                         len += vecp->i_len;
1623                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1624                 }
1625         }
1626
1627         ticket->t_res_num_ophdrs += headers;
1628         len += headers * sizeof(struct xlog_op_header);
1629
1630         return len;
1631 }
1632
1633 /*
1634  * If first write for transaction, insert start record  We can't be trying to
1635  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1636  */
1637 static int
1638 xlog_write_start_rec(
1639         struct xlog_op_header   *ophdr,
1640         struct xlog_ticket      *ticket)
1641 {
1642         if (!(ticket->t_flags & XLOG_TIC_INITED))
1643                 return 0;
1644
1645         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
1646         ophdr->oh_clientid = ticket->t_clientid;
1647         ophdr->oh_len = 0;
1648         ophdr->oh_flags = XLOG_START_TRANS;
1649         ophdr->oh_res2 = 0;
1650
1651         ticket->t_flags &= ~XLOG_TIC_INITED;
1652
1653         return sizeof(struct xlog_op_header);
1654 }
1655
1656 static xlog_op_header_t *
1657 xlog_write_setup_ophdr(
1658         struct log              *log,
1659         struct xlog_op_header   *ophdr,
1660         struct xlog_ticket      *ticket,
1661         uint                    flags)
1662 {
1663         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1664         ophdr->oh_clientid = ticket->t_clientid;
1665         ophdr->oh_res2 = 0;
1666
1667         /* are we copying a commit or unmount record? */
1668         ophdr->oh_flags = flags;
1669
1670         /*
1671          * We've seen logs corrupted with bad transaction client ids.  This
1672          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
1673          * and shut down the filesystem.
1674          */
1675         switch (ophdr->oh_clientid)  {
1676         case XFS_TRANSACTION:
1677         case XFS_VOLUME:
1678         case XFS_LOG:
1679                 break;
1680         default:
1681                 xfs_warn(log->l_mp,
1682                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1683                         ophdr->oh_clientid, ticket);
1684                 return NULL;
1685         }
1686
1687         return ophdr;
1688 }
1689
1690 /*
1691  * Set up the parameters of the region copy into the log. This has
1692  * to handle region write split across multiple log buffers - this
1693  * state is kept external to this function so that this code can
1694  * can be written in an obvious, self documenting manner.
1695  */
1696 static int
1697 xlog_write_setup_copy(
1698         struct xlog_ticket      *ticket,
1699         struct xlog_op_header   *ophdr,
1700         int                     space_available,
1701         int                     space_required,
1702         int                     *copy_off,
1703         int                     *copy_len,
1704         int                     *last_was_partial_copy,
1705         int                     *bytes_consumed)
1706 {
1707         int                     still_to_copy;
1708
1709         still_to_copy = space_required - *bytes_consumed;
1710         *copy_off = *bytes_consumed;
1711
1712         if (still_to_copy <= space_available) {
1713                 /* write of region completes here */
1714                 *copy_len = still_to_copy;
1715                 ophdr->oh_len = cpu_to_be32(*copy_len);
1716                 if (*last_was_partial_copy)
1717                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1718                 *last_was_partial_copy = 0;
1719                 *bytes_consumed = 0;
1720                 return 0;
1721         }
1722
1723         /* partial write of region, needs extra log op header reservation */
1724         *copy_len = space_available;
1725         ophdr->oh_len = cpu_to_be32(*copy_len);
1726         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1727         if (*last_was_partial_copy)
1728                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1729         *bytes_consumed += *copy_len;
1730         (*last_was_partial_copy)++;
1731
1732         /* account for new log op header */
1733         ticket->t_curr_res -= sizeof(struct xlog_op_header);
1734         ticket->t_res_num_ophdrs++;
1735
1736         return sizeof(struct xlog_op_header);
1737 }
1738
1739 static int
1740 xlog_write_copy_finish(
1741         struct log              *log,
1742         struct xlog_in_core     *iclog,
1743         uint                    flags,
1744         int                     *record_cnt,
1745         int                     *data_cnt,
1746         int                     *partial_copy,
1747         int                     *partial_copy_len,
1748         int                     log_offset,
1749         struct xlog_in_core     **commit_iclog)
1750 {
1751         if (*partial_copy) {
1752                 /*
1753                  * This iclog has already been marked WANT_SYNC by
1754                  * xlog_state_get_iclog_space.
1755                  */
1756                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1757                 *record_cnt = 0;
1758                 *data_cnt = 0;
1759                 return xlog_state_release_iclog(log, iclog);
1760         }
1761
1762         *partial_copy = 0;
1763         *partial_copy_len = 0;
1764
1765         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1766                 /* no more space in this iclog - push it. */
1767                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1768                 *record_cnt = 0;
1769                 *data_cnt = 0;
1770
1771                 spin_lock(&log->l_icloglock);
1772                 xlog_state_want_sync(log, iclog);
1773                 spin_unlock(&log->l_icloglock);
1774
1775                 if (!commit_iclog)
1776                         return xlog_state_release_iclog(log, iclog);
1777                 ASSERT(flags & XLOG_COMMIT_TRANS);
1778                 *commit_iclog = iclog;
1779         }
1780
1781         return 0;
1782 }
1783
1784 /*
1785  * Write some region out to in-core log
1786  *
1787  * This will be called when writing externally provided regions or when
1788  * writing out a commit record for a given transaction.
1789  *
1790  * General algorithm:
1791  *      1. Find total length of this write.  This may include adding to the
1792  *              lengths passed in.
1793  *      2. Check whether we violate the tickets reservation.
1794  *      3. While writing to this iclog
1795  *          A. Reserve as much space in this iclog as can get
1796  *          B. If this is first write, save away start lsn
1797  *          C. While writing this region:
1798  *              1. If first write of transaction, write start record
1799  *              2. Write log operation header (header per region)
1800  *              3. Find out if we can fit entire region into this iclog
1801  *              4. Potentially, verify destination memcpy ptr
1802  *              5. Memcpy (partial) region
1803  *              6. If partial copy, release iclog; otherwise, continue
1804  *                      copying more regions into current iclog
1805  *      4. Mark want sync bit (in simulation mode)
1806  *      5. Release iclog for potential flush to on-disk log.
1807  *
1808  * ERRORS:
1809  * 1.   Panic if reservation is overrun.  This should never happen since
1810  *      reservation amounts are generated internal to the filesystem.
1811  * NOTES:
1812  * 1. Tickets are single threaded data structures.
1813  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1814  *      syncing routine.  When a single log_write region needs to span
1815  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1816  *      on all log operation writes which don't contain the end of the
1817  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1818  *      operation which contains the end of the continued log_write region.
1819  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1820  *      we don't really know exactly how much space will be used.  As a result,
1821  *      we don't update ic_offset until the end when we know exactly how many
1822  *      bytes have been written out.
1823  */
1824 int
1825 xlog_write(
1826         struct log              *log,
1827         struct xfs_log_vec      *log_vector,
1828         struct xlog_ticket      *ticket,
1829         xfs_lsn_t               *start_lsn,
1830         struct xlog_in_core     **commit_iclog,
1831         uint                    flags)
1832 {
1833         struct xlog_in_core     *iclog = NULL;
1834         struct xfs_log_iovec    *vecp;
1835         struct xfs_log_vec      *lv;
1836         int                     len;
1837         int                     index;
1838         int                     partial_copy = 0;
1839         int                     partial_copy_len = 0;
1840         int                     contwr = 0;
1841         int                     record_cnt = 0;
1842         int                     data_cnt = 0;
1843         int                     error;
1844
1845         *start_lsn = 0;
1846
1847         len = xlog_write_calc_vec_length(ticket, log_vector);
1848         if (log->l_cilp) {
1849                 /*
1850                  * Region headers and bytes are already accounted for.
1851                  * We only need to take into account start records and
1852                  * split regions in this function.
1853                  */
1854                 if (ticket->t_flags & XLOG_TIC_INITED)
1855                         ticket->t_curr_res -= sizeof(xlog_op_header_t);
1856
1857                 /*
1858                  * Commit record headers need to be accounted for. These
1859                  * come in as separate writes so are easy to detect.
1860                  */
1861                 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1862                         ticket->t_curr_res -= sizeof(xlog_op_header_t);
1863         } else
1864                 ticket->t_curr_res -= len;
1865
1866         if (ticket->t_curr_res < 0)
1867                 xlog_print_tic_res(log->l_mp, ticket);
1868
1869         index = 0;
1870         lv = log_vector;
1871         vecp = lv->lv_iovecp;
1872         while (lv && index < lv->lv_niovecs) {
1873                 void            *ptr;
1874                 int             log_offset;
1875
1876                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
1877                                                    &contwr, &log_offset);
1878                 if (error)
1879                         return error;
1880
1881                 ASSERT(log_offset <= iclog->ic_size - 1);
1882                 ptr = iclog->ic_datap + log_offset;
1883
1884                 /* start_lsn is the first lsn written to. That's all we need. */
1885                 if (!*start_lsn)
1886                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
1887
1888                 /*
1889                  * This loop writes out as many regions as can fit in the amount
1890                  * of space which was allocated by xlog_state_get_iclog_space().
1891                  */
1892                 while (lv && index < lv->lv_niovecs) {
1893                         struct xfs_log_iovec    *reg = &vecp[index];
1894                         struct xlog_op_header   *ophdr;
1895                         int                     start_rec_copy;
1896                         int                     copy_len;
1897                         int                     copy_off;
1898
1899                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
1900                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
1901
1902                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
1903                         if (start_rec_copy) {
1904                                 record_cnt++;
1905                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
1906                                                    start_rec_copy);
1907                         }
1908
1909                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
1910                         if (!ophdr)
1911                                 return XFS_ERROR(EIO);
1912
1913                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
1914                                            sizeof(struct xlog_op_header));
1915
1916                         len += xlog_write_setup_copy(ticket, ophdr,
1917                                                      iclog->ic_size-log_offset,
1918                                                      reg->i_len,
1919                                                      &copy_off, &copy_len,
1920                                                      &partial_copy,
1921                                                      &partial_copy_len);
1922                         xlog_verify_dest_ptr(log, ptr);
1923
1924                         /* copy region */
1925                         ASSERT(copy_len >= 0);
1926                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
1927                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
1928
1929                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
1930                         record_cnt++;
1931                         data_cnt += contwr ? copy_len : 0;
1932
1933                         error = xlog_write_copy_finish(log, iclog, flags,
1934                                                        &record_cnt, &data_cnt,
1935                                                        &partial_copy,
1936                                                        &partial_copy_len,
1937                                                        log_offset,
1938                                                        commit_iclog);
1939                         if (error)
1940                                 return error;
1941
1942                         /*
1943                          * if we had a partial copy, we need to get more iclog
1944                          * space but we don't want to increment the region
1945                          * index because there is still more is this region to
1946                          * write.
1947                          *
1948                          * If we completed writing this region, and we flushed
1949                          * the iclog (indicated by resetting of the record
1950                          * count), then we also need to get more log space. If
1951                          * this was the last record, though, we are done and
1952                          * can just return.
1953                          */
1954                         if (partial_copy)
1955                                 break;
1956
1957                         if (++index == lv->lv_niovecs) {
1958                                 lv = lv->lv_next;
1959                                 index = 0;
1960                                 if (lv)
1961                                         vecp = lv->lv_iovecp;
1962                         }
1963                         if (record_cnt == 0) {
1964                                 if (!lv)
1965                                         return 0;
1966                                 break;
1967                         }
1968                 }
1969         }
1970
1971         ASSERT(len == 0);
1972
1973         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
1974         if (!commit_iclog)
1975                 return xlog_state_release_iclog(log, iclog);
1976
1977         ASSERT(flags & XLOG_COMMIT_TRANS);
1978         *commit_iclog = iclog;
1979         return 0;
1980 }
1981
1982
1983 /*****************************************************************************
1984  *
1985  *              State Machine functions
1986  *
1987  *****************************************************************************
1988  */
1989
1990 /* Clean iclogs starting from the head.  This ordering must be
1991  * maintained, so an iclog doesn't become ACTIVE beyond one that
1992  * is SYNCING.  This is also required to maintain the notion that we use
1993  * a ordered wait queue to hold off would be writers to the log when every
1994  * iclog is trying to sync to disk.
1995  *
1996  * State Change: DIRTY -> ACTIVE
1997  */
1998 STATIC void
1999 xlog_state_clean_log(xlog_t *log)
2000 {
2001         xlog_in_core_t  *iclog;
2002         int changed = 0;
2003
2004         iclog = log->l_iclog;
2005         do {
2006                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2007                         iclog->ic_state = XLOG_STATE_ACTIVE;
2008                         iclog->ic_offset       = 0;
2009                         ASSERT(iclog->ic_callback == NULL);
2010                         /*
2011                          * If the number of ops in this iclog indicate it just
2012                          * contains the dummy transaction, we can
2013                          * change state into IDLE (the second time around).
2014                          * Otherwise we should change the state into
2015                          * NEED a dummy.
2016                          * We don't need to cover the dummy.
2017                          */
2018                         if (!changed &&
2019                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2020                                         XLOG_COVER_OPS)) {
2021                                 changed = 1;
2022                         } else {
2023                                 /*
2024                                  * We have two dirty iclogs so start over
2025                                  * This could also be num of ops indicates
2026                                  * this is not the dummy going out.
2027                                  */
2028                                 changed = 2;
2029                         }
2030                         iclog->ic_header.h_num_logops = 0;
2031                         memset(iclog->ic_header.h_cycle_data, 0,
2032                               sizeof(iclog->ic_header.h_cycle_data));
2033                         iclog->ic_header.h_lsn = 0;
2034                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2035                         /* do nothing */;
2036                 else
2037                         break;  /* stop cleaning */
2038                 iclog = iclog->ic_next;
2039         } while (iclog != log->l_iclog);
2040
2041         /* log is locked when we are called */
2042         /*
2043          * Change state for the dummy log recording.
2044          * We usually go to NEED. But we go to NEED2 if the changed indicates
2045          * we are done writing the dummy record.
2046          * If we are done with the second dummy recored (DONE2), then
2047          * we go to IDLE.
2048          */
2049         if (changed) {
2050                 switch (log->l_covered_state) {
2051                 case XLOG_STATE_COVER_IDLE:
2052                 case XLOG_STATE_COVER_NEED:
2053                 case XLOG_STATE_COVER_NEED2:
2054                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2055                         break;
2056
2057                 case XLOG_STATE_COVER_DONE:
2058                         if (changed == 1)
2059                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2060                         else
2061                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2062                         break;
2063
2064                 case XLOG_STATE_COVER_DONE2:
2065                         if (changed == 1)
2066                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2067                         else
2068                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2069                         break;
2070
2071                 default:
2072                         ASSERT(0);
2073                 }
2074         }
2075 }       /* xlog_state_clean_log */
2076
2077 STATIC xfs_lsn_t
2078 xlog_get_lowest_lsn(
2079         xlog_t          *log)
2080 {
2081         xlog_in_core_t  *lsn_log;
2082         xfs_lsn_t       lowest_lsn, lsn;
2083
2084         lsn_log = log->l_iclog;
2085         lowest_lsn = 0;
2086         do {
2087             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2088                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2089                 if ((lsn && !lowest_lsn) ||
2090                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2091                         lowest_lsn = lsn;
2092                 }
2093             }
2094             lsn_log = lsn_log->ic_next;
2095         } while (lsn_log != log->l_iclog);
2096         return lowest_lsn;
2097 }
2098
2099
2100 STATIC void
2101 xlog_state_do_callback(
2102         xlog_t          *log,
2103         int             aborted,
2104         xlog_in_core_t  *ciclog)
2105 {
2106         xlog_in_core_t     *iclog;
2107         xlog_in_core_t     *first_iclog;        /* used to know when we've
2108                                                  * processed all iclogs once */
2109         xfs_log_callback_t *cb, *cb_next;
2110         int                flushcnt = 0;
2111         xfs_lsn_t          lowest_lsn;
2112         int                ioerrors;    /* counter: iclogs with errors */
2113         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2114         int                funcdidcallbacks; /* flag: function did callbacks */
2115         int                repeats;     /* for issuing console warnings if
2116                                          * looping too many times */
2117         int                wake = 0;
2118
2119         spin_lock(&log->l_icloglock);
2120         first_iclog = iclog = log->l_iclog;
2121         ioerrors = 0;
2122         funcdidcallbacks = 0;
2123         repeats = 0;
2124
2125         do {
2126                 /*
2127                  * Scan all iclogs starting with the one pointed to by the
2128                  * log.  Reset this starting point each time the log is
2129                  * unlocked (during callbacks).
2130                  *
2131                  * Keep looping through iclogs until one full pass is made
2132                  * without running any callbacks.
2133                  */
2134                 first_iclog = log->l_iclog;
2135                 iclog = log->l_iclog;
2136                 loopdidcallbacks = 0;
2137                 repeats++;
2138
2139                 do {
2140
2141                         /* skip all iclogs in the ACTIVE & DIRTY states */
2142                         if (iclog->ic_state &
2143                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2144                                 iclog = iclog->ic_next;
2145                                 continue;
2146                         }
2147
2148                         /*
2149                          * Between marking a filesystem SHUTDOWN and stopping
2150                          * the log, we do flush all iclogs to disk (if there
2151                          * wasn't a log I/O error). So, we do want things to
2152                          * go smoothly in case of just a SHUTDOWN  w/o a
2153                          * LOG_IO_ERROR.
2154                          */
2155                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2156                                 /*
2157                                  * Can only perform callbacks in order.  Since
2158                                  * this iclog is not in the DONE_SYNC/
2159                                  * DO_CALLBACK state, we skip the rest and
2160                                  * just try to clean up.  If we set our iclog
2161                                  * to DO_CALLBACK, we will not process it when
2162                                  * we retry since a previous iclog is in the
2163                                  * CALLBACK and the state cannot change since
2164                                  * we are holding the l_icloglock.
2165                                  */
2166                                 if (!(iclog->ic_state &
2167                                         (XLOG_STATE_DONE_SYNC |
2168                                                  XLOG_STATE_DO_CALLBACK))) {
2169                                         if (ciclog && (ciclog->ic_state ==
2170                                                         XLOG_STATE_DONE_SYNC)) {
2171                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2172                                         }
2173                                         break;
2174                                 }
2175                                 /*
2176                                  * We now have an iclog that is in either the
2177                                  * DO_CALLBACK or DONE_SYNC states. The other
2178                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2179                                  * caught by the above if and are going to
2180                                  * clean (i.e. we aren't doing their callbacks)
2181                                  * see the above if.
2182                                  */
2183
2184                                 /*
2185                                  * We will do one more check here to see if we
2186                                  * have chased our tail around.
2187                                  */
2188
2189                                 lowest_lsn = xlog_get_lowest_lsn(log);
2190                                 if (lowest_lsn &&
2191                                     XFS_LSN_CMP(lowest_lsn,
2192                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2193                                         iclog = iclog->ic_next;
2194                                         continue; /* Leave this iclog for
2195                                                    * another thread */
2196                                 }
2197
2198                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2199
2200
2201                                 /*
2202                                  * update the last_sync_lsn before we drop the
2203                                  * icloglock to ensure we are the only one that
2204                                  * can update it.
2205                                  */
2206                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2207                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2208                                 atomic64_set(&log->l_last_sync_lsn,
2209                                         be64_to_cpu(iclog->ic_header.h_lsn));
2210
2211                         } else
2212                                 ioerrors++;
2213
2214                         spin_unlock(&log->l_icloglock);
2215
2216                         /*
2217                          * Keep processing entries in the callback list until
2218                          * we come around and it is empty.  We need to
2219                          * atomically see that the list is empty and change the
2220                          * state to DIRTY so that we don't miss any more
2221                          * callbacks being added.
2222                          */
2223                         spin_lock(&iclog->ic_callback_lock);
2224                         cb = iclog->ic_callback;
2225                         while (cb) {
2226                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2227                                 iclog->ic_callback = NULL;
2228                                 spin_unlock(&iclog->ic_callback_lock);
2229
2230                                 /* perform callbacks in the order given */
2231                                 for (; cb; cb = cb_next) {
2232                                         cb_next = cb->cb_next;
2233                                         cb->cb_func(cb->cb_arg, aborted);
2234                                 }
2235                                 spin_lock(&iclog->ic_callback_lock);
2236                                 cb = iclog->ic_callback;
2237                         }
2238
2239                         loopdidcallbacks++;
2240                         funcdidcallbacks++;
2241
2242                         spin_lock(&log->l_icloglock);
2243                         ASSERT(iclog->ic_callback == NULL);
2244                         spin_unlock(&iclog->ic_callback_lock);
2245                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2246                                 iclog->ic_state = XLOG_STATE_DIRTY;
2247
2248                         /*
2249                          * Transition from DIRTY to ACTIVE if applicable.
2250                          * NOP if STATE_IOERROR.
2251                          */
2252                         xlog_state_clean_log(log);
2253
2254                         /* wake up threads waiting in xfs_log_force() */
2255                         wake_up_all(&iclog->ic_force_wait);
2256
2257                         iclog = iclog->ic_next;
2258                 } while (first_iclog != iclog);
2259
2260                 if (repeats > 5000) {
2261                         flushcnt += repeats;
2262                         repeats = 0;
2263                         xfs_warn(log->l_mp,
2264                                 "%s: possible infinite loop (%d iterations)",
2265                                 __func__, flushcnt);
2266                 }
2267         } while (!ioerrors && loopdidcallbacks);
2268
2269         /*
2270          * make one last gasp attempt to see if iclogs are being left in
2271          * limbo..
2272          */
2273 #ifdef DEBUG
2274         if (funcdidcallbacks) {
2275                 first_iclog = iclog = log->l_iclog;
2276                 do {
2277                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2278                         /*
2279                          * Terminate the loop if iclogs are found in states
2280                          * which will cause other threads to clean up iclogs.
2281                          *
2282                          * SYNCING - i/o completion will go through logs
2283                          * DONE_SYNC - interrupt thread should be waiting for
2284                          *              l_icloglock
2285                          * IOERROR - give up hope all ye who enter here
2286                          */
2287                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2288                             iclog->ic_state == XLOG_STATE_SYNCING ||
2289                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2290                             iclog->ic_state == XLOG_STATE_IOERROR )
2291                                 break;
2292                         iclog = iclog->ic_next;
2293                 } while (first_iclog != iclog);
2294         }
2295 #endif
2296
2297         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2298                 wake = 1;
2299         spin_unlock(&log->l_icloglock);
2300
2301         if (wake)
2302                 wake_up_all(&log->l_flush_wait);
2303 }
2304
2305
2306 /*
2307  * Finish transitioning this iclog to the dirty state.
2308  *
2309  * Make sure that we completely execute this routine only when this is
2310  * the last call to the iclog.  There is a good chance that iclog flushes,
2311  * when we reach the end of the physical log, get turned into 2 separate
2312  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2313  * routine.  By using the reference count bwritecnt, we guarantee that only
2314  * the second completion goes through.
2315  *
2316  * Callbacks could take time, so they are done outside the scope of the
2317  * global state machine log lock.
2318  */
2319 STATIC void
2320 xlog_state_done_syncing(
2321         xlog_in_core_t  *iclog,
2322         int             aborted)
2323 {
2324         xlog_t             *log = iclog->ic_log;
2325
2326         spin_lock(&log->l_icloglock);
2327
2328         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2329                iclog->ic_state == XLOG_STATE_IOERROR);
2330         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2331         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2332
2333
2334         /*
2335          * If we got an error, either on the first buffer, or in the case of
2336          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2337          * and none should ever be attempted to be written to disk
2338          * again.
2339          */
2340         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2341                 if (--iclog->ic_bwritecnt == 1) {
2342                         spin_unlock(&log->l_icloglock);
2343                         return;
2344                 }
2345                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2346         }
2347
2348         /*
2349          * Someone could be sleeping prior to writing out the next
2350          * iclog buffer, we wake them all, one will get to do the
2351          * I/O, the others get to wait for the result.
2352          */
2353         wake_up_all(&iclog->ic_write_wait);
2354         spin_unlock(&log->l_icloglock);
2355         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2356 }       /* xlog_state_done_syncing */
2357
2358
2359 /*
2360  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2361  * sleep.  We wait on the flush queue on the head iclog as that should be
2362  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2363  * we will wait here and all new writes will sleep until a sync completes.
2364  *
2365  * The in-core logs are used in a circular fashion. They are not used
2366  * out-of-order even when an iclog past the head is free.
2367  *
2368  * return:
2369  *      * log_offset where xlog_write() can start writing into the in-core
2370  *              log's data space.
2371  *      * in-core log pointer to which xlog_write() should write.
2372  *      * boolean indicating this is a continued write to an in-core log.
2373  *              If this is the last write, then the in-core log's offset field
2374  *              needs to be incremented, depending on the amount of data which
2375  *              is copied.
2376  */
2377 STATIC int
2378 xlog_state_get_iclog_space(xlog_t         *log,
2379                            int            len,
2380                            xlog_in_core_t **iclogp,
2381                            xlog_ticket_t  *ticket,
2382                            int            *continued_write,
2383                            int            *logoffsetp)
2384 {
2385         int               log_offset;
2386         xlog_rec_header_t *head;
2387         xlog_in_core_t    *iclog;
2388         int               error;
2389
2390 restart:
2391         spin_lock(&log->l_icloglock);
2392         if (XLOG_FORCED_SHUTDOWN(log)) {
2393                 spin_unlock(&log->l_icloglock);
2394                 return XFS_ERROR(EIO);
2395         }
2396
2397         iclog = log->l_iclog;
2398         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2399                 XFS_STATS_INC(xs_log_noiclogs);
2400
2401                 /* Wait for log writes to have flushed */
2402                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2403                 goto restart;
2404         }
2405
2406         head = &iclog->ic_header;
2407
2408         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2409         log_offset = iclog->ic_offset;
2410
2411         /* On the 1st write to an iclog, figure out lsn.  This works
2412          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2413          * committing to.  If the offset is set, that's how many blocks
2414          * must be written.
2415          */
2416         if (log_offset == 0) {
2417                 ticket->t_curr_res -= log->l_iclog_hsize;
2418                 xlog_tic_add_region(ticket,
2419                                     log->l_iclog_hsize,
2420                                     XLOG_REG_TYPE_LRHEADER);
2421                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2422                 head->h_lsn = cpu_to_be64(
2423                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2424                 ASSERT(log->l_curr_block >= 0);
2425         }
2426
2427         /* If there is enough room to write everything, then do it.  Otherwise,
2428          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2429          * bit is on, so this will get flushed out.  Don't update ic_offset
2430          * until you know exactly how many bytes get copied.  Therefore, wait
2431          * until later to update ic_offset.
2432          *
2433          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2434          * can fit into remaining data section.
2435          */
2436         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2437                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2438
2439                 /*
2440                  * If I'm the only one writing to this iclog, sync it to disk.
2441                  * We need to do an atomic compare and decrement here to avoid
2442                  * racing with concurrent atomic_dec_and_lock() calls in
2443                  * xlog_state_release_iclog() when there is more than one
2444                  * reference to the iclog.
2445                  */
2446                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2447                         /* we are the only one */
2448                         spin_unlock(&log->l_icloglock);
2449                         error = xlog_state_release_iclog(log, iclog);
2450                         if (error)
2451                                 return error;
2452                 } else {
2453                         spin_unlock(&log->l_icloglock);
2454                 }
2455                 goto restart;
2456         }
2457
2458         /* Do we have enough room to write the full amount in the remainder
2459          * of this iclog?  Or must we continue a write on the next iclog and
2460          * mark this iclog as completely taken?  In the case where we switch
2461          * iclogs (to mark it taken), this particular iclog will release/sync
2462          * to disk in xlog_write().
2463          */
2464         if (len <= iclog->ic_size - iclog->ic_offset) {
2465                 *continued_write = 0;
2466                 iclog->ic_offset += len;
2467         } else {
2468                 *continued_write = 1;
2469                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2470         }
2471         *iclogp = iclog;
2472
2473         ASSERT(iclog->ic_offset <= iclog->ic_size);
2474         spin_unlock(&log->l_icloglock);
2475
2476         *logoffsetp = log_offset;
2477         return 0;
2478 }       /* xlog_state_get_iclog_space */
2479
2480 /*
2481  * Atomically get the log space required for a log ticket.
2482  *
2483  * Once a ticket gets put onto the reserveq, it will only return after
2484  * the needed reservation is satisfied.
2485  *
2486  * This function is structured so that it has a lock free fast path. This is
2487  * necessary because every new transaction reservation will come through this
2488  * path. Hence any lock will be globally hot if we take it unconditionally on
2489  * every pass.
2490  *
2491  * As tickets are only ever moved on and off the reserveq under the
2492  * l_grant_reserve_lock, we only need to take that lock if we are going
2493  * to add the ticket to the queue and sleep. We can avoid taking the lock if the
2494  * ticket was never added to the reserveq because the t_queue list head will be
2495  * empty and we hold the only reference to it so it can safely be checked
2496  * unlocked.
2497  */
2498 STATIC int
2499 xlog_grant_log_space(xlog_t        *log,
2500                      xlog_ticket_t *tic)
2501 {
2502         int              free_bytes;
2503         int              need_bytes;
2504
2505 #ifdef DEBUG
2506         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2507                 panic("grant Recovery problem");
2508 #endif
2509
2510         trace_xfs_log_grant_enter(log, tic);
2511
2512         need_bytes = tic->t_unit_res;
2513         if (tic->t_flags & XFS_LOG_PERM_RESERV)
2514                 need_bytes *= tic->t_ocnt;
2515
2516         /* something is already sleeping; insert new transaction at end */
2517         if (!list_empty_careful(&log->l_reserveq)) {
2518                 spin_lock(&log->l_grant_reserve_lock);
2519                 /* recheck the queue now we are locked */
2520                 if (list_empty(&log->l_reserveq)) {
2521                         spin_unlock(&log->l_grant_reserve_lock);
2522                         goto redo;
2523                 }
2524                 list_add_tail(&tic->t_queue, &log->l_reserveq);
2525
2526                 trace_xfs_log_grant_sleep1(log, tic);
2527
2528                 /*
2529                  * Gotta check this before going to sleep, while we're
2530                  * holding the grant lock.
2531                  */
2532                 if (XLOG_FORCED_SHUTDOWN(log))
2533                         goto error_return;
2534
2535                 XFS_STATS_INC(xs_sleep_logspace);
2536                 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
2537
2538                 /*
2539                  * If we got an error, and the filesystem is shutting down,
2540                  * we'll catch it down below. So just continue...
2541                  */
2542                 trace_xfs_log_grant_wake1(log, tic);
2543         }
2544
2545 redo:
2546         if (XLOG_FORCED_SHUTDOWN(log))
2547                 goto error_return_unlocked;
2548
2549         free_bytes = xlog_space_left(log, &log->l_grant_reserve_head);
2550         if (free_bytes < need_bytes) {
2551                 spin_lock(&log->l_grant_reserve_lock);
2552                 if (list_empty(&tic->t_queue))
2553                         list_add_tail(&tic->t_queue, &log->l_reserveq);
2554
2555                 trace_xfs_log_grant_sleep2(log, tic);
2556
2557                 if (XLOG_FORCED_SHUTDOWN(log))
2558                         goto error_return;
2559
2560                 xlog_grant_push_ail(log, need_bytes);
2561
2562                 XFS_STATS_INC(xs_sleep_logspace);
2563                 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock);
2564
2565                 trace_xfs_log_grant_wake2(log, tic);
2566                 goto redo;
2567         }
2568
2569         if (!list_empty(&tic->t_queue)) {
2570                 spin_lock(&log->l_grant_reserve_lock);
2571                 list_del_init(&tic->t_queue);
2572                 spin_unlock(&log->l_grant_reserve_lock);
2573         }
2574
2575         /* we've got enough space */
2576         xlog_grant_add_space(log, &log->l_grant_reserve_head, need_bytes);
2577         xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2578         trace_xfs_log_grant_exit(log, tic);
2579         xlog_verify_grant_tail(log);
2580         return 0;
2581
2582 error_return_unlocked:
2583         spin_lock(&log->l_grant_reserve_lock);
2584 error_return:
2585         list_del_init(&tic->t_queue);
2586         spin_unlock(&log->l_grant_reserve_lock);
2587         trace_xfs_log_grant_error(log, tic);
2588
2589         /*
2590          * If we are failing, make sure the ticket doesn't have any
2591          * current reservations. We don't want to add this back when
2592          * the ticket/transaction gets cancelled.
2593          */
2594         tic->t_curr_res = 0;
2595         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2596         return XFS_ERROR(EIO);
2597 }       /* xlog_grant_log_space */
2598
2599
2600 /*
2601  * Replenish the byte reservation required by moving the grant write head.
2602  *
2603  * Similar to xlog_grant_log_space, the function is structured to have a lock
2604  * free fast path.
2605  */
2606 STATIC int
2607 xlog_regrant_write_log_space(xlog_t        *log,
2608                              xlog_ticket_t *tic)
2609 {
2610         int             free_bytes, need_bytes;
2611
2612         tic->t_curr_res = tic->t_unit_res;
2613         xlog_tic_reset_res(tic);
2614
2615         if (tic->t_cnt > 0)
2616                 return 0;
2617
2618 #ifdef DEBUG
2619         if (log->l_flags & XLOG_ACTIVE_RECOVERY)
2620                 panic("regrant Recovery problem");
2621 #endif
2622
2623         trace_xfs_log_regrant_write_enter(log, tic);
2624         if (XLOG_FORCED_SHUTDOWN(log))
2625                 goto error_return_unlocked;
2626
2627         /* If there are other waiters on the queue then give them a
2628          * chance at logspace before us. Wake up the first waiters,
2629          * if we do not wake up all the waiters then go to sleep waiting
2630          * for more free space, otherwise try to get some space for
2631          * this transaction.
2632          */
2633         need_bytes = tic->t_unit_res;
2634         if (!list_empty_careful(&log->l_writeq)) {
2635                 struct xlog_ticket *ntic;
2636
2637                 spin_lock(&log->l_grant_write_lock);
2638                 free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2639                 list_for_each_entry(ntic, &log->l_writeq, t_queue) {
2640                         ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV);
2641
2642                         if (free_bytes < ntic->t_unit_res)
2643                                 break;
2644                         free_bytes -= ntic->t_unit_res;
2645                         wake_up(&ntic->t_wait);
2646                 }
2647
2648                 if (ntic != list_first_entry(&log->l_writeq,
2649                                                 struct xlog_ticket, t_queue)) {
2650                         if (list_empty(&tic->t_queue))
2651                                 list_add_tail(&tic->t_queue, &log->l_writeq);
2652                         trace_xfs_log_regrant_write_sleep1(log, tic);
2653
2654                         xlog_grant_push_ail(log, need_bytes);
2655
2656                         XFS_STATS_INC(xs_sleep_logspace);
2657                         xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
2658                         trace_xfs_log_regrant_write_wake1(log, tic);
2659                 } else
2660                         spin_unlock(&log->l_grant_write_lock);
2661         }
2662
2663 redo:
2664         if (XLOG_FORCED_SHUTDOWN(log))
2665                 goto error_return_unlocked;
2666
2667         free_bytes = xlog_space_left(log, &log->l_grant_write_head);
2668         if (free_bytes < need_bytes) {
2669                 spin_lock(&log->l_grant_write_lock);
2670                 if (list_empty(&tic->t_queue))
2671                         list_add_tail(&tic->t_queue, &log->l_writeq);
2672
2673                 if (XLOG_FORCED_SHUTDOWN(log))
2674                         goto error_return;
2675
2676                 xlog_grant_push_ail(log, need_bytes);
2677
2678                 XFS_STATS_INC(xs_sleep_logspace);
2679                 trace_xfs_log_regrant_write_sleep2(log, tic);
2680                 xlog_wait(&tic->t_wait, &log->l_grant_write_lock);
2681
2682                 trace_xfs_log_regrant_write_wake2(log, tic);
2683                 goto redo;
2684         }
2685
2686         if (!list_empty(&tic->t_queue)) {
2687                 spin_lock(&log->l_grant_write_lock);
2688                 list_del_init(&tic->t_queue);
2689                 spin_unlock(&log->l_grant_write_lock);
2690         }
2691
2692         /* we've got enough space */
2693         xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes);
2694         trace_xfs_log_regrant_write_exit(log, tic);
2695         xlog_verify_grant_tail(log);
2696         return 0;
2697
2698
2699  error_return_unlocked:
2700         spin_lock(&log->l_grant_write_lock);
2701  error_return:
2702         list_del_init(&tic->t_queue);
2703         spin_unlock(&log->l_grant_write_lock);
2704         trace_xfs_log_regrant_write_error(log, tic);
2705
2706         /*
2707          * If we are failing, make sure the ticket doesn't have any
2708          * current reservations. We don't want to add this back when
2709          * the ticket/transaction gets cancelled.
2710          */
2711         tic->t_curr_res = 0;
2712         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
2713         return XFS_ERROR(EIO);
2714 }       /* xlog_regrant_write_log_space */
2715
2716
2717 /* The first cnt-1 times through here we don't need to
2718  * move the grant write head because the permanent
2719  * reservation has reserved cnt times the unit amount.
2720  * Release part of current permanent unit reservation and
2721  * reset current reservation to be one units worth.  Also
2722  * move grant reservation head forward.
2723  */
2724 STATIC void
2725 xlog_regrant_reserve_log_space(xlog_t        *log,
2726                                xlog_ticket_t *ticket)
2727 {
2728         trace_xfs_log_regrant_reserve_enter(log, ticket);
2729
2730         if (ticket->t_cnt > 0)
2731                 ticket->t_cnt--;
2732
2733         xlog_grant_sub_space(log, &log->l_grant_reserve_head,
2734                                         ticket->t_curr_res);
2735         xlog_grant_sub_space(log, &log->l_grant_write_head,
2736                                         ticket->t_curr_res);
2737         ticket->t_curr_res = ticket->t_unit_res;
2738         xlog_tic_reset_res(ticket);
2739
2740         trace_xfs_log_regrant_reserve_sub(log, ticket);
2741
2742         /* just return if we still have some of the pre-reserved space */
2743         if (ticket->t_cnt > 0)
2744                 return;
2745
2746         xlog_grant_add_space(log, &log->l_grant_reserve_head,
2747                                         ticket->t_unit_res);
2748
2749         trace_xfs_log_regrant_reserve_exit(log, ticket);
2750
2751         ticket->t_curr_res = ticket->t_unit_res;
2752         xlog_tic_reset_res(ticket);
2753 }       /* xlog_regrant_reserve_log_space */
2754
2755
2756 /*
2757  * Give back the space left from a reservation.
2758  *
2759  * All the information we need to make a correct determination of space left
2760  * is present.  For non-permanent reservations, things are quite easy.  The
2761  * count should have been decremented to zero.  We only need to deal with the
2762  * space remaining in the current reservation part of the ticket.  If the
2763  * ticket contains a permanent reservation, there may be left over space which
2764  * needs to be released.  A count of N means that N-1 refills of the current
2765  * reservation can be done before we need to ask for more space.  The first
2766  * one goes to fill up the first current reservation.  Once we run out of
2767  * space, the count will stay at zero and the only space remaining will be
2768  * in the current reservation field.
2769  */
2770 STATIC void
2771 xlog_ungrant_log_space(xlog_t        *log,
2772                        xlog_ticket_t *ticket)
2773 {
2774         int     bytes;
2775
2776         if (ticket->t_cnt > 0)
2777                 ticket->t_cnt--;
2778
2779         trace_xfs_log_ungrant_enter(log, ticket);
2780         trace_xfs_log_ungrant_sub(log, ticket);
2781
2782         /*
2783          * If this is a permanent reservation ticket, we may be able to free
2784          * up more space based on the remaining count.
2785          */
2786         bytes = ticket->t_curr_res;
2787         if (ticket->t_cnt > 0) {
2788                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2789                 bytes += ticket->t_unit_res*ticket->t_cnt;
2790         }
2791
2792         xlog_grant_sub_space(log, &log->l_grant_reserve_head, bytes);
2793         xlog_grant_sub_space(log, &log->l_grant_write_head, bytes);
2794
2795         trace_xfs_log_ungrant_exit(log, ticket);
2796
2797         xfs_log_move_tail(log->l_mp, 1);
2798 }       /* xlog_ungrant_log_space */
2799
2800
2801 /*
2802  * Flush iclog to disk if this is the last reference to the given iclog and
2803  * the WANT_SYNC bit is set.
2804  *
2805  * When this function is entered, the iclog is not necessarily in the
2806  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2807  *
2808  *
2809  */
2810 STATIC int
2811 xlog_state_release_iclog(
2812         xlog_t          *log,
2813         xlog_in_core_t  *iclog)
2814 {
2815         int             sync = 0;       /* do we sync? */
2816
2817         if (iclog->ic_state & XLOG_STATE_IOERROR)
2818                 return XFS_ERROR(EIO);
2819
2820         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2821         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2822                 return 0;
2823
2824         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2825                 spin_unlock(&log->l_icloglock);
2826                 return XFS_ERROR(EIO);
2827         }
2828         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2829                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2830
2831         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2832                 /* update tail before writing to iclog */
2833                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2834                 sync++;
2835                 iclog->ic_state = XLOG_STATE_SYNCING;
2836                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2837                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2838                 /* cycle incremented when incrementing curr_block */
2839         }
2840         spin_unlock(&log->l_icloglock);
2841
2842         /*
2843          * We let the log lock go, so it's possible that we hit a log I/O
2844          * error or some other SHUTDOWN condition that marks the iclog
2845          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2846          * this iclog has consistent data, so we ignore IOERROR
2847          * flags after this point.
2848          */
2849         if (sync)
2850                 return xlog_sync(log, iclog);
2851         return 0;
2852 }       /* xlog_state_release_iclog */
2853
2854
2855 /*
2856  * This routine will mark the current iclog in the ring as WANT_SYNC
2857  * and move the current iclog pointer to the next iclog in the ring.
2858  * When this routine is called from xlog_state_get_iclog_space(), the
2859  * exact size of the iclog has not yet been determined.  All we know is
2860  * that every data block.  We have run out of space in this log record.
2861  */
2862 STATIC void
2863 xlog_state_switch_iclogs(xlog_t         *log,
2864                          xlog_in_core_t *iclog,
2865                          int            eventual_size)
2866 {
2867         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2868         if (!eventual_size)
2869                 eventual_size = iclog->ic_offset;
2870         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2871         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2872         log->l_prev_block = log->l_curr_block;
2873         log->l_prev_cycle = log->l_curr_cycle;
2874
2875         /* roll log?: ic_offset changed later */
2876         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2877
2878         /* Round up to next log-sunit */
2879         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2880             log->l_mp->m_sb.sb_logsunit > 1) {
2881                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2882                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2883         }
2884
2885         if (log->l_curr_block >= log->l_logBBsize) {
2886                 log->l_curr_cycle++;
2887                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2888                         log->l_curr_cycle++;
2889                 log->l_curr_block -= log->l_logBBsize;
2890                 ASSERT(log->l_curr_block >= 0);
2891         }
2892         ASSERT(iclog == log->l_iclog);
2893         log->l_iclog = iclog->ic_next;
2894 }       /* xlog_state_switch_iclogs */
2895
2896 /*
2897  * Write out all data in the in-core log as of this exact moment in time.
2898  *
2899  * Data may be written to the in-core log during this call.  However,
2900  * we don't guarantee this data will be written out.  A change from past
2901  * implementation means this routine will *not* write out zero length LRs.
2902  *
2903  * Basically, we try and perform an intelligent scan of the in-core logs.
2904  * If we determine there is no flushable data, we just return.  There is no
2905  * flushable data if:
2906  *
2907  *      1. the current iclog is active and has no data; the previous iclog
2908  *              is in the active or dirty state.
2909  *      2. the current iclog is drity, and the previous iclog is in the
2910  *              active or dirty state.
2911  *
2912  * We may sleep if:
2913  *
2914  *      1. the current iclog is not in the active nor dirty state.
2915  *      2. the current iclog dirty, and the previous iclog is not in the
2916  *              active nor dirty state.
2917  *      3. the current iclog is active, and there is another thread writing
2918  *              to this particular iclog.
2919  *      4. a) the current iclog is active and has no other writers
2920  *         b) when we return from flushing out this iclog, it is still
2921  *              not in the active nor dirty state.
2922  */
2923 int
2924 _xfs_log_force(
2925         struct xfs_mount        *mp,
2926         uint                    flags,
2927         int                     *log_flushed)
2928 {
2929         struct log              *log = mp->m_log;
2930         struct xlog_in_core     *iclog;
2931         xfs_lsn_t               lsn;
2932
2933         XFS_STATS_INC(xs_log_force);
2934
2935         if (log->l_cilp)
2936                 xlog_cil_force(log);
2937
2938         spin_lock(&log->l_icloglock);
2939
2940         iclog = log->l_iclog;
2941         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2942                 spin_unlock(&log->l_icloglock);
2943                 return XFS_ERROR(EIO);
2944         }
2945
2946         /* If the head iclog is not active nor dirty, we just attach
2947          * ourselves to the head and go to sleep.
2948          */
2949         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2950             iclog->ic_state == XLOG_STATE_DIRTY) {
2951                 /*
2952                  * If the head is dirty or (active and empty), then
2953                  * we need to look at the previous iclog.  If the previous
2954                  * iclog is active or dirty we are done.  There is nothing
2955                  * to sync out.  Otherwise, we attach ourselves to the
2956                  * previous iclog and go to sleep.
2957                  */
2958                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2959                     (atomic_read(&iclog->ic_refcnt) == 0
2960                      && iclog->ic_offset == 0)) {
2961                         iclog = iclog->ic_prev;
2962                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2963                             iclog->ic_state == XLOG_STATE_DIRTY)
2964                                 goto no_sleep;
2965                         else
2966                                 goto maybe_sleep;
2967                 } else {
2968                         if (atomic_read(&iclog->ic_refcnt) == 0) {
2969                                 /* We are the only one with access to this
2970                                  * iclog.  Flush it out now.  There should
2971                                  * be a roundoff of zero to show that someone
2972                                  * has already taken care of the roundoff from
2973                                  * the previous sync.
2974                                  */
2975                                 atomic_inc(&iclog->ic_refcnt);
2976                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2977                                 xlog_state_switch_iclogs(log, iclog, 0);
2978                                 spin_unlock(&log->l_icloglock);
2979
2980                                 if (xlog_state_release_iclog(log, iclog))
2981                                         return XFS_ERROR(EIO);
2982
2983                                 if (log_flushed)
2984                                         *log_flushed = 1;
2985                                 spin_lock(&log->l_icloglock);
2986                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2987                                     iclog->ic_state != XLOG_STATE_DIRTY)
2988                                         goto maybe_sleep;
2989                                 else
2990                                         goto no_sleep;
2991                         } else {
2992                                 /* Someone else is writing to this iclog.
2993                                  * Use its call to flush out the data.  However,
2994                                  * the other thread may not force out this LR,
2995                                  * so we mark it WANT_SYNC.
2996                                  */
2997                                 xlog_state_switch_iclogs(log, iclog, 0);
2998                                 goto maybe_sleep;
2999                         }
3000                 }
3001         }
3002
3003         /* By the time we come around again, the iclog could've been filled
3004          * which would give it another lsn.  If we have a new lsn, just
3005          * return because the relevant data has been flushed.
3006          */
3007 maybe_sleep:
3008         if (flags & XFS_LOG_SYNC) {
3009                 /*
3010                  * We must check if we're shutting down here, before
3011                  * we wait, while we're holding the l_icloglock.
3012                  * Then we check again after waking up, in case our
3013                  * sleep was disturbed by a bad news.
3014                  */
3015                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3016                         spin_unlock(&log->l_icloglock);
3017                         return XFS_ERROR(EIO);
3018                 }
3019                 XFS_STATS_INC(xs_log_force_sleep);
3020                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3021                 /*
3022                  * No need to grab the log lock here since we're
3023                  * only deciding whether or not to return EIO
3024                  * and the memory read should be atomic.
3025                  */
3026                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3027                         return XFS_ERROR(EIO);
3028                 if (log_flushed)
3029                         *log_flushed = 1;
3030         } else {
3031
3032 no_sleep:
3033                 spin_unlock(&log->l_icloglock);
3034         }
3035         return 0;
3036 }
3037
3038 /*
3039  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3040  * about errors or whether the log was flushed or not. This is the normal
3041  * interface to use when trying to unpin items or move the log forward.
3042  */
3043 void
3044 xfs_log_force(
3045         xfs_mount_t     *mp,
3046         uint            flags)
3047 {
3048         int     error;
3049
3050         error = _xfs_log_force(mp, flags, NULL);
3051         if (error)
3052                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3053 }
3054
3055 /*
3056  * Force the in-core log to disk for a specific LSN.
3057  *
3058  * Find in-core log with lsn.
3059  *      If it is in the DIRTY state, just return.
3060  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3061  *              state and go to sleep or return.
3062  *      If it is in any other state, go to sleep or return.
3063  *
3064  * Synchronous forces are implemented with a signal variable. All callers
3065  * to force a given lsn to disk will wait on a the sv attached to the
3066  * specific in-core log.  When given in-core log finally completes its
3067  * write to disk, that thread will wake up all threads waiting on the
3068  * sv.
3069  */
3070 int
3071 _xfs_log_force_lsn(
3072         struct xfs_mount        *mp,
3073         xfs_lsn_t               lsn,
3074         uint                    flags,
3075         int                     *log_flushed)
3076 {
3077         struct log              *log = mp->m_log;
3078         struct xlog_in_core     *iclog;
3079         int                     already_slept = 0;
3080
3081         ASSERT(lsn != 0);
3082
3083         XFS_STATS_INC(xs_log_force);
3084
3085         if (log->l_cilp) {
3086                 lsn = xlog_cil_force_lsn(log,&n